L e c tu re N o te s in C o m p u te r S c ie n c e
E d ite d b y G . G o o s , J . H a rtm a n is a n d J . v a n L e e u w e n
1 8 0 3
3
B e r lin
H e id e lb e rg
N e w Y o rk
B a rc e lo n a
H o n g K o n g
L o n d o n
M ila n
P a r is
S in g a p o re
T o k y o
S te f a n o C a g n o n i e t a l. ( E d s .)
R e a l-W o rld A p p lic a tio n s
o f E v o lu tio n a ry C o m p u tin g
E v o W o rk sh o p s 2 0 0 0 : E v o IA S P , E v o S C O N D I,
E v o T e l, E v o S T IM , E v o R o b , a n d E v o F lig h t
E d in b u rg h , S c o tla n d , U K , A p ril 1 7 , 2 0 0 0
P ro c e e d in g s
13
S e rie s E d ito rs
G e r h a r d G o o s , K a r ls r u h e U n iv e r s ity , G e r m a n y
J u r is H a r tm a n is , C o rn e ll U n iv e r s ity , N Y , U S A
J a n v a n L e e u w e n , U tr e c h t U n iv e r s ity , T h e N e th e r la n d s
M a in V o lu m e E d ito r
S te fa n o C a g n o n i
U n iv e r s ity o f P a r m a
D e p a rtm e n t o f C o m p u te r E n g in e e rin g
P a rc o d e lle S c ie n z e 1 8 1 /a , 4 3 1 0 0 P a rm a , Ita ly
E -m a il: c a g n o n i@ c e .u n ip r.it
C a ta lo g in g -in -P u b lic a tio n d a ta a p p lie d fo r
D ie D e u ts c h e B ib lio th e k - C IP -E in h e its a u fn a h m e
R e
E v
2 0
B a
S p
a l w o rld a p p lic a tio n s
o W o rk sh o p s 2 0 0 0 : E
0 0 . S te fa n o C a g n o n i
rc e lo n a ; H o n g K o n g
rin g e r, 2 0 0 0
(L e c tu re n o te s in c o m
IS B N 3 -5 4 0 -6 7 3 5 3 -9
o f e v o lu
v o IA S P
. . . (e d .).
; L o n d o
tio
. . .
- B
n ;
n a ry c o m
, E d in b u
e rlin ; H
M ila n ;
p u
rg h
e id
P a r
tin g : p ro c e
, S c o tla n d ,
e lb e rg ; N e
is ; S in g a p o
e d in g s
U K , A
w Y o rk
re ; T o
/
p ril 1 7 ,
;
k y o :
p u te r s c ie n c e ; V o l. 1 8 0 3 )
C R S u b j e c t C l a s s i fi c a t i o n ( 1 9 9 8 ) : C . 2 , I . 4 , F . 3 , I . 2 , G . 2 , F . 2 , J . 2 , J . 1 , D . 1
IS S N 0 3 0 2 -9 7 4 3
IS B N 3 -5 4 0 -6 7 3 5 3 -9 S p rin g e r-V e rla g B e rlin H e id e lb e rg N e w Y o rk
T h is w o rk is s u b je c
c o n c e r n e d , s p e c i fi c a
re p ro d u c tio n o n m ic
o r p a rts th e re o f is p e
in its c u rre n t v e rs io n
lia b le fo r p ro s e c u tio
t to c o p y rig h t. A ll rig h ts a re re s e rv e d , w h e th e r th e w h o le o r p a rt o f th e m a te ria l is
lly th e rig h ts o f tra n s la tio n , re p rin tin g , re -u s e o f illu s tra tio n s , re c ita tio n , b ro a d c a s tin g ,
r o fi l m s o r i n a n y o t h e r w a y , a n d s t o r a g e i n d a t a b a n k s . D u p l i c a t i o n o f t h i s p u b l i c a t i o n
rm itte d o n ly u n d e r th e p ro v is io n s o f th e G e rm a n C o p y rig h t L a w o f S e p te m b e r 9 , 1 9 6 5 ,
, a n d p e r m is s io n f o r u s e m u s t a lw a y s b e o b ta in e d f r o m S p r in g e r- V e r la g . V io la tio n s a r e
n u n d e r th e G e rm a n C o p y rig h t L a w .
S p rin g e r-V e rla g is a c o m p a n y in th e B e rte ls m a n n S p rin g e r p u b lis h in g g ro u p .
c S p rin g e r-V e rla g B e rlin H e id e lb e rg 2 0 0 0
P rin te d in G e rm a n y
T y p e s e ttin g : C a m e ra -re a d y b y a u th o r
P rin te d o n a c id -fre e p a p e r
S P IN : 1 0 7 2 0 1 7 3
0 6 /3 1 4 2
5 4 3 2 1 0
i r a si gla ti fi l
d f
lti a r
p ta ti ( ) pr id s a la lt l
s, i spir d
t t r f a t ra ls l ti a d g ti i rita
,t
pr l s li g, a i la r i g, a d pti ia ti i a
r al
- rl
d a ppl
ia ti s.
spit s
a rl it iti s a
t , t at a
da t d a t t i ti f
p t rs, a d a tt r f r a ld fi iti f , a d i t
6 s,
t q stf r r a l
- rl
d a ppl
ia ti s f
l ga i t l
at
s.
dra a ti i r a s i
p t r p rf r a
si t l
a std a d ft 2 t
t r ga ris t a p siti f d a pr ss:
t
iq s
a
r a d
r a ppl
ia l,sti l
a ti gt gr t fit r sti t ir st d ,a d a l
l i g,
i tr ,
p
rf l pa ra dig s t
d is d.
I pa ra l
ll it
t r tia lr s l
ts, t
r ffi l
ds t
i
is
i ga ppl
id is i r a si gda
da ,a l g it t
pl it fa ppl
ia ti s
a d a ppl
ia ti d a i s.I pa rti l
a r,i d stria l
lr l a tfi l
ds,s
a s sig a l
a d i a g pr ssi g,
p t r isi , pa tt r r g iti , i d stria l tr l
,
tl
ia ti , s d l
i g a d ti ta l
i g, a d a r spa
gi ri g a r
pl i g
t
iq s t s l
pl r a l
- rl
d pr l s.
is l
ta i s t pr
di gs f
r s ps 2
: si
r sps
r al
- rl
d a ppl
ia ti s f
l
d
rr tl
pril 7 t 2
i di rg .
r s ps a r :
I
2
,t s
d
rp a
r sp
lti a r
p ta ti i I a g
a lsis a d ig a l r ssi g,
I2
,t first r p a
rs p
lti a r
p ta ti
i
st s,
tr la d ri s i I d str ,
l2
,t s
d
rp a
rs p
lti a r
l
ia ti s,
ti 2
, t first r p a
rs p
lti a r
d l
i ga d i ta l
i g,
2
,t
t ird
rp a
rs p
lti a r
tis, a d
l
ig t2
,t
first r p a
rs p
lti a r
r a tis.
r s ps 2
as l
di
ti
it t
t r a r rp a
ts: r
2
,t
rp a
fr
ti r gra
i g, l
d
pril t a d 6 t , a d I
2
,t
ird I t r a ti a l fr
la l st s: fr
ilg t a rd a r , l
d fr
pril7 t t
t.
l
dl
i t ta t
rs ft it r a ti a l
pr gra
itt s
ft si
r s ps i i ld d t
r s a r rs a d r s a r rs i t
sp ifi fi l
ds fit r sta d s r d t
ig q a l
it ft pa p rs.
pril2
a
t fa
a g i, i a rd
l
i,
i,
rg
it , a id
r , a rti a t s,
a rt, ir
a a i, g rt . .
rs,
a
t r, a d r
. ga rt
r
r s
ps 2
I
I
- ha ir
- ha ir
O
I ha ir
l - ha ir
l - ha ir
l - ha ir
I ha ir
ha ir
l
ight ha ir
a lha ir
l
ia tins ha ir
r s
ps 2
r
t fa n a gn ni( ni rsit f a rma Ita l)
i ar
l
i( ni rsit f irmingha m K)
n i( ni rsit f l
a sg
K)
rg mith ( ni rsit f a st ngl
ia K)
a i
rn ( ni rsit f a ing K)
a rtin O a t s ( ritish l m pl K)
mma a rt( ni rsit f in rgh K)
ir
a a n i( il
a n lt hni Ita l)
g rt . .
rs (
h
th rl
a n s)
n a ht r ( a pir ni rsit K)
rn
. ga rt ( th a nk ni rsit K)
r r
a na gitis a mi is
hn lgia l
a tina lInstit tin f h ssa lniki
r
i a nni
rni ni rsit f a rma Ita l
a rja n a n n kk r
h
th rl
an s
l
ista ir rmita g
a pir ni rsit K
l
fga ng a n ha f ni rsit f rtm n
rma n
an al
l.
r as
st rn s r
ni rsit
n ra
na rini il
a n lt hni Ita l
a him . rn a imlr hr slr
rma n
l rt r ggi ni rsit f a ia Ita l
t fa n a gn ni ni rsit f a rma Ita l
ri l
a ha kra rt a tina l ngin ring a ra t r
K
l
a l
a ri g
ni rsit f irmingha m K
ar
lm tti il
a n lt hni Ita l
is rr ia
ni rsit f is n
rt ga l
a s n a i a ni rsit f ihiga n
ar
rig r
ni rsit f r ss l
s
l
gi m
ha ng . a n
r t
K
g st n
i n
i n ni rsit h
th rl
an s
tr
lming ni rsit f h ffi l K
a ri lr a n
it rl
an
O s a r . a r ia ni rsit f ra na a pa in
rg r¨ l
rma n
a rk r n lr ni rsit fZ a gr
r a tia
Inma n a r
ni rsit f ss
K
r
hil s a n s ni rsit f ss
K
ra n is
rr ra ni rsit f ra na a pa in
a sa ki ira f ji
mp ta tina l
l
ing a ra t r
a nil a r
K
Kn
nt ni rsit f l
a sg
K
ik a k i ni rsit f ss
K
a ri K pp n ra nh fr I K
rma n
hn K a ta nf r ni rsit
Ka
h n a n a tina l ni rsit f inga p r inga p r
n i ni rsit f l
a sg
K
ln
tt n I I
ra n
an r a
r nima t a
ra n
O ra i igl
in
ni rsit f a pls Ita l
l
ia n il
lr ni rsit f irmingha m K
t fa n
l
fi a tina l s a r h
n ilIta l
t r r in ha l
m rs ni rsit
n
n a ht r a pir ni rsit K
i ar
l
i ni rsit f irmingha m K
im mith ni rsit fth
st f ngl
an
K
a thia s l
a n n a ra l
inga m
kia K
t r a nn
l
l
s
pl K
rt mith ni rsit fth
st f ngl
an
K
Ka
h n a n a tina l ni rsit f inga p r inga p r
n r a . . tta ma n i n tia Ita l
n
rr l
l ni rsit f rk K
a ns- iha l igt
I
rma n
. nr
i rp l ni rsit K
in a
ni rsit f irmingha m K
l
iZ a la l
a
rit- a tt ni rsit K
p
sr
s
a pir ni rsit K
t th
t rk f
s
l
ln
in
ltina r
mp ting
a pa n
s
p cil rp s I g
n ltin ith
l bl rd
. m lin . . st r . . rn l n .
r
t r sc pic isin f r
n id
. . . r
. rin n
.
.
st r n tic
ng . .
ing..
2
lst ring l
g rith ...................................
n . . ng
22
c n Int rpr t tin sing
. r h
. .
kls n
34
s s in ign l p c s.............................
44
inding l
f
rs s: h l
tr igh ch ppr ch....................
. . r
. rkins . . r m
. h ilr . . rt r
. .
ng . . rgh s
. .
m nski n . . l h
4
lt
il
ts nd
tr
n tic r gr
p t tin
ltin r
. . rrir
ntic
. .
b t sing
r hl
r ..............
ltin r
s rs
nd
c l
i tin f r
n id
b tb
ns f n tic
r gr
ing............................................................
. rlss n . rin n
. r hl
6
n th c lbil
it f n tic l
g rith s t
r
rg - c l
tr
lctin .................................................................
.
sr n
. .
rt
77
bining
ltin r , nn ctinist, nd
lssific tin
l
g rith s f r h p
n lsis ...........................................
. . sin n . O .
ngs
7
p ri nt l t r in tin f r s phil
br nic
rdin t s b
n tic l
g rith s, th i pl
th d, nd h ir brid ..............
. . ir
. . im kin . init n . sm n
7
p rs
.
n tic l
g rith
. i . .
ith
n
c l rch f r
. . ng
ling
b hp
r bl s..
7
istrib t d
rning ntr lf r ffic ign l
s..........................
. .
. rs n . ll n . ils
i
ris r dictin b
. hn n . i
r
ing
tr l l
7
r l t rks......
27
r jct r
ntr l
lr t rk nd Its sign
t
tin thr gh
ltin r
p ting................................................
. h ng n . i
3
ltin r
p t tin nd
nl
in r r gr
ingin
l
ti- d l
b st ntr l sign ..................................................
. lss n . r¨ l
47
nch rking st- ssign nt ch
sfr
l
ti- b jcti
ltin r l
g rith s................................................
.
k lkis n . i
t
tic nth sis f th th
p lg nd
r
t rs f r
ntr l
lr f r
hr - g lnt ith
i - c nd
l sing
n tic r gr
ing...................................................
. .
. .
n .
.
l
n . . nn tt
6
t
tic sign f l
ti ribl
ntr l st
i
ltin r
p t tin............................................................
. . n . .
n . . hr
7
lp nt f
r r nsf r r h r l d l
s f r il
p r t r r dictin .................................................
. . ng . ng . . tt ll . i h rs n . im ns n
n . .
s
t
.
.
tic l
id tin f r t c lInt rf c s scrib d in
rn
. n
r
n . q illr
lti
ri
d l
ing f
/I
t rk r ffic f r Intr sin
......... 2
t ctin ..
2 4
l
ti d l rf r nc r fils n th d pti
istrib t d
tb s
n g nt r bl ................................................... 224
. O t s . rn n .
r
r t c l nstr ctin sing
. h r ls n .
km n
n tic
rch
chniq
s.................
23
s rs
r dictin f
. rn
.
. rs n
r
nic tin
. .
q ir
nts f r igh- p d
ng
. n
r
n
rchit ct r f r
l
ti- g nt
. ll n . ils
irc its............... 247
. i lnt
rning
st s.......
2
p rs
n
b lnc
. . . r
r
st ring st .................................. 26 7
r . in l
n . . grt
st tic In stig tin f
rf r nc n obsh p ch d l
ing
r bl s ...............................................................
. rt n . ss
27 7
n nt l
g rith
ith
hr
n
l tin
lf r t l
rdin ss r bl s.....................................................
.
rkl n
. i n rf
2 7
n tic pr s nt tin nd
n lst r r ss r f r
b h p ch d l
ing r bl s..........................................
.
k
. iji . i
shi n . O k m r
2 7
pti ising n
ltin r l
g rith f r ch d l
ing...................
. rq h rt . hish lm n .
ht r
3 7
p rs
n- in
ltin f ntr lf r
r- gg d
b t sing n tic
r gr
ing...........................................................
. n rss n .
nss n . r hl n . rin
3
pti i d
l
l
isin r
ltin r
ft r
. lm
327
l
f- d pti
. ll n .
t tin in
rst
b t
t t nt n r tin b th
.........................................
ntr l
lrs..............................
sing
brid
ltin r - * ppr ch f r
rning
cti
h irs...............................................................
. tt n . . r
33
347
ig
p rs
p r is d
. .
rl
ltin r
th ds in r d n
.
n r n . ir
ic
sign
pti i tin .
3 7
n
ltin r l
g rith f r rg c l t
ring r bl s ith
ppl
ic tin t irl
in r
ch d l
ing.................................
.
rhi ri n . t n k
36 7
sign, I pl nt tin, nd ppl
ic tin f
lf r pti l ircr ft
sitining.............................................................. 3 2
. f lgrf . rnk .
i h n rgr n . t l n rg
r
d
.................................................. 3
S p e c ia l P u r p o s e I m a g e C o n v o lu tio n w ith E v o lv a b le
H a r d w a r e
a
J o e D u m o u lin , J a m e s A . F o s te r
a
N e w L ig h t In d u s trie s , L
C e n te r fo r S e c u re a n d D
c
M ic ro e le c tro n ic s R e s e
M o sc o w ID
d
D e p t. o f C o m p u te r S c ie
e
D e p t. o f E le c tric a l a n d
b
e m a i l :
j o e @ b r e s g a l . c o m ,
b ,c ,d
c ,e
, Ja m e s F . F re n z e l , S te v e M c G re w
a
td ., S p o k a n e , W A
e p e n d a b le S o ftw a re , U . Id a h o , M o s c o w , ID
a r c h a n d C o m m u n ic a tio n s I n s t., U . I d a h o ,
n c e , U . Id a h o , M o sc o w , ID
C o m p u te r E n g in e e rin g , U . Id a h o , M o s c o w , ID
f o s t e r @ c s . u i d a h o . e d u ,
s t e v e m @ i e a . c o m
j f f @ m r c . u i d a h o . e d u ,
A b str a c t.
I n th is p a p e r , w e in v e s tig a te a u n iq u e m e th o d o f in v e n tin g lin e a r e d g e e n h a n c e m e n t
o p e ra to r s u s in g e v o lu tio n a n d re c o n fig u r a b le h a r d w a r e . W e s h o w th a t th e te c h n iq u e is
m o tiv a te d b y th e d e s ir e f o r a to ta lly a u to m a te d o b je c t re c o g n itio n s y s te m . W e s h o w th a t
a n im p o r ta n t s te p in a u to m a tin g o b je c t r e c o g n itio n is to p ro v id e fle x ib le m e a n s to s m o o th
im a g e s , m a k in g f e a tu r e s m o r e o b v io u s a n d r e d u c in g in te r fe re n c e . N e x t w e d e m o n s tra te
a te c h n iq u e fo r b u ild in g a n e d g e e n h a n c e m e n t o p e r a to r u s in g e v o lu tio n a r y m e th o d s ,
im p le m e n tin g a n d te s tin g e a c h g e n e r a tio n u s in g th e X ilin x 6 2 0 0 f a m ily F P G A . F in a lly ,
w e p r e s e n t th e r e s u lts a n d c o n c lu d e b y m e n tio n in g s o m e a r e a s o f fu rth e r in v e s tig a tio n .
I n tr o d u c tio n
Im a g e e d g e e n h a n c e m e n t is a n im p o rta n t p a rt o f m o d e rn c o m p u te riz e d o b je c t re c o g n itio n
m e th o d s . E d g e e n h a n c e m e n t ty p ic a lly re q u ire s c o n v o lu tio n o p e ra to rs th a t p ro d u c e w e ig h te d
a v e ra g e tra n s fo rm a tio n s o n in d iv id u a l p ix e ls o f s o m e s o u rc e im a g e . E d g e e n h a n c e m e n t
o p e ra to rs te n d to b e d e v e lo p e d fo r p a rtic u la r ty p e s o f p ro b le m s a n d la rg e r o p e ra to rs te n d to b e
fo r v e ry s p e c ific p u rp o s e s . L a rg e irre g u la r p ix e l p a tte rn s re q u ire ra th e r la rg e a n d irre g u la r
e d g e e n h a n c e m e n t o p e ra to rs .
T h e s m o o th in g p ro c e s s a s s is ts a n e d g e d e te c tio n a lg o rith m to d is tin g u is h b e tw e e n
b a c k g ro u n d n o is e a n d a c tu a l o b je c ts in a n im a g e .
U n fo rtu n a te ly , s o ftw a re im p le m e n ta tio n s o f la rg e c o n v o lu tio n s a re e x tre m e ly s lo w .
C o n s e q u e n tly , c o n v o lu tio n s o f a n y re s p e c ta b le s iz e a re ty p ic a lly im p le m e n te d in s p e c ia liz e d
im a g e p ro c e s s in g h a rd w a re . A n o th e r p ro b le m w ith la rg e c o n v o lu tio n s is th a t it is o fte n v e ry
d iffic u lt to " d is c o v e r" g o o d o n e s . T h e p ro c e s s o f c re a tin g th e s e c o n v o lu tio n o p e ra to rs re q u ire s
tria l a n d e rro r, a n d e x p e rie n c e .
R e c e n tly , a t N e w L ig h t In d u s trie s , w e n e e d e d s p e c ia l p u rp o s e c o n v o lu tio n o p e ra to rs fo r u s e
in lo w -c o s t, c o m m e rc ia l o ff-th e -s h e lf h a rd w a re d e v ic e s . T h e s e c o n v o lu tio n s n e e d e d to b e v e ry
e ffic ie n t a n d im p le m e n ta b le in re c o n fig u ra b le h a rd w a re . U s in g g e n e tic a lg o rith m s a n d X ilin x
X C 6 0 0 0 te c h n o lo g y , w e im p le m e n te d a n e v o lu tio n a ry s y s te m th a t d e s ig n e d c o n v o lu tio n
o p e ra to rs fo r o u r a p p lic a tio n . T h e g e n e tic a lg o rith m e v a lu a te d o p e ra to rs b y c o m p a rin g th e
S . C a g n o n i e t a l. ( E d s .) : E v o W o r k s h o p s 2 0 0 0 , L N C S 1 8 0 3 , p p . 1 − 1 1 , 2 0 0 0 .
S p rin g e r-V e rla g B e rlin H e id e lb e rg 2 0 0 0
2
J . D u m o u lin e t a l.
c o n v o lu tio n o f o n e o rig in a l im a g e w
a b le to p ro d u c e .
T h e s e e x p e rim e n ts d e m o n s tra te d
e n h a n c in g c o n v o lu tio n k e rn e ls u s in g
d e s c rib e s o u r re s u lts a n d s h o w s a n
a lg o rith m . F in a lly , th is p a p e r p r
a p p lic a tio n s .
ith s e v e ra l p re p a re d im a g e s th a t th e c o n v o lu tio n s h o u ld b e
b o
e v
e v
e se
th th e
o lu tio
o lv e d
n ts s
v ia b
n a ry
c o n v
o m e
ility , a n d
te c h n iq u e
o lu tio n th
p o te n tia l
s o m e lim
s a n d p re
a t w a s u
fu tu re
ita tio n
p a re d
s e d in
re se a r
s , o f c re a tin g e d g e
im a g e s. T h is p a p e r
a n e d g e d e te c tin g
c h d ire c tio n s a n d
M o tiv a tio n
C o m p u te riz e d im a g e p ro c e s s in g p ro v id e s m a n y o p p o rtu n itie s fo r im p le m e n tin g
e v o lu tio n a ry p ro b le m -s o lv in g te c h n iq u e s . O u r p ro b le m is to id e n tify im p o rta n t fe a tu re s in a
d iffra c tio n p a tte rn th a t c h a n g e s o v e r tim e . T h is is a n a lo g o u s to id e n tify in g a n d tra c k in g
m u ltip le o b je c ts m o v in g th ro u g h a s e rie s o f fra m e s . C o n s id e r a c a m e ra a im e d a t a m o v in g
o b je c t o r a n o b je c t illu m in a te d b y a m o v in g lig h t s o u rc e . T h e p ro b le m is to id e n tify th e o b je c t
a s it m o v e s th ro u g h a s e rie s o f s till fra m e s . S in c e w e n e e d e d c o m p a c t, in e x p e n s iv e , e ffic ie n t
h a rd w a re to p e rfo rm th e id e n tific a tio n , w e c h o s e to im p le m e n t o u r s o lu tio n in re c o n fig u ra b le
s y s te m s u s in g X C 6 0 0 0 F P G A s fro m X ilin x .
O u r ta rg e t im a g e s w e re a s e rie s o f d iffra c tio n p a tte rn s re c o rd e d o v e r tim e w ith a m o v in g
lig h t s o u rc e . T h e d iffra c tio n p a tte rn s a re o p tic a lly d e riv e d F o u rie r T ra n s fo rm s o f fe a tu re s o n
th e s u rfa c e o f a n o b je c t. T h e fra m e s a re c a p tu re d in 8 -b it in g ra y -s c a le a n d th e o b je c ts to b e
re c o g n iz e d a re g e o m e tric " b lo b s " th a t b o th m o v e a n d c h a n g e s h a p e fro m fra m e to fra m e . W e
m u s t id e n tify e a c h o b je c t a s it m o v e s a n d c h a n g e s s h a p e o v e r a s e rie s o f fra m e s . W e c h o s e a
s im p le e d g e -d e te c tio n a lg o rith m [8 ] a s th e m e th o d fo r e x tra c tin g th e b o rd e r.
A k e y s te p in id e n tify in g o b je c t b o rd e rs in a p a rtic u la r fra m e is a p p ly in g a la rg e
c o n v o lu tio n o p e ra to r to th e im a g e . W e u s e d g e n e tic m e th o d s to d e v e lo p a p p ro p ria te
c o n v o lu tio n o p e ra to rs fo r o u r tra in in g s e t. T h e fo llo w in g s e c tio n s e x p la in th e a p p ro a c h a n d
th e re s u lts o f o u r in itia l e x p e rim e n ts .
B a c k g r o u n d
I m a g e P r o c e s s in g C o n s id e r a tio n s
T h e lite ra tu re o n im a g e p ro c e s s in g c o n ta in s m a n y e x a m p le s a n d te c h n iq u e s fo r p e rfo rm in g
im a g e e n h a n c e m e n t a n d im a g e s e g m e n ta tio n . G e n e ra lly , a s o n e m ig h t e x p e c t w ith a n y
c o m p u ta tio n a l ta s k , th e re a re tra d e o ffs b e tw e e n s p e e d a n d a c c u ra c y in th e s e g m e n ta tio n
p ro c e s s . W e u s e d o p tic a lly g e n e ra te d F o u rie r T ra n s fo rm s o f s u rfa c e fe a tu re s fo r th is
e x p e rim e n t. Id e n tific a tio n o f fe a tu re s in th e fre q u e n c y s p a c e o f th e s e im a g e s a m o u n te d to
id e n tify in g th e o b je c t b e in g s c a n n e d . W e n e e d e d to id e n tify n o t ju s t th e fe a tu re s o n m a n y
im a g e s , b u t th e p a th o f e a c h fe a tu re th ro u g h m u ltip le c o n s e c u tiv e fra m e s . W e in v e s tig a te d
s o m e w e ll-k n o w n m e th o d s fo r im a g e d e te c tio n th a t w e c o u ld u s e to fa c ilita te s e g m e n ta tio n
a n d o b je c t tra c k in g .
F o r th e e d g e d e te c tio n m e th o d , w e lo o k e d a t th re e a lg o rith m s : M a rr-H ild re th [8 ], th e
In fin ite S y m m e tric E x p o n e n tia l F ilte r (IS E F ) [9 ], a n d m o rp h o lo g ic a l b o u n d a ry e x tra c tio n [1 0 ].
M a rr-H ild re th e d g e d e te c to rs u s e a s m o o th in g a lg o rith m , u s u a lly a c o n v o lu tio n w ith a
G a u s s ia n fu n c tio n , fo llo w e d b y a z e ro -c ro s s in g b in a riz a tio n o f th e im a g e . W e c h o s e th is
m e th o d b e c a u s e it is s im p le , fa s t, a n d e a s y to im p le m e n t in v a rie d h a rd w a re a n d s o ftw a re .
T h e IS E F is a v e ry h ig h q u a lity e d g e d e te c to r o n im a g e s w ith h ig h fre q u e n c y u n ifo rm
n o is e . T h e IS E F a ls o c o m p u te s z e ro c ro s s in g s to fin d e d g e s , b u t it p e rfo rm s a m u c h m o re
c o m p le x s m o o th in g u s in g a b a n d -lim ite d L a p la c ia n o p e ra to r. T h e IS E F a lg o rith m w e
im p le m e n te d [1 1 ] u s e d a n e d g e fo llo w in g te c h n iq u e th a t c o n s u m e d m e m o ry a n d p ro b a b ly
S p e c ia l P u rp o s e Im a g e C o n v o lu tio n w ith E v o lv a b le H a rd w a re
3
c o u ld h a v e b e e n tim e -o p tim iz e d a s w e ll. W e fo u n d th is a lg o rith m to b e v e ry e ffe c tiv e , b u t
m u c h s lo w e r th a n M a rr-H ild re th . S in c e th e la tte r w a s a d e q u a te fo r o u r ta s k , w e d id n o t u s e
th e IS E F .
M o rp h o lo g ic a l B o u n d a ry E x tra c tio n u s e s s o m e s im p le s e t o p e ra tio n s to o u tlin e e d g e s o n
th e b o u n d a rie s o f o b je c ts in a n im a g e . In th e s im p le s t c a s e , th e e d g e e n h a n c e m e n t re s u lts
fro m e ro d in g a c o p y o f th e im a g e u s in g a s im p le k e rn e l a n d th e n s u b tra c tin g th e re s u ltin g
im a g e fro m th e o rig in a l im a g e . A n im a g e c o n ta in in g h ig h fre q u e n c y n o is e re q u ire s a n
a d d itio n a l o p e n in g s te p to c le a n th e fra m e b e fo re th e b o u n d a ry e x tra c tio n w ill w o rk c o rre c tly .
T h is a lg o rith m w a s n o t a s e ffe c tiv e fo r e x tra c tin g im a g e s fro m o u r im a g e s . T h e b o u n d a rie s
o f d iffe re n t fe a tu re s in th e im a g e te n d to b le e d to g e th e r if th e e ro s io n k e rn e ls a re n o t o f th e
p ro p e r s iz e fo r th e b o u n d a ry . If th e b o u n d a ry is d iffe re n t in d iffe re n t p a rts o f th e im a g e it m a y
b e n e c e s s a ry to e ro d e d iffe re n t p a rts o f th e im a g e w ith d iffe re n t k e rn e ls to g e t a n a c c u ra te
b o u n d a ry .
W e a ls o in v e s tig a te d s o m e th re s h o ld in g m e th o d s , b u t fo u n d th e m to b e v e ry lim ite d . O n e
m e th o d w e d id n o t e x p lo re w h ic h m ig h t p ro v e e ffe c tiv e is E d g e L e v e l T h re s h o ld in g , th o u g h
th is m e th o d p ro v e s to b e s lo w re la tiv e to th e m e th o d s w e c h o s e to in v e s tig a te .
W e d e c id e d to u s e th e M a rr-H ild re th m e th o d in itia lly , b e c a u s e it w a s b y fa r th e fa s te s t
m e th o d g iv e n o u r e n v iro n m e n t. T h e IS E F te c h n iq u e p ro v e d m o re a c c u ra te in o b je c t
s e g m e n ta tio n , b u t th e im p ro v e m e n t w a s n o t w a rra n te d g iv e n th a t th e p ro c e s s w a s s ig n ific a n tly
s lo w e r th a n M a rr-H ilre th . T h e M o rp h o lo g ic a l te c h n iq u e w e in v e s tig a te d w o rk e d a d e q u a te ly
o n ly w ith a g re a t d e a l o f “ h a n d -tw e a k in g ” th e p a ra m e te rs .
T h is a c tu a lly m a d e th e
m o rp h o lo g ic a l b o u n d a ry e x tra c tio n m e th o d a c a n d id a te fo r la te r e x p e rim e n ts w ith g e n e tic
m e th o d s . F o r th e c u rre n t e x p e rim e n t, th o u g h , w e fo u n d it d iffic u lt to c o n tro l th e re s u lts w h e n
w e trie d to a u to m a te th e a lg o rith m .
C o n v o lu tio n
C o n v o lu tio n o p e ra to rs a re c o m m o n ly u s e d in e d g e e n h a n c e m e n t te c h n iq u e s to in c re a s e
c o n tra s t o r e m p h a s iz e fe a tu re s w ith p a rtic u la r s h a p e s o n d ig itiz e d im a g e s . " E d g e s " a re u s u a lly
ra p id c h a n g e s in p ix e l v a lu e s . O n e w a y to th in k a b o u t a c o n v o lu tio n is a s a d iffe re n tia l
o p e ra to r th a t m e a s u re s th e ra te o f c h a n g e in s o m e d ire c tio n a lo n g th e im a g e . C o n v o lu tio n
o p e ra to rs fo r im a g e p ro c e s s in g a re u s u a lly e x p re s s e d a s o d d -v a lu e d s q u a re m a tric e s . T h e
o p e ra tio n is a p p lie d a c ro s s th e p ix e l fie ld b y m u ltip ly in g in d iv id u a l p ix e l v a lu e s b y th e m a trix
e le m e n ts a n d th e n a d d in g th e re s u lts to g e t a n e w p ix e l v a lu e . T h e n e w v a lu e re p la c e s th e
v a lu e fo r th e c e n te r p ix e l in th e fie ld . T h e o p e ra tio n is a p p lie d re p e a te d ly a c ro s s th e p ic tu re
u n til th e w h o le p ix e l fie ld h a s b e e n tra n s fo rm e d .
S im p le c o n v o lu tio n s a re o fte n 3 × 3 o r 5 × 5 m a tric e s . O n e s im p le a n d s tra ig h tfo rw a rd
c o n v o lu tio n is th e " d ire c tio n le s s " o r L a p la c ia n o p e ra to r. T h e L a p a c ia n o p e ra to r is d ire c tly
=
∂
2
+
∂
2
a n a lo g o u s to th e L a p la c ia n o f d iffe re n tia l a n a ly s is : ∇
2
L a p la c ia n c a n b e e x p re s s e d d is c re te ly a s a 3 × 3 m a trix
− 1
0
0
−
−
4
1
0
− 1
0
T h is is o n ly o n e c o m m o n e x a m p le o f a c o n v o lu tio n
b y g e n e tic m e th o d s la te r in th is p a p e r.
W h e n th e c o n v o lu tio n is a p p lie d to a 3 × 3 a re a o n
v a lu e fo r th e c e n te r p ix e l in a tra n s fo rm e d im a g e . T h
a s fo llo w s :
1
m a trix . W e w ill s h o w e x a m p le s c re a te d
∂ y 2
∂ x 2
U s in g th e d e fin itio n o f a d iffe re n c e , f ′ ≈ ( f ( x ) − f ( x − h ) ) h , a n d le ttin g h = 1 , th e
a p ix e l fie ld , th e o p e ra to r g iv e s a n e w
e c o n v o lu tio n is th e n m o v e d a c ro s s a n d
4
J . D u m o u lin e t a l.
d o w n th e p ix e l fie ld u n til a ll
o f d iffe re n t te c h n iq u e s fo r o
m a trix . In o u r e x p e rim e n ts , w
T h e c o n v o lu tio n o p e ra tio n
re s u lts . W e p re s e n t a little e
o n e h a s a s e c tio n o f a g ra y
p o rtio n o f a n im a g e .
2
1 0
1
5
1 2
0
2
8
1
0
F
se c t
th e
p a d
ig u r e
io n s o
c o n v o
d e d . A
3
6
4
6
2
g r
n in
c ia
n t
a y
g
n
h a
n v e rte d . A t th e e d g e s , th e re a re a n u m b e r
c a n n o t b e m a p p e d to th e c e n te r o f a 3 × 3
th e b o rd e rs o f th e p ix e l fie ld s .
a n p ro d u c e c o m p le x a n d o fte n in s c ru ta b le
a p p lie d to a s m a ll a rra y b e lo w . S u p p o s e
a t in F ig u re 1 . T h is c o rre s p o n d s to s o m e
0
2
4
4
o f a
e g in
a p la
lu tio
0
3
5
1 2
1 a . A s e c tio n
f th e im a g e , b
lu tio n w ith L
ls o a n y c o n v o
th e p ix e ls h a v e b e e n c o
p e ra tin g o n p ix e ls th a t
e ig n o re d th e v a lu e s a t
, th o u g h q u ite s im p le , c
x a m p le o f c o n v o lu tio n
s c a le im a g e s u c h a s th
0
0
0
1
0
2
s c a le
w ith
o f im
t su m
6
0
0
im
th e
a g e
s to
a g e
h ig
1 a
a n
0
. T h
h lig
. N
e g a
e
h t
o t
tiv
0
1 9
1 7
0
0
0
0
c o
e d
e
e
0
0
n v o
se c
th a t
n u m
0
lu t
tio
th
b e
io n
n .
e e
r is
0
9
0
0
0
o p e r a te s o n in c r e m e n ta l
F ig u r e 1 b . T h e r e s u lt o f
d g e s o f th e im a g e a r e 0
se t to z e r o .
C o n v o lu tio n is d e fin e d a s th e s u m o f th e e le m e n t-b y -e le m e n t p ro d u c ts o f tw o m a tric e s . S o
in th e c a s e o f th e L a p la c ia n o p e ra to r, a n d a n N × N m a trix , th e c o n v o lu tio n d e s c rib e s a n e w
m a trix w h o s e e le m e n ts a re d e fin e d a s : A ∗ ∆
i, j
A fte r m o v in g th is c o n v o lu tio n o p e ra to r a c
c o n tin u e s w ith th e n e x t 3 × 3 s e c tio n o f th e m a
s e c tio n o f th e m a trix a s s o c ia te d w ith lo c a l m a x
T h e L a p la c ia n is o n ly o n e o f m a n y w id e ly u
g iv e n a p a rtic u la r s h a p e to re c o g n iz e a n d a
c o n v o lu tio n k e rn e ls th a t w ill b e m o s t e ffe c tiv e
th e d e s ire d s h a p e o r s c a le . W h e n s h a p e s a re la
e ffe c tiv e (o r e ffic ie n t) c o n v o lu tio n k e rn e l to re
=
∑
1
p = − 1
ro s s th e s
trix . N o te
im a th a t a
se d c o n v o
p a rtic u la r
in e n h a n c
rg e o r irre
c o g n iz e th
∑
1
q = − 1
a
i+ p , j+ q
∆
p ,q
h a d e d a re a o f th e im a g e a b
h o w th e c o n v o lu tio n " e n h a n
re n o t o n e d g e s.
lu tio n o p e ra to rs (o r k e rn e ls ).
s c a le , th e re is a p a rtic u la r
in g th e im a g e to re c o g n iz e fe
g u la r, it c a n b e d iffic u lt to d e
e sh a p e .
o v e , o n e
c e s" o n e
U s u a lly ,
c la s s o f
a tu re s o f
v e lo p a n
E v o lu tio n a r y C o m p u tin g w ith X C 6 0 0 0 F P G A s
W e p e rfo rm e d th e im a g e e x tra c tio n a n d o rig in a l im a g e p ro c e s s in g te s tin g u s in g a M a tro x
G e n e s is im a g e g ra b b e r, w h ic h in c lu d e s a T M S 3 2 0 C 8 0 w ith p a ra lle l fix e d -p o in t M A C s a n d a
3 2 -b it d a ta b u s , a llo w in g u s to v e ry h a n d ily e v a lu a te th e d iffe re n t e d g e d e te c tio n m e th o d s
d is c u s s e d a b o v e . A t th e tim e o f its p u rc h a s e , th e d e v ic e a p p ro a c h e d $ 8 ,0 0 0 U S , fa r fro m th e
C O T S c rite ria th a t w e w e re lo o k in g fo r. F o r th is re a s o n , a n d to fa c ilita te th e e v o lu tio n a ry
a s p e c t o f o u r p ro je c t, w e c h o s e th e V C C F a t H O T P C I c a rd fo r o u r d e v e lo p m e n t s y s te m . It
w a s le s s th a n h a lf th e c o s t w h ile p ro v id in g th e n e c e s s a ry c o m p u tin g p o w e r.
D e v e lo p in g p ro g ra m s fo r F P G A s u s in g e v o lu tio n a ry m e th o d s re q u ire s s o m e a tte n tio n to th e
s p e c ia l c o n s tra in ts o f th e re c o n fig u ra b le h a rd w a re . F P G A -b a s e d d e s ig n s , e v o lu tio n a ry o r n o t,
a re p h y s ic a lly c o n s tra in e d b y g e o m e try a n d fu n c tio n u n it c a p a b ilitie s . F u n c tio n u n its w ith in
a n y F P G A p r o v id e b a s ic b in a r y lo g ic a l o p e r a tio n s ( e .g ., O R , A N D , X O R ) , b u t b e y o n d th is ,
th e re is a w id e v a ria tio n in th e p a rtic u la r fu n c tio n s th a t c a n b e im p le m e n te d w ith in a s in g le
fu n c tio n u n it o n th e F P G A .
S im ila rly , th e g e o m e try o f th e F P G A a s a w h o le h a s a n e ffe c t o n d e s ig n . T h e w a y in w h ic h
in d iv id u a l fu n c tio n u n its a re c o n n e c te d to o th e r fu n c tio n u n its a n d to th e e d g e s o f th e d e v ic e
a ls o c o n s tra in s th e ro u tin g o f c irc u itry in th e F P G A . W e h a v e c h o s e n th e X ilin x X C 6 0 0 0
s e rie s o f F P G A s fo r o u r re se a rc h fo r th e fo llo w in g re a s o n s :
• T h e re is a s ig n ific a n t b o d y o f re s e a rc h o n e v o lu tio n a ry h a rd w a re d e s ig n b u ild in g a ro u n d
th is c la s s o f F P G A d e v ic e s .
•
S p e c ia l P u rp o s e Im a g e C o n v o lu tio n w ith E v o lv a b le H a rd w a re
5
T h e c h ip d e s ig n is o p e n a n d c o n fig u ra tio n fo rm a ts a re a c c e s s ib le a t th e h a rd w a re le v e l.
F u n c tio n u n it ro u tin g is e x tre m e ly fle x ib le .
• T h e d e v ic e h a s a w e ll-d e fin e d P C I b u s in te rfa c e s ta n d a rd fro m th e m a n u fa c tu re r; a n d th is
m a k e s th e d e v ic e u s e fu l fo r e x p e rim e n tin g w ith d e s k to p c o m p u te r in te rfa c e s in a c o n s is te n t
m a n n e r w ith o th e r re s e a rc h e rs a n d m a n u fa c tu re rs .
• T h e m a n u fa c tu re r h a s d e v e lo p e d A P I s e ts th a t m a k e d e v ic e p ro g ra m m in g e a s ily a c c e s s ib le
to th e s o ftw a re d e v e lo p e r u s in g a d e s k to p P C . O th e r A P I s e ts a re a v a ila b le fo r d iffe re n t
d e v e lo p m e n t e n v iro n m e n ts .
T h e X C 6 0 0 0 g iv e s u s th e fle x ib ility to e x p e rim e n t w ith d iffe re n t e v o lu tio n a ry m e th o d s a n d
to ra p id ly b u ild a n d u tiliz e n e w to o ls fo r o u r re s e a rc h . M a n y re s e a rc h e rs d e s c rib e th e X C 6 2 0 0
a rc h ite c tu re a s it a p p lie s to o u r e x p e rim e n t. In p a rtic u la r, s e e [3 ], [4 ] a n d [5 ]. M a n y p o s s ib le
a p p ro a c h e s e x is t fo r e v o lu tio n a ry d e s ig n u s in g F P G A s . W e w ill d is c u s s re le v a n t a p p ro a c h e s to
o n e fe a tu re re c o g n itio n s u b s y s te m th a t h a v e e m e rg e d fro m o u r p re s e n t re s e a rc h . W h a t
d is tin g u is h e s d iffe re n t e v o lu tio n a ry d e s ig n a p p ro a c h e s is th e re p re s e n ta tio n a n d e v a lu a tio n o f
tria l s o lu tio n s .
•
E v o lu tio n a r y H a r d w a r e D e s ig n
R e p r e se n ta tio n
T h e k e y to p ro b le m s o lv in g w ith g e n e tic a lg o rith m s is to fin d a s u
th e p ro b le m . A g o o d re p re s e n ta tio n m u s t e n c o m p a s s th e fu ll ra n g e o f
m u s t e n a b le th e g e n e tic a lg o rith m to g e n e ra te n e w tria l s o lu tio n s
p ro b a b ility o f re ta in in g b e n e fic ia l fe a tu re s o f o ld tria l s o lu tio n s . M u
c h a n g e tria l s o lu tio n s in w a y s th a t le a d to re la tiv e ly s m a ll fitn e s s c h a n
o p e ra to rs s h o u ld te n d to a c c u m u la te u s e fu l fe a tu re s o f tw o o r m o re tr
tria l s o lu tio n . W e in v e s tig a te d tw o d iffe re n t re p re s e n ta tio n s o f F P G
m e th o d u s e s a c o d e d re p re s e n ta tio n o f p o s s ib le fu n c tio n s a n d a c
p o s s ib le ro u tin g s .
ita b le re p re s e n ta tio n o f
p o s s ib le s o lu tio n s , a n d
w ith a re la tiv e ly h ig h
ta tio n o p e ra to rs s h o u ld
g e s , a n d re c o m b in a tio n
ia l s o lu tio n s in to a n e w
A c o n fig u ra tio n s . O n e
o d e d re p re s e n ta tio n o f
F u n c tio n s
0
C o n s
C o n s
X A N
X O R
X X O
IN V
B U F
1
2
3
4
5
6
ta n
ta n
D
Y
R
t 0
t 1
Y
Y
R o u tin g :
1
2
3
4
N o rth
E a st
W e st
S o u th
F ig u r e 2 : C o d e d R e p r e s e n ta tio n I m p le m e n te d a s a L o o k -U p T a b le .
6
J . D u m o u lin e t a l.
C o d e d re p re s e n ta tio n s a re c o m m o n (s e e [4 ], [6 ]) in e v o lu tio n a ry p ro g ra m m in g , p a rtly
b e c a u s e th e y a re re la tiv e ly e a s y to d e s c rib e . In o u r re p re s e n ta tio n , o u r c h ro m o s o m e is a 2 D
a rra y o f in te g e rs th a t d e fin e d a c o n v o lu tio n k e rn e l n u m e ric a lly . T h e fitn e s s e v a lu a tio n s te p
c o n v e rts th e k e rn e l in to a s e t o f b it v a lu e s a n d th e n c o n v e rts th e b it v a lu e s in to a c o n fig u ra tio n .
T h e c o n fig u ra tio n p e rfo rm s th e c o n v o lu tio n b u t a llo w s u s to h a rd -c o d e (o r, m o re a c c u ra te ly ,
h a rd -w ire ) th e c o n v o lu tio n k e rn e l in to th e c irc u it d e s c rip tio n . T h is re d u c e s th e c irc u it s iz e a n d
s p e e d s u p th e e v a lu a tio n o f th e c o n v o lu tio n .
T h e p rim a ry a d v a n ta g e o f th is re p re s e n ta tio n is th a t it is a b s tra c te d fro m th e p a rtic u la r
F P G A a rc h ite c tu re , s o th a t it c a n b e re w ritte n to c o n fig u re d iffe re n t d e v ic e s (a s lo n g a s o n ly
c o m m o n fu n c tio n u n it/ro u tin g c o n fig u ra tio n s a re a llo w e d ). T h e p rim a ry d is a d v a n ta g e s o f th is
re p re s e n ta tio n a re :
• O n ly c o m m o n F P G A fu n c tio n c o n fig u ra tio n s a re a llo w e d , s o s p e c ia l fe a tu re s o f c e rta in
F P G A s ( i.e ., s e p a r a b le c lo c k in g , r e g is te r p r o te c tio n in th e X C 6 0 0 0 ) w ill n o t b e a v a ila b le to
th e e v o lu tio n a ry p ro c e s s . N o te th a t th is c a n b e a n a d v a n ta g e in s o m e c irc u m s ta n c e s , s u c h a s
w h e n it is d e s ira b le to c o n s tra in th e F P G A to c lo c k e d c irc u its .
• T h e a d d e d le v e l o f a b s tra c tio n (b e y o n d a b it-le v e l re p re s e n ta tio n ) a d d s a le v e l o f
c o m p le x ity to th e te s tin g p ro c e s s .
W e a ls o e v a lu a te d in trin s ic e v o lu tio n , in w h ic h w e d ire c tly u s e th e u n d e rly in g s tru c tu re o f
th e X C 6 0 0 0 in th e fu n c tio n u n it e n c o d in g . (F o r e x a m p le s o f in trin s ic e v o lu tio n , s e e [3 ].) In
th is re p re s e n ta tio n w e p a y p a rtic u la r a tte n tio n to th e a d d re s s /d a ta s tru c tu re o f th e F P G A
c o n fig u ra tio n . P ro g ra m m in g th e F P G A a t th is le v e l is a n a lo g o u s to p ro g ra m m in g a s ta n d a rd
C P U u s in g n a tiv e m a c h in e c o d e . T h e c o m p le x ity o f c re a tin g w o rk in g p ro g ra m s a t th is le v e l is
a v o id e d in th e e v o lu tio n a ry p ro c e s s . F P G A c o n fig u ra tio n s e v o lv e to a p p ro a c h th e p ro b le m
re p re s e n te d in th e fitn e s s s ta g e .
B y te 0
B y te 1
B y te 2
O u tp u t R o u tin g
F u n c tio n C o n fig / In p u t
F u n c tio n C o n fig / In p u t
F ig u r e 3 : F u n c tio n U n it C o n fig u r a tio n B y te s
A c irc u it c o n fig u ra tio n is re p re s e n te d a s a s trin g o f b y te s th a t to g e th e r d e fin e th e
c o n fig u ra tio n o f a fu n c tio n u n it o n th e X C 6 0 0 0 . T h e c o n fig u ra tio n o f a s in g le fu n c tio n u n it
c o m p ris e s th re e b y te s th a t d e fin e th e lo c a l ro u tin g a n d th e lo g ic im p le m e n te d in th e fu n c tio n
u n it.
O u r c h ro m o s o m e is a tw o -d im e n s io n a l a rra y o f th e s e th re e -b y te u n its . E a c h e le m e n t o f th e
a rra y re p re s e n ts a fu n c tio n u n it. T h e in d e x o f e a c h e le m e n t re p re s e n ts th e re la tiv e p o s itio n o f
th e fu n c tio n u n it w ith in th e F P G A , th e re b y a llo w in g u s to c o n s tru c t a n a d d re s s fo r e a c h
e le m e n t o f th e c o n fig u ra tio n . F o r d e ta ils a b o u t w h a t p a rtic u la r v a lu e s in d ic a te , s e e [1 ]. T h e
p rim a ry a d v a n ta g e s o f th is re p re s e n ta tio n a re :
• T h e re p re s e n ta tio n m a y b e d ire c tly in te rp re te d a s a " p ro g ra m " o f th e F P G A . V e ry little
p re lim in a ry p ro c e s s in g is n e e d e d to c re a te a lo a d a b le c o n fig u ra tio n .
• A ll p o s s ib le fu n c tio n s a re a llo w e d in p rin c ip le , b u t th e y c a n b e re s tric te d a t w ill b y m a s k in g
th e a p p ro p ria te b its o f e a c h fu n c tio n u n it c o n fig u ra tio n .
•
•
T h e p rim
T h e re su
u n d e rs ta
T h e c o n
m e th o d
m a n u fa c
a ry d is a d v a n ta g
ltin g e v o lv e d c o
n d .
fig u ra tio n is a lw
fo r o n e m a n u
tu re r’s F P G A .
e s o f a n in trin s ic re p re s e n ta tio n a re :
n fig u ra tio n c a n b e v e ry d iffic u lt fo r a h u m a n F P G A p ro g ra m m e r to
a y s s p e c ific to a p a rtic u la r F P G A . A p ro g ra m g e n e ra te d v ia th is
fa c tu re r’s F P G A c a n n o t b e e x p e c te d to ru n o n a d iffe re n t
S p e c ia l P u rp o s e Im a g e C o n v o lu tio n w ith E v o lv a b le H a rd w a re
T h e first re p
re a s o n s . U s in g
d e v e lo p m e n t.
e x p e rim e n ts to
o th e r m e th o d s
s im p le a n d a d e
re s e n ta tio n d isc u s s e d a b o v e is s u p e rio r to
th e firs t m e th o d , w e w ill n o t b e tie d to a n y
It is , h o w e v e r, m u c h m o re tim e -c o n s u m
d a te h a v e c o n c e n tra te d o n th e s e c o n d r
fo r c re a tin g F P G A c o n fig u ra tio n s (S e e [6
q u a te fo r th e ta s k .
th e s e c o n d fo r
p a rtic u la r F P G
in g to im p le m
e p re s e n ta tio n .
] in p a rtic u la r)
a n u m
A fo r
e n t th
T h e re
, b u t o
b e r o
th e p
e fir
a re ,
u r m
f p
u rp
s t,
o f
e th
7
ra c tic a l
o se s o f
so o u r
c o u rse ,
o d s a re
F itn e s s E v a lu a tio n a n d C r o s s o v e r
T h e e v a lu a tio n o f e a c h s o lu tio n c re a te d b y th e g e n e tic a lg o rith m is a n o th e r k e y a s p e c t o f
e v o lu tio n a ry d e s ig n o f p ro g ra m s fo r o u r e x p e rim e n t. W e te s t e a c h tria l s o lu tio n c re a te d b y o u r
g e n e tic a lg o rith m , a n d u s e th e te s t re s u lts to a s s ig n a fitn e s s v a lu e to th e tria l s o lu tio n .
G e n e ra lly , th e o b je c tiv e o f th e d e s ig n p ro c e s s is to d e v e lo p a d e s ig n th a t m e e ts c e rta in
p e rfo rm a n c e c rite ria . In e v o lu tio n a ry d e s ig n , h o w e v e r, it is im p o rta n t to u s e fitn e s s v a lu e s to
g u id e e v o lu tio n e ffic ie n tly to w a rd s th e fin a l o b je c tiv e .
T h e b a s ic flo w o f o u r F P G A e v o lu tio n a lg o rith m is :
1 . L o a d a f ile th a t r e p r e s e n ts a ll th e s ta tic c o n f ig u r a tio n p o r tio n s o f th e p r o g r a m ( e .g ., I /O
R e g is te rs , C o u n te rs , a n d o th e r c o n tro l s e c tio n s )
2 . C re a te a ra n d o m p o p u la tio n o f tria l s o lu tio n s .
3 . F o r e a c h tria l s o lu tio n , e v a lu a te a s fo llo w s :
a . C o n v e rt th e tria l s o lu tio n to a lo a d a b le c o n fig u ra tio n .
b . L o a d th e c o n fig u ra tio n .
c . L o a d th e in p u t re g is te rs .
d . W a it fo r s o m e fix e d a m o u n t o f tim e .
e . R e a d th e o u tp u t re g is te r.
f. C o m p a re th e re g is te r w ith a c a lc u la te d re s u lt b a s e d o n th e in p u t to g e t th e e rro r.
g . A s s ig n th e fitn e s s o f th e tria l s o lu tio n b a s e d o n s te p f.
4 . S e le c t tria l s o lu tio n s a c c o rd in g to fitn e s s to g e n e ra te a n e w p o p u la tio n .
5 . G e n e ra te n e w tria l s o lu tio n s b y a p p ly in g g e n e tic o p e ra to rs (m u ta tio n a n d re c o m b in a tio n ) to
th e s e le c te d tria l s o lu tio n s .
6 . A p p ly s te p s 3 th ro u g h 6 re p e a te d ly u n til a s a tis fa c to ry s o lu tio n e m e rg e s .
In s te p g a b o v e , c h a n g e s c a n b e m a d e to e m p h a s iz e o r d e -e m p h a s iz e c e rta in a s p e c ts o f th e
c a lc u la te d e rro r a t d iffe re n t s ta g e s in th e e v o lu tio n s o th a t, fo r e x a m p le , b a s ic fu n c tio n s a re
e v o lv e d a n d o p tim iz e d b e fo re th e y a re c o m b in e d in to m o re c o m p le x s tru c tu re s . F itn e s s
c a lc u la tio n c a n in c lu d e w e ig h tin g th e e rro r to g iv e m o re in flu e n c e to s o lu tio n s th a t m e e t
c e rta in c o n s tra in ts o r s u b – c rite ria . T h is is e x tre m e ly im p o rta n t in th e c a s e o f o u r a p p lic a tio n
s in c e w e w ill b e c o m p a rin g v e c to r v a lu e s re p re s e n tin g p a tte rn s o f fe a tu re s id e n tifie d b y th e
F P G A to re fe re n c e v e c to rs c a lc u la te d b y o th e r m e a n s .
In o rd e r to m a in ta in th e in te g rity o f w o rk in g p o rtio n s o f e a c h g e n o m e , w e d e v e lo p e d a
c ro s s o v e r o p e ra to r th a t p re s e rv e s re c ta n g u la r p o rtio n s o f th e g e n o m e d u rin g c ro s s o v e r. B y
d e fa u lt, th e 2 D in te g e r (3 D B in a ry ) g e n o m e in G a il w ill p e rfo rm b y te -le v e l c ro s s o v e r. T h is
h a s th e e ffe c t o f d e s tro y in g lo c a l o p tim a l a re a s in th e g e n o m e fro m g e n e ra tio n to g e n e ra tio n .
W e trie d to p re s e rv e lo c a l o p tim a l b e h a v io r in th e g e n o m e b y tre a tin g e a c h g e n o m e a s a 2 D
a rra y a n d e a c h c ro s s o v e r c o m p o n e n t a s a 2 D s u b -a rra y o f th e g e n o m e .
F o r e x a m p le , g iv e n tw o 9 × 9 g e n o m e s to c ro s s o v e r, th e c ro s s -o v e r a lg o rith m w o rk s a s
fo llo w s :
1 . R a n d o m ly s e le c t a s ta rtin g ro w a n d s ta rtin g c o lu m n . T h is w ill b e th e u p p e r le ft c o rn e r o f
th e c ro s s o v e r a rra y .
2 . R a n d o m ly s e le c t th e n u m b e r o f ro w s a n d c o lu m n s to c ro s s o v e r. T h is d e fin e s th e s iz e o f th e
c ro sso v e r a rra y .
8
J . D u m o u lin e t a l.
3 . S e
o f
4 . E x
5 . T h
to
L a
le c t tw o g e n o m e s to c ro s s o v e r a n d re m o v e
th e c a lc u la te d s iz e , fro m e a c h o f th e g e n o m
c h a n g e th e se s u b -a rra y s b e tw e e n th e g e n o m
e e v o lv e d c o n v o lu tio n s w e re c re a te d b y g e
" re p ro d u c e " a n im a g e th a t w a s e n h a n c e d w
p la c ia n ) in s p e c ia l p u rp o s e im a g e p ro c e s s in
a su b -a rr
e s.
e s.
n e tic a lly
ith a g e n
g h a rd w a
a y fro m
th e a p p ro p ria te p o s itio n a n d
d e v e lo p in g a c o n v o lu tio n to a tte m p t
e ra l-p u rp o s e c o n v o lu tio n (a p s e u d o re .
W e e v a lu a te d fitn e s s a s fo llo w s :
1 . G e t
2 . C o n
3 . C o m
(S e e
re su
4 . C o m
a g e n o m e .
v o lv e th e o r
p a re th e re
F ig u re 4 b
ltin g im a g e
p a re fitn e s s
ig in a l im a g e w ith th e g e n o m e .
s u ltin g im a g e w ith th e im a g e c
e lo w ) to g e t th e fitn e s s v a lu e . T
w ith e a c h p ix e l in th e o rig in a l im
e s p re s e rv in g th e lo w e st v a lu e (h
o n v o lv e d w ith th e 9 × 9 p s e u d o -L a p la c ia n
h e e v a lu a tio n c o m p a re s e a c h p ix e l in th e
a g e .
ig h e s t fitn e s s ) in d iv id u a ls .
R e s u lts
T h e e x p e rim e n ts
G A L IB 2 .4 .2 a n d th
9 × 9 2 -d im e n s io n a l
p ro b a b ility o f 2 0 % ,
th e G A p ro d u c e d v e
A s w e se e in th e
" te rrib le " re s u ltin g
fitn e s s e v a lu a tio n . B
re w a rd fitn e s s v a lu e
s h o w n in F ig u re 6 .
W e p re se n t so m
p ro d u c e d b e lo w . T h
F ig u re 4 .
w e p e rfo rm e d w e re c a rrie d o u t u s in g a g e n e tic a lg o rith m d e fin e d u s in g
e X IL IN X A P I lib ra ry fo r th e 6 2 0 0 p a rts . T h e G A w a s s e t u p to e v o lv e a
in te g e r g e n o m e . W e ra n th e G A w ith p o p u la tio n s iz e o f 5 0 , c ro s s o v e r
a n d m u ta tio n p ro b a b ility o f 5 % . W e n o tic e d th a t, u n d e r th e s e c o n d itio n s ,
ry s ta b le re s u lts a fte r a p p ro x im a te ly 1 0 0 g e n e ra tio n s .
p ic tu re s b e lo w , lo w fitn e s s m e m b e rs o f th e p o p u la tio n a c tu a lly p ro d u c e d
im a g e s . W e d e te rm in e d th a t th is w a s d u e to n u m e ric o v e rflo w d u rin g
e c a u se o f th is th e fitn e s s e v a lu a tio n w a s re -e x a m in e d a n d w e d e c id e d to
s c lo s e to h a lf th e m a x im u m v a lu e o f a lo n g in te g e r. O n e " g o o d " re s u lt is
e e x a m p le s o f e v o lv e d c o n v o lu tio n s a n d th e e n h a n c e d im a g e s th e y
e G A c re a te d a n d e v a lu a te d th e s e im a g e s u s in g th e tra in in g s e t im a g e in
− 1
0
0
0
− 1
0
0
0
− 1
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
− 1
0
0
0
8
0
0
0
0
0
0
0
0
− 1
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
− 1
0
0
0
0
0
0
− 1
0
0
0
0
0
0
− 1
F ig u r e 4 : A n o r ig in a l d iffr a c tio n p a tte r n fr o m th e tr a in in g s e t (le ft), a p s e u d o
L a p la c ia n o p e r a to r (r ig h t) a n d th e c o n v o lv e d im a g e u s e d fo r c o m p a r is o n (c e n te r ).
E v o lv e d c o n v o lu tio n o p e ra to rs p ro d u c e s o m e w id e ly v a ry in g re s u lts . B e lo w a re a fe w
e x a m p le s o f e v o lv e d o p e ra to rs a n d th e im a g e s th e y p ro d u c e d fro m th e o rig in a l im a g e in
F ig u re 4 .
S p e c ia l P u rp o s e Im a g e C o n v o lu tio n w ith E v o lv a b le H a rd w a re
1 1 0
− 6 6
4
− 2 3
− 3 9
− 3 2
1 1 7
− 1 1 5
− 2 0
− 5 7
8 3
− 7 1
− 5 7
− 3 1
1 2
7 4
3 3
1 0 7
2 2
9 1
− 6 2
− 4 1
3 5
− 5 1
5 0
− 3 9
− 1 9
− 1
6
9
− 1
5
1 1
9 6
1 2 0
0
− 6 4
1 1 0
6 0
6 6
− 6 3
3 4
0
0
0 5
7
3
8 5
− 1 1 9
8 0
E v o lv e d c o n v o lu tio n
w ith fitn e s s = 1 2 6 4 4 .
− 6 1
− 1 9
− 1 0 0
1 1 9
1 2 3
− 4 0
5 9
− 9 8
− 9 5
a n d
− 3 6
1 1 1
− 4 8
− 6 9
− 9 4
− 6 1
1 8
4 6
− 3 0
9 2
− 1 6
− 6 9
− 1 1 7
6 9
− 9 9
8 2
− 1 2 3
− 6
e n h a n c e d
− 8 0
− 1 3
− 3 0
1 6
4 9
2 7
4 4
− 3 6
7 1
1 2
7
8 6
2
− 4 7
− 5 4
4 5
im a g e
E v o lv e d c o n v o lu tio n a n d e n h a n c e d
w ith fitn e s s = 2 5 2 0 4 0 .
6
− 4 6
− 1 1 0
− 9 3
− 1 2 0
− 2 8
− 7 1
− 1 2 6
− 2 2
− 2 2
3 4
1 0 4
7 7
− 9
1 3
− 2 3
2 3
− 7 5
− 7
6 8
− 6 8
− 7
1 0 4
5 9
2 2
9 6
− 1 2 4
1 5
− 3
1 2 3
− 6 6
− 4
1 0 0
− 3 9
1 2 7
1 1 5
3 1
− 1 1 2
7 1
− 7 6
1 1 5
− 9 2
− 1 1 4
7 6
− 1 0
1 1 6
− 1 2 6
− 4 9
− 7 5
1 2 6
− 3 0
1
− 1 0
− 5 9
− 5 2
1 0 5
− 1 0
7
5 6
− 1 1 8
− 2
− 6 9
5 9
− 2 6
− 3 5
9
− 3 8
1 1 6
− 1 9
6 4
− 4 9
1 0 5
− 7 0
7 2
4 9
im a g e
F ig u r e 5 : T w o H ig h F itn e s s R e s u lts
In
d is tr
in tro
T h is
th e " h ig h -fitn
ib u tio n o f p ix e
d u c e d a h ig h -o
s u g g e s ts th a t la
e ss
l v
rd e
rg e
" c o n v o lu tio n
a lu e s th ro u g h
r ra n d o m n e ss
r c o n v o lu tio n
o p e
th e
a t th
a rra y
ra to rs,
d a rk a
e le v e
s m a y
w e
re a s
l o f
p ro v
c a n
o f
th e
id e
110
70
109
− 66
− 92
− 8
− 113
48
105
se e
th e o
c o n v
m o re
− 47
− 93
98
− 12
105
− 107
95
126
− 67
w h a t a p p e a rs to b e a ra n d o m
rig in a l fra m e . T h e s e o p e ra to rs
o lu tio n a rra y s iz e (9 × 9 p ix e ls ).
e ffe c tiv e m a tc h in g .
− 32
− 71
− 120
58
− 119
− 112
71
116
80
37
− 109
101
103
− 68
− 38
113
− 15
82
− 31
− 15
84
58
− 85
37
70
15
99
35
− 122
127
101
− 52
98
116
− 84
44
− 128
103
− 55
− 52
− 57
30
− 86
− 61
88
4
− 86
58
− 57
− 72
− 43
33
− 50
5
0
88
− 106
− 54
85
− 37
− 81
39
− 96
F ig u r e 6 : E v o lv e d c o n v o lu tio n a n d e n h a n c e d im a g e w ith fitn e s s = 2 0 1 8 8 5 1 8 9 7 .
A
w a s
F ig u
e d g e
lo w
g e n e
re 4 ,
d e te
fitn e s s re s u
ra te d in th
y e t it p ro v
c to r to fin d
lt th a t p r
e 1 0 0 th
id e s v e ry
fe a tu re s
o v
g e
g
in
id e d
n e ra
o o d
th e
b e tte r m a tc h in g
tio n . It m a tc h e s
d e ta il e x tra c tio n
s a m p le fra m e (s h
a p p e a rs
a lm o s t
a n d w a
o w n in
in F ig u r
n e g a tiv e
s in fa c t
F ig u re 4
e 6 a
ly w
u se d
) a n d
b o v
ith
in
o th
e . T h is
th e te m
a M a rr
e rs p ro
o p e ra
p la te
-H ild r
d u c e d
to r
in
e th
in
1 0
J . D u m o u lin e t a l.
tim e -s e q u e n c e in th e s a m e m a n n e r a s o u r e x a m p le . N o te th a t th e fitn e s s is " c lo s e " to h a lf th e
m a x im u m v a lu e o f a n u n s ig n e d 6 4 -b it in te g e r.
C o n c lu s io n s
W e h a v e a n u m b e r o f c o n c lu s io n s a n d p o s s ib le p a th s o f c o n tin u e d re s e a rc h fro m th e s e
in itia l re s u lts .
• I m p r o v e d F itn e s s E v a lu a tio n . W e n e e d a fitn e s s e v a lu a tio n m e th o d o lo g y m o re
a p p ro p ria te fo r th e v e ry la rg e n u m b e r o f c a lc u la tio n s re q u ire d to c o m p a re th e s e re s u lts
p ix e l-b y -p ix e l. O n e a p p ro a c h w o u ld a v e ra g e lo c a l re s u lts b e fo re d o in g th e s u m o f s q u a re s .
• L a r g e r C o n v o lu tio n K e r n e ls . S in c e w e a re w o rk in g w ith im a g e s o f 6 4 0 × 4 8 0 p ix e ls , w e
h a v e th e p o te n tia l to s e e b e tte r re s u lts w ith la rg e r c o n v o lu tio n k e rn e ls o r u s in g m u ltip le
k e rn e ls a n d lin e a rly c o m b in in g th e re s u ltin g im a g e s . P la n s a re n o w u n d e r w a y to 1 0 0 × 1 0 0
p ix e l k e rn e ls u s in g a c lu s te r c o m p u te r.
• T h e U tility o f th e F P G A v s . C lu s te r e d S y s te m s . T h e F P G A is v e ry u s e fu l fo r fitn e s s
e v a lu a tio n a n d e x e c u tio n o f s m a ll k e rn e ls b e c a u s e th e c o n v o lu tio n m a th e m a tic s c a n b e
d e s ig n e d in to th e c irc u it. T h e X C 6 2 0 0 in p a rtic u la r is v e ry u s e fu l w h e n w e c o n v o lv e a
fra m e u s in g m u ltip le k e rn e ls . F o r la rg e c o n v o lu tio n k e rn e ls , h o w e v e r, m u ltip le F P G A s a re
n e e d e d . T h is is b e c a u s e th e d a ta b u s o f th e F P G A b e c o m e s a lim ita tio n w h e n e v a lu a tin g
la rg e s e ts o f n u m b e rs . W e w ill b e e x p lo rin g th e u s e o f m a trix m a th o p e ra tio n s b u ilt in to
m o d e r n C O T S m ic r o p r o c e s s o r s ( e .g ., M M X , 3 D n o w ) to o p tim iz e c a lc u la tio n tim e f o r
m a trix o p e ra tio n s . W e h o p e to p ro d u c e a p e rfo rm a n c e c o m p a ris o n e v a lu a tin g C O T S
te c h n iq u e s fo r in te g e r m a trix c a lc u la tio n s in th e fu tu re .
• A p p lic a b ility to M o tio n D e te c tio n . F o r s o m e v e ry ru d im e n ta ry ty p e s o f o b je c ts , o u r
te c h n iq u e s c o u ld p ro v e u s e fu l fo r id e n tify in g " c u s to m " k e rn e ls th a t c o rre s p o n d to p a rtic u la r
o b je c ts . F o r e x a m p le , a ro b o tic v e h ic le c o u ld u s e th is m e th o d to is o la te u n re c o g n iz e d
o b je c ts a n d s a v e a lib ra ry o f d is c o v e re d te m p la te s fo r fu tu re re fe re n c e . W e h a v e n o
c o n c lu s iv e re s u lts a lo n g th is lin e , o n ly th e s u g g e s tio n o f v a lu e .
• O v e r a ll V ia b ility . T h e m e th o d o lo g y d e s c rib e d in th is p a p e r p ro v e d u s e fu l fo r o u r
p a rtic u la r a p p lic a tio n . W e h o p e to e x te n d th e m e th o d o lo g y to p ro v id e u s e fu l g e n e ra tio n o f
a u to m a tic c o n v o lu tio n k e rn e ls fo r o th e r g e n e ra l a p p lic a tio n s .
A c k n o w le d g e m e n ts
T h is w o rk w a s fu n d e d b y B M D O . J a m e s A . F o s te r w a s a ls o p a rtia lly fu n d e d b y D O D /O S T .
B ib lio g r a p h y a n d R e fe r e n c e s
P
C
C
R
G
[1 ] X IL IN X X C 6 2 0 0 F ie ld P ro g ra m m a b le G a te A rra y s , A p ril 2 4 , 1 9 9 7 , X IL IN X , L td
[2 ] R . M u rg a i, R . B ra y to n , A S a n g io v a n n i-V in c e n te lli; L o g ic S y n th e s is fo r F ie ld ro g ra m m a b le G a te A rra y s " 1 9 9 5 , K lu w e r A c a d e m ic P u b lis h e rs
[3 ] A . T h o m p s o n , I. H a rv e y a n d P . H u s b a n d s ; U n c o n s tra in e d E v o lu tio n a n d H a rd
o n s e q u e n c e s C S R P 3 9 7 , (in T o w a rd s E v o lv a b le H a rd w a re , S p rin g e r-V e rla g L e c tu re N o te s in
o m p u te r S c ie n c e , 1 9 9 6 )
[4 ] J . K o z a , S . B a d e , F . B e n n e tt III, M . K e a n e , J . H u tc h in g s , D . A n d re ; R a p id ly
e c o n fig u ra b le F ie ld -P ro g ra m m a b le G a te A rra y s fo r A c c e le ra tin g F itn e s s E v a lu a tio n in
e n e tic P ro g ra m m in g , P u b lis h e d in K o z a , J o h n R . (e d ito r). L a te B re a k in g P a p e rs a t th e
S p e c ia l P u rp o s e Im a g e C o n v o lu tio n w ith E v o lv a b le H a rd w a re
1 1
G e n e tic P ro g ra m m in g 1 9 9 7 C o n fe re n c e , S ta n fo rd U n iv e rs ity , J u ly 1 3 -1 6 , 1 9 9 7 . S ta n fo rd , C A :
S ta n fo rd U n iv e rsity B o o k s to re . P a g e s 1 2 1 - 1 3 1 .
[5 ] D . M o n ta n a , R . P o p p , S u ra j Iy e r, a n d G . V id a v e r; E v o lv a w a re : G e n e tic P ro g ra m m in g
fo r O p tim a l D e s ig n o f H a rd w a re -B a s e d A lg o rith m s , 1 9 9 8 , B B N T e c h n o lo g ie s , P ro c . In t.
C o n f. o n G e n e tic P ro g ra m m in g .
[6 ] J . M ille r, P . T h o m s o n ; E v o lv in g D ig ita l E le c tro n ic C irc u its fo r R e a l-V a lu e d F u n c tio n
G e n e ra tio n u s in g a G e n e tic A lg o rith m , 1 9 9 8 , N a p ie r U n iv e rs ity , P ro c . In t. C o n f. o n G e n e tic
P ro g ra m m in g .
[7 ] L . P a g ie , P . H o g e w e g ; E v o lu tio n a ry C o n se q u e n c e s o f C o e v o lv in g T a rg e ts,
E v o lu tio n a ry C o m p u ta tio n 5 (4 ):4 0 1 -4 1 8 , 1 9 9 8 .
[8 ] D . M a rr, E . H ild re th ; T h e o ry o f E d g e D e te c tio n , P ro c e e d in g s o f th e R o y a l S o c ie ty o f
L o n d o n , S e rie s B , V o l. 2 0 7 , p p .1 8 7 -2 1 7 , 1 9 8 0 .
[9 ] J . S h e n , S .C a s ta n ; A n O p tim a l L in e a r O p e ra to r fo r S te p E d g e D e te c tio n , C o m p u te r
V is io n , G ra p h ic s , a n d Im a g e P ro c e s s in g : G ra p h ic a l M o d e ls a n d U n d e rs ta n d in g , V o l.5 4 , 2 :p p .
1 1 2 -1 3 3 , 1 9 9 2 .
[1 0 ] J . S e rra ; Im a g e A n a ly s is a n d M a th e m a tic a l M o rp h o lo g y , A c a d e m ic P re s s , 1 9 8 8 .
[1 1 ] J .R . P a rk e r; A lg o rith m s fo r Im a g e P ro c e s s in g a n d C o m p u te r V is io n , J o h n W ile y a n d
S o n s, 1 9 9 7 .
S te r e o s c o p ic V is io n fo r a H u m a n o id R o b o t
U s in g
P r o g r a m m in g
C h r is to p h e r T .M . G r a a e , P e te r
a n d M a ts N o rd a h l
C o m p le x S y s te m s G ro u p , In s titu te o f P h y s ic a l
T h e o ry , C h a lm e rs U n iv e rs ity o f
T e c h n o lo g y , S -4 1 2 9 6 G ö te b o rg , S w e d e n
In th is
w e in tro d u c e a n e w
a d a p tiv e s te re o s c o p ic
V is io n . W e u s e g e n e tic p ro g ra m m in g , w h e re th e in p u t to th e in d iv id u a ls is ra w
d a ta fro m s te re o im a g e -p a irs a c q u ire d b y tw o C C D
th e in d iv id u a ls is th e d is p a rity m a p , w h ic h is tra n s fo r
u s in g tria n g u la tio n . T h e u s e d g e n e tic
in d iv id u a ls , a n d
th e re b y
h ig h P e rfo rm
T h e e v o lv e d in d iv id u a ls h a v e a n
1 .5
w h ic h is e q u iv a le n t to a n u n c e rta in ty o f
T h is w o rk is
b y a p p lic a tio n s to th e c o n tro
m a n o id ro b o ts T h e H u m a n o id
a t C h a lm e rs .
fro m
th e
1
a n d
T h e o u tp u t
m e d to a 3 D m a p o f
e n g in e e v o lv e s
a n c e o n w e a k
d is p a rity -e rro r o f
1 0 % o f th e tru e
l o f a u to n o m o u s
M o tiv a tio n
M a n is th e S ta n d a rd fo r
to o ls
h u m a
m a n s
o f h u
T
T h e
T
h u m
th a t
w ith
V
m a n
a ll in te ra c tio n s in o u r w o rld w h e re m o s t e n v iro n m e n ts ,
a n d m a c h in e s a re a d a p te d to th e a b ilitie s , m o tio n c a p a b ilitie s a n d g e o m e try o f
n s . W a lk in g ro b o ts h a v e a v e ry la rg e p o te n tia l in e n v iro n m e n ts c re a te d fo r
a s w e ll a s in m o re n a tu ra 1
T h e la rg e s t p o te n tia l is a s s o c ia te d w ith ro b o ts
h u m a n o id ro b o ts . It c o u ld b e m o re
m a n -lik e d im e n s io n s w a lk in g o n tw o le g s
to c o n tro l v a rio u s m a c h in e s b y th e s e ro b o ts th a n to re b u ild a ll m a c h in e s fo r
C o m p u te r c o n tro l
h is w o rk is p a rt o f th e H u m a n o id
a t C h a lm e rs U n iv e rs ity o f T e c h n o lo g y .
a s e rie s o f h u m a n o id e x p e rim e n ts , a ll o f w h ic h w ill b e p rim a rily
b y e v o lu tio n a ry a d a p tiv e m e th o d s .
h e fin a l g o a l o f th e re s e a rc h is to b u ild a h u m a n -s iz e d ro b o t b a s e d o n a
a n S k e le to n to e n s u re
a u th e n tic ity . T h e
o f th e
is
a s e c o n d -g e n e ra tio n p ro to ty p e o f a s m a ll h u m a n o id is b e in g d e v e lo p e d
E L V IS
a h e ig h t o f
6 0 c m .
is io n is th e m o s t im p o rta n t o f o u r fiv e
A s a n e x a m p le a
o f th e h u is d e v o te d to
b u t so m e
d e v o te o v e r h a lf o f th e ir
A c c o rd in g to a n e w
S t. L o u is .
e s tim a te b y n e u ro s c ie n tis t D a v id V a n E s s e n o f W a s h in g to n U n iv e rs ity in
S . C a g n o n i e t a l. ( E d s .) : E v o W o r k s h o p s 2 0 0 0 , L N C S 1 8 0 3 , p p . 1 2 − 2 1 , 2 0 0 0 .
S p rin g e r-V e rla g B e rlin H e id e lb e rg 2 0 0 0
S te re o s c o p ic V is io n fo r a H u m a n o id R o b o t
1 3
1 4
C .T .M . G r a a e , P . N o r d in , a n d M . N o r d a h l
S te re o s c o p ic V is io n fo r a H u m a n o id R o b o t
1 5
1 6
C .T .M . G r a a e , P . N o r d in , a n d M . N o r d a h l
S te re o s c o p ic V is io n fo r a H u m a n o id R o b o t
1 7
1 8
C .T .M . G r a a e , P . N o r d in , a n d M . N o r d a h l
S te re o s c o p ic V is io n fo r a H u m a n o id R o b o t
1 9
2 0
C .T .M . G r a a e , P . N o r d in , a n d M . N o r d a h l
S te re o s c o p ic V is io n fo r a H u m a n o id R o b o t
2 1
s r
en ,
c
s r g
H
, a nd Z Z Y o n
gr
pa rt
t f lctrica l gi ri ga d lctr ics
i rsit f i rp l
, i rp l
, 6 3 3 , .K.
q.h.wu@liv.ac.uk
pa rt
t f lctr ic
gi ri g
i rsit,
5 6 , . . ia
c.
ispa p rpr s tsa
l
g tic clst ri ga l
g rit c i i g a g tic a l
g rit
( ) it t cl
a ssica l a rd - a s clst ri g a l
g rit (
). Itpr c ss s pa rtiti
a tric s ra t r t a
s ts f c t r p its a d t s pr id s a
i pl
ta ti sc
f r t g tic p ra t r -r c
i a ti . r c pa ris
fp rf r a c
it t r isti g clst ri g a l
g rit s, a gra -l li a g q a tia ti pr l is c sid r d. p ri ta lr s l
ts s
t a tt pr p s d
al
g rit c
rg s
r q ickl t t gl a l pti
a d t s pr id s a tt r a
t ft dil a i
ic t tra diti a lclst ri g
al
g rit s a r a sil tra pp d i lca l pti a a d t g tic a ppr a c
is ti c s i g.
r
c
lsterin
etho ds pl
a a ita lro l
e in e pl
o ra to r da ta a na lsis. In the el
d
o fpa ttern reco nitio n [ ], the co n entio na l- ea ns clsterin a lo rith s ( s) ha e been idela ppl
ied. ro a dlspea in ,
s ca n be cl
a ssi ed into
r
s (H
) a nd
s (
) clsterin a lo rith s. H a rd
clsterin dea l
s ith a ssinin ea ch o b jectpo intto e a ctl o ne o fthe clsters,
herea s f
clsterin e tends this co nceptto a sso cia te ea ch o b jectpo intto
ea cho fthe clsters itha ea s re o fbel
o n in ness. he bel
o n in nesso fo b ject
po intsto the ha rd/ f
clstersist pica l
lrepresented b a e bership a tri
ca l
l
ed a ha rd/ f
pa rtitio n,respectiel.In enera l
,
s a i a t ndin the
o pti a lpa rtitio n a nd o pti iin a clsterin o b jectie f nctio n sin ca l
c lsba sed etho ds. H o e er, clsterin o b jectie f nctio ns a re hihl no n-l
inea r
a nd
l
ti- o da lf nctio ns. s a co nseq ence, b hil
l
-cl
i bin ,
s ca n be
ea sil tra pped into l
o ca le tre a a sso cia ted ith a
r pa rtitio n (i.e., a
pa rtitio n itho ne o r o re e pt ro s, ea nin tha tfe er tha n clsters ere
o bta ined in the na lpa rtitio n). o reo er,the a re a l
so sini ca ntlsensitie to
the initia lco nditio ns.
a to a chie e bo tha o ida nce o fl
o ca le tre a a nd ini a lsensitiit to
initia l
ia tio n is to se sto cha stic o pti ia tio n a ppro a ches, s ch a s
r
r
s ( s). n
is inspired b o r a nic e o ltio n a nd ha s been idel
bel
ie ed to be a n effectie l
o ba lo pti ia tio n a lo rith . In [3] [6 ], enetica l
l ided a ppro a ches ere de ned fo r the o pti a lclsterin pro bl
e s a nd
S . C a g n o n i e t a l. ( E d s .) : E v o W o r k s h o p s 2 0 0 0 , L N C S 1 8 0 3 , p p . 2 2 − 3 3 , 2 0 0 0 .
S p rin g e r-V e rla g B e rlin H e id e lb e rg 2 0 0 0
A F a s te r G e n e tic C lu s te rin g A lg o rith m
2 3
e peri ents ere perfo r ed ith different da ta sets. es l
ts sho ed tha t a n
ca n a el
io ra te the diffic l
t o fcho o sin a n initia l
ia tio n fo r the
s
a nd pro ides a ia bl
e a to a o id l
o ca le tre a . H o e er, a s sta ted in [3], a n
clsterin a ppro a ch ta es p to t o o rders o f a nit de o re ti e tha n
H
/
clsterin a lo rith s.
s a re itera tie sche es o pera tin o n a po p l
a tio n o fso -ca l
l
ed indiid al
s. a ch indiid a lin the po p l
a tio n co rrespo nds to a ia bl
e so ltio n to the
pro bl
e in ha nd. In pre io s o r o n enetica l
l ided clsterin ([4] [6 ]),
ea ch indiid a lin the po p l
a tio n ha s been desined to representa a tri o f
clster centers. l
tho h nder t o sets o fclsterin criteria the a ppin bet een a pa rtitio n a nd the co rrespo ndin seto fclster centers is e clsie, the
o a lo fa t pica lclsterin a ppro a ch is essentia l
l the o pti a lpa rtitio n ra ther
tha n the po sitio ns o fclster centers. In this pa per, a no el enetica l
l ided
clsterin a lo rith is de ned, hich ses pa rtitio n a trices a s the indiid al
s in po p l
a tio n. his h brid a lo rith co bines a enetic a lo rith ( )
iththe cl
a ssica lha rd - ea ns clsterin a lo rith (H
) a nd is therefo re
ter ed a s a enetic ha rd - ea ns clsterin a lo rith ( H
). H
ha s been a ppl
ied to a ra -l
e eli a e q a ntia tio n pro bl
e . o pa red ith
H
a nd
a ppro a ches, H
co n er es fa ster a nd a la s to a n
no nde enera te pa rtitio n hich is o r is nea rb the l
o ba lo pti
.
r
r
o nsider a seto f ecto rs X =
...
to be clstered into
ro ps
o fl
ie da ta . a ch
ℜ is a fea t re ecto r co nsistin o f rea l
- a led ea s re ents describin the fea t res o fthe o b jectrepresented b
. H a rd o r f
clsters o fthe o b jectpo ints ca n be represented b a ha rd/ f
e bership
a tri ca l
l
ed a ha rd/ f
pa rtitio n. he seto fa l
l
no nde enera te ha rd
pa rtitio n a trices is deno ted b
a nd de ned a s
=
=
ℜ
;
;
i
()
here
is a ha rd pa rtitio n a tri. he n ber o fpo ssibl
e ’
s, i.e.
the n ber o f a s o fclsterin
o b jects into no ne pt ro ps,is a tirl
in
n ber o fthe seco nd ind ien b (see [7 ])
(
H
!)
(− )
(2)
.
−
he clsterin o b jectie f nctio n fo r ha rd - ea ns (H
f nctio n
(
)=
·
(v
)
) pa rtitio n is the
(3)
2 4
L . M e n g , Q .H . W u , a n d Z .Z . Y o n g
here = [v v ...v ] is a a tri o fpro to t pe pa ra eters (clster centers)
v
ℜ ; a nd
(v
) is the
cl
idea n dista nce bet een the th fea t re
ecto r
a nd the ith clster pro to t pe v , hich is o f the fo r
(
) =
( − ) ( − ). his o b jectie f nctio n describes the a cc
l
a ted sq a red
erro r hen repl
a cin ea ch fea t re ecto r b the center o f clster to hich it
bel
o n s. H ence, itis a ct a l
l a ea s re o fdisto rtio n.
ini iin the clsterin o b jectie f nctio n ith respectto
l
ea ds to the
fo l
l
o in sets o fclsterin criteria . o r ea ch clster i(i [ ]), rstl,
(4)
v =
i.e.,the clstercenters a re po sitio ned a tthe center o f a ss o fthe fea t re ecto rs
bel
o n in to the clster. he seco nd set o fcriteria tha t ini ies the H
f nctio n sta tes tha t a n o b jectie po int sho l
d be a sso cia ted ith the cl
o sest
clster center:
ifi= in [ (v
)];
=
()
o ther ise.
In eq a tio ns( ) ( ),i= 2 ... deno testhe inde
a clster center a nd = 2 ... deno tes the inde n
H
a do ptsbo thsetso fthese clsterin criteria a nd
f nctio n (
) b al
terna tiel pda tin
a nd
sin
3
c
s r g
n bero fa clstero r
ber o fa n o b jectpo int.
ini iesthe o b jectie
eq a tio ns(4) a nd ( ).
gr
ta rtin
ith a n initia lco nditio n, a
e o les a po p l
a tio n to a rds s ccessiel better re io ns o fthe sea rch spa ce b
ea ns o fra ndo ied pro cesses o f
r
,
,a nd s
. he ien o pti ia tio n pro bl
e de nes
a n en iro n enttha tdel
iersa q a l
it info r a tio n (
ss
s) fo rne sea rch
po ints, a nd the sel
ectio n fa o rs tho se indiid a l
s o fhiher q a l
it to repro d ce
o re o ften tha n o rse indiid a l
s. he reco bina tio n echa nis s a l
l
o s fo r
iin o fpa renta linfo r a tio n hil
e pa ssin itto their o ffsprin ,a nd
ta tio n
intro d ces inno a tio n into the po p l
a tio n.
In brief, a
sea rch fo r the o pti a lindiid a lis t pica l
l i pl
e ented a s
fo l
l
o s.
. enera te a n initia lpo p l
a tio n;
2. a la te the tness a le o fea chindiid a lin the c rrentpo p l
a tio n;
3. el
ectpa irs o fpa rents;
4. enera te o ffsprin o fthe sel
ected pa rents ia reco bina tio n a nd
ta tio n;
. epl
a ce the pa rents ith their o ffsprin a nd crea te a ne
enera tio n;
6. H al
tthe pro cess ifa ter ina tio n criterio n is et. ther ise,pro ceed to step
3.
o red ce the sea rchspa ce dra stica l
l, e intro d ce H
into the t pica l
i pl
e enta tio n pro cess. rin e er
enera tio n, ith tnesse a la tio n
A F a s te r G e n e tic C lu s te rin g A lg o rith m
2 5
ta in pl
a ce,a sin l
eH
rea l
l
o ca tio n step is a ppl
ied to ea chindiid a lin
the po p l
a tio n. cco rdin to the pa rtitio n a tri represented b a n indiid a l
,
the seto fclster centers a re l
o ca ted sin eq a tio n (4) a nd then the pa rtitio n
a tri is pda ted b eq a tio n ( ),i.e.b a ssinin ea cho b jectpo intto itscl
o sest
clster. fter a l
lthe indiid a lpa rtitio ns a re pda ted sin H
,the three
ba sic enetic o pera tio ns - sel
ectio n, reco bina tio n a nd
ta tio n - sta rt.
In o rder to a ppla enetic a ppro a chto a ien pro bl
e ,a n ber o ff nda enta liss es
stbe a ddressed in a d a nce. he a re so ltio n representa tio n,
tness f nctio n, crea tio n o fthe initia lpo p l
a tio n a nd a s cceedin ne
enera tio n, i pl
e enta tio n sche es o f enetic o pera to rs, ter ina tio n criterio n, a nd
the
pa ra etersettin s. he resto fthis sectio n describes ea cho fthese iss es
in deta il
.
.
t
pr s t t
s entio ned in sectio n , ea chindiid a lin the po p l
a tio n is a ha rd pa rtitio n
a tri
. In ha rd clsterin , a n o b jectpo intbel
o n s to the cl
o sest
clstere clsiel. here iso nla do n a n co l n o fa ha rd pa rtitio n a tri.
H ereo f, it is po ssibl
e to si pif a
ha rd pa rtitio n a tri
into a n di ensio na l ecto r u ith the ith el
e entdescribes hich ro the l
ies do n
the ithco l n o fthe o riina l . he po ssibl
e a les o fthe el
e ents o fu ra n e
fro
to . his si pl
i ca tio n is a do pted in the pro po sed H
.
.2
t ss
ct
o al
l
o co pa riso n o fperfo r a nce ith the cl
a ssica lH
a s el
la s a n
e istin enetic clsterin a ppro a ch[3]),H
f nctio n (
) is sed
here a s the o b jectie f nctio n to be ini ied. nd the a tri
is ca l
c l
a ted
ithrespectto
sin eq a tio n (4). o l
a r elred ce the cha nce o f H
beco in st c a ta de enera te pa rtitio n, e ha e ta en the n ber o fe pt
clsters into co nsidera tio n. n a le o bta ined sin H
f nctio n is sca l
ed
ith a pena l
t fa cto r. ifferent fro the o ne in [3], o r o b jectie f nctio n is
rede ned a s fo l
l
o s:
(
)=
(
)
( +
)
(6 )
here is the to ta ln ber o fclsters a nd
deno tes the n ber o f
e pt clsters a nd is e a la ted ia co ntin the a l
l
- ro s in .
he o a l
o fa clsterin a ppro a chisto ini ie the o b jectie f nctio n hil
ea
fa o rs tterindiid a l
s. o co pro ise, e se the in erse o fa n indiid a l
’
s
sca l
ed H
f nctio n a le a sits tness a le.In a dditio n,a l
inea r tnesssca l
in
echa nis [ ] ha s been intro d ced to a inta in rea so na bl
e sel
ectio n press re.
.
t
t
o nsider tha t o o d cho ice fo r sta rtin co n
ra tio ns sho l
d be free o f o ert
bia ses. o r the ha rd pa rtitio n ecto rs u in the initia lpo p l
a tio n, ea ch el
e ent
2 6
L . M e n g , Q .H . W u , a n d Z .Z . Y o n g
p a re n t
0
1
:
th e w o rs t m e m b e r 1 k
o f c lu s te r i
:
:
:
l
:
:
p a re n t
_ 1
2 3 4
8
:
i
:
i
:
4 i’
:
:
0
4 6
1
7 8
:
j 2
:
1 2 5
:
j
:
:
:
k
:
:
:
l
3
:
:
o ffs p rin g
0
1
:
k
:
:
:
l
:
:
2 3 4
8
:
i’
:
i
:
i’
:
:
i. . ra p ica ld scripti
ft r c
s pa rtiti
.
t clst r fpa r t ’
_ 2
th e ra n d o m ly c h o s e n
m e m b e r fro m c lu s te r j
_ 1
m o v e d to c lu s te r i’
i a ti
stra t g
itis a ppl
id t t
is setto a ra ndo l enera ted n ber in the ra n e o f[ ].
do in so , e
a ct a l
l ra ndo l pa rtitio n the o b jectpo ints to initia lclsters.
.4
tc
p r t rs
In e er enera tio n, a
sel
ect pa rents fro the c rrent po p l
a tio n. heo retica l
l, the pro ba bil
it o fo ne indiid a lbein sel
ected is pro po rtio na lto its
tness a le rel
a tie to the o thers’ tness a les. fter bein sel
ected, pa rents
a re a ted to ie birth to their o ffsprin . ffsprin a re enera ted ia the o pera tio ns o freco bina tio n a nd
ta tio n.
.
ct . s to the sel
ectio n o pera to r,the s
s
rs s
sche e is a ppl
ied [9 ]. a sed o n the theo retica la nd e pirica la na lsis, a er
co nclded [9 ] tha tthis sche e is a n o pti a lseq entia lsa pl
in sche e hich,
fo r the rstti e, a ssin o ffsprin a cco rdin to the theo retica lspeci ca tio ns.
. c
t . bra nd-ne reco bina tio n stra te ha sbeen desined
fo r the clsterin pro bl
e s. Itrea l
l
o ca tes the o rst e ber in ea chclster o fa
pa rent’
spa rtitio n a cco rdin to its a tin pa rtner’
spa rtitio n. he o rst e ber
o fa certa in clsteristhe fa rthesto b jectpo intto the centero ftha tclster,a o n
tho se bel
o n in to it.
a rtic l
a rl, fo r the ith clster o fpa rent’
s pa rtitio n u , reco bina tio n is
ca rried o ta s fo l
l
o s:
A F a s te r G e n e tic C lu s te rin g A lg o rith m
2 7
. nd the o rst e ber o fthe ith clster o fu (s ppo se it’
s );
2. a cco rdin to the a tin pa rtner’
s pa rtitio n ecto r u , nd o t to hich
clster this o rst e ber
bel
o n s (s ppo se it’
s the th clster o fu );
3. stil
la cco rdin to u , ra ndo l cho o se a no ther o b ject po int fro the th
clster o fu (s ppo se itha ppens to be );
4. ba c to u , chec o t to hich clster o f u the cho sen o b ject po int
bel
o n s a nd si pl rea l
l
o ca te
to tha tclster.
hil
e stil
l
ho l
din the ra ndo nesspro pert,thisreco bina tio n is el
l ided.
In step 2, the a tin pa rtner’
s pa rtitio n u is referred to a ns er the fo l
l
o in
q estio n: hicho b jectpo ints sho l
d sha re the sa e clster iththe o rst e ber o fa certa in clster o fpa rtitio n u ? ince there is a tl
ea sto ne (itsel
f) hil
e
a be o re tha n o ne o b jectpo ints a a il
a bl
e, step 3 ra ndo l cho o se o ne o f
the . ina l
l,in step 4,the o rst e ber o fpa rtitio n u ’
s ithclster is rea l
l
oca ted to a ne clster s chtha titbel
o n s to the sa e clster a s this ra ndo l
cho sen o b jectpo int. o r better ndersta ndin , these i pl
e enta tio n steps a re
described ra phica l
lin i re .
he reco bina tio n pro cessisa ppl
ied to ea chpa rento fthe a tin pa irclster
b clster. ha tis to sa , ea ch clster is a dj sted independentl o fthe o thers.
c.
t t . fter e er reco bina tio n, ta tio n is i po sed o n ea ch el
eento fthe ne lco nstr cted pa rtitio n ith a
ta tio n pro ba bil
it p .
ta tio n sets the cho sen el
e ents to a ra ndo l enera ted inte er ra n in fro
to
.
.
r t
f
rt
r enetic clsterin a lo rith is a stea d -sta te
, hich repl
a ces o nl a
fra ctio n o f the po p l
a tio n ea ch enera tio n. he o tia tio n o f intro d cin a
stea d -sta te
is to eep a o o d ba l
a nce bet een e pl
o ita tio n o f the best
re io ns fo nd so fa r a nd co ntin ed e pl
o ra tio n fo r po tentia l
lbetter pa o ffa rea .
H o e er,a stea d -sta te
il
l
increa se the a ria nce a l
o n the ro thc r es
o findiid a l
s [ ]. o red ce the a ria nce,a irst-In- irst- t( I ) del
etio n
[ ] is e pl
o ed a nd th s per itthe se o fthis stea d -sta te
ith s a l
l
er
po p l
a tio ns. ith I
del
etio n the po p l
a tio n is si pl a rst-in- rst-o t
q e e ith ne indiid a l
s a dded to o ne end a nd del
eted indiid a l
s re o ed
fro the o ther end.
o reo er,el
itis is i pl
e ented to eep the best e ber o fthe po p l
a tio n.
.6
r
tr
tt gs
he enetic clsterin a ppro a chis co p ta tio na le pensie. he do ina ntco st
is tha to fca l
c l
a tin the tness a le fo r ea ch indiid a lin the po p l
a tio n fo r
e er enera tio n. his co st is a f nctio n o f , , a nd , i.e., the n ber o f
fea t res, n ber o fo b jectpo ints, a nd n ber o fcl
a sses to clster the o b ject
2 8
L . M e n g , Q .H . W u , a n d Z .Z . Y o n g
po ints. n increa se in a n o fthese pa ra etersres l
ts in a l
inea r increa se in ti e
per enera tio n.
ince the to ta l
co p ta tio na l
ti e isa l
so pro po rtio na l
to the po p l
a tio n sie,
s al
lpo p l
a tio ns a re preferred. hro ho tthe e peri ents ith H
,
e sed a po p l
a tio n sie o f3 .
In e er enera tio n,6 % indiid a l
s o fthe po p l
a tio n il
lnder o the three
ba sic enetic o pera tio nsto enera te o ffsprin . he pro ba bil
itieso fs cceedin reco bina tio n a nd
ta tio n a re .9 a nd . ,respectiel[ ]. hese pa ra eter
settin s ha e been fo nd to o ffer bestres l
ts.
he ter ina tio n criterio n sed here is the n ber o f enera tio ns. he a ppro a chsto ps hen the req ired enera tio n is crea ted.
p r
s
s clsterin is o ften a ppl
ied to i a e pro cessin a nd i a es a re rea l
- o rl
d
do a ins o f sini ca nt co pl
e it in ter s o f n ber o f o b ject po ints to be
clstered a nd n ber o f cl
a sses, a ra l
e eli a e q a ntia tio n pro bl
e is
co nsidered. In this a ppl
ica tio n ca se, the 2 6 2 6 bl
a c -a nd- hite ena i a e
is rstl diided e enl into s a l
lbl
o c s o f4 4piel
s. hen the ra l
e el
s
o fthe piel
s in ea ch o fthese bl
o c co po se a ecto r s ch tha tthere a re 49 6
i a e ecto rs o f 6 fea t res ( ra l
e el
s). he o a lo fthis i a e q a ntia tio n
pro bl
e is to clster these 49 6 i a e ecto rs into 2 6 cl
a sses. H ence, = 49 6 ,
= 2 6 ,a nd = 6 .
o this i a e q a ntia tio n pro bl
e , the cl
a ssica lH
, a present enetica l
l ided clsterin a lo rith [3],a nd o r o n H
ha e been
a ppl
ied independentl. nd the o b jectpo intspro cessed b a l
l
o fthese a lo rith s
ere the 49 6 i a e ecto rs co nsistin o f 6 pielra l
e el
s.
he cl
a ssica lH
a s i pl
e ented a s fo l
l
o s:
.
a ndo l pa rtitio n the i a e ecto rs into initia lclsters. a l
c l
a te the
co rrespo ndin seto fclster centers sin eq a tio n (4).
2. q a tio n ( ) is a ppl
ied s ch tha ta cco rdin to its 6 pielra l
e el
s, ea ch
i a e ecto r is a ssined to the cl
o sestclster center. ea n hil
e, the
cl
idea n dista nce bet een ea ch i a e ecto r a nd its cl
o sestclster center is
s
ed to the a cc
l
a ted sq a red erro r .
3. q a tio n (4) is a ppl
ied. a chclster center is pda ted a s the center o f a ss
o fa l
lthe i a e ecto rs bel
o n in to it.
teps 2 a nd 3 a re a ppl
ied a l
terna tiel ntilthe rel
a tie difference in the
a cc
l
a ted sq a red erro r o ft o s ccessie itera tio ns is l
ess tha n = .
−
(i.e. a bs(
)
).
t
ep
2
i
s
repea
t
ed
sin the l
a stset
rr
rr
o fclster centers. he res l
tin a cc
l
a ted sq a red erro r
a nd n ber o f
e pt clsters a re reco rded.
H al
l . desined the
a ppro a chfo l
l
o in the t pica li pl
e enta tio n
steps o fa
a nd the H
as al
so intro d ced in ea ch enera tio n. he
three enetic o pera to rs ere to rna entsel
ectio n,t o -po intreco bina tio n a nd
A F a s te r G e n e tic C lu s te rin g A lg o rith m
2 9
ra ndo
ta tio n. In
, indiid a l
s o fthe po p l
a tio ns ere sets o fclster
centers instea d o fpa rtitio n a trices.
o r co pa riso n, the
a ppro a ch ha s been repea ted. H o e er, there a re
so e differences bet een o r e peri ents a nd theirs a nd these a re hihl
ihted
a s fo l
l
o s:
.
hil
e the ha d sed -fo l
d to rna entsel
ectio n ith = 2, e ha e sed
sto cha stic niersa lsa pl
in sche e, a s in o r o n H H
.
2. hil
e the ha d sed a enera tio na l
ith a n el
itiststra te o fpa ssin
the t o ttestindiid a l
s to the ne t enera tio n, e ha e sed a stea d sta te
ith enera tio n a p o f .6 thro ho t. l
itis is i pl
e ented to
eep the ttestindiid a l
.
3. hil
e the ha d sed a bina r ra co de representa tio n fo r the indiid a l
s,
e ha e sed rea la le representa tio n.
ie tho se sed thro ho tthe e peri ents fo r o r enetic clsterin a lo rith , e ha e cho sen the sa e pa ra eter settin s in a l
lthe repea ted
a ppro a ches. he
ere a po p l
a tio n sie o f3 , a reco bina tio n pro ba bil
it o f
.9 , a nd a
ta tio n pro ba bil
it o f . . cepttha tbetter so ltio n ca n be
o bta ined ith l
a r er po p l
a tio n sie, these a le o ffer
the best perfo ra nce [3]. er a ppro a chsto pped a tthe 4th enera tio n.
s
s
ests o n clsterin i a e ena b the cl
a ssica lH
, the
a ppro a ch,
a nd o r H
a re nderta en respectiel. he e
ea s resfo rco pa rin
a lo rith perfo r a ncesa re the ea n sq a red erro r(
) a nd n bero fe pt
clsters, , a sso cia ted ith a pa rtitio n a tri. he ea n sq a red erro r is the
a le o fthe H
f nctio n
a era ed b the n ber o fi a e piel
s(
=
(2 6 2 6 )), hichindica tes the disto rtio n bet een the q a ntied i a e a nd
the o riina li a e. rin ea ch r n, the pa rtitio n ith the l
o esta cc
l
a ted
sq a red erro r is tra ced a nd the
a nd
a les a sso cia ted ith the o pti a l
pa rtitio n na l
l fo nd a re reco rded.
he reco rded a leso f
a nd a re a era ed o erthe to ta l
n bero fr ns,
i.e.
,4 a nd 4 fo r H
,
a nd H
respectiel. he a era e
a les a nd the sta nda rd de ia tio ns o f
a nd a re repo rted in a bl
e . he
a nd
a les a sso cia ted iththe er bestpa rtitio n e er o bta ined b ea ch
a lo rith a re a l
so inclded.
i re 2 sho sthe histo ra o fthe reco rded
a nd
a les fo rH
,
a nd H
r ns, respectiel. he histo ra is a sta tistica lrepresenta tio n o fthe distrib tio n o fl
o ca lo pti a lo bta ined b the a lo rith . ddin
res l
ts o f o re r ns did no tcha n e the distrib tio n f rther.
ro
a bl
e a nd i re 2, the fo l
l
o in o bser a tio ns a re dra n:
. U sin H
, a bro a d distrib tio n o fl
o ca lo pti a lis o bser ed. he sta nda rd de ia tio ns o f
a nd a re
ch hiher tha n tho se o fthe enetic
3 0
L . M e n g , Q .H . W u , a n d Z .Z . Y o n g
.
s l
ts f r tria l
s it t r
a ra g
st. d . f
ft
st
a ra g
st. d . f
ft
st
clst ri g a l
g rit
s
3 .3 7
.5 35 7 7 .7 27 5
5.36 35 3. 2
.26 7
25. 7 3 7 6 . 357 7 4. 5
4.
3. 5
6 .5 7
. 4
6
5
a ppro a ches. his indica tes tha tcl
a ssica lH
is sensitie to the initia l
distrib tio n o fclster centers a nd ea sil tra pped in l
o ca lo pti a , hil
e, o n
the o ther ha nd, sho s the ia bil
it o fa enetic a ppro a chto o erco e these
pro bl
e s.
2.
o th enetic ided clsterin a lo rith s o tperfo r the cl
a ssica lH
in the sense tha tthe end p in pa rtitio nsa sso cia ted ith chl
o erdisto rtio n a nd sini ca ntfe er e pt clsters. a rtitio ns o fsi il
ar
a les
ere repea tedl fo nd b
a nd H
. his indica tes tha tthese
res l
ts a re indeed nea rlo pti a l
.
3. In no ca se did H
res l
tin a de enera te pa rtitio n. n a era e, the
best
a le o bta ined b
H
is sl
ihtl l
o er tha n tha to bta ined
b
. o is fo r the er best
e er fo nd.
rin ea ch enetica l
l ided clsterin a ppro a ch sin either
or
H
, the
a le a nd e pt clster n ber o f the best pa rtitio n
fo nd p to a nd incldin ea ch s ccessie enera tio n a re reco rded a nd then
a era ed o er the to ta ln ber o fr ns. i re 3 sho s the res l
ta nta era e
a les o f
a nd
ith respectto the enera tio n n ber fo r bo th
a nd
H
. cco rdin to the t o
c r es, e see tha t, fo r bo th enetic
a ppro a ches, the initia lco n er ence ra tes a re er hih a nd a s the enera tio ns
pro ress co n er ence ra te decrea ses ra pidl. H o e er,in the ea rl
ier enera tio ns,
H
co n er es
ch fa ster tha n
a nd q ic l rea ch the desired reio n here the no nde enera te pa rtitio n a trices reside. H
ca n nd a n
a le l
o ertha n the l
o este erfo nd b H
a fterthe th enera tio n,
hil
e
do es ita fter enera tio n . s sta ted,fo r clsterin pro bl
e s the tness e a la tio n a te er
enera tio n is ti e co ns in . l
tho h
so o n
ca tches p a nd fro
enera tio n
the difference in
a les bet een t o
enetic clsterin a lo rith s is o itta bl
e, in the specia lca ses here the speed
a s el
la s perfo r a nce is req ired H
a pro ide a
ch fa ster a
to nd a n a ccepta bl
e so ltio n. rther o re, a s H
a ss res l
o ca lo pti a l
it a nd d e to its hil
l
-cl
i bin
etho d co n er es
ch fa ster tha n a n enetic
a ppro a ch, instea d o f a itin fo r the enetic a ppro a ches rea ch a n e a cto pti a l
so ltio n, e a sto p enetic sea rch a fter a necessa r n ber o f enera tio ns
a nd se H
to nd the co rrespo ndin l
o ca lo pti
.
A F a s te r G e n e tic C lu s te rin g A lg o rith m
6
3 1
c s
In this pa per a no elenetic clsterin a lo rith is pro po sed, hich co bines
a enetic a lo rith ( ) ith the cl
a ssica lha rd - ea ns clsterin a lo rith
(H
). U nl
ie o ther clsterin a lo rith s, H
pro cesses pa rtitio n
a trices ra ther tha n sets o fcenter po ints a nd th s a l
l
o s a ne i pl
e enta tio n
sche e fo r the enetic o pera to r -reco bina tio n. o r co pa riso n o fperfo r a nce
ith o ther present clsterin a lo rith s, e peri ents o n a ra -l
e eli a e
q a ntia tio n pro bl
e ha e been co nd cted. he res l
ts sho tha t H
co n er es
ch q ic er to the l
o ba lo pti
a nd pro ides a ia bl
e a to
so le the dil
e
a here the cl
a ssica lH
is fo nd ea sil ca htin l
o ca l
o pti a a nd a enetic a ppro a chreq ires l
a r e ti e co ns ptio n.
fr c s
. .
2.
3.
4.
5.
6.
7.
.
.
.
.
.
a d . .
al,
r
g ii
ri i s. ddis - sl ,
a di g, a ssa c s tts, 7 4.
. c
d rs,
g tic l d- a i a g q a tia ti a l
g rit ,”
r
g ii
., l
. 7 , pp. 547 -556 ,
6.
. . al
l
, I. .
rta d . .
d k, lst ri g it a g tica l
l g id d
pti i d a ppr a c ,”
r s.
i r
m
i , l
. 3, . 2, pp.
3- 2,
.
. c
d rs,
g tic - a s clst ri g a l
g rit
a ppl
id t c lr i a g
q a tia ti ,”
r
g i., l
. 3 , . 6 , pp. 5 - 6 6 ,
7.
. Kl
a
,
clst ri g it
lti a r a l
g rit s,” i r .
h
r
grss, l
. 2, pp. 3 2-323,
7.
. . a
a d . .
rt, lst ri g it
lti a r stra t gis,”
r
g i., l
. 27 , . 2, pp. 32 -32 ,
4.
. ra
it a d I. . t g , ds.,
f
hmi
i s, . .
pa rt
t f
rc , a ti a l r a
f ta da rds ppl
id a t
a tica l
ris. 55, 6 4.
. . l
d rg,
i
grihms i
rh
imi i
hi
ri g. ddis - sl
l
is i g
pa ,I c,
.
. . a k r,
d ci g ia s a d i ffici c i t s lcti a l
g rit ,” r .
.
f.
i
grihms, pp. 4-2 ,
7.
K. .
ga d . a r a ,
ra ti ga ps r isit d,”
i s f
i
grihms 2 , .
itl ( d.), pp. -2 . a il
,
: rga Ka f a ,
3.
. . c ra d l
p a d . K. l ,
a ic pa ra t r c di g f r g tic
al
g rit s,”
hi
r i g, l
. , . , pp. -2 ,
2.
3 2
L . M e n g , Q .H . W u , a n d Z .Z . Y o n g
7
no. of occurences among 100 runs
no. of occurences among 100 runs
3
2.5
2
1.5
1
0.5
0
60
80
100
120
140
160
6
5
4
3
2
1
0
-50
0
50
100
no. of empty clusters
150
200
0
50
100
no. of empty clusters
150
200
0
50
100
no. of empty clusters
150
200
MSE
(a )
12
no. of occurences among 40 runs
no. of occurences among 40 runs
3
2.5
2
1.5
1
0.5
0
60
80
100
120
140
10
8
6
4
2
0
-50
160
MSE
4
40
3.5
35
no. of occurences among 40 runs
no. of occurences among 40 runs
()
3
2.5
2
1.5
1
0.5
0
60
80
100
120
140
160
MSE
30
25
20
15
10
5
0
-50
(c)
i. . istri ti
ft
a sq a r
it t pti a l
pa rtiti sfi a l
lf d
a d
, r sp cti l.
rr r a d
pt clst r
clst ri gt
a i ag
r a ss cia t d
it
,
A F a s te r G e n e tic C lu s te rin g A lg o rith m
empty cluster number of the best partition
600
MSE of the best partition
500
400
300
200
100
0
0
10
20
30
generation
40
50
40
30
20
10
0
0
50
3 3
10
20
30
generation
40
50
10
20
30
generation
40
50
(a )
empty cluster number of the best partition
600
MSE of the best partition
500
400
300
200
100
0
0
10
20
30
generation
40
50
25
20
15
10
5
0
0
()
i. .
r
rg c pr p rtis f
a d
it r sp ctt t
g
ra ti
Scene Interpretation using Semantic Nets and
Evolutionary Computation?
D. Prabhu1, B. P. Buckles2 , and
F. E. Petry2
i2 Technologies, 1603 LBJ Freeway, Suite 780, Dallas TX 75234, USA and
Department of Electrical Engineering & Computer Science, Tulane University, New
Orleans, LA 70118, USA
1
2
Abstract. The tness function used in a GA must be measurable over
the representation of the solution by means of a computable function.
Often, the tness is an estimation of the nearness to an ideal solution or
the distance from a default solution. In image scene interpretation, the
solution takes the form of a set of labels corresponding to the components
of an image and its tness is dicult to conceptualize in terms of distance
from a default or nearness to an ideal. Here we describe a model in
which a semantic net is used to capture the salient properties of an ideal
labeling. Instantiating the nodes of the semantic net with the labels from
a candidate solution a chromosome provides a basis for estimating a
logical distance from a norm. This domain-independent model can be
applied to a broad range of scene-based image analysis tasks.
1
Introduction
We describe how a genetic algorithm GA can be employed to classifylabel
objects in a scene for which no prior truth data exists. Relationships among the
objects in a typical scene from the domain of discourse are encapsulated within a
semantic net. The method was validated using a test suite of images captured by
satellites. Mainly, these included the infrared band of North Atlantic scenes and
two bands of AVHRR data depicting regions of the Western U.S. The objective
of the North Atlantic image analysis was to identify currents such as the Gulf
Stream and eddies. The objective of using AVHRR images was to detect and
identify clouds by type.
2 Background
Classication or labeling of segments is the focus of this paper. Labeling a segment of an image is a particularly di cult subtask because there must be an
automatic method of assigning a gure of merit to a candidate solution. A very
?
This work was supported in part by a grant from NASAGoddard Space Flight
Center, NAG5-8570 and in part by DoD EPSCoR and the State of Louisiana
under grant F49620-98-1-0351.
S . C a g n o n i e t a l. ( E d s .) : E v o W o r k s h o p s 2 0 0 0 , L N C S 1 8 0 3 , p p . 3 4 − 4 3 , 2 0 0 0 .
S p rin g e r-V e rla g B e rlin H e id e lb e rg 2 0 0 0
S c e n e In te rp re ta tio n
3 5
general and computationally reasonable method for dening tness function for
image labeling is described that is based upon developing a semantic net for a
typical scene from the domain. The classes of the semantic net are dened by the
labels to be assigned to the segments. The predicates are relationships that exist
among objects bearing the corresponding labels. Given a candidate solution a
labeling for an image, a measure is described that estimates the conformance
of the labeling to the semantic net.
3 Methodology
In this section, the formal approach and extensions to our prior ad hoc model
using semantic nets. 1 is developed. A semantic network is a structure used to
represent knowledge as a combination of nodes interconnected with arcs.
3.1 Description of Scene Properties with Semantic Nets
Let CL be the set of possible classication labels or classes specic to the
application.
CL = fc1 c2 : : : cm g
1
Let S be the set of segments components in the given image:
S = fs1 s2 : : : sn g
2
Each of these components is assumed to be completely characterized by T , the
set of features specic to the application. Let,
T = ft1 t2 : : : th g
3
For any given image component, each of these features, ti , takes values from
some corresponding domain Di .
Assume that every statement about such application-specic knowledge that
is useful for the classication task can be described by using at most kmax number
of components drawn from the image. First, consider only those statements which
involve exactly k for some 1 k kmax components of the image. Let Ck be
the set of all k-tuples built from the indices of the elements of CL. i.e.,
Ck = f hj1 : : : jk i j 1 j1 : : : jk m g
4
Every image component s 2 S can be an instance of only one of the m classes
drawn from the set CL. Denote this by ISA, an instance function which maps
each image component to the index of a class.
ISA : S ,! f1 : : : mg
5
Thus, given a k-tuple of image components segments, ISA can be used to generate a corresponding k-tuple hj1 : : : jk i 2 Ck . Let Fk be a set of feature-value
3 6
D . P r a b h u , B .P . B u c k le s , a n d F .E . P e tr y
comparator functions such that each function f 2 Fk maps a k-tuple of feature
values drawn from the domain Di corresponding to some feature ti 2 T , to an
absolute comparison value.
Fk = f f j f : Dik ,! R g
6
Let Pk be the set of predicates such that each p 2 Pk is a fuzzy predicate mapping
a given absolute comparison value to a fuzzy truth-value in the continuous range
of 0 1 .
Pk = f p j p : R ,! 0 1 g
7
Predicates based on fuzzy logic are more appropriate than binary predicates for
this task in view of the heuristic knowledge employed. In addition, null 2 Pk ,
and nullx = 0 for all x.
For every k-tuple of image components, every feature-value based relationship inherited from the corresponding k-tuple of classes hj1 : : : jk i 2 Ck
among its elements can be quantied by using some specic feature-value comparator function and a corresponding predicate. For a given k-tuple of classes,
hj1 : : : jk i 2 Ck , there may exist multiple relationships among the elements of
the k-tuple of class-instances. Therefore, we dene a mapping, R1k such that
every given tuple of classes denotes a specic subset of the available predicates,
i.e.,
R1k : Ck ,! 2Pk
8
We also dene another functional mapping R2k such that every predicate is
mapped to a feature comparator function. i.e.,
R2k : Pk ,! Fk
9
Thus, every relationship involving k components can be uniquely denoted by the
tuple hhj1 : : : jk i pi where, hj1 : : : jk i 2 Ck and p 2 R1k hj1 : : : jk i . Consider a k-tuple of image componentssegments that is instantiated by virtue of
some ISA mapping to be a tuple of class-instances corresponding to hj1 : : : jk i.
Every relationship hhj1 : : : jk i pi inherited by the image components can be
quantied as follows: Suppose that f = R2k p and that f makes use of the feature ti 2 T using k-tuple of values from the domain Di . Now, this relationship
can be quantied by the composition pf x1 : : : xk where x1 : : : xk 2 Di are
the feature values of the image segments.
The set of all such relationships constitutes a semantic net, SNk . In other
words,
SNk = f hhj1 : : : jk i pi j hj1 : : : jk i 2 Ck ^ p 2 R1k hj1 : : : jk i g 10
This formal model for representing and utilizing the knowledge relating k elements can be summarized by FNk , a tness net.
FNk = hCL S T ISA Ck Pk R1k Fk R2k SNk Wk E i
11
where, Wk is a set of weights, one for each predicate p 2 Pk , and E is a tness
function described in section 3.3.
S c e n e In te rp re ta tio n
3 7
This model can now be generalized to represent all possible statements from
the knowledge base, i.e., for all values of k, 1 k kmax . Such a complete
model is given by:
FN = hCL S T ISA C P R1 F R2 SN W E i
12
where,
S
S
P = S1 k kmax Pk
C = S 1 k kmax Ck
R1 = S1 k kmax R1k F = S 1 k kmax Fk
R2 = S1 k kmax R2k SN = 1 k kmax SNk
W = 1 k kmax Wk
To illustrate the notions discussed above, consider a simple domain such as a
chair shown in Figure 1. Here we have, CL = fs b a lg, where the symbols de-
s4
s7
s1
s2
s6
s8
s3
s5
Fig. 1.
A chair and its segments
note, and index, the classes Seat", Back", Arm", and Leg" respectively. Also,
S = fs1 s2 s3 s4 s5 s6 s7 s8g from Figure 1. There are two measuresfeatures
for each segment: angle of rotation from the horizontal, t1 , and surface area, t2 .
While relationships of any degree except zero are permitted, for the sake of
simplicity we assume only binary relationships for this example, i.e., we consider
only the case of k = 2 and C = C2 . The set of feature comparator functions is
given by F = F2 = ff1 f2 g. f1 takes the angle-measures of two segments as
arguments and computes their relative orientation. f2 computes the dierence
in the surface area of the given argument pair of segments. Further, let the set
of predicates be P = P2 = fperpendicular-to parallel-to area-greater-than nullg.
The predicates perpendicular-to and parallel-to return 1 if the segments are mutually perpendicular and parallel respectively. Otherwise, they return 0. Similarly, the predicate area-greater-than returns 1, if the dierence computed by f2
is positive and 0 otherwise. Obviously, fuzzy predicates can be used instead to
assign values in the range 0 1. The mapping from class tuples to predicates,
R1 = R12, is simplied in this case since no class pair has more than one relationship. This is graphically shown in the top-half of Figure 2. The functional
mapping, R2 = R22 , can be constructed easily since predicates perpendicular-to
3 8
D . P r a b h u , B .P . B u c k le s , a n d F .E . P e tr y
and parallel-to use the function f1 and the predicate area-greater-than uses the
function f2 . In this simple example, the semantic net SN = SN2 closely correα >
α >
L e g e n d
:
p e r p e n d ic u la r -to
:
B a c k
S 1
Fig. 2.
a chair.
S e a t
S 2
S 3
L e g
S 4
S 5
A rm
S 6
α >
S 7
:
p a r a lle l-to
a r e a -g r e a te r -th a n
S 8
A sample semantic net and a candidate ISA mapping for a scene consisting of
sponds to R12 and the Figure 2 shows the useful parts of the semantic net for a
typical chair using the three predicates. Figure 2 also shows a particular instantiation, i.e., an ISA mapping, of the segments from the scene shown in Figure 1.
This mapping of fhs1,si hs2,ai hs3,li hs4,bi hs5,li hs6,li hs7,ai hs8,lig results
in the correct classication of the segments. Also, it is obvious that any other assignment of labels to the segments would result in a lower consistency evaluation.
Formally, the semantic net shown in Figure 2 is given by
SN = f hhb ai i hhb li nulli hhb si ?i hhb bi nulli hhs ai i
hhS Li ?i hhs si nulli hhl ai ?i hhl li ==i hha ai nulli g
where, the symbols ", ?", and ==" denote the predicates area-greater-than,
perpendicular-to
, and parallel-to respectively.
3.2 Representing Candidate Solutions in GAs
A candidate solution for the classication task takes the form of a vector of
indices hj1 : : : j : : : j i, containing one element for each segment s 2 S . This
vector of indices represents a possible ISA mapping for the segments in the set S .
In other words, ISAs = j where, 1 i n and 1 j m. For example, the
vector hs a l b l l a li represents the labeling shown in Figure 2. Similarly,
the vector hl a s b a l l li represents another candidate solution, albeit of
inferior quality.
i
n
i
i
i
i
S c e n e In te rp re ta tio n
3 9
3.3 Computing Fitness Using Semantic Net Description
Fitness is a quantitative measure of the consistency of an ISA relationship that a
candidate solution represents. The procedure described below may appear to be
computationally expensive. However, we have found that, in practice, there are
very few relationships involving more than two componentssegments and that a
large number of high-order relationships are reduced to null predicates. Further,
all the predicates can be precomputed for the segments in a given image and
the repeated tness computations need only do the summation of the various
predicate values using a table look-up.
E=
X X Xw
max
k
k
pf x1 : : : x
c
p
13
k
=1 hh k i i2 k h k i
s
SN
where hc i abbreviates hj1 : : : j i 2 C and hs i represents any k-tuple of image
segments instantiated to hj1 : : : j i via the ISA mapping. Also, f = R2 p ,
w 2 W is the weight corresponding to predicate p, and x1 : : : x 2 D are
the feature values of the image segments in hs i corresponding to some feature
t 2 T , depending on the feature comparator function f . For domains in which
knowledge is unevenly distributed in the semantic net, practice may dictate that
normalization over the set of predicates for each class or class-tuple be performed.
To illustrate tness, examine the best solution hs a l b l l a li taken from
Figure 2. Since there are four and two instances of the classes l and a respectively, we need to sum twenty-three predicate values, ignoring the null predicate.
Assuming unit weights for all predicates, on summation, the tness value for the
ideal solution can be seen to be twenty three, since all the predicates have a
value of 1 in this case. In contrast, consider the obviously sub-optimal solution
hl a s b a l l li. Here, only ten out of the twenty three predicate values have
a value of 1 resulting in a tness value of ten.
k
k
k
k
k
k
k
k
i
k
i
4 Experiments and Analysis
The goal in the oceanographic problem is to label the mesoscale features of the
North Atlantic from satellite images using a known set of classes. The classes
in this case are Gulf Stream North Wall"n , Gulf Stream South Wall"s ,
Warm Eddy"w , Cold Eddy"c , and Other"o , i.e., CL = fn s w c og.
We use edge-segmented images of the region in the infrared band 10.3-11.3
m . An infrared satellite image and its companion segmented image are shown
in Figures 3 and 4. For the image shown, we have, S = fs1 s2 : : : s35g. We
use only two measures for each segment in the image its position and length.
They are computed by using coordinates of the centroid based on the mass of
the segment and the two end-points.
The set of predicates P and the set of functions F are informally shown in
Table 1. The semantic net SN for the domain is shown in Table 2. The mappings
R1 and R2 are implicit in these tables. We use unit weights for all predicates.
4 0
D . P r a b h u , B .P . B u c k le s , a n d F .E . P e tr y
Fig. 3.
Table 1.
Original Infrared Image of the Gulf Stream
Description of predicates and functions for oceanic labeling
Predicate
Is-North-Ofi,j
Function
Comments
If AvgLati AvgLatj , Segment i is north of segment j
= 1 Otherwise, = 0.
Is-Neari,j
exp, X
X is distance between
segments i and j .
Is-Not-Neari,j
1 , Is-Neari j
Fuzzy complement of Is-Near.
Is-North-Of-And min f Is-North-Ofi j Segment i is north of segment j
-Fifty-Km-Fromi,j exp, jX , 50j g and is 50 km from it.
Arcs-Of-Circle is estimated based on
Arcs-Of-Circle-And min f Arcs-Of-Circlei j intersection of cords from segments.
-Less-Than-Hundred Less-Than-Hundred- Second predicate is computed as
-Km-Distanti,j
Km-Distanti j g
= 1, if X 100, and
= exp, jX , 100j, otherwise
ij
ij
ij
ij
ij
Also, it is noted that the predicates are computed a priori for all the segment
pairs and stored in a lookup table.
Candidate solutions are represented as vectors of labels. For the image shown
in Figure 4, any label vector h 1 2
2f
g for all
35 i such that
c c :::c
ci
n s w c o
S c e n e In te rp re ta tio n
Table 2. A semantic net for oceanic
4 1
segment labeling
Class tuple Predicate Name
h
i
Is-Neari,j
h
i
Arcs-Of-Circle-And
-Less-Than-Hundred-Km-Distanti,j
h
i
Is-North-Ofi,j
h
i
Is-Neari,j
h
i
Is-Neari,j
h
i
Is-North-Of-And-Fifty-Km-Fromi,j
h
i
Is-Neari,j
h
i
Is-North-Ofi,j
h
i
Is-Neari,j
h
i
Is-Neari,j
h
i
Arcs-Of-Circle-And
-Less-Than-Hundred-Km-Distanti,j
h
i
Is-Not-Neari,j
h
i
Is-Not-Neari,j
h
i
Is-Not-Neari,j
h
i
Is-Not-Neari,j
h
i
Is-Neari,j
Other tuples null
Legend:
w = Warm Eddy c = Cold Eddy o = Other
n = North Wall of Gulf Stream
s = South Wall of Gulf Stream
w w
w w
w n
w n
n n
n s
s s
s c
s c
c c
c c
o w
o n
o s
o c
o o
i, constitutes a feasible candidate solution. Such label vectors are encoded as
bit strings suitable for GA search. Table 3 shows the parameters used for the
GA runs. Each run with these settings was repeated 10 times, each starting
with a dierent initial random population. The accuracy of the best solution
generated by the GA in each run with respect to the fairly dicult image shown
in Figures 3 and 4 is compared with that of manual labeling and is listed in
Table 4. Figure 4 shows the best labeling obtained over all the runs.
5 Conclusions
Here we describe a domain-independent framework for labeling image segments
for scene interpretation. This approach is based on the abstract representation
of a typical scene from the domain of discourse. The abstraction form, i.e.,
semantic network, permits encoding the descriptions of relationships of arbitrary
degree among the instances of scene objects. A GA is used in searching the space
of candidate solutions for the best labeling. Fitness of a candidate solution is
4 2
D . P r a b h u , B .P . B u c k le s , a n d F .E . P e tr y
Table 3. Parameters of GA runs for oceanic
labeling
Description
Population size
Number of generations
Selection operator
Value
200
200
Proportional selection using stochastic
remainder sampling with replacement
Crossover operator
Uniform crossover allele level
Probability of crossover 0.600
Mutation operator
Bit mutation
Probability of mutation 0.005
Table 4. Accuracy of GA-generated oceanic
Run
1
2
3
4
5
Accuracy
80
57
66
83
83
labeling
Run Accuracy
6
63
7
71
8
77
9
71
10
69
estimated by evaluating the conformance of the solution to the relationships
depicted in the semantic net.
References
1. C. A. Ankenbrandt, B. P. Buckles, and F. E. Petry, Scene recognition using
genetic algorithms with semantic nets", Pattern Recognition Letters, vol. 11, no.
4, pp. 285293, 1990.
2. B. P. Buckles and F. E. Petry, Eds., Genetic Algorithms, IEEE Computer Society
Press, 1992.
3. B. Bhanu, S. Lee, and J. Ming, Self-optimizing image segmentation system using
a genetic algorithm", in Proceedings of the Fourth International Conference on
Genetic Algorithms, R.K. Belew and L.B. Booker, Eds., San Mateo, CA, 1991, pp.
362369, Morgan Kaufmann.
4. S. M. Bhandarkar and H. Zhang, Image segmentation using evolutionary computation", IEEE Trans. on Evolutionary Computation, vol. 3, no. 1, pp. 121, apr
1999.
5. R. Tonjes, S. Growe, J. Buckner, and C.-E. Liedtke, Knowledge-based interpretation of remote sensing images using semantic nets", Photogrammetric Engineering
& Remote Sensing, vol. 65, no. 7, pp. 811821, jul 1999.
6. J. Bala, K. DeJong, and P. Pachowicz, Using genetic algorithms to improve the
performance of classication rules produced by symbolic inductive methods", in
S c e n e In te rp re ta tio n
Fig. 4.
4 3
Best Labeling of the Gulf Stream found by the GA
Proceedings of 6th International Symposium Methodologies for Intelligent Systems
ISMIS'91, Z. W. Ras and M. Zemankova, Eds., Charlotte, NC, 16-19 Oct 1991,
pp. 286 295, Springer-Verlag, Berlin, Germany.
7. S. Truve, Using a genetic algorithm to solve constraint satisfaction problems
generated by an image interpreter", in Theory and Applications of Image Analysis. Selected Papers from the 7th Scandinavian Conference, Aalborg, Denmark,
P. Johansen and S. Olsen, Eds. Aug, 13-16 1991, pp. 133 147, World Scientic.
8. A. Hill and C. J. Taylor, Model-based image interpretation using genetic algorithms", Image and Vision Computing, vol. 10, no. 5, pp. 295 300, Jun 1992.
9. D. B. Fogel, Evolutionary programming for voice feature analysis", in Proceedings
of 23rd Asilomar Conference on Signals, Systems, and Computers, oct 1989, pp.
381 383.
Evolutionary Wavelet Bases in Signal Spaces
Adelino R. Ferreira da Silva
Universidade Nova de Lisboa,
Dept. de Eng. Electrotecnica,
2825 Monte de Caparica, Portugal
afs@mail.fct.unl.pt
Abstract. We introduce a test environment based on the optimization
of signals approximated in function spaces in order to compare the performance of di erent evolutionary algorithms. An evolutionary algorithm to
optimize signal representations by adaptively choosing a basis depending
on the signal is presented. We show how evolutionary algorithms can be
exploited to search larger waveform dictionaries for best basis selection
than those considered in current standard approaches.
1 Introduction
In order to facilitate an empirical comparison of the performance of dierent
evolutionary algorithms a test environment must be provided. Traditionally, sets
of test functions with specic topological properties, commonly known as tness
landscapes, have been proposed by several authors to be used in performance
benchmarking. In particular, the De Jong's test function set has been a standard
for genetic algorithm benchmarks since 1975. In most cases, the optimization
objective is formulated as a global function minimization problem.
In this paper, we depart from this view by considering the optimization of
functions approximated in function spaces. Series expansions of continuous-time
signals go back at least to Fourier's original expansion of periodic functions. A
basis is a set of linearly independent functions that can be used to produce all
admissible functions f t . The idea of representing a signal as a sum of elementary basis functions, or equivalently to nd orthogonal bases for certain function
spaces, is very powerful. However, classic approaches have limitations, in particular there are no "good" local Fourier series that have both time and frequency
localization. An alternative is the construction of wavelet bases, which use scaling instead of modulation in order to obtain an orthonormal basis for L2 R . An
entropy-based algorithm for best basis selection has been proposed in the literature 6. Under the specic conditions of its application, the standard best basis
SBB algorithm nds the optimum basis decomposition according to a specied cost functional. We show that this algorithm can be used to benchmark
evolutionary algorithms.
A second objective of this paper, is to show how evolutionary algorithms
can be exploited to search larger waveform dictionaries for best basis selection
than those considered in current standard approaches. We extend the scope of
S . C a g n o n i e t a l. ( E d s .) : E v o W o r k s h o p s 2 0 0 0 , L N C S 1 8 0 3 , p p . 4 4 − 5 3 , 2 0 0 0 .
S p rin g e r-V e rla g B e rlin H e id e lb e rg 2 0 0 0
E v o lu tio n a ry W a v e le t B a s e s in S ig n a l S p a c e s
4 5
the SBB algorithm by searching larger waveform dictionaries in order to nd
better adaptive signal representations. In Sect. 3, we present an evolutionary
algorithm for best basis selection. Adapted waveform analysis uses libraries of
orthonormal basis and an ecient functional to match a basis to a given signal
or family of signals. Two often used libraries are wavelet-packets and localized
trigonometric functions, since they support the expansion of the waveforms in
orthonormal basis whose elements have good time-frequency localization properties. These libraries constitute huge collections of basis from which we can
pick and choose the best matching basis. Flexible decompositions are important
for representing time-frequency atoms whose time-frequency localizations vary
widely. In this article, we propose the use of evolutionary algorithms 10 as the
main searching tool for best basis selection. The proposed approach generates
a population of solutions based on basis expansions of multi-lter, time-shifted,
wavelet-packet libraries. An evolutionary algorithm operates on the population
to evolve the best solution according to a given objective function. Libraries of
bases represent the population from which we want to select the best-t individuals. This optimization approach permits more exibility in searching for best
basis representations than traditional approaches.
2 Wavelet Spaces
In this section, we brie y review the framework of wavelet basis selection. Let x 2
R be an input signal and let 0 0 = R represent the signal space and B0 0 =
e1 : : : e be the standard basis of R . Wavelet packets split this original space
into two mutually orthogonal subspaces smoothly and recursively, i.e.,
n
n
n
n
= +1 2 +1 2 +1
1
for j = 0 1 : : : J k = 0 : : : 2 , 1, and J log2 n is the maximum i level of recursions specied by the user. Here, we have n = dim = n=2 . The wavelet
jk
j
k
j
k
j
j
j:
j
packet transforms, recursively split the frequency domain via the so-called conjugate quadrature lters. These splits naturally generate a set of subspaces with
the binary tree structure. Let = j k be an index to specify a node i.e., a
subspace with its basis set of this binary tree. The index j species the depth
of the binary tree this is an index of the width of frequency bands for wavelet
packets. The index k species the location of the frequency bands for wavelet
packets. Let = be such a collection of subspaces and let B = fB g be the
corresponding set of basis vectors where B = ! 1 : : : ! j is a set of basis
vectors that spans . Each basis vector in B is called a time-frequency atom,
and the whole set B is referred to as a time-frequency dictionary or a dictionary
of orthonormal bases. These dictionaries contain many orthonormal bases. If the
depth of the tree is J , each dictionary contains more than 22 J , dierent bases.
An important question is how to select from a large number of bases in the
dictionary a basis which performs "best" for one's task. In order to measure the
performance of each basis, we need a measure of eciency or tness of a basis for
n
1
4 6
A .R . F e r r e ir a d a S ilv a
the task at hand. For this purpose, several so-called information cost functionals
have been proposed. A commonly used information cost functional is entropy.
The entropy of a vector = f g is dened by,
d
H = ,
d
Xj
d k
k
k k2 j
dk
d
2
log2
j k kk j2
2
d
d
:
2
This cost functional was used as the objective function to drive the evolutionary
optimization approach outlined in Sect. 3. The goal of the optimization approach
is to nd an optimal, or quasi-optimal in some sense basis representation for a
given dataset.
The above considerations are the fundamentals behind the SBB algorithm
6 . However, it is possible to extend the library of bases in which the best
representations are searched for, by introducing additional degrees of freedom
that adjust the time-localization of the basis functions 2 . One such extension
is the shift-invariant wavelet packet transform. Actually, one well-known disadvantage of the discrete wavelet and wavelet packet transforms is the lack of shift
invariance. The added dimension in the case of shift-invariant decompositions
is a relative shift, between a given parent-node and its relative children nodes.
Shifted versions of these transforms for a given input signal, represent new bases
to be added to the library of bases, which may further improve our ability to
nd the "best" adapted basis. These modications of the wavelet transform and
wavelet packet decompositions lead to orthonormal best-basis representations
which are shift-invariant and characterized by lower information cost functionals 5 . Wavelet packet trees may be extended as joint wavelet packet trees to
prot from enlarged libraries of bases, thus increasing our chances of getting truly
adapted waveform representations. However, enlarged search spaces entail combinatorial explosion problems. We rely on evolutionary optimization approaches
to guide us on the search process.
3 Evolutionary Formulation
The two major major steps in applying any heuristic search algorithm to a particular problem are the specication of the representation and the evaluation
tness function. When dening an evolutionary algorithm one needs to choose
its components, such as mutation and recombination that suit the representation, selection mechanisms for selecting parents and survivors, and an initial
population. Each of these components have parameter values which determine
whether the algorithm will nd an near-optimum solution and whether it will
nd a solution eciently.
3.1 Representation
In the work reported here, a variable length integer sequence is used as the basic
genotype. The objective is to evolve wavelet decomposition trees through the
E v o lu tio n a ry W a v e le t B a s e s in S ig n a l S p a c e s
4 7
evolution of genetic sequences. The technique used to initialize the population is
based on generating an initial random integer sequence, according to the values
of the allele sets specied for the individual genes. The initial genotype sequence
which codies the wavelet tree matches the breadth-rst BF sequence required
to generate a complete binary tree, up to a pre-specied maximum depth. We
refer to these sequences as tree-mapped sequences. A well-built decomposition
tree for wavelet analysis purposes, is generated by imposing appropriate constraints to the genotype sequence as specied in Sect. 3.2. The imposition of
the constraints yields variable length code sequences after resizing. An alphabet
A = f0 1 2g is used to codify the wavelet tree nodes according to their types
as specied in Sect. 3.3, thus enabling us to map any tree structure into a code
sequence. The mapping of a code sequence to a complete BF tree traversal yields
an initial sequence with length = 2 , 1, for a tree of depth . The length
is also the number of nodes in a complete binary tree of depth L. When coding
a complete binary tree using a complete BF sequence the last level of terminal nodes is redundant. Therefore, we have used a codication based on resized
complete BF sequences to code genetic sequences. The chromossomes are constructed as follows. The rst gene assumes integer values 0 2 F, where F is the
set of possible lter types used in the implementation, as explained in Sect. 3.3.
The remaining genes are used to codify the wavelet decomposition tree.
L
l
L
l
g
3.2 Constraints
There are several methods for generating trees which can be used to initialize
the population. The full, grow and ramped half-and-half methods of tree generation were introduced in the eld of genetic programming 9. These methods are
based on tree depth. The ramped half-and-half method is the most commonly
used method of generating random parse trees because of its relative higher probability of generating subtrees of varying depth and size. However, these methods
do not produce a uniform sampling of the search space. In this work, we use
constrained genetic sequences for genome initialization. Two types of constraint
operators are used to guarantee that valid tree-mapped genetic sequences are
generated: 1 top-down operator, and 2 bottom-up operator. In addition, by
applying these operators we look for a uniform sampling of the tree search space.
In terms of binary tree data structures, the top-down constraint guarantees that
if a node has null code = 0 then its two sons 0 and 1 must have null code
0 = 0 and
1 = 0. The bottom-up constraint guarantees that if at least one
of the sons 0 and 1 of a node has non-null code, then the parent 0 must
have non-null code 6= 0. These constraint operators are biased in opposite
ways. Starting from a uniform random code sequence, the bottom-up constraint
operator constructs valid genetic sequences which are biased towards complete,
full-depth trees. By he same token, the top-down constraint operator constructs
valid genetic sequence which is biased towards null, minimum depth trees. To
get a more uniform sampling of the sequence space, for sequences of maximum
= 2 , 1, we use the following initialization procedure:
ti
ci
ci
ti
ti
ci
s
ti
ti
ci
L
ti
ti
0.0 0.1 0.2 0.3 0.4 0.5
A .R . F e r r e ir a d a S ilv a
0.000
0.002
0.004
4 8
0
50
100
150
200
250
0
InteriorNodes
2
4
6
8
Depth
Fig. 1. Histogram for the distribution of interior nodes in generated sequences left
panel histogram for the distribution of trees with specied depth right panel
1. initialize sequence to terminal code value = 0 2 A ,
2. get a random value 2 1 ,
3. initialize subsequence 1 = 1 2 with random interior code values,
= f1 2g 2 A ,
4. randomly select one of the two constraint operators, bottom-up or top-down,
to apply to .
s
ci
r
s
s
ci
r
s
s
By resizing pruning constrained code sequences we allow for genetic sequences of variable length, hence tree representations of variable depth. The left
panel in Fig. 1 shows a histogram for the number of interior nodes generated
by the initialization procedure, for a maximum specied depth = 8 of the
equivalent tree, and 3000 stochastic genetic sequences. The right panel in Fig. 1
presents a histogram for the distribution of trees with specied depth, generated
from the same stochastic samples with resizing.
L
3.3 Specication
The approach of organizing libraries of bases as a tree has been extended to
construct a joint tree, to guide the process of generating shifted wavelet packet
transforms. Libraries of bases represent the population from which we want to
select the best-t individuals. In the current formulation, the genotype sequence
G allows for three optimization parameters: best lter, best wavelet packet basis
and best shifted basis. The genetic representation is used to create an initial population and evolve potential solutions to the optimization problem. The genotype
is made up of the genes which guide the discrete wavelet decomposition of each
waveform, in accordance with the joint tree representation. A cost functional
is then applied to the wavelet coecients, and its value is used to derive the
tness of the individual. In terms of entropy, the optimization problem amounts
to evolve a minimum-entropy genotype. Therefore, the best individual is the
one with minimum evolved entropy in a given library space. Since we were able
to formulate the three subtasks, wavelet packet decomposition, shifted wavelet
transform, and wavelet lter to be applied in a common data structure, the original multiple optimization problem can be solved in terms of a single aggregate
E v o lu tio n a ry W a v e le t B a s e s in S ig n a l S p a c e s
4 9
functional. The rst gene 0 in G is responsible for the optimization of the lter
used in the decomposition. We have used in the implementation 16 possible types
of dierent lters, thus 0 = f0
15g. The lters considered in the implementation were the Haar lter, the Daubechies lters D4, D6 and D8, and several
biorthogonal lters commonly used in image analysis as implemented in 7. In
particular, the lter set included the 7 9 spline lter referred to in 1, the 7 9,
11 13, 10 6, 3 5, 6 2, and 3 9 lters dened in 11, the 7 9 "FBI-ngerprint"
lter, and the 10-tap lter listed in 3.
The analysis phase of the discrete shift wavelet packet transform is codied
in the genetic sequence G . The collection of wavelet packets comprises a library
of functions with a binary tree structure. To obtain the wavelet packet analysis of
a function, or data set in the discrete case, we rst nd its coecient sequence in
the root subspace, then follow the branches of the wavelet packet coecient tree
to nd the expansion in the descendent subspaces. Assigning to each tree node
a wavelet split value 2 f0 1 2g we may enumerate all possible binary tree
structures. The value w = 1 references unshifted interior nodes, i.e., nodes with
left and right children subtrees associated with unshifted decompositions. The
value w = 2 references time shifted interior nodes. The value w = 0 references
the leaves.
g
g
:::
=
=
=
=
=
=
si
=
=
s
s
4
s
Spaces of Test Signals
A well-known implementation of the SBB algorithm which can be used for comparison purposes is contained in the WaveLab package 4. We reference by evolutionary best basis EBB, the evolutionary formulation presented in Sect. 3
for best basis selection using multilter, time shifted wavelet packet bases. By
canonical best basis CBB, we mean an algorithm which is able to reproduce
approximate the results of the SBB algorithm, using optimization methodologies for best basis selection dierent from those conceived for the SBB algorithm.
In this sense, we may map the EBB algorithm into a CBB algorithm. The SBB
algorithm is based on building an information cost tree in order to minimize
some cost measure on the transformed coecients 6. The evolutionary algorithm proposed in Sect. 3 was applied to a set of test signals and the results
compared with the results produced by SBB, based on the entropy minimization criterion. For evaluation purposes, we will use the test signals depicted in
gure 2. Two of these signals are articial signals. The other two signals are
built from collected data. The signal HypChirps includes two hyperbolic chirps.
The signal MishMash includes a quadratic chirp, a linear chirp, and a sine, as
used in the WaveLab package. The signal Seismic is distributed throughout
the seismic industry as a test dataset. Finally, the signal Sunspots represent the
monthly sunspot numbers. The following basic parameters have been used in the
steady-state evolutionary algorithm 8, 12: 1 population size: 50 2 crossover
probability: 0.95 3 mutation probability: 0.02 4 replacement percentage: 0.6.
0
200
400
600
800
-3 -2 -1 0 1 2 3
MishMash
-1.5 -0.5 0.5 1.5
A .R . F e r r e ir a d a S ilv a
HypChirps
5 0
1000
0
200
200
400
600
600
800
1000
800
0 50 100 150 200
sunspots
0.0 0.5 1.0
0
400
Index
-1.0
seismic
Index
1000
0
Index
500
1000
1500
2000
Index
Fig. 2.
Test signals
4.1 Standard Wavelet Spaces
To reproduce the application conditions of the SBB algorithm, the EBB approach was restricted to handle unshifted wavelet packet decompositions and
use a specic lter. Hence, the CBB algorithm is a restricted version of the EBB
algorithm, which is used to reproduce the SBB results using a dierent methodology. Both the EBB and the CBB algorithms are evolutionary. The minimum
entropy values associated with the best basis selected by the SBB WaveLab
algorithm for these signals, using the Daubechies D8 lter and L = 9 decomposition levels, are reported in Table 1. The SBB entropy values are also depicted
by the dashed lines in Fig. 3. The CBB algorithm was applied to the same set
of test signals to evolve the best basis using the same entropy cost functional.
Fig. 3 shows the evolution of the median minimum entropy median best value
with the number of generations for each of the test signals. The entropy values
represent the median over 30 runs of the CBB algorithm. The median entropy
values after ngen = 80 generations are shown in Table 1 as well. The optimal
entropy values and the rate of convergence can be used as benchmarks to compare the performance of dierent evolutionary algorithms, or simply to tune the
value of the control parameters.
Table 1.
Comparative median minimum entropy values and reconstruction errors
Median Minimum Entropy L=9 Reconst. Error
SBB
CBB
EBB
EBB
D8 D8, ngen=80 ngen=80 ngen=80
HypChirps 3.7908
3.7908
2.7707
1.0 e-6
MishMash 4.4805
4.5162
3.5910
2.6 e-6
Seismic 2.6022
2.6025
1.9735
2.7 e-7
Sunspots 3.0059
3.0059
2.7749
3.9 e-8
Signal
E v o lu tio n a ry W a v e le t B a s e s in S ig n a l S p a c e s
40
60
80
100
Median minimum entropy
3.0
40
60
80
100
120
0
20
100
120
40
60
80
100
120
Evolution of the median minimum entropy by the CBB algorithm
3.2
5.0
4.5
EBB
WaveLab
EBB
WaveLab
3.5
4.0
3.6
Median minimum entropy
4.0
MishMash
2.8
40
60
80
100
120
60
80
60
80
100
120
120
3.00
EBB
WaveLab
2.90
2.8
2.4
40
100
3.10
sunspots
Generation
Fig. 4.
40
seismic
2.0
20
20
Generation
EBB
WaveLab
0
0
Generation
2.80
20
3.2
0
Median minimum entropy
Median minimum entropy
80
Generation
HypChirps
Median minimum entropy
60
CBB
WaveLab
Generation
Fig. 3.
40
sunspots
2.8
Median minimum entropy
20
seismic
2.6
20
4.9
0
Generation
CBB
WaveLab
0
4.7
120
Generation
3.02 3.04 3.06 3.08
20
3.2
0
CBB
WaveLab
4.5
CBB
WaveLab
5.1
MishMash
Median minimum entropy
3.80 3.85 3.90 3.95 4.00
Median minimum entropy
HypChirps
5 1
0
20
40
60
80
100
120
Generation
Evolution of the median minimum entropy by the EBB algorithm
5 2
A .R . F e r r e ir a d a S ilv a
HypChirps
MishMash
seismic
sunspots
Fig. 5.
Evolved trees by the EBB algorithm
4.2 Enlarged Wavelet Spaces
To show the application of the EBB technique we have applied it to evolve basis
for the optimization of three parameters: wavelet packet decomposition, shiftable
decomposition and lter to use. Fig. 4 presents the evolution of the minimum
entropy values generated by the EBB algorithm for the signals in the test set. The
values are the median values over 30 runs of the EBB algorithm. The dashed
horizontal line in Fig. 4 is the value of the best basis entropy generated by
SBB. We notice that the EBB algorithm is able to greatly reduce the minimum
entropy value used to assess best basis adaptability, compared to both the SBB
and the CBB algorithms. Table 1 references the numeric median values for the
EBB minimum entropy after ngen = 80 generations. The most selected lters
among the best evolved lters for each test signal, were the following: Brislawn
10-tap lter for signals HypChirps and MishMash, Villasenor 3 9 lter for signal
Seismic, and Villasenor 6 2 lter for signal Sunspots. Fig. 5 depicts typical best
evolved trees for a sample run of the EBB algorithm. Darker lines represent
shifted wavelet packet transforms. Thinner lines represent unshifted transforms.
Another important evaluation factor is the reconstruction error. Given a signal f , we reconstruct an approximate signal f^ from the transformed coecients
by applying the inverse shifted wavelet packet transform, and calculating the l2
error between these two signals k f , f^ k2 . Table 1, presents the numeric median
E v o lu tio n a ry W a v e le t B a s e s in S ig n a l S p a c e s
5 3
values of the reconstruction errors for each of the test signals, using the EBB
algorithm for 80 generations over 30 runs.
5 Conclusion
The approximation of signals in functions spaces was used to introduce a test
environment aimed at the comparative performance of dierent evolutionary algorithms. We have considered entropy as the optimization tness criterion to be
used. However, other cost functions may prove useful to extend the range of the
test environment. In particular, measures directed to the optimization of multiobjective criteria may be incorporated in the proposed framework. On the other
hand, the proposed test environment may be easily extended to incorporate two
dimensional signal spaces. In terms of signal processing, well-adapted signal expansions are important, for instance, in signal compression. For orthonormal
basis and additive cost measures, the standard algorithm for best basis selection
is e cient. However, with the introduction of overcomplete waveform dictionaries
the algorithm has increasing di culty in nding well-adapted signal representations. The proposed evolutionary approach oers more exibility in searching
for well-adapted signal representations than standard approaches.
References
1. M. Antonini, M. Barlaud, P. Mathieu, and I. Daubechies. Image coding using
wavelet transform. IEEE Trans. on Image Process., 12, April 1992.
2. G. Beylkin. On the representation of operators in bases of compactly supported
wavelets. Society for Industrial and Applied Mathematics, 66:1716 1740, December 1992.
3. C. M. Brislawn. Two-dimensional symmetric wavelet transform tutorial program.
Technical report, Los Alamos National Laboratory, December 1992.
4. J. Buckheit and D. L. Donoho. Wavelab and reproducible research. Technical
report, Department of Statistics, Stanford University, 1995.
5. I. Cohen, S. Raz, and D. Malah. Orthonormal shift-invariant wavelet packet decomposition and representation. Signal Processing, 573:251 270, March 1997.
6. R. R. Coifman and M. V. Wickerhauser. Entropy based methods for best basis
selection. IEEE Trans. on Inf. Theory, 382:719 746, 1992.
7. G. Davis. Baseline Wavelet Transform Coder Constrution Kit. Mathematics Department, Dartmouth College, January 1997.
8. D. E. Goldberg. Genetic Algorithms in Search, Optimization, and machine learning
and Filter Banks. Addison-Wesley, Reading, Massachusetts, 1989.
9. John R. Koza. Genetic Programming - On the Programming of Computers by
Means of Natural Selection. MIT Press, Cambridge, MA, 1992.
10. Z. Michalewicz. Genetic algorithms + data structures = evolution programs. Articial Intelligence. Springer-Verlag, New York, 1992.
11. J. Villasenor, B. Belzer, and J. Liao. Wavelet lter evaluation for image compression. IEEE Trans. on Image Process., 48:1053 1060, August 1995.
12. M. Wall. GAlib: A C++ Library of Genetic Algorithm Components. Mechanical
Engineering Department, Massachusetts Institute of Technology, August 1996.
Finding Golf Courses: The Ultra High Tech
Approach
Neal R. Harvey, Simon Perkins, Steven P. Brumby, James Theiler,
Reid B. Porter, A. Cody Young, Anil K. Varghese, John J. Szymanski and
Jeffrey J. Bloch
Space and Remote Sensing Sciences Group,
Los Alamos National Laboratory, Los Alamos, NM 87545, USA
Abstract. The search for a suitable golf course is a very important issue
in the travel plans of any modern manager. Modern management is also
infamous for its penchant for high-tech gadgetry. Here we combine these
two facets of modern management life. We aim to provide the cuttingedge manager with a method of finding golf courses from space!
In this paper, we present Genie: a hybrid evolutionary algorithm-based
system that tackles the general problem of finding features of interest in
multi-spectral remotely-sensed images, including, but not limited to, golf
courses. Using this system we are able to successfully locate golf courses
in 10-channel satellite images of several desirable US locations.
1
Introduction
There exist huge volumes of remotely-sensed multi-spectral data from an everincreasing number of earth-observing satellites. Exploitation of this data requires
the extraction of features of interest. In performing this task, there is a need for
suitable analysis tools. Creating and developing individual algorithms for specific
feature-detection tasks is important, yet extremely expensive, often requiring a
significant investment of time by highly skilled analysts. To this end we have
been developing a system for the automatic generation of useful feature-detection
algorithms using an evolutionary approach.
The beauty of an evolutionary approach is its flexibility: if we can derive a
fitness measure for a particular problem, then it might be possible to solve that
problem. Many varied problems have been successfully solved using evolutionary computation, including: optimization of dynamic routing in telecommunications networks [1], optimizing image processing filter parameters for archive film
restoration [2], designing protein sequences with desired structures [3] and many
others.
When taking an evolutionary approach, a critical issue is how one should
represent candidate solutions in order that they may be effectively manipulated.
We use a genetic programming (GP) method of representation of solutions, due
to the fact that each individual will represent a possible image processing algorithm. GP has previously been applied to image-processing problems, including:
S . C a g n o n i e t a l. ( E d s .) : E v o W o r k s h o p s 2 0 0 0 , L N C S 1 8 0 3 , p p . 5 4 - 6 4 , 2 0 0 0 .
S p rin g e r-V e rla g B e rlin H e id e lb e rg 2 0 0 0
F in d in g G o lf C o u rs e s
5 5
edge detection [4], face recognition [5], image segmentation [6], image compression [7] and feature extraction in remote sensing images [8–10]. The work of
Daida et al. Brumby et al. and Theiler et al. is of particular relevance since it
demonstrates that GP can be employed to successfully evolve algorithms for real
tasks in remote-sensing applications.
2
System Overview
We call our feature detection system “Genie” (GENetic Image Exploitation) [9,
10] Genie employs a classic evolutionary paradigm: a population of individuals
is maintained and each individual is assessed and assigned a fitness value. The
fitness of an individual is based on an objective measure of its performance
in its environment. After fitness determination, the evolutionary operators of
selection, crossover and mutation are applied to the population and the entire
process of fitness evaluation, selection, crossover and mutation is iterated until
some stopping condition is satisfied.
2.1
Training Data
The environment for each individual in the population consists of a set of training data. This training data consists of a data “cube” of multi-spectral data
together with some user-defined data defining “ground-truth”. Ground-truth,
in this context, is not what is traditionally referred to as ground-truth (this
being in-situ data collected at, or as close as possible to, the time the image
was taken). Here, ground-truth refers to what might normally be referred to as
“analyst-supplied interpretation” or “training data”. This training data for our
system is provided by a human analyst, using a Java-based tool called Aladdin.
Through Aladdin, the user can view a multi-spectral image in a variety of ways,
and can “mark up” training data by “painting’ directly on the image using the
mouse. Training data is ternary-valued with the possible values being “true”,
“false”, and “unknown”. True defines areas where the analyst is confident that
the feature of interest does exist. False defines areas where the analyst is confident that the feature of interest does not exist. Fig. 1 shows a screen capture of
an example session. Here the analyst has marked out golf courses as of interest.
2.2
Encoding Individuals
Each individual chromosome in the population consists of a fixed-length string
of genes. Each gene in Genie corresponds to a primitive image processing operation, and so the whole chromosome describes an algorithm consisting of a
sequence of primitive image processing steps.
Genes and Chromosomes A single gene consists of an operator name, plus
a variable number of input arguments, specifying where input is to come from;
5 6
N .R . H a r v e y e t a l.
Fig. 1. GUI for Training Data Mark-Up. Note that Aladdin relies heavily on color,
which does not show up well in this image. The light colored patches in the center-right
and upper-right parts of the image are two golf courses that have been marked up as
“true”. Most of the rest of the image has been marked up as “false”, except for a small
region around the golf courses which has been left as “unknown”.
output arguments, specifying where output is to be written to; and operator
parameters, modifying how the operator works. Different operators require different numbers of parameters. The operators used in Genie take one or more
distinct image planes as input, and generally produce a single image plane as
output. Input can be taken from any data planes in the training data image
cube. Output is written to one of a small number of scratch planes — temporary
workspaces where an image plane can be stored. Genes can also take input from
scratch planes, but only if that scratch plane has been written to by another
gene positioned earlier in the chromosome sequence.
The image processing algorithm that a given chromosome represents can
be thought of as a directed acyclic graph where the non-terminal nodes are
primitive image processing operations, and the terminal nodes are individual
image planes extracted from the multi-spectral image used as input. The scratch
planes are the ‘glue’ that combines together primitive operations into image
processing pipelines. Traditional GP ([11]) uses a variable sized (within limits)
tree representation for algorithms. Our representation differs in that it allows
for reuse of values computed by sub-trees since many nodes can access the same
scratch plane, i.e. the resulting algorithm is a graph rather than a tree. It also
differs in that the total number of nodes is fixed (although not all of these may
be actually used in the final graph), and crossover is carried out directly on the
linear representation.
We have restricted our “gene pool” to a set of useful primitive image processing operators. These include spectral, spatial, logical and thresholding operators.
Table 1 outlines these operators. For details regarding Laws textural operators,
the interested reader is referred to [12, 13].
The set of morphological operators is restricted to function-set processing
morphological operators, i.e. gray-scale morphological operators having a flat
structuring element. The sizes and shapes of the structuring elements used by
F in d in g G o lf C o u rs e s
5 7
Table 1. Image Processing Operators in the Gene Pool
Code
ADDP
SUBP
ADDS
SUBS
MULTP
DIVP
MULTS
DIVS
SQR
SQRT
LINSCL
LINCOMB
SOBEL
PREWITT
AND
OR
CL
LAWB
LAWD
LAWF
LAWH
Operator Description
Add Planes
Subtract Planes
Add Scalar
Subtract Scalar
Multiply Planes
Divide Planes
Multiply by Scalar
Divide by Scalar
Square
Square Root
Linear Scale
Linear Combination
Sobel Gradient
Prewitt Gradient
And Planes
Or Planes
Clip Low
Laws Textural Operator S3T
Laws Textural Operator E3T
Laws Textural Operator L3T
Laws Textural Operator S3T
× L3
× E3
× S3
× S3
Code
Operator Description
MEAN
VARIANCE
SKEWNESS
KURTOSIS
MEDIAN
SD
EROD
DIL
OPEN
CLOS
OPCL
CLOP
OPREC
CLREC
HDOME
HBASIN
CH
LAWC
LAWE
LAWG
Local Mean
Local Variance
Local Skewness
Local Kurtosis
Local Median
Local Standard Deviation
Erosion
Dilation
Opening
Closing
Open-Closing
Close-Opening
Open with Reconstruction
Close with Reconstruction
H-Dome
H-Basin
Clip High
Laws Textural Operator L3T × E3
Laws Textural Operator S3T × E3
Laws Textural Operator E3T × S3
these operators is also restricted to a pre-defined set of primitive shapes, which
includes, square, circle, diamond, horizontal cross and diagonal cross, and horizontal, diagonal and vertical lines. The shape and size of the structuring element
are defined by operator parameters. Other local neighborhood/windowing operators such as mean, median, etc. specify their kernels/windows in a similar way.
The spectral operators have been chosen to permit weighted sums, differences
and ratios of data and/or scratch planes.
We use a notation for genes that is most easily illustrated by an example: the
gene [ADDP rD0 rS1 wS2] applies pixel-by-pixel addition to two input planes,
read from data plane 0 and from scratch plane 1, and writes its output to scratch
plane 2. Any additional required operator parameters are listed after the input
and output arguments.
Note that although all chromosomes have the same fixed number of genes,
the effective size of the resulting algorithm graph may be smaller than this.
For instance, an operator may write to a scratch plane that is then overwritten
by another gene before anything reads from it. Genie performs an analysis of
chromosome graphs when they are created and only carries out those processing
steps that actually affect the final result. Therefore, in some respects, we could
refer to the fixed length of the chromosome as a “maximum” length.
2.3
Backends
Complete classification requires that we end up with a single binary-valued output plane from the algorithm. It would be possible to treat, say, the contents of
5 8
N .R . H a r v e y e t a l.
scratch plane 0 after running the chromosome algorithm, as the final output from
the algorithm (thresholding would be required to obtain a binary result). However, we have found it to be of great advantage to perform the final classification
using a non-evolutionary algorithm.
To do this, we first select a subset of the scratch planes and data planes
to be answer planes. Typically in our experiments this subset consists of just
the scratch planes. We then use the provided training data and the contents of
the answer planes to derive the Fisher Discriminant, which is the linear combination of the answer planes that maximizes the mean separation in spectral
terms between those pixels marked up as “true” and those pixels marked up
as “false”, normalized by the “total variance” in the projection defined by the
linear combination. See [14] for details of how this discriminant works.
The output of the discriminant-finding phase is a gray-scale image. This is
then reduced to a binary image by using Brent’s method [15] to find the threshold
value that minimizes the total number of misclassifications (false positives plus
false negatives) on the training data.
2.4
Fitness Evaluation
The fitness of a candidate solution is given by the degree of agreement between
the final binary output plane and the training data. This degree of agreement
is determined by the Hamming distance between the final binary output of the
algorithm and the training data, with only pixels marked as true or false contributing towards the metric. The Hamming distance is then normalized so that
a perfect score is 1000. To put this in a more formal/mathematical context. Let
H be the Hamming distance between the final binary output of the algorithm
and the training data, with only pixels marked as true or false contributing towards the metric, let N be the number of classified pixels in the training image
(i.e. pixels marked as either “true” or “false”) and let F be the fitness of the
candidate solution.
F = (1 − (H/N )) × 1000
2.5
(1)
Software Implementation
The genetic algorithm code has been implemented in object-oriented Perl. This
provides a convenient environment for the string manipulations required by the
evolutionary operations and simple access to the underlying operating system
(Linux). Chromosome fitness evaluation is the computationally intensive part of
the evolutionary process and for that reason we currently use RSI’s IDL language
and image processing environment. Within IDL, individual genes correspond to
single primitive image operators, which are coded as IDL procedures, with a chromosome representation being coded as an IDL batch executable. In the present
implementation, an IDL session is opened at the start of a run and communicates with the Perl code via a two-way unix pipe. This pipe is a low-bandwidth
connection. It is only the IDL session that needs to access the input and training
F in d in g G o lf C o u rs e s
5 9
data (possibly hundreds of Megabytes), which requires a high-bandwidth connection. The Aladdin training data mark-up tool was written in Java. Fig. 2
shows the software architecture of the system.
Fig. 2. Software Architecture of the System Described. Note that the feature depicted
on the right of this diagram represents the input data, training data and scratch planes
3
Why Golf Courses?
The usefulness of devising algorithms for the detection of golf courses may not,
at first, seem apparent (except to a manager, perhaps!). However, due to the
nature of golf courses and their characteristics in remotely-sensed data, they are
of great use in testing automatic feature-detection systems, such as described
here. They possess distinctive spectral and spatial characteristics and it is the
ability of feature-detection algorithms to utilize both these “domains” that we
seek to test. It is also useful that there exists a great deal of “ground truth”
data available: a great many golf courses, for the benefit of low-tech managers,
are marked on maps. In addition, golf courses usually possess a well-known,
particular type of vegetation and it is rare to find information regarding specific
vegetation types on maps. Fig. 3 (a) shows a map of NASA’s Moffet Field Air
Base, clearly showing the position of a golf course. Fig. 3 (b) shows a false col-
(a)
(b)
Fig. 3. (a) Map of NASA’s Moffet Field Air Base, showing a golf course (available at
http://george.arc.nasa.gov/jf/mfa/thesite2.html) (b) Image from remotely-sensed data
of NASA’s Moffet Field Air Base
6 0
N .R . H a r v e y e t a l.
our image of some remotely sensed data of the same region. The airfield and golf
course are clearly visible.
4
Remotely-Sensed Data
The remotely-sensed images referred to in this paper are 10-channel simulated
MTI data, produced from 224-channel AVIRIS data, each channel having 614 ×
512 pixels. The images displayed are false-color images (which have then been
converted to gray-scale in the printing process). The color mappings used are
the same for all images shown (an exception being Fig. 1 where the false-color
image has had a red and green overlay, corresponding to “false” and “true”
pixels, as marked by the human analyst). The particular color mappings used
here involve averaging bands A and B for the blue component, bands C and D
for the green component and bands E and F for the red component. In addition,
the images have been contrast enhanced. The choice of color mappings was
arbitrary, in that it was a personal decision made by the analyst, made in order
to best “highlight” the feature of interest, from his/her perspective and thus
enable him/her to provide the best possible training data. This choice of colormappings, together with a contrast-enhancement tool, are important and very
useful features of Aladdin. Table 2 provides details about MTI data.
Table 2. MTI Band Characteristics
Band
Wavelength (µm)
Color
SNR
Ground Sample Distance
A
B
C
D
E
F
G
H
I
O
J
K
L
M
N
0.45-0.52
0.52-0.60
0.62-0.68
0.76-0.86
0.86-0.89
0.91-0.97
0.99-1.04
1.36-1.39
1.55-1.75
2.08-2.35
3.50-4.10
4.87-5.07
8.00-8.40
8.40-8.85
10.2-10.7
blue/green
green/yellow
red
NIR
NIR
NIR
SWIR
SWIR
SWIR
SWIR
MWIR
MWIR
LWIR
LWIR
LWIR
120
120
120
120
500
300
600
4
700
600
250
500
800
1000
1200
5m
5m
5m
5m
20m
20m
20m
20m
20m
20m
20m
20m
20m
20m
20m
Figs. 3(a), 4(a) and 5(a) are data taken over an area of NASA’s Moffet Field
Air Base in California, USA. Fig. 3(a) is a sub-set of the data shown in Fig. 4(a).
Figs. 3(a) and 5(a) are non-adjacent regions of the original data. These sub-sets
of the data contain a lot of different features, but, of course, have a common
feature of interest: golf courses.
F in d in g G o lf C o u rs e s
5
6 1
Searching for Golf Courses
We reserve the data described above (Fig. 3(a)) for testing an evolved golfcourse finder algorithm and set the system the task of finding a golf course on
some other data. This data, showing the “truth” as marked out by an analyst,
is shown in Fig. 1. The golf course area has been marked as “true” and most
of the remaining data has been marked as “false”. The system was run for 400
generations, with a population of 100 chromosomes, each having a fixed length of
20 genes. At the end of the run the best individual had a fitness of 966 (a perfect
score would be 1000). This fitness score actually translates into a detection rate
of 0.9326 and a false alarm rate of 0.00018. The results of applying the best
overall algorithm found during the run to the data used in the training run are
shown in Fig. 4.
(a)
(b)
Fig. 4. (a) Image of training data (b) Result of applying algorithm found to training
data
It can be seen that the algorithm has been able to successfully detect the
golf course and has not detected any of the other features within the image.
In order to test the robustness of the algorithm found, it was applied to outof-training-sample data, as described previously, and shown in Fig. 3 (b). The
results are shown in Fig. 5.
It should be noted that the data shown in Fig. 5 covers a greater area than
shown by the map in Fig. 3 (a). It can be seen that the algorithm has successfully
found the golf course shown on the map. It can also be seen that the algorithm
has detected other golf courses. On closer examination of the data, it would
appear that further golf courses do, in fact, exist at those locations. It can also
be seen that the algorithm has not found any spurious features.
The “short” (redundant genes stripped out) version of the chromosome found
is detailed below.
[LAWG rD2 wS0] [OPREC rD3 wS3 5 1] [ADDP rS0 rS3 wS1] [ADDP rS1
rD6 wS1] [LAWE rD6 wS4] [LAWG rD6 wS0] [OPCL rS4 wS3 1 1] [DIL rS1 wS1
1 0] [OPREC rS1 wS1 5 0] [MEDIAN rS1 wS2 1] [LAWH rD2 wS4]
6 2
N .R . H a r v e y e t a l.
(a)
(b)
Fig. 5. (a) Image of out-of-training-sample data (b) Result of applying algorithm found
to out-of-training-sample data
A graphical representation of the algorithm found is shown in Fig. 6. Note
that the circles at the top of the graph indicate the data planes input to the
algorithm (in this case only 3 data planes out of a possible 10 have been selected),
the 5 circles in the center represent the scratch planes and the circle at the bottom
represents the final, binary output of the overall algorithm. The operations above
the line of scratch planes represent that part of the overall algorithm incorporated
in the chromosome. The operations below the line of scratch planes represent the
optimal linear combination of scratch planes and intelligent thresholding parts
of the overall algorithm.
It is interesting to have some kind of objective measure of the algorithm’s
performance on the out-of-training-sample data. To this end an analyst marked
up training data (i.e. true and false) for this data, with respect to the golf courses
present. This enabled determination of a fitness for the algorithm on this data as
well as detection and false alarm rates. The fitness of the algorithm was 926.6,
the detection rate was 0.8532 and false-alarm rate was 3.000E-05.
6
Comparison with Other Techniques
In order to compare the feature-extraction technique described here to a more
conventional technique, we used the Fisher discriminant, combined with the intelligent thresholding, as described previously, to try and extract the golf courses
in the images shown/described. This approach is based purely on spectral information. On application to the data used in the training run (Fig. 4(a)), this
“traditional” approach produced a result having a fitness of 757.228 (with respect to the training data/analyst-supplied interpretation), which translates into
a detection rate of 0.5159 and a false-alarm rate of 0.00141. On application to the
out-of-training-sample data, the result had a fitness of 872.323, which translates
into a detection rate of 0.7477 and false-alarm rate of 0.00305. Both of these
results are significantly below the performance of the results produced by the
Genie system described here.
F in d in g G o lf C o u rs e s
7
6 3
Conclusions
A system for the automatic generation of remote-sensing feature detection algorithms has been described. This system differs from previously described systems
in that it combines a hybrid system of evolutionary techniques and more traditional optimization methods. It’s effectiveness in searching for useful algorithms
has been shown, together with the robustness of the algorithms discovered. It has
also been shown to significantly out-perform more traditional, purely-spectral
approaches.
D 3
D 2
D 6
O p e n R e c .
la w G
A d d P la n e s
A d d P la n e s
D ila te
la w E
O p e n R e c .
O p e n -C lo s e
M e d ia n
la w H
S 1
S 4
- 3 .9 1 8 E - 7
M u ltip ly S c a la r
la w G
S 2
S 3
1 .1 7 2 E - 5
M u ltip ly S c a la r
3 .0 2 3 E - 6
M u ltip ly S c a la r
M u ltip ly S c a la r
S 0
6 .0 4 9 E - 6
M u ltip ly S c a la r
A d d P la n e s
- 1 .8 1 7
A d d S c a la r
0 .6 8 2
T h re s h o ld
O /P
Fig. 6. Graphical representation of algorithm found
5 .7 1 9 E - 7
6 4
N .R . H a r v e y e t a l.
References
1. Cox, L.A., Jr., Davis, L., Qiu, Y.: Dynamic anticipatory routing in circuit-switched
telecommunications networks, in Handbook of Genetic Algorithms, L. Davis, ed.,
pp. 124-143, Van Nostrand Reinhold, New York, 1991.
2. Harvey, N.R., Marshall, S.: GA Optimization of Spatio-Temporal Grey-Scale Soft
Morphological Filters with Applications in Archive Film Restoration. In: Poli, R.,
Voigt, H.-M., Cagnoni, S., Corne, D., Smith, G.D., Fogarty, T.C. (eds.): Evolutionary Image Analysis, Signal Processing and Telecommunications (1999) pp. 31–45
3. Dandekar, T., Argos, P.: Potential of genetic algorithms in protein folding and protein engineering simulations, Protein Engineering 5(7), pp. 637-645, 1992.
4. Harris, C., Buxton, B.: Evolving edge detectors, Research Note RN/96/3, University
College London, Dept. of Computer Science, London, 1996.
5. Teller, A., Veloso, M.: A controlled experiment: Evolution for learning difficult image
classification, in 7th Portuguese Conference on Artificial Intelligence, Volume 990
of Lecture Notes in Computer Science, Springer-Verlag, Berlin, 1995.
6. Poli, R., Cagoni, S.: Genetic programming with user-driven selection: Experiments
on the evolution of algorithms for image enhancement, in Genetic Programming
1997: Proceedings of the 2nd Annual Conference, J. R. Koza, et al., editors, Morgan
Kaufmann, San Francisco 1997.
7. Nordin, P., Banzhaf, W.: Programmatic compression of images and sound, in Genetic Programming 1997: Proceedings of the 2nd Annual Conference, J. R. Koza,
et al., editors,, Morgan Kaufmann, San Francisco, 1996.
8. Daida, J.M., Hommes, J.D., Bersano-Begey, T.F., Ross, S.J., Vesecky, J.F.: Algorithm discovery using the genetic programming paradigm: Extracting low-contrast
curvilinear features from SAR images of arctic ice, in Advances in Genetic Programming 2, P. J. Angeline and K. E. Kinnear, Jr., editors, chap. 21, MIT, Cambridge,
1996.
9. Brumby, S.P., Theiler, J., Perkins, S.J., Harvey, N.R., Szymanski, J.J., Bloch J.J.,
Mitchell, M.: Investigation of Image Feature Extraction by a Genetic Algorithm in
Proc. SPIE 3812, pp. 24–31, 1999.
10. Theiler, J., Harvey, N.R., Brumby, S.P, Szymanski, J.J., Alferink, S., Perkins, S.,
Porter, R., Bloch, J.J.: Evolving Retrieval Algorithms with a Genetic Programming
Scheme in Proc. SPIE 3812, in Press.
11. Koza, J.R.: Genetic programming: On the Programming of Computers by Means
of Natural Selection MIT Press, 1992
12. Laws, K.I.: Texture energy measures in Proc. Image Understanding Workshop,
Nov. 1979, pp. 47–51.
13. Pietikainen, M., Rosenfeld, A., Davis, L.S.: Experiments with Texture Classification using Averages of Local Pattern Matches IEEE Trans. on Systems, Man and
Cybernetics, Vol. SMC-13, No. 3, May/June 1983, pp. 421–426.
14. Bishop, C.M.: Neural Networks for Pattern Recognition, pp. 105–112, Oxford University Press, 1995.
15. Press, W.H., Teukolsky, S.A., Vetterling, W.T., Flannery, B.P.: Numerical Recipes
in C, 2nd Edition, Cambridge University Press, 1992, pp. 402–405..
L c
t
M
s G
ika rd Ka rl
sso ,
o
pl
r H
tc Pr r
t r o rdi a d
R
t
a ts o rda l
st s ro p, a l rs
i rsit o f c o l
o , -42 9 6 , ö t o r ,
d ,
a il
: o rdi ,tf
@ f .c a l rs.s
c
l
i ar
s st
as
s d to so l t pro l o f
so d l
o ca l
ia tio fo r a a to o o s
a o id ro o t, it t o icro p o s a s a rs. o d t r i t a lto t so d so rc , a
old
pro ra
as s d i a l
o o p o r a st r o sa plstr a ,
r t
tic pro ra
ts t l
a t stsa plpa irplsf d a ckfro t pr io s
it ra tio a s i p t.
pr cisio o ft
o l d pro ra s a s d p d t
o t
p ri ta ls t p. o r a sa to o t a fro a fi d dista c
t s al
lst rro r a s 8 .
ltti t dista c to t sa
so rc
a r t rro r a s 23 . o r a
a o ic a t a r i dista c s t rro r
a s p to 4 .
tr
ct
p rpo s o ft is pa p r is to i stiga t so d l
o ca l
ia tio
si g .
s sis it d d to b s d i a
a o id ro bo t q ipp d it t o icro p o s.
dl
o ca l
ia tio is p rfo r d b a s a l
lbi a r
a c i co d pro gra o a
pro c ss d st r o str a o fsa pld so d. i c t s st is it d d fo r a
a o id ro bo t,so
d gr o fsi il
a rit to
a so d l
o ca l
ia tio isd sir d
(s [ ] fo r a o t r st d o ft is pro bl i a ro bo tco t t). ca s o ft
l
i it d
po r o -bo a rd t ro bo tt co p ta tio a lr q ir
ts
d to
b
ii i d
s st
ds to g ra l
i fro a l
i it d s to ftra i i gda ta .
s st a l
so
ds to b a blto l
o ca l
i
a
diff r tki ds o fso ds, so ds
fro a l
l
dir ctio si t o rio ta l
pl
a ,so ds it diff r tit sit,so ds
fro a l
ldista c s,so ds
c o s a r pr s t,a d so ds it ba ckgro d
o is pr s t.
s a co pa riso ,
rst gi a s o rt d scriptio o ft
a a dito r
s st a d its so d l
o ca l
ia tio ca pa bil
itis.
t
o
.
a
i r s s
a bil
it to l
o ca l
i so ds is a i po rta tpa rto ft a dito r s st a d
as b
ss tia lto o r s r ia l
. o d pro pa ga t s t ro g a
di a s a
l
o git di a l a .
its a to t a rdr
itpa ss s t o t r a r,ca l
ld t
pi a , a d t a dito r ca a l
.
l
o git di a l a i t a ir is tra sfrr d
S . C a g n o n i e t a l. ( E d s .) : E v o W o r k s h o p s 2 0 0 0 , L N C S 1 8 0 3 , p p . 6 5 − 7 6 , 2 0 0 0 .
S p rin g e r-V e rla g B e rlin H e id e lb e rg 2 0 0 0
6 6
R . K a rls s o n , P . N o rd in , a n d M . N o rd a h l
to t a rdr , o r t pa ic
bra .
sig a lis a pl
i d o its a b
t o ssicls, ic
o rk a s l rs fro t t pa ic
bra to t s a l
lr
a r a o ft o a l i do .
o
to ft o a l i do is tra sfrr d ito a
l
o git di a l a i t l
iq id- l
ld co c la .
l
o git di a l a pro pa ga t sdo
t l
o gstr tc d ba sil
ar
bra
i t co c la .
bra is a rro r a d t i r, a d t s o r s siti
to ig r fr q
cis, a r t o a l i do a d r spo ds to l
o r fr q
cis
f rt r do
t
bra . t
35
a d5
ro s a r distrib t d
al
o gt ba sil
ar
bra . a c a s a c a ra ct ristic fr q
c rl
a t d to t
fr q
c ca si g a i a ldispl
ac
to ft ba sil
ar
bra a tits l
o ca tio .
is o rga ia tio ,
r
ig bo ri g
ro s a si il
a r r spo s , is ca l
ld
to o to p . It a s a fr q
c ra g fro 2 H to 2
H . rq
cis l
o r
t a 2 H a r co d d i t r so fti o f ra lri g,b p a s l
o cki gb t
t so d a a d t
ro
ri g. t
2 H a d4
H bo t p a s
l
o cki ga d to o to p a r s d, ilo l to o to p is s d a bo 4
H .
.2
a
s
ca ia i
t s o i stiga t t c s fo r so d l
o ca l
ia tio pr s t rig t a t t
a rdr s. o sid r a d
a d i a a c o ic ( c o -fr ) c a b r it
o
so d so rc .
so d a fro t so d so rc pro pa ga t s t ro g
t a ir it a l
o cit o f34 / s a d r a c s t d
a d a d its pi a ,
a dito r ca a l
s a d a rdr s. Ift d
a d is fa ci g t so d so rc
a ta a glit il
lta k t so d l
o g r to r a c o
a r t a t o t r. is
diff r c i ti o fa rria lis k o
a s t it ra ra lti d l
a (I ). Ift
so d is co ti o s t I
ca i st a d b d t r i d b co pa ri gp a s s o f
t so d sig a l
s a tt a rdr s.
If
ko t rl
a ti p a s s a d t
a l gt s o ft
a s,
ca
o ld tr i
iq l ic
a l
a gs t o t r if k o t a to
a is
a la s d l
a d b lss t a
al
fa a l gt it r sp ctto t o t r.
is
p ts a l
i ito t fr q
c.
t so d co s fro t lfto r rig tt
d l
a is a i a l
.
dista c b t
t a rs is a ppro i a t l2 c , ic
a s t a tt
a l gt
ds to b l
o g r t a 4 c . is co rr spo ds to
fr q
cis s a l
lr t a 5 H , fo r a ir it a so d l
o cit o f34 / s.
o t r c t a tca b s d to d t r i t dir ctio to a so d so rc
is t it ra ra lit sit diff r c (II ), si c t
a d ca sts a so d s a do
a d r d c s t a pl
it d o fa pa ssi g so d a . o d a s l
o grta
t
idt o ft
a d a r stro gl diffra ct d, a d fr q
cis l
a rg r t a
7
H a r stro gl s a d d.
pi a a l
so l
t r t a co stic sp ctr
i a dir ctio -d p d t a .
is l
t r is ca l
ld a d-r l
a t d tra sfr f ctio (H
). Itca b s
as a
f ctio t a t a ps it sit to it sit a d ta k s t fr q
c o ft so d
a d t dir ctio to t so d so rc a sa rg
ts.
H
pro id sa o t r
po t tia lc fo r so d l
o ca l
ia tio .
or l
o ca l
ia tio i t
o rio ta lpl
a
I
a d II a r
a jo r c s fo r
a s,
il t H
is a c fo r l
o ca l
ia tio i t
rtica lpl
a s. I
S o u n d L o c a liz a tio n fo r a H u m a n o id R o b o t
6 7
a st I
is s d fo r fr q
cis s a l
lrt a 2 kH , ilt II is s d
fo r fr q
cis a bo 2 kH .
s rs l
ts a gr fa irl l
l it t a rg
ts
a bo .
o r d ta il
so
a so d l
o ca l
ia tio a d its a cc ra c ca b fo d i
[2]. I t
o rio ta lpl
a ,
a s a k a rro ro f4 to
d p di go t
dir ctio to t so d so rc . o ds stra ig ta a d o f s a r a sistto l
o ca l
i ,
ilitis a rd r to l
o ca l
i so ds i t
dia pl
a . Itis
o r diffic l
t
to l
o ca l
i so ds ifo
is d a fo o
a r.
titis po ssibl to s o l t
H
as c .
2
p r
t s tp
o icro p o s r s d to r co rd so ds i st r o fro diff r tdir ctio s
i t
o rio ta lpl
a . o r so d l
o ca l
ia tio to b a s si pla d ffici ta s
po ssibl
o l d pro gra st a tco l
dl
o ca l
i so d si go la ra sa pld
so d sig a la s i p t.
pro gra
a s it ra t d it f dba ck o ft o tp t
fro t pr io s it ra tio a d t
tpa ir o fso d sa pls a s i p t.
2.
c r i
s
s
o d r co rdi gs i st r o
r
ad
it t o
icro p o s pl
ac d i t
d
a d o ft ro bo t.
sig a lfro t
icro p o s a s pr a pl
i d,
a d sa pld a d r co rd d i st r o b a
so d ca rd. l
lr co rdi gs r
a d i a o ffic it ba ckgro d o is a i l fro co p t r fa s.
c o ic o fsa pl
i g fr q
c a d r so ltio l
i its
ic c s ca b
s d fo r so d l
o ca l
ia tio , si c t sa pl
i g la ds to a discr tia tio
rro r.
is c o ic a l
so d t r i s t fr q
c ra g .
a rd a r r stricts s to
o r 6 bits r pr s ta tio .
discr tia tio
rro r is a t o st a l
ft r so ltio
i ti a d sa pl a l . o ca l
c l
at t I
a d II a diff r c b t
sig a l
s is i o l d, so t
a i a lrro r is t r so ltio i ti a d sa pl
al .
s o fI
fo r so d l
o ca l
ia tio r q ir s ig ti r so ltio ; t
sa plr so ltio is lss i po rta t. Ift so d a rri s fro i it t I
i
r
is
d p ds o t a glto t so d a ppro i a t la s
=
t dista c b t
t a rs, is t so d l
o cit a d is t a glto t
so d so rc .
rro r is t
gi b :
=
o o bta i a rro r o f
fo r a so d fro stra ig ta a d t sa pl
i gti
−
stb a ss o rta s46
s,co rr spo di gto a sa pl
i gfr q
c o f22 kH .
rq
cis p to 5 H ca b s d to ca l
c l
at I
fo r
a s, ilt
l
i itis 9
H fo r a ro bo t it 9 c b t
its a rs.
II o t o t r
a d is a i ls siti to t sa plr so ltio . I pra ctic ,fr q
cis a bo
7 H ca b s d to ca l
c l
a t II fo r
a s, ilt l
i itfo r t ro bo t
is 3
H . o s t H
a l
a rg ra g o ffr q
cis
stb pr s ta d
a ig r so ltio i t sa pl a l is
d d.
6 8
R . K a rls s o n , P . N o rd in , a n d M . N o rd a h l
a ki g t s fa cts ito a cco t, a 6 kH sa pl
i g fr q
c
a s d cid d
po , ba s d o a co pro is b t
r d ci g t discr tia tio
rro r, a d
r d ci gt a o to fi p tda ta fo r t g tic pro gra .
ra lso ds r r co rd d si gdiff r t p ri ta ls t ps.
a
si gi g o ic a s r co rd d it a
pt
a d o ft ro bo t(a c bic a l i
bo , o p
i t to p a d ba ck).
o s ts o f 6
H sa to o t a s r
r co rd d it diff r tpi a . i co
a ds spo k b a
a
o ic
r
al
so r co rd d.
s r co rdi gs a r r fr d to a s r co rdi gs
, a I a d II
a d o
a dsi t r sto ft pa p r.
sa pls r r co rd d si g 6 bits.
I t s t p fo r r co rdi g
t
a d a s pt.
pi a
r ad o f
pl
a stic, a d
r si il
a r to
a pi a . o a dito r ca a l
s r pr s t.
al
a si gi ga to
a sr co rd d fro
6
ldistrib t d dir ctio s i
t
o rio ta lpl
a . o r t r co rdi g a I t
ad as l
ld it i s l
a ti g
a t ria l
. dito r ca a l
s it a dia t r o f7
a d a d pt o f
r
a d o to fpl
a stic. o
r t s al
lpi a .
6 H sa to o t a
as
tra s itt d fro a l
o dsp a k r .2
t rs fro t
a d. o ds r r co rd d
fro
6 dir ctio s q a l
l spa c d i t
o rio ta lpl
a .
I t r co rdi g o
a ds t
ad as l
ld it i s l
a ti g a t ria l
,
a d t ro bo t a s q ipp d it a rs a d o f o d l
i gcl
a a d o d ld a ft r
a a rs.
a dito r ca a l
s r a ppro i a t l 2
l
o g.
a ccid t,
t a rca a l
sdiff r d so
a ti idt . o ds r r co rd d it t d
a d pl
ac d i
diff r tpo sitio s i t ro o .I a c r co rdi gs r a d
fro
6
l distrib t d dir ctio s i t
o rio ta lpl
a .
so d so rc
a s pl
a c d it r o , t o , t r o r fo r
t rs fro t d
a d.
r co rd d so ds r a a l o ic gii g t co
a ds a l
k, sta d, fo r a rd,
ba ck, sto p, rig t, lfta d gra b.
I t r co rdi g a II t sa
a d a s i r co rdi g o
a ds a s s d
( c pt t a t t
a r ca a l
s r
ad q al
l id ). a to o t a s it a
fr q
c of 6
H
r r co rd d it t
a d pl
a c d a t t r diff r t
l
o ca tio s i t ro o ,
,2
a d 3 fro t so d so rc . t
so ds
fro
6 dir ctio s
l distrib t d i t
o rio ta lpl
a
r r co rd d. I
t o t r t o po sitio s so ds fro
i dir ctio s a s r co rd d.
2.2
ic pr ra
i
rs
ca ia i
pro gra s o l d s d a ra sa pld so d sig a la s i p t.
pro gra
a s it ra t d it f dba ck o ft o tp t fro t pr io s it ra tio a d t
t pa ir o f so d sa pls a s i p t.
i diid a lpro gra ca b
i d
a s a f ctio
o s i p ts a d o tp ts a r a rra s.
i p tco sists o ft o
o ris, t pr io s r a lo tp ts, t o sa pls a d a co sta t.
o tp t
co sists o ft t o
o ris a d t o r a lo tp ts.
o ris r pr s tt o tp tfro t i diid a l s d a s i p ti t
tit ra tio .
’
r a lo tp ts’a r t o tp ts s d to a la t t a glto
t so d so rc .
s a r s d a s f dba cki t sa
a as t
o ris.
sa pls a r a pa ir o fi p tsa pls fro t lfta d rig tso d c a
l
.
S o u n d L o c a liz a tio n fo r a H u m a n o id R o b o t
6 9
I itia l
l t i p ts i fo r o f
o ris a d pr io s r a lo tp ts a r s t
to ro . ri g s bs q
tit ra tio s,
o ris a d pr io s r a lo tp ts, a r
s d a s f dba ck.
rstti a i diid a lis r itis gi t
rstpa ir o f
sa pls i t
t ss ca s a s i p t, a d a
pa ir o fsa pls is t
s ppl
id
i a c it ra tio .
ca d scrib o t i diid a l
s a r s d i ps do -co d , s
gr .
sa p
sa p
ss as
ss as
r
)
)
)
)
a
sa p
sa p
sa p
a
a
a
s o fa
i ps do -co d .
ss as
ss as
s
s r
sa p
s )
)
)
r
pr
sr a
p s
i diid a lto d t r i t dir ctio to a so
d so rc is s o
t o r a lo tp ts ’
o t ’a d ’
o t2’a r it rpr t d a s t
-a d co o rdi a t i a ca rt sia co o rdi a t s st , a d t a gl is ca l
c l
a t d fro
t is i fo r a tio . is r pr s ta tio is s gg st d b t g o tr o ft pro bl (a a g l
a r r pr s ta tio
as al
so trid).
b r o fit ra tio s
d d ca b sti a t d fro t p sics o ft
pro bl . o s t I
t i diid a l ds to s a tla sta
b ro fsa pls
co rr spo di gto t ti d l
a .
a i a lti d l
a is a ppro i a t l t
ti itta k s t so d to tra lt dista c b t
t a rs, a ppro i a t l
.26 s fo r a 9 c ro bo t a d. is co rr spo ds to 4.2 sa pls fo r a sa pl
ig
fr q
c of 6
H . o s t II t i diid a l ds to s a tla sto
a l gt . o r a so d o f4 H t is co rr spo ds to 4 sa pls. 4 H is
l
l
b l
o t fr q
ciso f ig stit sit i t
a sp a ki g o ic . i c
ig r fr q
cis a r o stit r sti gfo r t H
o o r sa pls t a to
a la t t II a r l
ik l to b
d d.
2.
s s
l
i ar
s st
o li g bi a r
a c i co d
a s s d [3].
s st
o p ra t s it a fo r i diid a lto r a
t s lctio
si g st a d sta t .
d
s
r s d, to g t r it a a rit
tic f ctio s t.
7 0
R . K a rls s o n , P . N o rd in , a n d M . N o rd a h l
t ss ca s co sists o fa
b r o fpa irs o fsa pls.
s
r c os
ra do l fro a l
o g r a rra o fsa pls fro a r co rdi g co rr spo di g to a
sp ci c dir ctio to t so d so rc . l
ldir ctio s r gi
q a l ig t
i t tra i i g s t.
t ss ca s s r c o s
a c ti a i diid a l a s
a la t d, to i pro g ra l
ia tio .
t sts t a s c o s ra do lfro t l
a rg r s to fsa pls i t sa
a a s t tra i i gs t.
a la tio
a s do
o
t ss ca s s c o s t
sa
a a s t o s s d fo r tra i i g.
al
ida tio s t a s s d to c ck o
l
lt g tic pro gra g ra l
i s to a diff r ts tt a t tra i i gs t.
rro r o f a i diid a l a s ca l
c l
a t d a s diff r c , t s o rt st a
a ro d t circl, b t
t co rr cta gla d t a glca l
c l
a t d fro t
o tp to ft g tic pro gra .
t ss a s d
d b t fo l
l
o i g pr ssio :
i
=
− i
r
is t
b r o f t ss ca s s i t tra i i gs ta d t i d
spo ds to a sp ci c t ss ca s .
3
co rr -
s ts
s cc ss o ft
o l d pro gra
a rid sig i ca tl it t
p ri ta l
s t p.
rs l
ts fro
p ri ts I to
ar s o
i g r s 2 to 7 . ig r s 2 to
5s o t
t ss o ft b sti diid a l
a la t d o t tra i i g,t sta d
al
ida tio s ts.
al
so s o t a ra g a d
dia t ssfo rt po p l
a tio
asa
o l,a d t pro gra l gt o ft b stg tic pro gra a d t a ra g
l gt o ft g tic pro gra s i t po p l
a tio .
i dia gra s i g r 6 a d g r 7 s o t a glca l
c l
a t d a ft r
a c it ra tio d ri g t
c tio o ft g tic pro gra i
p ri ts I
a d .I gr 6 o
gra p is s o
i a c dia gra a d i g r 7 t r
gra p s co rr spo di g to so ds co i g fro diff r tdista c s a r s o . I
t
i dia gra s t d sir d a s r ra g s fro -9
i t
rstto + 9
i
t l
a st it a spa ci g o f22.5 .
a la tio s r do
o a t sts tt a t
a s id tica lto t tra i i gs t.
a bl s o s t pa ra t rs t a t r a rid b t
p ri ts.
fo l
l
o i gpa ra t rs r id tica li a l
l p ri ts:
()
g tic pro gra s r tra i d o so ds co i gfro t dir ctio s
= , 22 5 , 45 , 6 7 5 a d 9 .
(2) l
lsa pls r a pl
i d i al
l
p ri ts i s c a a t a t
t
a i a lsa pla pl
it d
a s id tica li a l
l t ss ca s s. ft r t is a pl
i ca tio a s co d a pl
i ca tio
a s p rfo r d i a l
l p ri ts c ptfo r
p ri t .
l
a sta pl
i ca tio
a s c o s ra do l b t
. a d
fo r a c t ss ca s ,to a o id t a tt g tic pro gra la r d to r co g i a c
r co rd d so d l.
S o u n d L o c a liz a tio n fo r a H u m a n o id R o b o t
p. a ta fo r:
rr.
rr.
r.
ra i i
st Va l
.
ra i . st
I
a I
a I 2
- 7
(9 )
(9 )
II
a
a
a I 4
34 - 6 2
(27 )
( 8) (9 )
III
a I
a I
a 8
6 (9 )
( 8) (9 )
IV
a I
a I
a
3
(9 )
(9 )
( 8)
V
a II
a II a I 23
24 - 28
(27 )
(27 ) (9 )
a
rr.
Va l
.
-
tti s:
l
o o ps
2
4
3
4
5
4 - 56 4
38
-5
7 1
.
42 - 6 3 [36 , 45] 27
45 - 53 [36 , 45] 32
7
(45)
a
.
ta blis a s
a r o fpa ra t r s tti gs a d r s l
ts fo r p rits Ito . o l t o to fo r s o t so d ls s d a s tra i i g,t sta d
al
ida tio s ts, a d t
b r o fso d ls s d. I co l
to s
t
t ss o ft b sti diid a lca l
c l
a t d o t tra i i g,t sta d a l
ida tio s ts
ar s o .
fo l
l
o i gco l
s o st
b r o fit ra tio s o ft g tic
pro gra s. I
p ri t I a d
t
b r o fit ra tio s a s c o s ra do l i t gi it r a l
.
b rs i t l
a stco l
so t
b r
o fg ra tio s a c
p ri t a s r .
()
po p l
a tio co sist d o f4
i diid a l
s diid d ito 4 d
s.
()
9 % cro sso r pro ba bil
it a s s d.
pro ba bil
it fo r
ta tio s
to o cc r a s t sa .
pro ba bil
it fo r o o l
o go s cro sso r a s 4 % a d
t r a s o igra tio b t
d
s.
ra t o fcro sso r b t
d
s
as % .
()
o l co sta ta a il
a blto t g tic pro gra s a s -43 6
o i diid a l a s a l
l
o d to b l
o g r t a 2 b t s (32 bl
o cks).
2 .
(6 ) I
p ri tIa d IIt
t ss a s a la t d o t
t ss ca s s fro
a c r co rd d so d l. I t o t r p ri ts t a t
b r as
.
t
t ss a s a la t d o t t sts ta d tra i i g s tt t t ss ca s s
r ta k fro
a c r co rd d so d li t t sta d a l
ida tio s ts.
o r a ls cc ss r o bta i d
t tra i i g s t r fro
co rdi g
.
rro r fo r so d l
o ca l
ia tio o t tra i i gs t a s
rs al
lr t a
s 2 . is r s l
t a s o bta i d
al
li diid a l
s r
a la t d o
a ctl
t sa
t ss ca s s.
rro r o f 42
a s o bta i d
t
t ss ca s s
r c a gd b t
ac
a la tio o fa i diid a l
. is i dica t s t a tt
g tica l
l o l d pro gra o rtra i d a d
o ri d t
t ss ca s s i st a d
o fso li gt r a lpro bl o fso d l
o ca l
ia tio .
7 2
R . K a rls s o n , P . N o rd in , a n d M . N o rd a h l
130
length (bytes)
120
110
100
90
80
70
60
0
50
100
150
200
generations
250
300
350
400
0
50
100
150
200
generations
250
300
350
400
100
error (degrees)
80
60
40
20
0
s l
ts fro
p ri tI.
pp r dia ra s o s t l t i t s o ft
sti diid a l(so l
id) a d t a ra l t (da s d).
l
o r dia ra s o s t
fit ss o ft
sti diid a lo t tra i i s t(so l
id) a d t t sts t(do tt d), t
a ra fit ss (da s d) a d t
dia fit ss (da s do t).
140
length (bytes)
120
100
80
60
40
0
20
40
60
80
generations
100
120
140
160
0
20
40
60
80
generations
100
120
140
160
error (degrees)
100
80
60
40
20
t
t
t
t
s l
ts fro
p ri tII.
pp r dia ra s o s t l t i
ts o f
sti diid a l(so l
id) a d t a ra l t (do tt d).
l
o r dia ra s o s
fit ss o ft
sti diid a lo t tra i i s t(so l
id), t t sts t(do tt d) a d
al
ida tio s t(da s d l
o r). Ita l
so s o s t a ra fit ss (da s d pp r) a d
dia fit ss (da s do t).
S o u n d L o c a liz a tio n fo r a H u m a n o id R o b o t
7 3
130
length (bytes)
120
110
100
90
80
70
0
50
100
150
200
generations
250
300
350
400
0
50
100
150
200
generations
250
300
350
400
120
error (degrees)
100
80
60
40
20
0
t
t
t
t
s l
ts fro
p ri tIII.
pp r dia ra s o s t l t i
ts o f
sti diid a l(so l
id) a d t a ra l t (do tt d).
l
o r dia ra s o s
fit ss o ft
sti diid a lo t tra i i s t(so l
id), t t sts t(do tt d) a d
al
ida tio s t(da s d l
o r). Ita l
so s o s t a ra fit ss (da s d pp r) a d
dia fit ss (da s do t).
140
length (bytes)
120
100
80
60
40
0
50
100
150
generations
200
250
300
0
50
100
150
generations
200
250
300
error (degrees)
80
60
40
20
0
t
t
t
t
s l
ts fro
p ri tIV.
pp r dia ra s o s t l t i
ts o f
sti diid a l(so l
id) a d t a ra l t (do tt d).
l
o r dia ra s o s
fit ss o ft
sti diid a lo t tra i i s t(so l
id), t t sts t(do tt d) a d
al
ida tio s t(da s d l
o r). Ita l
so s o s t a ra fit ss (da s d pp r) a d
dia fit ss (da s do t).
7 4
R . K a rls s o n , P . N o rd in , a n d M . N o rd a h l
0
-50
-100
-150
100
0
50
-20
0
-40
-50
0
20
40
60
20
0
-100
-60
0
20
40
60
-80
20
60
0
40
-20
20
0
20
40
60
0
20
40
60
0
20
40
60
-20
-40
-60
0
20
40
60
-40
80
100
60
0
40
-100
0
20
40
60
0
100
80
60
20
0
20
40
60
-200
40
0
20
40
60
20
6
s l
ts fo r p ri tIV. a c o ft dia ra s s o s t a l ca l
c l
atd
fro t o tp ts o ft
sti diid a lo o
fit ss ca s fro t t sts t.
a l
is ca l
c l
a t d a ft r a c
c tio o ft i diid a li t l
o o p c ptfo r t first.
t sts t a s id tica lto t tra i i s t. I a c o ft
i dia ra s t
o rio ta l
l
i s o t co rr cta s r.
50
0
-50
-100
0
20
40
60
50
100
100
50
50
0
0
-50
-50
-100
0
20
40
60
-100
50
100
0
0
50
-50
-50
0
-100
0
20
40
60
100
-100
0
20
40
60
80
20
40
60
0
20
40
60
0
20
40
60
100
60
50
-50
0
80
40
0
-50
60
20
0
20
40
60
0
0
20
40
60
40
7
s l
ts fo r p ri tIV. a c o ft dia ra s s o s t a l ca l
c l
atd
fro t o tp ts o ft
sti diid a lo t r fit ss ca s s fro t t sts t.
t r fit ss ca s s co rr spo d fro so ds fro t sa
dir ctio
tfro diff r t
dista c s.
dista c s
r
(so l
id),2 (do tt d) a d 3 (da s do t).
a l
is ca l
c l
a t d a ft r a c
c tio o f t i diid a li t l
oop
c pt fo r t first.
I a c dia ra , t t sts t
r id tica lto t tra i i s t. I a c o ft
i
dia ra s t
o rio ta ll
i s o s t co rr cta s r.
S o u n d L o c a liz a tio n fo r a H u m a n o id R o b o t
7 5
sc ss
rstr a rk t a t ca tio s o l
d b o bs r d
co pa ri g t r s l
ts
fro t diff r t p ri ts, si c t diff r t tra i i g s ts r r co rd d
it diff r to t r a rs.
rs l
ts o f[4] s o t a tift s a p o ft
a
pi a is c a g d o g , o r a bil
it to l
o ca l
i so ds is dra a tica l
lr d c d.
t s sta rt it
p ri t .
fa ct t a t t
t ss s fo r t b st
i diid a l a la t d o t tra i i ga d t sts ts a r a lo st q a l
ls al
l(s
ta bl ), a st a tt g tic pro gra g ra l
i sto a l
lso dsi t tra i i g
s t. is is a l
so t ca s fo r t o t r p ri ts c ptfo r p ri tII, a s
ca b s
i g r s 2 to 5.
I ta bl o
ca s t a tt
o l d pro gra i p ri t do s o t
g ra l
i to a
a sp a ki g o ic ( o
a ds)
tra i d o sa to o t
a s.
is is o ts rprisi g si c itis a rd fo r t g tic pro gra
to
l
o ca l
i a sa to o t a so d co i gfro diff r tdir ctio s a d dista c s,
s
g r 7 . is g r a l
so s o s t a tt g tic pro gra do s o tg ra l
i
t at l
lto so ds co i g fro diff r tdista c s.
diffic l
t is pro ba bl
d to c o s fro
al
l
s a d f r it r i t ro o .
ra i i go a
a
o ic a s a l
so diffic l
t, s
g r 3.
b st o l d
pro gra
a s a t ss o f4 .
t ss ca b co pa r d to t a to bta i d fro
a pro gra t a ta la s gi s t co sta to tp t a l = .
t
t ss
is 5 , ic s o s t a tso
la r i g a s o cc rr d.
I gr 6 o
ca s t a tt a glfro o
i diid a la ris p rio dica l
l
a la t d o o
t ss ca s , ilco
rgi g to t rig ta s r.
p rio d is t sa
a s t a to ft sa to o t a t i diid a lis a lt d o .
is i dica t s t a tt g tic pro gra
s s t II i so
a . is p rio dic
pa tt r a s lss cla r
t g tic pro gra s r
o l d o so d co i g
fro diff r tdista c s,s
g r 7.
r a so fo rt is a b t a tt g tic
pro gra s
stla r to co p
it c o s, ic
a k s it o r diffic l
tto s
t II .
r
a so
t a titis po ssiblto
o l a s al
l a c i co d pro gra t a t
ca it rpr ta
pro c ss d str a o fsa pld st r o so d it o ti s rtio
o fdo a i k o ldg o r o t r str ct r .
o l d pro gra ca l
o ca l
i dir ctio o fa so d it a r i gs cc ss.
a rti cia lso d t i diid a l
s a
si il
a r p rfo r a c to
a s
as r d as a g l
a r r so ltio , b t c r s a rc r a i s fo r t is p rfo r a c to o l
d
d r o r r al
istic s tti gs a d
co ditio s.
c
ts
is r s a rc
a s s ppo rt d b
l
op
t.
K, t
dis
o a rd fo r
c
ol
o gica l
7 6
R . K a rls s o n , P . N o rd in , a n d M . N o rd a h l
r c s
. Iri . .
s
i i
i i
i r
r i
s
r
i
, t sis, I , a
rid
9 9 3.
2. l
a rt .
i
ri
h s h h si s
i i , I
r ss, 9 9 7 .
3. a
a f ., t.a l
.
i r r
i , dp kt. rl
a a d or a
a f a
l
is rs, H id
r a d a
ra sico , 9 9 8.
4. H o f a
. ., t.a l
., la r i so d l
o ca l
ia tio
it
a rs,
r
rsi
, 5, pp. 47 -42 , 9 9 8.
t
c
iit f
t V r L r -Sc
dr a s
osr a d
tic
tr
. a ra si a
rit s
cti
urt
r an
s a rc
nt r f r rtificia lInt l
l
ig nc
676
Ka is rsl
a ut rn, r a n
moser@dfki.de
pa rt nt f
put r cinc a nd ut a tin
India n Institut f cinc
a nga lr - 56
2, India
mnm@csa.iisc.ernet.in
. a tur
lctin is a r pr ising pti isa tin stra t g
f r a tt rn
c gnitin s st s. ut, a s a n
-c plt ta sk, itis tr
l difficul
tt ca rr ut. a ststudis t r f r
r ra t r l
i it d
in it r t ca rdina l
it ft f a tur spa c r t nu
r fpa tt rns
util
is d t a ss ss t f a tur su s tp rf r a nc .
is stud
a in s t sca l
a il
it f istri ut d
n tic l
g rit s
t r l
a rg -sca l a tur
lctin. s d a in fa ppl
ica tin, a cl
a ssifica tin s st f r ptica l a ra ct rs is c s n.
s st is ta ilr d
t cl
a ssif a nd- ritt n digits, in ling7 6
ina r f a tur s. u t t
a stn ss ft in stiga t d pr l , t is stud f r s a st p int n
r a ls in a tur
lctin f r cl
a ssifica tin.
pr s nt a s t f cust isa tins f
s t a t pr id f r a n a ppl
ica tin f kn n c nc pts t
a tur
lctin pr l s f pra ctica l
int r st.
l
i ita tins f
s in t d a in f a tur
lctin a r
unr a ld a nd i pr
nts a r sugg st d.
id l us d stra t g t
a cc lra t t pti isa tin pr c ss, ra ining t a pl
ing, a s s r d
t fa ilin t is d a in fa ppl
ica tin.
p ri nts n uns n a l
ida tin da ta sugg st t a t istri ut d
s
a r ca pa l fr ducingt pr l c pl it significa ntl.
r sul
ts
s
t a tt cl
a ssifica tin a ccura c ca n
a inta in d
ilr ducing
t f a tur spa c ca rdina l
it
a ut5 % . n tic l
g rit s a r d nstra t d t sca l l
l
t r l
a rg -sca lpr l sin a tur
lctin.
tr
cti
r c td l
op
ti I fo r a tio
c ol
o g a s ro ug ta o uta tr
do us flo o fi fo r a tio . o pa is,o rga isa tio s a d i diidua l
s ar l
it ra l
l
g tti gdro
d i a flo o d o f a sur
ts.
a tt r
co g itio ,a d i pa rticul
a rpa tt r cl
a ssi ca tio ,o ff rspro isi g
a s to a rds a o r r
d a o f d al
i g it i fo r a tio . I o rd r to
i pro t pro c sso fcl
a ssi ca tio , a tur
lctio
a
a ppl
id to i cr a s
t t ro ug puto fp rti ti fo r a tio . Itdo s o to lr duc t a o u to f
S . C a g n o n i e t a l. ( E d s .) : E v o W o r k s h o p s 2 0 0 0 , L N C S 1 8 0 3 , p p . 7 7 − 8 6 , 2 0 0 0 .
S p rin g e r-V e rla g B e rlin H e id e lb e rg 2 0 0 0
7 8
A . M o s e r a n d M .N . M u r ty
da ta ta k ito a cco u tfo ra I fo r a tio
ro c ssi gta sk, ut a a l
so i pro
t o ra l
lpro c ss a ccura c .
Ho
r pro isi g t s o s r a tio s a
, t r a r s rio us c a l
lgs
to
t: a tur
lctio is k o
to
a
-co plt pro l . is fa ct
a k s it tr
l difficul
tto
a i l
a rg -sca l,r a l
- o rl
d do a i s.
ic t
a rl studis o f a tur
lctio
t o ds i t s
tis, t
o tio o f a t“ l
a rg -sca l” a ctua l
l a s a s c a g d.
rst p ri ts
d al
t it ra t r sm llpro l s o fso
t s o ffa tur s. it t i cr a s o f
co puta tio a lpo r, studis o flr s l ta sks (
- 5 fa tur s) ca
po ssi l. o da , do a i s it o r t a 5 fa tur s ca
sa id to
trul
c al
l gi g;su s qu tl, suc pro l s s a l
l ca l
ld r lr s l.
r a r a sica l
l t o a s o fca rr i g o ut a
a tur
lctio pro c ss:
il rm
sa d r
rm
s [ ].
fo r r cl
a ss o fa ppro a c s r l
is
o g ra lsta tistica lpro p rtis o ft pro l do a i a d a stra cts fro pa rticul
a r cl
a ssi r o d l
s. ra pp r t o ds o t o t r a d o pti is t p rfo r a c o fa co cr t cl
a ssi r.
r a s il
t r t o ds t pica l
ld l
i r or
g ra lr sul
ts, ra pp r t o ds a r co sid r d to il
d a tt r cl
a ssi ca tio
p rfo r a c fo r sp ci c ta sks [2].
us,
a i i ga ta ig cl
a ssi ra ccura c , ra pp r t o dss
to
t
o r pro isi gca dida t s. utt a sta ss ss
tti fo r cl
a ssi r p rfo r a c ltsuc a ppro a c sa pp a r i fa si lfo r r a l
- o rl
d a ppl
ica tio s a tt
rstgl
a c .H o
r,a d a c s i t r a l o f istri ut d rti cia lI t l
l
ig c
pro id
a s to a ppl ra pp r t o ds to c a l
l gi gr a l
- o rl
d pro l s.
fo l
l
o i gpa g s il
ld scri a succ ssfula ppl
ica tio o f istri ut d
tic l
go rit s -a ra pp r t o d -to a r l
a rg -sca lpro l i a tur
lctio fo r a
s st . irst, t g ra lcusto isa tio o f
s fo r a tur
lctio is o utl
i d. Itis fo l
l
o d
a d scriptio o fo ur ta il
o ri go f
s
fo r r l
a rg -sca l a tur
lctio .
p ri ta ls ctio d o stra t s t
us ful ss o ft s tup. i a l
l, t r sul
ts a r su
a ris d a d po ssi l t sio s a r sugg st d.
tic
rit
st
ct
tr s
s a pro d to
ca pa lo fd a l
i g it
-co plt pro l s i a rio us
l
ds. idlck a d l
a sk sugg st d t ir a ppl
ica tio to
a tur
lctio fo r
cl
a ssi ca tio [3]. t r ra pp r (a d il
t r)
t o ds a
a ppl
id i
a rio us studis [ ]. ut sp cia l
l
i gfa c d it r d a di gl
a rg sca lta sks,
s a pp a r to
t
stc o ic du to t ir i r tpa ra l
ll
is
a d o i a ld a d fo r a ckgro u d k o ldg .
.
si
i p
s
I t ir ilsto
pa p r, idlck a d l
a sk us d a i pl
to g ra t
pro isi gfa tur su s ts.
a i o ft irstud
a sto r duc t fa tur spa c
ca rdi a l
it a d a tt sa
ti to k p t pr dicti a ccura c a cc pta l ig .
O n th e S c a la b ility o f G e n e tic A lg o rith m s
7 9
o a ppl
s, t t o r s a rc rs “ co r” t pa tt r s it a i a r
a sk
t ro ug
ic t cl
a ssi r p rc i s t sa pls - t fa tur su s t. ’’
i dica t s t pr s c a d a ’’t a s c o fa fa tur .
c ro o so
is
o ta i d
tru ca ti g t ro s o ft fa tur
a sk. idlck a d l
a sk ’
s
o d las
succ ssful
l a do pt d a d r
d
a o t r r s a rc rs.
co
o l a ppl
id t ss fu ctio fo r a fa tur su s t co sists o f a
co
i a tio o fcl
a ssi r a ccura c a d su s tco pl it :
( )=
()
( )−
I t is fo r ul
a,
d o t s a a ccura c sti a t fo r t i sta tia t d cl
a ssi r, a d ( ) is a
a sur fo r t co pl it o ft fa tur s t- usua l
lt
u
r o futil
is d fa tur s. urt r o r ,
il
ds t fa tur spa c ca rdi a l
it,
a d
isa pu is
tfa cto rto
ig t
ul
tiplo jcti so ft t ssfu ctio .
u
r o ffa tur s us d
a su s tis it d d to la d t a l
go rit
to r gio s o fs a l
lco pl it.
.
i
is ri u
s
s tup a s d scri d a o
o rks pr tt
l
lfo r fa tur spa c s o fco pa ra ti l s a l
lca rdi a l
it. ut
i g co fro t d it
r l
a rg -sca ldo a i s, t a sta ss ss
tti fo r fa tur su s ts pro i its t is a ppro a c .
u c t a l pl
o d istri ut d
s fo r a tur
tra ctio i o rd r to
stud
ig r-o rd r pro l s [4] .
us d a icro ra i istri ut d
to
a cc lra t t
al
ua tio pro c ss, il
di gup to l
i a r sp dups.
U si gsuc co
tio a l icro ra i
s, a fa tur su s tis ra t d
s di gitto a r o t a l
ua to r.
su s t’
s pr dicti a ccura c is sti a t d t r
ru i ga co plt cl
a ssi r.
r sul
ts a r t
r tur d to a a st r o d
a d us d to guid t
.
c tl, istri ut d
ri l
s r sugg st d i o rd r to a cc lra t
t
al
ua tio pro c ss fo r cl
a ssi r o pti isa tio s i furt r [5].
us o ft is
al
ua tio sc
fa cil
ita t s fo r a stud o f r l
a rg sca ldo a i s i a tt r
co g itio . is o d lca
cia l
l a ppl
id to
a tur
lctio .
.
s
u is
ro to da ’
s p rsp cti , pa ststudis a ppli g
lctio
r l
i it d i a rio us a s:
a jo rit o fr s a rc rsi stiga t d do
t a tis l
o
fa tur s [3,6 – ].
tic
l
go rit
s to
a i s o fra t rs a l
lco
ug unc t a l
. ca l
lt ir c ntri utin
rth r s a rch
tin a d l
a ssifi ca ti
si g
tic l
grith s , t
in fa ctca rr
tra cti pr c ss.
a tur
pl it,
a tur
uta
lcat r
8 0
A . M o s e r a n d M .N . M u r ty
tudis o rki go l
a rg sca la d r l
a rg sca ldo a i sr l
id o a ra t r
s al
lu
ro ftra i i gda ta [9 – ],([ 2]).
r- tti gt d cist us r
quit l
ik l. a l
ida tio
a s usua l
l o tr po rt d.
o pa ra ti studis o f
tic l
go rit s a d o t r t o ds r it r
as d o s al
l
-sca lpro l s [ , 3, 4], o r t circu sta c s fo r
s do
o ta pp a r to
cia l[ 2].
ais i
su s qu t s ctio s il
ld scri
a s to stud
r l
a rg -sca l do a tur
lctio
ila o idi gt a o l
i ita tio s.
3
r s
r
r
c
is
a ppli g
s to do a i s o fl
a rg co pl it i
a tur
lctio , t
a jo r c a l
l g is t o r
li gti co pl it o ft pro c ss.
.
r
s
i
o ug t ti r quir
ts fo r t
t ss a l
ua tio ca
dra a tica l
ld cr a s d
irtu o f istri ut d ( rtica l
)
s, t dura tio o ft a ss ss
t
pro c ss r a i s t t r a t i gl
i ita tio o ft s a rc pro c ss.
t us
a s to i
tgo o d so l
utio s i f r c cls a s co pa r d to o t r l
ds o fa ppl
ica tio .
crucia ltra d o ff is to k p t
al
ua tio pro c ss a s t si a s
r quir d - a d a s co cis a s po ssi l.
.
I p cs
s
si
I o ur stud , s ra l a sur
ts r ta k i o rd r to co p up it t s
r quir
ts. o sto ft s
a sur
ts a
r po rt d fo r ;itis t ir
ca r fulco
i a tio t a tfa cil
ita t s fo r a succ ssfula ppl
ica tio s o f
s to
r
l
a rg -sca l a tur
lctio .
mi
r rs [ 5]pro id a ffici t a s to
fo rc co
rg c o f
t s a rc pro c ss. is is c ssa r a s t ti r quir
ts to a ss ss t
i diidua l
s pr
tt
fro co
rgi gi a “ a tura l
” a .
ft ra a turi gp a s o fa o uto - u dr d g ra tio s,t s lctio pr ssur a s i cr a s d i o ur s tup i o rd r to la d t a l
go rit to a - c ssa rill
o ca l
-o pti u . is a sa c i d
r duci gt
uta tio ra t a d
i cr a si g cro sso r ra t a d i fo r a tio it rc a g
t
i diidua l
s.
i
d scri s furt r d ta il
s o t pa ra t r s tti gs.
I o rd r to
a c t i itia l
isa tio pro c ss, t ra do
g ra to r a s
a ipul
a t d a cco rdi gto a
pl
icit ii lis i
i s. I its usua li sta tia tio ,t i itia lpo pul
a tio o ft
is o ta i d
to ssi ga fa ir co i fo r
r i diidua la d
r fa tur . I diidua l
si r l
a rg -sca la ppl
ica tio s il
lr fl ctt is c o ic i t a tt
il
l s ttld a ro u d t 5 % l l
o fco pl it.
al
go rit
i gfo rc d to co
rg fa st o ul
d t us a
a po o r i itia lco ra g o ft s a rc spa c .
O n th e S c a la b ility o f G e n e tic A lg o rith m s
U si g a I itia l
isa tio
ia s, t pro a il
it fo r a ’’to o ccur i t
po pul
a tio is c o s a cco rdi gto
( )=
( )= (
−
+
o
)
8 1
i itia l
o
o
I t s qua tio s, t I itia l
isa tio
ia s ra g s fro
to
- it r
li a rl o r
o
tia l
l. o i d o t s t
u
r o fi diidua l
si t
po pul
a tio a d it
u
r o fi diidua l (
i
o i );
ca
a djust d usi gt fr pa ra t r n.
s t co pl it o ft i diidua l
s ca
p ct d to r fl ctt is ia s,t
s a rc spa c co ra g a
l
a rg d
t is a sur
t.
pa rt fro
t sta da rd uta tio o p ra to r, s
r
s m
i
as
a ppl
id. ro ppi g t a ssu ptio o f qua l uta tio pro a il
itis
a d
( d o ti gt pro a il
it o f’
i’to
c a g d ito ’
j’
) fa cil
ita t s fo r
a
pl
o ra tio o fa ro a d r ra g o ft s a rc spa c .
t di g t s to fusua l
l a ppl
id t ss fu ctio a s itro duc d i t
fo r r s ctio , l mi l
is m
i s r a ppl
id:
( )=
( )−
(
()
) −
()
op
a s t a t t is
a sur
t o ul
d
co ura g t a l
go rit
to
pl
o r r gio s o fs a l
lco pl it pr fra l.
o a o id o sitio a l ia s, i mi l rss r[ 6 ] a s us d:
i diidua l
s
a r co
i d
to ssi ga ia s d co i fo r
r g . sid s t is a sur t,
a d
a ppi gs r t st d to
co ura g t pro ductio o f
i diidua l
si l
o r a d upp r ra g s o fco pl it o ft s a rc spa c .
l
so sugg st d i [ 6 ] a d id l us d i
a tur
lctio studis, ra ii g t a pl
ig (
) a s trid i o rd r to a cc lra t t p rfo r a c
a ss ss
t.
id a is to ra do ls lcta su s to ft cl
a ssi r tra i i g
da ta fo r sti a ti gt cl
a ssi ca tio a ccura c o ft su s ts. s du to
t r -sa pl
i gt
t ss a sur c a g s i t co a rs o ft s a rc pro c ss, i diidua l
s i rit d fro fo r r g ra tio s a to
r - al
ua t d.
I t l
it ra tur ,
a s r po rt d to sp d up t ss a l
ua tio
it o ut
itro duci gto o uc o f o is ito t s a rc pro c ss
p ri
ts
o al
ida t t us ful ss o ft custo isa tio s d scri d a o ,t sts r co duct d,ta ckl
i ga r l
a rg sca lta ski a tur
lctio .
ti -co su i g
p ri ts r co duct d a tt
r a
s a rc
t r fo r rti cia lI t l
l
ig c , Ka is rsl
a ut r , r a .
8 2
A . M o s e r a n d M .N . M u r ty
.
up
I o ur stud
i stiga t d a a tur
ur cl
a ssi r a s ta il
o r d to ca t go ris
lctio pro l fo r a
a d ritt digits.
s st .
s
la r i gda ta co sist d o f
i a r pa tt r s it 24 32
pi l
s, diid d ito t cl
a ss s o f qua lca rdi a l
it.
da ta s t a s
us d
s ra l
ti s fo r i t
co
u it, o str c tl
ra ka s a d urt
[ 7 ] a d a ra d i[ ].
I o rd r to i stiga t t us ful ss o f s i t is r l
a rg -sca ldo a i ,
t la r i gda ta a s ra do lspl
itito a 5 % tra i i gs t,a 3 % t sts t,a d
a 2 % al
ida tio s ta tt
gi i go f a c ru .
sa pls r o r a l
is d
to qua lsi i a pr pro c ssi g st p, usi g a stra ig tfo r a rd sca l
i g pro c ss
ic pr s r d t
sp ct a tio o ft pa tt r s.
ssi r o ra t a s to ffa tur s,t t sts t a s ca t go ris d a cco rdi g
to t tra i i g s t usi g a
- a r st- ig o ur l
a ssi r.
ra lstudis
pro d t is o d lto
a ro ust, ffici t a d
l
l
-p rfo r i g a ppro a c fo r
-a ppl
ica tio s[ 9 ]. o o la
ista c usi gt
o p ra to rfa cil
ita t d fo r
a it-pa ra l
ll a ppi go ft pa tt r s. o ug co sid ra l ffo rt a d
sp t
to a cc lra t t cl
a ssi ca tio pro c ss,t ti r quir d to a ss ssa si glfa tur
su s ta o u t d to o - u dr d- ft s co ds o a U
U
o rksta tio .
o il
l
ustra t t i pa cts o ft is gur : i t po pul
a tio to co sisto fo u dr d i diidua l
sa d t fa tur spa c ca rdi a l
it to
7 6 ,a si t pro l
i stiga t d
r,a
a usti s a rc usi g a si gl a c i
o ul
d r quir
a o ut
a rs!
uri g t o pti isa tio pro c ss, t
al
ida tio s t a s ful
l k pta sid . It
a s us d to judg t g ra l
it o ft r sul
ts a ft r t
t r i a t d. u to
t a str so urc r quir
ts,co pa ra ti studis it o t r a tur
lctio
t o ds r o tco duct d. H o
r, t ca r ful p ri ta ls tup it a
s pa ra t a l
ida tio s tfa cil
ita t s fo r a o jcti
a sur
to ft
t o d’
s
p rfo r a c .
istri ut d
rtica l
a s ru o t irt U
o rksta tio s o f
tp
- ,
-2 a d U
. Its a da pti l
o ad al
a ci gpro id d
fo r a
ffici t us o f t
t ro g o us ul
ti-us r
iro
t. o r furt r
d ta il
s a o utt co c pto f rtica l s t r a d r is r frr d to [5] a d [2 ].
po pul
a tio o fo - u dr d i diidua l
s a s a ita i d. o s lctt pa r ts o ft
g ra tio ,a co
i d d t r i istic a d ra do stra t g a s
us d:
sti diidua l
s r co pid dir ctl ito t
tg ra tio , a d
t r a i i g pl
ac s i t
a ti g po o l r l
ld
usi g a fo rtu
l
a s d o sca ld t ss a l
u s. is sca l
i g a s do
suc t a tt diff r c s
i cl
a ssi ca tio a ccura c
r a pl
i d.
i i a l t ss a l
u
a s a dd d to
p r itfo r a s lctio o fi diidua l
s it l
o pr dicti a ccura c .
fo l
l
o ig
O n th e S c a la b ility o f G e n e tic A lg o rith m s
qua tio s o s o t
t ss a l
u
to
t a ra g t ss a d
t
( )=
( ) − (2 ¯−
t ss
8 3
( ) o fa i diidua lp is sca ld, gi
a i a l t ss i t po pul
a tio :
)+
t ss if
t r is
a l su
a ris s t
o st i po rta t pa ra t r s tti gs i t co a rs
o fa t pica lru .
d o ts t
u
r o fg ra tio s pa ss d, a d cro ss ia s
i dica t s t pro a il
it fo r t
i o ia lcro sso rto c o o s a g fro pa r t
o
fo r c il
do .
g
-5
( lra ti
5 -75
(gr i g
7 5( a t ri g
(c
.
rg c
a ra t r
in itn ss
uta tin a t
cr ss r a t
cr ss ia s
in itn ss
uta tin a t
cr ss r a t
cr ss ia s
in itn ss
uta tin a t
cr ss r a t
cr ss ia s
in itn ss
uta tin a t
cr ss r a t
cr ss ia s
al
.4
2-3%
7 5%
%
.2
%
%
7 5%
.2
.5%
%
6 %
ff ct
d ra t s lctin pr ssur
f cus n plra tin
d ra t c r
s
int ra ctin
l g n int ra ctin
incr a s d s lctin pr ssur
plra tin a nd c
ina tin
d ra t g n int ra ctin
f cus n a turing
incr a s d c r
s
int ra ctin
incr a s d g n int ra ctin
ig s lctin pr ssur
%
nf rc c n rg nc
% ig c r
s
int ra ctin
5%
ig g n int ra ctin
. na ic p ra t rs
su s
si
r ii
pi
ra i i g t a pl
i g(
) as
us d
a r s a rc rs i t
l
d o f a tur
lctio to co p up it t o r
l
i gti r quir
ts. is t c iqu a s
as l
ltrid fo r t purpo s s
o ft is stud .
tra i i gs ts a
r -sa pld a tt
% a d3 % l l
.
i r s o st
ol
utio o ft
st t ss a l
u s o t t stda ta , t a tis
t da ta us d i t o pti isa tio pro c ss its l
f.
tra i i gs t a s r -sa pld
i 3 % ra tio s. o t t a tdu to t is r -sa pl
i gt
t ss fu ctio
a ris fro
g ra tio to g ra tio .
sca
s
i t gur , o pro gr ssi t r so ft p rfo r a c crit rio
o t t stda ta is o ta i d.
a sur
ts a ki g us o ft
al
ida tio da ta
tur d o utto
o r disa stro us.
o is itro duc d to t s a rc pro c ss
s o d to
to o a stfo r t is ig -di sio a lda ta s ti co ju ctio
it t
ig u
r o ftra i i gi sta c s. us,
as o t
furt rl co sid r d
i t is stud .
8 4
A . M o s e r a n d M .N . M u r ty
91.5
91
Accuracy
90.5
90
89.5
89
88.5
88
0
20
i. .
40
60
80
100
Generation number
120
stfitn ss a l
u s using ra ining
140
160
180
t a pl
ing
rc
p c
r
U si gt sta da rd o p ra to rs,t
as o s r d
to tra c i diidua l
s a rt 5 % l l
o fco pl it pr fra l. is a
l
ad to t p cul
ia ritis o ft pro l do a i ,suc a s ig o - o o to o usit
o ft
t ss fu ctio , ig i s sitiit o ft
l
a ssi r to c a g s i fa tur spa c ca rdi a l
it,t l
i it d a turi gp a s a d s a rp p rfo r a c p a ks
i ra g s o fl
o co pl it.
o gt
a sur
ts sugg st d i t fo r rs ctio ,a s c ro o us uta tio tur d o utto
t
o st ff cti
a sto ro a d t
a i d r gio .
i co rpo ra tio o fa pu is
tfa cto r ito t
t ss fu ctio a l
o
as o t
ca pa lo fguidi g t a l
go rit to o t r r gio s o fit r st.
us o f
a ppi gs i suppl
t o fso
o ft cro sso r o p ra tio s a s o s r d to
la d to pr a tur co
rg c :
al
go rit go tstuck i su o pti a lr gio s
it pr dicti a ccura c
l
o t ful
lfa tur s t.
c
ii
urt r p ri ts usi gt ful
ltra i i gs tfo r a ccura c sti a tio
r co duct d.
r sul
ts s o t a t
s sca l l
lto do a i s o fl
a rg
co pl it i a tur
lctio :
a s o fd a ic o p ra to rs, t
co rg d it i a o utt o - u dr d g ra tio s.
u
r o futil
is d fa tur s
co ul
d
r duc d
a o ut5 %
ilpr s r i gt pr dicti a ccura c o ft
cl
a ssi r.
i r 2 s o st d l
op
to ft p rfo r a c o t u s
al
ida tio
pa tt r s: ft r r a c i ga c rta i d gr o f a turit,t ca l
cul
a t d fa tur su -
O n th e S c a la b ility o f G e n e tic A lg o rith m s
8 5
s ts d l
i r cl
a ssi ca tio a ccura cis co pa ra lto o r
tt r t a t ful
l
fa tur s t. is r a rka lr sul
tco ul
d
r pro duc d r l
ia li t co s cuti ru s.
93.2
Calculated feature set
93
Accuracy (%)
92.8
92.6
92.4
Full feature set
92.2
92
91.8
91.6
0
20
40
50
60
80
100
120
Generation number
140 150 160
180
200
i. . a l
ida tin p rf r a nc
r
t
k
is r po rta i d a ti stiga ti gt us ful ss o f
tic l
go rit s fo r r
l
a rg -sca l a tur
lctio .
ra lpa rtia l
l al
r a d r po rt d
a sur
ts
r co
i d i o rd r to co p
it t co pl pro l .
o
l
i ita tio s o f s fo r r l
a rg -sca l a tur
lctio pro l s r
o sr da d
a sto o rco
t
r sugg st d. p ri tss o d t a t
s sca l l
lto do a i s o fl
a rg co pl it i a tur
lctio .
id l a ppl
id a ppro i a tio
t o d, ra i i g t a pl
i g, a s o s rd to fa ilgro ssli o ur a ppl
ica tio . o l
l
o i gt
a sic id a o ft is a ppro a c ,
o r suita lcusto isa tio s s o ul
d
trid i futur studis. I st a d o fra do lr -sa pl
i gt tra i i gs t,a ro to t p
lctio pro c ss a
ca rrid
o utsi ul
ta o usl to
a tur
lctio . a ra d ia ppl
id t is id a succ ssful
l
to s a l
l
-sca l a tt r
co g itio pro l s [ ]. is co
i a tio a pp a rs to
pro isi gfo r r l
a rg -sca lta sks, a s l
l
.
utur o rk il
la i a ta stud o ffurt r a tur
lctio
t o ds i r
l
a rg -sca ldo a i s.
pro l o f a tur r a tio
il
l a ddr ss d i o rd r
to
a c t s a rc pro c ss i furt r.
8 6
A . M o s e r a n d M .N . M u r ty
fr c s
.
2.
3.
4.
5.
6.
7.
.
.
.
.
2.
3.
4.
5.
6.
7.
.
.
2 .
. a s a nd . iu. a tur s lctin f r cl
a ssifica tin. I t ll
ig t a ta
a lsis,
7.
.
n, . K a i, a nd K. fl g r. Irr l a ntf a tur s a nd t su s ts lctin
pr l . rc di gs fth I t r a ti a l
fr c
a chi
a r i g, ,
4.
. idlckia nd . kl
a nsk . n t n g n tic a l
g rit s f r l
a rg -sca lf a tur
s lctin. a tt r
c g iti
tt rs, :335–347 ,
.
unc ,
d a n, i, a i ia - un, .
l
a nd, a nd . n d . urt r r s a rc n f a tur s lctin a nd cl
a ssifica tin using g n tic a l
g rit s. rc di gs
fth 5th I t r a ti a l
fr c f
tic l
grith s,
3.
.
s r.
distri ut d rtica lg n tic a l
g rit
f r f a tur s lctin. ifth
I t r a ti a l
fr c
c
t a lsis a d
c g iti
s a rch
r ,
.
. lt ing r.
a tur s lctin
g n tic a l
g rit s. II
rt ri s, 36 ,
3.
. ra ka s a nd . . urt. a tur s lctin t i pr c cl
a ssifica tin a ccura c
usinga g n tic a l
g rit .
r a l fth I dia I stit t f ci c ,
7.
. K. a in a nd . Z ngk r. a tur s lctin: a l
ua tin, a ppl
ica tin a nd s a l
l
sa plp rf r a nc . I
ra sa cti s
a tt r
a lsis a d a chi I t l
l
ig c , (2),
7.
. .
it , . . ga rt,a nd I. .
ns n.
n tic f a tur s lctin f r cl
ust ringa nd cl
a ssifica tin. rc di gs fth I
llq i
tic l
grith s
i I a g rc ssi g&
isi ; I
igst
4/ 3,
4.
. u rra - a l
c d a nd .
itl .
n tic s a rc f r f a tur s lctin: c pa ris n t n
a nd
I . rc di gs fth
si
tic
lgrith s,
.
. a nga nd . na a r. a tur su s ts lctin usinga g n tic a l
g rit . a t r
tra cti
str cti a d lcti a ta i i g rs cti ,
.
. . rri, . udil
, . a t f, a nd . Kittlr.
pa ra ti stud ft c niqu s
frl
a rg -sca lf a tur r ductin. a tt r
c g iti i ra ctic I ,
4.
I. . I a a nd . a fa i. n
prica lc pa ris n t n gl a la nd gr d -l
ik
s a rc f r f a tur s lctin. rc di gs fth lrida I s a rch
si ,
4.
. I.
a ng a nd . . ipp a nn. sing g n tic a l
g rit s t i pr
pa tt rn
cl
a ssifica tin p rf r a nc . d a c s i
ra lI f r a ti
rc ssi g, 3,
.
. a sl , . . ul
l
, a nd . . a rtin. n
ri
fg n tic a l
g rit s;pa rt
2: s a rc t pics.
i rsit
ti g, 5(4):5 –6 ,
3.
. Z . ril
l
, . . r n, a nd . . a rtin. a stg n tic s lctin ff a tur s f r
n ura ln t rk cl
a ssifi rs. I
ra sa cti s f
ra l t rks, 3(2),
2.
. ra ka s a nd . . urt. r ingsu spa c pa tt rn r c gnitin
t ds a nd
t ir n ura l
-n t rk
d l
s. I
ra sa cti s
ra l t rks, ( ),
7.
. . a ra d i. a tt r
rs ta ti a d rt t
lcti i lassifi ca ti .
a st r
sis, pa rt nt f
put r cinc a nd ut a tin,India n Institut
f cinc , a nga lr ,
.
. lstr
, . K istin n, a nd . ja .
ura la nd sta tistica lcl
a ssifi rs -ta n
a nd t ca s studis. I
ra sa cti s
ra l t rks, ( ),
7.
.
s r. istri ut d g n tic a l
g rit s f r f a tur s lctin. i l a h sis
i rsit f a is rsl
a tr
r a ,
.
ii
ti
ssi c ti
r
rit
sis
cti ist
sfr
p
a ul . o sin
pa rt nto f o put r cinc
a rdiff ni rsit
K
P.L.Rosin@cs.cf.ac.uk
a nd
H nr . o ng sa
cho o lo f o puting
h ffi l
d H al
l
a
ni rsit
K
H.Nyongesa@shu.ac.uk
c . his pa p r pr s nts a n in stiga tin int th cl
a ssifica tin f
a difficul
tda ta s tc nta ining l
a rg intra -cl
a ss a ria il
it utl int rcl
a ss a ria il
it. ta nda rd cl
a ssifi rs a r
a k a nd fa ilt a chi sa tisfa ct r r sul
ts h
r, itis pr p s d tha ta c m ina tin fsuch a k
cl
a ssifi rs ca n impr
ra l
lp rf rma nc . h pa p r a l
s intr duc s
a n l l
utina r a ppr a ch t fu
rulg n ra tin f r cl
a ssifica tin
pr lms.
I tr
cti
his pa p r d scrib s a s ris o f p ri nts in ta c l
inga difficul
tcl
a ssi ca tio n
pro bl . h da ta co nsists o f a rio us b a ns a nd s ds, a pls o f hich a r
sho n in gur . l
tho ugh so
o f th o b jcts a r l
a rg r tha n o th rs ( .g.
a lo nds co pa r d to lntil
s)
a r int r st d in cl
a ssif ing th
ba s d o n
th ir sha p a l
o n itho utusinginfo r a tio n a bo utth ir si . his co rr spo nds
to th situa tio n h r th dista nc b t n th o b jcts a nd th ca ra is no t
d, a nd so th ir a ppa r nti a g d si s o ul
d a r . h difficul
t o fth ta s
l
is in th r l
a ti l s a l
lint r-cl
a ss diff r nc in sha p a nd th high intra cl
a ss diff r nc s. In o th r o rds,a l
lth o b jcts l
o o si il
a r, a pp a ringro ughl
l
l
iptica l
. l
tho ughth sha p so fso
o b jcts( .g.a lo nds) a r fa irlco nsist nt
o th rs a r co nsid ra bl( .g. co rn rn l
s).
h ba sisfo rcl
a ssif ingth o b jcts il
l
b a s to fsha p pro p rtis a sur d
fro th ir sil
ho u tt s. inc si info r a tio n is to b disca rd d th pro p rtis
n d to b in a ria ntto sca l
ing. i is , in a ria nc to po sitio n a nd o rinta tio n
S . C a g n o n i e t a l. ( E d s .) : E v o W o r k s h o p s 2 0 0 0 , L N C S 1 8 0 3 , p p . 8 7 − 9 6 , 2 0 0 0 .
S p rin g e r-V e rla g B e rlin H e id e lb e rg 2 0 0 0
8 8
P .L . R o s in a n d H .O . N y o n g e s a
cha ng s is n c ssa r . urth r o r ,it a b us fulto incl
ud in a ria nc to a dditio na ltra nsfo r a tio ns o f th sha p . o r insta nc , if th d t r ining sha p
fa cto r o fa cl
a ss is its si il
a rit to a n l
l
ips th n th a sp ctra tio a b irr l
a nt. h co put r isio n l
it ra tur pro id s a a rit o fsha p
a sur s [ ].
s lctio n o fth s , in co bina tio n ith so
n sha p pro p rtis d l
op d
b
o sin [2], ha b n a ppl
id to g n ra t a s to f 7
a sur
nts o f a ch
sa pl. h ca n b diid d into subgro upsa cco rdingto th irpro p rtisa nd/ o r
al
go rith s:
s
–fo ur a ttribut s in a ria ntto ro ta tio n,tra nsl
a tio n,a nd sca l
ing
(in a ria ntund r si il
a rit tra nsfo r a tio ns).
s 2 – thr a ttribut s in a ria ntto ro ta tio n, tra nsl
a tio n, sca l
ing,
a nd s
(in a ria ntund r a ffin tra nsfo r a tio ns).
s
r – fo ur sta nda rd a ttribut s – cc ntricit, circul
a rit, co pa ctn ss,
a nd co n it (in a ria ntund r si il
a rit tra nsfo r a tio ns).
r
– thr
a sur
nts o f l
l
ipticit, r cta ngul
a rit, a nd tria ngul
a rit (in a ria ntund r a ffin tra nsfo r a tio ns, si il
a rit tra nsfo r a tio ns
a nd str tchinga l
o ngth a s, a nd a ffin tra nsfo r a tio ns r sp cti l).
r
2 –thr a l
t rna ti
a sur
nts o f l
l
ipticit, r cta ngul
a rit,
a nd tria ngul
a rit.
i. . a mpls fda ta ; r s c nta in: ( ) a l
m nds, (2) chickp a s, (3) c ff
a ns,
(4) lntil
s, (5) p a nuts, (6 ) c rn k rn l
s, (7 ) pumpkin s ds, ( ) ra isins, ( ) sunfl
r
s ds.
In this pa p r
ha in stiga t d fo ur diff r nt cl
a ssi ca tio n t chniqu s,
a nd co bin d th in a n a tt ptto i pro o ra l
lcl
a ssi ca tio n.
o i plnta tio ns o fd cisio n tr s a r us d. h rst tho d is th
l
l no n 4.5
C o m b in in g E v o lu tio n a ry , C o n n e c tio n is t, a n d F u z z y C la s s ific a tio n A lg o rith m s
8 9
a chin la rning tho d d l
op d b
uinl
a n [3]. 4.5induc s tr s tha tpa rtitio n fa tur spa c into quia lnc cl
a ss s usinga is-pa ra l
llh p rpl
a n s ( .g.
in 2 this o ul
d co nsisto fho rio nta la nd rtica ll
in s). h s co nd a ppro a ch
is
[4] hich is a g n ra l
isa tio n in tha t, ra th r tha n ch c ingth a l
u o fa
singla ttribut a t a chno d ,itt sts a l
in a r co bina tio n o fa ttribut s. a tur
spa c isco ns qu ntlpa rtitio n d b o bl
iqu h p rpl
a n s. h third
tho d isa n
ns blo fn ura ln t o r s tha ta r tra in d o n diff r ntfa tur s pa tt rns, incl
udingo n n t o r tha tinco rpo ra t s a l
lpa tt rns. ina l
l, ha a l
so o l d
fu
ruls using o l
utio na r pro gra
ing.
2
rif
r i
f
ssi c ti
cti ist
cisi
r
ura l
n t o r sa r co nn ctio nists st stha ta r id lus d fo rr gr ssio n a nd
functio n a ppro i a tio n. h co pris inputno d s, hich a r us d to pro id
tra iningda ta to th n t o r a nd o utputno d s us d to r pr s ntth d p nd nt
a ria bls. Input- utput a pping is a chi d b a djusting co nn ctio n
ights
b t n th inputa nd o utputno d s,but o r usua l
l,thro ugha n int r dia t
l
a r o fno d s. his cha ra ct ristic ca n b
o di d fo r cl
a ssi ca tio n pro bl s b
sp cif ing d sira bln t o r o utputs to b bina r a l
u s. ura ln t o r s a r
rl
ia bl us d in cl
a ssi ca tio n o fco pl da ta .
u
cl
a ssi ca tio n isa rul-ba s d a ppro a chin hichI - H
rulsa r us d
to ca t go ris da ta . h ruls r l
a t g n ra l
is d o r i pr cis gro upings o finput
da ta ,a nd th d cisio n o fa gi n rulr pr s nts a d gr o fb l
o ngingto a gi n
o utput cl
a ss. his t p o fcl
a ssi r is pa rticul
a rl us ful h n itis n c ssa r
to pro id int rpr ta bil
it to th cl
a ssi ca tio n s st in th fo r o fl
inguistic
ruls. H o
r, th pro c ss o fcr a ting cl
a ssi ca tio n ruls is o ft n difficul
ta nd
ti -co nsu ing.
ra l
studis ha a tt pt d to co p
iththis pro bl using
la rninga l
go rith s,a nd in pa rticul
a r th us o fn ura ln t o r s.
cisio n tr s a r a
l
lsta bl
ish d t chniqu ba s d o n structuringth da ta ,
a cco rdingto info r a tio n th o r , into utua l
l cl
usi crisp r gio ns. l
a ssi ca tio n is g n ra l
l p rfo r d b sta rting a tth tr ’
s ro o tno d a nd ch c ing
th a l
u o fa singla ttribut , a nd d p ndingo n its a l
u s th a ppro pria t l
in
is fo l
l
o d. his pro c ss is r p a t d a tsucc dingno d s untila la fis r a ch d
a nd a cl
a ssi ca tio n is a ssign d. n ra ting th o pti a ld cisio n tr is g n ral
l
-ha rd, a nd th r fo r a sub-o pti a ltr is induc d inst a d, usinggr d
o r ra ndo is d hil
lcl
i bing a l
go rith s fo r insta nc . h r sul
ting tr is o ft n
prun d; subtr s a r r pl
a c d b la fno d s ifthis r duc s th
p ct d rro r
ra t s. his r sul
ts in po t ntia l
ls al
lr a nd o r a ccura t tr s.
3
ti
r
ssi c ti
ol
utio na r t chniqu s ha no tb n pr io usla ppl
id to fu
cl
a ssi ca tio n
pro bl s. h cl
o s str l
a t d o r studis us g n tic a l
go rith s to o pti is
9 0
P .L . R o s in a n d H .O . N y o n g e s a
fu
cl
a ssi ca tio n rulsa nd th ir
b rship functio ns. n disa d a nta g ith
this a ppro a chis tha titis in a ria bln c ssa r to pr -sp cif th structur o fth
ruls, hich o ft n r sul
ts in sub-o pti a lcl
a ssi ca tio n. In this pa p r,
ha
pro po s d a n
t chniqu in hich fu
cl
a ssi ca tio n ruls o fa rbitra r si
a nd structur ca n b g n ra t d using g n tic pro gra
ing. his is d sira bl
fo r co pl pro bl s, ith l
a rg nu b rs o finputfa tur s, fo r hich itis no t
fa sibl to fo r ul
a t th structur o fruls a nua l
l. urth r o r , ith such
l
a rg nu b rs o f fa tur s it is usua l
l th ca s tha t c rta in fa tur s a r no t
signi ca nt in cl
a ssi ca tio n o fdiff r nt o utput cl
a ss s. H nc , in this ca s ,
ca n sa tha tg n tic pro gra
ingis us d fo r unco nstra in d ruldisco r a nd
o pti isa tio n.
n tic pro gra
ingis a n o l
utio na r t chniqu in principlto H o l
l
a nd’
s
g n tic a l
go rith s. h a in diff r nc s a r ,( ) th structur o fa g n tic pro gra is a tr , (2) th no d s o fth tr s a r functio ns (o r t r ina l
s), hich
na bls th tr s to int rpr t d a s pro gra s a nd (3) th si o f a ch tr in a
po pul
a tio n is a ria bl, unl
i
o stg n tic a l
go rith s h r a l
lindiidua l
s ar
th sa
si . th r is , sta nda rd o p ra to rs a ppl
ica bl to g n tic a l
go rith s
a r us d in g n tic pro gra
ing.
In this stud , th no r a l
is d input spa c
a s pa rtio n d into thr fu
b rship functio ns,
r a nd s
[5]. t , Z , b th fu
b rship functio ns.
a ssu d si pl rul co nstructs co pris d o f t o
inputs a nd o n o utput. h no n-t r ina lno d s o fth
tr s r pr s ntth s
si plruls, hich a r co bin d to fo r th co pl cl
a ssi ca tio n rul. a ch
si pl rul is a l
ua t d b
a tching its inputs a ga instth fu
a nt c d nts
a nd th o utputis o bta in d usinga n
o p ra to r,na l I . hus,a no d
pr ss d a s Z ( , ) is int rpr t d a s:
I (
)
H
I (
)
hr
is th d gr o fb l
o ngingo f to th fu
b rship functio n Z .
his t p o frulco nstructis pr fra blto dir ctco bina tio n o fth inputpa ra t rs b ca us ita ssists in int rpr ta bil
it o fth cl
a ssi ca tio n s st . h r
a r nin diff r ntfu
ruls hich ca n b fo r d fro th co bina tio n o fth
thr
b rship functio ns.
h stud us d
so urc co d (l
il
gp) d l
o p d a t ichiga n ta t nirsit [6 ]. his a s us d to
o l a s to fco pl fu
cl
a ssi ca tio n ruls,
o n fo r a ch o utputcl
a ss o fda ta . h functio n s tis co pris d o fth si pl
ruls d scrib d a bo , hilth t r ina ls tco pris d ra ndo co nsta nts a nd
co
o n a rith tic o p ra to rs. h tn ss o fth tr s o n th ir a l
ua t d o utputs r d t r in d a ga instta rg ts o f . fo r th co rr ctda ta cl
a ss a nd .
o th r is . s a n a pl, th fo l
l
o ingis a s a l
lpo rtio n o fa co pl rul:
(ZP (NP a b) (ZN d c) )
Itca n b int rpr t d a s:
I d is Z
I a is
c is
b is
H
H
t p =
t p2 =
I (
I (
)
)
C o m b in in g E v o lu tio n a ry , C o n n e c tio n is t, a n d F u z z y C la s s ific a tio n A lg o rith m s
9 1
I t p is Z
t p2 is
H
o ut= I (
)
t p
t p2
ti
c
s
ca us n ura ln t o r s a r tra in d o n l
i it d sa pl s ts o f r pr s nta ti
da ta th r a r a la s signi ca nt rro rs in th g n ra l
isa tio n o fco pl functio ns. n a to o rco
this pro bl is to tra in ul
tipln ura ln t o r s o n
ind p nd ntda ta s ts a nd th n us o tingsch s to d t r in a n o ra l
lcl
a ssi ca tio n [7 ]. h t o po pul
a r sch s a r co
o nl no n a s ns bla nd
o dul
a r n t o r s. In ns bls o r co
itt s r dunda ntn t o r s a r tra in d
to ca rr o utth sa
ta s , ith o tingsch s b inga ppl
id to d t r in a n
o ra l
lcl
a ssi ca tio n. n th o th r ha nd, itis po intlss to tra in id ntica ln ura l
n t o r sa nd co nsid ra tio n is thus o ft n gi n to usingdiff r ntto po l
o gis,da ta
s ts o r tra ining a l
go rith s. In o dul
a r n t o r s th cl
a ssi ca tio n pro bl is
d co po s d into subta s s. o r a pl,n ura ln t o r s a r tra in d to r spo nd
to o n cl
a ss o r a gro up o fcl
a ss s.
h o utputo fa cl
a ssi ca tio n n t o r ca n b us d to indica t th d gr to
hichth inputfa tur s a r a tch d to th diff r ntcl
a ss s. h r fo r a si pl
a ppro a chto n t o r co bina tio n isto su th a ctia tio n l l
so fco rr spo nding
o utputno d s. r n
ntto this sch
is to sca lth o utputl l
s ithin a
n t o r suchtha tth su to o n . hisa l
l
o sth co ntributio nsa cro ssn t o r s
to b
o r co pa ra bl.
n al
t rna ti a ppro a chis ba s d o n th co nfusio n a tri,a ta blco nta ining
ntris , hich indica t th fr qu nc tha t da ta sa pls fro cl
a ss
r
l
a b ld a s cl
a ss . uch a ta bl is us fulfo r a na lsing th p rfo r a nc o f a
cl
a ssi r. ur a ppro a ch is ba s d o n th cl
a ssi ca tio n a ccura cis fo r a ch cl
a ss
gi n b
h co ntributio ns o f a ch cl
a ssi r a r
ight d b th s
p ct d a ccura cis.
n a plo fa co nfusio n a tri is sho n in a bl . Itca n b s n tha tth
cl
a ssi r is ca pa blo fco nsist ntl co rr ctl cl
a ssif ing a l
linsta nc s o fcl
a ss ,
buto nl 7 5% o fcl
a ss2.
p ri
ts
h p ri nts co pa r d th p rfo r a nc o ffo ur cl
a ssi rs o n th sha p cl
a ssi ca tio n pro bl . h tra ining a nd t sting da ta bo th co nsisto f 3 sa pls
a ch co nta ining s nt n co ntinuo us a ttribut s o f sha p . o l
l
o ing sta nda rd
pra ctic th da ta a s no r a l
is d prio r to pr s nta tio n to th n ura la nd fu
cl
a ssi rs, h r a s this a s no tn c ssa r fo r th d cisio n tr s.
h rsts to f p ri nts co nc ntra t d o n a l
ua ting a chindiidua lcl
a ssi r. a plso fth iro utputs a r sho n in igur s 2 to 4. h irp rfo r a nc s
9 2
P .L . R o s in a n d H .O . N y o n g e s a
2 3 4 56 7
2
3
4
5
6
7
%
ccura c
.
3
2 5
3 2
23 3
3
2
7 5.
3.3
7 .6
.
32
25.
6
4.2
2 33
7 .7
3 7 6 .5
. nfusin ma tri
a r sho n in a bl2. h bo tto thr ntris co rr spo nd to a l
t rna ti
a ns
to d riinga cl
a ssi ca tio n ba s d o n n ura lo utputno d a ctia tio n l l
s. h s
a r sta nda rd inn r-ta -a l
l(
),a nd usingth t o d cisio n tr s (
a nd
4.5). Itis o fint r stto no t tha tth si pl inn r-ta -a l
lp rfo r s b tt r
tha n th o r co pl d cisio n tr s in co bining th o utputs o fth n ura l
n t o r s.
h s co nd s to f p ri nts co pa r d a rio us o ting sch s a ppl
id to
n ura ln t o r ns bls. h n ura ln t o r s r tra in d o n th
subgro ups o fsha p pro p rtis,a nd a n a dditio na ln ura ln t o r a s tra in d o n th
co bin d s nt n pro p rtis. h o tingsch s ith r s lcto nl th a iu a ctia tio n l l ithin a n t o r ( inn r-ta -a l
l
),o r a l
t rna ti la l
lth
a ctia tio ns l l
s a r us d. h l l
s a
rstb sca ld, a nd
ight d b th
co nfusio n a tri,a nd a r th n su
d o r th n t o r s. h r sul
tin a bl3
sho s tha tb tt r p rfo r a nc is a chi d h n th n t o r ns blincl
ud s
th n t o r tra in d ith a l
lpro p rtis. urth r o r , th us o fth co nfusio n
a tri g n ra l
l i pro s p rfo r a nc . H o
r, th o dul
a r a ppro a ch s pro id l
ittli pro
nto r th ba sic singln t o r cl
a si ca tio n r sul
t.
h na ls t o f p ri nts in stiga t d h bridisa tio n o fthr cl
a ssi rs,
na l,n ura ln t o r ,fu a nd 4.5,usingth co nfusio n a tri
tho d.
his do s no t r l o n po t ntia l
l inco
nsura t l
i l
iho o d a l
u s tha t a
b pro duc d b diff r ntcl
a ssi rs.
s fro
a bl4tha tfurth r ga ins in
p rfo r a nc ha b n a chi d,indica tingtha tth n ura l
n t o r a nd d cisio n
tr pro id us fulco pl
i nta r info r a tio n.
6
c si s
his pa p r ha s pr s nt d a co pa riso n o fdiff r ntcl
a ssi ca tio n t chniqu s o n a
difficul
tsha p a na lsis pro bl . i il
a r to o th r r po rts in th l
it ra tur [ ,9 ]
itha s b n sho n th r is no signi ca ntdiff r nc s b t n th indiidua lt chniqu so n o urcl
a ssi ca tio n pro bl .H o
r, ha sho n tha ti pro
nts
ca n b a chi d thro ugh diff r ntco bina tio ns o fth s t chniqu s [ ].
C o m b in in g E v o lu tio n a ry , C o n n e c tio n is t, a n d F u z z y C la s s ific a tio n A lg o rith m s
1
T I
1
9 3
1
1
1
1
1
11
1
1
7
3
7
7
3
3
7
2
73
3
252 3
2
32 5
2
5
25
2
3 5 22 5 2
2
5
5
2 3 22
8 3
4
55
3 5 38 5 4 5 6
5
26 5
6 88
55
4
5
4 68 48 6
6 6 6
86 8 6 684 4444
9
8 8
4
8
8
68
6
7
2
2
9
9
5
4
9
6 9
99
6
6
99
6
8 9 89
9
89
E F
i. . n a mpl fth pa rtitining ff a tur spa c f r th
a n cl
a ssifica tin ta sk
th
d cisin tr using nl t pr p rtis. h h ri nta la is pr p rt is
l
l
ipticit, a nd th
rtica la is pr p rt is tria ngul
a rit.
H
%
4.5
u
+
+
+
4.5
53
54
44
57
52
5
. ccura c f r singlcl
a ssifi rs
9 4
P .L . R o s in a n d H .O . N y o n g e s a
Rule 4:
property 16 <= 0.225411
-> class 6
Rule 7:
property
property
property
property
property
property
property
property
property
-> class 8
1 > 63.7441
1 <= 65.4694
4 > 0.161205
4 <= 0.179735
11 <= 0.967232
13 > 0.710997
15 <= 0.061688
16 > 0.225411
16 <= 0.329981
Rule 8:
property 1 > 65.4694
property 4 <= 0.179735
property 15 <= 0.061688
-> class 6
i. .
t pica ls t fruls g n ra t d
4.5
(ZP (NP b
(PN (- (PN a 0.83860)
(NP (ZP (- (- (ZN d c)
(NZ -0.02504 d))
(PZ a f)) h)
(PZ e h))) g))
(ZP a
(ZP (- (ZN d c)
(NP d
(ZN d c)))
(ZP (NP f b)
(ZN d c)))))
i. . p rtin fa c mpl fu
rul in pr fi n ta tin. pp r ca s ltt rs
d n t fu
ruls, a nd l r ca s ltt rs d n t inputsha p pr p rt a l
u s.
C o m b in in g E v o lu tio n a ry , C o n n e c tio n is t, a n d F u z z y C la s s ific a tio n A lg o rith m s
%
u gr ups u gr ups +
r thing
3
46
4
46
55
5
44
55
4
56
42
4
43
5
5
56
5
5
H
um
um
um
um
, um fa ctia tins
, um fa ctia tins ith sca l
ing
, um f
, um fa ctia tins, ight d
, um fa ctia tins ith sca l
ing, ight d
fa ctia tins
fa ctia tins, ith sca l
ing
fa ctia tins, ight d
fa ctia tins, ith sca l
ing, ight d
. ccura c f r
9 5
tingsch m s
H
%
+
+
4.5
u
u
+
4.5
+
4.5+
.H
rid
63
5
57
6
l
a ssifi rs
n a ppro a ch to g n ra t fu
cl
a si ca tio n ruls usingg n tic pro gra ing,o n o th r ha nd,ha s no td o nstra t d sa tisfa cto r r sul
t. his a b du
to th l
i it d pa rtitio ning o fth fu
input spa c into o nl thr
b rship functio ns i pl nt d in th stud a nd furth r o r il
lin stiga t th
po t ntia lo fthis tho d.
fr c s
.
2.
3.
4.
5.
6.
7.
. nka , . H l
a a c, a nd .
l. I g r ssi g
lsis
hi
isi . ha pma n a nd H a l
l
,
3.
. . sin.
a suringsha p : l
l
ipticit,r cta ngul
a rit,a nd tria ngul
a rit. s b itt f r bli ti .
. . uinl
a n. 4.5 rgr s f r
hi
r i g.
rga n Ka ufma nn,
3.
.K. urth , . Ka sif, a nd . a l rg.
st m f r inductin f l
iqu d cisin
tr s.
r l f rtifi i lI t llig
s rh, 2: –33,
4.
. . ink ns a nd H . .
ng sa .
n tic a l
g rithms f r fu
c ntr l- pa rti.
I
r trl h r
l., 42: 6 – 7 6 ,
5.
. Z ngk r a nd . unch. htt / / isl
. s. s . /
/ s ft r/ lil-g.
6.
rl t rs. pring r- rl
a g,
. . . ha rk , dit r.
bi i g rtifi i l
.
9 6
P .L . R o s in a n d H .O . N y o n g e s a
. H . h n, . untin, . h , . utja hj, . mm r,a nd .
l. p rtpr dictin,s m l
ic la rning,a nd n ura ln t rks: n p rim nt n gr h und ra cing.
I
rt, :2 –27 ,
4.
. . ul
hl
l
a nd, . . H i rt, . . H a dda d,a nd . a rsl . c mpa ris n fcl
a ssifica tin in a rtificia lint l
l
ig nc ,inductin rsus n ura ln t rks. h
tri s
I t llig t b rt r
st s, 3 : 7 – 2 ,
5.
. .Xu, . Kr a k, a nd . . u n.
th ds f c m ining mul
tipl cl
a ssifi rs
a nd th ir a ppl
ica tin t ha nd riting r c gnitin. I
r s. st s
b r ti s, 22:4 –435,
2.
Experimental Determination of Drosophila
Embryonic Coordinates by Genetic Algorithms
the Simplex Method and Their Hybrid
Alexander V Spirov
Dmitry L Timakin
John Reinitz
and David Kosman
The Sechenov Institute of Evolutionary Physiology and Biochemistryflflfl Thorez
AveflflStflPetersburgflflflflflflflflRussiafl
Deptflof Automation and Control SystemsflPolytechnic Universityflflfl Polytechnic
StflStflPetersburgflflflflflflflflRussia
Deptflof Biochemistry and Molecular BiologyflBox flflflfl MtflSinai Medical Schoolfl
One Gustave LflLevy PlaceflNew YorkflNY flflflflfl USA
Abstract Modern largeflscale flfunctional genomicsfl projects are inconfl
ceivable without the automated processing and computerflaided analysis
of imagesfl The project we are engaged in is aimed at the construction
of heuristic models of segment determination in the fruit fly
s
s flThe current emphasis in our work is the automated transfl
formation of gene expression data in confocally scanned images into an
electronic database of expressionfl We have developed and tested profl
grams which use genetic algorithms for the elastic deformation of such
imagesfl In additionfl genetic algorithms and the simplex methodfl both
separately and in concertfl were used for experimental determination of
s
embryonic curvilinear coordinatesflComparative tests demonfl
strate that the hybrid approach performs bestflThe intrinsic curvilinear
coordinates of the embryo found by our optimization procedures appear
to be well approximated by lines of isoconcentration of a known morfl
phogenflBicoidfl
Introduction
Computer Aided Analysis of Biological Images
The ongoing revolution in molecular genetics has progressed from the large scale
automated characterization of genomic sequence to the characterization of the
biological function of the genome These investigations mark the beginning of
the era of functional genomics
A key feature of genomic scale approaches
is the automated treatment of large amounts of data Both current and future
work in the eld is impossible without the automated processing and computer
aided analysis of images in connection with updating interactive electronic image
databases
A key aspect of such processing involves the segmentation of individual im
ages and the registration of serial images Many problems involving the recog
nition classi cation segmentation and registration of images can be formulated
as optimization problems These optimization problems are typically di cult
S . C a g n o n i e t a l. ( E d s .) : E v o W o r k s h o p s 2 0 0 0 , L N C S 1 8 0 3 , p p . 9 7 − 1 0 6 , 2 0 0 0 .
S p rin g e r-V e rla g B e rlin H e id e lb e rg 2 0 0 0
9 8
A .V . S p ir o v e t a l.
involving multiple minima and a complex search space topologyflContemporary
approaches based on evolutionary computations are a promising avenue for the
solution of such problems flEvoIASPflflflfl
Here we describe a new method for the determination of intrinsic biological
coordinates in embryos of the fruit fly Drosophila melanogaster by means of
genetic algorithms flGAsflflGAsflthe simplex methodfland a hybrid of both were
applied to the problemfland we flnd that hybrid methods perform the bestflOur
results indicate that these coordinates may be determined by a morphogenetic
gradient of the protein Bicoidfla result of some biological interestfl
Stripe Straightening Search of Intrinsic Coordinates of Early
Embryo
Early in the development the fruit fly embryo is shaped roughly like a hollow
prolate ellipsoidflcomposed of a shell of nuclei which are not separated by cell
membranesfl Deviations from the elipsoidal shape reveal the future polarity of
the animalfls bodyflThe more pointed end on the long axis makes anterior flheadfl
structuresfland the rounder end posterior fltailfl structuresflFrom a lateral flsidefl
perspectiveflone long edge of the embryo is flat and will will make dorsal flflbackflfl
structuresflwhile the other long edge is rounded and makes ventral flflundersideflfl
structuresflIn this paper we follow the standard biological convention and show
embryos with anterior to the left and flif a lateral viewfl dorsal up flFigflflflfl
Figfl fl shows that so called pairflrule stripes flearly markers of the future
segmental pattern flflflfl are not parallel and straightfl but have a crescentfllike
formflThe curvature of the stripes is highest at the terminifland minimal at the
central partflEach stripe specifles an anteriorflposterior flAflPfllocationfland these
stripes can be regarded as contours in an intrinsic coordinate system that is being
created by the embryo itselffl Another set of embryonic determinants exists for
the dorsoflventral flDflVfl axisflIf the image is smoothly transformed such that the
curvilinear coordinates are plotted orthogonallyflthe stripes appear straightflso
the determination of these coordinates can be viewed as a flstripe straighteningfl
procedureflOur task is to understand and characterize this curvilinear coordinate
system as it relates to the AflP axisfl
Two coupled objectives of this study arefl
flfl To characterize the intrinsic embryonic curvilinear coordinatesfl
flfl To use carefully characterized and tested computational procedures for the
purpose of automatically processing large numbers of imagesfl
Methods and Approaches
The work reported here is part of a large scale project to construct a model of
segment determination in the fruit fly D melanogaster based on coarseflgrained
chemical kinetic equations flflflfl The acquisition and mapping of gene expression
data at a heretofore unprecedented level of precision is an integral part of this
E x p e rim e n ta l D e te rm in a tio n o f D ro s o p h ila E m b ry o n ic C o o rd in a te s
9 9
Fig
Image of early flblastoderm stagefl fly embryo with crescentfllike stripesfl in
Cartesian physical coordinatesfl This is a confocally scanned image of an embryo
stained by indirect fluorescence flimmunostaining with polyclonal antisera against the
segmentation proteinflflEach small dot is an individual nucleusfl
projectflThe current emphasis in our work is on the automated transformation
of gene expression data in confocally scanned images into an electronic database
of expressionfl
Images of Drosophila Genes Expression
Transformations of embryonic coordinates begin with data expressed in terms
of the average fluorescence level flproportional to gene expression levelfl of each
nucleusflwhere segmentation proteins exert their biological functionflThis data
was obtained as followsfl
Antibodies for flfl protein products of segmentation genes were raised and over
flflfl images were prepared and scanned flflflfl These images were computationally
treated by means of the Khoros package flflflfl Embryos were rotated and cropped
automatically such that the physical long axis of the embryo was parallel with
the x axis and the short physical axis with the y axisfl
Nextfl the images were segmented flflfl Kosman et alfl in preparationflfl About
flflflfl segmented and identifled nuclei are obtained from each imageflEach nucleus
is labeled numericallyfl and the x and y coordinates of its centroid are foundfl
together with the average fluorescence level over that nucleusfl The segmented
1 0 0
A .V . S p ir o v e t a l.
data takes the form of tables in ASCII text formatflThe result is the conversion of
an image to a set of numerical data which is then suitable for further processingfl
Stripe straightening algorithm
In Figflfl the crescentfllike pairflrule stripes of an embryo in near saggital projecfl
tion are shownflWe assume that the center of a pairflrule stripe follows a curve
of constant AflP positionfl The origin of the image coordinate system is at the
top leftflwith image coordinates for width w increasing to the right and height
h increasing downfl
Our goal here is to flnd the true AflP and DflV coordinates on the imageflWe
approximate the true coordinate system by a Taylor series as followsflWe denote
the true AflP coordinate by x
fl and the true DflV coordinate by yfl
fl We pick the
origin of flfl
x yfl
fl and the origin of new image coordinates flx yfl so that they are
the same and as close as possiblefl
We note that there is an AflP position at which a stripe is exactly vertical
on its whole lengthflThe center of that stripe deflnes x
fl fl flflwhich is the yfl
fl axisfl
Each pairflrule stripe other than the one at xfl fl fl is curvedfland we imagine the
x
fl axis to intersect each of the stripes at the point where it is exactly verticalfl
Now we pick new image coordinates x and y such that they have the same origin
and orientation as the flfl
x yfl
fl coordinatesflthat is
x fl w − w y fl −h − h
flflfl
We now turn our attention to xflFor
fl
nowflwe can assume that yfl fl yflEven
if we donflt do thatfl two important things will be true about the relationship
between flx yfl and flfl
x yflflflflfl
fl
The y and yfl axes are coextensivefland flflfl The loci
yfl fl const are orthogonal to the y and yfl axes as they cross y fl yfl fl flflBoth of
these important points follow from the existence of the vertical stripeflWe would
like to write x
fl in terms of x and yflso that
x
fl fl f flx yfl
flflfl
We expand in a Taylor series to third order around the originflThat gives
x
fl fl
f flfl flfl
fl
x
fl
fl flfl flfl
x fl flfl flfl
xy fl flfl flfl
fl flfl flfl
x fl flfl flfl
x y fl flfl flfl
fl flfl flfl
y
y
x
xy
flflfl
Now consider the terms and what they meanflf flfl flfl fl fl by deflnitionflWe
picked flfl
x yfl
fl such that at the origin
fl fl and
fl flflFor pure y terms we can
say more than thatflThe fact that the y and the yfl axes are coextensive means
fl
fl fl as wellfl Thus far we have shown
that f flfl yfl fl fl
yfl and so
that flve of the ten terms of the Taylor expansion vanishfl
E x p e rim e n ta l D e te rm in a tio n o f D ro s o p h ila E m b ry o n ic C o o rd in a te s
The unit vector e in the x
fl direction is proportional to
flso
1 0 1
measures
the change in length of e as we move along the x axisflThis means that
fl flfl
flThis term can be thought of as the rate of change in size of
Now consider
the unit vector e fl
along the yflaxisflAlong the yflaxis where x fl x
fl fl flfl
fl fl
yflso that derivatives of this quantity with respect to y vanishfland
hence this term of the series vanishesflThis has eliminated all but three terms
from the seriesflso now we write the flrst order model of image transformation
as
x
fl fl x fl Axy fl Bx y fl Cx
flflfl
All of these terms have a clear interpretationfl The xy term is the main onefl
it gives quadratic DflV curvature that increases with distance from the xflaxisfl
The x y term gives residual DflV asymmetry and the x term gives residual AflP
asymmetryfl Lastlyfl if one expresses the above equation in terms of w and hfl
expansion will bring back lower order terms in h and w when expanding
x
fl fl w − w fl Aflw − w flfl−h − h fl fl Bflw − w fl fl−h − h flfl Cflw − w fl flflfl
in terms of w and hfl
We tested this flflst order model and found that in more then half of cases
it is insufl cient for straightening stripesfl We expanded the model empiricallyfl
with the result that an empirical extension of the flflst order model is given by
x
fl fl Aflw − w flfl−h − h fl fl Bflw − w fl fl−h − h flfl
Cflw − w fl fl−h − h fl fl Dflw − w flfl−h − h fl
flflfl
We can treat of these additional fourth order members as followsflCx y is a
correction term for parabolic splayflwhile Dxy serves to correct DflV asymmetryfl
In generalflthe situation is typical of a polynomial approximation problemfl there
is one polynomial that is best but there are a number of distinct ones that can
approximate it very wellfl
Preliminary calculations have shown that the best outcome is achieved with
an independent deformation of the anterior and posterior half of an embryoflIn
summary it requires the determination of fl parameters of a deformation plus an
evaluation of values w flh flh fl
Genetic Algorithms Technique Simplex Method and Their
Hybrid
GA Search The optimization problem of flnding the coefl cient values for
proper elastic transformations was initially implemented with GAsfl We have
reduced the problem to the determination of factors AflBflC and D of equation
flfl
We use the following cost functionflEach embryofls image under consideration
was subdivided into a series of longitudinal stripsflThen each strip is subdivided
into boxes and the mean value of the product flEVEN SKIPPED proteinfl is
1 0 2
A .V . S p ir o v e t a l.
calculated for each boxflEach row of means gives the local proflle of even skipped
gene expression along each stripflThe cost function is computed by comparing
each proflle and summing the squares of diflerences between the stripsflThe task
of the GA is to minimize this cost functionfl
Following the classical GA algorithmflthe program generates a population of
floatingflpoint chromosomesfl Initial chromosomes are randomly generatedfl Affl
ter that the program evaluates every chromosome as described abovefl thenfl
according to the truncation strategyflthe average score is calculatedflCopies of
chromosomes with above average scores replace all chromosomes with a score
less than averagefl
On the next step a predetermined proportion of the chromosome population
undergoes mutationflso that one of the coefl cients gets a small incrementflThis
cycle is repeatedflall chromosomes are consecutively evaluatedflthe average score
is calculated and the winnersfloflspring substitutes for the losers in the process
of reproductionfluntil an accepted level of stripe straightening is achievedfl
Simplex Search We also solved the optimization problem by the downhill
simplex method in multidimensions of Nelder and Mead flflflflThe method requires
only function evaluationflnot derivativesflThis is an important speed advantage
over gradient methodsfl since calculation of the gradients requires many more
evaluations of the cost functionfl
A simplex is the geometric flgure in N dimensions of N fl fl vertices and all
their interconnecting line segmentsflThe NelderflMead method starts with such
a set of N fl fl points deflning an initial simplexflThe downhill simplex method
operates by moving the point of the simplex where the function is largest through
the opposite face of the simplex to a lower pointfland so on until it reaches the
vicinity of an extremumfl
Hybrid Procedure Initial experience indicated that that the simplex method
is fast but does not give high quality answersflwhile GAs give excellent answers
but are slowflWe noted that both multiple simplex runs and GA search perform
numerous evaluations for many random points in search spaceflIf we use small
increments as mutations we will perform practically the same search by using
GAs or the simplex methodfl If sofl we could use a set of chromosomes from
the GA technique as a starting simplex for NelderflMead optimizationfl In the
hybrid algorithmflwe implement a simple evolution strategy with floatingflpoint
chromosomes with small mutational incrementsflSelection and reproduction are
performed as described aboveflIn additionflfrom the very beginning the program
links pointers to mutant oflspring so as to achieve complete lists of N fl fl pointers
on N fl fl relativesfl
These flclans of mutantsfl are ready for simplex procedurefl Following the
completion of the flrst list of N fl fl pointers the program starts to perform not
only mutationfl selectionflreproduction proceduresflbut also the simplex procefl
dure for the lowest scoring members of flcompletefl clansflThe more clans achieve
completion the more species undergo simplex procedureflIn summaryflGAs must
E x p e rim e n ta l D e te rm in a tio n o f D ro s o p h ila E m b ry o n ic C o o rd in a te s
1 0 3
provide search of global optima together with local onesflwhile simplex provides
fast downhill movingfl
Results and Discussion
Search Space Features for Stripe Straightening Problem
The above described task of image elastic deformations turned out to be a diffl
flcult numerical problemflThis is caused flrst of all by the unusual geometry of
search spacefl Figfl fl gives a picture of its features through crossections of the
search space for one typical embryo under considerationfl
Fig
Search space features for one of crossections flA fl Dfl for typical imagefl This
is surface plot where vertical flZfl axis is evaluation oneflwhile X and Y are A and D
coefl cients of expression flfl
As we can see this cross section includes two groovesflone of which is deeper
than anotherflThe sharp rectangular walls in Figflfl are caused by penalty confl
ditionsflThe omission of the penalties gives a smoother surface with one groovefl
which corresponds to the deeper groove flnot shownflflIn turnflpenalties are abfl
solutely needed to avoid highly nonlinear folding of an image instead of smooth
deformationsfl
The bottom of both grooves have several local minimaflAs a result the simplex
search gives in serial searches tens of such local extremafl GA search is more
1 0 4
A .V . S p ir o v e t a l.
eflective and it flnds the best local extremum on the bottom of the deepest
grooveflHowever to jump from the shallow to the deeper groove is still a difl cult
task for GA search as wellflTo overcome this we use large population sizes or a
series of runs to achieve the best solutionfl
The Results of Stripe Straightening
After completion of the stripe straightening procedure with flfl coefl cients flw
and two sets of Afl Bfl Cfl D and h fl h fl for about two hundred images from
the stages when all seven crescent stripes are visible we could compare found
coefl cient setsfl These coefl cient sets show considerable diversityfl so that we
fail to elucidate a general formula of appropriate elastic deformations to achieve
satisfactory stripe straighteningflHoweverflthe resulting transformation of coorfl
dinates are very similar for most of the imagesfl Typical example is shown in
Figflflfl
Fig
Typical example of curvilinear coordinates found by our optimization procefl
duresfl
On the contraryflcomparison of coordinate curves for anterior and posterior
halves of embryos reveal small but quite evident diflerences flCfflFigflfl and Figflflflfl
A biological subject of interest is the source of the pairflrule stripesflcurvaturefl
It is known that in Drosophila segmentation the maternally expressed protein
BICOID forms an anteriorflposterior morphogenetic gradient in the egg which
controls all following segmentation events flflflfl It is interesting that contour lines
of a flDflconcentration map of the BICOID gradient closely coincide with curvifl
linear coordinates determined by our methodflThe full biological implications of
this observation will be reported elsewherefl
E x p e rim e n ta l D e te rm in a tio n o f D ro s o p h ila E m b ry o n ic C o o rd in a te s
1 0 5
E ectivity and Cost of GAs Simplex and Hybrid Techniques
In a table fl the results of the comparative tests on elastic transformation of
a typical image by means of the simplexflmethodflGAsfland GAs with simplex
flthe hybrid techniquefl are presentedflTo flnd parameters of optimization for a
simplex and GAs giving the most eflective optimizationfl careful tuning of the
limits in variation of a mutational increment and the range of variability of
the initial population was carried outfl The hybrid method was tested at the
same values of parametersflas GA techniqueflThe result of testing was compared
according to the time required for calculation and according to the standard
deviation of the resultfl On each tested procedure flflfl independent runs were
carried outfl Inspection of the table reveals that the simplexflmethod is fastestfl
but also the least preciseflThe GA technique is the most preciseflbut also requires
flfl times as much computingflOur results indicate that the hybrid technique is
approximately twice faster than GAs at the same accuracyfl
Table
Comparison of eflectiveness of three approaches on calculation time flin a
summarized amount of evaluationsfl and on a divergence flin a standard deviation valfl
uesfl
Method
Simplex Method
GA Technique
Hybrid Technique
Time in evaluations Standard Deviation
flflflflfl
flflflflflflflfl
flflflflflfl
flflflflflflfl
flflflflflfl
flflflflflflfl
The problems we encountered with our optimization task arefl
flfl An abundance of local minima very close to the global minimumflThe simfl
plex can get stuck in very smallflscale holesfl even with starting conditions
exceptionally close to a known solutionflWe need to allow the optimizer to
move from a position that is very close to the global minimum towards the
global minimumfl It seems that this last stage is the difl cult part for the
NealderflMead simplex methodfl It is possible to produce good scores with
a number of distinct end points in polynomial parameter spaceflsuggesting
that the problem is probably overflspecifledfl
flfl This is the polynomial approximation problemflThere will be one polynomial
that is the best for the stripe straightening but there are a number of distinct
ones that can approximate it very wellfl
As to the flrst itemfl the successful approach to a solution of this problem
is to employ hybrid techniquesfl Genetic algorithms alone are usually slow in
optimization problemsfl since they are too coarseflgrained to obtain a solution
quicklyfl On the other handfl downhill algorithms are usually fast flin terms of
processor cyclesflflif they are close to the solutionflbut tend to get stuck in local
minimaflCombining both kind of algorithms manages to avoid local minimafland
flnds solutions accuratelyfl
1 0 6
A .V . S p ir o v e t a l.
Conclusions
In the task of optimization of parameters of elastic transformation a simplexfl
method is fastestflbut also the least preciseflgiving the greatest divergenceflThe
GA technique is the most preciseflbut also requires at least flfl times more timefl
The hybrid technique is twice faster than GAs at the same accuracyfl
The intrinsic curvilinear coordinates of an embryo found by our procedures of
optimization appears to be approximated by contour lines of a map of a gradient
of the morphogen bicoidfl It is in the good agreement with known ideas about
a governing role of this gradient in consequent processes of segmentation of an
early fly embryofl
Acknowledgements
This work is supported by INTASflgrant No flflflflflflflflflRussian Foundation for
Basic Researchesfl grant No flflflflflflflflflflflfl USA National Institutes of Healthfl
grant ROflflRRflflflflflfland CRDFflgrant No RBflflflflflflAflSflwishes to thank Timfl
othy Bowler for stimulating discussions and KingflWai Chu for help with profl
grammingfl
References
flfl AkamflMflfl The molecular basis for metameric pattern in the Drosophila embryofl
Development
flflflflflfl flflflflfl
flfl KosmanflDfland ReinitzflJflflRapid preparation of a panel of polyclonal antibodies to
Drosophila segmentation proteinsflDevelopmentflGenesfland Evolution
flflflflflfl
flflflflflflflfl
flfl KosmanflDflflReinitzflJfland Sharp DflHflflflflflflflAutomated assay of gene expression at
cellular resolutionfl In AltmanflRflflDunkerflKflflHunterflLfland KleinflTfleditorsflProfl
ceedings of the flflflfl Paciflc Symposium on Biocomputingfl pages flflflflfl Singaporefl
World Scientiflc Pressfl
flfl LanderflEflSflflThe new genomicsflGlobal view of biologyflScience
flflflflflfl flflflfl
flfl PressflWflHflfl FlanneryflBflPflfl TeukolskyflSflAfl and VetterlingflWflTflfl flflflflfl Numerical
Recipes in CflThe Art of Scientiflc ComputingflCambridgefl Cambridge University
Pressfl
flfl Rasure Jfland Young MflflAn open environment for image processing software develfl
opmentfl Infl flflflfl SPIEflISET Symposium on Electronic Imagingfl Vflflflflfl of SPIE
ProcessingsflSPIEflflflflflfl
flfl ReinitzflJflfl KosmanflDflfl VanarioflAlonsoflCflEfl SharpflDflflStripe forming architecture
of the gap gene systemflDevelopmental Genetics
flflflflflfl flflflflflfl
flfl SanchezflCflfl LachaizeflCflfl JanodyflFflflet alflflGrasping at molecular interactions and
genetic networks in Drosophila melanogaster using FlyNetsflan Internet databasefl
Nucleic Acids Research
flflflflflfl flflflflflfl
A Genetic Algorithm with Local Search
for Solving Job Problems
½ ¾
½
½
¾
!"# $
% & &
' ( )
& *
+"), - ( ( .( &*
+ .
Æ
!"# $%℄ $'℄
(
(
)
) * + ,
+ * , +
Æ -"
!" ..# $/℄ $0℄
1 !"
2 3# $4℄
* 3
5
#
3 %# 2 5
3 * !"
6 '# 3
5
* Æ 7
S. Cagnoni et al. (Eds.): EvoWorkshops 2000, LNCS 1803, pp. 107-116, 2000.
Springer-Verlag Berlin Heidelberg 2000
108
L.W. Cai, Q.H. Wu, and Z.Z. Yong
Æ
℄
℄
!℄ "℄ #
$% $ &℄ '
(
$
$
' $
'
' Æ )
*
+ ' '
, ' '
' -'
(
. '
+
'
, (
/
$ "
* '
$ ' '
& 0
'
* $
$
' '
1 '
)
2
3
2 '
)
3
4
2
( '
½
)
'
℄
-'
2
(
' /
(
' 5
2
-' 5
'
$
-'
, $ 5' (
$
,
-'
24
6
½ ¾
℄
)
%
'
. ( /
%
-'
,
$ $ 4
7 8 9 $
7"8 #
6
6 ½
7&8 7 8 7"8 '
7 8 ½ 6
´ ·½µ 7 6 8
6
2
6
6
'
:
A Genetic Algorithm with Local Search for Solving Job Shop Problems
½¾
¾¾
½
¾
½½
¾½
109
½
¾
!
" # ½ ¾ ℄
!
%&' ( ) % ' ½ ½ " ½
% ' * " ´ ·½µ % " & &' " + * ½ " ½ ´ ·½µ % "
& &' ½ "
%,' %&' % ' " + "
-
½½ ½¾ ½
¾½ ¾¾ ¾
½ ¾
. .
/
110
L.W. Cai, Q.H. Wu, and Z.Z. Yong
Æ
!
" # $ %
& '
' ! "
!
$ (
%
)
% * (
) "
+
(
%
#
'
'
-
,
'
,
"
.
,
.
.
,
.
'
#
' # , , - . - ' . - ' ' , , ' . - .
, # ' ' . - ' - , . . , ' , , - . -
+ !
'
'
-
' # - - ' - ' ' ' -
/
'
, , . . , , . .
,
, # ' ' - ' - ' - -
/
,
. , . . , , , .
" '
- # - - ' - ' ' ' -
* ,
0
1
. # ' ' - ' - ' - -
A Genetic Algorithm with Local Search for Solving Job Shop Problems
111
!
! "
#
$ #
%" &
" # '
#
( "
) %"
*
+
, - ./ (
. " 0/
1
$ "
- # ' "
# '
" " Æ %"
" " # '
!
"
112
L.W. Cai, Q.H. Wu, and Z.Z. Yong
!
" #
$
%
#
$
#
& &
'
(
'
#
)
*+ * &
,
-
/+
)
.
/ &
. 0
,
-
)
.
1 & . 2 .
1
-
+
$
!
-
1
9
'
3( ! 3(4&
5 6
)
$7
/"
/"
8
8
1
%
%
A Genetic Algorithm with Local Search for Solving Job Shop Problems
Ü
Ü
Ü
Ü
Ü
¿
½
¾
¼
1200
1170
population size=168
pc=0.7, pm=0.2
1140
1110
Makespan
GA
1080
1050
GA/OSLS
1020
990
GA/MSLS
960
930
900
0
20
40
60
80 100 120 140 160 180 200 220 240 260 280 300 320 340 360 380 400
Generation
!"
$ !"
#
% !&"
'( !&"
)
!$&"
#
##
*
113
114
L.W. Cai, Q.H. Wu, and Z.Z. Yong
℄
ß !" #
ß $$ $ !" $ #
ß %$& %$ !" & #
' ( )
" $$" %$&# &&" *+$ & (, - (
" ' '
)' ' , - ' ) ./0"
1 ) ' 2" '
$,3 $,% (, 4 ) - '
./0 ' ) 56
6 7 7 ' ./0, '
( 8 , '
) ./0 . 9 :0: , :'
' ) ' ) :0: )
' , ' )
%" + (, 4 " ; 9
; 9 ) ! ,
0 ./0#
( '
! ' , 0
, ' ) ℄
( ./0, ' '
' Æ , '
; 9 ')' , 4 '
' ' )' ,
( ' ) ,
)
) ' , ' ( ' )
,
, 6 < . ' . / = ( )
6 ) ' = ( ) 6 '
,
! "#$
% & ' ( ) * ! +
! ,- . "/%
A Genetic Algorithm with Local Search for Solving Job Shop Problems
115
!" #"$ " %&" !
% ' ( ) * + , & &
! #-$ .%&/.
" 0 ) * 12
3 &
* &45 ( ' 5 !
. 6 7 * + 8 ( &
+ 7 9 : ' + * *; 0:)1
. &/. %
/ )< & = +
( + 7 9 : ' + *&
*; 0:)1 >&/
! ' ? 4 @A 2 5 < &, , &
5 ( :))) : ' + ), 5 '&
/ : B* %% &%%.
4 @A ' ? 2 +
( :))) : ' + ), 5 '
/ : B* //&!>
> ;
? 7 ' , + &
( + 7 9 : ' + *
*; 0:)1 %/%&%/
9 0 ' * + '&
C 1 " #$ "&-%
- * &
: + 5 / #$ -"& -
? 7 ; * 6 & , +
9 : ' + *
) 5D : ,
* D *3):*Æ
" %.&"
% 9 7 3 : ( &A ) 6
'E ;6 5 .
* +
-
%
-"
>
%> ">
>
!
!
-
>
.
%
. -
.
-- -/
"
"> "
-- -"
-.
> %>
!
%
">
7..
% " .
!
%
""
% %%
"> "%
-/
/
.
>
! %"
> !
-/
>
>
!
%% "
"% ""
-"
>
/ %
%
%" %
! -/
116
L.W. Cai, Q.H. Wu, and Z.Z. Yong
Distributed Learning Control of Trac Signals
Y. J. Cao, N. Ireson, L. Bull and R. Miles
Intelligent Computer Systems Centre
Faculty of Computer Studies and Mathematics
University of the West of England, Bristol, BS16 1QY, UK
Abstract. This paper presents a distributed learning control strategy
for trac signals. The strategy uses a fully distributed architecture in
which there is eectively only one low level of control. Such strategy
is aimed at incorporating computational intelligence techniques into the
control system to increase the response time of the controller. The idea is
implemented by employing learning classi er systems and TCP IP based
communication server, which supports the communication service in the
control system. Simulation results in a simpli ed trac network show
that the control strategy can determine useful control rules within the
dynamic trac environment, and thus improve the trac conditions.
1 Introduction
Trac control in large cities is a dicult and non-trivial optimization problem. Most of the existing automated urban trac control systems, such as
TRANSYT1, SCATS2, LVA3 and SCOOT4, have a centralized structure,
i.e. information gathering and processing, as well as control computations, are
carried out in a centralized manner, in which case eciency is decreased due to
the large volume and the heterogeneous character of information 5. To achieve
global optimality, hierarchical control algorithms are generally employed. However, these algorithms have a slow speed of reaction and it has been recognized
that incorporating some computational intelligence into lower levels can remove
some burdens of algorithm calculation and decision making from higher levels
6. Recently, there is a growing body of work concerned with the use of evolutionary computing techniques for the control of trac signals. Montana and
Czerwinski 7 proposed a mechanism to control the whole network of junctions
using genetic programming 8. They evolved mobile creatures" represented as
rooted trees which return true or false, based on whether or not the creature
wished to alter the trac signal it has just examined. Cao et al has developed an
intelligent local trac junction controller using learning classier systems and
fuzzy logic 9 and showed that the local controller can determine useful junction control rules within the dynamic environment. Mikami and Kakazu used
S . C a g n o n i e t a l. ( E d s .) : E v o W o r k s h o p s 2 0 0 0 , L N C S 1 8 0 3 , p p . 1 1 7 − 1 2 6 , 2 0 0 0 .
S p rin g e r-V e rla g B e rlin H e id e lb e rg 2 0 0 0
1 1 8
Y .J . C a o e t a l.
a combination of local learning by a stochastic reinforcement learning method
with a global search via a genetic algorithm. The reinforcement learning was
intended to optimize the trac ow around each crossroad, while the genetic
algorithm was intended to introduce a global optimization criterion to each of
the local learning processes 12 . Escazut and Fogarty proposed an approach to
generate a rule for each junction using classi er systems in biologically inspired
con gurations 13 .
This paper is devoted to developing a distributed learning control strategy
for trac signals. A fully distributed architecture has been developed in which
each subsystem is solely responsible for one aspect of the system and where a
coherent global control plan emerges from the interactions of the subsystems
no hierarchical structure is included. Such an approach is aimed at increasing
the speed of response of the local controller to changes in the environment. To
do this, we have developed an agent-alike controller, which is implemented by
employing learning classi er systems 14, 15 and TCPIP based communication
server, supporting the communication in the control system. Simulation results
in a simpli ed trac network show that the control strategy can determine useful
control rules within the dynamic trac environment, and therefore improve the
trac conditions.
2 An Agent-alike Controller
Optimization of a group of trac signals over an area is a large and multiagent type real-time planning problem without precise reference model given.
To do this planning, each signal should learn not only to acquire its control
plans individually through reinforcement learning but also to cooperate with
each other. This requires communication between the agents. If each signal
simultaneously communicates with each other and controls its phases according
to the change of the global trac ow, the total volume of the area will be
well optimised. However, to provide the ecient communication is a dicult
task, caused by the inecient accounting of interactions between subsystems in
decentralized case and the complex communication structure in the hierarchical
case.
In this work, we developed an agent-alike controller, consisting of a learning
classi er system and a communication server, as shown in Figure 1. Rule-based
controller as classi er systems lie midway between neural network and symbolic
processing systems that can combine the bene ts of both. The limitation of
specifying a single classi er system is that while it may work for a simple controller, the method does not scale up to complex control systems. Work done so
far has addressed this problem by dividing a complex system into its simplest,
physical sub-systems, specifying rule-based controller for each of these and thus
creating a multi-agent system.
For trac control problem, we associate an agent to each junction of the
trac network. So, the whole control strategy developed in this work contains a
number of distributed, communicating agents, where each agent has a classi er
D is trib u te d L e a rn in g C o n tro l o f T ra ffic S ig n a ls
1 1 9
A g e n t
C la s s ifie r S y s te m
C o m m u n ic a tio n
S e rv e r
Figure 1: Structure of the agent-alike controller
system providing the control strategy and a communication server which is
used to connect the agent to the user interface, the application and to other
agents.
The two elements of the agent, classi er system and communication server,
are separate since as messages are passed around the agent network, the communication server acts independently of the classi er system to route the message
to its neighbours. Another reason for keeping the communication server distinct
from the classi er system is that communication is likely to be implementation
speci c even in the test applications. Thus it is necessary when specifying the
communication server to consider the general requirements of setting up and
maintaining the communication in a distributed learning system rather than
those in a speci c software and hardware implementation.
2.1 Classier systems
A classi er system is a learning system in which a set population of conditionaction rules called classiers compete to control the system and gain credit
based on the system's receipt of reinforcement from the environment. A classier's cumulative credit, termed strength, determines its in uence in the control
competition and in an evolutionary process using a genetic algorithm in which
new, plausibly better, classi ers are generated from strong existing ones, and
weak classi ers are discarded.
A classi er c is a condition-action pair
c = condition : action
with the interpretation of the following decision rule: if a current observed state
matches the condition, then execute the action. The condition is a string of
characters from the ternary alphabet f 0 1 g, where acts as a wildcard
allowing generalization. The action is represented by a binary string and both
conditions and actions are initialized randomly. The real-valued strength of a
classi er is estimated in terms of rewards obtained according to a payo function.
Action selection is implemented by a competition mechanism, where a strength
1 2 0
Y .J . C a o e t a l.
C o m m u n ic a tio n
C o n fig u ra tio n
P a ra m e te rs
C o m m u n ic a tio n S e rv e r
2 . C re a te S e rv e r
3 . C o n n e c t
(C o m m u n ic a tio n S e rv e r )
C o n n e c to r
4 . R e q u e s t
C o n n e c tio n
1 . C re a te M
6 . A d d C o n n e c tio n
C o n n e c to r M
5 . R e q u e s t
C o n n e c tio n
U s e r I n te r fa c e ,
A p p lic a tio n o r
A g e n t N e ig h b o u r
o n ito r
o n ito r
7 . A c c e p t C o n n e c tio n
C o m m u n ic a tio n S e rv ic e
8 . In itia lis e S tre a m s
C o n n e c tio n
Figure 2: Initialisation structure of communication server
proportionate selection method is usually used. To modify classier strengths,
the given credit assignment algorithm is used, e.g. the Bucket brigade 14 .
To create new classiers a standard GA is applied, with three basic genetic
operators: selection, crossover and mutation. The GA is invoked periodically
and each time it replaces low strength classiers with the o spring of the selected
tter ones the reader is referred to 14 for full details .
2.2 Communication Server
The communication server provides the service for each agent connecting to the
user interface, application and other agents. All these channels might involve
two-way communication. During the initialisation the agents open a communication channel and await a connection message. The channel is tested to ensure
the communication is setup correctly as although the conguration parameters
have been previously checked for consistency, the parameters may be inconsistent
with the physical communication process, also this process might be faulty.
Although the term socket is used in the specication as the medium to
connect communication channels in implementation other methods can be used,
such as calls to remote objects, when using RMI or DCOM. The basis of the
communication initialisation and run-time processes are not a ected. The creation of the communication object and binding in a remote registry on a given
hostname and port replaces the creation of a server socket and calls to the
remote object replace read and write calls to the sockets.
Note that it is possible for the communication server to create separate
processes to listen on the communication channel for messages, this allows the
agent to be reactive to external messages.
The communication with neighbours requires a single channel for incoming
messages, and separate channels from sending to each neighbor except if the
messages are broadcast on sent via a proxy . The initialisation of the communication server, shown in Figure 2, involves the following steps:
D is trib u te d L e a rn in g C o n tro l o f T ra ffic S ig n a ls
1 2 1
1. The Communication Server object creates the specic Communication Services
Application, User Interface or Neighbourhood as specied by the conguration.
2. The Communication Server object create a monitor which maintains the list of
current connections
3. The Communication Server object passes the Communication Service object
and connection conguration information to the Connector object which, for
connection with the User Interface and Application and incoming channel from
the neighbouring agents, opens a Server Socket on the specied port and waits
for a request to connect. For the outgoing channel to the neighbouring agents the
Connector object intermittently requests a connection to the neighbours specied
port..
4. The User Interface, Application or Neighbouring Agent sends a request to connect.
5. The request to connect is accepted by the neighbour's server socket.
6. The Connector sends the Communication Service object and open socket to the
Communication Monitor.
7. The Communication Monitor object tests the communication channel, if the
test succeeds the Communication Service is passed to the Connection object,
otherwise the socket is closed and the failure reported.
8. The Connection object starts the thread to handle the connection and passes
the input and output streams to the Communication Service object.
3 How to Control Trac Signals
To control a trac network, we associate an agent to each junction of the trac
network. The agents are initialised according to the trac network conguration
and user-specied parameters. For the simulated 2 2 trac network, shown in
Figure 3, four agents, i.e., agents I, II, III, and IV, associating with junctions I,
II, III, and IV, are need to provide comprehensive control of the network. Agent I
has the neighbouring agents II and III, and agent II has the neighbouring agents
I and IV, etc. The communication server in each agent provides the control
actions of its neighbouring agents, and these information is used to construct
control rules for its junction.
The classier system employed is a version of Wilson's zeroth-level" system
ZCS 15. ZCS is a Michigan-style classier system, without internal memory.
In order to avoid the genetic algorithm manipulating unnecessarily long rules, we
extend the binary string representation in ZCS to a more general representation,
which uses 0 to L L 10 for each variable bit position instead of the binary
code. This reduces the string length signicantly and appears to benet multiple
variable problems. For these hybrid strings, mutation in the GA is performed
by changing an allele to a randomly determined number between 0 and L other
than itself 16.
1 2 2
Y .J . C a o e t a l.
I
II
III
IV
Figure 3: The simulated trac environment
3.1 Individuals
The classiers have the representation shown in Figure 4. The condition part of
each classier consists of six bits, which re ects the scalar level of queue length
from each direction and the previous actions of the neighbouring agents. In this
application, the scalar level of the queue length is set to 4, which ranges from 0 to
3, corresponding to the four linguistic variables, fzero small medium large g.
The action part indicates the required state of the signal. For instance, for
junction I, the rule 130201:1 says that if the queue from directions east and west
are small 1 and zero 0, but the queue from directions south and north are
large 3 and medium 2, and the previous neighbourhood junction controllers'
actions are vertically red 0 junction II and green 1 junction III, then the
trac light stays green vertically 1 for a xed period of time.
3.2 Evaluation of actions
We assume that the junction controller can observe the performance around it,
let the evaluated performance be P . Trac volume sensors are set at each of
the intersections. They are able to count the numbers of the cars that come
from all directions, pass through the intersection and stop at the intersection.
In this study, the evaluation function we use to reward the individuals is the
average queue at the specic junction. Let qi denote the queue length from
direction
i at the intersection i = 1 2 3 4, then the evaluation function is:
1 P4
q . We thus attempted to minimize this measure. Let us identify
f =
i=1 i
4
the k-th cycle by a subscript k, then fk for the cycle k is calculated by observing
the sensor from the beginning of the k-th cycle to the end of this cycle. Thus,
the evaluated performance of the action performed at the k-th cycle is computed
as Pk = fk,1 , fk . Specically, if Pk 0, the matched classiers containing the
performed action should be rewarded, otherwise penalized.
D is trib u te d L e a rn in g C o n tro l o f T ra ffic S ig n a ls
A c tio n
C o n d itio n
T ra ffic c o n d itio n
fro m e a st
T ra ffic c o n d itio n
fro m n o rth
T ra ffic c o n d itio n
fro m s o u th
1 2 3
T ra ffic c o n d itio n
fro m w e st
T ra ffic lig h t s ta te
N e ig h b o u rh o o d a c tio n s
Figure 4: Structure of the classier system rules
3.3 Reinforcement learning
After the controller has produced an action, the environment judges the output,
and accordingly, gives payo in the following manner:
Rewards: The objective of the local signal controller is to minimize the
average queue length, . We have found the performance-based rewards
are helpful in the environments we used in our experiments. The reward
P
function we used was = 14 4=1 100 , 4 + , ,1 3 , where
denotes the queue length of the th direction at the th cycle.
Punishments: We use punishments i.e., negative rewards . We found the
the use of appropriate punishments results in improved performance in a
xed number of cycles , at least in the environments used in our experiments. We also found that large punishments could lead to instability of
the classi ers and slow convergences of the rules. The appropriate punishments should be determined by trial tests.
fi
r
qki
q
qki
ki
i
qk
i
i
k
4 Simulation Results
For the trac network shown in Figure 3, we developed a simpli ed trac simulator, which is similar to the one used in 12. The simulator is composed of
four four-armed junctions and squared roads. Each end of a road is assumed to
be connected to external trac, and cars are assumed to arrive at those roads
according to a Poisson distribution. Each intersection has two complementary"
signals: when the horizontal signal is red, the vertical signal is green and vice
versa. Each of the cars attempts to attain the same maximum speed. When a
car passes an intersection, it changes its direction according to the probability
associated with that intersection. Speci cally, let , = 1 2 3, be the next
directions for a car, that is, f g = f right, forward, left g. At each of the intersections, the probabilities f g are previously given, where corresponds to
di
i
di
pdi
pdi
1 2 4
Y .J . C a o e t a l.
7.5
7
Traffic Speed
6.5
6
5.5
5
4.5
0
1000
2000
a
3000
4000
5000
6000
Time Steps
7000
8000
9000
10000
9000
10000
Number of Cars = 30
6.5
6
5.5
Traffic Speed
5
4.5
4
3.5
3
2.5
0
1000
2000
b
3000
4000
5000
6000
Time Steps
7000
8000
Number of Cars = 60
the probability of selecting an action d for the car passing through the intersection. Roads are not endless, thus only a limited number of cars is allowed to be
on the road at a given time. If a car reaches the end of the road, then the car is
simply removed from the simulation, and another car is generated, entering on
a randomly selected road.
For comparison purpose, two types of control strategies are employed: random control strategy and the developed distributed learning control DLC strategy. The random control strategy determines the tra c light's state 0 or 1
randomly at 50 of probability whilst distributed learning control DLC strategy determines the tra c light's state according to the action of the winning
classier of the agent. The parameters used for the DLC were as follows:
i
Population size: 100
Mutation probability: 0.05
Crossover probability: 0.85
Selection method: Roulette wheel selection
D is trib u te d L e a rn in g C o n tro l o f T ra ffic S ig n a ls
1 2 5
5.5
5
4.5
Traffic Speed
4
3.5
3
2.5
2
1.5
0
1000
2000
c
3000
4000
5000
6000
Time Steps
7000
8000
9000
10000
Number of Cars = 90
Figure 5: Comparison performance of the control strategies
As the major task is to test whether the proposed DLC can learn some good
rules in the trac network, experiments were carried out for three di erent types
of trac conditions. In these simulations, the mean arrival rates for the cars
are the same but the number of cars in the area is limited to 30, 60, and 90,
corresponding to a sparse, medium, and crowded trac condition. In all cases,
the DLC strategy is found to learn how to reduce the average queue length
and improve the trac speed in the network. For example, Figure 5 shows
the average performances of the random control strategy and DLC strategy
respectively over 10 runs in all cases, where the solid line represents DLC strategy
and the dotted line represents random control strategy. It can be seen that the
DLC strategy consistently learns and improves the trac speed over 10,000
iterations.
5 Conclusion and Future Work
In this paper we have presented a distributed learning control strategy for trac
signals. The simulation results on a simplied trac environment are encouraging since we have shown that the developed control strategy can learn to coordinate and determine useful control rules within a dynamic environment. This
preliminary work needs, of course, a number of extensions. We are currently
extending this work in a number of directions, particularly examining ways of
improving the learning capability of classier systems and the performances in
much more complicated trac network.
6 Acknowledgment
This work was carried out as part of the ESPRIT Framework V Vintage project
ESPRIT 25.569.
1 2 6
Y .J . C a o e t a l.
References
1 Robertson, D. I.: TRANSYT A trac network study tool. Transport and Research Laboratory, Crowthorne, England 1969
2 Luk, J. Y., Sims, A. G. and Lowrie, P. R.: SCATS application and eld comparison
with TRANSYT optimized xed time system. In Proc. IEE Int. Conf. Road Trac
Signalling, London 1982
3 Lowrie, P. R.: The Sydney coordinated adaptive trac system. In Proc. IEE Int.
Conf. Road Trac Signalling, London 1982
4 Hunt, P. B., Robertson, D. I., Bretherton, R. D. and Winston, R. I.: SCOOTA
trac responsive method of co-ordinating trac signals. Transport and Research
Laboratory, Crowthorne, England 1982
5 Scemama, G.: Trac control practices in urban areas. Ann. Rev. Report of the
Natl Res. Inst. on Transport and Safety. Paris, France 1990
6 Al-Khalili, A. J.: Urban trac control a general approach. IEEE Trans. on Syst.
Man and Cyber. 15, 1985 260271
7 Montana, D. J. and Czerwinski, S.: Evolving control laws for a network of trac
signals. Proc. of 1st Annual Conf. on Genetic Programming, 1996 333338
8 Koza, J. R: Genetic Programming. MIT Press, Cambridge, MA 1992
9 Cao, Y. J., Ireson, N. I., Bull, L. and Miles, R.: Design of Trac Junction Controller Using a Classier System and Fuzzy Logic. In Computational Intelligence:
Theory and Applications, Reusch, B. ed , Lecture Notes in Computer Sciences,
1625, Springer Verlag, 1999 342353
10 Cao, Y. J. and Wu, Q. H.: An improved evolutionary programming approach to
economic dispatch. International Journal of Engineering Intelligent Systems, 6,
2 , 1998 187194
11 Cao, Y. J. and Wu, Q. H.: Optimisation of control parameters in genetic algorithms: a stochastic approach. International Journal of Systems Science, 20, 2 ,
1999 551559
12 Mikami, S. and Kakazu, K.: Genetic reinforcement learning for cooperative trafc signal control. Proceedings of the IEEE World Congress on Computational
Intelligence, 1994 223229
13 Escazut, C. and Fogarty, T. C.: Coevolving classier systems to control trac
signals. In Koza, J. R ed : Late breaking papers at the Genetic Programming
1997 Conference, Stanford University, 1997 5156
14 Holland, J. H.: Adaptation in Natural and Articial Systems. MIT Press, Cambridge, MA 1992
15 Wilson, S. W.: ZCS: A zeroth level classier system. Evolutionary Computation,
2, 1994 118
16 Cao, Y. J. and Wu, Q. H.: A mixed-variable evolutionary programming for optimisation of mechanical design. International Journal of Engineering Intelligent
Systems, 7, 2 , 1999 7782
T im e S e r ie s P r e d ic tio n b y G r o w in g L a te r a l D e la y
N e u r a l N e tw o r k s
L ip to n C h a n a n d Y u n L i
C e n tre fo r S y s te m s a n d C o n tro l, a n d D e p a rtm e n t o f E le c tric a l a n d E le c tro n ic s E n g in e e rin g ,
U n iv e r s ity o f G la s g o w , G la s g o w G 1 2 8 L T , U .K .
E m a i l :
L . C h a n @ e l e c . g l a . a c . u k
A b s tr a c t. T im e -s e rie s p re d ic tio n a n d fo re c a s
s c ie n c e a n d e c o n o m ic s . N e u ra l n e tw o rk s
p ro b le m s . H o w e v e r, th e d e s ig n o f th e s e n e tw o
u n d e rs ta n d in g to o b ta in u s e fu l re s u lts . In th is
b a s e d in n o v a tiv e te c h n iq u e to g ro w n e tw o
s im p lify th e ta s k o f tim e -s e rie s p re d ic tio n . A
th is n e tw o rk is a ls o g iv e n to ta k e a d v a n ta g
n e tw o rk is n o t re s tric te d to tim e -s e rie s p re d
m o d e llin g d y n a m ic s y s te m s .
1 .
tin g is m u c h u s e d in e n g in e e rin g ,
a re o fte n u s e d fo r th is ty p e o f
rk s re q u ire s m u c h e x p e rie n c e a n d
p a p e r, a n e v o lu tio n a ry c o m p u tin g
rk a rc h ite c tu re is d e v e lo p e d to
n e ffic ie n t tra in in g a lg o rith m fo r
e o f th e n e tw o rk d e s ig n . T h is
ic tio n a n d c a n a ls o b e u s e d fo r
I n tr o d u c tio n
D y n a m ic m o d e llin g a d d re s s e s th e m o d e llin g p ro b le m fro m d a ta o f a d y n a m ic s y s te m .
A d y n a m ic s y s te m is a s y s te m w h ic h h a s in te rn a l s ta te s re p re s e n te d in a n a b s tra c t
p h a s e o r tim e s p a c e . Its fu tu re s ta te a n d o u tp u ts d e p e n d s o n its c u rre n t s ta te . T h e y
c a n b e m a th e m a tic a lly d e s c rib e d b y a n in itia l v a lu e p ro b le m [1 ]. A n e x a m p le o f
d y n a m ic m o d e llin g is th e m o d e llin g o f tim e s e rie s d a ta , w h e re p re d ic tio n s h a s to b e
m a d e o n th e fu tu re v a lu e s o f th e tim e s e rie s b a s e d o n c u rre n t v a lu e s o f th e s e rie s .
T h is ty p e o f m o d e llin g trie s to c a p tu re th e g e o m e try a n d g e o m e tric a l in v a ria n ts o f a
d y n a m ic s y s te m fro m p a s t o u tp u ts o f th e s y s te m [2 ]. T h e u s e o f p a s t o u tp u ts , d e la y
c o -o rd in a te s , to m o d e l d y n a m ic s y s te m s c a n b e tra c e d b a c k a s fa r a s 1 9 2 7 to th e w o rk
o f Y u le , w h o u s e d a u to -re g re s s io n (A R ) to c re a te a p re d ic tiv e m o d e l fo r s u n s p o t
c y c le s [1 ], [3 ].
T h e m o s t p o p u la r m e th o d o f m o d e llin g tim e -s e rie s d a ta to d a y is th e s ta tis tic a l
m e th o d o f B o x -J e n k in s . T h e B o x -J e n k in s m e th o d o lo g y s e a rc h fo r a n a d e q u a te m o d e l
fro m A R , m o v in g a v e ra g e (M A ), a u to -re g re s s io n m o v in g a v e ra g e (A R M A ), a n d a u to re g re s s io n in te g ra te d m o v in g a v e ra g e (A R IM A ) [5 ]. T h is m o d e llin g is a th re e -s ta g e
p ro c e s s : id e n tific a tio n , e s tim a tio n , a n d d ia g n o s tic s .
T h e id e n tific a tio n in v o lv e s th e u s e o f s a m p le a u to c o rre la tio n fu n c tio n s (S A C F )
a n d s a m p le p a rtia l a u to c o rre la tio n fu n c tio n s (S P A C F ) to a n a ly s is th e lin e a r
re la tio n s h ip s o f th e tim e s e rie s w ith its la g g e d im a g e s . T h e e s tim a tio n p ro c e s s
in v o lv e s fin d in g a m o d e l, o n e o f A R , M A , A R M A , a n d A R IM A , w ith a th e o re tic a l
A C F a n d P A C F s im ila r to th e S A C F a n d S P A C F o f th e tim e s e rie s . T h e th ird s ta g e ,
d ia g n o s tic s , in v o lv e s re s id u a l a n a ly s is , g o o d n e s s o f fit s ta tis tic s , a n d c ro s s -v a lid a tio n .
T h e lim ita tio n o f th is m e th o d is th a t h u m a n d e c is io n a n d a s s o c ia te d e rro rs a re in h e re n t
S . C a g n o n i e t a l. ( E d s .) : E v o W o r k s h o p s 2 0 0 0 , L N C S 1 8 0 3 , p p . 1 2 7 − 1 3 8 , 2 0 0 0 .
S p rin g e r-V e rla g B e rlin H e id e lb e rg 2 0 0 0
1 2 8
L . C h a n a n d Y . L i
in e v e ry s ta g e o f th e p ro c e s s [6 ]. A ls o th e S A C F a n d S P A C F fu n c tio n s m e a s u re o n ly
lin e a r re la tio n s h ip s .
N e u ra l n e tw o rk s h a v e b e e n s h o w n to b e u n iv e rs a l fu n c tio n a p p ro x im a to rs [4 ]. B y
u s in g d e la y c o -o rd in a te s a s th e in p u ts o f a n e u ra l n e tw o rk , it c a n b e u s e d a s a n o n lin e a r a p p ro x im a to r fo r a d e la y d iffe re n tia l e q u a tio n . W ith a n a p p ro p ria te s tru c tu re
a n d le a rn in g s tra te g y , th is d e la y d iffe re n tia l e q u a tio n c a n b e tu n e d to h a v e s im ila r
b e h a v io u r to th e d y n a m ic s y s te m b e in g m o d e lle d .
N e u ra l n e tw o rk s a re c o m m o n ly u s e d fo r tim e -s e rie s p re d ic tio n s a ls o . D iffe re n t
n e tw o rk a rc h ite c tu re s h a v e b e e n e m p lo y e d to ta c k le th e p re d ic tio n o f tim e -s e rie s d a ta ,
fo r e x a m p le th e m u lti-la y e r p e rc e p tro n s [7 ], th e fin ite -d u ra tio n im p u ls e re s p o n s e
(F IR ) n e tw o rk s [8 ], a n d th e re c u rre n t n e tw o rk s [9 ]. G o o d p re d ic tio n re s u lts h a v e b e e n
o b ta in e d fro m s u c h n e tw o rk s . T h e d iffic u lty w ith th e s e a rc h ite c tu re s is th a t th e y a re
s ta tic , fix e d b e fo re th e tra in in g is b e g u n . T h a t is , th e d e s ig n e r n e e d s to d e c id e o n th e
n u m b e r o f d e la y c o -o rd in a te s to u s e , th e n u m b e r o f h id d e n n e u ro n s to h a v e , e tc .
In th is p a p e r, a n o v e l n e tw o rk a rc h ite c tu re a n d tra in in g s tra te g y is p ro p o s e d fo r th e
m o d e llin g o f d y n a m ic s y s te m s w h ic h a lle v ia te s th e d e s ig n e r o f m u c h o f th e s e
d e c is io n s . T h e n e tw o rk s tru c tu re is n o t s ta tic a n d c h a n g e s d u rin g th e tra in in g p ro c e s s
w h ic h m a k e s u s e o f e v o lu tio n a ry a lg o rith m s (E A ). A d e s c rip tio n o f th e n o v e l
n e tw o rk a rc h ite c tu re is g iv e n in S e c tio n 2 . T h e e v o lu tio n a ry te c h n iq u e fo r g ro w in g
a n d tra in in g th e n e tw o rk is d e s c rib e d in S e c tio n 3 . In S e c tio n 4 , re s u lts o f p re d ic tin g
c h a o tic tim e s e rie s u s in g th is te c h n iq u e a n d n e tw o rk a re s h o w n .
S e c tio n 5
d e m o n s tra te s th e a p p lic a tio n o f tim e s e rie s p re d ic tio n o n re a l d a ta , a n d s h o w s a
m e th o d o f im p ro v in g p re d ic tio n . F in a lly , th e p a p e r is c o n c lu d e d in S e c tio n 6 .
2 .
L a te r a l D e la y N e u r a l N e tw o r k
A n e u ra l n e tw o rk re q u ire s m e m o ry to h a v e d y n a m ic b e h a v io u r [8 ]. T h is m e m o ry c a n
b e d e la y e le m e n ts in th e a rc h ite c tu re o r th e u s e d e la y c o -o rd in a te in p u ts . T h e n u m b e r
o f d e la y c o -o rd in a te in p u ts re la te s to th e e m b e d d in g d im e n s io n o f th e s y s te m , a n d th e
n u m b e r o f h id d e n n e u ro n s is d ic ta te d b y n e c e s s a ry d e g re e s o f fre e d o m [1 0 ]. B o th o f
th e s e h a v e to b e e s tim a te d fo r s ta tic n e tw o rk a rc h ite c tu re s .
T h e p rin c ip le o f g ro w in g th e n e tw o rk s tru c tu re o n e h id d e n n e u ro n a t a tim e h a s
s h o w n to b e a fa s t a n d e ffic ie n t m e th o d o f a p p ro x im a tin g a fu n c tio n w ith a n e u ra l
n e tw o rk [1 1 ]. S e v e ra l s u c h in c re m e n ta l le a rn in g a lg o rith m s h a v e b e e n im p le m e n te d
[1 1 ], [1 2 ]. It is d e s ira b le to u s e th e s e in c re m e n ta l a p p ro a c h e s to s p e e d th e le a rn in g
p ro c e s s a n d re d u c e th e d im e n s io n a lity o f o p tim is a tio n . Y e t th e re a re tw o s tru c tu ra l
p a ra m e te rs th a t n e e d s to b e e s tim a te d , n a m e ly , n u m b e rs o f d e la y s a n d h id d e n n e u ro n s .
T h e d e s ig n o f th e la te ra l d e la y n e u ra l n e tw o rk (L D N N ) c o m b in e s th e s e tw o
p a ra m e te rs in to o n e fo r th e im p le m e n ta tio n o f in c re m e n ta l le a rn in g a lg o rith m s . T h is
a llo w s fo r th e in c re a s e d s p e e d a n d e ffic ie n c y o f m o d e llin g o f d y n a m ic s y s te m s .
T h e c o m b in a tio n o f d e la y e le m e n ts a n d h id d e n n e u ro n s is s h o w n in F ig . 1 o f th e
n e tw o rk a rc h ite c tu re . T h is a rc h ite c tu re is in c re m e n ta l w ith th e h id d e n n e u ro n s
fo rm in g a o n e -w a y c h a in o f d e la y e d e le m e n ts . T h e s im p lic ity o f d e s ig n is fa c ilita te d
b y n o t re q u irin g s tru c tu ra l d e c is io n s . Im p ro v in g th e p e rfo rm a n c e o f n e tw o rk is
a c h ie v e d b y a d d in g a n e u ro n to th e la te ra l c h a in . T h e re a re v e ry fe w s y n a p tic w e ig h ts
a s s o c ia te d w ith e a c h n e u ro n , th u s th e c o m p le x ity o f th e w e ig h t o p tim is a tio n is v e ry
lo w .
T im e S e rie s P re d ic tio n b y G ro w in g L a te ra l D e la y N e u ra l N e tw o rk s
x
1
x
2
z
-1
z
-1
1 2 9
yˆ
. . .
. . .
x
m
z
-1
F ig . 1 . A rc h ite c tu re o f L D N N .
T h e L D N N
s y s te m in p u ts
p re d ic tio n , th e
th e o u tp u t is
w h e re x n is th e
3 .
3 .1 .
c a n
a n d
re is
th e p
th
n d
b e u s e d in d y n a m ic
i t s o u t p u t yˆ b e i n g
o n ly o n e in p u t w h ic
re d ic te d fu tu re v a lu
a ta in th e s e rie s .
m o
th e
h th
e o
d e llin g w
p re d ic te
is th e c u
f tim e -s e
ith its in p
d s y s te m
rre n t v a lu
r ie s , i.e .,
u ts x ∈ ℜ m
o u tp u t. I
e o f th e tim
x = [x n ] a
b e in g th
n tim e -s e
e -s e rie s ,
n d xˆ n + 1 =
e m
rie s
a n d
yˆ ,
E v o lu tio n a r y N e tw o r k G r o w in g a n d T r a in in g
E A -o n ly T r a in in g
D u e to th e L D N N a rc h ite c tu re ’s s im p lic ity a n d in c re m e n ta l d e s ig n , tra in in g
a lg o rith m s c a n b e d e v is e d to ta k e a d v a n ta g e o f th e s e p ro p e rtie s . E A s c a n b e u s e d to
tra in th e n e tw o rk in c re m e n ta lly b y firs t o p tim is in g th e s y n a p tic w e ig h ts o f th e
n e tw o r k w ith o n ly o n e h id d e n n e u r o n , i.e ., n o d e la y e le m e n ts y e t. T h e b e s t s o lu tio n
fo u n d is th e n u s e d to “ h o t-s ta rt” s u b s e q u e n t E A o p tim is a tio n s o f th e n e tw o rk w ith a n
a d d e d h id d e n n e u ro n . T h is c a n b e ite ra te d u n til th e d e s ire d a c c u ra c y is re a c h e d o r
u n til o v e r-fittin g b e g in s to o c c u r.
U s in g E A fo r tra in in g in th is w a y c a n p ro d u c e g o o d re s u lts , b u t th e d im e n s io n a lity
o f th e o p tim is a tio n s p a c e in c re a s e s ra p id ly . If th e re is a w e ig h t a s s o c ia te d w ith e a c h
s y n a p tic c o n n e c tio n a n d o n e fo r th e th re s h o ld o f e a c h n e u ro n , th e n th e d im e n s io n a lity
in v o lv e d is (m + 3 )n − 1 w h e re m is th e n u m b e r o f in p u ts a n d n is th e n u m b e r o f h id d e n
n e u ro n s.
1 3 0
L . C h a n a n d Y . L i
3 .2 .
O r th o g o n a l T r a in in g w ith E A
T h e d im e n s io n a lity o f th e o p tim is a tio n c a n b e g re a tly re d u c e d b
o p tim is e d w e ig h ts o f th e p re v io u s n e tw o rk fix e d a n d o n ly o p tim is in g
th e n e w ly a d d e d n e u ro n . A le a rn in g a lg o rith m b a s e d o n th is id e a is
L D N N h e re .
B e lic z y n s k i g a v e a n in c re m e n ta l a lg o rith m
fo r tra in in g o n
p e rc e p tro n s [1 1 ]. T h is a lg o rith m is s lig h tly m o d ifie d to a c c o m m o d a te
L D N N s a n d la te ra l d e la y s ; th e p ro o fs c a n b e fo u n d in h is w o rk [1 1 ].
T h u s , firs t a s s u m e th a t th e tim e s e rie s to b e m o d e lle d is d e fin e d b y
in p u t-o u tp u t p a irs :
{ ( x 1, f( a 1, x 1) ) , ( x 2, f( a 2, x 2) ) , …
y k e e p in g th e
th e w e ig h ts o f
g iv e n fo r th is
e -h id d e n -la y e r
th e le a rn in g o f
a fin ite s e t o f
, (x N, f(a N, x N))}
(1 )
w h e re f : ℜ
→ ℜ is th e s y s te m o u tp u t (th e n e x t d a ta ), x j is a n in p u t (th e c u rre n t
c
d a ta ), x j ∈ ℜ , a j is th e in te rn a l s ta te o f th e s y s te m , a j ∈ ℜ , j = 1 , … , N , a n d N is th e
n u m b e r o f in p u t-o u tp u t p a irs . It fo llo w s th a t fo r tim e s e rie s ,
c ,1
x
= f ( a j, x j)
j+ 1
, N − 1 . D e fin e th e v e c to r X
w h e re j = 1 , …
o rd e r, a s
g k(
e la
w h
o rd
aˆ
, x
y e d in
e re n
in g to
j,k
[
) d e n o
p u t to
≤ n m ax
th e s tr
j
aˆ
j,k
, f ( a
f (a 1, x 1 ), f (a 2 , x 2 ),
T
x 2 , x 3 , , x N +1 ]
[
te th e fu n c tio
th e h id d e n n e
a n d n m ax is th
u c tu re o f th e
0
=
g
ℜ
T
a n d o u tp u t v e c to r
F ( X ) =
=
o f in p u ts , i.e ., th e tim e s e r ie s d a ta in
, x N] ∈
X = [ x 1, x 2, …
L e t
is th e d
… , N ,
A c c
(2 )
n o f
u ro n
e m a
n e tw
N
N
(3 )
, x
N
)
]T
∈ ℜ
N
(4 )
a h id d e n n e u ro n , w h e re g k : ℜ
→ ℜ , aˆ j , k
, fo r e v e ry n e u ro n k = 1 , … , n , a n d tim e j = 1 ,
x im u m n u m b e r o f h id d e n n e u ro n s a llo w e d .
o rk
1 ,1
if j = 1 , o r k = 1
k − 1
( aˆ
, x
j − 1 ,k − 1
j− 1
)
(5 )
o th e rw is e
w h e re k = 1 , … , n , a n d j = 1 , … , N . T h e d e la y s o f th e n e tw o rk c o n s titu te s th e
i n t e r n a l s t a t e o f t h e n e t w o r k , t h e r e b y t h e n e t w o r k s t a t e v e c t o r aˆ j a n d t h e s t a t e m a t r i x
Aˆ
n
c a n b e d e fin e d a s fo llo w s
aˆ
= [ aˆ
j
Aˆ
n
j 1,
, aˆ
j,2
= [ aˆ 1 , aˆ 2 ,
,
, a ˆ
, a ˆ
N
j,n
]T ∈ ℜ
]T ∈ ℜ
(6 )
n
(7 )
N ,n
D e f in e th e v e c to r G k( X ) w h ic h is c o m p o s e d o f th e o u tp u ts o f th e k
fo r th e w h o le tim e s e rie s
th
h id d e n n e u ro n
T im e S e rie s P re d ic tio n b y G ro w in g L a te ra l D e la y N e u ra l N e tw o rk s
G
k
( X ) = [ g
= [ aˆ
k
( aˆ
2 ,k + 1
, x 1 ), g
, aˆ 3 , k + 1 ,
1 ,k
k
( aˆ
, aˆ
, x 2 ) , , g k ( aˆ N , k , x
T
ˆ
+ 1 , g k ( a N ,k , x N ) ]
2 ,k
N ,k
a n d d e fin e th e m a trix
H
A ls o le t W
n
n
( X ) = [G 1( X ),G
( X ),
2
, G
( X )] ∈ ℜ
n
N
)]
T
∈ ℜ
1 3 1
N
(8 )
(9 )
N ,n
d e n o te th e v e c to r o f w e ig h ts o f th e o u tp u t n e u ro n
W
, w n] T ∈
= [ w 1, w 2, …
n
ℜ
n
(1 0 )
w h e re w k is th e w e ig h t o f th e s y n a p tic c o n n e c tio n fro m n e u ro n k to th e o u tp u t
n e u ro n , k = 1 , ,n .
T h u s th e n e tw o r k ’ s p r e d ic tio n s , F n( X ) , o f th e tim e s e r ie s X is
F
(X ) = H
n
(X )W
n
n
∈
ℜ
N
(1 1 )
N o w th e tra in in g e rro r in n e tw o rk p re d ic tio n c a n b e d e fin e d a s
E
n
F ( X )
F ( X ) − F
( X ) =
fo r n = 0
fo r 1 ≤ n ≤ n
( X )
n
(1 2 )
m a x
a n d th e m e a n s q u a re d tra in in g e rro r is
e n ( X ) =
F ro m
E
n
( X )
[1 1 ],
E
n
( X )
1
E
N
n
( X )
2
(1 3 )
is n o n -in c re a s in g , a n d th e m a x im u m
o c c u rs w h e n th e n e w ly -a d d e d n e u ro n , g
su p E
g ∈ G
n
( X )
G
n + 1
( X )
G
n + 1
( X )
T
n + 1
ra te o f d e c re a s e in
, is c h o s e n s u c h th a t
(1 4 )
is a c h ie v e d . T h is is p r o v id e d th a t n m ax < N . I n p r a c tic e , N h a s to b e la r g e f o r tim e s e r ie s p r e d ic tio n a n d n m ax < < N . T h e q u a n tity in E q u a tio n 1 4 , is th e s c a la r p r o d u c t o f
th e tw o v e c to rs , a n d b y m a x im is in g it, th e tw o v e c to rs a re m a d e q u a s i-p a ra lle l. T h e
e rro r is a ls o o rth o g o n a l to th e o u tp u t o f e v e ry h id d e n n e u ro n in th e n e tw o rk .
T h e re a re tw o o p tim is a tio n s to b e p e rfo rm e d fo r e a c h n e w n e u ro n b e in g a d d e d .
T h e o p tim is a tio n o f th e n e w ly a d d e d h id d e n n e u r o n , g n, a n d W n, th e w e ig h ts o f th e
o u tp u t n e u ro n . T h e s e a re o p tim is e d u s in g E A s . T h e d im e n s io n a lity o f th e s e s e a rc h
a r e m + 1 f o r g n a n d n + 1 f o r W n, w h e r e m is th e n u m b e r o f in p u ts a n d n is th e to ta l
n u m b e r o f h id d e n n e u ro n s . T h e w h o le o rth o g o n a l in c re m e n ta l tra in in g p ro c e s s is
s u m m a ris e d in th e flo w c h a rt in F ig . 2 .
T h e h i d d e n - l a y e r n e u r o n f u n c t i o n , g k ( aˆ j , k , x j) , c a n u s e d i f f e r e n t a c t i v a t i o n
fu n c tio n s . N e ith e r o f th e tw o tra in in g a lg o rith m s d is c u s s e d in th is s e c tio n h a v e m a n y
re s tric tio n s o n th e a c tiv a tio n fu n c tio n s u s e d fo r th e h id d e n n e u ro n s . F o r a ll th e tim e s e rie s p re d ic tio n s m a d e in th is w o rk , s ig m o id fu n c tio n s h a v e b e e n u s e d a s a c tiv a tio n
fu n c tio n s fo r th e h id d e n n e u ro n s . A lin e a r fu n c tio n is c h o s e n fo r th e o u tp u t n e u ro n ’s
1 3 2
L . C h a n a n d Y . L i
a c tiv a tio n . T h e u s e o f s ig m o
b o u n d e d , th u s th e tim e -s e rie s
T h e tra in in g a lg o rith m s in
th e tra in in g p ro c e s s . T h u s th
a n d th e tim e -s e rie s th a t c a n b
id a c tiv
b e in g m
th is s e c
e s y s te m
e p re d ic
a tio n f
o d e lle
tio n a r
s th a t
te d a re
u n c tio n
d m u st
e s ta tic
c a n b e
s ta tio n
s m e
a ls o
; th e
m o d
a ry .
a n s th a t th e n e tw o rk o u tp u t is
b e b o u n d e d .
n e tw o rk d o e s n o t c h a n g e a fte r
e lle d a re a u to n o m o u s s y s te m s
S T A R T
n = 0
F in d g
n + 1
g iv in g
su p E
g ∈ G
U p d a t e s t a t e m a t r i x Aˆ
n
F in d W
n + 1
n
( X )
fo r m in e
w ∈ ℜ
G
n + 1
G
n + 1
T
n + 1
( X )
( X )
u s in g E A
( X ) u s in g E A
n = n + 1
N o
D e s ire d a c c u ra c y
re a c h e d ?
Y e s
E N D
F ig . 2 . F lo w c h a rt o f th e o rth o g o n a l in c re m e n ta l tra in in g p ro c e s s .
T h e re s u lts s h o w n in F ig . 3 a re th e tim e s ta k e n to tra in a n a d d itio n a l n e u ro n b y th e
tw o tra in in g a lg o rith m s a s th e n u m b e r o f n e u ro n s in c re a s e s . T h e s e tim e s a re th e
a v e ra g e tim e s o f 5 n e tw o rk tra in in g s fo r th e p re d ic tio n o f th e M a c k a y -G la s s c h a o tic
tim e -s e rie s . T h e a d d itio n a l tra in in g tim e ta k e n b y th e o rth o g o n a l tra in in g a lg o rith m ,
w ith re d u c e d c o m p le x ity , s ta y s n e a r c o n s ta n t – ris in g o n ly s lig h tly a s th e n u m b e r o f
h id d e n n e u ro n s in c re a s e s . W h e re a s , w ith th e E A -o n ly tra in in g , th e e x tra tim e ta k e n
p e r n e u ro n in c re a s e s e x p o n e n tia lly w ith th e a d d itio n o f h id d e n n e u ro n s .
E x p e rim e n ta tio n s h o w s th a t th e E A -o n ly a lg o rith m o fte n p ro d u c e s v e ry g o o d
p re d ic tio n a c c u ra c y w ith v e ry fe w n e u ro n s , w h ile th e o rth o g o n a l a lg o rith m m a y n e e d
m a n y m o re n e u ro n s to a c h ie v e th e s a m e le v e l o f a c c u ra c y . D u e to th e lo w
c o m p u ta tio n a l p o w e r re q u ire d b y th e o rth o g o n a l tra in in g a lg o rith m , th e re s t o f th e
p re d ic tio n re s u lts in th is p a p e r a re o b ta in e d u s in g th is tra in in g m e th o d .
T im e S e rie s P re d ic tio n b y G ro w in g L a te ra l D e la y N e u ra l N e tw o rk s
A d d itio n a l tra in in g tim e p e r n e u ro n
1 0 0 0
8 0 0
T im e ta k e n (s )
1 3 3
6 0 0
O rth o g o n a l
tra in in g
4 0 0
E A tra in in g
2 0 0
0
2
3
4
H id d e n n e u ro n s
5
F ig . 3 . C o m p a ris o n o f tra in in g tim e s ta k e n b y th e tw o tra in in g a lg o rith m s w h e n a d d in g n e w
n e u ro n s.
4 .
P r e d ic tio n o f C h a o tic S e r ie s
4 .1 .
L o g is tic M a p
P re d ic tio n o f L o g is tic M a p
0
-0 .3
2 s te p
a h e a d
L o g (E rro r)
-0 .6
-0 .9
1 s te p
a h e a d
-1 .2
-1 .5
-1 .8
1
2
3
4
5
F ig . 4 . R M S v a lid a tio n e rro rs o f L D N N
lo g is tic m a p .
6
7
8
9
1 0 H id d e n n e u ro n s
p re d ic tio n 1 s te p a n d 2 s te p s in to th e fu tu re o f th e
T h e lo g is tic m a p is a c h a o tic s e rie s g iv e n b y th e e q u a tio n
x
U s in g th e o r
m a p 1 s te p in to
re s u ltin g v a lid a
h id d e n n e u ro n s
th a t th e tra in in
m o d e l’s g e n e ra
n e u ro n n e tw o rk
n + 1
= 4 x n (1 − x n )
(1 5 )
th o g o n a l tra in in g a lg o rith m , a L D N N is tra in e d to p re d ic t th e lo g is tic
th e fu tu re , a n d a n o th e r L D N N to p re d ic t 2 s te p s in to th e fu tu re . T h e
tio n e rro rs o f p re d ic tio n , o f d a ta u n s e e n d u rin g tra in in g , fo r u p to 1 0
a re p lo tte d in F ig . 4 . T h e u s e o f in c re m e n ta l te c h n iq u e s g u a ra n te e s
g e rro r m o n o to n ic a lly d e c re a s e s . T h e v a lid a tio n e rro rs s h o w th e
lis a tio n c a p a b ility . T h e p re d ic tio n v a lid a tio n v a lu e s o f th e 1 0 h id d e n
a re c o m p a re d w ith th e a c tu a l v a lu e s in F ig . 5 , s h o w in g a c lo s e m a tc h .
1 3 4
L . C h a n a n d Y . L i
L o g is t ic M a p P r e d ic t io n s ( 1 s t e p a h e a d )
1 .2
1
0 .8
0 .6
0 .4
0 .2
0
-0 .2
0
2 0
A c tu a l
1 .2
1
0 .8
0 .6
0 .4
0 .2
0
-0 .2
4 0
6 0
8 0
1 0 0
1 2 0
1 4 0
T im e
P r e d ic t io n
L o g is tic M a p P re d ic tio n s (2 s te p s a h e a d )
0
2 0
A c tu a l
4 0
6 0
8 0
1 0 0
P re d ic tio n
1 2 0
1 4 0
T im e
F ig . 5 . 1 s te p a n d 2 s te p s a h e a d lo g is tic m a p p re d ic tio n re s u lts w ith 1 0 h id d e n n e u ro n s .
4 .2 .
M a c k a y -G la s s S e r ie s
T h e M a c k a y -G la s s d iffe re n tia l e q u a tio n
q u a s i-p e rio d ic . It is d e fin e d a s
x = − b x
t
+
a x
1 + x
p ro d u c e s c h a o tic
t − τ
1 0
t − τ
tim e -s e rie s th a t a re
(1 6 )
T h e d e c re a s e o f p re d ic tio n v a lid a tio n e rro rs w ith th e in c re a s e o f h id d e n n e u ro n s
c a n b e s e e n in F ig . 6 fo r th e p re d ic tio n o f th e M a c k a y -G la s s s e rie s u p to 5 s te p s in to
th e fu tu re . T h e p re d ic tio n v a lid a tio n re s u lts a re s h o w n in re la tio n to th e a c tu a l d a ta
v a lu e s fo r 1 , 3 , a n d 5 s te p s a h e a d p re d ic tio n in F ig . 7 . It c a n b e s e e n th a t p re d ic tio n
re s u lts a re v e ry c lo s e to th e a c tu a l d a ta v a lu e s , th o u g h th e p re d ic t re s u lts g e t m o re
e rra tic th e fu rth e r in to th e fu rth e r th e p re d ic tio n is .
A ll th e s e p re d ic tio n s a re m a d e w ith o n ly o n e in p u t to th e L D N N . It m a y b e
p o s s ib le to im p ro v e o n th e s e re s u lts b y e x p lic itly e m b e d d in g o th e r d is ta n t d e la y c o o r d in a te s a s in p u ts to th e n e tw o r k , f o r e x a m p le x t– 5, x t– 15, e tc .
T im e S e rie s P re d ic tio n b y G ro w in g L a te ra l D e la y N e u ra l N e tw o rk s
P re d ic tio n o f M a c k a y -G la s s S e rie s
5 s te
a h e a
4 s te
a h e a
3 s te
a h e a
2 s te
a h e a
1 s te
a h e a
-0 .6
L o g (E rro r)
-0 .9
-1 .2
-1 .5
-1 .8
-2 .1
1
2
3
4
5
6
7
8
9
1 0
1 3 5
p
d
p
d
p
d
p
d
p
d
H id d e n n e u ro n s
F ig . 6 . R M S v a lid a tio n e rro rs o f p re d ic tio n s o f th e M a c k a y -G la s s s e rie s a t s e v e ra l s te p s in to th e
fu tu re .
M -G P re d ic tio n s (3 s te p a h e a d )
M -G P re d ic tio n s (1 s te p a h e a d )
1 .5
1 .5
1 .2
1 .2
0 .9
0 .9
0 .6
0 .6
0 .3
0 .3
0
0
0
5 0
A c tu a l
1 0 0
1 5 0
2 0 0
2 5 0
P re d ic tio n
3 0 0
T im e
0
5 0
1 0 0
A c tu a l
1 5 0
2 0 0
2 5 0
P re d ic tio n
3 0 0
T im e
M -G P re d ic tio n s (5 s te p a h e a d )
1 .5
1 .2
0 .9
0 .6
0 .3
0
0
5 0
1 0 0
A c tu a l
1 5 0
2 0 0
P re d ic tio n
2 5 0
3 0 0
T im e
F ig . 7 . 1 , 3 a n d 5 s te p s a h e a d M a c k a y -G la s s p re d ic tio n re s u lts w ith 1 0 h id d e n n e u ro n s .
5 .
S u
n u
u s
a p
P r e d ic tio n o f S u n s p o t N u m b e r s
n s p o ts n u m
m b e rs fo rm
e d to tra in
p lic a tio n . T
b e rs a re
a c y c le
a L D N
h e d a ta
in d ic a to
o f a p p r
N , d e m
u se d a re
rs fo
o x im
o n s
y e a
r th e le
a te ly 1
tra tin g
rly n u m
v e l
1 y
th e
b e r
o f s o la r
e a rs. H
u se o f
s fro m 1
a c tiv ity in
e re , y e a rly
th is n e tw
8 5 1 to 1 9 9
th e S
su n s
o rk
8 [1 3
u n . T h e su n sp o t
p o t n u m b e rs a re
a n d its e a s e o f
].
1 3 6
n e
tra
a b
c o
a d
a n
L . C h a n a n d Y . L i
F ig . 8 s h o w s th e R M S e rro rs
u ro n s in c re a s e s . U s in g in c re
in in g e rro rs , a n d th e a n a ly s is
ility o f th e L D N N m o d e l. F
n tin u a lly in c re a s in g w h e n th e
d itio n o f h id d e n n e u ro n s w ill m
d in c re a s e o v e r-fittin g to th e tra
o f p re d ic tio n v a lid a tio n a s th e
m e n ta l tra in in g g u a ra n te e s th e
o f v a lid a tio n e rro rs w ill s h o w
ro m F ig . 8 , th e v a lid a tio n e rr
n u m b e r o f h id d e n n e u ro n s re a
o s t lik e ly d e c re a s e th e m o d e l’s g
in in g d a ta .
m b e r o f h id d e n
c re a s e in R M S
e g e n e ra lis a tio n
a p p e a r to s ta rt
d 6 -7 . F u rth e r
ra lis a tio n a b ility
P re d ic tio n o f Y e a rly S u n s p o t N u m b e rs
1 .7 5
1 .6 5
L o g (E rro r)
n u
d e
th
o rs
c h e
e n e
2 s te
a h e a
1 s te
a h e a
1 .5 5
1 .4 5
p
d
p
d
1 .3 5
1 .2 5
1
2
3
4
5
6
7
8
9
1 0 H id d e n n e u ro n s
F ig . 8 . R M S v a lid a tio n e rro rs fo r th e p re d ic tio n o f y e a rly s u n s p o t n u m b e rs .
T o im p ro v e o n th e s e re s u lts , o n e c a n c h o o s e to u s e a d iffe re n t a c tiv a tio n fu n c tio n
fo r s u b s e q u e n tly a d d e d h id d e n n e u ro n s o n th e d e te c tio n o f o v e r-fittin g . A lte rn a tiv e ly ,
e x p lic itly e m b e d d in g d e la y c o -o rd in a te s a s in p u ts to th e n e tw o rk c a n b e e m p lo y e d .
T h e la tte r a p p ro a c h is u s e d in th is c a s e a n d th e re s u lts o f F ig . 8 c a n b e u s e d to s u g g e s t
w h ic h d e la y c o -o rd in a te to e m b e d . S in c e o v e r-fittin g o c c u rs a t a b o u t th e a d d itio n o f
th
th
th e 7 h id d e n n e u ro n a n d h e n c e th e b e s t n e tw o rk o u tp u ts d e p e n d o n o n ly u p to th e 7
th
d e la y c o - o r d in a te , th e 8 d e la y c o - o r d in a te is c h o s e n , i.e .,
xˆ
xˆ
w h e re a
n
= f1 (a
n + 1
n + 2
=
f2 (a
n
n
, x n , x
n − 7
, x n , x
n − 7
(1 7 )
)
(1 8 )
)
is th e in te rn a l s ta te o f th e m o d e l, a n d x n is th e n
v a lu e o f th e tim e s e rie s .
P re d ic tio n o f Y e a rly S u n s p o t N u m b e rs
1 .7 5
1 .6 5
L o g (E rro r)
th
2 s te p
a h e a d
1 .5 5
1 s te p
a h e a d
1 .4 5
1 .3 5
1 .2 5
1
2
3
4
5
6
7
8
9
1 0 H id d e n n e u ro n s
F ig . 9 . R M S v a lid a tio n e rro rs w ith th e e x p lic it e m b e d d in g o f a n e x tra d e la y c o -o rd in a te .
T im e S e rie s P re d ic tio n b y G ro w in g L a te ra l D e la y N e u ra l N e tw o rk s
T h
n e u ro
d o e s
a n d v
re sp e
e re s u ltin
n s c a n b
n o t o c c u r
a lid a tio n
c tiv e ly .
g d e c re a s
e s e e n in
so so o n w
re s u lts fo
e in
F ig
ith
r 1
R M
. 9
th is
a n d
S v a lid a tio n
w h e n u s in g a
a rra n g e m e n t
2 s te p s a h e a d
e rro
n e
a n d
p re
rs w ith
x tra d e
g e n e ra
d ic tio n
in
la y
lis
s a
c re a s in g n u m
c o -o rd in a te
a tio n is g o o d
re s h o w n in
2 0 0
2 0 0
1 6 0
1 6 0
1 2 0
1 2 0
8 0
8 0
4 0
4 0
1 3 7
b e r o f h id d e
. O v e r-fittin
. T h e tra in in
F ig . 1 0 a n d 1
n
g
g
1
0
0
0
2 0
4 0
6 0
A c tu a l
8 0
1 0 0
0
1 2 0
1 0
T ra in in g re s u lt
2 0
3 0
4 0
A c tu a l
V a lid a t io n r e s u lt
F ig . 1 0 . T ra in in g a n d v a lid a tio n re s u lts fo r 1 s te p a h e a d p re d ic tio n o f s u n s p o t n u m b e rs .
2 0 0
2 0 0
1 6 0
1 6 0
1 2 0
1 2 0
8 0
8 0
4 0
4 0
0
0
0
2 0
4 0
A c tu a l
6 0
8 0
1 0 0
1 2 0
T ra in in g re s u lt
0
1 0
2 0
3 0
A c tu a l
V a lid a tio n re s u lt
4 0
F ig . 1 1 . T ra in in g a n d v a lid a tio n re s u lts fo r 2 s te p a h e a d p re d ic tio n o f s u n s p o t n u m b e rs .
A
6 .
C o n c lu s io n s
n o v e l a rc h ite c tu re a n d g ro w in g te c h n iq u e a re p ro p o
s im p lifie s th e d e s ig n p ro c e s s o f n e u ra l n e tw o rk s fo r tim
e v o lu tio n a ry in c re m e n ta l tra in in g a lg o rith m p re s e n te d fo r
s h o w s its e lf to b e a fa s t a n d e ffic ie n t m e th o d o f a p p ro x
d y n a m ic s y s te m . T h e tra in in g a lg o rith m a llo w s fo r d
a c tiv a tio n s fu n c tio n s to b e u s e d in th e n e tw o rk . It is p o s s ib
s e d in th is p a p e r w h ic h
e -s e rie s p re d ic tio n . T h e
th is n e tw o rk a rc h ite c tu re
im a tin g th e n e tw o rk to a
iffe re n t c o m b in a tio n s o f
le th a t a c h a n g e in h id d e n
1 3 8
L . C h a n a n d Y . L i
n e
c o
In
o f
e x
a s
u ro n a c tiv a tio n fu n c tio n w ill in c re a s e th e ra te o f e rro r re d u c tio n , w h e n th is ra te is
n v e rg in g .
a d d itio n to th e p re d ic tio n o f tim e -s e rie s , th e n e tw o rk m a y b e u s e d fo r th e m o d e llin g
d y n a m ic s y s te m s . Im p ro v e d re s u lts m a y b e o b ta in e d w ith th e a d d itio n o f
o g e n o u s v a ria b le s o r e x p lic it e m b e d d in g o f d e la y c o -o rd in a te s . W ith s ta tic tra in in g
g iv e n in th is p a p e r, a u to n o m o u s s y s te m s a n d s ta tio n a ry tim e -s e rie s c a n b e
m o d e lle d . T h is m e th o d o lo g y is v a lid a te d b y th e a p p lic a tio n to a re a l e x a m p le , n a m e ly
th e p re d ic tio n o f s u n s p o t n u m b e rs . Its e a s e a n d e ffic ie n c y is d e m o n s tra te d a lo n g w ith
e x p lic it e m b e d d in g o f d e la y c o -o rd in a te s to im p ro v e re s u lts . F o r th e m o d e llin g o f
n o n -a u to n o m o u s s y s te m s a n d n o n -s ta tio n a ry tim e -s e rie s , re a l-tim e a d a p ta tio n o f th is
n e tw o rk m a y b e re q u ire d , w h ic h is a s u b je c t o f o n g o in g re s e a rc h h e re in C S C ,
U n iv e rs ity o f G la s g o w .
7 .
R e fe r e n c e s
1 . M e is s J .: N o n lin e a r S c ie n c e F A Q , M a y 1 9 9 9 , V e r . 1 .3 .1 . I n te r n e t F A Q C o n s o r tiu m ,
h ttp ://w w w .f a q s .o r g /f a q s /s c i/n o n lin e a r - f a q /. O n lin e .
2 . P a c k a r d N .H ., e t a l.: G e o m e tr y f r o m a T im e S e r ie s . P h y s ic a l R e v ie w L e tte r s , V o l. 4 5 , N o . 9 .
(1 9 8 0 ) 7 1 2 -7 1 6
3 . Y u le G .U .: P h ilo s o p h ic a l T r a n s a c tio n s o f th e R o y a l S o c ie ty o f L o n d o n A , V o l. 2 2 6 . ( 1 9 2 7 )
2 6 7
4 . H o r n ik K ., S tin c h c o m b e M ., a n d W h ite H .: M u ltila y e r F e e d f o r w a r d N e tw o r k s a r e U n iv e r s a l
A p p ro x im a to rs . N e u ra l N e tw o rk s , V o l. 2 . (1 9 8 9 ) 3 5 9 -3 6 6
5 . N e e r c h a l N .K .: T im e D o m a in , A u g u s t 1 9 9 9 . T im e S e r ie s T u to r : A n I n te r a c tiv e I n tr o d u c tio n
to T im e S e r ie s A n a ly s is , h ttp ://m a th .u m b c .e d u /~ n a g a r a j/. O n lin e .
6 . E lk a te b M .M ., S o la im a n K ., a n d A l- T u r k i Y .: A c o m p a r a tiv e s tu d y o f m e d iu m - w e a th e r d e p e n d e n t lo a d fo re c a s tin g u s in g e n h a n c e d a rtific ia l/fu z z y n e u ra l n e tw o rk a n d s ta tis tic a l
te c h n iq u e s . N e u ro c o m p u tin g , V o l. 2 3 . (1 9 9 8 ) 3 -1 3
7 . C o n w a y A .J ., e t a l.: A n e u r a l n e tw o r k p r e d ic tio n o f s o la r c y c le 2 3 . J o u r n a l o f G e o p h y s ic a l
R e s e a rc h , V o l. 1 0 3 , N o . A 1 2 . (1 9 9 8 ) 2 9 7 3 3 -2 9 7 4 2
8 . H a y k in S .: N e u r a l N e tw o r k s . M a c m illa n . ( 1 9 9 4 )
9 . M a ts u o k a M ., G o le a M ., a n d S a k a k ib a r a Y .: C o lu m n a r R e c u r r e n t N e u r a l N e tw o r k a n d T im e
S e rie s A n a ly s is . F u jits u S c ie n tific & T e c h n ic a l J o u rn a l, V o l. 3 2 , N o . 2 . (1 9 9 6 ) 1 8 3 -1 9 1
1 0 .L o w e D ., a n d H a z a r ik a N .: C o m p le x ity m o d e llin g a n d s ta b ility c h a r a c te r is a tio n f o r lo n g
te r m ite r a te d tim e s e r ie s p r e d ic tio n . I E E C o n f e r e n c e P u b lic a tio n , N o .4 4 0 . ( 1 9 9 7 ) 5 3 - 5 8
1 1 .B e lic z y n s k i B .: I n c r e m e n ta l A p p r o x im a tio n b y O n e - H id d e n - L a y e r N e u r a l N e tw o r k s :
D is c re te F u n c tio n s R a p p ro c h e m e n t. IE E E
In te rn a tio n a l S y m p o s iu m
o n In d u s tria l
E le c tr o n ic s V o l.1 . ( 1 9 9 6 ) 3 9 2 - 3 9 7
1 2 .F r itz k e B .: F a s t le a r n in g w ith in c r e m e n ta l R B F n e tw o r k s . N e u r a l P r o c e s s in g L e tte r s 1 .
(1 9 9 4 ) 2 -5
1 3 .S u n s p o t N u m b e r s , O c to b e r 1 9 9 9 . S o la r - T e r r e s tr ia l P h y s ic s D iv is io n o f th e N a tio n a l
G e o p h y s ic a l D a ta C e n te r , h ttp ://w w w .n g d c .n o a a .g o v /s tp /s tp .h tm l. O n lin e .
T r a je c to r y C o n tr o lle r N e tw o r k a n d I ts D e s ig n
A u to m a tio n th r o u g h E v o lu tio n a r y C o m p u tin g
G re g o ry C h o n g a n d Y u n L i
C e n t r e f o rS y s t e m s &
C o n tr o l , D e p a r t m e n t o f E l e c t r o n i c s &
E le c t r i c a l E n g i n e e r i n g
U n i v e r s i t y o f G l sa g o w , G l a s g o w , G 1 2 8 L T , U K .
g r e g c c y @ e l e c . g l a . a c . u k
A b s t r a c t . C la s s i c a l c o n t r o l l e r s a r e h i g h l y p o p u l a r i n in d u s t r i a l a p p l i c a t
io n s .
H o w e v e r , m o s t c o n t r o l l e r s a r e t u n e d m a n u a l l y in a t r i a l a n d e r r o r p r o c e s s
t h o u g h c o m p u t e r s im u l a t i o n . T h i s i s p a r t i c u l a r l y d i f f i c u l t w h e n t h e s y s te m t o b e
c o n t r o l l e d is n o n l i n e a r . T o a d d r e s s t h is p r o b l e m a n d h e l p d e s i g n o f i n d u s t r i a l
c o n t r o l l e r s f o r a w i d e r ra n g e o f o p e r a t i n g t ra j e c t o r y , t h i s p a p e r p r o p o s e s a
t r a j e c t o r y c o n t r o l l e r n e t w o r k ( T C N ) te c h n i q u e b a s e d o n l i n e a r a p p r o x i m a t i o n
m o d e l ( L A M ) t e c h n i q u e . I n a T C N, e a ch c on t r o l l e r c an b e o f a s i m p l e f o r m ,
w h ic h m a y b e ob ta in e d s tr a ig h tfo r w a r d ly v ia c la s s ic a l d e s ig n or e v o lu tio n a ry
m e a n s . T o c o - o dr i n a t e t h e o v e r a l l c o n t r o l l e r p e r f o r m a n c e , t h e s c h e d u l i n g o f h t e
T C N i s e v o l v e d t h r o u g h t h e e n t i r e o p e r a t i n g e n v e l o p e . S i n c e p la n t s t e p
r e s p o n s e d a t a a r e o f te n r e a d i l y a v a i l a b l e i n e n g i n e e r i n g p r a c t i c e ,
th e de s ig n o f
s u c h T C N is f u l l y a u t o m a t e d u s i n g a n e v o l u t i o n a r y a l g o r i t h m w i t h o u t t h e n e e d
o f m o d e l i d e n t i f i c a t i o n . T h i s i s l il u s t r a t e d a n d v a l i d a t e d t h r o u g h a n o n l i n e a r
c o n t r o l ex a m p l e .
1
I n tr o d u c tio n
A
d y n a m ic e n g in e e rin g s y s te m is u s u a lly n o n lin e a r a n d c o m p le x in p ra c tic e . P la n t
d y n a m ic s m a y v a ry s ig n ific a n tly w ith c h a n g e s o f o p e ra tin g c o n d itio n s . T h e re fo re , th e
u s e o f a s in g le n o m in a l lin e a r m o d e l u n d e r o n e o p e ra tin g c o n d itio n , a n d h e n c e
c o n tro lle rs d e s ig n e d o u t o f s u c h a p la n t m o d e l, a re o fte n u n re lia b le a n d in a d e q u a te to
re p re s e n t a p ra c tic a l s y s te m . T h e re c e n tly d e v e lo p e d lo c a l c o n tro lle r n e tw o rk
te c h n iq u e s [5 ], h a v e p ro v id e d s o m e e ffe c tiv e s o lu tio n s to th e s e p ro b le m s , b u t th e y a re
b a s e d o n lo c a lly lin e a ris e d m o d e ls .
T o a d d re s s th e s e p ro b le m s m o re c o m p le te ly fo r a w id e r ra n g e o f o p e ra tin g
tra je c to rie s a n d to m a k e u s e o f p la n t s te p -re s p o n s e d a ta th a t a re o fte n re a d ily a v a ila b le
in e n g in e e rin g p ra c tic e , th is p a p e r p ro p o s e s a tra je c to ry c o n tro lle r n e tw o rk (T C N )
te c h n iq u e b a s e d o n lin e a r a p p ro x im a tio n m o d e l (L A M ) te c h n iq u e [2 ]. S u c h a L A M
n e tw o rk is o b ta in a b le d ire c tly fro m
p la n t s te p -re s p o n s e b y fittin g n o n lin e a r
tra je c to rie s b e tw e e n tw o o p e ra tin g le v e ls . A s p re lim in a rie s to d e s ig n , th is m o d e llin g
te c h n iq u e is o u tlin e d in S e c tio n 2 .
In a T C N , e a c h c o n tro lle r c a n b e o f a s im p le fo rm , s u c h a s a p ro p o rtio n a l p lu s
in te g ra l p lu s d e riv a tiv e (P ID ) c o n tro lle r, w h ic h m a y b e o b ta in e d s tra ig h tfo rw a rd ly v ia
S . C a g n o n i e t a l. ( E d s .) : E v o W o r k s h o p s 2 0 0 0 , L N C S 1 8 0 3 , p p . 1 3 9 − 1 4 6 , 2 0 0 0 .
S p rin g e r-V e rla g B e rlin H e id e lb e rg 2 0 0 0
1 4 0
G . C h o n g a n d Y . L i
c la s s ic a l
p e rfo rm a n
e n v e lo p e .
te c h n iq u e
S e c tio n 5 .
2
d e s ig n
c e , th e
T h is is
th ro u g h
o r
sc
d e
a
e v o lu tio
h e d u lin g
ta ile d in
n o n lin e a
n a r
o f
S e c
r c
y
m
th e
tio n
o n tr
e a n s
T C N
3 . S
o l e x
. T o
is e v
e c tio n
a m p le
c o -o r
o lv e d
4 illu
. F in a
d in a
th r
s tra
lly ,
te
o u g
te s
c o n
th e
o v e ra ll
th e e n tire
a n d v a lid a te s
c lu s io n s a re
h
L in e a r A p p r o x im a tio n M o d e l fo r N o n lin e a r S y s te m
H e re , th e L A M to a p
T h e p la n t u s e d fo r th e
m o d e ls liq u id -le v e l fo
a ls o b e fo u n d in th e
e q u a tio n s , th e s y s te m
=
h 2
a te a n o n lin e a r
le is a tw in -ta n k
c h e m ic a l a n d d
to ry . B a s e d o n
re is g iv e n b y :
− sg n (h1 − h
h 1
p ro x im
e x a m p
u n d in
la b o ra
s tru c tu
sg n (h1 − h
2
)
c1a
A
1
2
)
c1a
A
2 g h1 − h
p la n
c o u
ia ry
th e
2
−
c 2a
A
s tra
n lin
T h e
lli’s
te d
e a r
sc
m
th ro u
h y d ra
a le d d
a ss-b a
2 g h1 − h
1
t is illu
p le d n o
p la n ts .
B e rn o u
Q
+ A
0
2
2 g (h
2
− H
2
)
0
tro lle r
ra tin g
T C N
w n in
M o d e llin g
g h a
u lic
o w n
la n c
0 v i
0 0
1
c o n
o p e
th e
d ra
n e x a m
s y s te m
m o d e l
e a n d f
p le .
th a t
c a n
lo w
(1 )
T h e s y s t e m i n p u t i s t h e v o l t a g e a p p l i e d t o t h e p u m p , v i, a n d t h e s y s t e m o u t p u t i s t h e
liq u id le v e l in ta n k 2 , h 2. T h e c o e ff ic ie n ts o f th e tw in ta n k a r e ta b u la te d in T a b le 1 .
T h e n o n -lin e a rity o f th e p la n t m o d e l is c le a rly p lo tte d a s s h o w n in F ig . 1 .
T a b le 1 . N o n lin e a r s y s te m
p a ra m e te rs
H e ig h t o f w a te r in ta n k 1
h 1(m )
H e ig h t o f w a te r in ta n k 2
h 2(m )
m in im u m
H
h e ig h t o f w a te r in ta n k
= 0 .0 3 m
0
2
C ro s s s e c tio n a l a re a o f ta n k 1 & 2
A = 0 .0 1 m
D is c h a rg e c o e ffic ie n t o f o rific e 1
c 1 = 0 .5 3
D is c h a rg e c o e ffic ie n t o f o rific e 2
c 2 = 0 .6 3
C ro s s s e c tio n a l a re a o f o rific e 1
a
= 0 .0 0 0 0 3 9 6 m
2
1
C ro s s s e c tio n a l a re a o f o rific e 2
a
= 0 .0 0 0 0 3 8 6 m
2
2
G ra v ita tio n a l c o n s ta n t
g = 9 .8 1 m
P e r-v o lt P u m p F lo w ra te
Q
F lo w ra te fro m
D is c h a rg e ra te
s
-2
= 0 .0 0 0 0 0 7 (m
Q
3
1
(m
s )
Q
(m
3
o
s )
ta n k 1 to ta n k 2
i
-1
-1
3
s
-1
V
-1
)
T ra je c to ry C o n tro lle r N e tw o rk
1 4 1
output (h2) versus input(vi )
0.18
h2, Liquid level (m )
0.16
0.14
Tank 2
0.12
0.1
0.08
0.06
0.04
0.02
0
0
2
4
6
8
Pump voltage, vi (volt)
F ig . 1 . N o n -lin e a rity o f th e p la n t. h
v o lta g e v i is a p p lie d to th e p u m p
2
is th e s te a d y s ta te liq u id le v e l o f ta n k 2 w h e n th e in p u t
If th re e o p e ra tin g p o in ts o n th e n o n lin e
b a s e d o n th e ra te o f c h a n g e o r e q u a lly d
liq u id le v e l. T h e s im p le d iv is io n is u s e d
L A M m o d e ls . T h e th re e o p e ra tin g p o in ts
re s p o n s e s fro m th e L A M a t th e s e p o in ts a
a r
iv
b
a r
re
tra je c to ry
id e d fro m
e c a u se o f
e 0 .0 5 m , 0
s h o w n in F
a re
th e
th e
.1 m
ig .
u se
fu ll
tra je
a n d
2 .
d , th
o p e r
c to ry
0 .1 5
e y
a tin
c a
m .
c a n b e c
g ra n g e
p a b ility
G e n e ra te
Step response from LAM
Tank 2 Liquid level(m)
0.16
0.15m
0.14
0.12
0.1
0.1m
0.08
0.06
0.05m
0.04
0.02
0
0
200
400
600
Time (sec)
F ig . 2 . S te p re s p o n s e s a t th re e o p e ra tin g p o in ts o f a L A M
800
n e tw o rk
1000
h o
o f
o f
d s
se
th
th
te
n
e
p
e
1 4 2
G . C h o n g a n d Y . L i
3
E v o lv in g a T C N
M a
s im
n o n
n o n
n y c o n tro l s y s te m d e s ig n m e th o d s a re b a s e d o n lin e a r s y s te m s a n a ly s is .
p le P ID c o n tro l s y s te m is d e s ig n fo r a L A M . T o a p p ly th e T C N te c h n
lin e a r p la n ts , c o n tro lle rs m u s t b e d e s ig n e d fo r th e e n tire L A M n e tw o rk
lin e a r s y s te m .
T h is is to p ro v id e a d e q u a te p e rfo rm a n c e a c ro s s th e o p e ra tin g e n v e lo p e
s y s te m . T h re e P ID c o n tro lle rs in th e T C N a re s c h e d u le d o r s w itc h b e tw e e n
s h o w n in F ig . 3 . D u rin g c o n tro lle r o p e ra tio n , a v a ria b le in d ic a tin g o p e ra tin g
m o n ito re d a n d d iffe re n t c o n tro lle rs (o r c o n tro lle r p a ra m e te rs ) a re a c tiv a te d a c
to th is s c h e d u lin g v a ria b le . In th is d e s ig n , th e p la n t o u tp u t y (t) is u s e d a s s c h
v a ria b le to s c h e d u le th e o u tp u t o f th e c o n tro lle rs .
H e re , a
iq u e to
o f th e
o f th e
th e m a s
p o in t is
c o rd in g
e d u lin g
S c h e d u lin g v a ria b le
u
C o n tro lle r1
r
e
2
u
3
C o n tro lle r2
u
1
In te rp o la tin g
C o n tro lle r
O u tp u t
u
N o n -lin e a r
p la n t
A c tu a to r
C o n tro lle r3
A n ti W in d -u p
F ig . 3 . M u ltip le c o n tro lle rs b a s e d tra je c to ry c o n tro lle r n e tw o rk .
T h e T C N u s e s a lin e a r in te rp o la tio n o r w e ig h tin g s c h e d u le a s s h o w n in F ig . 4 .
Weight
100%
Controller 1
P1
Controller 2
y(t)
P2
F ig . 4 . A s im p le in te rp o la tio n s c h e d u le in fo rm in g a T C N
Controller 3
P3
Operating Level
y
T ra je c to ry C o n tro lle r N e tw o rk
1 4 3
T h e r e f o r e , a t a n y o u t p u t l e v e l y ( t ) , t h e i n d i v i d u a l c o n t r o l l e r o u t p u t s u i( t ) a r e
in te r p o la te d g iv in g a f in a l c o n tr o llin g o u tp u t u ( t) u s in g e q u a tio n ( 2 ) , w h e r e P 1= 0 .0 5 m ,
P 2= 0 .1 m a n d P 3= 0 .1 5 m .
P
P
i+ 1
i+
y (t) − P
− y (t)
× u i(t) +
P i+ 1 − P
1 − P i
u (t) =
u
u
i
× u
i+ 1
(t)
if P
i
≤ y (t) ≤ P
i+ 1
i
if y (t) < P
i
if y (t) > P
n
i
n
(2 )
w h e r e i= 1 ,…
,n -1 a n d n is th e to ta l n u m b e r o f lin e a r c o n tro lle r
4
H e re , in te rp o la tio n m a y a ls o a p p lie d to th e c o n tro lle r p a ra m e te rs K
P
, K
i
a n d K d.
D e s ig n E x a m p le a n d V a lid a tio n
4 .1 G e n e r a t in g T r a j e c t o r y C o n t r o lle r s f r o m
S te p R e sp o n se s
In d iv id u a l P ID c o n tro lle rs fro m a s te p -re s p o n s e tra je c to ry to e a c h o f th e th re e
o p e ra tin g p o in ts a re g e n e ra te d fro m th e P ID e a s y ™ d e s ig n a u to m a tio n p a c k a g e [4 ], a s
s h o w n in F ig . 5 .
F ig . 5 . D ire c t d e s ig n fro m
p la n t re s p o n s e u s in g P ID e a s y ™
P ID e a s y ™ a n a ly s e s s te p re s p o n s e d a ta a n d g e n e ra te s a n a p p ro p ria te P ID c o n tro lle r
fro m th e m . A t e a c h o p e ra tin g p o in t, fittin g th e s te p re s p o n s e g e n e ra te d fro m a L A M
p ro d u c e s fa s t g e n e ra tio n o f P ID c o n tro lle r. T h e c lo s e d lo o p re s p o n s e s a t th e s e
1 4 4
G . C h o n g a n d Y . L i
o p e ra tin g p o in ts a re p lo tte d o n th e s a m e g ra p h s h o w n in F ig . 6 . N o te th a t th e fa s t
g e n e ra tio n o f P ID c o n tro lle rs u s in g lin e a r P ID e a s y ™ te c h n iq u e is te s te d a g a in s t th e
n o n lin e a r p la n t. T h is re v e a ls th e n e e d o f n e tw o rk tu n in g .
Performance of each individual trajectory PID controller
0.18
Tank 2 Liquid Level (m)
0.16
0.15m
0.14
0.125m
0.12
0.1
0.1m
0.08
0.075m
0.06
0.05m
0.04
0.02
0
0
200
400
600
800
1000
1200
Times(sec)
F ig . 6 . P e rfo rm a n c e o f e a c h In d iv id u a l tra je c to ry P ID c o n tro lle r
4 .2 .
N e tw o r k in g T h r o u g h E v o lu tio n
A
g e n e tic a lg o rith m (G
p ro b le m s b y e m u la tin g
u s in g th e e v o lu tio n a ry
o p tim a l s o lu tio n s . O n e
fitn e s s fu n c tio n n e e d s
m in im is e d is th e s u m
o p e ra tin g p o in ts w ith in
A ) p ro v id e s g lo b a lly o p tim a l s o
n a tu ra l e v o lu tio n . A p o p u la tio n
o p e ra to rs o f c ro s s o v e r, m u ta tio
a d v a n ta g e o f a G A fo r o p tim is
n o t to b e d iffe re n tia b le . H e re ,
m a tio n o f a ll e rro rs a c ro s s th e
a g iv e n tim e p e rio d m .
J =
H e re ,
sh o w n
0 .1 5 m
p o in ts
th e
in
. T
. E
re a re n = 5 re f
F ig .7 . T h e 5
h e s e 5 p o in ts
a c h re fe re n c e
∑
n
lu tio n s to e n g in e e rin g d e s ig
o f p o te n tia l s o lu tio n s e v o lv e
n a n d s e le c tio n to a p p ro a c
a tio n is th a t th e o b je c tiv e o
th e o b je c tiv e fu n c tio n to b
e n tire T C N a t n re fe re n c
n
s
h
r
e
e
m
∑ e (t)
re f = 1 t= 0
e re n c e le v e ls u s e d to e v a
re fe re n c e le v e ls a re s e t
c o v e r th e w h o le tra je c to
is te s te d fo r a p e rio d o f m
lu a te
to 0
ry a n
= 1 0
e (t)= | r (t)– y (t) |.
(3 )
th e
.0 5 m
d in
0 0 s
e rr
, 0
c lu
e c ,
o r tra c k in g p e rfo rm a n c e a s
.0 7 5 m , 0 .1 m , 0 .1 2 5 m a n d
d in g tw o u n s e e n o p e ra tin g
w h e re
(4 )
re p re s e n ts th e tra c k in g e rro r b e tw e e n th e c lo s e d -lo o p o u tp u t y (t) a n d th e c o m m a n d
r (t).
T ra je c to ry C o n tro lle r N e tw o rk
Controller 1
Controller 2
1 4 5
Controller 3
100%
=Reference points
R1
R2
R3
R4
R5
Operating Level
F ig . 7 . E v a lu a tio n p o in ts in th e o p e ra tin g e n v e lo p e .
T o e v o
p a ra m e
s im u lta
T h e c lo
a p o p u
d e p ic te
th e n o n
lv e T
te rs o
n e o u s
s e d lo
la tio n
d in F
lin e a r
C N b a s e d o n th e fa s t g e n e ra te d P ID
f th e th re e lin e a r c o n tro lle rs a n d th e
ly , a t th e s e e n a n d u n s e e n o p e ra tin g p
o p re s p o n s e s o f th e fin a lly e v o lv e d T C
s iz e o f 5 0 a r e s h o w n in F ig .8 f o r
ig .7 . I t c a n b e s e e n th a t th e lin e a r T C N
c o n tro l p ro b le m .
c o n tro lle rs o u t o f L A
s c h e d u lin g w e ig h ts , a
o in ts a lo n g in o p e ra tin g
N a t th e e n d o f 5 0 g e n e
a ll o f th e te s te d o p e ra
p ro v id e d a n e x c e lle n t
M , a ll th e
re e v o lv e d
e n v e lo p e .
ra tio n w ith
tin g le v e ls
s o lu tio n to
Closed loop responses of TCN
0.18
0.16
Tank 2 Liquid Level (m)
0.15m
0.14
0.125m
0.12
0.1
0.1m
0.08
0.075m
0.06
0.05m
0.04
0.02
0
0
200
400
600
800
1000
1200
Time(sec)
F ig . 8 . C lo s e d lo o p re s p o n s e s o f th e T C N
0 .1 2 5 m a n d 0 .0 7 5 m .
a t o p e ra tin g p o in ts in c lu d in g th e u n s e e n o n e s a t
1 4 6
G . C h o n g a n d Y . L i
5
D is c u s s io n a n d C o n c lu s io n
T o a s s is t c o n tro l s y s te m d e s ig n fo r a w id e ra n g e o f o p e ra tin g e n v e lo p
p la n ts , th is p a p e r h a s d e v e lo p e d a tra je c to ry c o n tro lle r n e tw o rk (T
b a s e d o n lin e a r a p p ro x im a tio n m o d e l (L A M ) te c h n iq u e . T h e e x a m p le
lin e a r T C N u s e d to c o n tro l a n o n lin e a r s y s te m p e rfo rm s w e ll in th e e
e n v e lo p e . T h is o ffe rs p o te n tia l b e n e fits a n d s im p lic ity fo r c o n tro
s y s te m s . T h e re s u lts s h o w th a t th e G A b a s e d a u to m a tic c o n tro lle r n e tw
n o n lin e a r s y s te m s is p o s s ib le a n d u s e fu l. S u c h a n e tw o rk is e a s ily
s a m p le d re s p o n s e d a ta .
e fo r n o n lin e a r
C N ) te c h n iq u e
s h o w s th a t th e
n tire o p e ra tin g
l o f n o n lin e a r
o rk d e s ig n fo r
d e s ig n e d fro m
R e fe r e n c e s
1 .
2 .
3 .
4 .
5 .
6 .
7 .
8 .
G .J . G r a y , D .J . M u r r a y S m ith , Y . L i, K .C . S h a r m a n , T . W e in b r e n n e r : N o n lin e a r m o d e l
s tr u c tu r e id e n tif ic a tio n u s in g g e n e tic p r o g r a m m in g . C o n tr o l E n g in e e r in g P r a c tic e , V o l.6 ,
N o .1 1 . (1 9 9 8 ) 1 3 4 1 -1 3 5 2
Y . L i a n d K .C . T a n : L in e a r a p p ro x im a tio n m o d e l n e tw o r k a n d its f o r m a tio n v ia
e v o lu tio n a ry c o m p u ta tio n , A c a d e m y P ro c e e d in g s in E n g in e e rin g S c ie n c e s (S A D H A N A ),
In d ia n A c a d e m y o f S c ie n c e s , In v ite d p a p e r (1 9 9 9 )
D .E . G o ld b e r g : G e n e tic A lg o r ith m in S e a r c h , O p tim is a tio n a n d M a c h in e L e a r n in g ,
A d d is o n -W e s le y , R e a d in g (1 9 8 9 )
Y . L i, W . F e n g , K .C . T a n , X .K . Z h u , X . G u a n a n d K .H . A n g : P I D e a s y ™ a n d a u to m a te d
g e n e ra tio n o f o p tim a l P ID
c o n tro lle rs , T h e T h ird A s ia -P a c ific C o n fe re n c e o n
M e a s u re m e n t a n d C o n tro l, D u n h u a n g , C h in a , P le n a ry p a p e r. (1 9 9 8 ) 2 9 -3 3
G .J . G r a y , Y . L i, D .J . M u r r a y - S m ith a n d K .C . S h a r m a n : S p e c if ic a tio n o f a c o n tr o l s y s te m
fitn e s s fu n c tio n u s in g c o n s tra in ts fo r g e n e tic a lg o rith m b a s e d d e s ig n m e th o d s , P ro c . F irs t
I E E /I E E E I n t. C o n f . o n G A in E n g . S y s t.: I n n o v a tio n s a n d A p p l., S h e f f ie ld . ( 1 9 9 5 ) 5 3 0 5 3 5
Y . F a th i: A lin e a r a p p ro x im a tio n m o d e l fo r th e p a ra m e te r d e s ig n p ro b le m , E u ro p e a n
J o u r n a l O f O p e r a tio n a l R e s e a r c h , V o l.9 7 , N o .3 . ( 1 9 9 7 ) 5 6 1 - 5 7 0
K la tt a n d E n g e ll: G a in -s c h e d u lin g tra je c to ry c o n tro l o f a c o n tin u o u s s tirre d ta n k re a c to r,
C o m p u te r s & C h e m ic a l E n g in e e r in g , V o l.2 2 , N o .4 - 5 . ( 1 9 9 8 ) 4 9 1 - 5 0 2
G . C o rrig a , A . G iu a , G . U s a i: A n im p lic it g a in -s c h e d u lin g c o n tro lle r fo r c ra n e s , IE E E
T r a n s a c tio n s O n C o n tr o l S y s te m s T e c h n o lo g y , V o l.6 , N o .1 . ( 1 9 9 8 ) 1 5 - 2 0
E v o lu tio n a r y C o m p u ta tio n a n d N o n lin e a r P r o g r a m m in g
i n M u l t i - mo d e l - ro b u s t C o n t r o l D e s i g n
D o ro th e a K o lo s s a
d o r o t h e a . k o l o s s a @
d a i m l e r c h r y s l e r . c o m
G e o rg G rü b e l
*
g e o r g . g r u e b e l @ i e e e . o r g
A b s tr a c t. A n a lg o rith m ic p a ra m e te r tu n in g m e th o d o lo g y fo r c o n tro lle r d e s ig n
o f c o m p le x s y s te m s is n e e d e d . T h is m e th o d o lo g y s h o u ld o ffe r d e s ig n e rs a g re a t
d e g re e o f fle x ib ility a n d g iv e in s ig h t in to th e p o te n tia ls o f th e c o n tro lle r
s tru c tu re a n d th e c o n s e q u e n c e s o f th e d e s ig n d e c is io n s th a t a re m a d e . S u c h a
m e th o d is p ro p o s e d h e re . F o r a n e x p lo ra to ry p h a s e a n e w p a re to -ra n k e d g e n e tic
a lg o rith m is p ro p o s e d to g e n e ra te a n e v e n ly d is p e rs e d s e t o f n e a r o p tim a l,
g lo b a l, s o lu tio n s . B y p a ir-w is e p re fe re n c e s ta te m e n ts o n d e s ig n a lte rn a tiv e s a
lin e a r p ro g ra m is s e t u p a s a fo rm a l m e a n s fo r s e le c tin g th e s o lu tio n w ith b e s t
o v e ra ll d e s ig n e r s a tis fa c tio n . In a fo llo w in g in te ra c tiv e d e s ig n p h a s e u s in g
n o n lin e a r p ro g ra m m in g te c h n iq u e s w ith a p rio ri d e c is io n s o n a llo w e d q u a lity
le v e ls , a b e s t tu n in g c o m p ro m is e in c o m p e tin g re q u ire m e n ts s a tis fa c tio n is
s e a rc h e d fo r w h ile g u a ra n te e in g p a re to -o p tim a lity . In p a rtic u la r, th is tw o -p h a s e
tu n in g a p p ro a c h a llo w s th e d e s ig n e r to b a la n c e n o m in a l c o n tro l p e rfo rm a n c e
a n d m u lti-m o d e l c o n tro l ro b u s tn e s s .
1
I n tr o d u c tio n
C o n tro l e n g in e e rin g w o rk is m a in ly o c c u p ie d w ith a d a p tin g a c o n tro l s y s te m
a rc h ite c tu re w ith g iv e n c o n tro l la w s tru c tu re , s e n s o rs a n d a c tu a to rs , to n e e d s o f
c h a n g e d p ro d u c t re q u ire m e n ts o r n e w p ro d u c t v e rs io n s . T h is is c a lle d ‘in c re m e n ta l
d e s ig n ’. In p ra c tic e th is o c c u rs m u c h m o re o fte n th a n s ta rtin g c o n tro l s y s te m d e s ig n
a fre s h . T h e e s s e n c e o f in c re m e n ta l d e s ig n is a d a p ta tio n b y tu n in g th e c o n tro l la w
p a ra m e te rs , p a rtia l re p la c e m e n t o r a u g m e n ta tio n o f th e c o n tro l la w s tru c tu re b y
d y n a m ic c o m p e n s a to rs , filte rs , a n d s ig n a l lim ite rs , a n d tu n in g th e o v e ra ll s tru c tu re in
c o n c u rre n c e w ith th e b a s ic c o n tro l la w p a ra m e te rs .
C o m m o n in d u s tria l p ra c tic e is h a rd w a re -in -th e -lo o p , m a n u a l, tu n in g c a lle d
‘c a lib ra tin g ’. H o w e v e r, s in c e th e re m a y b e v e ry m a n y p a ra m e te rs to b e tu n e d m a n u a l
tu n in g is n o t e ffic ie n t n e ith e r in re q u ire d e n g in e e rin g c o s ts n o r in e x p lo itin g th e fu ll
p o te n tia l o f th e c h o s e n s y s te m a rc h ite c tu re w ith re s p e c t to m u ltiv a ria te re q u ire m e n ts .
T h e re fo re , ‘v irtu a l p ro d u c t e n g in e e rin g ’ b a s e d o n h ig h fid e lity s y s te m m o d e l
s im u la tio n s is m o re a n d m o re b e c o m in g th e e n g in e e rin g life s ty le o f c h o ic e . C A C S D ,
i.e ., C o m p u te r A u to m a te d C o n tr o l S y s te m D e s ig n , is th e d is c ip lin e to p r o v id e th e
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _
*
E v o S c o n d i N o d e , D L R - In s titu te o f R o b o tic s a n d M e c h a tro n ic s , O b e rp fa ffe n h o fe n
S . C a g n o n i e t a l. ( E d s .) : E v o W o r k s h o p s 2 0 0 0 , L N C S 1 8 0 3 , p p . 1 4 7 − 1 5 7 , 2 0 0 0 .
S p rin g e r-V e rla g B e rlin H e id e lb e rg 2 0 0 0
1 4 8
D . K o lo s s a a n d G . G rü b e l
p e rtin
a u to m
p a ra m
p a rtic
e n t
a te d
e te r
u la r,
te c h n
sy m
se a r
C A C
o lo
b o
c h
S D
g y in c o
lic a l / n
v ia n o n
s u p p o rts
n tro l
u m e ric
lin e a r
th e d e
synthesis
model / rules
G O
paretooptimal
algorith.
tuning
T
Q
D
in g ,
m d
m in
p u ta
m a k in g u s e o f re c
y n a m ic s m o d e lin g
g o r e v o lu tio n a ry
tio n lo o p d e p ic te d in
executable
s/w specification
G
controller
synthesis
e n g in e e r
a l s y s te
p ro g ra m
s ig n c o m
G O
K
q u a lity fu n c tio n s
4 = ^PD[ /0 +
`
plant (p)
models
commands /
disturbances
evaluation
cases
simulation
analysis
in
ic
In
].
G
controller
model
C
e n t a d v a n c e s
a n d a lg o rith m
c o m p u ta tio n .
F ig u re 1 , c f. [1
characteristics
extraction
L, H
requirements
capture
Q
C
D
quality functions
characteristics
design directors
T
tuner parameters
K
controller parameters
L, H quality scalings
F ig . 1 . C A C S D g e n e ric c o m p u ta tio n lo o p to s u p p o rt in c re m e n ta l c o n tro l d e s ig n
T h is C A C S D c o m p u ta tio n lo o p is g e n e ric in th a t it a llo w s to in c o rp o ra te a n y
c o n tro lle r s tru c tu re to b e tu n e d . T u n in g p a ra m e te rs T m a y re la te to c o n tro lle r
p a r a m e te r s e ith e r d ir e c tly , K = T , e .g ., in P I D c o n tr o l [ 2 ] , o r in d ir e c tly v ia a n a n a ly tic
s y n th e s is m e th o d , K = f ( T , s y n th e s is m o d e l) , e .g ., in H ∞ [ 3 ] c o n tr o l, a s w e ll a s f u z z y
c o n tro l, K = f(T , fu z z y ru le s ), w h e re in th e la tte r c a s e p o s s ib le tu n in g p a ra m e te rs a re
th e s c a lin g fa c to rs o f th e m e m b e rs h ip fu n c tio n s a n d th e w e ig h tin g fa c to rs o f th e fu z z y
c o n tr o l r u le s [ 4 ] . F u r th e r m o r e , it a llo w s to a p p ly a n y a n a ly s is m e th o d , e .g ., lin e a r
m e th o d s in s ta te s p a c e a n d fre q u e n c y d o m a in a s w e ll a s n o n -lin e a r tim e s im u la tio n , to
c o p e w ith n o n -c o m m e n s u ra b le c o n tro l q u a lity e v a lu a tio n s s im u lta n e o u s ly .
T o a s s e s s re q u ire m e n ts s a tis fa c tio n , d e s ig n re q u ire m e n ts a re c a p tu re d fo rm a lly b y
q u a lity fu n c tio n s . T h e fu z z y -ty p e in te rv a l fo rm u la tio n o f q u a lity fu n c tio n s , c f. S e c tio n
2 , a llo w s to d e a l w ith q u a lity le v e ls , e .g ., le v e ls o f ‘ g o o d ’ , ‘ a c c e p ta b le ’ , ‘ b a d ’ ,
re q u ire m e n t s a tis fa c tio n . T h e d a ta o b ta in e d b y e v a lu a tio n o f a ll th e q u a lity fu n c tio n s
fe e d a tu n in g a lg o rith m to c o m p u te p a re to -o p tim a l tu n in g p a ra m e te r v a lu e s . F o r th is
k in d o f d a ta -d riv e n tu n in g b o th e v o lu tio n a ry a lg o rith m s , c f. S e c tio n 3 , o r n o n -lin e a r
p ro g ra m m in g a lg o rith m s , c f. S e c tio n 4 , c a n b e u s e d .
P a r e to - o p tim a lity le n d s its e lf n o t to a u n iq u e s o lu tio n . T h e r e f o r e , in S e c tio n 3 .1 a
n e w m u lti-o b je c tiv e g e n e tic a lg o rith m is p ro p o s e d , w h ic h h a s th e s p e c ia l p ro p e rty th a t
it y ie ld s e v e n ly d is p e rs e d s o lu tio n s in o r n e a r to th e p a re to -o p tim a l s e t, th u s m a k in g
b e s t u s e o f e v o lu tio n a ry c o m p u ta tio n to p ro d u c e a ric h s e t o f d e s ig n a lte rn a tiv e s .
H a v in g m a n y a lte rn a tiv e s a v a ila b le to c h o o s e fro m , s e le c tio n o f th e b e s t c a n d id a te
n e e d s to f o llo w a fo rm a l a p p ro a c h . T h is is d e a lt w ith in S e c tio n 3 .2 .
N o n lin e a r p ro g ra m m in g fo rm u la tio n s in S e c tio n 4 a re u s e d to g e n e ra te d e d ic a te d
p a re to -o p tim a l d e s ig n a lte rn a tiv e s e ith e r to a tta in o p tim a l d e s ig n e r s a tis fa c tio n o r to
ite ra te q u a n tita tiv e c o m p ro m is e s in c o m p e tin g r e q u ir e m e n ts s a tis fa c tio n .
E v o lu tio n a ry C o m p u ta tio n a n d N o n lin e a r P ro g ra m m in g
T h is s u g g e s ts a tw o -p h a s e tu n in g p ro c e d u re to a c h ie
ro b u s tn e s s , S e c tio n 5 . In p h a s e o n e a g lo b a l m u lti-o b je c tiv e d
m o d e l in s ta n tia tio n is c a rrie d o u t u s in g e v o lu tio n a ry c o m p u
in te ra c tiv e n o n lin e a r p ro g ra m m in g c o m p u ta tio n s a re a p p lie d
c o n tro l b e h a v io r w ith o ff-n o m in a l b e h a v io r c h a ra c te riz e d b y
p la n t m o d e l in s ta n tia tio n s . B y th is tu n in g o ff-n o m in a l b e h a v
‘a c c e p ta b le ’ w h ile n o m in a l b e h a v io r is to b e k e p t w ith in th e ‘
2
v e m u lti-m o d e l c o
e s ig n fo r a n o m in a l
ta tio n , a n d in p h a s e
to c o m p ro m is e n o m
a n u m b e r o f o ff-n o m
io r is to b e c o m e a t
g o o d ’ q u a lity le v e l.
1 4 9
n tr
p la
tw
in
in
le a
o l
n t
o
a l
a l
st
R e q u ir e m e n ts C a p tu r e a n d S a tis fa c tio n A s s e s s m e n t
F o r d e s ig n a s s e s s m e n t d e s ig n c h a ra c te ris tic s lik e s y s te m d a m p in g , s te a d y s ta te e rro r,
g a in a n d p h a s e m a rg in s , o r m a x im u m c o n tro l ra te , h a v e to b e tra n s fo rm e d in to a
q u a lity v a lu e w h ic h in d ic a te s th e d e g re e to w h ic h re q u ire m e n ts a re m e t. T w o k in d s o f
m a th e m a tic a l fo rm u la tio n s a re c o m m o n ly in u s e : p o s itiv e d e fin ite ‘th e -s m a lle r-th e b e tte r ’ f u n c tio n a ls o f tim e a n d f r e q u e n c y r e s p o n s e s , e .g ., [ 2 ] , w h ic h o u g h t to b e
m in im iz e d , a n d in e q u a litie s o n th e d e s ig n c h a ra c te ris tic s , w h ic h o u g h t to b e s a tis fie d
a s c o n s tr a in ts , e .g ., [ 5 ] , [ 3 ] .
A d v a n ta g e s o f th e tw o a p p ro a c h e s fo r q u a lity m o d e lin g c a n b e c o m b in e d b y th e s m a lle r-th e -b e tte r in te rv a l q u a lity fu n c tio n s , w h e re re q u ire m e n t s a tis fa c tio n is
c o n s id e re d ‘g o o d ’ fo r o n e ra n g e w ith fu n c tio n v a lu e z e ro , ‘a c c e p ta b le ’ in a ra n g e w ith
fu n c tio n v a lu e n o t g re a te r th a n o n e , a n d ‘b a d ’ o u ts id e a lim itin g ra n g e .
S u c h a n in te rv a l q u a lity fu n c tio n , q (c ), is m a th e m a tic a lly d e fin e d o n th e d e s ig n
c h a r a c t e r i s t i c s , c , b y t h e m a x - o p e r a t o r ( 1 ) w i t h f o u r i n t e r v a l v a l u e s EO < J O < J K < EK
c o m p lia n t w ith ‘b a d ’, ‘a c c e p ta b le ’ a n d ‘g o o d ’ c h a ra c te ris tic s v a lu e s , c f. F ig u re 2 ,
T( F ) = m a x
fu n
o p e
q u a
se t
c o m
h a n
{
/( F ) , 0 , + ( F ) } ,
/( F ) = ( F − J O ) / ( EO − J O ) ,
EO < J O
+ ( F ) = ( F − J K ) / ( EK − J K ) ,
J O < J K < EK .
(1 )
R e q u ire m e n t s a tis fa c tio n is a s s e s s e d a s th e b e tte r, th e s m a lle r a v a lu e th is q u a lity
c tio n a s s u m e s . F u rth e rm o re , th e m a x -fo rm u la tio n fits to fu z z y lo g ic A N D ra tio n w ith m a x -o p e ra to r [7 ] to m a k e o v e ra ll ‘g o o d ’, ‘a c c e p ta b le ’, ‘b a d ’ s y s te m
lity s ta te m e n ts in th e v e in o f fu z z y lo g ic . It a ls o a llo w s to c o m b in e a n e n u m e ra te d
o f c o m m e n s u r a b l e q u a l i t y c h a r a c t e r i s t i c s FN , /N ( FN ) , + N ( FN ) , t o f o r m a
p o u n d q u a lity f u n c tio n f o r , e .g ., ta k in g a c c o u n t o f a ll e ig e n v a lu e s c o n c u r r e n tly o r
d l i n g a l l v a l u e s FN : = F ( W N ) o f a d i s c r e t i z e d t i m e r e s p o n s e a s a n e n t i t y i n
re q u ire m e n ts c a p tu re fo r ro b u s t tra c k in g th u m b p rin t p e rfo rm a n c e [6 ]. W ith H := 0 , th e
‘g o o d ’ in te rv a l is o p e n to th e rig h t, s e e th e e x a m p le o f F ig u re 2 , w ith L := 0 it is o p e n
to th e le ft.
1 5 0
D . K o lo s s a a n d G . G rü b e l
T ζ
1
acceptable
bad
good
L
H
b l = 0 .3
ζ
g l = 0 .7
F ig . 2 . Q u a lity f u n c tio n f o r e ig e n v a lu e d a m p in g , w ith ζ > 0 .7 ‘ g o o d ’ a n d > 0 .3 ‘ a c c e p ta b le ’
3
P a r e to -T u n in g b y G e n e tic A lg o r ith m s a n d D e s ig n P r e fe r e n c e
C o n tro l s y s te m s p a ra m e te r tu n in g is a lw a y s a m u lti-o b je c tiv e p ro b le m
w ith
c o m p e tin g re q u ire m e n ts o f c o n tro l p e rfo rm a n c e v e rs u s c o n tro l e ffo rt a n d ro b u s tn e s s .
H e n c e th e d e s ig n e r’s p rim e ta s k is to s e a rc h fo r a s u ita b le tra d e o ff w h ile g e n e ra tin g
fe a s ib le s o lu tio n s . T h is s e a rc h o u g h t to b e c o n fin e d to th e s e t o f ‘b e s t a c h ie v a b le ’
c o m p ro m is e s o lu tio n s k n o w n a s p a re to -o p tim a l s o lu tio n s . G e n e ra lly , a d e s ig n
a lte rn a tiv e a is s a id to b e p a re to p re fe rre d to a n a lte rn a tiv e a if a ll q u a lity m e a s u re s q
o f a a re b e tte r (s m a lle r) th a n o r e q u a l to th o s e o f a , w ith a t le a s t o n e b e in g s tric tly
b e tte r. T h u s , a p a re to -o p tim a l, o r n o n -d o m in a te d , s o lu tio n is o n e w h e re n o q u a lity
m e a s u re c a n b e im p ro v e d w ith o u t c a u s in g d e g ra d a tio n o f a t le a s t o n e o th e r q u a lity
m e a su re .
T o m a k e a n in fo rm e d tra d e o ff d e c is io n , th e d e s ig n e r n e e d s a ric h s e t o f d e s ig n
a lte rn a tiv e s a s w e ll a s fo rm a l m e th o d s to s u p p o rt a s y s te m a tic s e le c tio n p ro c e s s .
S u c h a m e th o d o lo g y is p ro p o s e d h e re . A n e w p a re to -ra n k e d g e n e tic a lg o rith m
g e n e ra te s a n e v e n ly d is p e rs e d s e t o f d e s ig n a lte rn a tiv e s o f n e a r o p tim a l s o lu tio n s ,
g iv in g th e d e s ig n e r a g lo b a l o v e rv ie w o f w h a t c a n b e a c c o m p lis h e d w ith th e u s e d
c o n tro lle r s tru c tu re , a n d b y th e m e th o d o f p re fe re n c e -d ire c te d d e s ig n th a t s o lu tio n ,
w h ic h re s u lts in g re a te s t o v e ra ll d e s ig n e r s a tis fa c tio n , is s e le c te d .
L
L
3 .1
M
M
A N e w G e n e tic A lg o r ith m
to G e n e r a te a n E v e n ly D is p e r s e d S e t o f S o lu tio n s
A
n a tu ra l w a y o f fin d in g th e p a re to s e t b y g e n e tic a lg o rith m s w a s p ro p o s e d b y
F o n s e c a a n d F le m in g [8 ]. T h is a p p ro a c h m e a s u re s th e fitn e s s o f a n in d iv id u a l b y th e
n u m b e r o f o th e r in d iv id u a ls th a t d o m in a te it in th e s e n s e o f p a re to p re fe re n c e .
A c c o rd in g ly , a p o p u la tio n is ra n k e d , w h e re th e b e s t s o lu tio n s w ill b e th e n o n d o m in a te d o n e s . T h u s , th e n o n -d o m in a te d s o lu tio n s w ill a lw a y s b e m o s t lik e ly to b e
s e le c te d , le a d in g to a c o n v e rg e n c e o f th e p o p u la tio n to th e p a re to s e t.
F ig u re 3 g iv e s a n e x a m p le o f h o w a p o p u la tio n w o u ld b e ra n k e d w ith th is
is th e g o a l. In th is
a lg o rith m , w h e re m in im iz a tio n o f tw o q u a lity m e a s u re s C , C
e x a m p le , th e re a re th re e n o n -d o m in a te d s o lu tio n s w h ic h a re ra n k e d w ith a z e ro ; a ll
E v o lu tio n a ry C o m p u ta tio n a n d N o n lin e a r P ro g ra m m in g
o th e r s o lu tio n s a re d o m in a te d a n d th e ir ra n k
s o lu tio n s a re b e tte r in th e p a re to s e n s e .
is d e te rm in e d
b y
h o w
m a n y
1 5 1
o th e r
&
&
F ig . 3 . P a re to ra n k in g a c c o rd in g to F o n s e c a a n d F le m in g [8 ]
H o w e v e r, d u e to g e n e tic d rift, th e p o p u la tio n te n d s to c o n v e rg e to a s in g le p o in t o n
th e p a re to s u rfa c e . S in c e it is th e a im o f a n e x p lo ra to ry d e s ig n p h a s e to p ro v id e th e
d e s ig n e r w ith a ric h b a s is o f a lte rn a tiv e s fo r p o s s ib le tra d e o ff d e c is io n s , fin d in g o n ly
o n e p a re to s o lu tio n is u n s a tis fa c to ry . T h u s , g e n e tic d rift s h o u ld b e a v o id e d , w h ic h c a n
b e a c c o m p lis h e d b y m e th o d s lik e m u ltip le s u b -p o p u la tio n s o r p e n a liz in g o v e rc ro w d e d
n e ig h b o rh o o d s (fitn e s s s h a rin g ) [3 ].
A v a ria n t o f fitn e s s s h a rin g is p ro p o s e d h e re fo r th e p u rp o s e o f fin d in g a n e v e n ly
d is p e r s e d n e a r-p a re to -o p tim a l s e t o f s o lu tio n s . T h e id e a is to in c o rp o ra te th e
p ro x im ity o f o th e r in d iv id u a ls in to th e fitn e s s o f o n e in d iv id u a l s u c h th a t in d iv id u a ls
in re m o te re g io n s o f th e s e a rc h s p a c e w ill e n jo y a n a d v a n ta g e o v e r th o s e in m o re
o v e rc ro w d e d re g io n s . S in c e s o m e in d iv id u a ls in th e c ro w d e d re g io n s s h o u ld re m a in ,
s o a s to n o t d is tra c t th e g e n e tic a lg o rith m fro m p ro m is in g re g io n s o f th e s e a rc h s p a c e ,
p e n a ltie s fo r c ro w d in g a re lim ite d to v a lu e s le s s th a n 1 . A s th e fitn e s s s c o re fro m
p a re to ra n k in g is a n in te g e r, th is e n s u re s th a t th e in d iv id u a ls a re a lw a y s ra n k e d firs t
b y p a re to d o m in a tio n b u t a m o n g in d iv id u a ls o f th e s a m e p a re to ra n k , th e in d iv id u a ls
a re fo rc e d to s p re a d o u t e v e n ly .
)/(c
-c
) th e
H a v in g m a p p e d th e n c r ite r ia v a lu e s to [ 0 ,1 ] b y c = ( c - c
p e n a lty to a tta in a n e v e n ly d is p e rs e d s e t is c o m p u te d b y th e fo llo w in g fo rm u la :
L
S= (1 − δ ) ⋅
2
2
∑ FL , FXUUHQW − FL , FORVHVW ⋅ 1
L
LPLQ
Q.
LPD[
LPLQ
(2 )
L= 1 . . Q
w h e r e δ i s c h o s e n s m a l l b u t g r e a t e r t h a n z e r o HJ δ = 0 1 FLM s i g n i f i e s t h e L
c r i t e r i o n v a l u e o f t h e M i n d i v i d u a l , ’ FXUUHQW’ i s t h e i n d i v i d u a l f o r w h i c h t h e p e n a l t y i s
c a l c u l a t e d a n d ’ FORVHVW’ i s t h e i n d i v i d u a l w h i c h i s t h e c l o s e s t p r e c e d i n g t h e c u r r e n t o n e
in a lis t o f a ll in d iv id u a ls s o rte d b y p a re to -ra n k . T h is lis t is s o rte d in a s c e n d in g o rd e r,
s o th a t th e b e s t in d iv id u a ls c o m e firs t, a n d b e tw e e n tw o in d iv id u a ls o f th e s a m e ra n k ,
th e p o s itio n in th e lis t is d e c id e d b y ra n d o m .
T h u s , th e p e n a lty is a lw a y s s m a lle r th a n 1 a n d th e a d ju s te d fitn e s s v a lu e is a lw a y s
s tric tly p o s itiv e , w h ic h is n e c e s s a ry fo r s o m e g e n e tic s e a rc h a lg o rith m s .
T h is v a ria n t o f a m u ltio b je c tiv e g e n e tic a lg o rith m h a s b e e n im p le m e n te d b y m e a n s
o f th e G e n e tic a n d E v o lu tio n a ry A lg o rith m T o o lb o x [9 ]. R e s u lts s h o w th a t th is
WK
WK
1 5 2
D . K o lo s s a a n d G . G rü b e l
ra n k in g p ro c e d u re a s s u re s c o n v e rg e n c e to w a rd s th e e n tire p a re to s e t a s o p p o s e d to
e ith e r c o n v e rg in g to ju s t a p a rt o f it o r fa v o rin g re m o te a re a s o f th e s e a rc h s p a c e w ith
le s s th a n o p tim a l c rite ria v a lu e s . T h e e v o lu tio n a ry o p tim iz a tio n is c a rrie d o u t u n til
e ith e r a c e rta in n u m b e r o f g e n e ra tio n s is re a c h e d o r a liv e lin e s s c rite rio n is m e t. B u t
s in c e n o a n a ly tic a l c o n v e rg e n c e c rite rio n is u s e d th e re is n o g u a ra n te e fo r o b ta in in g
e x a c t o p tim a l s o lu tio n s ra th e r th a n a tta in in g a n e a r p a re to -o p tim a l s e t.
O n e e x a m p le o f h o w th is m u ltio b je c tiv e g e n e tic a lg o rith m fa re s w ith p e n a ltie s (2 )
c a n b e s e e n in F ig u re 4 . T h e o p tim iz a tio n re s u lts a re s h o w n to g e th e r w ith th e
b o u n d a ry o f th e s e t o f a ll p o s s ib le s o lu tio n s . A s m e a s u re fo r th e d e g re e o f s p re a d
a lo n g th e p a re to s u rfa c e th e a v e ra g e s ta n d a rd d e v ia tio n in th e o b ta in e d p a re to v a lu e
s e t, c o m p u te d o v e r 1 0 0 0 ru n s a n d s c a le d to a m a x im u m o f 1 , is s h o w n in T a b le 1 .
T a b le 1 . A v e ra g e s ta n d a rd d e v ia tio n in c o m p u te d p a re to v a lu e s e t
W ith o u t P e n a ltie s
0 .1 4 9 3
A v e ra g e S ta n d a rd D e v ia tio n
W ith P e n a ltie s
0 .2 4 1 7
All Generations
4000
: In itia l P o p u la tio n
3000
2000
1000
0
1
2
3
4
5
6
7
8
6
7
8
Final Generation
4000
3000
2000
1000
0
1
2
3
4
5
F ig . 4 . E x a m p le : c o n v e rg e n c e to e v e n ly d is p e rs e d p a re to -o p tim a l v a lu e s e t
3 .2
D e s ig n S e le c tio n fo r B e s t D e s ig n e r S a tis fa c tio n
A fte r a n e
d e s ig n e r s
b y a n o rm
th e s m a lle
v e n ly
a tis fa c
a liz e d
r th e v
d is p e rs e d s e t o f s o lu tio n s is fo u n d , th a t s o lu tio n w h ic h g iv e s h ig h e s t
tio n is to b e s e le c te d . C o m m o n ly , d e s ig n e r s a tis fa c tio n is m e a s u re d
w e ig h te d -s u m v a lu e fu n c tio n v (q ) a n d a d e s ig n is ju d g e d th e b e tte r
a lu e v (q ):
Y = ∑ Z ⋅ T ( 7 ) , w i t h : ∑ ZN = 1 , ∀ N : ZN > 0 , TN ≥ 0 .
N
(3 )
E v o lu tio n a ry C o m p u ta tio n a n d N o n lin e a r P ro g ra m m in g
T h e p ro b le m ,
c o m p lia n c e w ith
m in im a l v a lu e v (
[1 0 ] c o p e s w ith
n u m b e r o f p a ir-w
If a d e s ig n e r m
th is im p lie s
h o w e v e r, is h o w to a ttrib u te
d e s ig n e r’s in te n tio n s to fo rm a
q ). T h e a p o s te rio ri a p p ro a c h o
th is p ro b le m b y s o -c a lle d im
is e p re fe re n c e s ta te m e n ts to b e
a k e s a p re fe re n c e s ta te m e n t “ T
∑
Y( 7 ) =
L
∑
Z ⋅ T (7) <
N
LN
L
N
a p rio ri th e n u m e ric a l w e ig
lly d e c id e o n th e ‘b e s t’ s o lu
f p re fe re n c e -d ire c te d d e s ig n
p re c is e v a lu e fu n c tio n s b a
m a d e b y th e d e s ig n e r.
u n i n g T L i s s u p e r i o r t o T M” t h
Z ⋅ T (7) ,
N
MN
1 5 3
h ts w N in
tio n w ith
s e le c tio n
se d o n a
e n b y (3 )
(4 )
M
N
w h ic h c a n b e re w ritte n a s
∑
∑
Z ( T ( 7 ) − T ( 7) ) > 0 ,
N
MN
M
LN
L
N
W ith k n o w n q , q
a d m is s ib le w e ig h ts w
w h e th e r a c o m p a tib le
w e ll. T h is is fo rm a lly
LN
, a se t
. In [1 0
p re fe re
d e c id e d
MN
N
m in ∑
Z
o f
] th
n c e
b y
O
− 1 ,N
( 72 *
O
− 1
(5 )
N
p re fe re n c e s ta te m e
is c o n s titu te s a n ‘im
o rd e rin g e x is ts a m
s o lv a b ility o f th e fo
Z ( T2 *
N
ZN = 1 , ∀ N : ZN > 0 .
) − T2 *
n ts
p re
o n g
llo w
,
O N
(5 )
c is e
o th
in g
d e s c rib e s
v a lu e fu n
e r d e s ig n
lin e a r p ro
a su b sp a c e fo r
c tio n ’ to c h e c k
a lte rn a tiv e s a s
g ra m :
(6 )
( 72 * )
O
N
V. W.
∑ Z ( T (7 ) − T (7)) > 0
N
MN
M
LN
L
N
:
∑ Z ( T2 *
N
O
− 1 ,N
( 72 *
O
− 1
) − T2 *
,
O N
( 72 * ) ) > 0
O
N
∑ ZN = 1 , ∀ N : ZN > 0 ,
N
, q
w h e re fo r j = (2 * l-1 ) a n d i = (2 * l) th e q u a lity fu n c tio n v a lu e s q
o f a p a ir o f a d d itio n a l in fe rio r a n d s u p e rio r d e s ig n a lte rn a tiv e s , re s p e c tiv e ly .
B a s e d o n a fe w c o m p a tib le p re fe re n c e s ta te m e n ts (5 ) th is fo rm a lis m
p a rtia lly o rd e r th e s e t o f d e s ig n a lte rn a tiv e s fo r s e le c tin g ‘th e b e s t’ o n e . T h e
v a lu e fu n c tio n is a m e a n s to p ru n e in fe rio r p a th s d u rin g d e s ig n s p a c e s e
re d u c in g th e c o m p le x ity o f th e s e le c tio n p ro c e s s . H a v in g fin a lly s e le c te d
a v a ila b le d e s ig n a lte rn a tiv e , a c o m p a tib le s e t o f w e ig h ts c a n b e c o m p u te
m in im iz in g th e d iffe re n c e o f th e v a lu e fu n c tio n o f th e c h o s e n a lte rn a tiv e to
i.e . s u p e r io r , v a lu e o f d e s ig n e r s a tis f a c tio n , w h ic h is c h a r a c te r iz e d b y q = 0 .
ON
N
ON
a re th o s e
a llo w s to
im p re c is e
a rc h th u s
th e b e s t
d b y (6 ),
its id e a l,
1 5 4
4
D . K o lo s s a a n d G . G rü b e l
P a r e to -T u n in g b y I n te r a c tiv e N o n lin e a r P r o g r a m m in g
T h e g e n e tic a lg o rith m w ith p a re to p re fe rre d ra n k in g , a s p r o p o s e d in S e c tio n 3 .1 , fits
w e ll to th e n a tu re o f e v o lu tio n a ry c o m p u ta tio n s in c e o n re tu rn to th e m a n y fu n c tio n
e v a lu a tio n s th a t e v o lu tio n a ry c o m p u ta tio n re q u ire s it y ie ld s a ric h s e t o f s o lu tio n s
e v e n ly d is p e rs e d in o r c lo s e to th e e n tire p a re to -o p tim a l s e t. F u rth e rm o re , g e n e tic
a lg o rith m s c o p e w e ll w ith a la rg e n u m b e r o f p a ra m e te rs a n d w ith a la rg e s e a rc h s p a c e ,
w h ic h m a k e s th e m lik e ly to fin d th e g lo b a l in s te a d o f a lo c a l s o lu tio n in m u ltim o d a l
p ro b le m s . T o g e th e r w ith fo rm a l d e c is io n s u p p o rt, a s d e a lt w ith in S e c tio n 3 .2 , th is is
w e ll s u ite d fo r s e le c tin g a p o s te rio ri a d e s ig n c a n d id a te w ith b e s t d e s ig n e r s a tis fa c tio n
a m o n g a n u m b e r o f g lo b a l d e s ig n a lte rn a tiv e s .
O n th e o th e r s id e , n o n lin e a r p ro g ra m m in g a p p ro a c h e s to a tta in p a re to -o p tim a l
s o lu tio n s a re b a s e d o n a n a n a ly tic a l o p tim a lity c o n d itio n , w h ic h m a k e s th e m v e ry
e ffic ie n t to c o m p u te ju s t o n e , a p rio ri d e d ic a te d , p a re to -o p tim a l s o lu tio n in th e lo c a l
n e ig h b o rh o o d o f w h e re th e a lg o rith m s g e ts s ta rte d . F u rth e rm o re , th e n e c e s s a ry
K a ru s h -K u h n -T u c k e r o p tim a lity c o n d itio n s y ie ld a n u m e ric a l c o n v e rg e n c e c o n d itio n
th a t a llo w s to a tta in a p a re to o p tim u m w ith h ig h a c c u ra c y . T h is m a k e s n o n lin e a r
p ro g ra m m in g a lg o rith m s s u ita b le fo r ‘fin e tu n in g ’. If p a ra m e te riz e d in a d e c is io n
in tu itiv e w a y , in te ra c tiv e , d e c la ra tiv e s e a rc h to a tta in a s p e c ific , ‘b e s t’, c o m p ro m is in g
s o lu tio n fo r re q u ire m e n t s a tis fa c tio n u n d e r lo c a l d e s ig n c o n flic ts b e c o m e s fe a s ib le .
N o n lin e a r P ro g ra m m in g c a n b e u s e d to c o m p u te p a re to -o p tim a l tu n in g v a lu e s T b y
s o lv in g [1 1 ] th e n o rm a liz e d w e ig h te d -s u m m in im iz a tio n p ro b le m , c f. (3 )
m i n ∑ ZN TN ,
7
V. W. ∑ ZN = 1 , ∀ N : ZN > 0 , T ≥ 0 ,
N
N
(7 )
o r th e m in -m a x o p tim iz a tio n p ro b le m
{ T ( 7) / G }
m in m a x
7
L
L
L
, V. W. ≤ T ( 7 ) ≤ G , N = { L, M} .
N
M
(8 )
O p tim iz a tio n (7 ) o r (8 ) is a s u ffic ie n t c o n d itio n to y ie ld a p a re to -o p tim a l s o lu tio n
w ith p a ra m e te riz e d s p e c ific p ro p e rtie s . T h e re a re s ta n d a rd a lg o rith m s o f n o n lin e a r
p r o g r a m m in g , lik e S Q P , to b e u s e d s ta r tin g w ith a ( g lo b a l) s o lu tio n b y S e c tio n 3 .1 .
F o rm u la tio n (7 ) y ie ld s th e s o lu tio n o f o p tim a l d e s ig n e r s a tis fa c tio n w ith re s p e c t to
a ttr ib u te d w e ig h ts a s f o u n d , e .g ., b y th e a p o s te r io r i d e c is io n p r o c e d u r e o f S e c tio n 3 .2 .
A p rio ri s e le c tio n o f w e ig h ts is n o t d e c is io n in tu itiv e [1 1 ] to a tta in s p e c ific p ro p e rtie s .
F o rm u la tio n (8 ) y ie ld s a p a re to -o p tim a l s o lu tio n d e p e n d in g o n p a ra m e te rs d . T h is
fo rm u la tio n is w e ll s u ite d fo r c o m p ro m is in g c o m p e tin g re q u ire m e n ts b y u s in g th e s e
p a ra m e te rs ite ra tiv e ly a s a p rio ri ‘d e s ig n d ir e c to r s ’ to b a la n c e re q u ire m e n ts
s a tis fa c tio n w ith in a fe a s ib le s o lu tio n s e t. A d e c is io n in tu itiv e a p p ro a c h to c h o o s e
d e s ig n d ire c to rs d is n o w p ro p o s e d , w h ic h is in s p ire d b y th e p e rfo rm a n c e v e c to r
d e c is io n s y s te m a tic s d u e to K re is s e lm e ie r [1 2 ].
In itia liz a tio n S te p : S ta rt b y a g lo b a l s o lu tio n w ith d e s ig n e r s a tis fa c tio n a c c o rd in g to
S e c tio n 3 a n d p a re to -o p tim iz e th is s o lu tio n b y s o lv in g th e u n c o n s tra in e d m in -m a x
p ro b le m w ith d = { 1 } , j ∈ ∅ . T h is y ie ld s a b a la n c e d s o lu tio n w h e re a ll q u a lity
fu n c tio n s g e t p a re to -m in im a l a n d th e v a lu e fu n c tio n fo r d e s ig n e r s a tis fa c tio n is fu rth e r
im p ro v e d if th e s ta rt s o lu tio n is n o t y e t a n o p tim iz e d o n e .
L
E v o lu tio n a ry C o m p u ta tio n a n d N o n lin e a r P ro g ra m m in g
1 5 5
B y s o lv in g th e u n c o n s tra in e d m in -m a x p ro b le m , th e q u e s tio n fo r th e m a in c o n flic ts
*
in re q u ire m e n ts s a tis fa c tio n c a n b e a n s w e re d : If T = T is a m in im iz e r, th e n
m a x { TL ( 7 * ) } = TF* 1 = TF* 2 = α
*
> 0 ,
(9 )
w h i c h m e a n s t h a t t h e v a l u e s o f q u a l i t y f u n c t i o n s TF1 , TF 2 , . . . b e l o n g i n g t o t h e m o s t
c o m p e tin g re q u ire m e n ts a re e q u a l a n d th a t th e y h a v e th e la rg e s t v a lu e a m o n g a ll
q u a lity fu n c tio n s . M o re o v e r, α ≤ 1 c h a ra c te riz e s a fe a s ib le (‘a c c e p ta b le ’) s o lu tio n ,
w h ic h g iv e s ro o m fo r c o m p ro m is in g th e m o s t c o m p e tin g re q u ire m e n ts w ith in th e s e t
o f p a re to -o p tim a l a lte rn a tiv e s .
Ite r a tiv e C o m p r o m is in g S te p s : S ta rtin g w ith a p a re to -o p tim a l s o lu tio n , s a tis fa c tio n
o f c o m p e tin g re q u ire m e n ts c a n n o t b e im p ro v e d s im u lta n e o u s ly . T h is m e a n s th a t
l o w e r i n g t h e v a l u e o f o n e q u a l i t y f u n c t i o n , q F 7 , c a n b e a c h i e v e d o n l y a t t h e
e x p e n s e o f a h i g h e r v a l u e o f a n o t h e r , q F 7 , a n d v i c e v e r s a . D i f f e r e n t c o m p r o m i s e
s o l u t i o n s q F v e r s u s q F c a n b e a c h i e v e d b y d i f f e r e n t c h o i c e s o f t h e d e s i g n d i r e c t o r s d
in a n ite ra tiv e p ro c e d u re :
W i t h a g i v e n p a r e t o - o p t i m a l s o l u t i o n 4 ( ν − 1 ) = { TL( ν − 1 ) , TF( ν − 1 ) } , f o r t h e n e x t
i t e r a t i o n s t e p d e c i d e w h i c h o f t h e m o s t c o n f l i c t i n g q u a l i t y f u n c t i o n s { q F q F` s h a l l b e
i m p r o v e d , s a y q F. T h i s c h o i c e m a y b e m a d e d e p e n d e n t o n t h e w e i g h t s t h a t a r e
a s s o c ia te d to th e s e q u a lity fu n c tio n s v ia th e fo rm a lis m o f S e c tio n 3 . T h e n ,
c o n c a te n a te
N ( ν ) ∈ { L, F1 } , a n d c h o o s e G N = 1 a n d G F 2 s u c h t h a t TF( ν 2
− 1 )
< G νF 2
(≤ 1 ). S o lv in g th e c o n s tr a in e d m in -m a x p ro b le m ,
m i n m a x { TN( Y ) ( 7 ) } ,
7
N
V. W. TF 2 ( 7 ) ≤ G F( 2Y ) ,
(1 0 )
th e
a re
to
p ro
re q
n
n a tta in s th e b e s t p o s s ib le s o lu tio n in th e s e n s e th a t a ll q u a lity fu n c tio n s o f in te re s t
m in im iz e d u p to th e c o n s tra in t o f th e q u a n tifie d lim it o f d e g ra d a tio n o n e d e c la re s
b e a c c e p ta b le fo r th e m a in c o n flic tin g q u a lity fu n c tio n . T h u s in a n ite ra tiv e
c e d u re o n e c a n s e a rc h fo r a ‘b e s t’ c o m p ro m is e s a tis fa c tio n o f c o m p e tin g
u ire m e n ts .
Ite ra tiv e c o m p ro m is in g is b e s t c a rrie d o u t in a n in te r a c tiv e m o d e o f w o r k in g w h ic h
e e d s fa s t a lg o rith m s to e x e c u te th e C A C S D c o m p u ta tio n lo o p o f F ig u re 1 . In
d d itio n it n e e d s v is u a l d e c is io n s u p p o rt o n v a rio u s in fo rm a tio n le v e ls to b e s t g ra s p
e s ig n p ro b le m c o m p le x ity . In p a rtic u la r, a g ra p h ic a l u s e r in te rfa c e [1 3 ] w ith a
a ra lle l c o o rd in a te s d is p la y o f th e m a n y q u a lity fu n c tio n s , u s e d a s in te ra c tiv e s te e rin g
id to d e te c t c o m p ro m is e c o n flic ts a n d to c h o o s e d e s ig n d ire c to rs a t ru n tim e , g re a tly
n h a n c e s e n g in e e rin g p ro d u c tiv ity . A N D E C S _ M O P S is s u c h a n e n v iro n m e n t [1 4 ].
a
d
p
a
e
5
P a r e to -O p tim a l M u lti-m o d e l R o b u s tn e s s T u n in g
F e e d b a c k c o n tro l s u ffe rs fro m p o te n tia l s ta b ility p ro b le m s , b u t if p ro p e rly d e s ig n e d
fe e d b a c k re d u c e s p a ra m e te r s e n s itiv ity . T h e re fo re d e s ig n o f c o n tro lle rs , w h ic h a re
s ta b ility a n d p e rfo rm a n c e ro b u s t w ith re s p e c t to o ff-n o m in a l o p e ra tio n , is o f p rim e
1 5 6
D . K o lo s s a a n d G . G rü b e l
c o n c e rn . A n a ly tic a l ro b u s t-c o n tro l th e o ry , lik e µ -s y n th e s is , re lie s o n a n a ly tic a l
s ta b ility c rite ria a n d p e rtin e n t (lin e a r) p la n t m o d e l a n d u n c e rta in ty d e s c rip tio n s . T h u s ,
it is re s tric te d to p ro b le m s w ith s p e c ific , c o m m e n s u ra b le , p e rfo rm a n c e m e a s u re s .
A c o m p le te ly g e n e ra l a p p ro a c h to ro b u s t c o n tro l d e s ig n is s o -c a lle d m u lti-m o d e l
d e s ig n a s im p lie d b y th e C A C S D tu n in g lo o p , c f. F ig u re 1 . It is a p p lic a b le to a n y k in d
o f (n o n -lin e a r) p la n t m o d e ls a n d n o n -c o m m e n s u ra b le p e rfo rm a n c e m e a s u re s s in c e it
re lie s o n ly o n th e d a ta o f th e p e rfo rm a n c e m e a s u re s a n d n o t o n th e ir a n a ly tic a l
d e s c rip tio n . S tru c tu ra l in d e p e n d e n c e m a k e s th is k in d o f ro b u s t c o n tro l a p p ro a c h
a p p lic a b le to a n y ty p e o f c o n tr o lle r , i.e ., P I D , o b s e r v e r f e e d b a c k , f u z z y c o n tr o l, e tc .
T h e id e a o f ro b u s t m u lti-m o d e l d e s ig n is to s ta te th e d e s ig n p ro b le m fo r a n o m in a l
p la n t m o d e l in s ta n tia tio n re fle c tin g n o m in a l o p e ra tio n c o n d itio n s a n d n o m in a l s y s te m
p a ra m e te rs w ith in p e rtin e n t to le ra n c e b a n d s . T h e n , th e s a m e p ro b le m is s ta te d fo r a
n u m b e r o f o f f - n o m in a l m o d e l in s ta n tia tio n s r e f le c tin g w o r s t c a s e p la n t b e h a v io r , e .g .,
fa s t, lig h tly d a m p e d , a n d s lo w , o v e r-d a m p e d , b e h a v io r w ith in th e ra n g e o f a s s u m e d
o p e ra tio n c o n d itio n s a n d p a ra m e te r u n c e rta in ty in te rv a ls . T h e q u a lity fu n c tio n s o f a ll
th e s e fo rm u la tio n s a re c o n c a te n a te d to a s in g le m u ltio b je c tiv e p ro b le m fo r w h ic h a
s a tis fy in g p a re to -o p tim a l s o lu tio n is to b e fo u n d . T h is a p p ro a c h is h ig h ly c o m p e titiv e
in c o m p a ris o n to o th e r (a n a ly tic ) ro b u s t c o n tro l a p p ro a c h e s [1 5 ].
T h e s o lu tio n a p p ro a c h e s th a t a re d e a lt w ith in S e c tio n s 2 , 3 , 4 , a re p a rtic u la rly
s u ita b le to b e c o m b in e d fo r th is ty p e o f m u lti-m o d e l, m u lti-o b je c tiv e , ro b u s t c o n tro l
d e s ig n in fo rm o f a tw o -p h a s e d e s ig n p r o c e d u r e :
In p h a s e 1 , o n ly a n o m in a l m o d e l in s ta n tia tio n is c o n s id e re d a n d in te rv a l q u a lity
fu n c tio n s fo rm u la te d a c c o rd in g to S e c tio n 2 a re o p tim iz e d b y th e m u ltio b je c tiv e
g e n e tic a lg o rith m o f S e c tio n 3 .1 to y ie ld a ric h s e t o f g lo b a l, p a re to -o p tim a l, d e s ig n
a lte r n a tiv e s a s b a s is f o r d e s ig n e r p r e f e r e n c e s e le c tio n , S e c tio n 3 .2 .
In p h a s e 2 , th e ‘b e s t’ n o m in a l p e rfo rm a n c e a c h ie v e d in p h a s e 1 is e m b e d d e d in
‘g o o d ’ in te rv a ls b y re -s c a lin g th e q u a lity le v e ls a s re q u ire d . T h is s o lu tio n is u s e d to
s ta rt fu rth e r tu n in g u n d e r th e a s p e c t o f ro b u s tn e s s : T h e o ff-n o m in a l d e s ig n c a s e s a re
a d d e d to th e n o m in a l c a s e a n d in te ra c tiv e n o n lin e a r p ro g ra m m in g ite ra tio n s a c c o rd in g
to S e c tio n 4 a re s im u lta n e o u s ly a p p lie d to a ll d e s ig n c a s e s to ro b u s tify th e re s u lt o f
th e firs t, n o m in a l, d e s ig n p h a s e .
T h u s , th e u s e r is a llo w e d to m a k e q u a n tita tiv e tra d e -o ff d e c is io n s c o n c e rn in g
n o m in a l v e rs u s ro b u s t p e rfo rm a n c e . In th e s e d e c is io n s o ff-n o m in a l c o n tro l b e h a v io r,
c h a ra c te riz e d b y w o rs t-c a s e p la n t m o d e l in s ta n tia tio n s , s h o u ld b e c o m e a t le a s t
‘a c c e p ta b le ’, w h ile n o m in a l b e h a v io r is to b e k e p t w ith in a ‘g o o d ’ q u a lity le v e l.
6
C o n c lu s io n
A p a ra m e te r tu n in g m e th o d o lo g y to s u p p o rt c o n tro l d e s ig n a u to m a tio n is d e s c rib e d . It
u s e s a m u ltio b je c tiv e g e n e tic a lg o rith m w ith fitn e s s s h a rin g to fin d a ric h s e t o f g lo b a l
s o lu tio n s e v e n ly d is p e rs e d in o r n e a r to th e p a re to -o p tim a l s e t, fro m w h ic h a d e s ig n
c a n d id a te fo r b e s t d e s ig n e r s a tis fa c tio n is fo rm a lly s e le c te d v ia p a ir-w is e p re fe re n c e
s ta te m e n ts . T h e n m in -m a x n o n lin e a r p ro g ra m m in g is a p p lie d fo r c o m p ro m is e tu n in g
to a tta in a p a re to -o p tim a l s o lu tio n w ith b e s t tra d e o ffs in re q u ire m e n ts s a tis fa c tio n . A n
o n -lin e in te ra c tiv e m o d e o f w o rk in g u s in g n o n lin e a r p ro g ra m m in g in th e
c o m p ro m is in g p h a s e is s u p p o rte d b y a s y s te m a tic s fo r c h o o s in g ‘d e s ig n d ire c to rs ’ a s
E v o lu tio n a ry C o m p u ta tio n a n d N o n lin e a r P ro g ra m m in g
a llo
T o g
p h a
n o n
w a b le
e th e r
se a p p
-c o m m
u p p
w ith
ro a c
e n s
e r b o u n d s to
re q u ire m e n
h is w e ll s u
u ra b le p e rfo
lim it th e
ts c a p tu re
ite d to q u
rm a n c e a n
e x
b y
a n
d
p e n se o
fu z z y tita tiv e
ro b u s tn
n e
ty p
m u
e ss
is w
e in
ltire q
illin g to p
te rv a l q u a
m o d e l-ro b
u ire m e n ts
1 5 7
a y in m a k in g tra d e o ffs .
lity fu n c tio n s , th is tw o u s t c o n tro l d e s ig n w ith
.
R e fe r e n c e s
1 . G r ü b e l, G .: P e r s p e c tiv e s o f C A C S D : E m b e d d in g th e C o n tr o l S y s te m D e s ig n P r o c e s s in to a
V irtu a l E n g in e e rin g E n v iro n m e n t. P ro c . IE E E In t. S y m p o s iu m o n C o m p u te r A id e d C o n tro l
S y s te m D e s ig n , H a p u n a -B e a c h , H a w a ii (1 9 9 9 ) 2 9 7 -3 0 2
2 . F e n g , W ., L i, Y .: P e r f o r m a n c e I n d ic e s in E v o lu tio n a r y C A C S D A u to m a tio n w ith
A p p lic a tio n to B a tc h P ID G e n e ra tio n . P ro c . IE E E In t. S y m p o s iu m o n C o m p u te r A id e d
C o n tro l S y s te m D e s ig n , H a p u n a -B e a c h , H a w a ii (1 9 9 9 ) 4 8 6 -4 9 1
3 . C h ip p e r f ie ld , A .J ., D a k e v , N .V ., F le m in g , P .J ., W h id b o r n e , J .F .: M u ltio b je c tiv e R o b u s t
C o n tro l U s in g E v o lu tio n a ry A lg o rith m s . P ro c . IE E In t. C o n f. In d u s tria l T e c h n o lo g y (1 9 9 6 )
2 6 9 -2 7 3
4 . J o o s , H .- D ., S c h lo th a n e , M ., G r ü b e l, G .: M u lti- O b je c tiv e D e s ig n o f C o n tr o lle r s w ith F u z z y
L o g ic . P ro c . IE E E /IF A C J o in t S y m p o s iu m o n C o m p u te r A id e d C o n tro l S y s te m D e s ig n ,
T u c so n , A Z (1 9 9 4 ) 7 5 -8 2
5 . Z a k ia n , V ., A l-N a ib , U .: D e s ig n o f D y n a m ic a l a n d C o n tr o l S y s te m s b y th e M e th o d o f
In e q u a litie s . P ro c . In s titu te o f E le c tric a l E n g in e e rs , V o l. 1 2 0 , N o . 1 1 . (1 9 7 3 ) 1 4 2 1 -1 4 2 7
6 . T a n , K .C ., L e e , T .H ., K h o r , E .F .: C o n tr o l S y s te m D e s ig n A u to m a tio n w ith R o b u s t T r a c k in g
T h u m b p rin t P e rfo rm a n c e U s in g a M u lti-O b je c tiv e E v o lu tio n a ry A lg o rith m . P ro c . IE E E
S y m p o s iu m o n C o m p u te r-A id e d C o n tro l S y s te m D e s ig n , H a p u n a -B e a c h , H a w a ii (1 9 9 9 )
4 9 8 -5 0 3
7 . K ie n itz , K .H .: C o n tro lle r D e s ig n U s in g F u z z y L o g ic – A C a s e S tu d y . A u to m a tic a , V o l. 2 9 ,
N o . 2 . (1 9 9 3 ) 5 4 9 -5 5 4
8 . F o n s e c a , C .M ., F le m in g , P .J .: A n O v e r v ie w o f E v o lu tio n a r y A lg o r ith m s in M u ltio b je c tiv e
O p tim iz a tio n . E v o lu tio n a ry C o m p u tin g , V o l. 3 , N o . 1 . (1 9 9 5 ) 1 -1 6
9 . P o h lh e im , H .: G e n e tic a n d E v o lu tio n a ry A lg o rith m T o o lb o x fo r U s e w ith M a tla b D o c u m e n ta tio n . T e c h n ic a l R e p o rt, T e c h n ic a l U n iv e rs ity Ilm e n a u , (1 9 9 6 )
1 0 . d 'A m b r o s i o , J . G . , B i r m i n g h a m , W . P . : P r e f e r e n c e - d i r e c t e d D e s i g n . A r t i f i c i a l I n t e l l i g e n c e f o r
E n g in e e rin g D e s ig n , A n a ly s is a n d M a n u fa c tu rin g , V o l. 9 . (1 9 9 5 ) 2 1 9 -2 3 0
1 1 . M ie ttin e n , K .M .: N o n lin e a r M u ltio b je c tiv e O p tim iz a tio n , K lu w e r A c a d e m ic P u b lis h e r s ,
(1 9 9 8 )
1 2 . K r e is s e lm e ie r , G ., S te in h a u s e r , R .: A p p lic a tio n o f V e c to r P e r f o r m a n c e O p tim iz a tio n to
R o b u s t C o n tro l L o o p D e s ig n o f a F ig h te r A irc ra ft. In t. J o u rn a l C o n tro l, V o l. 3 7 , N o . 2 .
(1 9 8 3 ) 2 5 1 -2 8 4 .
1 3 . F in s te r w a ld e r , R ., J o o s , H .- D ., V a r g a , A .: A G r a p h ic a l U s e r I n te r f a c e f o r F lig h t C o n tr o l
D e v e lo p m e n t. P ro c . IE E E S y m p o s iu m o n C o m p u te r-A id e d C o n tro l S y s te m D e s ig n ,
H a p u n a -B e a c h , H a w a ii (1 9 9 9 ) 4 3 9 -4 4 4
1 4 . G r ü b e l, G ., F in s te r w a ld e r , R ., G r a m lic h , G ., J o o s , H .- D ., L e w a ld , S .: A N D E C S : A
C o m p u ta tio n E n v iro n m e n t fo r C o n tro l A p p lic a tio n s o f O p tim iz a tio n . In : C o n tro l
A p p lic a tio n s o f O p tim iz a tio n , R . B u lir s c h , D . K r a f t, e d s ., I n t. S e r ie s o f N u m e r ic a l
M a th e m a tic s , V o l. 1 1 5 . B irk h ä u s e r V e rla g , B a s e l (1 9 9 4 ) 2 3 7 -2 5 4
1 5 . G r ü b e l, G .: A n o th e r V ie w o n th e D e s ig n C h a lle n g e A c h ie v e m e n ts . I n : R o b u s t F lig h t
C o n tr o l – A D e s ig n C h a lle n g e , M a g n i, J .F ., B e n n a n i, S ., T e r lo u w , J .C ., E d s ., L e c tu r e N o te s
in C o n tro l a n d In fo rm a tio n S c ie n c e s 2 2 4 , S p rin g e r V e rla g , B e rlin H e id e lb e rg N e w Y o rk
(1 9 9 7 ) 6 0 3 -6 0 9
Benchmarking Cost-Assignment Schemes for
Multi-objective Evolutionary Algorithms
Konstantinos Koukoulakis, Dr Yun Li
Department of Electronics and Electrical Engineering, University of Glasgow
Abstract.
Currently there exist various cost-assignment schemes that perform the necessary
scalarization of the objective values when applied to a multi-objective optimization
problem. Of course, the final decision depends highly on the nature of the problem
but given the multiplicity of the schemes combined with the fact that what the user
ultimately needs is a single compromise solution it is evident that elaborating the
selection of the method is not a trivial task. This paper intends to address this problem
by extending the benchmarks of optimality and reach time given in [1] to
mutliobjective optimization problems. A number of existing cost-assignment schemes
are evaluated using such benchmarks.
1. Introduction
Having in mind the number of existing approaches to cost-assignment one could
presume that the next step would be an appropriate choice. The concept of Pareto
dominance has proven to be a great aid towards the formulation of the various
schemes but further thinking reveals that what the user would like to have is simply a
single compromise solution and not all of the solutions that form the Pareto-optimal
set. As stated in [2], “although a Pareto-optimal solution should always be a better
compromise solution than any solution it dominates, not all Pareto-optimal solutions
may constitute acceptable compromise solutions”. Therefore, what is needed is a
performance index that can be used for the evaluation of the suitability of each
scheme in the context of a specific problem.
To address this issue, two benchmarks used presented in Section 2. Section 3
describes a benchmark problem used for the comparison. Section 4 presents the
evolutionary algorithm employed along with an outline of the various cost-assignment
methods. Comparison is made between a number of evolutionary algorithms in
Section 5. Conclusions are drawn in section 6.
2. The Benchmarks
Two benchmarks used for the evaluation of the different approaches have been
defined in [1]. In this section they are going to be briefly presented.
S. Cagnoni et al. (Eds.): EvoWorkshops 2000, LNCS 1803, pp. 158-167, 2000.
Springer-Verlag Berlin Heidelberg 2000
Benchmarking Cost-Assignment Schemes
159
The first benchmark is called ‘optimality’. Suppose that we have a test function
f(x): X F, where X Rn, F Rm,
where R
⊆n represents the search space⊆in n dimensions,
m
⊆
R represents the space of all the possible objective values,
n is the number of parameters,
m is the number of the objectives
f ∈ F is the collection of the individual objective elements
Also, consider the theoretical objective vector fo = { f(x) } that contains the
objective values that can ultimately be reached. Finally, consider an objective reached
as in Eq. 1.
⎛ ^ ⎞ ^ ^
f ⎜ xo ⎟ = f 0 , xo ∈ X
⎝ ⎠
(1)
with ^xo, representing a corresponding solution found.
The optimality measures how close an objective reached is to the theoretical
objective vector and is calculated using the formula in Eq.2.
^
⎛ ⎞
Optimality⎜ f o ⎟ = 1 −
⎝ ⎠α
^
fo − f o
α
_
∈ [0,1]
f− f
(2)
_ α
-
where f and f- are the upper and lower bounds of f respectively.
Any norm can be used to evaluate the optimality of an objective and this paper
uses the Euclidean metric (a=2) for this purpose.
The above formula needs to be refined when the problem addressed is a nondominant or non-commensurate one since no such concept as ‘overall optimality’ can
be assessed in a problem of this kind. Since this is the case for this paper, the
‘distance to demands’ method explained in [3], is used here.
The second benchmark this short study uses is one that measures the convergence
of the algorithm and is called ‘reach time’. The reach time is defined as the total
number of function evaluations performed by the algorithm by which the optimality
of the best individual first reaches b.
Re ach _ time b = C b
(3)
For the purposes of the tests, b is set to 0.999, a certainly high value that may not
always be reached by the algorithm. Because of that, a single algorithm terminates
when either the set optimality threshold is reached or 20n generations of size 20nxm
have been evolved. Those termination conditions are identical to the ones used in [1]
with the latter one meaning that the algorithm is not supposed to perform worse than
an O(n2) algorithm in terms of computational time.
160
K. Koukoulakis and Y. Li
3. The Problem
A set of two objective functions (Fonsceca and Fleming, 1995) was chosen for the
evaluation of the cost-schemes. The functions of Eq.4 and Eq.5 were chosen in an
effort to produce as “standard” a Pareto optimal front as possible.
2
⎛ n ⎛
1 ⎞ ⎞⎟
f1 ( x) = 1 − exp⎜ − ∑ ⎜ xi −
⎟
⎜ i =1 ⎝
n ⎠ ⎟⎠ (4)
⎝
2
⎛ n ⎛
1 ⎞ ⎞⎟
⎜
f 2 (x ) = 1 − exp − ∑ ⎜ xi +
⎟
⎜ i =1 ⎝
n ⎠ ⎟⎠
⎝
(5)
Each individual consists of a real-valued vector of n parameters. For the purposes
of this paper n was set to the value of 2, with each parameter coded in the interval [-2,
2).
The individuals that form the Pareto-optimal set belong on the line shown in Fig. 1.
Functions f1 and f2 are plotted for n=2 in figures 2 and 3 respectively.
Fig. 1. Pareto optimal front
Fig. 2. f1 plotted for n=2
Fig. 3. f2 plotted for n=2
Benchmarking Cost-Assignment Schemes
161
4.The Algorithm
4.1 Selection, Crossover and Mutation
The evolutionary algorithm that was used is quite a simple and straightforward
one. It uses a binary tournament selection scheme to form the mating pool of the
individuals after, of course, the cost assignment procedure has taken place. Each
individual in the mating pool then randomly mates with another one using arithmetic
crossover since each individual consists of a real-valued vector. Arithmetic crossover,
as described in [4], is a canonical intermediate recombination operator, which
produces the i-th component of the offspring by averaging it with some weight as
defined in Eq.6.
xi/ = αx1i + (1 − α )x2 i
(6)
Next, the offspring are evaluated and then refined by the simulated annealing (SA)
technique. The SA positive mutation cycles where conducted using a non-linear
Boltzmann learning schedule as the one employed in [5]. For the purposes of
mutation, the creep mutation operator (Davis, 1989) was employed. As suggested in
[6], entrapment must be alleviated in the case of this operator when used with a
bounded small random amount for mutation. As such is the case here, entrapment can
be said to have been partially overcome by the probabilistic nature of the SA
technique, which maintains a probability of retaining lesser-valued individuals.
After SA, the parents are merged with the offspring the new population is formed
with binary tournament selection.
The parameter settings of the algorithm, most of which are suggested in [5], are
listed in table 1. The optimality threshold was set to a high value as the tests were
intended to prohibit the algorithm to reach it so that a clearer picture of each scheme’s
behavior could be obtained.
Table 1. Parameter settings
Optimality threshold
Number of parameters
Weight vector
Priority vector
Goal vector
Number of generations
Population size
Arithmetic crossover
constant
Creep mutation
probability
Tournament size
Boltzmann constant
Initial temperature
Final temperature
Initial annealing factor
Transient constant
0.999
2
{ 1, 1 }
{ 1, 0=top }
{ 0.0, 0.0 }
160
320
0.5
0.05
2
5E-06
1E05
1
0.3
16
162
K. Koukoulakis and Y. Li
4.2 The Cost-Assignment Schemes
Each scheme is successively described below with a minimization problem
assumed.
4.2.1 The Weighted Sum Method
According to this approach, all of the objectives are weighted by positive
coefficients defined by the user and are added together to obtain the cost.
Φ : ℜn → ℜ
m
(7)
f ( x ) = ∑ wk f k (x )
k =1
,where Ö denotes the cost assignment scheme and x is the parameter vector.
It must be noted that the same weights that are used here are also used to weight
the objective vectors prior to the calculation of the norms of the optimality benchmark
measure as suggested in [1].
4.2.2 The Minimax Method
This method tries to minimize the maximum weighted difference between the
objectives and the goals, with the weights and the goals supplied by the user.
Φ : ℜn → ℜ
f (x ) − g k
f ( x ) = max k
k =1Km
wk
(8)
4.2.3 The Target Vector Method
This approach minimizes the distance of the objective vector from the goal vector
using a defined distance measure. Again the user supplies the goals. The Euclidean
metric was used as the distance measure in this case.
Φ : ℜn → ℜ
f ( x ) = [ f ( x ) − g ]W −1
(9)
α
4.2.4 The Lexicographic Method
Here, the objectives are assigned distinct priorities and the selection proceeds with
the comparison of the individuals with respect to the objective of the highest priority
Benchmarking Cost-Assignment Schemes
163
with any ties resolved by a successive comparison with respect to the objective with
the second-highest priority, until the lowest priority objective is reached.
Φ : ℜn → {0,1, K µ − 1}, µ = pop _ size
f ( xi ) = ∑ l ( f (x j )l < f ( xi )), µ = pop _ size
µ
(10)
j =1
where l(condition) evaluates to unity if condition is true and
f (x j )l < f ( x i ) ⇔ ∃ p ∈ {1K m } :
∀k ∈{p,L, m}, f k (x j ) ≤ f k (xi ) ∧ f p (x j ) < f p (xi )
(11)
4.2.5 Pareto Ranking (Goldberg’s Approach)
According to the definition in [6] all non-dominated individuals are assigned a cost
of one and then removed from contention with the next set of non-dominated
individuals assigned a cost of two until the whole population has been ranked.
Φ : ℜ n → {1, K µ }, µ = pop _ size
1 ⇐ not ( f (x j ) p < f ( xi )), ∀j ∈ {1K µ }
f (xi ) = {
φ ⇐ not ( f (x j ) p < f ( xi )), ∀ j ∈ {1K µ }
(12)
\ {l : Φ( f ( xl )) < φ }
,where the p< condition denotes partial domination of the individual j over the
individual i and is true if and only if
∀k ∈ {1K m}
f k (x j ) ≤ f k ( xi ) ∧ ∃k ∈{1K m} : f k (x j ) < f k ( xi )
(13)
4.2.6 Pareto Ranking (Fonseca and Fleming’s Approach)
Proposed in 1993, this approach ranks an individual according to the number of
individuals that dominate him.
Φ : ℜ n → {0,1, K µ − 1}, µ = pop _ size
µ
⎞
⎛
f (xi ) = ∑l ⎜ f (x j ) p < f (xi )⎟, µ = pop _ size
⎠
j =1 ⎝
(14)
164
K. Koukoulakis and Y. Li
4.2.7 Pareto Ranking (With Goals and Priorities)
This approach combines the pareto-optimality concept with goal and priority
information. Equal priorities may be assigned to different objectives with both the
priorities and the goals supplied by the user. Individuals are compared as in the
lexicographic method but it is also affected from whether the individuals attain the
goals set or not.
Φ : ℜ n → {0,1, K µ − 1}, µ = pop _ size
f ( xi ) = ∑ l ⎛⎜ f (x j ) p f ( xi ) ⎞⎟, µ = pop _ size
g
⎠
j =1 ⎝
µ
(15)
,where
the condition within the brackets denotes preferrability of the j-th individual over
the i-th individual and
g is the preference vector, a vector that contains the goals of each objective
grouped by priority
As for the evaluation of the condition it is deemed too detailed to mention here but
is fully described in [2].
5. Comparison Results
For each method, 10 experiments were carried out each with a random initial
population, with an experiment terminating either when the optimality threshold has
been reached or 400mn2 generations have been evolved. A discussion of the results
obtained follows.
5.1 Pareto Front Sampling and Diversity
A cost-assignment scheme is considered successful if it has managed to offer a
diverse sample of the pareto-optimal front as quickly as possible. With this in mind, a
short discussion for each scheme tested follows.
A snapshot of the population in which the most optimal individual was found for
each scheme can be seen at figures 4-10. Remember that the goals were set to 0.0 for
both objectives.
The weighted sum approach was unable to sample the concave region of the line,
focusing entirely on the zero-cost line f2 = -f1.
With identical, equally weighted goals, the minimax scheme failed to sample the
pareto front. Nevertheless, with appropriate goal and weight settings it can prove
successful, but surely less successful than the pareto-based approach with goals and
priorities, which provides a better sampling in a quick and more efficient manner
using only the goal information available.
The target vector scheme has introduced better sampling diversity than the
minimax approach in roughly the same time. The diversity is even better than the
pareto-based approach with goals and priorities but the latter scheme is significantly
quicker in providing its results.
Benchmarking Cost-Assignment Schemes
165
Fonseca and Fleming’s approach along with Goldberg’s original one has indeed
most quickly sampled a very good proportion of the pareto-optimal front with the
former being better at that. It must be noted that they have performed better at that
than the last scheme without using any information available. This leads us to the
conclusion that in the case of unattainable goals both of these schemes can offer a
better sample of the front than the last scheme in some applications.
The last approach has performed very well using both the goal and priority
information. It is interesting to compare it with the lexicographic cost-assignment
scheme, which only uses priorities. It can clearly be seen that the latter scheme has
driven the population to the minimisation of f2, which has a higher priority over f1.
So, it can be said that the lexicographic method needs an aid for better results and the
most obvious one is niching combined with mating restriction.
5.2 Optimality
As far as optimality is concerned, as can be seen in Fig. 11, Goldberg’s and
Fonseca and Fleming’s approaches have both quickly given optimal solutions without
using the goal information and also have the added bonus of good diversity. Of course
with attainable goals, the pareto-based approach with goals and priorities should be
the quickest cost-assignment scheme to offer the most optimal solution.
5.3 Reach Time
As no method managed to reach the high optimality threshold of 0.999 (max =
1.0), all of them had a reach time of 400mn2 = 51200.
6. Conclusions
The cost-assignment scheme acts as the driving force of the algorithm. Performing
the scalarization of the objectives, it is the determining factor of evaluation. The
purpose of this paper was to expose the magnitude of its impact on (a) the quality of
the sampling of the Pareto-optimal front and (b) on the speed at which this quality is
achieved.
The goals were deliberately set unattainable because they were intended to be a
means to push the population towards the front. The optimality benchmark itself
worked in accord with the goal settings so that it could act as an observer of the
algorithm’s behaviour rather than a strict evaluator.
It is thus concluded that for a given problem, the sampling of the Pareto-optimal
front is generally easier achieved with the Pareto-based cost schemes. As for the rest
of the schemes, it is thought that their usefulness can only be experienced with proper
tuning of their associated parameters.
As a further study, it would be interesting to test all of the schemes in the context
of a harder problem, that is a problem with a more diverse front. Finally, it is believed
that this testing should employ a wide range of weights so that the promising
aggregating ‘target vector’ scheme can be examined more closely.
166
K. Koukoulakis and Y. Li
Fig. 4. Pareto ranking (Fonseca and Fleming),
Gen. 46
Fig. 5. Pareto ranking (Goldberg), Gen.41
Fig. 6. Lexicographic, Gen. 158
Fig. 7. Target vector, Gen. 156
Fig. 8. Minimax, Gen. 145
Fig. 9. Weighted sum, Gen. 67
Benchmarking Cost-Assignment Schemes
Fig. 10. Pareto ranking
(goals and priorities), Gen.
47
167
Fig. 11. Maximum optimality
References
1.Benchmarks
for testing evolutionary algorithms, The Third Asia-Pacific Conference on
Measurement and Control, Dunhuang, China, 31 Aug. - 4 Sept 1998, 134-138. (W. Feng , T.
Brune, L. Chan, M. Chowdhury, C.K. Kuek and Y. Li).
2.Back T., Fogel D. B. and Michalewicz Z., Handbook of Evolutionary Computation (New
York, Oxford: Oxford University Press, Bristol, Philadelphia: Institute Of Physics
Publishing, 1997)
3.Michalewicz Z., Genetic Algorithms + Data structures = Evolution programs. (Berlin:
Springer-Verlag, 1992)
4.Michalewicz Z., Nazhiyath G. and Michalewicz M, A note on the usefulness of geometrical
crossover for numerical optimization problems, Proc 5th Ann. Conf. on Evolutionary
Programming ed L. J. Fogel, P. J. Angeline and T. Back (Cambridge, MA: MIT Press, 1996)
5.Tan K.C., Evolutionary methods for Modelling and Control of Linear and Nonlinear
Systems, Ph.D. thesis (Department of Electronics and Electrical Engineering, University of
Glasgow, 1997)
6.Goldberg D. E., Genetic algorithms in Search, Optimization and Machine Learning
(Reading, MA: Addison-Wesley, 1989)
7.Davis L., Adapting operator probabilities in genetic algorithms, Proc 3rd Int. Conf. on GAs
(Fairfax, VA, June 1989) ed J. D. Schaffer (San Mateo, CA: Morgan Kaufmann) pp 61-69
8.Fonseca C. M. and Fleming P. J., Multiobjective genetic algorithms made easy: selection
sharing and mating restriction (First Int. Conf. on GAs in Eng. Systems: Innovations and
Applications, Sheffield, UK, 1995) pp 45-52
A u to m a tic S y n th e s is o f B o th th e T o p o lo g y a n d
P a r a m e te r s fo r a C o n tr o lle r fo r a T h r e e -L a g P la n t w ith a
F i v e - S e c o n d D e l a y Us i n g G e n e t i c P r o g r a m m i n g
J o h n R . K o z a
S ta n fo rd U n iv e rs ity , S ta n fo rd , C a lifo rn ia
k o z a @ s t a n f o r d . e d u
M a r tin A . K e a n e
E c o n o m e tr ic s I n c ., C h ic a g o , I llin o is
m a k e a n e @ i x . n e t c o m . c o m
J e sse n Y u
G e n e tic P r o g r a m m in g I n c ., L o s A lto s , C a lif o r n ia
j y u @ c s . s t a n f o r d . e d u
W illia m
M y d lo w e c
G e n e tic P r o g r a m m in g I n c ., L o s A lto s , C a lif o r n ia
m y d @ c s . s t a n f o r d . e d u
F o r r e st H
B e n n e tt III
G e n e tic P ro g ra m m in g In c .
(C u rre n tly , F X P a lo A lto L a b o ra to ry , P a lo A lto , C a lifo rn ia )
f o r r e s t @ e v o l u t e . c o m
A b str a c t
T h is p a p e r d e s c rib e s h o w th e p ro c e s s o f s y n th e s iz in g th e d e s ig n o f b o th th e
to p o lo g y a n d th e n u m e ric a l p a ra m e te r v a lu e s (tu n in g ) fo r a c o n tro lle r c a n
b e a u to m a te d b y u s in g g e n e tic p ro g ra m m in g . G e n e tic p ro g ra m m in g c a n b e
u s e d to a u to m a tic a lly m a k e th e d e c is io n s c o n c e rn in g th e to ta l n u m b e r o f
s ig n a l p ro c e s s in g b lo c k s to b e e m p lo y e d in a c o n tro lle r, th e ty p e o f e a c h
b lo c k , th e to p o lo g ic a l in te rc o n n e c tio n s b e tw e e n th e b lo c k s , a n d th e v a lu e s
o f a ll p a ra m e te rs fo r a ll b lo c k s re q u irin g p a ra m e te rs . In s y n th e s iz in g th e
d e s ig n o f c o n tro lle rs , g e n e tic p ro g ra m m in g c a n s im u lta n e o u s ly o p tim iz e
p re s p e c ifie d p e rfo rm a n c e m e tric s (s u c h a s m in im iz in g th e tim e re q u ire d to
b rin g th e p la n t o u tp u t to th e d e s ire d v a lu e ), s a tis fy tim e -d o m a in c o n s tra in ts
(s u c h a s o v e rs h o o t a n d d is tu rb a n c e re je c tio n ), a n d s a tis fy fre q u e n c y d o m a in
c o n s tra in ts . E v o lu tio n a ry m e th o d s h a v e th e a d v a n ta g e o f n o t b e in g
e n c u m b e re d b y p re c o n c e p tio n s th a t lim it its s e a rc h to w e ll-tra v e le d p a th s .
G e n e tic p ro g ra m m in g is a p p lie d to a n illu s tra tiv e p ro b le m in v o lv in g th e
d e s ig n o f a c o n tro lle r fo r a th re e -la g p la n t w ith a s ig n ific a n t (fiv e -s e c o n d )
tim e d e la y in th e e x te rn a l fe e d b a c k fro m th e p la n t to th e c o n tro lle r. A d e la y
in th e fe e d b a c k m a k e s th e d e s ig n o f a n e ffe c tiv e c o n tro lle r e s p e c ia lly
d iffic u lt.
S . C a g n o n i e t a l. ( E d s .) : E v o W o r k s h o p s 2 0 0 0 , L N C S 1 8 0 3 , p p . 1 6 8 − 1 7 7 , 2 0 0 0 .
S p rin g e r-V e rla g B e rlin H e id e lb e rg 2 0 0 0
A u to m a tic S y n th e s is o f B o th th e T o p o lo g y a n d P a ra m e te rs
1
1 6 9
I n tr o d u c tio n
T h e p ro c e s s o f c re a tin g (s y n th e s iz in g ) th e d e s ig n o f a c o n tro lle r e n ta ils m a k in g
d e c is io n s c o n c e rn in g th e to ta l n u m b e r o f p ro c e s s in g b lo c k s to b e e m p lo y e d in th e
c o n tr o lle r , th e ty p e o f e a c h s ig n a l p r o c e s s in g b lo c k ( e .g ., le a d , la g , g a in , in te g r a to r ,
d iffe re n tia to r, a d d e r, in v e rte r, s u b tra c to r, a n d m u ltip lie r), th e v a lu e s o f a ll p a ra m e te rs
fo r a ll b lo c k s re q u irin g p a ra m e te rs , a n d th e to p o lo g ic a l in te rc o n n e c tio n s b e tw e e n th e
s ig n a l p ro c e s s in g b lo c k s . T h e la tte r in c lu d e s th e q u e s tio n o f w h e th e r o r n o t to e m p lo y
in te r n a l f e e d b a c k ( i.e ., f e e d b a c k in s id e th e c o n tr o lle r ) .
T h e p ro b le m o f s y n th e s iz in g a c o n tro lle r to s a tis fy p re s p e c ifie d re q u ire m e n ts is
s o m e tim e s s o lv a b le b y a n a ly tic te c h n iq u e s (o fte n o rie n te d to w a rd p ro d u c in g
c o n v e n tio n a l P ID c o n tro lle rs ). H o w e v e r, a s B o y d a n d B a rra tt s ta te d in L in e a r
C o n tr o lle r D e s ig n : L im its o f P e r fo r m a n c e (1 9 9 1 ),
" T h e c h a lle n g e fo r c o n tro lle r d e s ig n is to p ro d u c tiv e ly u s e th e e n o rm o u s
c o m p u tin g p o w e r a v a ila b le . M a n y c u rre n t m e th o d s o f c o m p u te r-a id e d
c o n t r o l l e r d e s i g n s i m p l y a u t o m a t e p r o c e d u r e s d e v e l o p e d i n t h e 1 9 3 0 ’s
t h r o u g h t h e 1 9 5 0 's … "
T h is p a p e r d e s c rib e s h o w g e n e tic p ro g ra m m in g c a n b e u s e d to a u to m a tic a lly c re a te
b o th th e to p o lo g y a n d th e n u m e r ic a l p a r a m e te r v a lu e s ( i.e ., th e tu n in g ) f o r a c o n tr o lle r
d ire c tly fro m a h ig h -le v e l s ta te m e n t o f th e re q u ire m e n ts o f th e c o n tro lle r. G e n e tic
p ro g ra m m in g c a n , if d e s ire d , s im u lta n e o u s ly o p tim iz e p re s p e c ifie d p e rfo rm a n c e
m e tric s (s u c h a s m in im iz in g th e tim e re q u ire d to b rin g th e p la n t o u tp u t to th e d e s ire d
v a lu e a s m e a s u re d b y , s a y , th e in te g ra l o f th e tim e -w e ig h te d a b s o lu te e rro r), s a tis fy
tim e -d o m a in c o n s tra in ts (in v o lv in g , s a y , o v e rs h o o t a n d d is tu rb a n c e re je c tio n ), a n d
s a tis fy fre q u e n c y d o m a in c o n s tra in ts . E v o lu tio n a ry m e th o d s h a v e th e a d v a n ta g e o f
n o t b e in g e n c u m b e re d b y p re c o n c e p tio n s th a t lim it th e ir s e a rc h to w e ll-tra v e le d p a th s .
S e c tio n 2 d e s c rib e s a n illu s tra tiv e p ro b le m o f c o n tro lle r s y n th e s is . S e c tio n 3
p ro v id e s g e n e ra l b a c k g ro u n d o n g e n e tic p ro g ra m m in g . S e c tio n 4 d e s c rib e s h o w
g e n e tic p ro g ra m m in g is a p p lie d to c o n tro l p ro b le m s . S e c tio n 5 d e s c rib e s th e
p re p a ra to ry s te p s n e c e s s a ry to a p p ly g e n e tic p ro g ra m m in g to th e illu s tra tiv e c o n tro l
p ro b le m . S e c tio n 6 p re s e n ts th e re s u lts .
2
I llu s tr a tiv e P r o b le m
T h e illu s tra tiv e p ro b le m e n ta ils c re a tio n o f b o th th e to p o lo g y a n d p a ra m e te r v a lu e s fo r
a c o n tro lle r fo r a th re e -la g p la n t w ith a s ig n ific a n t (fiv e -s e c o n d ) tim e d e la y in th e
e x te rn a l fe e d b a c k fro m th e p la n t o u tp u t to th e c o n tro lle r s u c h th a t p la n t o u tp u t re a c h e s
th e le v e l o f th e re fe re n c e s ig n a l in m in im a l tim e (a s m e a s u re d b y th e in te g ra l o f th e
tim e -w e ig h te d a b s o lu te e rro r), s u c h th a t th e o v e rs h o o t in re s p o n s e to a s te p in p u t is
le s s th a n 2 % , a n d s u c h th a t th e c o n tro lle r is ro b u s t in th e fa c e o f d is tu rb a n c e (a d d e d
in to th e c o n tro lle r o u tp u t). T h e d e la y in th e fe e d b a c k m a k e s th e d e s ig n o f a n e ffe c tiv e
c o n tro lle r e s p e c ia lly d iffic u lt (A s tro m a n d H a g g lu n d 1 9 9 5 ). T h e tra n s fe r fu n c tio n o f
th e p la n t is
A
G (s ) =
K e
− 5 s
(1 + τ s )
3
c o n tro lle r p re s e n te d in A s tro m a n d H a g g lu n d 1 9 9 5 (p a g e 2 2 5 ) d e liv e rs c re d ib le
p e rfo rm a n c e o n th is p ro b le m fo r v a lu e s o f K = 1 a n d τ = 1 .
1 7 0
b y
lim
se v
b y
J .R . K o z a e t a l.
T o m a k e th e
th e c o n tro lle r
ite d to th e ra n
e ra l d iffe re n t
A s tro m a n d H
3
p ro
p re
g e
c o m
a g g
b le m
s e n te
b e tw
b in a
lu n d
m o
d in
e e n
tio n
w a s
re re a lis tic
A s tro m a n
-4 0 a n d + 4
s o f v a lu e s
in te n d e d o
, w e a d d e d
d H a g g lu n d
0 v o lts . T h
fo r K a n d
n ly fo r K =
a n a d d itio n a l
1 9 9 5 ) th a t th e
e p la n t in th is
τ (w h e re a s th e
1 a n d τ = 1 ).
c o n s tra in t
in p u t to th
p a p e r o p e r
c o n tro lle r
(sa
e p
a te
d e
tis fie d
la n t is
s o v e r
s ig n e d
B a c k g r o u n d o n G e n e tic P r o g r a m m in g
4
G e n e tic p ro g ra m m in g is a n a u to m a tic te c h n iq u e fo r g e n e ra tin g c o m p u te r p ro g ra m s to
s o lv e , o r a p p ro x im a te ly s o lv e , p ro b le m s .
G e n e tic p ro g ra m m in g (K o z a 1 9 9 2 ; K o z a a n d R ic e 1 9 9 2 ) is a n e x te n s io n o f th e
g e n e tic a lg o rith m (H o lla n d 1 9 7 5 ). G e n e tic p ro g ra m m in g is c a p a b le (K o z a 1 9 9 4 a ,
1 9 9 4 b ) o f e v o lv in g re u s a b le , p a ra m e triz e d , h ie ra rc h ic a lly -c a lle d a u to m a tic a lly d e fin e d
fu n c tio n s (A D F s ) s o th a t a n o v e ra ll p ro g ra m c o n s is ts o f a m a in re s u lt-p ro d u c in g
b ra n c h a n d o n e o r m o re re u s a b le a n d p a ra m e te riz a b le a u to m a tic a lly d e fin e d fu n c tio n s
(fu n c tio n -d e fin in g b ra n c h e s ). In a d d itio n , a rc h ite c tu re -a lte rin g o p e ra tio n s (K o z a ,
B e n n e tt, A n d re , a n d K e a n e 1 9 9 9 ; K o z a , B e n n e tt, A n d re , K e a n e , a n d B ra v e 1 9 9 9 )
e n a b le g e n e tic p ro g ra m m in g to a u to m a tic a lly d e te rm in e th e n u m b e r o f a u to m a tic a lly
d e fin e d fu n c tio n s , th e n u m b e r o f a rg u m e n ts th a t e a c h p o s s e s s e s , a n d th e n a tu re o f th e
h ie ra rc h ic a l re fe re n c e s , if a n y , a m o n g s u c h a u to m a tic a lly d e fin e d fu n c tio n s .
G e n e tic p ro g ra m m in g o fte n c re a te s n o v e l d e s ig n s b e c a u s e it is a p ro b a b ilis tic
p ro c e s s th a t is n o t e n c u m b e re d b y th e p re c o n c e p tio n s th a t o fte n c h a n n e l h u m a n
th in k in g d o w n fa m ilia r p a th s . F o r e x a m p le , g e n e tic p ro g ra m m in g is c a p a b le o f
s y n th e s iz in g th e d e s ig n o f b o th th e to p o lo g y a n d s iz in g fo r a w id e v a rie ty o f a n a lo g
e l e c t r i c a l c i r c u i t s f r o m a h i g h - l e v e l s t a t e m e n t o f t h e c i r c u i t ’s d e s i r e d b e h a v i o r a n d
c h a ra c te ris tic s (K o z a , B e n n e tt, A n d re , a n d K e a n e 1 9 9 9 ; K o z a , B e n n e tt, A n d re , K e a n e ,
a n d B ra v e 1 9 9 9 ). F iv e o f th e e v o lv e d a n a lo g c irc u its in th a t b o o k in frin g e o n
p re v io u s ly is s u e d p a te n ts w h ile fiv e o th e rs d e liv e r th e s a m e fu n c tio n a lity a s p re v io u s ly
p a te n te d in v e n tio n s in a n o v e l w a y .
A d d itio n a l in fo rm a tio n o n c u rre n t re s e a rc h in g e n e tic p ro g ra m m in g c a n b e fo u n d
in B a n z h a f, N o rd in , K e lle r, a n d F ra n c o n e 1 9 9 8 ; L a n g d o n 1 9 9 8 ; R y a n 1 9 9 9 ; K in n e a r
1 9 9 4 ; A n g e l i n e a n d K i n n e a r 1 9 9 6 ; S p e c t o r , L a n g d o n , O ’R e i l l y , a n d A n g e l i n e 1 9 9 9 ;
K o z a , G o ld b e rg , F o g e l, a n d R io lo 1 9 9 6 ; K o z a , D e b , D o rig o , F o g e l, G a rz o n , Ib a , a n d
R io lo 1 9 9 7 ; K o z a , B a n z h a f, C h e lla p illa , D e b , D o rig o , F o g e l, G a rz o n , G o ld b e rg , Ib a ,
a n d R io lo 1 9 9 8 ; B a n z h a f, P o li, S c h o e n a u e r, a n d F o g a rty 1 9 9 8 ; B a n z h a f, D a id a , E ib e n ,
G a rz o n , H o n a v a r, J a k ie la , a n d S m ith 1 9 9 9 ; P o li, N o rd in , L a n g d o n , a n d F o g a rty 1 9 9 9 ;
a t w e b s ite s s u c h a s w w w . g e n e t i c - p r o g r a m m i n g . o r g ; a n d in th e G e n e tic
P r o g r a m m in g a n d E v o lv a b le M a c h in e s jo u rn a l (fro m K lu w e r A c a d e m ic P u b lis h e rs ).
G e n e tic P r o g r a m m in g a n d C o n tr o l
C o n tro lle rs
p ro c e s s in g
o u tp u t(s ), a
in s id e th e c
b o th th e to
b e tw e e n th
g e rm a n e to
c a n b e re p re s e n te d b y
fu n c tio n s , in w h ic h e x
n d in w h ic h c y c le s in
o n tro lle r. G e n e tic p ro g
p o lo g y a n d p a ra m e te r
e p ro g ra m tre e s u s e d
c o n tro lle rs .
b lo c k d ia g ra m
te rn a l p o in ts
th e b lo c k d ia
ra m m in g c a n
v a lu e s fo r a
in g e n e tic p
s in w h ic h th e b lo c k s re p re
r e p r e s e n t t h e c o n t r o l l e r ’s i n
g ra m c o rre s p o n d to in te rn a
b e e x te n d e d to th e p ro b le m
c o n tro lle r b y e s ta b lis h in g
ro g ra m m in g a n d th e b lo c k
s e n t s ig n a l
p u t(s ) a n d
l fe e d b a c k
o f c re a tin g
a m a p p in g
d ia g ra m s
A u to m a tic S y n th e s is o f B o th th e T o p o lo g y a n d P a ra m e te rs
1 7 1
T h e n u m b e r o f re s u lt-p ro d u c in g b ra n c h e s in th e to -b e -e v o lv e d c o n tro lle r e q u a ls th e
n u m b e r o f c o n tro l v a ria b le s th a t a re to b e p a s s e d fro m th e c o n tro lle r to th e p la n t. E a c h
re s u lt-p ro d u c in g b ra n c h is a c o m p o s itio n o f th e fu n c tio n s a n d te rm in a ls fro m a
re p e rto ire (b e lo w ) o f fu n c tio n s a n d te rm in a ls .
P ro g ra m tre e s in th e p o p u la tio n d u rin g th e in itia l ra n d o m g e n e ra tio n (g e n e ra tio n 0 )
c o n s is t o n ly o f re s u lt-p ro d u c in g b ra n c h (e s ). A u to m a tic a lly d e fin e d fu n c tio n s a re
in tro d u c e d in c re m e n ta lly (a n d s p a rin g ly ) in to th e p o p u la tio n o n s u b s e q u e n t
g e n e ra tio n s b y m e a n s o f th e a rc h ite c tu re -a lte rin g o p e ra tio n s . E a c h a u to m a tic a lly
d e fin e d fu n c tio n is a c o m p o s itio n o f th e fu n c tio n s a n d te rm in a ls a p p ro p ria te fo r
c o n tro l p ro b le m s , re fe re n c e s to e x is tin g a u to m a tic a lly d e fin e d fu n c tio n s , a n d
(p o s s ib ly ) d u m m y v a ria b le s (fo rm a l p a ra m e te rs ) th a t p e rm it p a ra m e te riz a tio n o f th e
a u to m a tic a lly d e fin e d fu n c tio n . A u to m a tic a lly d e fin e d fu n c tio n s p ro v id e a m e c h a n is m
fo r in te rn a l fe e d b a c k (re c u rs io n ) w ith in th e to -b e -e v o lv e d c o n tro lle r. A u to m a tic a lly
d e fin e d fu n c tio n s a ls o p ro v id e a m e c h a n is m fo r re u s in g u s e fu l s u b s tru c tu re s .
E a c h b ra n c h o f e a c h p ro g ra m tre e in th e in itia l ra n d o m p o p u la tio n is c re a te d in
a c c o rd a n c e w ith a c o n s tra in e d s y n ta c tic s tru c tu re . E a c h g e n e tic o p e ra tio n e x e c u te d b y
g e n e tic p ro g ra m m in g (c ro s s o v e r, m u ta tio n , re p ro d u c tio n , o r a rc h ite c tu re -a lte rin g
o p e ra tio n ) p ro d u c e s o ffs p rin g th a t c o m p ly w ith th e c o n s tra in e d s y n ta c tic s tru c tu re .
G e n e tic p ro g ra m m in g h a s re c e n tly b e e n u s e d to c re a te a c o n tro lle r fo r a p a rtic u la r
tw o -la g p la n t a n d a th re e -la g p la n t (K o z a , K e a n e , Y u , B e n n e tt, a n d M y d lo w e c 2 0 0 0 ).
B o th o f th e s e g e n e tic a lly e v o lv e d c o n tro lle rs o u tp e rfo rm e d th e c o n tro lle rs d e s ig n e d b y
e x p e rts in th e fie ld o f c o n tro l u s in g th e c rite ria o rig in a lly s p e c ifie d b y th e e x p e rts .
5
P r e p a r a to r y S te p s
S ix m a jo r p re p a ra
d e te rm in e th e a rc h
th e fu n c tio n s , (4 ) d
a n d (6 ) c h o o s e th e
5 .1
S in c e th e re
th e c o n tro lle
o n e re s u lt-p
(g e n e ra tio n
g e n e ra tio n s ,
d e fin e d fu n c
to ry s te
ite c tu re
e fin e th
te rm in a
p s
o f
e f
tio
a re re q u ire d
th e p ro g ra m
itn e s s m e a s u r
n c rite rio n a n
b e fo re
tre e s , (2
e , (5 ) c h
d m e th o
P ro g ra m A rc h ite c tu re
is o n e re s u lt-p ro d u c in g b ra n c h in th
r a n d th is p ro b le m in v o lv e s a o n e -o u
ro d u c in g b ra n c h . E a c h p ro g ra m tr
0 ) h a s n o a u to m a tic a lly d e fin e d
a rc h ite c tu re -a lte rin g o p e ra tio n s m
tio n s (u p to a m a x im u m o f fiv e p e r p
a p p ly in g g e
) id e n tify th
o o s e c o n tro
d o f re s u lt d
e p ro g ra m
tp u t c o n tro
e e in th e
fu n c tio n s .
a y in s e rt
ro g ra m tre
n e tic p ro g ra m m in g : (1 )
e te rm in a ls , (3 ) id e n tify
l p a ra m e te rs fo r th e ru n ,
e s ig n a tio n .
tre e fo r e a c h
lle r, e a c h p ro g
in itia l ra n d o m
H o w e v e r, in
a n d d e le te a
e ).
o u tp u t fro m
ra m tre e h a s
g e n e ra tio n
su b se q u e n t
u to m a tic a lly
5 .2
T e rm in a l S e t
A c o n s tra in e d s y n ta c tic s tru c tu re p e rm its o n ly a s in g le p e rtu rb a b le n u m e ric a l v a lu e to
a p p e a r a s th e a rg u m e n t fo r e s ta b lis h in g e a c h n u m e ric a l p a ra m e te r v a lu e fo r e a c h
s ig n a l p ro c e s s in g b lo c k re q u irin g a p a ra m e te r v a lu e . T h e s e n u m e ric a l v a lu e s in itia lly
r a n g e f r o m - 5 .0 to + 5 .0 . T h e s e n u m e r ic a l v a lu e s a r e p e r tu r b e d d u r in g th e r u n b y a
G a u s s ia n m u ta tio n o p e ra tio n th a t o p e ra te s o n ly o n n u m e ric a l v a lu e s . N u m e ric a l
c o n s ta n ts a re la te r in te rp re te d o n a lo g a rith m ic s c a le s o th a t th e y re p re s e n t v a lu e s in a
ra n g e o f 1 0 o rd e rs o f m a g n itu d e (K o z a , B e n n e tt, A n d re , a n d K e a n e 1 9 9 9 ).
T h e r e m a i n i n g t e r m i n a l s a r e t i m e - d o m a i n s i g n a l s . T h e t e r m i n a l s e t , T, f o r t h e
re s u lt-p ro d u c in g b ra n c h a n d a n y a u to m a tic a lly d e fin e d fu n c tio n s (e x c e p t fo r th e
p e rtu rb a b le n u m e ric a l v a lu e s m e n tio n e d a b o v e ) is
1 7 2
J .R . K o z a e t a l.
T= { R E F
C
S p a c e
(a lth o u g h
K o z a , K e
5 .3
T h e
te rm
a u to
F=
E R E
O N S
d o e
th e
a n e ,
N C E
T A N
s n o
m e a
Y u ,
_ S I G N A L , C O N T R O L
T _ 0 } .
t p e rm it a d e ta ile d d e s
n in g o f th e a b o v e te rm
B e n n e tt, a n d M y d lo w e
L E R _ O U T P U T , P L A N T _ O U T P U T ,
c rip tio n o f th e v a rio u s te rm in a ls u s e d h e re in
in a ls s h o u ld b e c le a r fro m th e ir n a m e s ). S e e
c 2 0 0 0 fo r d e ta ils .
F u n c tio n S e t
fu n c tio n s a re s ig n a l p ro c e s s in g fu n c tio n s th a t o p e ra te o n tim e -d o m a in s ig
i n a l s i n T) . T h e f u n c t i o n s e t , F, f o r t h e r e s u l t - p r o d u c i n g b r a n c h
m a tic a lly d e fin e d fu n c tio n s is
{ G A I N , I N V E R T E R , L E A D , L A G , L A G 2 ,
D I F F E R E N T I A L _ I N P U T _ I N T E G R A T O R , D I F F E R E N T I A T O R ,
A D D _ S I G N A L , S U B _ S I G N A L , A D D _ 3 _ S I G N A L , D E L A Y , A D F 0 , …
A D F 4 } .
A D F 0 , … , A D F 4 d e n o te a u to m a tic a lly d e fin e d fu n c tio n s a d d e d d u rin g th
a rc h ite c tu re -a lte rin g o p e ra tio n s .
T h e fu n c tio n a lity o f e a c h o f th e a b o v e s ig n a l p ro c e s s in g fu n c tio n s is s u g g
th e ir n a m e s a n d is d e s c rib e d in d e ta il in K o z a , K e a n e , Y u , B e n n e tt, a n d M
2 0 0 0 .
n a ls (th e
a n d a n y
,
e ru n b y
e s te d b y
y d lo w e c
5 .4
F itn e s s
G e n e tic p ro g ra m m in g is a p ro b a b ilis tic a lg o rith m th a t s e a rc h e s th e s p a c e o f
c o m p o s itio n s o f th e a v a ila b le fu n c tio n s a n d te rm in a ls . T h e s e a rc h is g u id e d b y a
fitn e s s m e a s u re . T h e fitn e s s m e a s u re is a m a th e m a tic a l im p le m e n ta tio n o f th e h ig h le v e l re q u ire m e n ts o f th e p ro b le m . T h e fitn e s s m e a s u re is c o u c h e d in te rm s o f “ w h a t
n e e d s to b e d o n e ” n o t “ h o w to d o it.”
T h e fitn e s s m e a s u re m a y in c o rp o ra te a n y m e a s u ra b le , o b s e rv a b le , o r c a lc u la b le
b e h a v io r o r c h a ra c te ris tic o r c o m b in a tio n o f b e h a v io rs o r c h a ra c te ris tic s . T h e fitn e s s
m e a s u re fo r m o s t p ro b le m s o f c o n tro lle r d e s ig n is m u lti-o b je c tiv e in th e s e n s e th a t
th e re a re s e v e ra l d iffe re n t (u s u a lly c o n flic tin g ) re q u ire m e n ts fo r th e c o n tro lle r.
T h e f itn e s s o f e a c h in d iv id u a l is d e te r m in e d b y e x e c u tin g th e p r o g r a m tr e e ( i.e .,
th e re s u lt-p ro d u c in g b ra n c h a n d a n y a u to m a tic a lly d e fin e d fu n c tio n s th a t m a y b e
in v o k e d ) to p ro d u c e a n in te rc o n n e c te d s e q u e n c e o f s ig n a l p ro c e s s in g b lo c k s th a t is ,
a b lo c k d ia g ra m fo r th e c o n tro lle r. A S P IC E n e tlis t is th e n c o n s tru c te d fro m th e b lo c k
d ia g ra m . T h e S P IC E n e tlis t fo r th e re s u ltin g c o n tro lle r is w ra p p e d in s id e a n
a p p ro p ria te s e t o f S P IC E c o m m a n d s . T h e c o n tro lle r is th e n s im u la te d u s in g o u r
m o d if ie d v e r s io n o f th e S P I C E s im u la to r . T h e 2 1 7 ,0 0 0 - lin e S P I C E 3 s im u la to r
(Q u a rle s , N e w to n , P e d e rs o n , a n d S a n g io v a n n i-V in c e n te lli 1 9 9 4 ) is a n in d u s tria ls tre n g th s im u la to r. It is ru n a s a s u b m o d u le w ith in o u r g e n e tic p ro g ra m m in g s y s te m .
T h e S P IC E s im u la to r re tu rn s ta b u la r o u tp u t a n d o th e r in fo rm a tio n fro m w h ic h th e
fitn e s s o f th e in d iv id u a l is th e n c o m p u te d .
T h e fitn e s s o f a c o n tro lle r is m e a s u re d u s in g 1 3 e le m e n ts c o n s is tin g o f 1 2 tim e d o m a in -b a s e d e le m e n ts b a s e d o n a m o d ifie d in te g ra l o f tim e -w e ig h te d a b s o lu te e rro r
(IT A E ) a n d o n e tim e -d o m a in -b a s e d e le m e n t m e a s u rin g d is tu rb a n c e re je c tio n .
T h e f itn e s s o f a n in d iv id u a l c o n tr o lle r is th e s u m ( i.e ., lin e a r c o m b in a tio n ) o f th e
d e trim e n ta l c o n trib u tio n s o f th e s e 1 3 e le m e n ts o f th e fitn e s s m e a s u re . T h e s m a lle r th e
s u m , th e b e tte r.
T h e firs t 1 2 e le m e n ts o f th e fitn e s s m e a s u re e v a lu a te h o w q u ic k ly th e c o n tro lle r
c a u s e s t h e p l a n t t o r e a c h t h e r e f e r e n c e s i g n a l a n d t h e c o n t r o l l e r 's s u c c e s s i n a v o i d i n g
A u to m a tic S y n th e s is o f B o th th e T o p o lo g y a n d P a ra m e te rs
1 7 3
o v e rs h o o t. T w o re fe re n c e s ig n a ls a re u s e d . T h e firs t re fe re n c e s ig n a l is a s te p
fu n c tio n th a t ris e s fro m 0 to 1 v o lts a t t = 1 0 0 m illis e c o n d s w h ile th e s e c o n d ris e s fro m
0 to 1 m ic ro v o lts a t t = 1 0 0 m illis e c o n d s . T h e tw o s te p fu n c tio n s a re u s e d to d e a l w ith
th e n o n -lin e a rity c a u s e d b y th e lim ite r. T w o v a lu e s o f th e tim e c o n s ta n t, τ , a re u s e d
( n a m e ly 0 .5 a n d 1 .0 ) . T h r e e v a lu e s o f K a r e u s e d , n a m e ly 0 .9 , 1 .0 , a n d 1 .1 . E x p o s in g
g e n e tic p ro g ra m m in g to d iffe re n t c o m b in a tio n s o f v a lu e s o f s te p s iz e , K , a n d τ
p ro d u c e s a ro b u s t c o n tro lle rs a n d a ls o p re v e n ts g e n e tic p ro g ra m m in g fro m e n g a g in g in
p o le e lim in a tio n . F o r e a c h o f th e s e 1 2 fitn e s s c a s e s , a tra n s ie n t a n a ly s is is p e rfo rm e d
in th e tim e d o m a in u s in g th e S P IC E s im u la to r. T a b le 1 s h o w s th e e le m e n ts o f th e
fitn e s s m e a s u re in its le ft-m o s t fo u r c o lu m n s .
T h e c o n trib u tio n to fitn e s s fo r e a c h o f th e s e 1 2 e le m e n ts o f th e fitn e s s m e a s u re is
b a s e d o n th e in te g ra l o f tim e -w e ig h te d a b s o lu te e rro r (IT A E )
3 6
∫ (t − 5 ) e (t) A (e (t))B C d t .
t= 5
B e c a u s e o f th e b u ilt-in fiv e -s e c o n d tim e d e la y , th e in te g ra tio n ru n s fro m tim e t = 5
s e c o n d s to t = 3 6 s e c o n d s . H e re e (t) is th e d iffe re n c e (e rro r) a t tim e t b e tw e e n th e
d e la y e d p la n t o u tp u t a n d th e re fe re n c e s ig n a l. T h e in te g ra l o f tim e -w e ig h te d a b s o lu te
e rro r p e n a liz e s d iffe re n c e s th a t o c c u r la te r m o re h e a v ily th a n d iffe re n c e s th a t o c c u r
e a rlie r.
W e m o d ifie d th e in te g ra l o f tim e -w e ig h te d a b s o lu te e rro r in fo u r w a y s . F irs t, w e
u s e d a d is c re te a p p ro x im a tio n to th e in te g ra l b y c o n s id e rin g 1 2 0 3 0 0 -m illis e c o n d tim e
s te p s b e tw e e n t = 5 to t = 3 6 s e c o n d s . S e c o n d , w e m u ltip lie d e a c h fitn e s s c a s e b y th e
re c ip ro c a l o f th e a m p litu d e o f th e re fe re n c e s ig n a ls s o th a t b o th re fe re n c e s ig n a ls (1
m ic ro v o lt a n d 1 v o lt) a re e q u a lly in flu e n tia l. S p e c ific a lly , B is a fa c to r th a t is u s e d to
n o rm a liz e th e c o n trib u tio n s a s s o c ia te d w ith th e tw o s te p fu n c tio n s . B m u ltip lie s th e
d iffe re n c e e (t) a s s o c ia te d w ith th e 1 -v o lt s te p fu n c tio n b y 1 a n d m u ltip lie s th e
6
d iffe re n c e e (t) a s s o c ia te d w ith th e 1 -m ic ro v o lt s te p fu n c tio n b y 1 0 . T h ird , th e in te g ra l
c o n ta in s a n a d d itio n a l w e ig h t, A , th a t v a rie s w ith e (t). T h e fu n c tio n A w e ig h ts a ll
v a r ia tio n u p to 1 0 2 % o f th e r e f e r e n c e s ig n a l b y a f a c to r o f 1 .0 , a n d h e a v ily p e n a liz e s
o v e r s h o o ts o v e r 2 % b y a f a c to r 1 0 .0 . F o u r th , th e in te g r a l c o n ta in s a s p e c ia l w e ig h t, C ,
w h ic h is 5 .0 f o r th e tw o f itn e s s c a s e s f o r w h ic h K = 1 a n d τ = 1 , a n d 1 .0 o th e r w is e .
T h e 1 3 th e le m e n t o f th e fitn e s s m e a s u re is b a s e d o n d is tu rb a n c e re je c tio n . T h e
p e n a lty is c o m p u te d b a s e d o n a tim e - d o m a in a n a ly s is f o r 3 6 .0 s e c o n d s . I n th is
a n a ly s is , th e re fe re n c e s ig n a l is h e ld a t a v a lu e o f 0 . A d is tu rb a n c e s ig n a l c o n s is tin g o f
a u n it s te p is a d d e d to th e C O N T R O L L E R _ O U T P U T a t tim e t = 0 a n d th e re s u ltin g
d is tu rb e d s ig n a l is p ro v id e d a s in p u t to th e p la n t. T h e d e trim e n ta l c o n trib u tio n to
fitn e s s is 5 0 0 /3 6 tim e s th e tim e re q u ire d to b rin g th e p la n t o u tp u t to w ith in 2 0
m illiv o lts o f th e r e f e r e n c e s ig n a l o f 0 v o lts ( i.e ., to r e d u c e th e e f f e c t to w ith in 2 % o f
th e 1 -v o lt d is tu rb a n c e s ig n a l) a s s u m in g th a t th e p la n t s e ttle s to w ith in th is ra n g e w ith in
3 6 s e c o n d s . If th e p la n t d o e s n o t s e ttle to w ith in th is ra n g e w ith in 3 6 s e c o n d s , th e
d e trim e n ta l c o n trib u tio n to fitn e s s is 5 0 0 p lu s th e a b s o lu te v a lu e o f th e p la n t o u tp u t in
v o lts tim e s 5 0 0 . F o r e x a m p le , if th e e ffe c t o f th e d is tu rb a n c e w a s n e v e r re d u c e d b e lo w
1 v o lts , th e d e trim e n ta l c o n trib u tio n to fitn e s s w o u ld b e 1 0 0 0 .
A c o n tro lle r th a t c a n n o t b e s im u la te d b y S P IC E is a s s ig n e d a h ig h p e n a lty v a lu e o f
fitn e s s (1 0 8 ).
1 7 4
5 .5
T h e
te rm
1 0 0
p a ra
ra n g
J .R . K o z a e t a l.
C o n tro l P a ra m
p o p u la tio n s iz e , M , w a
in a ls ) w a s e s ta b lis h e d
p o in ts w a s e s ta b lis h e
m e te rs fo r c o n tro llin g
e o f p ro b le m s (K o z a , B
5 .6
T e rm
T h e ru n w a s m a n
s u c c e s s iv e b e s t-o
s in g le b e s t-s o -fa r
e te rs
s 5 0 0 ,0 0 0 . A m a x
fo r e a c h re s u lt-p ro
d fo r e a c h a u to m
th e ru n s a re th e d
e n n e tt, A n d re , a n d
im u m s iz e o f 1 5 0
d u c in g b ra n c h a n
a tic a lly d e fin e d
e fa u lt v a lu e s th a t
K e a n e 1 9 9 9 ).
p o in ts
d a m a
fu n c tio
w e a p
(fu
x im
n .
p ly
n c tio n s a n d
u m s iz e o f
T h e o th e r
to a b ro a d
in a tio n
u a lly m o n ito re d a n d m a n u a lly te rm in a te d w h e n th e fitn e s s o f m a n y
f-g e n e ra tio n in d iv id u a ls a p p e a re d to h a v e re a c h e d a p la te a u . T h e
in d iv id u a l is h a rv e s te d a n d d e s ig n a te d a s th e re s u lt o f th e ru n .
5 .7
P a ra lle l Im p le m e n ta tio n
T h is p ro b le m w a s ru n o n a h o m e -b u ilt B e o w u lf-s ty le (S te rlin g , S a lm o n , B e c k e r, a n d
S a v a re s e 1 9 9 9 ; B e n n e tt, K o z a , S h ip m a n , a n d S tiffe lm a n 1 9 9 9 ) p a ra lle l c lu s te r
c o m p u te r s y s te m c o n s is tin g o f 1 ,0 0 0 3 5 0 M H z P e n tiu m
II p ro c e sso rs (e a c h
a c c o m p a n ie d b y 6 4 m e g a b y te s o f R A M ). T h e s y s te m h a s a 3 5 0 M H z P e n tiu m II
c o m p u te r a s h o s t. T h e p ro c e s s in g n o d e s a re c o n n e c te d w ith a 1 0 0 m e g a b it-p e r-s e c o n d
E th e rn e t. T h e p ro c e s s in g n o d e s a n d th e h o s t u s e th e L in u x o p e ra tin g s y s te m . T h e
d is trib u te d g e n e tic a lg o rith m w ith u n s y n c h ro n iz e d g e n e ra tio n s a n d s e m i-is o la te d
s u b p o p u la tio n s w a s u s e d w ith a s u b p o p u la tio n s iz e o f Q = 5 0 0 a t e a c h o f D = 1 ,0 0 0
d e m e s . T w o p ro c e s s o rs a re h o u s e d in e a c h o f th e 5 0 0 p h y s ic a l b o x e s o f th e s y s te m . A s
e a c h p ro c e s s o r (a s y n c h ro n o u s ly ) c o m p le te s a g e n e ra tio n , fo u r b o a tlo a d s o f e m ig ra n ts
fro m e a c h s u b p o p u la tio n (s e le c te d p ro b a b ilis tic a lly b a s e d o n fitn e s s ) a re d is p a tc h e d to
e a c h o f th e fo u r to ro id a lly a d ja c e n t p ro c e s s o rs . T h e m ig ra tio n ra te is 2 % (b u t 1 0 % if
th e to ro id a lly a d ja c e n t n o d e is in th e s a m e p h y s ic a l b o x ).
6
R e s u lts
T h e b e s t in d iv id u a l in g e n e r a tio n 0 h a s a f itn e s s o f 1 9 2 6 .4 9 8 .
T h e b e s t-o f-ru n c o n tro lle r e m e rg e d in g e n e ra tio n 1 2 9 (fig u re 1 ). T h is b e s t-o f-ru n
c o n tr o lle r h a s a f itn e s s o f 5 2 2 .6 0 5 . T h e r e s u lt- p r o d u c in g b r a n c h o f th is b e s t- o f - r u n
in d iv id u a l h a s 1 1 9 p o in ts (fu n c tio n s a n d te rm in a ls ) a n d 9 5 , 9 3 , a n d 7 0 p o in ts ,
re s p e c tiv e ly , in its th re e a u to m a tic a lly d e fin e d fu n c tio n s .
N o te th a t g e n e tic
p r o g r a m m in g e m p lo y e d a 4 .8 s e c o n d d e la y ( c o m p a r a b le to th e f iv e - s e c o n d p la n t
d e la y ) in th e tra n s fe r fu n c tio n o f th e e v o lv e d p re -filte r. T h is b e s t-o f-ru n c o n tro lle r
fro m g e n e ra tio n 1 2 9 h a s a b e tte r v a lu e o f fitn e s s fo r a s te p s iz e o f 1 v o lt, a n in te rn a l
g a in , K , o f 1 .0 , a n d a tim e - c o n s ta n t, τ ,o f 1 .0 ( th e s p e c if ic c a s e c o n s id e r e d b y A s tr o m
a n d H a g g lu n d 1 9 9 5 ).
F ig u re 2 c o m p a re s th e tim e -d o m a in re s p o n s e to s te p in p u t o f th e b e s t-o f-ru n
c o n tro lle r fro m g e n e ra tio n 1 2 9 (tria n g le s ) w ith th e c o n tro lle r in A s tro m a n d H a g g lu n d
1 9 9 5 ( s q u a r e s ) f o r a s te p s iz e o f 1 v o lt, a n in te r n a l g a in , K , o f 1 .0 , a n d a tim e - c o n s ta n t,
τ ,o f 1 .0 .
F ig u re 3 c o m p a re s th e d is tu rb a n c e re je c tio n o f th e b e s t-o f-ru n c o n tro lle r fro m
g e n e ra tio n 1 2 9 (tria n g le s ) w ith th e c o n tro lle r in A s tro m a n d H a g g lu n d 1 9 9 5 (s q u a re s )
f o r a s te p s iz e o f 1 v o lt, a n in te r n a l g a in , K , o f 1 .0 , a n d a tim e - c o n s ta n t, τ ,o f 1 .0 .
A u to m a tic S y n th e s is o f B o th th e T o p o lo g y a n d P a ra m e te rs
Reference
Signal
0 .718 (1 + 7.00 s + 16 .74 s 2 ) + 0.282 e 4.80531s (1 + 4 .308 s )(1 + 1 .00106 s )
0 .1215
s (1 + 0.0317 s )(1 + 0 .01669 s )
+
1 7 5
Plant
Output
Control
Variable
Plant
-
(1 + 3 .96 s + 4.205 s 2 )(1 + 0 .238 s + 0.0837 s 2 )
F ig u re 1 B e s t-o f-ru n c o n tro lle r fro m
T a b le
th e A s tro
C = 5 .0 .
A s tro m a
1 c o
a n
A ll
n d H
m
m p
d H
1 2
a g g
a re s th e fitn
a g g lu n d 1 9
e n trie s a re
lu n d 1 9 9 5 c
g e n e ra tio n 1 2 9 fo r th re e -la g p la n t w ith fiv e -s e c o n d d e la y .
e s s o f th e b e s t-o f-ru n c o n tro lle r fro m g e n e ra tio n 1 2 9 a n d
9 5 . T w o o f th e e n trie s a re d iv id e d b y th e s p e c ia l w e ig h t
b e tte r fo r th e g e n e tic a lly e v o lv e d c o n tro lle r th a n fo r th e
o n tro lle r.
T a b le 1 F itn e s s o f tw o c o n tr o lle r s fo r th r e e -la g p la n t w ith
B e s t-o f-ru
T im e
P la n t
E le m e n t
S te p
g e n e ra tio n
c o n s ta n
in te rn a l
s iz e
1 2 9
G a in , K
(v o lts )
t, τ
0
1
0 .9
1 .0
1 3 .7
1
1
0 .9
0 .5
2 5 .6
2
1
1 .0
1 .0
3 4 .0 / 5 =
3
1
1 .0
0 .5
1 8 .6
4
1
1 .1
1 .0
4 .4
5
1
1 .1
0 .5
1 6 .3
6
1 0 -6
0 .9
1 .0
1 3 .2
7
1 0 -6
0 .9
0 .5
2 5 .5
8
1 0 -6
1 .0
1 .0
3 0 .7 / 5 =
9
1 0 -6
1 .0
0 .5
1 8 .5
1 0
1 0 -6
1 .1
1 .0
4 .3
1 1
1 0 -6
1 .1
0 .5
1 6 .2
D is tu rb a n c e
1
1
1
3 0 2
n
fiv e -s e c o n d d e la y .
A s tro m a n d
H a g g lu n d
c o n tro lle r
2 7 .4
3 8 .2
6 .8
2 2 .9
2 9 .3
2 5 .4
2 2 .7
2 7 .4
3 8 .2
6 .1
2 2 .9
2 9 .3
2 5 .4
2 2 .7
3 7 3
R e fe r e n c e s
A n g e lin e , P e te r J . a n d K in n e a r, K e n n e th E . J r. (e d ito rs ).
P r o g r a m m in g 2 . C a m b rid g e , M A : T h e M IT P re s s .
A s tro m , K a rl J . a n d H a g g lu n d , T o re . 1 9 9 5 . P ID C o n tr o
T u n in g . 2 n d E d itio n . R e s e a rc h T ria n g le P a rk , N C : In s tru
B a n z h a f , W o lf g a n g , D a id a , J a s o n , E ib e n , A . E ., G a r z o n ,
J a k ie la , M a rk , a n d S m ith , R o b e rt E . (e d ito rs ). 1 9 9 9 . G
th e G e n e tic a n d E v o lu tio n a r y C o m p u ta tio n C o n fe r e n c e ,
F lo r id a U S A . S a n F ra n c is c o , C A : M o rg a n K a u fm a n n .
B a n z h a f , W o lf g a n g , N o r d in , P e te r , K e lle r , R o b e r t E ., a n d
G e n e tic P r o g r a m m in g – A n In tr o d u c tio n . S a n F ra n c is c
a n d H e id e lb e rg : d p u n k t.
1 9 9 6 . A d v a n c e s in G e n e tic
lle r s : T
m e n t S
M a x H
E C C O
J u ly 1 3
h e o r y , D e s ig n , a n d
o c ie ty o f A m e ric a .
., H o n a v a r , V a s a n t,
-9 9 : P r o c e e d in g s o f
-1 7 , 1 9 9 9 , O r la n d o ,
F ra n c o n e , F ra n k D . 1 9 9 8 .
o , C A : M o rg a n K a u fm a n n
1 7 6
J .R . K o z a e t a l.
B a n z h a f, W o lfg a n g , P o li, R ic c a rd o , S c h o e n a u e r, M a rc , a n d F o g a rty , T e re n c e C . 1 9 9 8 .
G e n e tic P r o g r a m m in g : F ir s t E u r o p e a n W o r k s h o p . E u r o G P ’9 8 . P a r is , F r a n c e , A p r il
1 9 9 8 P r o c e e d in g s . P a r is , F r a n c e . A p r il l9 9 8 . L e c tu re N o te s in C o m p u te r S c ie n c e .
V o lu m e 1 3 9 1 . B e rlin , G e rm a n y : S p rin g e r-V e rla g .
B e n n e tt, F o r r e s t H I I I , K o z a , J o h n R ., S h ip m a n , J a m e s , a n d S tif f e lm a n , O s c a r . 1 9 9 9 .
B u ild in g a p a r a lle l c o m p u te r s y s te m f o r $ 1 8 ,0 0 0 th a t p e r f o r m s a h a lf p e ta - f lo p p e r
d a y . I n B a n z h a f , W o lf g a n g , D a id a , J a s o n , E ib e n , A . E ., G a r z o n , M a x H ., H o n a v a r ,
V a s a n t, J a k ie la , M a rk , a n d S m ith , R o b e rt E . (e d ito rs ). 1 9 9 9 . G E C C O -9 9 :
P r o c e e d in g s o f th e G e n e tic a n d E v o lu tio n a r y C o m p u ta tio n C o n fe r e n c e , J u ly 1 3 -1 7 ,
1 9 9 9 , O r la n d o , F lo r id a U S A . S a n F ra n c is c o , C A : M o rg a n K a u fm a n n . 1 4 8 4 - 1 4 9 0 .
B o y d , S . P . a n d B a rra tt, C . H . 1 9 9 1 . L in e a r C o n tr o lle r D e s ig n : L im its o f P e r fo r m a n c e .
E n g le w o o d C liffs , N J : P re n tic e H a ll.
H o lla n d , J o h n H . 1 9 7 5 . A d a p ta tio n in N a tu r a l a n d A r tific ia l S y s te m s . A n n A rb o r, M I:
U n iv e rs ity o f M ic h ig a n P re s s .
K in n e a r, K e n n e th E . J r. (e d ito r). 1 9 9 4 . A d v a n c e s in G e n e tic P r o g r a m m in g .
C a m b rid g e , M A : T h e M IT P re s s .
K o z a , J o h n R . 1 9 9 2 . G e n e tic P r o g r a m m in g : O n th e P r o g r a m m in g o f C o m p u te r s b y
M e a n s o f N a tu r a l S e le c tio n . C a m b rid g e , M A : M IT P re s s .
K o z a , J o h n R . 1 9 9 4 a . G e n e tic P r o g r a m m in g II: A u to m a tic D is c o v e r y o f R e u s a b le
P r o g r a m s . C a m b rid g e , M A : M IT P re s s .
K o z a , J o h n R . 1 9 9 4 b . G e n e tic P r o g r a m m in g II V id e o ta p e : T h e N e x t G e n e r a tio n .
C a m b rid g e , M A : M IT P re s s .
K o z a , J o h n R ., B a n z h a f , W o lf g a n g , C h e lla p illa , K u m a r , D e b , K a ly a n m o y , D o r ig o ,
M a r c o , F o g e l, D a v id B ., G a r z o n , M a x H ., G o ld b e r g , D a v id E ., I b a , H ito s h i, a n d
R io lo , R ic k . (e d ito rs ). 1 9 9 8 . G e n e tic P r o g r a m m in g 1 9 9 8 : P r o c e e d in g s o f th e T h ir d
A n n u a l C o n fe r e n c e . S a n F ra n c is c o , C A : M o rg a n K a u fm a n n .
K o z a , J o h n R ., B e n n e tt I I I , F o r r e s t H , A n d r e , D a v id , a n d K e a n e , M a r tin A . 1 9 9 9 .
G e n e tic P r o g r a m m in g III: D a r w in ia n In v e n tio n a n d P r o b le m S o lv in g . S a n
F ra n c is c o , C A : M o rg a n K a u fm a n n . F o rth c o m in g .
K o z a , J o h n R ., B e n n e tt I I I , F o r r e s t H , A n d r e , D a v id , K e a n e , M a r tin A ., a n d B r a v e
S c o tt. 1 9 9 9 . G e n e tic P r o g r a m m in g III V id e o ta p e : H u m a n -C o m p e titiv e M a c h in e
In te llig e n c e . S a n F ra n c is c o , C A : M o rg a n K a u fm a n n .
K o z a , J o h n R ., D e b , K a ly a n m o y , D o r ig o , M a r c o , F o g e l, D a v id B ., G a r z o n , M a x , I b a ,
H ito s h i, a n d R io lo , R . L . (e d ito rs ). 1 9 9 7 . G e n e tic P r o g r a m m in g 1 9 9 7 : P r o c e e d in g s
o f th e S e c o n d A n n u a l C o n fe r e n c e S a n F ra n c is c o , C A : M o rg a n K a u fm a n n .
K o z a , J o h n R ., G o ld b e r g , D a v id E ., F o g e l, D a v id B ., a n d R io lo , R ic k L . ( e d ito r s ) .
1 9 9 6 . G e n e tic P r o g r a m m in g 1 9 9 6 : P r o c e e d in g s o f th e F ir s t A n n u a l C o n fe r e n c e .
C a m b rid g e , M A : M IT P re s s .
K o z a , J o h n R ., K e a n e , M a r tin A ., Y u , J e s s e n , B e n n e tt, F o r r e s t H I I I , a n d M y d lo w e c ,
W illia m . 2 0 0 0 . A u to m a tic c re a tio n o f h u m a n -c o m p e titiv e p ro g ra m s a n d c o n tro lle rs
b y m e a n s o f g e n e tic p ro g ra m m in g . G e n e tic P r o g r a m m in g a n d E v o lv a b le M a c h in e s .
(1 ) 1 2 1 - 1 6 4 .
K o z a , J o h n R ., a n d R ic e , J a m e s P . 1 9 9 2 . G e n e tic P r o g r a m m in g : T h e M o v ie .
C a m b rid g e , M A : M IT P re s s .
L a n g d o n , W illia m B . 1 9 9 8 . G e n e tic P r o g r a m m in g a n d D a ta S tr u c tu r e s : G e n e tic
P r o g r a m m in g + D a ta S tr u c tu r e s = A u to m a tic P r o g r a m m in g ! A m s te rd a m : K lu w e r.
A u to m a tic S y n th e s is o f B o th th e T o p o lo g y a n d P a ra m e te rs
1 7 7
P o li, R ic c a r d o , N o r d in , P e te r , L a n g d o n , W illia m B ., a n d F o g a r ty , T e r e n c e C . 1 9 9 9 .
G e n e tic P r o g r a m m in g : S e c o n d E u r o p e a n W o r k s h o p . E u r o G P ’9 9 . P r o c e e d in g s .
L e c tu re N o te s in C o m p u te r S c ie n c e . V o lu m e 1 5 9 8 . B e rlin : S p rin g e r-V e rla g .
Q u a r le s , T h o m a s , N e w to n , A . R ., P e d e r s o n , D . O ., a n d S a n g io v a n n i- V in c e n te lli, A .
1 9 9 4 . S P I C E 3 V e r s io n 3 F 5 U s e r ’s M a n u a l . D e p a r tm e n t o f E le c tr ic a l E n g in e e r in g
a n d C o m p u te r S c ie n c e , U n iv . o f C a lifo rn ia . B e rk e le y , C A . M a rc h 1 9 9 4 .
R y a n , C o n o r. 1 9 9 9 . A u to m a tic R e -e n g in e e r in g o f S o ftw a r e U s in g G e n e tic
P r o g r a m m in g . A m s te rd a m : K lu w e r A c a d e m ic P u b lis h e rs .
S p e c t o r , L e e , L a n g d o n , W i l l i a m B . , O ’R e i l l y , U n a - M a y , a n d A n g e l i n e , P e t e r ( e d i t o r s ) .
1 9 9 9 . A d v a n c e s in G e n e tic P r o g r a m m in g 3 . C a m b rid g e , M A : M IT P re s s .
S te r lin g , T h o m a s L ., S a lm o n , J o h n , B e c k e r , D . J ., a n d S a v a r e s e , D . F . 1 9 9 9 . H o w to
B u ild a B e o w u lf: A G u id e to Im p le m e n ta tio n a n d A p p lic a tio n o f P C C lu s te r s .
C a m b rid g e , M A : M IT P re s s .
1.2
1
Voltage
0.8
0.6
0.4
0.2
0
0
5
10
15
20
25
30
35
40
Time
F ig u re 2 C o m p a ris o n fo r s te p in p u t.
1
0.8
Volt age
0.6
0.4
0.2
0
-0.2
0
5
10
15
20
25
30
35
Time
F ig u re 3 C o m p a ris o n fo r d is tu rb a n c e re je c tio n .
40
A u to m a tic D e s ig n o f M u ltiv a r ia b le Q F T C o n tr o l S y s te m
v ia E v o lu tio n a r y C o m p u ta tio n
K . C . T a n , T . H . L e e , a n d E . F . K h o r
D e p a rtm
N a tio
1 0 K e n t R
{ e le ta n k c ,
e n t
n a l U
id g e
e le le
o f E le c tric a l
n iv e rs ity o f
C re s c e n t S in
e th , e n g p 8 6 2
E n g in
S in g a
g a p o r
6 } @ n
e e
p o
e 1
u s
rin g
re
1 9 2 6 0
.e d u .s g
A b str a c t. T h is p a p e r p ro p o s e s a m u lti-o b je c tiv e e v o lu tio n a ry a u to m a te d
d e s ig n m e th o d o lo g y fo r m u ltiv a ria b le Q F T c o n tro l s y s te m s . U n lik e
e x is tin g m a n u a l o r c o n v e x o p tim is a tio n b a s e d Q F T d e s ig n a p p ro a c h e s ,
t h e ’i n t e l l i g e n t ’ e v o l u t i o n a r y t e c h n i q u e i s c a p a b l e o f a u t o m a t i c a l l y
e v o lv in g b o th th e n o m in a l c o n tro lle r a n d p re -filte r s im u lta n e o u s ly to
m e e t a ll p e rfo rm a n c e re q u ire m e n ts in Q F T , w ith o u t g o in g th ro u g h th e
c o n s e rv a tiv e a n d s e q u e n tia l d e s ig n s ta g e s fo r e a c h o f th e m u ltiv a ria b le
s u b -s y s te m s . In a d d itio n , it a v o id s th e n e e d o f m a n u a l Q F T b o u n d
c o m p u ta tio n a n d tria l-a n d -e rro r lo o p -s h a p in g d e s ig n p ro c e d u re s , w h ic h
is p a rtic u la rly u s e fu l fo r u n s ta b le o r n o n -m in im u m p h a s e p la n ts fo r
w h ic h s ta b ilis in g c o n tro lle rs m a y b e d iffic u lt to b e s y n th e s is e d .
E ffe c tiv e n e s s o f th e p ro p o s e d Q F T d e s ig n m e th o d o lo g y is v a lid a te d
u p o n a b e n c h m a rk m u ltiv a ria b le s y s te m , w h ic h o ffe rs a s e t o f lo w -o rd e r
P a re to o p tim a l c o n tro lle rs th a t s a tis fy a ll th e re q u ire d c lo s e d -lo o p
p e rfo rm a n c e s u n d e r p ra c tic a l c o n s tra in ts .
1
I n tr o d u c tio n
Q u a n tita tiv e F e e d b a c k T h e o ry (Q F T ) is w e ll-k n o w n a s a n e ffic ie n t fre q u e n c y d o m a in
c o n tro lle r d e s ig n m e th o d o lo g y th a t u tilis e s N ic h o ls c h a rt to a c h ie v e a d e s ire d ro b u s t
d e s ig n o v e r s p e c ifie d ra n g e s o f s tru c tu re d p la n t p a ra m e te r u n c e rta in tie s w ith a n d
w ith o u t c o n tro l e ffe c to r fa ilu re s [1 -3 ]. T h e b a s ic id e a o f Q F T is to c o n v e rt d e s ig n
s p e c ific a tio n o n c lo s e d -lo o p re s p o n s e a n d p la n t u n c e rta in ty in to ro b u s t s ta b ility a n d
p e rfo rm a n c e b o u n d s o n o p e n -lo o p tra n s m is s io n o f th e n o m in a l s y s te m a s s h o w n in
F ig . 1 . A fix e d s tru c tu re c o n tro lle r G (s ) a n d p re -filte r F (s ) is th e n s y n th e s iz e d u s in g
g a in -p h a s e lo o p -s h a p in g te c h n iq u e s o th a t th e tw o -d e g re e -fre e d o m o u tp u t fe e d b a c k
s y s te m is c o n tro lle d w ith in s p e c ific a tio n fo r a n y m e m b e r o f th e p la n t te m p la te s . F o r
m u lti-in p u t m u lti-o u tp u t (M IM O ) s y s te m s , c o n v e n tio n a l Q F T m e th o d re q u ire s th e
d e s ig n p ro c e s s to b e tu rn e d in to a s e q u e n c e o f m u lti-in p u t s in g le -o u tp u t (M IS O )
p ro b le m s b e fo re a n y Q F T d e s ig n p ro c e d u re c a n b e p e rfo rm e d [1 -3 ]. G iv in g a
c o m b i n e d s o l u t i o n o f c o n t r o l l e r G ( s ) = d i a g [ g i( s ) ] a n d p r e - f i l t e r F ( s ) = [ f ij] , " i , j = 1 ,
2 , … , m fo r a n m in p u t a n d m o u tp u t c o n tro l p ro b le m , th e s o lu tio n o f th e firs t s e t o f
M IS O p ro b le m is th e firs t tra n s fe r fu n c tio n o f th e d ia g o n a l c o n tro lle r g 1(s ) a n d f11(s ).
S. Cagnoni et al. (Eds.): EvoWorkshops 2000, LNCS 1803, pp. 178-194, 2000.
' Springer-Verlag Berlin Heidelberg 2000
Automatic Design of Multivariable QFT Control System
T
f1
f2
M
in
m
h e n a
(s) w
2 ( s ) , …
IS O
p u ts /o
n e w se t o
h ic h re s u
, f2m (s ) a
e q u iv a le n
u tp u ts , w
R (s)
f M IS O
lts to th
n d e tc .,
t lo o p s
h ic h c a n
p ro b le m is d e f
e la tte r s o lu tio
H e n c e , th is d e
a n d its s iz e
b e v e ry te d io u
in e d b a s e d o
n o f th e c o n
s ig n m e th o d
g ro w s e x p o
s to th e d e s ig
n th e p la n t, g 1(s ), f11(s ), f12(
tro lle r g 2(s ) a n d p re -filte r
le a d s to a n o v e ra ll d e s ig n
n e n tia lly to th e m n u m b
n e r [3 ].
179
s ) ,… ,
f21(s ),
o f m 2
e r o f
Y (s)
F (s)
G (s)
P (s)
F ig . 1 . A ty p ic a l o u tp u t fe e d b a c k m u ltiv a ria b le c o n tro l s y s te m
in Q F T
A s p o in te d o u t in [4 ], th e re a re n o o th e r d e s ig n m e th o d s th a n th e m a n u a l tria la n d -e rro r p ro c e s s c o u ld b e e m p lo y e d to d e te rm in e th e s e s e rie s o f lo o p s . M o re o v e r,
th is s e q u e n tia l M IS O d e s ig n p ro c e d u re m a y b e c o n s e rv a tiv e s in c e th e s o lu tio n o f a s e t
o f M IS O is h ig h ly d e p e n d e n t a n d re s tric te d to th e fo rm e r s e ts o f M IS O s o lu tio n . T h e
u s e r m a y th u s h a v e to re p e a t o r re -re s ta rt th e d e s ig n p ro c e d u re fro m th e firs t M IS O
lo o p s y n th e s is , if th e re e x is t a n y s e t o f M IS O s o lu tio n s th a t is u n fe a s ib le d u e to
im p ro p e r o r o v e rly d e s ig n o f th e p re v io u s s e ts o f M IS O lo o p s . B e s id e s , Q F T b o u n d s
in N ic h o ls c h a rt fo r e a c h s p e c ific a tio n o f a ll fre q u e n c y p o in ts m u s t b e a c q u ire d b e fo re
th e d e s ig n , w h ic h is o fte n a n e x h a u s tiv e tria l-a n d -e rro r p ro c e s s . T h e re a s o n is th a t, fo r
e v e ry fre q u e n c y p o in t w ith s u ffic ie n tly s m a ll fre q u e n c y in te rv a l, th e te m p la te n e e d s to
b e m a n u a lly s h ifte d u p o r d o w n o n th e N ic h o ls c h a rt u n til th e g a in v a ria tio n o f th e
te m p la te is e q u a l to th e g a in v a ria tio n a llo w e d fo r a n y p a rtic u la r ro b u s t s p e c ific a tio n
a t th a t fre q u e n c y . In a d d itio n , o n ly th e c o n tro lle r c a n b e s y n th e s iz e d v ia Q F T b o u n d
c o m p u ta tio n u s in g th e c o n v e n tio n a l lo o p -s h a p in g m e th o d . A n o th e r in d e p e n d e n t
d e s ig n ta s k h a s to b e a c c o m p lis h e d in o rd e r to o b ta in th e p re -filte r w ith in a tw o -s ta g e
d e s ig n fra m e w o rk fo r e a c h s e t o f M IS O s o lu tio n .
T h e a c q u is itio n o f a n o p tim a l Q F T c o n tro lle r is in fa c t a m u lti-o b je c tiv e m u ltim o d a l d e s ig n o p tim is a tio n p ro b le m th a t in v o lv e s s im u lta n e o u s ly d e te rm in in g m u ltip le
c o n tro lle r a n d p re -filte r p a ra m e te rs to s a tis fy d iffe re n t c o m p e tin g p e rfo rm a n c e
re q u ire m e n ts , s u c h a s s e n s itiv ity b o u n d s , c ro s s -c o u p lin g b o u n d s , ro b u s t m a rg in a n d
e tc ., T o s o lv e th e s e p r o b le m s , a fe w a n a ly tic a l/m a th e m a tic s o r ie n te d o p tim is a tio n o r
'a u t o m a t i c d e s i g n ' t e c h n i q u e s h a v e r e c e n t l y b e e n i n v e s t i g a t e d a n d d e v e l o p e d [ 5 - 7 ] .
T h e s e c o n v e x b a s e d o p tim is a tio n a p p ro a c h e s , h o w e v e r, im p o s e m a n y u n p ra c tic a l o r
u n re a lis tic a s s u m p tio n s th a t o fte n le a d to v e ry c o n s e rv a tiv e d e s ig n s o r a re h a rd in
fin d in g th e g lo b a l P a re to o p tim a l s o lu tio n s in th e m u lti-o b je c tiv e m u lti-d im e n s io n a l
d e s ig n s p a c e . W ith th e v ie w o f ta c k lin g th e s e d ra w b a c k s a n d a u to m a tin g th e Q F T
d e s i g n p r o c e d u r e , c o m p u t e r i s e d 'i n t e l l i g e n t ' t r i a l - a n d - e r r o r b a s e d m e t h o d o l o g y b a s e d
o n e v o lu tio n a ry o p tim is a tio n h a s b e e n p ro p o s e d a n d s u c c e s s fu lly a p p lie d to in d u s tria l
o r b e n c h m a rk a p p lic a tio n s [8 -1 0 ].
T h i s p a p e r f u r t h e r d e v e l o p s t h e m u l t i - o b j e c t i v e 'i n t e l l i g e n t ' a u t o m a t e d Q F T
d e s ig n m e th o d o lo g y to M IM O c o n tro l s y s te m u s in g a h ig h p e rfo rm a n c e e v o lu tio n a ry
a lg o rith m to o lb o x [1 1 ]. U n lik e e x is tin g m e th o d s , th e e v o lu tio n a ry Q F T d e s ig n
a p p ro a c h is c a p a b le o f c o n c u rre n tly e v o lv in g th e c o n tro lle r a n d p re -filte r fo r th e e n tire
180
K.C. Tan, T.H. Lee, and E.F. Khor
s e t o f M IS O s u b -s y s te m s to m e e t a ll p e rfo rm a n c e re q u ire m e n ts in Q F T , w ith o u t g o in g
th ro u g h th e c o n s e rv a tiv e a n d s e q u e n tia l d e s ig n s ta g e s fo r e a c h o f th e M IS O s u b s y s te m s . B e s id e s , th e e v o lu tio n a ry to o lb o x is b u ilt w ith c o m p re h e n s iv e u s e r in te rfa c e
a n d p o w e rfu l g ra p h ic a l d is p la y s fo r e a s y a s s e s s m e n t o f v a rio u s s im u la tio n re s u lts o r
tra d e -o ffs a m o n g th e d iffe re n t d e s ig n s p e c ific a tio n s . T h e p a p e r is o rg a n iz e d a s
fo llo w s : T h e v a rio u s Q F T d e s ig n s p e c ific a tio n s a n d th e ro le o f th e M O E A to o lb o x in
th e m u ltiv a ria b le Q F T d e s ig n a re g iv e n in S e c tio n 2 . V a lid a tio n o f th e p ro p o s e d
m e th o d o lo g y a g a in s t a b e n c h m a rk M IM O s y s te m is illu s tra te d in S e c tio n 3 .
C o n c lu s io n s a re d ra w n in S e c tio n 4 .
2 .
E v o lu tio n a r y A u to m a te d M u ltiv a r ia b le Q F T D e s ig n
2 .1 M u lt i-O b j e c t iv e
A s m e n tio n e d in th e
o b je c tiv e s n e e d to b e
to th e c o n v e n tio n a l tw
a re fo rm u la te d a s a m
to c o n c u rre n tly d e s ig
s a tis fy a ll th e re q u ire d
Q F T D e s ig n S p e c ific a tio n s
In tro d u c tio n , th e re a re a n u m b e r o f u s u a
s a tis fie d c o n c u rre n tly in m u ltiv a ria b le Q F
o -s ta g e lo o p -s h a p in g a p p ro a c h , th e s e p e rf
u lti-o b je c tiv e d e s ig n o p tim is a tio n p ro b le m
n th e n o m in a l c o n tro lle r G (s ) a n d p re -f
s p e c ific a tio n s a s d e s c rib e d b e lo w :
lly c o n flic tin g d e
T d e s ig n s . In c o n
o rm a n c e re q u ire m
h e re . T h e a im is
ilte r F (s ) in o rd e
s ig n
tra s t
e n ts
th u s
r to
(i) S ta b ility (R H S P )
T h e c o s t o f s ta b ility , R H S P , is in c lu d e d to e n s u re s ta b ility o f th e c lo s e d -lo o p s y s te m ,
w h ic h c o u ld b e e v a lu a te d b y s o lv in g th e ro o ts o f th e c h a ra c te ris tic p o ly n o m ia l.
C le a rly , a s ta b le c lo s e d -lo o p s y s te m fo r a ll th e p la n t te m p la te s ¨ re q u ire s a z e ro
v a lu e o f R H S P ,
P G F
> 0 , ∀ P i ∈ ℘
R H S P =
N r r e a l p o l e i
(1 )
I + P i G
i
In o rd e r to e n s u re in te rn a l s ta b ility a n d to g u a ra n te e n o u n s ta b le p o le a n d n o n m in im u m p h a s e z e ro c a n c e lla tio n s , it is d e s ire d th a t a m in im u m p h a s e a n d s ta b le
c o n tro lle r b e d e s ig n e d . T h is im p lie s th a t th e s e a rc h ra n g e fo r a p o ly n o m ia l c o e ffic ie n t
s e t i s l i m i t e d t o e i t h e r t h e f i r s t o r t h e t h i r d ’q u a d r a n t ’, i . e . , a l l c o e f f i c i e n t s i n t h e
n u m e ra to r o r d e n o m in a to r m u s t b e o f th e s a m e s ig n [1 0 ]. T h e p o le s a n d z e ro s o f th e
c o n tro lle rs c a n b e c a lc u la te d e x p lic itly to a v o id R H P c a n c e lla tio n s o r a lte rn a tiv e ly ,
th e H o ro w itz m e th o d fo r Q F T d e s ig n o f u n s ta b le a n d n o n -m in im u m p h a s e p la n ts c a n
b e u s e d , i.e ., Q F T b o u n d s fo r a n u n s ta b le /n o n -m in im u m p h a s e n o m in a l p la n t c a n b e
tra n s la te d to th o s e fo r a s ta b le a n d m in im u m p h a s e p la n t, if n e c e s s a ry .
∑
(ii) R o b u s t U p p e r a n d L o w e r T r
T h e c o s t o f u p p e r tra c k in g p e rfo
lo o p tra n s fe r fu n c tio n , g iv e n b y
u p p e r tra c k in g b o u n d a s s h o w n
a t e a c h fre q u e n c y p o in t a s g iv e n
a c k in g P e r fo r m a n c
r m a n c e o f t h e i th d
E R R U T (i,i), i s i n c l u
in F ig . 2 . It is c o m
b y ,
E R R U T
(i,i)
=
∑
e (E
ia g o
d e d
p u te
R R U
n a l e
to a d
d a s
T &
le m
d re
th e
E
e n
ss
su
R R L T )
t o f M IM O c lo s e d th e s p e c ific a tio n o f
m o f a b s o lu te e rro r
n
e
k = 1
( i,i ) u t
(ω
k
)
(2 )
Automatic Design of Multivariable QFT Control System
w h e re n is th e to ta l n u m b e r o f in te
b e tw e e n th e u p p e r b o u n d o f th e
C L (i,i)U a n d t h e p r e - s p e c i f i e d u p p e r
b o u n d o f th e c lo s e d -lo o p s y s te m
b o u n d o r le s s th a n th e p re -s p e c ifie d
e q u a l to z e ro a s illu s tra te d in F ig . 2
e a c h f r e q u e n c y wk r e p r e s e n t s t h e m
re s te d fre q u
( i,i) e le m e n
tra c k in g b o
is g re a te r
lo w e r tra c k
, fo r w h ic h
a g n itu d e o f
181
e n c y p o i n t s ; e ( i , i ) u t ( wk ) i s t h e d i f f e r e n c e
t o f th e c lo s e d -lo o p tra n s fe r fu n c tio n
u n d T ( i , i ) U a t f r e q u e n c y wk , i f t h e u p p e r
th a n th e p re -s p e c ifie d u p p e r tra c k in g
i n g b o u n d T ( i , i ) L ; o t h e r w i s e , e ( i , i ) u t ( wk ) i s
th e le n g th o f th e v e rtic a l d o tte d lin e s a t
e ( i , i ) u t ( wk ) .
F i g . 2 . C o m p u t a t i o n o f u p p e r t r a c k i n g p e r f o r m a n c e f o r t h e i th d i a g o n a l e l e m e n t
T h e c o s t f o r l o w e r t r a c k i n g p e r f o r m a n c e o f t h e i th d i a g o n a l e l e m e n t o f t h e
c l o s e d - l o o p t r a n s f e r f u n c t i o n , g i v e n b y E R R L T (i,i), c a n b e d e f i n e d a s t h e s u m o f
a b s o lu te e rro r a t e a c h fre q u e n c y p o in t,
E R R L T
( i,i )
=
∑
n
e
( i,i ) lt
(ω
k
)
(3 )
k = 1
w h e r e n i s t h e n u m b e r o f f r e q u e n c y p o i n t s ; e (i,i)lt i s t h e
b o u n d o f t h e c l o s e d - l o o p s y s t e m C L (i,i)L a n d t h e p r e - s p
T (i,i)L , i f t h e l o w e r b o u n d o f t h e c l o s e d - l o o p s y s t e m i s
u p p e r t r a c k i n g b o u n d T (i,i)U o r l e s s t h a n t h e p r e - s p e c i f i e
O t h e r w i s e , e (i,i)lt i s e q u a l t o z e r o a s i l l u s t r a t e d i n F i g . 3 .
d iffe r
e c ifie
g re a te
d lo w
e n c e b e tw e
d lo w e r tra
r th a n th e
e r tra c k in g
e n th e lo w e r
c k in g b o u n d
p re -s p e c ifie d
b o u n d T (i,i)L ;
F i g . 3 . C o m p u t a t i o n o f l o w e r t r a c k i n g p e r f o r m a n c e f o r t h e i th d i a g o n a l e l e m e n t
182
K.C. Tan, T.H. Lee, and E.F. Khor
(iii) C r o s s -c o u p lin g P e r fo r m a n c e (E R R U C )
A p a rt fro m a d d re s s in g th e tra c k in g p e rfo rm a n c e o f d ia g o n a l e le m e n ts , it is a ls o
e s s e n tia l to re d u c e th e c o u p lin g e ffe c t o f th e o ff-d ia g o n a l tra n s fe r fu n c tio n s fo r a ll th e
p la n t te m p la te s in M IM O Q F T c o n tro l s y s te m d e s ig n . S in c e th e o b je c tiv e is to re d u c e
th e g a in a n d b a n d w id th o f th e o ff-d ia g o n a l tra n s fe r fu n c tio n , o n ly th e u p p e r b o u n d s in
th e fre q u e n c y re s p o n s e n e e d to b e p re s c rib e d [4 ]. T h e u p p e r b o u n d o f c o u p lin g e ffe c t
r e p r e s e n t e d b y t r a n s f e r f u n c t i o n T (i,j)U f o r t h e o f f - d i a g o n a l t r a n s f e r f u n c t i o n s ( i , j )
w h e re i j c a n b e d e fin e d a c c o rd in g to th e a llo w a b le g a in K a n d th e b a n d w id th
b e tw e e n w 1 a n d w 2, w h ic h is s h o w n in F ig . 4 a n d ta k e s th e fo rm o f
1
K (
s )
w 1
(4 )
T (i, j)U ( s ) =
1
1
s + 1
s + 1
w 1
w 2
T h e c o s t o f c ro s s -c o u p lin g e ffe c t fo r th e o ff-d ia g o n a l e le m e n ts o f M IM O c lo s e d -lo o p
s y s t e m , g i v e n b y E R R U C (i,j), i s i n c l u d e t o a d d r e s s t h e s p e c i f i c a t i o n o f u p p e r c r o s s c o u p lin g b o u n d a n d is c o m p u te d a s th e s u m o f a b s o lu te e rro r a t e a c h fre q u e n c y p o in t,
E R R U C
w h e re n
th e u p p e
c o u p lin g
p re -sp e c
e (i,j))u c i s
is th
r b o
b o u
ifie d
e q u a
e
(ω
( i, j) u c
k
( i, j)
=
∑
n
e
( i, j) u c
(ω
k
) , fo r i j
(5 )
k = 1
e n
u n d
n d
u p
l to
u m b e r o f fre q u e n c y p o in
o f th e c lo s e d -lo o p s y s te
T (i,j)U , i f t h e u p p e r b o u n d
p e r c ro s s -c o u p lin g b o u n
z e ro a s g iv e n b y ,
C L (i, j)U ( ω k ) − T (i, j)U ( ω k
) =
0
)
m
t s ; e ( i , j ) u c ( wk ) a t wk
C L (i,j)U a n d t h e
o f th e c lo s e d -lo o p
d T (i,j)U a s i l l u s t r a
,C L
( i, j)U
(ω
k
is
p re
sy
te d
) > T
o th e r w is e
th e
-sp e
s te m
in
( i, j)U
(ω
d iffe
c ifie
is g
F ig .
k
)
,
re n
d u
re a
4 ;
c e b e
p p e r
te r th
O th e
tw
c r
a n
rw
fo r i ≠ j
e e n
o ssth e
is e ,
(6 )
F ig . 4 . C o m p u ta tio n o f u p p e r c ro s s -c o u p lin g p e rfo rm a n c e fo r o ff-d ia g o n a l e le m e n ts
Automatic Design of Multivariable QFT Control System
(iv ) R o b u s t M a r g in (R M )
P ra c tic a l c o n tro l a p p lic a tio n s o fte n in v o lv e n e g le c te d
d y n a m ic s a t th e h ig h fre q u e n c y . B e c a u s e o f th e s e m
m o d e l u s e d in c o n tro l s y s te m d e s ig n is o fte n in a c c u
n e g le c t o r d u e to th e la c k o f u n d e rs ta n d in g o f th e p h y
a d d re s s th e s e u n m o d e lle d u n c e rta in tie s , th e u n c e rta in
m u l t i p l i c a t i v e p l a n t u n c e r t a i n t y , P ip ( s ) = P ( s ) { I + W iI (
u n c e rta in tie s o r
is s in g d y n a m ic s ,
ra te e ith e r th ro u g
s ic a l p ro c e s s [1 2 ]
fe e d b a c k s y s te m
− 1
s ) ∆ iI } a s s h o w n
u n m o d e
th e n o m
h d e lib e
. In o rd e
w ith in v
in F ig .
183
lle d
in a l
ra te
r to
e rse
5 is
c o n s d ire d in th e Q F T d e s ig n . T h e ro b u s t m a rg in s p e c ific a tio n th a t a d d re s s e s th e
c lo s e d -lo o p s ta b ility d u e to th e in v e rs e m u ltip lic a tiv e p la n t u n c e rta in ty fo r a n
u n c e r t a i n t y w e i g h t i n g f u n c t i o n W iI c a n b e d e f i n e d a s [ 1 2 ] ,
R M
IM
1
=
I + L i( jω )
<
W
1
( jω )
iI
, ∀ ω
(7 )
w h e r e L i ( j ω ) i s t h e i th o p e n - l o o p t r a n s f e r f u n c t i o n w i t h t h e j th l o o p b e i n g c l o s e d i n a n
M IM O s y s te m , w h ic h is s im p ly th e lo o p tra n s m is s io n P ( jω )G ( jω ) in a n S IS O
s y s te m .
W
iI
(s)
∆
P
iI
ip
Y (s)
G (s)
P (s)
F ig . 5 . F e e d b a c k s y s te m
(v )
T h
m a
m a
6 ,
p la
e
x
g
w
T h e
o f S
fro m
n
S e n s itiv ity R e je c tio n (R S )
ro b u s t s e n s itiv ity re je c tio
im u m a m p litu d e o f th e re g u
n itu d e . A g e n e ra l s tru c tu re
h ic h d e p ic ts th e p a rtic u la r
t o u tp u t. T h e m a th e m a tic a l
w ith in v e rs e m u ltip lic a tiv e u n c e rta in ty
n
is to fin d a Q F T c o n tro lle r th a t m in im is e s
la te d o u tp u t o v e r a ll p o s s ib le d is tu rb a n c e s o f b o u n
to re p re s e n t th e d is tu rb a n c e re je c tio n is g iv e n in
c a s e w h e re th e d is tu rb a n c e e n te rs th e s y s te m a t
re p re s e n ta tio n is g iv e n b y ,
Y
− 1
S =
= {I + P (s )G (s )}
D
m a trix S (s ) is k n o w n a s th e d is tu r b a n c e re je c tio n . T h e m a x im u m s in g u la r v a
d e te rm in e s th e d is tu rb a n c e a tte n u a tio n s in c e S is in fa c t th e c lo s e d -lo o p tra n
d is tu rb a n c e D to th e p la n t o u tp u t Y .
D (s)
G (s)
P (s)
Y (s)
W
s
(s)
F ig . 6 . F o rm u la tio n o f a s e n s itiv ity re je c tio n p ro b le m
th e
d e d
F ig .
th e
(8 )
lu e s
sfe r
184
K.C. Tan, T.H. Lee, and E.F. Khor
T h e d is tu rb a n c e a tte n u a tio n s p e c ific a tio n fo r th e c lo s e d -lo o p s y s te m
w ritte n a s ,
σ (S ) ≤ W
− 1
s
⇒
∞
W
s
S
∞
m a y th u s b e
< 1
(9 )
w h e re σ
d e fin e s th e la rg e s t s in g u la r v a lu e a n d W s th e d e s ire d d is tu rb a n c e
a tte n u a tio n fa c to r, w h ic h is a fu n c tio n o f fre q u e n c y to a llo w a d iffe re n t a tte n u a tio n
fa c to r a t e a c h fre q u e n c y .
(v i) H ig h F r
T h e h ig h fre
tra n s m is s io n
n o is e a n d th
a c tu a to r s a tu
is g iv e n a s ,
e q u e n c y G a in R
q u e n c y g a in p e
L (s ) a t th e h ig
e u n m o d e lle d
ra tio n a n d in s ta
o ll-o ff (H
rfo rm a n c
h fre q u e n
h ig h -fre q
b ility . T h
F G )
e , H F
c y in
u e n c y
e h ig h
G , is
o rd e
d y n
fre q
in c lu
r to a
a m ic s
u e n c y
d e d to
v o id th
/h a rm o
g a in o
re d u
e h ig
n ic s
f lo o
c e th
h -fre
th a t
p tra
lim s rL ( s )
s →
w h e re r is
is to b e o p
h ig h fre q u
s tru c tu re g
th e re la
tim is e d
e n c y g
iv e n a s
tiv e o rd e r o f L (s ). S in
, th is p e rfo rm a n c e re q
a in o f th e c o n tro lle r
[9 ],
b ( i,i) n s n
G ( i,i) ( s ) =
a ( i,i) m s m
e g a in o f
q u e n c y se
m a y re su
n s m is s io n
lo o p
n so r
lt in
L (s)
(1 0 )
∞
c e o n ly th e c o n tro lle r in th e lo o p tra n s m is s io n
u ire m e n t is e q u iv a le n t to th e m in im iz a tio n o f
o r th e m a g n itu d e o f b n/a m fo r a c o n tro lle r
+ b
+ a
( i,i) n − 1
( i,i) m − 1
n − 1
s
s
m − 1
+ Lb
+ La
( i,i ) 0
(1 1 )
( i,i)0
w h e r e n a n d m is th e o r d e r o f th e n u m e r a to r a n d d e n o m in a to r fo r th e ( i,i) e le m e n t o f
d ia g o n a l c o n tro lle r G (s ), re s p e c tiv e ly .
2 .2
E v o lu tio n a r y A lg o r ith m
T o o lb o x a n d I ts R o le in Q F T D e s ig n
A lth o u g h th e m u lti-o b je c tiv e o p tim is a tio n b a s e d Q F T d e s ig n m e th o d h a s th e m e rit o f
a v o id in g c o n v e n tio n a l in d e p e n d e n t tw o -s ta g e c o n tro lle r s y n th e s is o r th e te d io u s
s e q u e n tia l d e s ig n s to d e te rm in e th e s e rie s o f lo o p s fo r e a c h o f th e M IS O s u b -s y s te m ,
th e a p p ro a c h n e e d s to s e a rc h fo r m u ltip le o p tim is e d c o n tro lle r a n d p re -filte r
c o e ffic ie n ts to s a tis fy a s e t o f n o n -c o m m e n s u ra b le a n d o fte n c o m p e tin g d e s ig n
s p e c ific a tio n s . S u c h a n o p tim is a tio n p ro b le m is o fte n s e m i-in fin ite a n d g e n e ra lly n o t
e v e ry w h e re d iffe re n tia b le [9 ]. It is th u s h a rd to b e s o lv e d v ia tra d itio n a l n u m e ric a l
a p p ro a c h e s th a t o fte n re ly o n a d iffe re n tia b le p e rfo rm a n c e in d e x , w h ic h fo rm s th e
m a jo r o b s ta c le fo r th e d e v e lo p m e n t o f a g e n e ra lis e d n u m e ric a l o p tim is a tio n p a c k a g e
fo r Q F T c o n tro l a p p lic a tio n s .
T h is p a p e r p ro p o s e s a n e v o lu tio n a ry a u to m a te d d e s ig n m e th o d o lo g y fo r th e
m u lti-o b je c tiv e Q F T c o n tro l o p tim is a tio n p ro b le m . F ig . 7 s h o w s a g e n e ra l
a rc h ite c tu re fo r th e c o m p u te r a id e d c o n tro l s y s te m d e s ig n (C A C S D ) a u to m a tio n o f
M IM O Q F T c o n tro l s y s te m u s in g a m u lti-o b je c tiv e e v o lu tio n a ry a lg o rith m (M O E A )
to o lb o x . T h e d e s ig n c y c le a c c o m m o d a te s th re e d iffe re n t m o d u le s : th e in te ra c tiv e
h u m a n d e c is io n -m a k in g m o d u le (c o n tro l e n g in e e r), th e o p tim is a tio n m o d u le (M O E A
to o lb o x ) a n d th e Q F T c o n tro l m o d u le (s y s te m a n d s p e c ific a tio n s ). A c c o rd in g to th e
s y s te m p e rfo rm a n c e re q u ire m e n ts a n d a -p r io r i k n o w le d g e o n th e p ro b le m o n -h a n d if
a n y , c o n tro l e n g in e e rs m a y s p e c ify o r s e le c t th e d e s ire d Q F T s p e c ific a tio n s a s
Automatic Design of Multivariable QFT Control System
185
d is c u s s e d in p re v io u s s e c tio n s to fo rm a m u lti-o b je c tiv e fu n c tio n , w h ic h n e e d n o t
n e c e s s a ry b e c o n v e x o r d iffe re n tia b le . B a s e d o n th e s e d e s ig n s p e c ific a tio n s , re s p o n s e s
o f th e c o n tro l s y s te m c o n s is ts o f th e s e t o f in p u t/o u tp u t s ig n a ls , th e p la n t te m p la te a s
w e ll a s th e c a n d id a te c o n tro lle r G (s ) a n d p re -filte r F (s ) re c o m m e n d e d fro m th e
o p tim is a tio n m o d u le a re s im u la te d a s to d e te rm in e th e d iffe re n t c o s t v a lu e s fo r e a c h
d e s ig n s p e c ific a tio n in th e m u lti-o b je c tiv e fu n c tio n .
A c c o rd in g to th e e v a lu a tio n re s u lts o f th e m u lti-o b je c tiv e fu n c tio n in th e c o n tro l
m o d u le a n d th e d e s ig n g u id a n c e s u c h a s g o a l o r p rio rity in fo rm a tio n fro m th e
d e c is io n -m a k in g m o d u le , th e o p tim is a tio n m o d u le (M O E A to o lb o x ) a u to m a te s th e
Q F T d e s i g n p r o c e s s a n d i n t e l l i g e n t l y s e a r c h e s f o r t h e ’o p t i m a l ’ c o n t r o l l e r a n d p r e filte r p a ra m e te rs s im u lta n e o u s ly th a t b e s t s a tis fy th e s e t o f Q F T p e rfo rm a n c e
s p e c ific a tio n s . O n -lin e o p tim is a tio n p ro g re s s a n d s im u la tio n re s u lts , s u c h a s th e
d e s ig n tra d e -o ffs o r c o n v e rg e n c e a re d is p la y e d g ra p h ic a lly a n d fe e d b a c k to th e
d e c is io n -m a k in g m o d u le . In th is w a y , th e o v e ra ll Q F T d e s ig n e n v iro n m e n t is
s u p e rv is e d a n d m o n ito re d e ffe c tiv e ly , w h ic h h e lp s c o n tro l e n g in e e rs to m a k e
a p p ro p ria te a c tio n s s u c h a s e x a m in in g th e c o m p e tin g d e s ig n tra d e -o ffs , a lte rin g th e
d e s ig n s p e c ific a tio n s , a d ju s tin g g o a l s e ttin g s th a t a re to o s trin g e n t o r g e n e ro u s , o r
e v e n m o d ify in g th e Q F T c o n tro l a n d s y s te m s tru c tu re if n e c e s s a ry . T h is m a n -m a c h in e
in te ra c tiv e d e s ig n a n d o p tim is a tio n p ro c e s s m a y b e p ro c e e d e d u n til th e c o n tro l
e n g in e e r is s a tis fie d w ith th e re q u ire d p e rfo rm a n c e s o r a fte r th e d e s ig n s p e c ific a tio n s
h a v e b e e n m e t. S u c h a n e v o lu tio n a ry a u to m a te d a p p ro a c h a llo w s th e Q F T d e s ig n
p ro b le m a s w e ll a s th e in te ra c tio n w ith o p tim is a tio n p ro c e s s to b e c lo s e ly lin k e d to th e
e n v iro n m e n t o f th a t p a rtic u la r a p p lic a tio n . C o n tro l e n g in e e r, fo r m o s t o f th e p a rt, is
n o t re q u ire d to d e a l w ith a n y d e ta ils th a t a re re la te d to th e o p tim is a tio n a lg o rith m o r to
g o th ro u g h th e m a n u a l tria l-a n d -e rro r tw o -s ta g e a n d s e q u e n tia l d e s ig n a s a d o p te d in
c o n v e n tio n a l Q F T d e s ig n m e th o d s .
T h e M O E A
to o lb o x [1 1 ] h a s b e e n d e v e lo p e d u n d e r th e M a tla b [1 3 ]
p ro g ra m m in g e n v iro n m e n t, w h ic h is e ffe c tiv e fo r g lo b a l o p tim is a tio n a n d a s s e s s m e n t
o f m u lti-o b je c tiv e d e s ig n tra d e -o ff s c e n a rio s , a id in g a t d e c is io n -m a k in g fo r a n o p tim a l
s o lu tio n th a t b e s t m e e ts a ll d e s ig n s p e c ific a tio n s . It is a ls o c a p a b le o f h a n d lin g
p ro b le m s w ith c o n s tra in ts a n d in c o rp o ra tin g a d v a n c e d g o a l a n d p rio rity in fo rm a tio n
w ith lo g ic a l A N D /O R o p e ra tio n s fo r h ig h e r-d e c is io n s u p p o rt. B e s id e s , it is fu lly
fu n c tio n e d w ith g ra p h ic a l u s e r in te rfa c e (G U I) a n d is re a d y fo r im m e d ia te u s e w ith
m in im a l k n o w le d g e o n e v o lu tio n a ry c o m p u tin g o r M a tla b p ro g ra m m in g . T h e to o lb o x
a ls o a llo w s th e d iffe re n t re p re s e n ta tio n o f s im u la tio n re s u lts in v a rio u s fo rm a ts , s u c h
a s te x t file s o r g ra p h ic a l d is p la y s fo r th e p u rp o s e o f o n -lin e v ie w in g a n d a n a ly s is .
W ith th e to o lb o x , d e s ig n e r m e re ly n e e d s to g iv e a m o d e l file re la tin g to h is /h e r
p a rtic u la r o p tim is a tio n p ro b le m , a n d c o n fig u re s th e p ro b le m b a s e d o n a fe w s im p le
G U I s e tu p s . F u rth e r d e s c rip tio n s o f th e to o lb o x a n d G U Is m a y b e re fe rre d to [1 1 ] o r
th e tu to ria ls in th e to o lb o x , w h ic h is fre e ly a v a ila b le fo r d o w n lo a d in g a t
h ttp ://w e b .s in g n e t.c o m .s g /~ k a y c h e n /m o e a .h tm .
186
K.C. Tan, T.H. Lee, and E.F. Khor
Setting/modifing
objective functions
Design performance
Decision-making
Module
Goals and priorities
QFT De sign
Spe c if ic at io ns
Tracking
bounds
Coupling
bounds
Robust
margin
Sensitivity
rejection
High freq.
gain
Objective
vector
Multi-objective function evaluation
Robust
stability
Multi-objective
Optimisation
Module
Test signals, weighting functions
Results
Gragh ic al
Displ ays
QFT design
parameters
System
response
QFT Control Module
R(s)
F(s)
G (s)
Plant
Templates
Y (s)
...
Fig. 7 . A general evolutionary design automated QFT control framework
3.
A Be nc h mark M I M O
QFT De sign Pro b l e m
The benchmark MIMO QFT control problem given in [14] is studied in this section,
which is shown in Fig. 1 with the MIMO uncertain plant sets given as,
3 0.5a º
ª a
«/ ( s )
/ (s) »
P ( s) «
(12)
»
8 »
« 1
«
/ (s) »
¬/ ( s )
¼
where / ( s ) s 2 0.03as 10 and a [6, 8] . Apart from the few design speciciations
studied by [14], additional performance requirements such as robust tracking and
cross-coupling specifications are included here for wider consideration of the QFT
design objectives, which subsequently adds to the design difficulty and complexity.
The specification of high frequency gain [9, 10] is also incorporated to avoid any
high-frequency sensor noise and unmodelled high-frequency dynamics/harmonics.
The various closed-loop performance requirements for this MIMO QFT design are
formulated as follows:
(i)
Robust Tracking Bounds for diagonal transfer functions:
T (i ,i ) L (Z ) d CL(i ,i ) ( jZ ) d T(i,i )U (Z ) , for i = 1, 2
(13)
Upper Tracking Model:
T(1,1)U (Z )
T( 2,2)U (Z )
1.9 u104 ( jZ ) 6.4 u105
( jZ )3 2.3 u102 ( jZ )2 1.9 u104 ( jZ ) 6.4 u105
6.4 u103 ( jZ ) 3.4 u105
( jZ ) 3 1.5 u102 ( jZ ) 2 8 u103 ( jZ ) 3.4 u105
Lower Tracking Model:
(14a)
(14b)
Automatic Design of Multivariable QFT Control System
1 u10 6
T(1,1) L (Z )
( jZ ) 3 3 u10 2 ( jZ ) 2 3 u10 4 ( jZ ) 1 u10 6
2.5 u105
T(2,2) L (Z )
( jZ )3 2.3 u102 ( jZ )2 1.5 u104 ( jZ ) 2.5 u105
(ii) Robust Cross-Coupling Bounds for off-diagonal transfer functions:
CL(i , j ) ( jZ ) d T(i , j )U (Z ) ,for i j, and i, j = 1,2
187
(15a)
(15b)
(16)
where,
T(1, 2 )U (Z )
0 .0032 ( jZ )
>0.016 ( jZ ) 1@ >0.016 ( jZ ) 1@
6.3 u 10 3 ( jZ )
>0.016 ( jZ ) 1@ >0.016 ( jZ ) 1@
T( 2,1)U (Z )
(iii) Robust Sensitivity Rejections for full matrix transfer functions:
S i , j ( jZ ) ai , j ( jZ ) , for Z 10
(17a)
(17b)
(18)
where, ai,j = 0.01w , for i = j; ai,j = 0.005w , for i j
(iv) Robust Stability Margin:
1
1.8
1 Li ,i ( jZ )
,for " i 1,2 , and Z ! 0
(19)
The performance bounds of QFT are computed within a wide frequency range of
10-2 rad/s to 103 rad/s. Without loss of generality, the structure of the diagonal
controller G (s) is chosen in the form of a general transfer function [9] as given by,
4
¦b s
m
¦a s
n
m
G
i ,i
( s)
, bm , a n , for i = 1, 2
m 0
4
(20)
n
n 0
Note that the controller can also be designed by refining position of poles and zeros
directly or by using other structures such as the realisable (non-ideal) PID structure if
desired. The filter is fixed to a full matrix first-order transfer function as it is relevant
to the tracking and cross-coupling bound in the frequency response. Since the
resultant pre-filter must satisfy lim[ F ( s )] 1 for a step forcing function [9], the
so 0
structure of pre-filter F(s) is chosen as a full matrix first-order transfer function as
given by,
1
cn , for " i, j = 1, 2
(21)
Fi , j ( s )
2
n
cn s
1
¦
j n
188
K.C. Tan, T.H. Lee, and E.F. Khor
A p a rt fro m m o st d e fa
w ith a p o p u la tio n a n d g e n
e v o lu tio n a ry Q F T d e s ig n
p e rfo rm a n c e re q u ire m e n ts
d e te rm in a tio n o f th e g o a l a
p e rfo rm a n c e re q u ire m e n ts ,
c o m m itm e n t’ d e s ig n [9 ].
p e rfo rm a n c e s p e c ific a tio n s
e v o lu tio n a ry o p tim is a tio n a
u lt s e ttin g s , th e e v o lu tio
e ra tio n s iz e o f 2 0 0 a n d
o p tim is a tio n p ro c e s s , g
m a y b e in c lu d e d o p tio n a
n d p rio rity m a y b e a s u b
it m a y b e u n n e c e s s a ry a n
In p rin c ip le , a n y n u
c a n b e a d d e d to th e
p p ro a c h if n e c e s s a ry .
n a ry to o lb o x h a s b e e n c o n fig u re
1 0 0 , re s p e c tiv e ly . T o g u id e th
o a l a n d p rio rity fo r e a c h o f th
lly a s s h o w n in F ig . 8 . A lth o u g
je c tiv e m a tte r a n d d e p e n d s o n th
d c a n b e ig n o re d fo r a ‘m in im u m
m b e r o r c o m b in a tio n o f Q F
d e s ig n u s in g th e m u lti-o b je c tiv
d
e
e
h
e
T
e
F ig . 8 . S e ttin g s o f th e M O E A to o lb o x fo r th e b e n c h m a rk Q F T d e s ig n p ro b le m
A p o w e rfu l fe a tu re o f th e e v o lu tio n a ry Q F T d e s ig n is th a t it a llo w s o n -lin e
e x a m in a tio n o f d iffe re n t tra d e -o ffs a m o n g th e m u ltip le c o n flic tin g s p e c ific a tio n s ,
m o d ific a tio n o f e x is tin g o b je c tiv e s a n d c o n s tra in ts , a n d z o o m in to a n y re g io n o f
in te re s t b e fo re s e le c tin g o n e fin a l s e t o f c o n tro lle r a n d p re -filte r fo r re a l tim e
im p le m e n ta tio n . T h e tra d e -o ff g ra p h o f th e re s u lta n t Q F T c o n tro l s y s te m is s h o w n in
F ig . 9 , w h e re e a c h lin e re p re s e n tin g a s o lu tio n fo u n d b y th e e v o lu tio n a ry o p tim is a tio n .
T h e c o s t o f o b je c tiv e s s u c h a s s ta b ility (R H S P ), ro b u s t tra c k in g a n d c ro s s c o u p lin g
p e rfo rm a n c e s (E R R U T a n d E R R L T ) a re la b e lle d a s o b je c tiv e s 1 -7 , w h ic h a re a ll e q u a l
to z e ro a s d e s ire d a c c o rd in g to th e g o a l s e ttin g s in F ig . 8 . T h e x -a x is s h o w s th e d e s ig n
s p e c ific a tio n s , th e y -a x is s h o w s th e n o rm a lis e d c o s t fo r e a c h o b je c tiv e a n d th e c ro s s m a rk s h o w s th e d e s ire d g o a l s e ttin g fo r e a c h p e rfo rm a n c e re q u ire m e n t. C le a rly , tra d e o ffs b e tw e e n a d ja c e n t s p e c ific a tio n s re s u lts in th e c ro s s in g o f th e lin e s b e tw e e n th e m ,
w h e re a s c o n c u rre n t lin e s th a t d o n o t a c ro s s e a c h o th e r in d ic a tin g th e s p e c ific a tio n s d o
n o t c o m p e te w ith o n e a n o th e r. F o r e x a m p le , th e ro b u s t s e n s itiv ity o b je c tiv e o f 1 2
(R S 2 1 ) a n d 1 3 (R S 2 2 ) a re n o t c o m p e tin g w ith e a c h o th e r, w h e re a s th e ro b u s t m a rg in
o b je c tiv e 8 (R M 1 ) a n d 9 (R M 2 ) a p p e a r to c o m p e te h e a v ily , a s e x p e c te d . T h e
in fo rm a tio n c o n ta in e d in th is tra d e -o ff g ra p h a ls o s u g g e s ts th a t lo w e r g o a l s e ttin g s fo r
ro b u s t s e n s itiv ity (o b je c tiv e s 1 0 -1 3 ) a re p o s s ib le , w h ic h c a n b e fu rth e r o p tim is e d to
a rriv e a t a n e v e n b e tte r ro b u s t p e rfo rm a n c e .
Automatic Design of Multivariable QFT Control System
189
F ig . 9 . T ra d e -o ff g ra p h o f th e e v o lu tio n a ry d e s ig n e d Q F T c o n tro l s y s te m
N o te th a t th e e v o lu tio n a ry Q F T d e s ig n a ls o a llo w s e n g in e e rs to d iv e rt th e
e v o lu tio n to o th e r fo c u s e d tra d e -o ff re g io n o r to m o d ify a n y p re fe re n c e s o n th e
c u rre n t s p e c ific a tio n s e ttin g s a fte r o b s e rv a tio n fo r a n u m b e r o f g e n e ra tio n s . F o r
e x a m p l e , t h e d e s i g n e r c a n c h a n g e h i s p r e f e r e n c e a n d d e c i d e t o r e d u c e t h e 9 th g o a l
s e ttin g fo r r o b u s t m a r g in ( R M 2 ) fr o m 1 .8 to 1 .3 . F ig . 1 0 illu s tr a te s th e b e h a v io u r o f
th e e v o lu tio n u p o n th e m o d ific a tio n o f th is g o a l s e ttin g a fte r th e e v o lu tio n a ry Q F T
d e s ig n in F ig . 9 . D u e to th e s u d d e n c h a n g e o f a tig h te r g o a l s e ttin g , in itia lly n o n e o f
th e in d iv id u a ls m a n a g e to m e e t a ll th e re q u ire d s p e c ific a tio n s a s s h o w n in F ig . 1 0 (a ).
A fte r c o n tin u in g th e e v o lu tio n fo r 2 g e n e ra tio n s , th e p o p u la tio n m o v e s to w a rd s
s a tis fy in g th e o b je c tiv e o f R M 2 a s s h o w n in F ig . 1 0 (b ) a t th e p e rfo rm a n c e e x p e n s e o f
o th e r o b je c tiv e s s in c e th e y a re h ig h ly c o rre la te d a n d c o m p e tin g to e a c h o th e r. T h e
e v o lu tio n c o n tin u e s a n d a g a in le a d s to th e s a tis fa c tio n o f a ll th e re q u ire d g o a l s e ttin g s
in c lu d in g th e s tric te r s e ttin g o f o b je c tiv e R M 2 a s s h o w n in F ig . 1 0 (c ). C le a rly , th is
m a n -m a c h in e in te ra c tiv e d e s ig n a p p ro a c h h a s e n a b le d Q F T d e s ig n e rs to d iv e rt th e
e v o lu tio n in to a n y in te re s te d tra d e -o ff re g io n s o r to m o d ify c e rta in s p e c ific a tio n s a n d
p re fe re n c e s o n -lin e , w ith o u t th e n e e d o f re s ta rtin g th e e n tire d e s ig n p ro c e s s a s
re q u ire d b y c o n v e n tio n a l Q F T d e s ig n m e th o d s .
N o rm a lis e d c o s ts
1
0 .8
O n -lin e
M o d ific a tio n
0 .6
0 .4
0 .2
0
2
4
6
8
1 0
1 2
1 4
O b je c tiv e s
(a ) O n -lin e g o a l m o d ific a tio n o f ro b u s t m a rg in o b je c tiv e (R M 2 )
190
K.C. Tan, T.H. Lee, and E.F. Khor
1
1
N o rm a liz e d c o s ts
N o rm a liz e d c o s ts
0 .8
0 .6
0 .4
0 .2
0 .8
0 .6
0 .4
0 .2
0
0
2
4
6
8
1 0
1 2
2
1 4
4
6
8
1 0
1 2
1 4
O b je c tiv e s
O b je c tiv e s
(b ) A fte r 2 g e n e ra tio n s
(c ) A fte r a n o th e r 2 g e n e ra tio n s
F ig . 1 0 . E ffe c ts o f th e e v o lu tio n u p o n th e o n -lin e m o d ific a tio n o f g o a l s e ttin g
F ig . 1 1 s h o w s th e ro b u s t tra c k in g p e rfo rm a n c e s in th e fre q u e n c y d o m a in fo r th e
tw o d ia g o n a l e le m e n ts o f th e c lo s e d -lo o p s y s te m . It c a n b e s e e n th a t a ll th e fre q u e n c y
re s p o n s e s o f C L U a n d C L L fo r b o th th e d ia g o n a l c h a n n e ls a re lo c a te d s u c c e s s fu lly
w ith in th e ir re s p e c tiv e p re -s p e c ifie d tra c k in g b o u n d s o f T U a n d T L . B e s id e s , th e
c o u p lin g e ffe c t fro m th e o ff-d ia g o n a l e le m e n ts o f th e c lo s e d -lo o p s y s te m fo r a ll th e
p la n t te m p la te s h a s a ls o b e e n re d u c e d s a tis fa c to ry a n d s u c c e s s fu lly b o u n d e d b y th e
u p p e r c o u p lin g b o u n d w ith m in im a l g a in a n d b a n d w id th o f th e o ff-d ia g o n a l tra n s fe r
fu n c tio n s a s s h o w n in F ig . 1 2 .
1 0
1 0
T
( 1 ,1 )U
0
T
0
( 2 ,2 ) U
-1 0
C L
-2 0
M a g n itu b e (d B )
M a g n itu b e (d B )
-1 0
( 1 ,1 )
-3 0
T
( 1 ,1 ) L
-4 0
-5 0
C L
( 2 ,2 )
T
( 2 ,2 ) L
-3 0
-4 0
-5 0
-6 0
-7 0
-6 0
-7 0
-2 0
-8 0
1 0
-2
1 0
-1
1 0
0
1 0
1
1 0
2
F re q u e n c y (ra d /s e c )
( a ) D ia g o n a l e le m e n t o f ( 1 ,1 )
1 0
3
-9 0
1 0
-2
1 0
-1
1 0
0
1 0
1
1 0
2
F re q u e n c y (ra d /s e c )
( b ) D ia g o n a l e le m e n t o f ( 2 ,2 )
F ig . 1 1 . T h e tra c k in g p e rfo rm a n c e in th e fre q u e n c y d o m a in
1 0
3
Automatic Design of Multivariable QFT Control System
0
-1 0
-1 0
-2 0
T
( 1 ,2 )U
C L
-4 0
( 1 ,2 )
-5 0
-6 0
-7 0
-7 0
-8 0
-9 0
-1 0 0
1 0
-1
1 0
0
1 0
1
1 0
2
1 0
( 2 ,1 )
-6 0
-9 0
-2
C L
-5 0
-8 0
1 0
( 2 ,1 ) U
-4 0
M a g n itu b e (d B )
-3 0
T
-3 0
-2 0
M a g n itu b e (d B )
191
3
1 0
-2
1 0
-1
1 0
0
1 0
1
1 0
2
1 0
3
F re q u e n c y (ra d /s e c )
F re q u e n c y (ra d /s e c )
( a ) O ff -d ia g o n a l e le m e n t o f ( 1 ,2 )
( b ) O ff-d ia g o n a l e le m e n t o f ( 2 ,1 )
F ig . 1 2 . T h e c ro s s -c o u p lin g p e rfo rm a n c e in th e fre q u e n c y d o m a in
F ig s . 1 3 a n d 1
la n t te m p la te s
n e d c o n tro lle r
rm a n c e s h a v e
in g b o u n d s , a s
th e p
d e s ig
p e rfo
tra c k
4 sh o
in th
a n d p
b e e n
d e s ire
w
th e u n
e tim e d
re -filte r.
s a tis fie d
d .
it s te p
o m a in
C le a rly
su c c e
tra
fo
, a
ssf
c k in
r a
ll th
u lly
g a n d c o u p lin g p e rfo rm
ra n d o m s e le c te d s e t o f
e tim e d o m a in tra c k in g
a n d w ith in th e re q u ire
a n c
e v o
a n d
d p
e s fo r a ll
lu tio n a ry
c o u p lin g
re s c rib e d
1 .4
1 .4
T
1 .2
T
( 1 ,1 )U
1 .2
( 2 ,2 ) U
1
1
0 .8
C L
0 .6
T
0 .4
M a g n itu d e
M a g n itu d e
0 .8
( 1 ,1 )
( 1 ,1 )L
C L
( 2 ,2 )
0 .4
T
0 .2
0 .2
0
0 .6
( 2 ,2 )L
0
0
0 .0 2
0 .0 4
0 .0 6
0 .0 8
0 .1
0 .1 2
0 .1 4
0 .1 6
T im e (s e c )
( a ) D ia g o n a l e le m e n t o f ( 1 ,1 )
0 .1 8
0 .2
0
0 .0 2
0 .0 4
0 .0 6
0 .0 8
0 .1
0 .1 2
0 .1 4
0 .1 6
0 .1 8
T im e (S e c )
( b ) D ia g o n a l e le m e n t o f ( 2 ,2 )
F ig . 1 3 . T h e tra c k in g re s p o n s e s o f th e d ia g o n a l e le m e n ts in th e c lo s e d -lo o p s y s te m
0 .2
192
K.C. Tan, T.H. Lee, and E.F. Khor
0 .1 5
0 .0 6
0 .1
0 .0 4
0 .0 5
0 .0 2
0
- 0 .0 5
M a g n itu d e
M a g n itu d e
0
- 0 .1
- 0 .1 5
- 0 .2
C L
- 0 .0 2
- 0 .0 4
- 0 .0 6
( 1 ,2 )
- 0 .2 5
C L
( 2 ,1 )
- 0 .0 8
- 0 .3
- 0 .3 5
0
0 .0 2
0 .0 4 0 .0 6
0 .0 8
0 .1
0 .1 2
0 .1 4
0 .1 6
0 .1 8
- 0 .1
0 .2
0
0 .0 2
0 .0 8
0 .0 4 0 .0 6
T im e (S e c )
0 .1
0 .1 2
0 .1 4
0 .1 6
0 .1 8 0 .2
T im e (s e c )
( a ) O ff -d ia g o n a l e le m e n t o f ( 1 ,2 )
( b ) O ff-d ia g o n a l e le m e n t o f ( 2 ,1 )
F ig . 1 4 . T h e c o u p lin g re s p o n s e s o f th e o ff-d ia g o n a l e le m e n ts in th e c lo s e d -lo o p
s y s te m
T o illu s tra te ro b u s tn e s s o f th e e v o lu tio n a ry
d is tu rb a n c e re je c tio n , a u n it s te p d is tu rb a n c e s ig n a l
O u tp u t d is tu rb a n c e re s p o n s e s fo r a ll th e fin a l P
e le m e n t o f th e c lo s e d -lo o p tra n s fe r m a trix a re illu s
s te p d is tu rb a n c e h a s b e e n s u c c e s s fu lly a tte n u a te
d iffe re n t v a lu e s o f p a ra m e te r u n c e rta in tie s , a s
s p e c ific a tio n o f ro b u s t s e n s itiv ity re je c tio n .
.
h
it
e
e
0 .3
0 .2
0 .6
0 .1
M a g n itu d e
M a g n itu d e
n
0 .4
1
0 .8
0 .4
Y
0 .2
( 1 ,1 )
D
0
- 0 .1
- 0 .2
- 0 .3
0
Y
- 0 .4
- 0 .2
- 0 .4
d e s ig n e d Q F T c o n tro l s y s te m o
w a s a p p lie d to th e M IM O s y s te m
a re to o p tim a l c o n tro lle rs a t e a c
tra te d in F ig . 1 5 . C le a rly , th e u n
d to z e ro e v e n tu a lly fo r a ll th
q u a n tifie d b y th e p e rfo rm a n c
( 1 ,2 )
D
0
- 0 .5
0 .0 2
0 .0 4
0 .0 6
0 .0 8
0 .1
0 .1 2
0 .1 4
0 .1 6
0 .1 8
0 .2
T im e (s e c )
(a ) D is tu rb a n c e re s p o n s e o f e le m e n t (1 , 1 )
- 0 .6
0
0 .0 2
0 .0 4
0 .0 6
0 .0 8
0 .1
0 .1 2
0 .1 4
0 .1 6
0 .1 8
0 .2
T im e (s e c )
(b ) D is tu rb a n c e re s p o n s e o f e le m e n t (1 , 2 )
Automatic Design of Multivariable QFT Control System
193
1
0 .1 5
0 .1
0 .0 5
- 0 .0 5
Y
( 2 ,1 )
( 2 ,2 )
D
0
D
- 0 .1
Y
M a g n itu d e
M a g n itu d e
0 .5
0
- 0 .1 5
- 0 .2
0
0 .0 2
0 .0 4
0 .0 6 0 .0 8
0 .1
0 .1 2
0 .1 4 0 .1 6
0 .1 8
0 .2
- 0 .5
0
0 .0 2
0 .0 4
0 .0 6
0 .0 8
(c ) D is tu rb a n c e re s p o n s e o f e le m e n t (2 , 1 )
0 .1
0 .1 2
0 .1 4 0 .1 6
0 .1 8 0 .2
T im e (s e c )
T im e (s e c )
(d ) D is tu rb a n c e re s p o n s e o f e le m e n t (2 , 2 )
F ig . 1 5 . O u tp u t re s p o n s e s fo r th e u n it s te p d is tu rb a n c e in th e tim e d o m a in
4
C o n c lu s io n
T h is p a p e r h a s a n a ly s e d d iffic u ltie s in e x is tin g Q F T d e s ig n te c h n iq u e s fo r
m u ltiv a ria b le c o n tro l s y s te m s . T o a d d re s s th e s e d e s ig n d e fic ie n c ie s , a n a u to m a te d
m u ltiv a ria b le Q F T d e s ig n m e th o d o lo g y u s in g a h ig h p e rfo rm a n c e M O E A to o lb o x h a s
b e e n p r o p o s e d . U n l i k e e x i s t i n g d e s i g n m e t h o d s , t h e ’i n t e l l i g e n t ’ d e s i g n a p p r o a c h i s
c a p a b le o f a u to m a tic a lly e v o lv in g b o th n o m in a l c o n tro lle r a n d
p re -filte r
s im u lta n e o u s ly to m e e t a ll p e rfo rm a n c e re q u ire m e n ts in Q F T , w ith o u t g o in g th ro u g h
th e s e q u e n tia l d e s ig n s ta g e s fo r e a c h o f th e m u ltiv a ria b le s u b -s y s te m s . B e s id e s , th e
a p p ro a c h a ls o a v o id s th e n e e d o f m a n u a l Q F T b o u n d c o m p u ta tio n a n d tria l-a n d -e rro r
lo o p -s h a p in g p ro c e d u re s a s re q u ire d b y c o n v e n tio n a l m e a n s . It is s h o w n th a t c o n tro l
e n g in e e rs ’ e x p e rtis e s a s w e ll a s g o a l a n d p rio rity in fo rm a tio n c a n b e e a s ily in c lu d e d
a n d m o d ifie d o n -lin e a c c o rd in g to th e e v o lv in g tra d e -o ffs , in s te a d o f re p e a tin g o r
re s ta rtin g th e w h o le d e s ig n p ro c e s s . It is o b v io u s th a t th e p ro p o s e d e v o lu tio n a ry Q F T
d e s ig n fra m e w o rk is fu lly e x p a n d a b le to o n -lin e d e s ig n o p tim is a tio n a n d
im p le m e n ta tio n . T h is c a n b e re a lis e d e ith e r v ia th e h a rd a n d s o ftw a re s y s te m s s u c h a s
d S P A C E [1 5 ] o r M IR C O S [1 6 ] fo r g ra p h ic a l p ro g ra m m in g a n d re a l-tim e o p e ra tio n to
p ro v id e n e c e s s a ry lin k a g e s b e tw e e n th e to o lb o x a n d th e p h y s ic a l e n v iro n m e n ts . A p a rt
fro m th e d e v e lo p m e n ts fo r o n -lin e a d a p ta tio n , th e m u lti-o b je c tiv e e v o lu tio n a ry Q F T
d e s ig n p a ra d ig m is c u rre n tly b e in g e x te n d e d to ro b u s t c o n tro l o f n o n lin e a r s y s te m s
a n d to in c o rp o ra te o th e r d e s ig n s p e c ific a tio n s s u c h a s e c o n o m ic a l c o s t c o n s id e ra tio n .
P ro g re s s a n d re s u lts w ill b e re p o rte d in d u e c o u rs e .
R e fe r e n c e s
1 .
2 .
3 .
Y a n iv , O ., H o r o w itz , I .: A Q u a n tita tiv e d e s ig n m e th o d f o r M
s y s te m h a v in g u n c e rta in p la n ts . In t. J . C o n tr o l, v o l. 4 3 , n o . 2
Y a n iv , O ., S c h w a r tz , B .: A C r ite r io n fo r lo o p s ta b ility in th
o f M IM O fe e d b a c k s y s te m s . In t. J . C o n tr o l, v o l. 5 3 , n o . 3 , p
H o u p is , C . H .: Q u a n tita tiv e fe e d b a c k th e o r y ( Q F T ) te c h n iq
e d . (1 9 9 6 ). T h e C o n tr o l H a n d b o o k , C R C P re s s & IE E E P re s
IM O lin e a r fe e d b a c k
, p p . 4 0 1 -4 2 1 , 1 9 8 6 .
e H o ro w itz S y n th e s is
p . 5 2 7 -5 3 9 , 1 9 9 0 .
u e . I n L e v in e , W . S .,
s, p p . 7 0 1 -7 1 7 , 1 9 9 3 .
194
4 .
5 .
6 .
7 .
8 .
9 .
1 0 .
1 1 .
1 2 .
1 3 .
1 4 .
1 5 .
1 6 .
K.C. Tan, T.H. Lee, and E.F. Khor
S n e ll, S . A ., H e s s , R . A .: R o b u s t, d e c o u p le d , flig h t c o n tr o l d e s ig n w ith r a te
s a tu ra tin g a c tu a to rs . C o n f. a n d E x h ib it. O n A IA A A tm o s p h e r ic F lig h t M e c h a n ic s ,
p p . 7 3 3 -7 4 5 , 1 9 9 7 .
T h o m p s o n , D . F ., a n d N w o k a h , O . D . I .: A n a ly tic a l lo o p -s h a p in g m e th o d s in
q u a n tita tiv e fe e d b a c k th e o ry , J . D y n a m ic S y s te m s , M e a s u r e m e n t a n d C o n tr o l,
v o l. 1 1 6 , p p . 1 6 9 -1 7 7 , 1 9 9 4 .
B r y a n t, G . F ., a n d H a lik ia s , G . D .: O p tim a l lo o p -s h a p in g fo r s y s te m s w ith la r g e
p a ra m e te r u n c e rta in ty v ia lin e a r p ro g ra m m in g , In t. J . C o n tr o l, v o l. 6 2 , n o . 3 , p p .
5 5 7 -5 6 8 , 1 9 9 5 .
C h a it, Y .: Q F T lo o p -s h a p in g a n d m in im is a tio n o f th e h ig h -f r e q u e n c y g a in v ia
c o n v e x o p tim is a tio n , P r o c . S y m . Q u a n tita tiv e F e e d b a c k T h e o r y a n d o th e r F r e q .
D o m a in M e th o d a n d A p p lic a tio n s , G la s g o w , S c o tla n d , p p . 1 3 -2 8 , 1 9 9 7 .
C h e n , W . H ., B a lla n c e , D . J . L i, Y .: A u to m a tic lo o p -s h a p in g in Q F T u s in g
g e n e tic a lg o r ith m s . P r o c . o f 3 rd A s ia - P a c ific C o n f. o n C o n t. & M e a s ., p p . 6 3 -6 7 ,
1 9 9 8 .
T a n , K . C ., L e e , T . H . K h o r , E . F .: C o n tr o l s y s te m d e s ig n a u to m a tio n w ith r o b u s t
tra c k in g th u m b p rin t p e rfo rm a n c e u s in g a m u lti-o b je c tiv e e v o lu tio n a ry a lg o rith m " ,
I E E E I n t . C o n f . C o n t r o l A p p l . a n d S y s . D e s i g n , H a w a i i , 2 2 - 2 6 th A u g u s t , p p . 4 9 8 5 0 3 , 1 9 9 9 .
C h e n , W . H ., B a lla n c e , D . J ., F e n g , W ., a n d L i, Y .: G e n e tic a lg o r ith m e n a b le d
c o m p u te r-a u to m a te d d e s ig n o f Q F T c o n tro l s y s te m s , IE E E In t. C o n f. C o n tr o l
A p p l . a n d S y s . D e s i g n , H a w a i i , 2 2 - 2 6 th A u g u s t , p p . 4 9 2 - 4 9 7 , 1 9 9 9 .
T a n , K . C ., W a n g , Q . G ., L e e , T . H ., K h o o , T . T ., a n d K h o r , E . F .: A M u ltio b je c tiv e
E v o lu tio n a r y
A lg o r ith m
T o o lb o x
fo r
M a tla b ,
( h ttp ://v la b .e e .n u s .e d u .s g /~ k c ta n /m o e a .h tm ) , 1 9 9 9 .
S k o g e s ta d , S ., P o s tle th w a ite , I .: M u ltiv a r ia b le F e e d b a c k C o n tr o l: A n a ly s is a n d
D e s ig n . J o h n W ile y & S o n s L td , W e s t S u s s e x . E n g la n d , 1 9 9 6 .
T h e M a th W o r k s , I n c .: U s in g M A T L A B , v e r s io n 5 , 1 9 9 8 .
B o r g h e s a n i, C ., C h a it, Y . a n d Y a n iv , O .: Q u a n tita tiv e F e e d b a c k T h e o r y T o o lb o x
U s e r M a n u a l, T h e M a th W o rk In c , 1 9 9 5 .
H a n s e lm a n n , H .: A u to m o tiv e c o n tr o l: F r o m c o n c e p t to e x p e r im e n t to p r o d u c t,
IE E E In t. C o n f. C o n tr . A p p l. a n d S y s . D e s ., D e a rb o rn , 1 9 9 6 .
R e b e s c h ie ß , S .: M I R C O S - M ic r o c o n tr o lle r -b a s e d r e a l tim e c o n tr o l s y s te m
to o lb o x fo r u s e w ith M a tla b /S im u lin k , IE E E In t. C o n f. C o n tr . A p p l. a n d S y s .
D e s ig n , H a w a ii, U S A , p p . 2 6 7 -2 7 2 , 1 9 9 9 .
p
H
t
s r
a ng ,H Z ng ,K I
h
i
tta l
l,Z
r r s r r
r
p r t r r icti
icha rdso n ,
i o nso n ,a nd
H
pa rtm nt f lctrica l ngin ringa nd lctr nics
ni rsit f i rp l
, i rp l
, 6 3 , . .
q.h.wu@liv.ac.uk
lctric
r s a rch Instit t
ingh , ijing
5, . . hina
ngin ring a nd
chn lg
h N a tina l rid
mpa n pl
c, . .
c . his pa p r d scrib s a n
th rma lm d l f il
-imm rs d,
f rc d-a ir c ld p
r tra nsf rm rs a nd a m th d lg f r m d lc nstr ctin singint l
l
ig ntla rninga ppl
id t n-sit m a s r m nts. h
m d ld l
i rs th a l fb tt m- ila nd t p- ilt mp ra t r s f r th rma lp rf rma nc pr dictin a nd n-l
in m nit ring fp
r tra nsf rmrs. h r s l
ts bta in d sing th n
th rma lm d la r c mpa r d
ith th r s l
ts fa tra ditina lth rma lm d la nd th r s l
ts d ri d
fr m a rtificia ln ra ln t rks.
I tr
cti
n-l
in o nito ringo fpo r tra nsfo r rso p ns th po ssi il
it fo r t ndingth
o p ra tingti o fpo rtra nsfo r rs,r d cingth risko f p nsi fa ilr s a nd
pro idingpo t ntia lfo r cha ngingth a int na nc stra t g [ ] [3]. h s f l
l
if o fa tra nsfo r r is d t r in d in pa rt th a il
it o fth tra nsfo r r to
dissipa t th int rna l
l g n ra t d h a t to its s rro ndings. o ns q ntl, th
co pa riso n o fa ct a la nd pr dict d o p ra ting t p ra t r s ca n pro id a s nsiti
a s r o ftra nsfo r r co nditio n a nd ightindica t a no r a lo p ra tio n.
od l
ing tra nsfo r r th r a ld na ics is r ga rd d a s o n o fth o sti po rta ntiss sa nd co nstr ctio n o fa n a cc ra t th r a l o d lisa n i po rta nta sp ct
o ftra nsfo r rco nditio n o nito ring. h g n ra l
la cc pt d
tho ds [5][6 ],ca n
s d to pr dict o n s o f c ss t p ra t r in a tra nsfo r r. H o
r, th
co n ntio na lca l
c l
a tio n o fint rna ltra nsfo r r t p ra t r is no to nl a co pl
ica t d a nd diffic l
tta sk ta l
so la dsto a co ns r a ti sti a t a s d o n so
a ss ptio ns o fth o p ra ting co nditio ns. Its a il
it to pr dictth tra nsfo r r
t p ra t r nd r r a l
istic l
o a dingco nditio ns is th r fo r so
ha tl
i it d.
In this pa p r, t o diff r ntint l
l
ig ntla rning tho ds, g n tic a l
go rith
( ) a nd a rti cia ln ra ln t o rk (
), a r s d to co nstr ctth r a l o dl
s fro th o n-sit
as r
nts. h st d sho s tha t int l
l
ig nt la rning
tho ds ca n pr dicto nl
in tra nsfo r r t p ra t r s in r a lti
ith gr a t r
S . C a g n o n i e t a l. ( E d s .) : E v o W o r k s h o p s 2 0 0 0 , L N C S 1 8 0 3 , p p . 1 9 5 − 2 0 4 , 2 0 0 0 .
S p rin g e r-V e rla g B e rlin H e id e lb e rg 2 0 0 0
1 9 6
W .H . T a n g e t a l.
a cc ra c tha n tha to ta in d singth tra ditio na l o d l
s. h d l
o p d th ra l o d lco l
d
s d a s th a sis o fa n int l
l
ig ntpro t ctio n s st ,a s l
l
a s a n l ntin a r o t s p r isinga nd co ntro ls st .
r
s
o nsid ring th o p ra ting r gi s o fco o l
ing s st s a nd th a ct a l a s r nts a a il
a l, t o i po rta ntt p ra t r
as r
nts, o tto -o ilt p rat r (
) a nd to p-o ilt p ra t r (
), a r cho s n fo r th p rpo s o f
tra nsfo r r co nditio n o nito ring. In th fo l
l
o ing,t o th r a l o d l
s fo r t p ra t r pr dictio n a nd co nditio n o nito ringo ftra nsfo r rs a r d scri d.
.
-s
s
o r np p d (o r ‘na t ra l
’
) o ilco o l
ing o f po r tra nsfo r rs ( ), th o il
t p ra t r a tth to p o f indings is a ppro i a t l q a lto th
insid
th ta nk. H o
r, fo r fo rc d o ilcirc l
a tio n ( ), th
is th s
o fth
o ilt p ra t r a tth o tto o fth inding,
,a nd th diff r nc
t n
o ilt p ra t r s a tth to p a nd o tto o fth inding[4] [7 ].
h
is d scri d a s fo l
l
o s:
=
a nd th
is r pr s nt d
=
+
+
+
()
+
:
+
+
+ (t−
)
(2)
hr
= a intt p ra t r
= st a d sta t o ilt p ra t r a t th o tto o fth inding ith th
o p ra tingl
o ad
= o ilt p ra t r ris a o a inta tth o tto o fth inding nd r
th ra t d l
o ad
= th ra tio o fl
o ad l
o ss (a tra t d l
o a d) to no -l
o ad l
o ss
= th ra tio o fth o p ra tingl
o a d c rr ntto ra t d l
o a d c rr nt, =
= st a d sta t o il
t p ra t r a tth to p o fth inding ithth o p ra ting
l
o ad
lt p ra t r ris a o a inta tth to p o fth inding nd r th
t = oi
ra t d l
o ad
= po n ntr l
a t d to o ilt p ra t r ris d to to ta ll
o ss s
= po n ntr l
a t d to indings t p ra t r ris d to th l
o a d c rr nt.
l
lth pa ra t rs in o l d in this o d l
,s cha s , a nd ,a r
in d thro gh p ri nto r
p rinc .
s al
ld t r-
D e v e lo p m e n t o f P o w e r T ra n s fo rm e r T h e rm a l M o d e ls
.
r
si
-s
s
pp ic
1 9 7
p r i
In co ntra stto th indingt p ra t r , th tra nsint
a nd
( a nd
r
sp
ct
i
l
)
ca
nno
ti
di
a
t
l
r
a
ch
t
h
co
rr
spo
ndi
ng
st
a
d
st
at al
t
s nd r cha nging l
o a ds, sinc th ir th r a lti co nsta nts a r in th o rd r
o fho rs [ ].
cho o s a r c rsi fo r o fth o d l sing th pr io s sa pls to r pr s ntth
a nd
r sp cti l, hich ca n r fl ctcha ng s o f
th th r a lti co nsta nts d to diff r nto p ra ting co nditio ns o fth po r
tra nsfo r r.
a n a ssist d o ilco o l
ingo ftra nsfo r rs is a ctia t d a to a tica l
l a cco rding
to th o ilt p ra t r . h n th o ilt p ra t r incr a s sa nd c dsa c rta in
a l , ,fa ns il
l s itch d o n. o disting ishth diff r ntth r a ld na ics
a ppro pria t to th p rio ds h n th fa ns a r o n a nd h n th a r o ff, a t o pic
o d lis intro d c d. h fo l
l
o ing o d lis pl
o d to pr dictth
a nd
tha tca n
pr ss d
th co
ina tio n o fa r c rsi
o d la nd a
fa n f nctio n.
h
a tti insta nt is:
()=
( − 2) +
+ A
+
( − 2) +
( − )+
( − 2)
+
( − 2) +
( − 2) +
()
(3)
hr
()>
o th r is
()=
h
is d scri d a s:
t( )
=
t
( − 2)( ( − 2) −
+ At t ( − ) +
t t(
)+
t
− 2) +
+
t
t
+
( − 2)
t
()
t
(4)
hr
()
t
=
t+
(
t
−
n− n)
t
t( ) >
o th r is
In th o d l(3) (4),th co fficints iths script‘ ’d no t tho s a sso cia t d ith
a nd ‘ ’ ith
. a cht p ra t r is d scri d
a t o -pic
od l
. h co fficints in th o d l
s a r sho n in a l , in hich o n s to f
co fficints a r a sso cia t d ith th no r a lco nditio n itho tfa n o p ra tio n a nd
th o th r a r r l
a t d to th ti p rio d d ring hich fa ns a r s itch d o n.
ig
.
ic
I t ig t
ri
r ig
s
is a po rf ln
rica lopti ia tio n t chniq , hichis ro o t d in th
cha nis o f o ltio n a nd na t ra l
g n tics.
sd ri th irstr ngths si l
a ting
1 9 8
W .H . T a n g e t a l.
th na t ra ls a rcha nd s lctio n pro c ss a sso cia t d ith na t ra lg n tics. s t
o fg n s in its ‘chro o so s’d t r in s
r o rga nis s id ntit, hich is r frr d to a s a ‘string’in
. a t ra ls lctio n ta k s pl
a c in s ch a a tha t
th s cha ra ct ristics a r i pl
icitl s lct d ia th s r ia lo fth tt stcrit rio n. his a l
go rith , th
o stpo p l
a r fo r a to f hich is th ina r g n tic
al
go rith ,sta rts iths ttingo jcti f nctio ns a s d o n th ph sica l o d lo f
pro l s to ca l
c l
a t tn ss a l s, a nd th r a ft r a s r a ch ina r co d d
string’
s str ngth ith its tn ss a l . h stro ng r strings a d a nc a nd a t
ith o th r stro ng r strings to pro d c o ffspring. ina l
l, th
sts r i s. n
o fth i po rta nt a d a nta g s is tha t
co l
d
a l to nd o t th gl
o al
ini
o f tn ss inst a d o fa l
o ca lso ltio n. h
ha s ga in d po p l
a rit
in r c nt a rs a s a ro sto pti ia tio n to o lfo r a a rit o fpro l s in ngin ring,scinc , co no ics, na nc , tc.[9 ] [ ]. h
il
lt pica l
l pl
o
thr o p ra to rs: r pro d ctio n, cro sso r a nd
ta tio n [ 2]. pica l
l, a ch o f
th s o p ra to rs is a ppl
id to th po p l
a tio n o nc p r g n ra tio n, a nd s a l
l
s ra lg n ra tio ns a r r q ir d to a chi sa tisfa cto r r s l
ts.
pro d ctio n isa pro c ssin hicha n o l
d stringisca rrid thro ghinto a n
po p l
a tio n d p ndingo n its p rfo r a nc ind (i. . tn ss) a l . ho s strings
ith high r tn ss a l s t nd to ha a high r pro a il
it o fco ntri tingo n
o r o r o ffsprings to th n tg n ra tio n.
h cro sso r o p ra to r pro id s ra ndo info r a tio n cha ng . Itis a i d
to a rds o ling tt r il
ding l
o cks (sch a ta ith sho rtd ning lngths
a nd high a ra g tn ss a l s). ro sso r po ints a r ra ndo l cho s n. h
fr q nc o fcro sso ris go rn d
a s r s lct d cro sso rra t o rpro a il
it
o fcro sso r.Incr a singcro sso rra t incr a s sr co
ina tio n o f il
ding l
o cks,
t ith a n incr a singpro a il
it o fl
o o singgo o d strings.
h
ta tio n o p ra to r is si pla n o cca sio na lra ndo a l
t ra tio n o fa string
po sitio n ( a s d o n a pro a il
it o f ta tio n). In a ina r co d , this in o l s
cha nging a to a a nd ic
rsa . h
ta tio n o p ra to r h l
ps in a o iding
th po ssi il
it o f ista kinga l
o ca l ini
fo r a gl
o a l ini
.
ta tio n is
s al
l s d spa ringl. h n
ta tio n is s d in co nj nctio n ithr pro d ctio n
a nd cro sso r,iti pro s th gl
o a lna t r o fth
s a rch.
h
is
pl
o d to s a rch fo r th o pti a lpa ra t rs o fth th r a l
od l
. It is a ina r co d d i pl nta tio n ritt n in is a l + + 5. a nd
d l
o p d a t th ni rsit o f i rpo o l
. Its co p ta tio n pro c ss is l
ist d a s
fo l
l
o s:
( ) Initia l
i th po p l
a tio n.
(2)
a la t th tn ss o f r stringin th po p l
a tio n.
(3) K p th
ststringin th po p l
a tio n.
(4) a k a s lctio n o n th po p l
a tio n a tra ndo .
(5) a k cro sso r o n s lct d strings ith pro a il
it .
(6 ) a k
ta tio n o n s lct d strings ith pro a il
it
.
(7 )
a la t th tn ss o f r stringin th n po p l
a tio n.
( ) ak l
itis .
(9 )
p a t(4) to ( ) ntilth t r ina tio n co nditio n is t.
D e v e lo p m e n t o f P o w e r T ra n s fo rm e r T h e rm a l M o d e ls
1 9 9
In th co p ta tio n, th cro sso r pro a il
it, , a nd th
ta tio n pro a il
it, , th si o fth po p l
a tio n a nd th a i
n
r o fg n ra tio ns
a r s lct d ‘a prio ri’
.
.
i
ss
ci
h tn ss f nctio n sho l
d r fl ctth d sir d cha ra ct risticso fth s st
ing
o pti i d. Itis r i po rta ntto s a n a ppro pria t tn ss f nctio n sinc this
dri s th o ltio n pro c ss. o a pplth
to th po r tra nsfo r r th r a l
od l
s fo r l
a t d in ctio n 2,a tn ss f nctio n a nd o th r r l a ntpa ra t rs
o fth ina r g n tic a l
go rith sho l
d
d t r in d. h rro r t n th
a s r d a ria ls a nd th o d lo tp ts is d n d a s tn ss. h s,th tn ss
f nctio n sho l
d a tla stco nta in t o t r s co rr spo ndingto th tn ss o f
a nd
r sp cti l. o r a ch indiid a l(string) o f a g n ra tio n, its to ta l
tn ss a l is ca l
c l
a t d a s fo l
l
o s:
=
+
(5)
t
n
n
=
(k) +
t (k)
n
=
n
(
(k) −
(k)) +
( t (k) −
t (k))
hr
a nd t a r th tn ss fo r
a nd
co po n nts o fth o d l
r sp cti l;
(k) a nd
(k
)
a
r
t
h
rro
rs
t n th r a l a s r t
nta nd th o d lo tp t nd r th a ppro pria t s r ic co nditio n (a int
t p ra t r , l
o a d ra tio ) r sp cti l;
(k) a nd
t (k) r pr s nta gro p o f
as r
nts a nd is th to ta ln
r o f as r
ntgro ps. (k) a nd
(k
)
ca
n
a
si
l
ca
l
c
l
a
t
d
t
h
a
o
o
d
l
s.
t
4
4
.
s ts
sis
r
rs
h ca l
c l
a tio n a s a s d o n th da ta pro id d
th a tio na l rid o pa n (
). h da ta co r o th na t ra lo il
, na t ra la ir (
) a nd fo rc d
o il
,fo rc d a ir (
) o p ra tingr gi s hicha r s itch d o r a to a tica l
l.
o t p s o f o d lha
n in stiga t d a s d scri d a o :( ) o n ntio na l
st a d -sta t o d l
; (2) ra nsint-sta t o d l ith r c rsio n o fca l
c l
a t d o tp ts a nd ff ct o ffa ns, hich a ct a l
l fa l
linto t o gro ps. h rst r fl cts
th st a d -sta t r l
a tio nship t n th
,
,a intt p ra t r a nd
tra nsfo r r l
o a d ra tio ; th s co nd co nc rns th t p ra t r a ria tio n o r a
p rio d ti a s a n in rtia lr spo ns .
s sta t d a o ,th
a nd
a r s lct d a s th o d lo tp ts. h
diff r nc a o ng th o d l
sl
is in th co nstr ctio n o fth fo r l
a fo r o th
2 0 0
W .H . T a n g e t a l.
th
o tto -o ila nd to p-o ilt p ra t r s. h
a s r d pa ra t rs, pro id d
, a r a intt p ra t r
, ra tio o fl
o ad ,
,
t a nd
co prising 4 gro ps o f a s r
nts ith a n sa plint r a lo f
in t .
h o d lpa ra t rs ha
n o pti i d singth ina r g n tic a l
go rith
a s d o n 4 gro ps o fr a l a s r
nts,a nd th n ri d o n th r a ining
56 gro ps. h ca l
c l
atd r s l
tsd ri d fro
od l
( ) a nd tho s fro
od l
(2) a r l
ist d in a l .
4
.
i
-s
h st a d -sta t o d lha s
n t st d rst. Itis a ss
d tha tth
a nd
ca n i
dia t l r a chth ir st a d -sta t a l s
a nd r l
a t to o nl o n s r ic co nditio n, pr ss d
q a tio ns ( ) a nd (2) in
s ctio n 2. .
h o d lpa ra t rs to
d t r in d in o l , , , ra t d
a nd
ra t d
l, th pa ra t rs a r o ta in d
pirica l
l. h r fo r ,
t. o r a l
to g th r ith l
o a d ra tio
a nd a intt p ra t r
, th r spo ns s o fth
pirica l o d lca n
ca l
c l
atd
q a tio ns ( ) a nd (2). h co pa riso ns
t n th o d lo tp tsa nd r a l
o tp tsa r sho n in ig r s a nd 2 r sp cti l. It a
no t d tha tth st a d -sta t o d lpro d c s a n nsa tisfa cto r
tn ss (in a l ) to th
p ri nta lda ta .
r si -s
pp ic
p r i
s disc ss d in
s ctio n 2.2, th
a nd
ca nno t i
dia t l r a ch th co rr spo nding
st a d -sta t a l s nd r cha ngingl
o a ds,a s th ir th r a lti co nsta nts a r o f
th o rd r o fho rs. h r fo r , th ir tra nsint ha io r sho l
d
r pr s nt d
in a r c rsi fo r
singpr io s sa pls.
h rs l
tso ta in d singth o d l
sr l
a tingto fa n o p ra tio n ithr c rsio n
o fca l
c l
a t d o tp ts a r sho n in ig r s 3 a nd 4. s p ct d, th s r s l
ts
pr s ntth sta gr
nt ithth
as r
nts,co pa r d itho th r o d l
s.
Ifa o r pr cis fa n o p ra tio n ti co l
d
pro id d, itis p ct d tha ta n
n o r a cc ra t r s l
t o l
d
a chi d.
co pa riso n ith p ri nta l a s r
nts indica t s th tra nsint-sta t
o d l ith r c rsio n o fth o tp ts ha s a tt r p rfo r a nc tha n tha to fth
st a d -sta t o d l
. inc itdo s no tr q ir th
a nd
as r
nts,
this il
lc rta inl o fgr a tpra ctica la l a s a pr dicti to o l nd r a id r
a rit tra nsfo r r l
o a dinga nd a intco nditio ns.
rti ci
r
t r
s
ha s
n id l s d to p rfo r co pl f nctio ns in a rio s l
ds o f
a ppl
ica tio n incldingpa tt rn r co gnitio n,cl
a ssi ca tio n a nd co ntro l
s st s. h
t chniq is a s d o n th th o r o f io l
o gica ln r o ss st s a nd in o l s
s lctio n o finp t, o tp t, n t o rk to po l
o g a nd
ight d co nn ctio ns o fth
no d s [ 3]. In o r ca s ,th n t o rkinp ts a r cho s n a s l
o a d ra tio a nd a int
D e v e lo p m e n t o f P o w e r T ra n s fo rm e r T h e rm a l M o d e ls
2 0 1
n iro n ntt p ra t r a nd th n t o rk o tp tis th
or
o fth
tra nsfo r r.
h tra ining po chs a r 5 . l
lth n ro n f nctio ns a r cho s n a s th
sig o id f nctio n c ptth o tp tn ro n hich pl
o sa p r l
in a rf nctio n.
7 % o f to ta lda ta a r s d to tra in th n t o rk a nd th r a ining 3 % a r
r s r d fo r a la tio n o fth p rfo r a nc o fth
a pping.
h dir ct a pping pl
o s th
inp ts:
[ ( − )
( − 2)
( − 3)
( − )
( − 2)
( − 3)]
a nd th o tp t t ( ) o r ( ). 6 9
is cho s n fo r th dir ct a pping.
h rs l
ts a r sho n in ig r s 5a nd 6 .
h
o d lpro id s a go o d a pping t n th inp ts a nd o tp ts.
Ho
r itdo s no tpo ss ss a n ph sica l a ning a nd itha s
n no t d tha t,
fo r dir ct a pping,th o d lo nlr pr s nts th inp ta nd o tp tr l
a tio nship
a cc ra t l ithin th ra ng co r d
th tra iningda ta . tsid this ra ng ,th
o d lr spo ns a nd
a s r d da ta do no ta gr
ith a ch o th r sa tisfa cto ril.
It o nl d o nstra t s th po ssi il
it o fo n st p pr dictio n o fth tra nsfo r r
t p ra t r s. l
so a sth o d l
do sno tha a n ph sica l a ning,itpo ss ss s
l
ittlpo t ntia lfo r co nditio n o nito ring.
6
c si s
h a ppl
ica tio n o fth o d l
sto th
la rning tho d,pr s nt d in thispa p r
sho tha t high pr cisio n r s l
ts r a tta in d h n co pa r d to th r s l
ts
d ri d fro th o d l
s itho tr c rsio n a nd th o d l
so f
.Itca n s n
tha tth tra nsint-sta t o d l ith r c rsio n o fca l
c l
a t d o tp ts, inclding
th ff cto ffa n o p ra tio n, ha s a po t ntia lfo r r pr s nting a cc ra t l th r a l
tra nsfo r r th r a ld na ics. o thth o d lpa ra t rs a nd r spo ns s in th
o d lappl
ica lto fa n o p ra tio n ca n
s d fo rpr dictingth r a l
distri tio n
a nd o nito ringth s r ic co nditio n o ftra nsfo r rs.
r c s
. . H . r a n a na , . . a tt ns, ra nsf rm r c nditin m nit ring r a l
iing a n
int rgra t d a da pti a na lsis s st m, I
2 s ssin,3
g st-5 pt mb r.
2. . ibfrid, . n rr, . ir ck, n-l
in m nt ring f p
r tra nsf rm rstr nds, n d lpm nta nd first p rinc s, I
s ssin.
3. I. . mp, a rtia ldischa rg pl
a nt-m nit ringt chn lg : r s nta nd f t r d lpm nts,
rc di gs ci c ,
as r
ta d ch
g, l
. 42, n . ,
an ar
5, pp.4- .
4. a m s, . ta l
,
d l
-ba s d m nit ring ftra nsf rm rs, a ssa ch s tts Instit t
f chn lg , a b ra t r f r lctr ma gn tic a nd lctr nic st ms,
5.
5. ra nsf rm rs mmitt fth I
r ngin ring cit,
g id t
a di g i ra i i
rs d tra sf r r,
td 57. 5
, h Instit t
f lctrica la nd lctr nics ngin rs,Inc.,
,345 a st47 th tr t,N
rk,
N
7,
.
2 0 2
W .H . T a n g e t a l.
6 . Int rna tina llctr t chnica lc mmissin,
5
a di gg id f r i i
rs d
r tra sf r rs, q i a t t
/
5 6
( hi s
ta da rd),
,
n` , iss .
7 . hin s sta nda rds b a rd,
5,
r tra sf r rs a rt :
ra t r
ris ,
5.
. . . , p ra tin fp
rtra nsf rm rs, hi s
ctric
r b ish r, n
3.
. . H.
, . . a,
r s st m ptima lr a cti dispa tch sing
ltina r
pr gra mming,
ra s
r st s, l
. , n .3,
5, pp. 243- 24.
. .H .
, . . a , . .
n, ptima l
r a cti p
rdispa tch singa n a da pti
g n tic a l
g rithm, t r a ti a
r a f ctrica
ra d
rg
st s,
l
.2 , n . ,
, pp.56 3-56 .
. . . a, . H .
, n ra t r pa ra m t r id ntifica tin sing
ltina r pr gra mming, t r a ti a
r a f ctrica
ra d
rg
st s, l
. 7,
n .6 ,
5, pp.47 -423.
2. . il
l
a , .N l
a n, ppl
ica tin fg n tic a l
g rithms t m t r pa ra m t r d t rmina tin f r tra nsintt rq ca l
c l
a tins,
ra sa cti s
d str
i
ca ti s, l
.33, n .5, pt mb r/ ct b r
7 , pp. 27 3- 2 2.
3. . .Z a ma n, p rim nta l
t sting fth a rtificia l
n ra l
n t rkba s d pr t ctin
fp
r tra nsf rm rs,
ra sa cti s
r
i r, l
. 3, n .2, pirl
, pp.5 -5 7 .
7RSRLO WHPSHUDWXUH GHJUHH
PHDVXUHG WRSRLO WHPS
ZLWKRXW UHFXUVLRQ
i. .
1XPEHU RI VDPSOHV 6DPSOH LQWHUYDO PLQXWH
p- ilt mp ra t r
fth m d l ith tr c rsin
D e v e lo p m e n t o f P o w e r T ra n s fo rm e r T h e rm a l M o d e ls
2 0 3
%RWWRPRLO WHPSHUDWXUH GHJUHH
PHDVXUHG ERWWRPRLO WHPS
ZLWKRXW UHFXUVLRQ
1XPEHU RI VDPSOHV 6DPSOH LQWHUYDO PLQXWH
tt m- ilt mp ra t r
i. .
fth m d l ith tr c rsin
PHDVXUHG WRSRLO WHPS
ZLWK UHFXUVLRQ RI FDOFXODWHG YDOXHV DQG HIIHFW RI IDQV
7RSRLO WHPSHUDWXUH GHJUHH
1XPEHU RI VDPSOHV 6DPSOH LQWHUYDO PLQXWH
i. . p- ilt mp ra t r
ff ct ffa ns
f th m d l ith r c rsin f ca l
c l
atd
tp ts a nd
%RWWRPRLOWHPSHUDWXUH GHJUHH
PHDVXUHGERWWRPRLOWHPS
ZLWKUHFXUVLRQRIFDOFXODWHGYDOXHVDQGHIIHFWRIIDQV
1XPEHURIVDPSOHV 6DPSOHLQWHUYDOPLQXWH
i. . tt m- ilt mp ra t r
ff ct ffa ns
fth m d l ith r c rsin fca l
c l
atd
tp ts a nd
W .H . T a n g e t a l.
PHDVXUHG WRSRLO WHPS
$11 PDSSLQJ
7RSRLO WHPSHUDWXUH GHJUHH
1XPEHU RI VDPSOHV 6DPSOH LQWHUYDO PLQXWH
p- ilt mp ra t r
i. .
fth
N N ma pping
%RWWRPRLO WHPSHUDWXUH GHJUHH
2 0 4
PHDVXUHG ERWWRPRLO WHPS
$11 PDSSLQJ
1XPEHU RI VDPSOHV 6DPSOH LQWHUYDO PLQXWH
tt m- ilt mp ra t r
i.6 .
.
fth
N N ma pping
s l
ts fth m d l
s c nstr ct d b
d l
()
itn ss
t
4.
. 3 . 3 5.7 4 .6
5 32
d l
(2)
b tt m
2. 3 2 .
t
tp
mpirica la l s
.
.23
t
t
.5 .2 .2
t
t
25.2 .2 - . 3 .
.
t
.
.
t
t
.2
t
. 2 -3.6 4- . 4 . 2
5
.6
=
5,
=
. h si fp p l
a tin is 5 a nd th ma im m n mb r f
g n ra tins is 2
. h t ta lgr ps fda ta f r ca l
c l
a tin is 4 .
h d finitin fpa ra m t rs a r sh n in q a tins (3) (4).
Automatic Validation of Protocol Interfaces
Described in VHDL
Fulvio Corno, Matteo Sonza Reorda, Giovanni Squillero
Politecnico di Torino
Dipartimento di Automatica e Informatica
Corso Duca degli Abruzzi 24 I-10129, Torino, Italy
{corno, sonza, squillero}@polito.it
Abstract. In present days, most of the design activity is performed at a high
level of abstraction, thus designers need to be sure that their designs are
syntactically and semantically correct before starting the automatic synthesis
process. The goal of this paper is to propose an automatic input pattern
generation tool able to assist designers in the generation of a test bench for
difficult parts of small- or medium- sized digital protocol interfaces. The
proposed approach exploit a Genetic Algorithm connected to a commercial
simulator for cultivating a set of input sequence able to execute given
statements in the interface description. The proposed approach has been
evaluated on the new ITC 99 benchmark set, a collection of circuits offering a
wide spectrum of complexity. Experimental results show that some portions of
the circuits remained uncovered, and the subsequent manual analysis allowed
identifying design redundancies.
1
Introduction
In the past years, the design flow of protocol interfaces, and Application Specific
Integrated Circuits (ASICs) in general, experienced radical changes. Due to the
maturity of automatic logic synthesis tools most of the design activity is now
performed at high level of abstraction, such as register transfer level (RT), instead of
low level such as gate. The new methodology dramatically increases designer
productivity since high-level descriptions are more readable and considerably smaller.
One important step of the new design flow consists of design validation, i.e., the
verification that the design is syntactically and semantically correct before starting
automatic logic synthesis. Although many techniques have already been proposed in
the CAD literature (e.g., static checks, formal verification [HuCh98], mutation testing
[AHRo98]), none has gained enough popularity to compete with the current industrial
practice of validation by simulation. Verification engineers resort to extensive
simulation of each design, and of the complete system, in order to gain confidence
over its correctness.
This situation is far from ideal, and designers need to face many difficulties. At the
present days, simulation technology is effective enough for synthesized circuits. But
when it comes to mixed-signal circuits, or to circuits containing embedded cores, or to
S . C a g n o n i e t a l. ( E d s .) : E v o W o r k s h o p s 2 0 0 0 , L N C S 1 8 0 3 , p p . 2 0 5 − 2 1 4 , 2 0 0 0 .
S p rin g e r-V e rla g B e rlin H e id e lb e rg 2 0 0 0
2 0 6
F . C o rn o , M . S o n z a R e o rd a , a n d G . S q u ille ro
a complete system composed of a network of several interconnected interfaces,
simulation is unable to provide the needed versatility.
Even restricting our attention to medium-sized, digital interfaces, the fundamental
issue of measuring the test bench quality can be considered still unanswered. Many
metrics have been proposed to evaluate the thoroughness of a given set of input
stimuli, often adopted from the software testing domain [Beiz90], ranging from
statement or branch coverage, state coverage (for finite state machine controllers),
condition coverage (for complex conditionals), to the more complex path coverage.
Many variants have been developed, mainly to cater for observability [DKGe96] and
for the inherent parallelism of hardware descriptions [TAZa99], that are not taken into
account by standard metrics. Since no well established metric is yet widely accepted
for validation, some authors also propose to measure the quality of validation patterns
with the stuck-at fault coverage.
Several products (normally integrated into existing simulation environments) are
now available that provide the user with the possibility of evaluating the coverage of
given input stimuli with respect to a selected metric. Designers can therefore pinpoint
the parts of their design that are poorly tested, and develop new patterns specifically
addressing them. Currently, this is a very time consuming and difficult task, since all
the details of the design must be understood for generating suitable input sequences.
The right trade-off between designer s time and validation accuracy is often difficult
to find, and this often results in under-verified circuits. Moreover, in the generation of
test vectors the designer may be biased by his knowledge of the desired system or
module behavior, so that he often fails in identifying input sequences really able to
activate possible critical points in the description.
When faced with this problem, the CAD research community traditionally invested
in formal verification [GDNe91] [HuCh98], in the hope that circuits can be proven
correct by mathematical means. Although formal verification tools give good results
on some domains, they still have too many limitations or they require too much
expertise to be used as a mainstream validation tool. Designers are left waiting for the
perfect formal verification system, while few or no innovative tools help them with
simulation-based validation.
The goal of this paper is to propose GIP-PI (Genetic Input Pattern generator for a
Protocol Interface). GIP-PI is an automatic input pattern generation tool able to assist
designers in the generation of a test bench for difficult parts of small- or mediumsized digital protocol interfaces. The proposed approach belongs to a brand new
framework that can be called approximate validation, which explicitly relinquishes
exactness in order to gain the ability of dealing with realistic designs. This philosophy
has already been successfully applied in different areas: validation of the
implementation of protocol interfaces [CSSq99c]; automatic test pattern generation
[CSSq99a]; low-level [CSSq98] and mixed-level [CSSq99b] equivalence validation.
Although the goal of this paper is completely different from the previous applications,
all these methodologies share a common quality: being able to deal with real circuits
exploiting an evolutionary algorithm.
GIP-PI employs a Genetic Algorithm, interacting with a VHDL simulator, for
deriving an input sequence able to execute a given statement, or branch, in the highlevel description. Whenever the test bench quality, as measured by one of the
proposed metrics, is too low, our tool can be used to generate test patterns that are
A u to m a tic V a lid a tio n o f P ro to c o l In te rfa c e s D e s c rib e d in V H D L
2 0 7
able to stimulate the parts of the design that are responsible for the low metric. The
designer must manually analyze only those parts of the description that the tool failed
to cover. Experimental results show that only a small fraction of difficult statements
remain uncovered, and that many of them, upon closer inspection, indeed contain
design errors or redundancies.
While no metric is yet widely accepted by validation teams, we aimed at evaluating
the effectiveness of our approach using some pre-defined metric. The algorithm is
quite easily adapted to different metrics, but for the sake of the experiments we
adopted branch coverage as a reference. We developed a prototypical system for
generating test patterns based on branch coverage, applicable to synthesizable VHDL
descriptions. We aim at addressing moderately sized circuits, that usually can not be
handled by formal approaches, and at working directly on the VHDL description,
without requiring any transformation nor imposing syntax limitations.
The approach has been evaluated on the new ITC 99 benchmark set [ITC99], a
collection of circuits described in high-level (RT) VHDL that offers a wide spectrum
of complexity. Manually derived validation suites did not adequately cover all parts of
the designs, and new sequences have been generated by the tool to increase the
overall coverage. Experimental results show that some portions of the circuits
remained uncovered, and the subsequent manual analysis allowed identifying design
redundancies.
Section 2 gives an overview over the proposed approach for test bench generation,
experimental are presented in Section 3 and Section 4 concludes the paper.
2
RT-level Test Bench Generation
The goal of test bench generation is to develop a set of input sequences that attain
the maximum value of a predefined validation metric. Despite this implementation of
GIP-PI is tuned for simulating high-level VHDL network interfaces only, the
proposed method could be easily extended to deal with any simulable descriptions.
For instance, given a protocol specification in ESTELLE, and with a reduced effort,
GIP-PI could eventually generate a set of stimuli (events) to validate the protocol
description itself.
2.1 Adopted Metric
Most available tools grade input patterns according to metrics derived from
software testing [Beiz90]: statement coverage and branch coverage are the most
widely known, but state/transition coverage (reaching all the states/transitions of a
controller) and condition coverage (controlling all clauses of complex conditionals)
are also used in hardware validation. Path coverage, although often advocated as the
most precise one, is seldom used due to its complexity, and because it loses
meaningfulness when multiple execution threads run concurrently in parallel
processes. Some recent work extends those metrics to take also into account
observability [DGKe96] and the structure of arithmetic units [TAZa99]. Those
2 0 8
F . C o rn o , M . S o n z a R e o rd a , a n d G . S q u ille ro
extensions are essential when the sequences have to be used as test patterns to cover
stuck-at faults, but for validation they have lower importance since internal values are
available.
The metric we adopt in this paper is branch coverage, although the tool can be
easily adapted to more sophisticated measures. Also, since synthesizable VHDL is a
structured language, complete statement coverage implies complete branch coverage,
and the tool takes advantage of this simplification.
2.2
Overall Approach
The adopted approach is an evolution of the one presented in [CPSo97], where a
Genetic Algorithm uses a simulator to measure the effectiveness of the sequences it
generates. Instead of trying to justify values across behavioral statements, that would
require solving Boolean and arithmetic constraints [FADe99], thanks to the nature of
Genetic Algorithms we just need to simulate some sequences and analyze the
propagation of values. Each sequence is therefore associated with the value returned
by a fitness function, that measures how much it is able to enhance the value of the
validation metric, and the Genetic Algorithm evolves and recombines sequences to
increase their fitness.
The fitness function needs to be carefully defined, and accurately computed. In
particular, the fitness function can not be just the value of the validation metric: it
must also contain some terms that indicate how to increase the covered branches, not
just to count the already covered ones. In a sense, the fitness function includes a
dominant term, that measures the accomplished tasks (covered branches), and
secondary terms, that describe sub-objectives to be met in order to cover new
branches.
The computation of such function is accomplished by analyzing the simulation
trace of the sequence, and by properly weighting the executed assignments,
statements, and branches according to the target statements. In the implementation, to
avoid arbitrary limitations in the VHDL syntax, simulation is delegated to a
commercial simulator that runs an instrumented version of the VHDL code and
records the simulation trace in the transcript file. Such trace is then interpreted
according to control- and data-dependencies, that are extracted from a static analysis
of the design description. Figure 1 shows a simplified view of the overall system
architecture.
2.3
VHDL Analysis
The goal of the algorithm is to achieve complete coverage, but for efficiency
reasons we do not consider each statement separately, and we group them into basic
blocks [ASUl86]: a basic block is a set of VHDL statements that are guaranteed to be
executed sequentially, i.e., they reside inside a process and do not contain any
intermediate entry point nor any control statement (i f , c a s e , ). All the operations
required for code instrumentation, dependency analysis, branch coverage evaluation,
and fitness function computation are performed at the level of basic blocks.
A u to m a tic V a lid a tio n o f P ro to c o l In te rfa c e s D e s c rib e d in V H D L
2 0 9
Since the Genetic Algorithm exploits the knowledge about data and control
dependencies, we need to extract that information from the VHDL code: for this
reason, we build a database containing a simplified structure and semantics of the
design. The database is structured as follows:
•
The hierarchy of component instantiations inside different entities is flattened
(C1 and C2 in the figure 2). A dictionary of signal equivalencies is also built, that
allow us to uniquely identify signals that span multiple hierarchical levels.
9+'/
$1$/<=(5
,167580(17(5
seqs
*(1(7,&
$/*25,7+0
&200(5&,$/
6,08/$725
trace
Figure 1: System architecture
•
All VHDL processes occurring in the flattened circuit are given a unique
identifier (Pi in the figure 2). This operation also converts standalone concurrent
statements into their equivalent process. The design is thus represented as a
network of processes interconnected by signals.
E1
E2
P1
P3
P4
C1
P5
P6
C2
P2
')*
A
C
BB1
BB2
BB3
BB4
BB5
BB6
BB7
P7
+
B
Figure 2: Abstract representation of RT-level designs
•
•
Each process is analyzed to define its interface, in terms of signals that it reads
and writes.
The sequential part of each process is analyzed, its control flow graph (CFG) is
extracted, and statements are grouped in basic blocks (BBs). The control
structure of the process is described as a control flow of basic blocks (figure 2
reports the CFG for process P2).
2 1 0
•
F . C o rn o , M . S o n z a R e o rd a , a n d G . S q u ille ro
A dependency matrix between basic blocks is computed, by assigning a
probability that a basic block will be executed, given that another block has just
been executed. These correlation probabilities take into account the branching
and looping nature of the control flow.
Each basic block is entered, and the data flow graph (DFG) of the operations that
occur inside each basic block is extracted. Since a basic block consists of
multiple statements and/or conditions, multiple dependencies are associated to a
single block. Fig. 2 shows the DFG for the basic block BB4 of process P2.
•
2.4
Genetic Algorithm
The Genetic Algorithm (GA) is based on encoding potential test sequences as
variable length bit matrices. A number of such sequences are randomly generated and
constitute an initial population: the goal of the GA is to evolve this population to
increase its fitness value. The fitness function measures the closeness of a sequence to
the goal. Currently, in GIP-PI the genetic algorithm is run several times, each time for
a different target. Moreover, the fitness function assumes two different forms:
•
In the initial phase, all basic blocks are considered simultaneously. The goal is to
generate a set of sequences S that activate most of the blocks, for identifying
easy-to-execute blocks. The fitness function is simply the number of activated
blocks over the total number of blocks:
activated _blocks(S )
fitness(S ) =
tot _ blocks
•
Subsequently, GIP-PI is targeting a specific block T and the goal is to generate a
sequence S able to cover it. In this phase sequences are targeted and the executed
blocks are weighted by their correlation probability with respect to the target,
measured as the weighted average of the execution counts of the basic blocks in
the input cone (taking into account both control and data dependencies, thus
potentially spanning several processes) of the target. The adopted weights take
into account the probabilities of conditional execution that were statically
computed in the database.
fitness(S , T ) =
∑ correlation(b, T )
b∈ covered_ bb ( S )
In both phases, a saturation mechanism prevents easy-to-execute but not-sorelevant statements from diverting the attention of the GA.
Moreover, during each run of the GA, an heuristic mechanism detects individuals
that may be useful in a subsequent run. When such an individual is found, it is saved
and later it is inserted in the initial population of the correct run of the GA. These
predefined individuals may never exceed 5% of the initial population.
The GA in GIP-PI evolves a population of µ individuals and in each generation λ
new sequences are first generated, then selection is performed on the whole set of
µ +λ individuals. Individuals are selected for reproduction using a roulette wheel
mechanism based on their linearized fitness. In p of the cases, the new individual is
built mutating a single parent: the original sequence can be shortened, or enlarged, or
some bits may be flipped. In 1-p of the cases, the new individual is built mating two
A u to m a tic V a lid a tio n o f P ro to c o l In te rfa c e s D e s c rib e d in V H D L
2 1 1
different parents: the offspring sequence can inherit the beginning from one parent
and the end from the other, or some entire bit column from each parent.
3
Experimental results
To test the effectiveness of the tool in generating test benches, we selected a set of
VHDL benchmarks from the ITC 99 benchmark set [ITC99]. The first columns report
some data about the RT-level descriptions, in terms of VHDL lines, VHDL processes
(with hierarchy unflattened), and overall number of extracted basic blocks (BB). To
have a better idea about circuit size, some characteristics of the synthesized netlists
are reported in the last columns: number of Primary Inputs, Primary Outputs, FlipFlops, and combinational gates.
These benchmarks have been publicly released in September 1999 at the IEEE
International Test Conference, and there are no published results, yet, to compare
with. We compared our results against a pure random approach, to evaluate the
effectiveness of the GA, but it was so easily overcome that results are not reported.
The implementation consists of about 4,700 lines of C code for VHDL code
analysis and instrumentation, linked to the LEDA LPI interface [LEDA95], and of
2,700 lines of C code for the Genetic Algorithm and the interface to the simulator. All
experiments were run on a Sun Ultra 5 running at 333 MHz with 256MB of memory.
We adopt a population of µ = 30 individuals, with λ = 20 new individuals in each
generation. The mutation probability was set to p = 0.3.
2 1 2
F . C o rn o , M . S o n z a R e o rd a , a n d G . S q u ille ro
Circ
b01
b02
b03
b04
b05
b06
b07
b08
b09
b10
b11
b12
b13
b14
b15
b20
b21
VHDL
Lines Proc
111
1
71
1
142
1
103
1
333
3
129
1
93
1
90
1
104
1
168
1
119
1
570
4
297
5
510
1
672
3
1,085
3
1,089
3
BB
28
17
27
23
94
25
21
14
16
38
37
118
74
244
171
491
491
GATE
PI PO FF
2
2
5
1
1
4
4
4 30
11
8 66
1 36 34
2
6
9
1
8 49
9
4 21
1
1 28
11
6 17
7
6 31
5
6 121
10 10 53
32 54 245
36 70 449
32 22 490
32 22 490
Gate
46
28
149
597
963
60
420
167
159
189
481
1,036
339
4,775
8,893
9,419
9,803
Table 1: Benchmark characteristics
In Table 2, we report the experiments we obtained with our prototypical tool in
terms of percent number of covered branches, number of generated vectors, and
required CPU time. These data demonstrate that for most descriptions, our method is
able to reach a complete or very high branch coverage. There are a few circuits (e.g.,
b05 and b12) where the obtained coverage is low: this is due to the specific
characteristics of these circuits, which include highly nested conditional statements
that the current version of our algorithm can hardly go through.
It is worth noting that a manual analysis of the branches left uncovered proved that
many of them were effectively unreachable, in most cases due to e l s e or d e f a u l t
statements that are required by the synthesis tool not to infer sequential logic, but are
redundant since all cases have already been considered in previous tests.
A u to m a tic V a lid a tio n o f P ro to c o l In te rfa c e s D e s c rib e d in V H D L
CIRCUIT
b01
b02
b03
b04
b05
b06
b07
b08
b09
b10
b11
b12
b13
b14
b15
b20
b21
2 1 3
Cov %
#VECT CPU [s]
100,00
259
439.0
100,00
114
41.3
100,00
174
55.5
100,00
83
425.1
52,13
68
2014.9
100,00
125
52.4
95,24
351
920.6
100,00
1,005
971.1
100,00
958
511.5
100,00
364
122.1
94,59
1,222 1,410.2
36,44
155 1,022.3
100,00
3,303 4,203.9
93,03
4,597 7,875.5
91,81
2,838 9,369.2
93,48
7,784 27,286.5
93,69
6,376 28,878.9
Table 2: Experimental results
4
Conclusions
This paper presented an automatic input pattern generation tool able to assist
designers in the generation of a test bench for difficult parts of small- or mediumsized digital protocol interfaces. The approach resorts to a Genetic Algorithm that
interacts with a simulator to generate new sequences able to increase the coverage of
the test bench with respect to a predefined validation coverage metric.
The methodology has been tested on the new ITC 99 benchmark set [ITC99].
Experimental results prove that the method is able to increase the quality of the
validation process both over manual simulation and pseudo-random sequence
generation. However, the proposed method could be easily extended to deal with any
simulable descriptions, like the ESTELLE specification of a network protocol.
The tool results have also been useful as a feedback for better understanding the
most difficult parts of the design from the validation point of view.
5
References
[AHRo98] G. Al-Hayek, C. Robach: From Design Validation to Hardware Testing: A Unified
Approach, JETTA: The Journal of Electronic Testing, Kluwer, No. 14, 1999, pp.
133-140
2 1 4
F . C o rn o , M . S o n z a R e o rd a , a n d G . S q u ille ro
[ASUl86]
[Beiz90]
[CPSo97]
[CSSq98]
[CSSq99a]
[CSSq99b]
[CSSq99c]
[DGKe96]
[FADe99]
[FDKe98]
[GDNe91]
[HuCh98]
[ITC99]
[LEDA95]
[TAZa99]
A.V. Aho, R. Sethi, J.D. Ullman, Compilers, Principles, Techniques, and Tools,
Addison-Wesley Publishing Company, 1986
B. Beizer, Software Testing Techniques (2nd ed.), Van Nostrand Rheinold, New
York, 1990
F. Corno, P. Prinetto, M. Sonza Reorda: Testability analysis and ATPG on
behavioral RT-level VHDL, Proc. IEEE International Test Conference, 1997, pp.
753-759
F. Corno, M. Sonza Reorda, G. Squillero, VEGA: A Verification Tool Based on
Genetic Algorithms, Intl. Conf. on Circuit Design, 1998, pp. 321-326
F. Corno, M. Sonza Reorda, G. Squillero, Improved Test Pattern Generation on
RT-level VHDL descriptions, ITSW 99: International Test Synthesis Workshop,
1999
F. Corno, M. Sonza Reorda, G. Squillero, Simulation-Based Sequential
Equivalence Checking of RTL VHDL, ICECS’99: 6th IEEE Intl. Conf. on
Electronics, Circuits and Systems, 1999
F. Corno, M. Sonza Reorda, G. Squillero, Approximate Equivalence Verification
for Protocol Interface Implementation via Genetic Algorithms, Evolutionary Image
Analysis, Signal Processing and Telecommunications First European Workshops,
EvoIASP’99 and EuroEcTel’99, 1999, pp. 182-192
S. Devadas, A. Ghosh, K. Keutzer: An Observability-Based Code Coverage Metric
for Functional Simulation, Proc. ICCAD 96
F. Fallah, P. Ashar, S. Devadas: Simulation Vector Generation from HDL
Descriptions for Observability-Enhanced Statement Coverage, Proc. 36th DAC,
New Orleans, 1999, pp. 666-671
F. Fallah, S. Devadas, K. Keutzer: OCCOM: Efficient Computation of
Observability-Based Code Coverage Metrics for Functional Verification, Proc.
35th DAC, 1998
A. Ghosh, S. Devadas, A.R. Newton, Sequential Logic Testing and Verification,
Kluwer, 1991
S.-Y. Huang, K.-T. Cheng, Formal Equivalence Checking and Design Debugging,
Kluwer, 1998
h t t p : / / w w w . i t c t e s t w e e k . o r g / b e n c h m a r k s . h t m l
LVS System User s Manual, LEDA Languages for Design Automation, Meylan
(F), April 1995
P.A. Thaker, V.D. Agrawal, M.E. Zaghloul: Validation Vector Grade (VVG): A
New Coverage Metric for Validation and Test, VTS 99: IEEE VLSI Test
Symposium, 1999, pp. 182-188
ti M d i
r I tr si
il
ippo
/
N t r Tr
D t cti
c
ri
- ni rsit o f imo nt rinta l, o rso o rsa l
ino 54,
5
lssa ndria ( L), Ita l
neri@di.unito.it, neri@al.unipmn.it
c . h d t ctio n o fintrusio ns o r co mput r n t o rks ca n b
ca stto th ta sk o fd t cting a no ma l
o us pa tt rns o fn t o rk tra ffic. In
this ca s , pa tt rns o f no rma ltra ffic ha
to b d t rmin d a nd co mpa r d a ga inst th curr nt n t o rk tra ffic. a ta mining s st ms ba s d
o n n tic l
go rithms ca n co ntribut po rfuls a rcht chniqu s fo r th
a cquisitio n o fpa tt rns o fth n t o rk tra ffic fro m th l
a rg a mo unto f
da ta ma d a a il
a blb a uditto o l
s.
In this pa p r
co mpa r mo d l
s o fda ta tra ffic a cquir d b a s st m
ba s d o n a distribut d g n tic a l
go rithm ith th o n s a cquir d b a
s st m ba s d o n gr d h uristics. l
so
discuss r pr s nta tio n cha ng
o fth n t o rk da ta a nd its impa cto r th p rfo rma nc s o fth tra ffic
mo d l
s.
tr d cti
h ris i th
mb r o fco mp t r br a -i s, irt a l
l o cc rri g a ta
sit ,
d t rmi s a stro gr q stfo r pl
o iti gco mp t r s c rit t ch iq s to pro t ct th sit a ss ts.
a rit o f a ppro a ch s to itr sio d t ctio do
ist
[
i g, 9 7]. o m o fth m pl
o itsig a t r s o f o
a tta c s fo r d t cti g
h a itr sio o cc rs. h s th a r ba s d o a mo d lo fpo ssiblmis s s o f
a r so rc [K ma r a d pa ffo rd, 9 9 4]. th r a ppro a ch s to itr sio d t ctio
cha ra ct ri th o rma lsa g o fth r so rc s d r mo ito ri g.
itr sio is
s sp ct d h a sig i ca tshiftfro m th r so rc ’
s o rma l sa g is d t ct d.
his a ppro a chs ms to b mo r pro misi gb ca s o fits po t tia la bil
it to d t ct
o
itr sio s. H o
r, ita l
so i o l s ma jo r cha l
l g s b ca s o f
th
d to a cq ir a mo d lofth o rma ls g ra l o ghto a l
l
o a tho ri d
s rsto o r itho tra isi ga l
a rms,b tsp ci c o ghto r co g i d a tho ri d sa g s [ a a d ro dl , 9 9 7, ho sh ta l
., 9 9 9 ,
ta l
., 9 9 9 ]. o r
th sco p o fthis o r ,
fo c s to th ta s o fd t cti gitr sio s b a a l i g
l
o gs o f t o r s tra ffic.
a
a sp cts o f d pl
o i g i -pra ctic a ppro a ch s ba s d o da ta mi i g,
ho
r,a r stil
lo p [ a a d ro dl , 9 9 ].
co c tra t h r o s lcti gi fo rma ti da ta r pr s ta tio a d cl
a ssi ca tio p rfo rma c s. s la r i gm tho ds,
pl
o it d t o r lba s d s st ms:
a h ristic o , I
[ o h , 9 9 5], a d a sto cha stic o
(ba s d o g tic
S . C a g n o n i e t a l. ( E d s .) : E v o W o r k s h o p s 2 0 0 0 , L N C S 1 8 0 3 , p p . 2 1 4 − 2 2 3 , 2 0 0 0 .
S p rin g e r-V e rla g B e rlin H e id e lb e rg 2 0 0 0
E v o lu tiv e M o d e lin g o f T C P /IP N e tw o rk T ra ffic fo r In tru s io n D e te c tio n
2 1 5
al
go rithms),
[ io rda a a d
ri, 9 9 5, ria d a itta , 9 9 6]. h rst
s st m ha s b
s lct d b ca s o fits pr io s s [
ta l
., 9 9 9 ]; it il
l
th s a ct a s b chma r . h s co d s st m ha s b
s lct d b ca s its i tri sica l
l sto cha stic b ha io r a d sho l
d al
l
o th a cq isitio o fmo r ro b st
mo d l
s [ ria d a itta , 9 9 6].
I th fo l
l
o i g, a d scriptio o fth s st m
( ctio 2) a d o fth
p rim ts p rfo rm d i th I a d
co t ts ( ctio 3 a d ctio
4) a r r po rt d. i a l
l, th co clsio s a r dra .
The Systems Regal and Ripper
o r spa c r a so
il
lpro id h r a a bstra ct d scriptio o f bo th la r i g s st ms
a d I
a s th ir f l
ld scriptio s ha a l
r ad b
p bl
ish d. h s st m
is a a il
a blfo r fr fro m th a tho r.
[ io rda a a d
ri, 9 9 5, ria d a itta , 9 9 6] is a la r i g s st m, ba s d o a distrib t d g tic a l
go rithm ( ). Itta s a s i p ta s to f
da ta (tra i i gi sta c s) a d o tp ts a s to fs mbo l
ic cl
a ssi ca tio r ls cha ra ct rii gth i p tda ta . s s a l
, la r i gis a chi d b s a rchi ga spa c o f
ca dida t cl
a ssi ca tio r ls.
h l
a g ag
s d to r pr s tcl
a ssi ca tio r lsisa H o r cl
a s l
a g ag
i hich t rms ca b a ria bls o r disj ctio s o fco sta ts, a d ga tio o cc rs i a r strict d fo rm [ icha l
s i, 9 3].
a mplo fa a to mic pr ssio
co ta i i g a disj cti t rm is
r(
r ), hich is s ma tica l
l
q ia l t to
r(
) r
r( r ). ch fo rm l
a s a r r pr s t d
a s bitstri gs tha ta r a ct a l
l th po p l
a tio i diid a l
s pro c ss d b th
.
l
a ssica lg tic o p ra to rs, o p ra ti g o bi a r stri gs, ith th a dditio o f
ta s o ri t d s
a
a d
ra
cro sso rs a r
pl
o it d, i a a da pti
a , i sid th s st m (fo r d ta il
s s [ io rda a a d
ri, 9 9 5].
is a distrib t d g tic a l
go rithm tha t ff cti lco mbi s th h o r o f ich s a d p cis o f io l
o gica l o ltio to g th r ithpa ra l
llpro c ssi g. h s st m a rchit ct r is ma d b a s to f t d d impl
tic l
go rithms (
) [ ol
db rg, 9 9 ], hich co o p ra t s to si a d scriptio spa c ,
a d b a
p r iso r pro c ss tha tco o rdi a t s th
s ffo rts b a ssig i gto
a ch o fth m a diff r tr gio o fth ca dida t r l spa c to b s a rch d. I
pra ctic this is a chi d b d a mica l
l d isi g s bs ts o fth da ta s tto b
cha ra ct ri d b a ch
.
h s st m I
[ o h , 9 9 5] is ba s d o th it ra t d a ppl
ica tio o f
a gr d h ristic, simil
a r to th I fo rma tio
ai m as r [ il
a , 9 9 3], to
b il
d co j cti cl
a ssi ca tio r ls. t a chit ra tio ,tho s tra i i gi sta c s
co rr ctlcl
a ssi d b th fo d r lsa r r mo d a d th a l
go rithm co c tra t
o la r i g a cl
a ssi ca tio r lfo r th r ma i i g o . h s st m o tp ts a
o rd r d l
ist o f cl
a ssi ca tio r ls (po ssibl a sso cia t d to ma
cl
a ss s) to b
a ppl
id i tha tsa m o rd r to cl
a ssif a
i sta c .
it r sti gfa t r s o f
th m tho d is tha tit pl
o its o -l
i r lpr i g hili cr m ta l
l b il
di g
a
cl
a ssi ca tio r lto a o id o r tti g.
2 1 6
3
F . N e ri
tr si d t cti
t tc t st
i t
r
ti
p r ti
a la tio o f
o r a itr sio d t ctio ta s b
pl
o iti g da ta
fro m th I fo rma tio
pl
o ra tio
ho o to t ro jct (I ) is r po rt d i this
s ctio . h I ma d a a il
a bl t o r l
o gs pro d c d b ’
tcpd mp’fo r a l
a ti gda ta mi i gto o lo r l
a rg s to fda ta . h s l
o gs r co l
lct d a tth
ga t a b t
a
t rpris
a d th o tsid - t o r (I t r t). I th
I co t t, d t cti gitr sio s m a s to r co g i th po ssiblo cc rr c o f
a tho ri d (’
ba d’
) da ta pa c ts it rla d ith th a tho ri d (’
go o d’
)o s
o r th
t or
d r mo ito ri g. h I ’
s pro jctma s a a il
a blfo r tor l
o gs: o
is g a ra t o tto co ta i a
itr sio a tt mpts, h r a s th
o th r o s do i cld bo th o rma ltra ffic a d itr sio s a tt mpts. I th I
co t t, o cl
a ssi ca tio fo r a ch da ta pa c tis r q st d, i st a d a o ra l
l
cl
a ssi ca tio o fa b ch o fth
t o r tra ffic, a s co ta i i g a tta c s o r o t, is
d sir d.
a ppro a ch to itr sio d t ctio , ba s d o a o ma l d t ctio , ha s
b
s lct d.
pro c d a s fo l
l
o s. I da ta ca b pa rtitio d, o th ba sis o fth ir I
a ddr ss s,ito pa c ts iti gth r fr c i sta l
l
a tio ( tgo i g), t ri gth
i sta l
l
a tio (I co mi g) a d bro a dca st d fro m ho stto ho sti sid th i sta l
l
a tio
(I t rl
a ). hr mo d l
s o fth pa c ttra ffic, o
fo r a ch dir ctio , ha b
b il
tfro m th itr sio -fr da ta s t. h , th s mo d l
s ha b
a ppl
id to
th thr da ta s ts co ta i i gitr sio s.
p ctto o bs r a sig i ca t a ria tio i th cl
a ssi ca tio ra t b t
itr sio -fr l
o gs a d l
o gs co ta i i g
itr sio s b ca s o fth a
r a cha ra ct ristics o fth tra ffic pro d c d b th
itr si b ha io r. Ifthis o l
d a ct a l
lo cc r, co l
d a ss rttha tth la r d
tra ffic mo d l
s co rr ctlca pt r th ss tia lcha ra ct ristics o fth itr sio -fr
tra ffic. o t tha t h th r this a ppro a ch sho l
d or ,
co l
d co cld tha t
a itr si a tt mptis ha pp i g b ca s o fa diff r t
a tra
a tt r .
p rim ts ha b
p rfo rm d bo th ith I
a d
.
h
I
is a ppl
id to th I da ta , th cl
a ssi ca tio ra t a pp a ri g
i a bl b co m s id t[
ta l
., 9 9 9 ]. a chta bl tr r pr s ts a cl
a ssi ca tio
rro r a s m a s r d o r th fo r t o r l
o gs, o
fo r a ch l
i ,a d
ith r sp ctto th thr cl
a ss l
ab l
s:
tgo i g, I t rl
a , a d I co mi g. h
co rr ctcl
a ssi ca tio ra t s ca b o bta i d b s btra cti ga
tr ’
s a l fro m
. h s th rstro sho th miscl
a ssi d o rma ltra ffic pa c ts, h r a s th
o th r o s sho s th miscl
a ssi d pa c ts d ri ga itr sio a tt mpt.
hs rs l
ts ha b
o bta i d b a ppli g I
to th da ta a s a a il
a blfro m th tcpd mp d ls(s
pp di ). o pr pro c ssi go rth da ta ,
s cha s fa t r co str ctio ,ha s b
a ppl
id. h p rim ta l di gs sho s
tha tth a cq ir d mo d l
s do o t hibit r diff r tcl
a ssi ca tio ra t h
a ppl
id to l
o gs co ta i i g itr sio s ith r sp ctto itr sio -fr l
o gs. h s
di gs ma s gg sttha tth
pl
o it d da ta r pr s ta tio is to o d ta ild ith
r sp ctto th ca pa bil
it o fth la r i gs st m. I t r , this ca s s th la r d
mo d l
s to miss th i fo rma tio cha ra ct rii gitr sio -fr tra ffic.
E v o lu tiv e M o d e lin g o f T C P /IP N e tw o rk T ra ffic fo r In tru s io n D e te c tio n
. p rim nta l
r sul
ts o fa ppling I
to I da ta s ts usingth ra
r pr s nta tio n. a ch ta bl ntr sta t s a cl
a ssifica tio n rro r.
2 1 7
da ta
a ta s t
int rl
a n inco mingo utgo ing
no rma ltra ffic
.4
.4
.4
intrusio n
.23
.7
.4
intrusio n2
.9
.7
.5
intrusio n3
.
.4
.4
. p rim nta l
r sul
tso fa ppling I
da ta r pr s nta tio n.
to I
da ta s tsusinga co mpr ss d
a ta s t
int rl
a n inco mingo utgo ing
no rma ltra ffic
.2
.5
.4
intrusio n
.
.
.2
intrusio n2
.3
.3
.2
intrusio n3
.
.2
.2
ol
l
o i g this o bs r a tio ,
d l
o p a mo r co mpa ct r pr s ta tio fo r
th pa c ts tha tco sists i ma ppi g a s bs to ffa t r ’
s a l s ito a si gl
a l , th s r d ci g th ca rdi a l
it o fpo ssibl fa t r s a l s (s
pp di
).
pl
o iti g this r pr s ta tio , I
’
s p rfo rma c s b co m th o s
r po rt d i a bl 2 a d
’
s p rfo rma c s pl
o iti g th sa m co mpa ct
da ta r pr s ta tio a pp a r i a bl3. h o bs r d g r s sho a mo r sta bl
cl
a ssi ca tio b ha io r o fth mo d l
s a cro ss diff r ttra ffic co ditio s. l
so a
mo r disti ctcl
a ssi ca tio p rfo rma c b t
th itr sio -fr l
o g a d th
l
o gsi cldi gitr sio sis id t. co mpr ssio -ba s d r pr s ta tio isth a
a la bl a o fi cr a si gcl
a ssi ca tio p rfo rma c s itho titro d ci gco mpl fa t r tha tma i o l s a dditio a lpro c ssi go rh a d.
a la tio o f
th ff ctca s d b th a dditio o fco mpl fa t r s to th ra
t o r da ta
r pr s ta tio ha s b
p rfo rm d i [
ta l
., 9 9 9 ].
o r th sa o fcl
a rit, a
a mpl o fr lcha ra ct rii g itr sio -fr I co mi gpa c ts,la r d b
,a pp a rs i ig r . h I co mi gpa c ts
a r cha ra ct ri d i t rm o fth a l so fth fa t r sfro m th ir
/ I h a d r.
his r ls cc ssf l
lco rs 7 349 I co mi gpa c ts itho tb i gfo o ld b a
. p rim nta l
r sul
ts o fa ppling
da ta r pr s nta tio n.
Lto I
da ta s ts usinga co mpr ss d
a ta s t
int rl
a n inco mingo utgo ing
no rma ltra ffic
.2
.4
.4
intrusio n
.2
.5
.
intrusio n2
.6
.
.2
intrusio n3
.2
.5
.
2 1 8
F . N e ri
I
srcprt( ,[[ ,2 ],[4, ],[ 5 ,2 ],[ 5 ]]) a nd
dstprt( ,[
24]) a nd fla g( ,[ ,pt]) a nd
s q ( ,[[ , 5 ],[2 ,3 ],[5 ,5 ],[
]]) a nd
s q2( ,[[5 , ],[2 ,3 ],[5 ,2
]]) a nd
a ck( ,[[ ,3
],[5 ,
]]) a nd
in( ,[[ ,2
],[ 3
]]) a nd
buf( ,[ =5 2])
H
Inco ming a ck t( )
o ra g : (Int rl
a n, Inco ming, utgo ing) = ( , 7 349 , )
i. . a mplo fa rulcha ra ct riingpa rto fth inco mingtra ffic. h ruld scrib s
7 349 inco mingpa ck ts itho utco nfusing th m ith a n o utgo ing o r int rl
a n pa ck t.
I t rl
a o r tgo i go s. d scriptio o fth pr dica t s a pp a ri gi th r l
is pro id d i pp di .
tr si d t cti i t
t cti
ti
r r
tr si
al
so p rfo rm d a a dditio a l a la tio
o f o r a ppro a ch o r
tor l
o gs fro m 9 9
I tr sio
t ctio
a la tio
ro gra mm
[ ippma
ta l
., 9 9 9 ] ho s o b jcti
a s to s r
a d a la t r s a rch i
itr sio d t ctio . sta da rd s to fda ta to b a dit d, hichi cld s a id
a rit o fitr sio s sim l
a t d i a mil
ita r
t or
iro m t, a s pro id d.
pl
o it d da ta a a il
a blfro m th K ’
9 9 I tr sio
t ctio
o t st.
h ra tra i i gda ta a s a bo tfo r giga b t s o fco mpr ss d bi a r
d mp da ta fro m s
s o f t o r tra ffic. his a s pro c ss d ito a bo t
mil
l
io co
ctio r co rds. imil
a rl, th t o
s o f t st da ta il
d d
a ro d t o mil
l
io co
ctio r co rds. co
ctio is a s q
c of
pa c ts sta rti g a d di g a tso m
l
ld
d tim s, b t
hich da ta flo s
fro m a so rc I a ddr ssto a ta rg tI a ddr ss d rso m
l
ld
d pro to co l
.
a ch co
ctio is l
a b ld a s ith r o rma l
, o r a s a a tta c , ith a ctl o
sp ci c a tta c t p . a chco
ctio r co rd co sists o fa bo t
b t s. tta c s
fa l
lito fo r ma i ca t go ris:
: d ia l
-o f-s r ic , .g. s flo o d;
2 : a tho ri d a cc ss fro m a r mo t ma chi , .g. g ssi gpa ss o rd;
2 : a tho ri d a cc ss to l
o ca ls p r s r (ro o t) priilg s, .g., a rio s
“ b ff r o rflo ” a tta c s;
ro b : s r il
l
a c a d o th r pro bi g, .g., po rtsca i g.
I pra ctic t o da ta ls co ta i i gcl
a ssi d co
ctio s a r a a il
a bl: o
ha s
to b s d fo r a cq iri ga mo d lo fth tra ffic a d th o th r o
fo r t sti g its
Info rma tio n a bo ut K
’
9 9 Intrusio n
t ctio n
http:/ /
. psil
o n.co m/ kdd9 / ta sk.html
.
o nt st is a a il
a bl o n-l
in
at
E v o lu tiv e M o d e lin g o f T C P /IP N e tw o rk T ra ffic fo r In tru s io n D e te c tio n
2 1 9
p rfo rma c s. h disti ctio is impo rta tb ca s th t st lco ta i s a tta c
t p s o to cc rri gi th la r i g l. his is it d d to ma th ta s mo r
r al
istic. I g r 2 a d g r 3,p rfo rma c s o f I
pls ta - a r i g
1
Probe
Dos
U2r
R2l
0.8
0.6
0.4
0.2
0
0
0.02
0.04
0.06
0.08
0.1
0.12
0.14
0.16
0.18
0.2
t ctio n p rfo rma nc s hibit d b
I
pl
us
ta -L a rning o n th
t st da ta . n t nd d r pr s nta tio n o f th da ta a nd a co mpl la rning
a ppro a ch (m ta -l lla rning) ha b n pl
o it d.
i. .
(a s s d i [
ta l
., 9 9 9 ]) a d
o r
’
s da ta a r r sp cti l
sho . I this ca s p rfo rma c s a r sho
b
pl
o iti g c i r p ra ti g
r s a s do
h
a la ti gda ta mi i gto o l
s.I th g r s,th a is r pr s ts th fa l
s al
a rm ra t , i. . th p rc ta g o f’ o rma l
’co
ctio s l
a b ld
a sitr sio s, h r a sth a isr pr s tsth d t ctio ra t ,i. .th p rc ta g
o fitr sio s tha tha b
co rr ctlr co g i d. a chl
i is a sso cia t d to a diffr ta tta c t p . his i d o fgra phis s d to sho ho a cl
a ssi rs b ha io r
d gra d s h ”r l
a i g” its ma tchi gco ditio s: g ra l
l mo r itr sio a r
d t ct d hila l
so co ri gmo r fa l
s al
a rms. H o
r i th ca s o fa s to f
s mbo l
ic cl
a ssi ca tio r ls (itis a ct a l
l b tt r to sa i th ca s o f I
a d
) th r a r o co ditio s to b r l
a d tha tis th r a so o fth
fla tl
i a tth to p.
h r po rt d p rfo rma c s ha b
o bta i d o th co
ctio s o cc rri g i th t st l. h r po rt d gra phs sho simil
a r d t ctio p rfo rma c s,
b t
th mo d l
s a cq ir d b th s st ms, fo r ro b a d
mo t - o - o ca l
( 2l
) a tta c s t p s. I st a d,
’
s mo d lp rfo rms sl
ightl b tt r o
t p a tta c s b t o rsto
s r- o - o o t( 2r) a tta c s.
tco sid r, o ,th mo d l
i ga ppro a ch s pl
o it d b th t o s st ms.
a d to l
fo [
ta l
., 9 9 9 ] r
I
o ra
t d d da ta r pr s ta tio
o fth tcp co
ctio i cldi g, i a dditio to th ba sic tcp fa t r s, d ri d
i fo rma tio s ch a s: th
mb r o fco
ctio s to th sa m ho st i th pa st
2 2 0
F . N e ri
1
Probe
Dos
U2r
R2l
0.8
0.6
0.4
0.2
0
0
0.02
0.04
i. . t ctio n p rfo rma nc s
ta -L a rningha s b n us d).
0.06
0.08
0.1
0.12
0.14
0.16
0.18
0.2
hibitb
Lo n
t stda ta (no a dditio na l
co mpr ss d da ta r pr s nta tio n ha s b n pl
o it d.
t o s co ds (’
co t’
), th
mb r o fco
ctio s to th sa m s r ic , a s th
c rr tco
ctio , i th pa stt o s co ds (’
sr -co t’
). h s fa t r s ha
b
cho s o th ba sis o fth a tho rs p rtis . pr pro c ssi g o fth ra
t or l
o gs is r q ir d i o rd r to
pl
o its this fa t r s.
ra lcl
a ssi rs
(r ls ts) fo r a ch a tta c t p ha b
o bta i d.
t al
l m ta -la r i g,
i. . la r i g a t th cl
a ssi r l l
, ha s b
a ppl
id to pro d c th r po rt d
p rfo rma c s.
,o th co tra r ,ha sb
r a ft ra ppli ga co mpr ssio ma ppi g
to th fa t r a l s, a s d scrib d i pp di .
l th ba sic fa t r s o fa
co
ctio ha b
co sid r d s ch a s: ’
d ra tio ’
, sta ti g th l gth
( mb r o fs co ds) o fth co
ctio , ’
pro to co l
-t p ’
, sta ti g th t p o fth
pro to co l( .g. tcp, dp, tc.), o r ’
src-b t s’
, sta ti g th
mb r o fda ta b t s
fro m so rc to d sti a tio . o a dditio a lm ta -la r i gpha s is c ssa r .
c si s
r po rt p rim tstha tsho th po t tia l
it o fa distrib t d g tic la r r
to th mo d l
i g o f t o r da ta .
o diff r ts t- ps to d a l ith d t cti g
itr sio s ha b
pl
o r d.
a a l d a da ta pa c tr pr s ta tio
pl
o iti gco mpr ssio o fth fa tr’
s a l s i th ffo rtto r d c th co mpl it o fa cq iri g mo d lo fth
tra ffic.
b l
i this b i ga impo rta tr q isit fo r th a to ma tic mo d l
ig
a d th o -l
i d pl
o m to fitr sio d t ctio s st m.
h p rim ta lr s l
ts s ppo rt s o fth co mpr ssio o fth fa t r a l s
a s a a la blm tho d to i cr a s d t ctio p rfo rma c s hila o idi gth s
o fd ri d a d co mpl fa t r s tha ti o l a dditio a lco mp ta tio a lo rh a d.
E v o lu tiv e M o d e lin g o f T C P /IP N e tw o rk T ra ffic fo r In tru s io n D e te c tio n
c
d
2 2 1
ts
a tha s to th a o
impro i gth pa p r.
mo s r i
pp di .
r
d t r pr s t ti
rs ho s
ti
s f lco mm
p r ti
ts h l
p dm i
t tr
h I da ta (a a il
a bl o l
i a thttp:/ / iris.cs. ml
. d ) ha b
co l
lct d
b m a s o fth
til
it. a i g ito a cco tpria c co c r s, th
da ta po rtio o f a ch pa c tha s b
dro pp d. o r a ch pa c ti th da ta s ts
th fo l
l
o i ga ttrib t s a r a a il
a bl:
tim - co
rt d to flo a ti gpts co ds .. hr* 36 + mi * 6 + s cs.
a ddr a d po rt- (j stg trid o f . .256 .256 .po rt) h rstt o
l
ds o fth src
a d d sta ddr ss ma
p th fa a ddr ss, so th co
rt d a ddr ss a s ma d
a s: + * 256 .
fla g - a dd d a ” ” fo r dp da ta (o l ha s l ) - m a s pa c t a s a
a m s r r r q sto r r spo s . h I # a d r sto fda ta is i th ”o p” l
d.
(s tcpd mp d scriptio )
-m a s th r
r o po rts...fro m ”fra gm t d
pa c ts”.
s q - th da ta s q
c
mb r o fth pa c t.
s q2 - th da ta s q
c
mb r o fth da ta p ct d i r t r .
b f- th
mb r o fb t s o fth r c i r b ff r spa c a a il
a bl.
a c - th s q
c
mb r o fth
tda ta p ct d fro m th o th r dir ctio
o this co
ctio .
i -th
mb r o fb t s o fr c i b ff r spa c a a il
a blfro m th o th r dir ctio o this co
ctio .
l - ifa dp pa c t, th l gth.
o p - o ptio a li fo s ch a s (df) ... do o tfra gm t.
a rtic l
a r a tt tio ha s to b ta
h d al
i g ith l
ds l
i ’
o p’tha tco ta i s a l
a rg a mo to f a l s.
pp
di .
d t
c
pr ss d
t r r pr s t ti
o m fa t r s o fth I da ta ma a ss m a l
a rg s to f a l s ith r co ti o s o r discr t . h s l
a rg s ts do impa cto r cl
a ssi ca tio p rfo rma c s o f
th la r d mo d l
s b ca s o fth itri sic diffic l
t o fa cq iri g r l ha i g
a g ra lsco p . h , a r d ctio o fth ra g o fpo t tia la l s is d sira bl
to i cr a s bo th th g ra l
it o fth la r d mo d la d to r d c th la r i g
co mp ta tio a lco mpl it.
al
t r a ti a ppro a ch to this pro blm co sists i a ddi g/ b il
di g mo r
co mpl fa t r s, co mbi i gth ba sic o s,to th o rigi a lda ta r pr s ta tio .
do o tfo l
l
o thisa ppro a chi this o r ,b ca s
b l
i tha tth pr io s
a ppro a chis simplr a d sho l
d b th rstto b a a l d.
2 2 2
F . N e ri
rigina l a l
u
srcpo rt 5
5 srcpo rt
... skipp d t st...
srcpo rt 2
... skipp d t t...
o p co nta ins ” ”
o p co nta ins ”
o ma in”
o p co nta ins
Y
H
L
. o mpr ssio n ma ppinga ppl
id
al
u
srcpo rt=
srcpo rt=
... skipp d t t...
srcpo rt=
... skipp d t t...
o p=
o p=2
o p=3
h n d al
ing ith I
n t o rk da ta .
s a i sta c o fr d ci gth ra g o fth fa t r a l s,co sid rs tha tth
fa t r ’
srcpo rt’(s
pp di
fo r a d scriptio ) ma
irt a l
l a ss m a
it g r mb r fro m to 6 5536 . l
so , th fa t r ’
o p’ma a ss m h dr ds
o fdiscr t a l s. a i g ito a cco tba sic o ldg a bo tth do ma i ,
ma a l
ld l
o p d th r d ctio ma ppi g sho
i a bl4. his ma ppi g is
o tto b co sid r d a s th b sto
b ta s a pro o ftha ta simplr d ctio o f
th fa t r a l s ma po siti l impa cto r th r co g itio ca pa bil
itis.
r c s
[ o h n, 9 9 5] o h n, . ( 9 9 5). a st ff cti rulinductio n. In r di gs fI t rti
hi
ri g
fr
, La k a ho ,
. o rga n Ka ufma nn.
[ nning, 9 7 ]
nning, . ( 9 7 ). n intrusio n d t ctio n mo d l
.I
r s ti
ft r
gi ri g, - 3(2):222–232.
[ ho sh ta l
., 9 9 9 ] ho sh, ., ch a rt ba rd, ., a nd cha t , . ( 9 9 9 ). L a rning
pro gra m b ha io rpro fils fo rintrusio n d t ctio n. In
I
rksh
I tr si
t ti
d
t rk
it ri g.
I
sso cia tio n.
[ io rda na a nd
ri, 9 9 5] io rda na , . a nd
ri, . ( 9 9 5). a rch-int nsi co nc pt
inductio n.
ti r
m t ti , 3 (4):37 5–46 .
[ ol
db rg, 9 9 ] o l
db rg, . ( 9 9 ).
ti
grithms i
rh
timi ti
d
hi
r i g. ddiso n- sl , a ding, a .
[Kuma r a nd pa ffo rd, 9 9 4] Kuma r, . a nd pa ffo rd, . ( 9 9 4).
pa tt rn ma tching
mo d lfo r misus d t ctio n. In
ti
m tr
rit
fr ,pa g s –2 ,
al
timo r .
[La n a nd ro dl , 9 9 7 ] La n , . a nd ro dl , . ( 9 9 7 ). n a ppl
ica tio n o fma chin
la rningto a no ma ld t ctio n. In ti
I f rm ti
st ms
rit
fr ,
al
timo r .
[La n a nd ro dl , 9 9 ] La n , .a nd ro dl , .( 9 9 ). ppro a ch sto o nl
in la rninga nd co nc ptua l
driftfo r us rid ntifica tio n in co mput rs curit. chnica l
r po rt,
a nd th
La bo ra to r , urdu
ni rsit, o a st
9 - 2.
[L
ta l
., 9 9 9 ] L , ., to l
fo , ., a nd o k, K. ( 9 9 9 ). ining in a da ta -flo
niro nm nt: p rinc in n t o rk intrusio n d t ctio n. In
dg is
r
d
t iig
’ , pa g s 4– 24.
r ss.
E v o lu tiv e M o d e lin g o f T C P /IP N e tw o rk T ra ffic fo r In tru s io n D e te c tio n
2 2 3
[Lippma nn ta l
., 9 9 9 ] Lippma nn, ., unningha m, ., rid, ., ra f, I., K nda l
l
,
K.,
bst r, .,a nd Z issma nn, . ( 9 9 9 ). sul
ts o fth
9 9 o ffl in intrusio n d t ctio n a l
ua tio n. In
t d
s i I tr si
t ti
I ’ ,
. La fa tt , I . urdu
ni rsit.
[ icha l
ski, 9 3] icha l
ski, . ( 9 3). th o r a nd m tho do l
o g o finducti la rning. In icha l
ski, ., a rbo n l
l
, ., a nd itch l
l
, ., dito rs,
hi
ri g
rtifi i I t ig
r h, o l
um I, pa g s 3– 34. o rga n Ka ufma nn,Lo s
l
to s,
.
[ ria nd a itta , 9 9 6 ]
ri, . a nd a itta , L. ( 9 9 6 ).
pl
o ringth po r o fg n tic
s a rchin la rnings mbo l
ic cl
a ssifi rs. I
r s.
tt r
sis d
hi
I t ig ,
I- : 35– 42.
[ uinl
a n, 9 9 3] uinl
a n, . . ( 9 9 3). 4.
rgrms f r
hi
r i g. o rga n
Ka ufma nn, a l
ifo rnia .
M u ltim o d a l P e r fo r m a n c e P r o file s o n th e A d a p tiv e
D is tr ib u te d D a ta b a s e M a n a g e m e n t P r o b le m
M . O a te s 1, D . C o rn e 2, a n d R . L o a d e r
1
2
2
B ritis h T e le c o m A d a s tra l P a rk , M a rtle s h a m H e a th , S u ffo lk , E n g la n d , IP 5 3 R E
D e p a rtm e n t o f C o m p u te r S c ie n c e , U n iv e rs ity o f R e a d in g , R e a d in g , R G 6 6 A Y
A b s tr a c t. P re v io u s p u b lic a tio n s b y th e a u th o rs h a v e d e m o n s tra te d a b im o d a l
p e rfo rm a n c e p ro file fo r s im p le e v o lu tio n a ry s e a rc h o n v a ria n ts o f th e A d a p tiv e
D is trib u te d D a ta b a s e M a n a g e m e n t P ro b le m (A D D M P ) a n d o th e r p ro b le m s o v e r
a ra n g e o f e v a lu a tio n lim its . T h is p a p e r e x a m in e s a n a n o m a ly s e e n in o n e o f
th e s e p ro file s a n d to g e th e r w ith re s u lts fro m a ra n g e o f o th e r p ro b le m s , s h o w s
th a t w ith s u ffic ie n tly h ig h e v a lu a tio n lim its , a m u ltim o d a l p e rfo rm a n c e p ro file is
a p p a re n t in s e a rc h s p a c e s w ith s ig n ific a n t n u m b e rs o f d e c e p tiv e lo c a l o p tim a .
T h is is p a rtic u la rly a p p a re n t in th e p e rfo rm a n c e p ro file o f th e H ie ra rc h ia l If a n d
o n ly If p ro b le m (H -IF F ) w h e re th e re g u la r s tru c tu re o f th e s e a rc h s p a c e
p ro d u c e s s e v e ra l d is tin c t p e a k s a n d tro u g h s in th e p e rfo rm a n c e p ro file , p o s s ib ly
in d ic a tiv e o f a ra n g e o f s p e c ific ‘fitn e s s b a rrie rs ’ w h ic h a re s u rm o u n ta b le b y
s p e c ific ra te s o f m u ta tio n . T h is o b s e rv a tio n c o u ld p ro v e im p o rta n t in g e n e ra l
E A p a ra m e te r tu n in g o v e r a ra n g e o f p ro b le m s w ith s im ila r c h a ra c te ris tic s .
F u rth e r, th e e x is te n c e o f o p tim a l m u ta tio n ra te s in d u c in g a m in im u m in s ta n d a rd
d e v ia tio n o f ru n -tim e , is o f c ritic a l im p o rta n c e in th e a p p lic a tio n o f E A s to re a ltim e , re a l-w o rld p ro b le m s .
1 I n tr o d u c tio n
M a n y r e a l w o r ld , r e a l tim e a p p lic a tio n s o f E v o lu tio n a r y A lg o r ith m s ( E A s ) [ 1 ,3 ,4 ,6 ]
re q u ire th e s e a rc h p ro c e s s to re lia b ly p ro d u c e q u a lity s o lu tio n s in a fix e d n u m b e r o f
e v a lu a tio n s . T o in c re a s e th e lik e lih o o d o f th is , it is o fte n e s s e n tia l to tu n e th e
a lg o rith m b y s e le c tio n o f s u ita b le p a ra m e te r v a lu e s s u c h a s p o p u la tio n s iz e a n d
m u ta tio n ra te . O n e s u c h a p p lic a tio n w h ic h h a s b e e n s tu d ie d e x te n s iv e ly b y th e a u th o rs
[ 1 1 ,1 2 ,1 3 ,1 4 ,1 5 ,1 6 ,1 7 ] is th e A d a p tiv e D is tr ib u te d D a ta b a s e M a n a g e m e n t P r o b le m
(A D D M P ) w h ic h a tte m p ts to b a la n c e a n u m b e r o f u s e r lo a d s o n to a ra n g e o f a v a ila b le
s e rv e rs o v e r a c o m m u n ic a tio n s n e tw o rk to m a x im is e a g iv e n q u a lity o f s e rv ic e m e tric .
T h e p ro b le m is re d u c e d to o n e o f c o m b in a to ria l o p tim is a tio n , w h e re a ‘s o lu tio n
v e c to r’ g e n e ra te d b y th e E A d e fin e s fo r e a c h c lie n t n o d e (d e te rm in e d b y lo c u s ), w h ic h
s e rv e r to u s e (d e te rm in e d b y a lle le v a lu e ) u s in g a n a tu ra l k -a ry re p re s e n ta tio n (w h e re
a lle le s a re in te g e rs in th e ra n g e 1 th ro u g h k ). T h is e ffe c tiv e ly d e te rm in e s a ro u te
th ro u g h th e c o m m u n ic a tio n s n e tw o rk , a n d th e c o m b in e d d e la y im p o s e d b y th e s e rv e r
p ro c e s s in g tim e a n d th e c o m m u n ic a tio n s la te n c y is c a lc u la te d b y a p e rfo rm a n c e m o d e l
u tilis in g th e p rin c ip le s o f M M 1 q u e u in g a n d L ittle ’s L a w . T h e p ro b le m h a s b e e n
lik e n e d to a fo r m o f c o m p le x ‘b in -p a c k in g ’ a n d fu r th e r d e ta ils a r e a v a ila b le in [ 1 2 ,1 3 ]
w ith e x a m p le c o d e a n d d a ta s e ts a v a ila b le a t h ttp ://w w w .d c s .n a p ie r /a c /u k /e v o n e t p a r t
S . C a g n o n i e t a l. ( E d s .) : E v o W o r k s h o p s 2 0 0 0 , L N C S 1 8 0 3 , p p . 2 2 4 − 2 3 4 , 2 0 0 0 .
S p rin g e r-V e rla g B e rlin H e id e lb e rg 2 0 0 0
M u ltim o d a l P e rfo rm a n c e P ro file s
2 2 5
o f th e E C T E L N E T w e b s ite (E C T E L N E T is th e T e le c o m m u n ic a tio n s s u b g ro u p o f th e
E u ro p e a n N e tw o rk o f E x c e lle n c e in E v o lu tio n a ry C o m p u ta tio n )
C lie n t :
W h ic h S e rv e r to u s e :
1
2
1
3
4
4
3
5
4
6
3
7
2
8
2
9
3
4
1 0
1
F ig . 1 . E x a m p le s o lu tio n v e c to r.
M a n y in s ta n c e s o f th e A D D M P e x is t, w ith d iffe re n t u s e r lo a d p a tte rn s , d iffe re n t s e rv e r
p ro file s a n d d iffe re n t c o m m u n ic a tio n s n e tw o rk to p o lo g ie s , a n d s o m e o f th e s e a re
d is c u s s e d in [1 3 ]. S o m e o f th e s e in s ta n c e s h a v e b e e n s h o w n to p re s e n t c o n s id e ra b le
d iffic u lty to a r a n g e o f a lg o r ith m s [ 1 1 ,1 4 ] a n d in g e n e r a l, a n E v o lu tio n a r y a lg o r ith m ,
c o rre c tly tu n e d , h a s b e e n s h o w n to g iv e th e m o s t c o n s is te n t p e rfo rm a n c e . P re v io u s ly
p u b lis h e d s tu d ie s o f o n e o f th e s e in s ta n c e s [ 1 6 ,1 7 ] s h o w a th r e e d im e n s io n a l
p e rfo rm a n c e p ro file fo r a s im p le g e n e ra tio n a l B re e d e r E A [7 ] u s in g 5 0 % e litis m ,
u n ifo rm c ro s s o v e r [1 9 ], a n d u n ifo rm ly d is trib u te d a lle le re p la c e m e n t m u ta tio n a t a
fix e d ra te p e r g e n e . E x p e rim e n ts w e re ru n o v e r a w id e ra n g e o f m u ta tio n ra te s , fro m
1 E -7 to 0 .8 3 , th e la tte r e ffe c tiv e ly d e g e n e r a tin g th e E A in to r a n d o m s e a r c h , a n d a
w id e ra n g e o f p o p u la tio n s iz e s ty p ic a lly 1 0 to 5 0 0 m e m b e rs in s te p s o f 1 0 . E a c h ru n
w a s a llo w e d 2 0 ,0 0 0 e v a lu a tio n s a n d e a c h e x p e r im e n t w a s r e p e a te d 5 0 tim e s . D u r in g
e a c h ru n , a n o te w a s m a d e o f th e fitn e s s o f th e b e s t s o lu tio n fo u n d , a n d th e e v a lu a tio n
n u m b e r a t w h ic h it w a s firs t fo u n d . F o r th e 5 0 ru n s a ll w ith th e s a m e e x p e rim e n ta l
p a ra m e te rs , th e m e a n o f th e ‘tim e to b e s t s o lu tio n ’ w a s p lo tte d a g a in s t p o p u la tio n s iz e
a n d m u ta tio n ra te , a n d th e re s u ltin g g ra p h is s h o w n in F ig u re 2 . A s w a s c o m m e n te d o n
in [1 6 ], th is p ro d u c e s , fo r a g iv e n , lo w p o p u la tio n s iz e , a b im o d a l p e rfo rm a n c e p ro file ,
w h e re th e n u m b e r o f e v a lu a tio n s ta k e n to firs t fin d th e b e s t s o lu tio n to b e fo u n d in th e
ru n firs t ris e s , th e n fa lls , th e n ris e s a g a in b e fo re fa llin g in to ra n d o m s e a rc h . T h e s e
fe a tu re s a re in te rp re te d a s fo llo w s : th e lo w v a lu e s o f e v a lu a tio n s u s e d a t lo w m u ta tio n
ra te s re p re s e n t p re m a tu re c o n v e rg e n c e o n p o o r s o lu tio n s , ra p id ly e x h a u s tin g th e
lim ite d d iv e rs ity a v a ila b le in th e in itia l p o p u la tio n w h ic h c a n n o t b e s u p p le m e n te d b y
m u ta tio n d u e to its lo w ra te ; a s m u ta tio n ra te s in c re a s e , th e E A is a b le to p ro g re s s
fu rth e r w ith its s e a rc h , a s d e p le te d a lle le v a lu e s a re re -in je c te d in to th e p o p u la tio n
w ith in c re a s in g fre q u e n c y b y th e in c re a s in g m u ta tio n ra te ; a p o in t is re a c h e d w h e re
th e re is a s u ffic ie n tly h ig h ra te o f m u ta tio n to a llo w th e s e a rc h to ty p ic a lly u tilis e in
e x c e s s o f 8 0 % o f th e a v a ila b le e v a lu a tio n s in te rm s o f fitn e s s im p ro v e m e n t – th is is
th e le ft h a n d p e a k o n F ig u re 2 ; fu rth e r in c re a s e s in m u ta tio n ra te a llo w g o o d s o lu tio n s
to th e p ro b le m to b e fo u n d in fe w e r e v a lu a tio n s , le a d in g to a p o in t w h e re th e m u ta tio n
ra te a llo w s g o o d s o lu tio n s to b e fo u n d in a m in im u m o f e v a lu a tio n s – th e tro u g h in
F ig u re 2 ; a s m u ta tio n ra te is in c re a s e d fu rth e r, p e rfo rm a n c e d e te rio ra te s a s th e
m u ta tio n ra te b e g in s to d e s tro y p ro g re s s a lm o s t a s fa s t a s it is m a d e – th e rig h t h a n d
p e a k o n F ig u re 2 ; fin a lly a t v e ry h ig h ra te s o f m u ta tio n , in e x c e s s o f 4 0 % ,
p e rfo rm a n c e d e te rio ra te s to ra n d o m s e a rc h w ith th e m e a n n u m b e r o f e v a lu a tio n s u s e d
te n d in g to 5 0 % o f th o s e a llo w e d . T h is in te rp re ta tio n is s u p p o rte d b y e x a m in a tio n o f
th e m e a n fitn e s s e s o f s o lu tio n s fo u n d a s m u ta tio n ra te s a re in c re a s e d w h ic h s h o w :
p o o r fitn e s s a t lo w m u ta tio n ra te s ; a s te a d y im p ro v e m e n t in fitn e s s fo r m u ta tio n ra te s
2 2 6
M . O a te s , D . C o rn e , a n d R . L o a d e r
a b o v e th e firs t p e a k in F ig u re 2 ; a p la te a u o f g o o d fitn e s s c o in c id in g
m in im a l e v a lu a tio n s in F ig u re 2 ; a n d a d e te rio ra tio n in fitn e s s
a p p ro a c h th e rig h t h a n d s id e o f F ig u re 2 . T h is c o -in c id e n c e o f tro
a n d p e a k in fitn e s s o c c u rs a t m u ta tio n ra te s a ro u n d 1 / L (w
c h r o m o s o m e le n g th , h e r e e q u iv a le n t to a r o u n d 2 .5 % ) , w h ic h
th e o re tic a l s tu d ie s b y , a m o n g s t o th e rs , B a e c k [1 ], M ü h le n b e in [8 ] a
s tu d ie s in c lu d in g [ 2 ,8 ,9 ,1 8 ] b y D e b e t a l, v a n N im w e g e n a n d C r u tc
a n d th is a n d o th e r fe a tu r e s a r e d is c u s s e d in m o r e d e ta il in [ 1 6 ,1 7 ] .
w ith th e tro u g h o f
a s m u ta tio n ra te s
u g h in e v a lu a tio n s
h e re L = b in a ry
is in s u p p o rt o f
n d m a n y e m p iric a l
h fie ld , O a te s e t a l,
O f p a rtic u la r in te re s t h o w e v e r in th is p a p e r is th e a n o m a ly w h ic h c a n b e s e e
le ft h a n d e d g e o f th e rig h t h a n d p e a k in F ig u re 2 a t lo w p o p u la tio n s iz e s (to w
b a c k o f th e fig u re ). H e re , th e ris e in m e a n e v a lu a tio n n u m b e r w ith in c re a s in g
c a n b e s e e n to b e n o t s m o o th , w ith a rid g e fe a tu re a p p a re n t o v e r a ra n g e
p o p u la tio n s iz e s . S u c h a n a n o m a ly w a s n o t s e e n a t lo w e r e v a lu a tio n lim its
e x te n s iv e in v e s tig a tio n s o f s im p le ‘u n im o d a l’ s e a rc h s p a c e s s u c h a s p re s e n te
‘M a x -O n e s ’ p ro b le m . T h e re m a in d e r o f th is p a p e r e x p lo re s th is p h e n o m e n o n
d e ta il, w ith s e c tio n 2 d e s c rib in g th re e o th e r te s t p ro b le m s a n d th e ir e x p e rim
u p ; s e c tio n 3 d e s c r ib in g th e r e s u lts a t b o th 2 0 ,0 0 0 a n d 1 m illio n e v a lu a tio n s ;
d ra w in g s o m e p re lim in a ry c o n c lu s io n s fro m th e s e re s u lts a n d s e c tio n s
a c k n o w le d g in g s u p p o rt a n d re fe re n c e d m a te ria l.
n o n th e
a rd s th e
m u ta tio n
o f lo w e r
, n o r o n
d b y th e
in m o re
e n ta l s e t
s e c tio n 4
5 a n d 6
2 M e th o d
T o e x a m in e th is
ra n g e o f ‘s ta n d
v a lu e s , a n d a s e
lim it a n d m o re r
a n o m a ly in
a rd ’ te s t p
c o n d s e rie s
e fin e d s c a le
m o r
ro b le
o f e
o f m
e d e ta il,
m s o v e
x p e rim e
u ta tio n
a s e rie s o f e x p e rim e n ts w a s c o n d u c te d w ith a
r s im ila r ra n g e s o f e x p e rim e n ta l p a ra m e te r
n ts c o n d u c te d w ith a m u c h h ig h e r e v a lu a tio n
ra te s .
T h e firs t p ro b le m to b e lo o k e d a t w a s th e ‘M a x -O n e s ’ p ro b le m w h e re in a b in a ry
s trin g , fitn e s s is c a lc u la te d to b e th e n u m b e r o f ‘1 ’s p re s e n t in th e c h ro m o s o m e . W ith
s ta n d a rd re p re s e n ta tio n , s te a d y s ta te , s in g le 3 w a y to u rn a m e n t s e le c tio n [3 ], u n ifo rm
c ro s s o v e r a n d s im p le p e r g e n e m u ta tio n , th is p re s e n ts a s im p le u n im o d a l s e a rc h
la n d s c a p e w h ic h c a n b e a s c e n d e d b y e v e n th e s im p le s t ‘h illc lim b in g ’ a lg o rith m . 5 0
tria ls w e re ru n fo r e a c h p a ra m e te r s e ttin g , w h ic h ra n g e d fro m p o p u la tio n s iz e s o f 2 to
1 0 0 a n d p e r g e n e m u ta tio n r a te s fr o m 1 E -7 to .8 3 . M e a n e v a lu a tio n s to fir s t fin d th e
b e s t s o lu tio n fo u n d in 2 0 ,0 0 0 e v a lu a tio n s w e r e n o te d . C h r o m o s o m e le n g th fo r r e s u lts
s h o w n h e re w a s 5 0 b its , h o w e v e r s im ila r e x p e rim e n ts h a v e a ls o b e e n d o n e a t 3 3 , 3 0 0
a n d 1 6 3 0 b its , g iv in g c o rre s p o n d in g re s u lts .
T h e s e c o n d p ro b le m to b e in v e s tig a te d w a s a 6 4 b it im p le m e n ta tio n o f W a ts o n ’s H I F F ( H ie r a r c h ic a l I f a n d o n ly I f) p r o b le m [ 2 0 ,2 1 ] , w h ic h in c r e a s in g ly r e w a r d s e v e r
la rg e r a lig n e d b lo c k s o f c o n tig u o u s ‘1 ’s o r ‘0 ’s a n d c a n b e re p re s e n te d a s :
M u ltim o d a l P e rfo rm a n c e P ro file s
f(B ) =
1 ,
|B | + f ( B L ) + f ( B
f(B L ) + f(B R ),
R
2 2 7
i f |B | = 1
i f ( | B | > 1 ) a n d ( "i { b i = 0 } o r "i { b i = 1 } ) ,
o th e rw is e
),
w h e r e B i s a b l o c k o f b i t s , { b 1 , b 2 , … b n } , |B | i s t h e s i z e o f t h e b l o c k = n , b i i s t h e i t h e l e m e n t o f B , a n d B L
a n d B R a r e t h e l e f t a n d r i g h t h a l v e s o f B ( i . e . B L = { b 1 , … b n /2 } , B R = { b n /2 + 1 , … b n } . N m u s t b e a n i n t e g e r
p o w e r o f 2 .
H -IF F th e re fo re h a s tw o g lo b a l o p tim a , o n e a t a ll ‘1 ’s a n d o n e a t a ll ‘0 ’s . T h e re a re
s e c o n d a ry o p tim a a t s trin g s o f 3 2 ‘1 ’s fo llo w e d b y 3 2 ‘0 ’s a n d v ic e v e rs a , a n d a ra n g e
o f s u b -o p tim a a t c o m b in a tio n s o f a lig n e d b lo c k s o f ‘1 ’s a n d ‘0 ’s e a c h o f le n g th 1 6 , 8 ,
4 a n d 2 . T h e ‘s e a rc h la n d s c a p e ’ c a n b e c o n s id e re d to b e ru g g e d b u t in a h ig h ly
s tru c tu re d fa s h io n . S im p le h illc lim b e rs p e rfo rm in a d e q u a te ly o n th is s u rfa c e a s th e
b a s in o f a ttra c tio n o f th e g lo b a l o p tim a is v e ry s m a ll, a n d th e re a re m a n y lo c a l o p tim a
w h e n s e e n b y s in g le p o in t m u ta tio n a lo n e . E v e n s ta n d a rd o n e -p o in t a n d tw o -p o in t
c ro s s o v e r o p e ra to rs h a v e b e e n s e e n to re q u ire v e ry h ig h p o p u la tio n s iz e s to a c h ie v e
re a s o n a b le a n d c o n s is te n t p e rfo rm a n c e o n th is p ro b le m u n le s s a n a p p ro p ria te d iv e rs ity
m a in te n a n c e te c h n iq u e is e m p lo y e d .
T h e th ird p ro b le m re p o rte d o n is th is p a p e r is a 5 0 -8 K a u fm a n N K la n d s c a p e [5 ]
g e n e ra te d b y d e fin in g a ta b le o f ra n d o m n u m b e rs o f d im e n s io n 5 0 b y 5 1 2 . F o r a
b in a ry c h ro m o s o m e o f le n g th N = 5 0 , s ta rtin g a t e a c h o f th e 5 0 g e n e p o s itio n s , K + 1
(h e re 9 ) w e ig h te d c o n s e c u tiv e g e n e s a re u s e d to g e n e ra te a n in d e x (in th e ra n g e 0 5 1 1 ), a n d th e 5 0 v a lu e s s o in d e x e d fro m th e ta b le a re s u m m e d to g iv e a fitn e s s v a lu e .
F o r N K la n d s c a p e s w ith a K v a lu e o f 0 , e a c h g e n e p o s itio n c o n trib u te s in d iv id u a lly to
fitn e s s , h e n c e p ro d u c in g a u n im o d a l la n d s c a p e to s in g le p o in t m u ta tio n h illc lim b e rs .
H o w e v e r a s K is in c re a s e d , a n y s in g le p o in t m u ta tio n w ill a ffe c t K + 1 in d e x e s , ra p id ly
in tro d u c in g p o s itio n a l lin k a g e in to th e p ro b le m a n d p ro d u c in g a ru g g e d a n d
in c re a s in g ly u n s tru c tu re d la n d s c a p e .
T h e A D D M P , M a x -O n e s , H -IF F a n d N K 5 0 -8 p ro b le m s w e re
e v a lu a tio n s in th e firs t s e rie s o f e x p e rim e n ts , w ith re s u lts s h o w n
re s p e c tiv e ly . A s e c o n d s e rie s o f e x p e rim e n ts w a s th e n p e rfo
1 ,0 0 0 ,0 0 0 e v a lu a tio n s , fo r A D D M P ( F ig u r e 6 ) o v e r a r a n g e o f
3 0 0 in s te p s o f 1 0 ) a n d fo r M a x -O n e s a n d H -IF F a t a fix e d
(F ig u re s 7 a n d 8 ). F o r e a s ie r c o m p a ris o n o f re s u lts , th e A D D M P
ru n s a t a p o p u la tio n s iz e o f 2 0 a re s h o w n m o re c le a rly in F ig u re
a n d m in u s o n e s ta n d a rd d e v ia tio n s e e n o v e r th e 5 0 tria ls a n d
s h o w n in e a c h o f th e fix e d p o p u la tio n s iz e g ra p h s .
ll tr ia lle d a t 2 0 ,0 0 0
F ig u r e s 2 ,3 ,4 a n d 5
e d w ith a lim it o f
o p u la tio n s iz e s (1 0 p o p u la tio n s iz e o f 2 0
1 ,0 0 0 ,0 0 0 e v a lu a tio n
9 . P lo ts s h o w in g p lu s
m e a n fitn e s s a re a ls o
U n
w e
c ro
th e
D D
t G
o in
b le
le s s
re c
sso v
o b v
o th
o n d
e r.
io u
e rw is e
u c te d
T h e H
s in a p p
s ta te d , a ll e x
w ith a s te a d
-IF F e x p e rim
ro p ria te n e s s
p e rim
y s ta
e n ts
o f th e
e n ts
te , 3
w e re
u n if
(w ith
w a y
c o n d
o rm o
th e e x c
s in g le
u c te d u
p e ra to r
e p tio n o f
to u rn a m
s in g o n e
o n th is p
A
e n
-p
ro
a
in
rm
p
M P a t 2 0 K e v a ls )
A u s in g u n ifo rm
t c ro s s o v e r d u e to
m .
M . O a te s , D . C o rn e , a n d R . L o a d e r
Mean Evaluations
12000
20000
18000
16000
10000
10
50
90
130
170
210
250
8000
6000
20
80
140
200
12000
10000
8000
260
6000
320
4000
380
330
2000
440
370
0
500
410
290
4000
14000
Pop Size
14000
18000-20000
16000-18000
14000-16000
12000-14000
10000-12000
8000-10000
6000-8000
4000-6000
2000-4000
0-2000
2E 0
-0
8E 7
3. 07
2E
1. -0
28 6
E
5. -0
12 5
E
0. -0
00 5
0
0. 20
00 5
0
0. 81
00 9
3
0. 27
01 7
3
0. 10
05 7
2
0. 42
20 9
9
0. 71
83 5
88
61
16000
16000-18000
14000-16000
12000-14000
10000-12000
8000-10000
6000-8000
4000-6000
2000-4000
0-2000
Mean Evaluations
18000
Pop Size
2 2 8
2000
450
490
Mutation
2E 0
-0
8E 7
3. -0 7
2E
1. -0
28 6
E
5. -0
12 5
0. E-0
00 5
0
0. 20
00 5
0
0. 81
00 9
3
0. 27
01 7
3
0. 10
05 7
2
0. 42
20 9
9
0. 71
83 5
88
61
0
Mutation Rate
F ig . 2 . A D D M P a t 2 0 K e v a lu a tio n s
F ig . 4 . H -IF F 6 4 a t 2 0 K e v a lu a tio n s
18000
16000
12000
14000
2
16
30
8000
6000
44
58
4000
Pop Size
10000
12000
10000
8000
2
14
26
38
50
6000
4000
62
72
Mutation Rate
F ig . 3 . O n e M a x a t 2 0 K e v a lu a tio n s
86
98
0
3.
2E
2. 06
56
E05
0.
00
02
05
0.
00
16
38
0.
01
31
07
0.
10
48
5
0.
8
83
88
61
100
0
0
74
-0
7
86
2000
4E
2000
0
4E
-0
3. 7
2E
2. -06
56
E
0. -05
00
02
0
0.
00 5
16
3
0.
01 8
31
0
0.
10 7
48
5
0.
83 8
88
61
Mean Evaluations
14000
16000-18000
14000-16000
12000-14000
10000-12000
8000-10000
6000-8000
4000-6000
2000-4000
0-2000
16000
Mean Evaluations
18000
18000-20000
16000-18000
14000-16000
12000-14000
10000-12000
8000-10000
6000-8000
4000-6000
2000-4000
0-2000
Mutation
F ig . 5 . N K 5 0 -8 a t 2 0 K e v a lu a tio n
Pop Size
20000
M e a n E v a ls
+sd
-s d
F itn e s s
1000000
Mean Evaluations
Pop
150
Size
10
8 7 5 0 0 0 -1 0 0 0 0 0 0
7 5 0 0 0 0 -8 7 5 0 0 0
6 2 5 0 0 0 -7 5 0 0 0 0
5 0 0 0 0 0 -6 2 5 0 0 0
3 7 5 0 0 0 -5 0 0 0 0 0
2 5 0 0 0 0 -3 7 5 0 0 0
1 2 5 0 0 0 -2 5 0 0 0 0
0 -1 2 5 0 0 0
290
800000
350
400000
300
200000
250
0
200
-2 0 0 0 0 0
150
M u ta tio n R a te
e v a lu a tio n s
F ig . 8 . H -IF F 6 4 a t 1 M
e v a ls , P o p s iz e = 2 0
3.
2E
-0
6
2.
56
E05
0.
00
02
05
0.
00
16
38
0.
01
31
07
0.
10
48
58
0.
83
88
61
4E
0
40
-0
7
0
M u ta t io n R a t e
e v a ls , p o p s iz e = 2 0
0
-2 0 0 0 0 0
M u ta t io n R a t e
F ig . 9 . A D D M P a t 1 M
e v a ls , p o p s iz e = 2 0
M e a n E v a ls
+sd
-s d
F itn e s s
2 2 9
F ig . 7 . O n e M a x a t 1 M
200000
1
42
400000
0.
1
44
600000
0.
01
200000
46
0.
00
1
400000
48
0.
00
01
R a w E v a ls
+sd
-sd
F itn e s s
600000
M u ltim o d a l P e rfo rm a n c e P ro file s
800000
3500
3000
2500
2000
1500
1000
500
0
800000
1E
-0
7
0.
00
00
01
0.
00
00
1
50
Mean Evaluations
1000000
1000000
Mean Evaluations
400
600000
M u t a t io n R a te
F ig . 6 . A D D M P a t 1 M
450
1.
00
E
3. -0
36 7
E1. 0
13 7
E
3. -0
81 6
E
1. -0
28 6
E4. 0
31 5
E
1. -0
45 5
E
4. -0
87 4
E1. 0
64 4
E
5. -0
51 3
E
1. -0
85 3
E
6. -0
23 2
E
2. -0
10 2
E
7. -0
05 1
E01
1. 0
6E
5. -06
12
E
0. -05
00
16
3
0.
05 8
24
29
Mean Evaluations
1000000
875000
750000
625000
500000
375000
250000
125000
0
2 3 0
M . O a te s , D . C o rn e , a n d R . L o a d e r
3 R e s u lts
A n in te rp re ta tio n o f F ig u re 2 h a s a lre a d y b e e n g iv e n in S e c tio n 1 o f th is p a p e r, a n d a
s im ila r p e r fo r m a n c e p r o file a t 2 0 ,0 0 0 e v a lu a tio n s c a n a ls o b e s e e n fo r th e M a x -O n e s
p ro b le m in F ig u re 3 . A g a in , a t lo w p o p u la tio n s iz e s a n d lo w m u ta tio n ra te s , th e s e a rc h
is s e e n to s ta ll a fte r v e ry fe w e v a lu a tio n s , a s d iv e rs ity in th e in itia l p o p u la tio n is
ra p id ly d e p le te d . V e ry lo w ra te s o f m u ta tio n p re v e n t re -o c c u rre n c e o f p o te n tia lly
u s e fu l a lle le s a n d th u s th e m e a n n u m b e r o f e v a lu a tio n s u s e d is v e ry lo w . A s m u ta tio n
ra te s in c re a s e , th e s e a rc h is a b le to p ro g re s s to h ig h e r m e a n n u m b e rs o f e v a lu a tio n s ,
u n til a p o in t is re a c h e d w h e re m u ta tio n ra te s a llo w , o n a v e ra g e , m o s t o f th e
e v a lu a tio n s to b e u tilis e d in th e s e a rc h fo r g o o d s o lu tio n s . A s m u ta tio n ra te s in c re a s e
fu rth e r, th e s e g o o d s o lu tio n s a re fo u n d in le s s a n d le s s e v a lu a tio n s , u n til th e tro u g h
p o in t in F ig u re 3 is re a c h e d w h e re g lo b a l o p tim u m s o lu tio n s a re fo u n d in th e
m in im u m n u m b e r o f e v a lu a tio n s . B e y o n d th is ra te o f m u ta tio n , p e rfo rm a n c e
d e te rio ra te s e v e n tu a lly d e g e n e ra tin g in to ra n d o m s e a rc h . H o w e v e r in c o n tra s t to
A D D M P (F ig 2 ), th is c a n b e s e e n to d e te rio ra te s m o o th ly , w ith n o a n o m a ly o n th e le ft
h a n d e d g e o f th e rig h t h a n d p e a k . T h e m u ta tio n ra te in d u c in g th e firs t p e a k c a n b e
s e e n to b e e ffe c tiv e ly in d e p e n d e n t o f p o p u la tio n s iz e (o v e r its v is ib le ra n g e ) a n d
o c c u r s a t a r a te a r o u n d .0 0 0 1 . T h e tr o u g h r a te is c e n tr e d a r o u n d a m u ta tio n r a te o f .0 5
(= 5 % c h a n c e o f a p p ly in g a N e w R a n d o m A lle le (la te r re fe rre d to a s N R A ), th is c o u ld
b e w r itte n a s 2 .5 % c h a n c e o f a p p ly in g ‘g u a r a n te e d flip ’ m u ta tio n
1 / 5 0 ).
F ig u r e 4 s h o w s th e p e r fo r m a n c e p r o file fo r th e 6 4 b it H -I F F p r o b le m a g a in a t 2 0 ,0 0 0
e v a lu a tio n s . H e re , in s ta rk c o n tra s t to F ig u re 3 , a th ird p e a k c a n b e s e e n in th e
p e rfo rm a n c e p ro file , p a rtic u la rly a t lo w p o p u la tio n s iz e s . It is b e lie v e d th a t th e le ft
h a n d tro u g h c o rre s p o n d s to a lo w m u ta tio n ra te w h ic h re p re s e n ts a lo w e r b o u n d th a t is
u s e fu l fo r fin d in g a c e rta in s e t o f s u b -o p tim a in th e p ro b le m s p a c e – fo r e x a m p le
b lo c k s o f s iz e 4 . H o w e v e r, a s th e m u ta tio n ra te is in c re a s e d , th is ra te b e c o m e s s u b o p tim a l fo r th is b lo c k s iz e , a n d a n o th e r m u ta tio n ra te is fo u n d w h ic h p ro v e s u s e fu l in
fin d in g la rg e r b lo c k s iz e s – p o s s ib ly o f s iz e 8 o r 1 6 . T h is s u p p o s itio n is s u p p o rte d b y
th e fa c t th a t a p lo t o f m e a n fitn e s s e s fo u n d a g a in s t m u ta tio n ra te s h o w s m a rk e d s te p
im p ro v e m e n ts b e tw e e n e a c h o f th e s e tro u g h s in e v a lu a tio n s u s e d , im p ly in g th a t th e
E A w a s s u d d e n ly a b le to fin d b e tte r o p tim a a t e a c h o f th e s e s p e c ific m u ta tio n ra te
b o u n d a rie s . T h is o b s e rv a tio n c a n b e s e e n in F ig u re 8 a n d is n o w th e s u b je c t o f fu rth e r
in v e s tig a tio n w h ic h w ill b e re p o rte d o n in d u e c o u rs e . It is lik e ly th a t th e re e x is t o th e r
‘o p tim a l’ m u ta tio n ra te s fo r o th e r b lo c k s iz e s in th is p ro b le m , b u t th e s e a re n o t
d is tin g u is h a b le a t th is s c a le o f re s o lu tio n o n th e m u ta tio n ra te a x is a n d th is is a g a in
fu rth e r e x p lo re d in th e d is c u s s io n o f F ig u re 8 . A s m u ta tio n is in c re a s e d b e y o n d a n
u p p e r th re s h o ld , p e rfo rm a n c e is s e e n to d e te rio ra te , a s in th e p re v io u s tw o p ro b le m
c a s e s , in to ra n d o m s e a rc h . T h e m u ta tio n ra te in d u c in g th e firs t a n d s e c o n d p e a k s a re
s e e n to b e .0 0 0 1 a n d .0 1 3 r e s p e c tiv e ly .
F ig u r e 5 s h o w s th e 2 0 ,0 0 0 e v a lu a tio n p e r fo r m a n c e p r o file fo r th e N K 5 0 -8 p r o b le m .
H e re it c a n b e s e e n , a s w a s a ls o v is ib le in th e A D D M P p ro file (F ig u re 2 ), th a t th e
d e te rio ra tio n o f p e rfo rm a n c e fro m th e o p tim u m tro u g h is n o t s m o o th , w ith a ‘rid g e -
M u ltim o d a l P e rfo rm a n c e P ro file s
2 3 1
lik e ’ a n o m a ly a g a in p re s e n t o n th e le ft h a n d e d g e o f th e rig h tm o s t p e a k o v e r a ra n g e o f
lo w e r p o p u la tio n s iz e s .
F ig u re 6 s h o w s a p e rfo rm a n c e p ro file fo r th e s a m e A D D M P in s ta n c e a s F ig u re 2 ,
h o w e v e r th is tim e th e E A is a llo w e d 1 ,0 0 0 ,0 0 0 e v a lu a tio n s o v e r a p o p u la tio n r a n g e o f
1 0 th ro u g h 3 0 0 in s te p s o f 1 0 . In fa c t re s u lts w e re o b ta in e d a t e a c h o f fifty th o u s a n d ,
o n e h u n d re d th o u s a n d , tw o h u n d re d th o u s a n d a n d fiv e h u n d re d th o u s a n d e v a lu a tio n s ,
w ith th e o b s e rv e d a n o m a ly s im p ly b e c o m in g m o re d is tin c t a s e v a lu a tio n lim it
in c re a s e d . In d e e d , th e m u ta tio n ra te in d u c in g th e firs t p e a k , firs t tro u g h a n d a n o m a ly
fe a tu re s w e re s e e n to d e c re a s e w ith in c re a s e d e v a lu a tio n lim it, im p ly in g a re la tio n s h ip
to to ta l n u m b e r o f m u ta tio n s u tilis e d a s w e ll a s a c tu a l m u ta tio n ra te . T h is o b s e rv a tio n
is a ls o c u rre n tly th e s u b je c t o f fu rth e r in v e s tig a tio n a n d w ill b e re p o rte d o n in d u e
c o u rse .
A s c a n c le a rly b e s e e n in F ig u re 6 , th e m u ta tio
d r o p p e d fr o m 4 E -4 a t 2 0 ,0 0 0 e v a lu a tio n s ( F ig u r e 2
T h e s e c o n d a ry a n d te rtia ry p e a k s ‘e m e rg e d ’ fro m
p e a k w ith h ig h e r e v a lu a tio n s lim its u n til h e re a t 1
b e s e e n to b e p e rfo rm a n c e la n d s c a p e fe a tu re s in th
ra n g e o f p o p u la tio n s iz e s . In d e e d , th is p lo t s h o w s
IF F p ro b le m (s e e F ig u re s 4 a n d 8 ).
n ra te in d u c in g th e fir
) to 1 .3 E -5 a t 1 m illio n
th e le ft h a n d e d g e o f th
m illio n e v a lu a tio n s th e y
e ir o w n rig h t, p e rs is te n t
s im ila ritie s w ith re s u lts
st p e a k h a s
e v a lu a tio n s .
e rig h t h a n d
c a n c le a rly
o v e r a w id e
fro m th e H -
F ig u re 7 s h o w s th e 2 d im e n s io n a l p lo t a t a fix e d p o p u la tio n s iz e o f 2 0 o n th e ‘M a x O n e s ’ p r o b le m a t 1 ,0 0 0 ,0 0 0 e v a lu a tio n s . H e r e , th e p o s itio n o f th e fir s t p e a k c a n b e
s e e n to b e a t a c o n s id e ra b ly lo w e r m u ta tio n ra te th a n in F ig u re 3 . T h e tro u g h o f lo w
e v a lu a tio n s is s e e n to b e m u c h w id e r, in d ic a tin g a w id e r ra n g e o f m u ta tio n ra te s
c a p a b le o f fin d in g th e g lo b a l o p tim u m in a lo w n u m b e r o f e v a lu a tio n s . F ig u re 7 a ls o
s h o w s p lo ts o f th e m e a n n u m b e r o f e v a lu a tio n s p lu s a n d m in u s 1 s ta n d a rd d e v ia tio n o f
re s u lts o v e r th e 5 0 ru n s . A s c a n b e s e e n , w h e re m u ta tio n ra te s a re to o lo w , p ro c e s s
v a ria tio n is h ig h , b u t a s m u ta tio n ra te s a p p ro a c h o p tim u m (th e rig h t h a n d e d g e o f th e
le ft h a n d p e a k ), th is v a ria tio n is re d u c e d . O n c e m u ta tio n ra te s b e c o m e c o u n te r
p ro d u c tiv e (a b o v e 2 0 % N R A ), d iv e rs ity is s e e n to in c re a s e a g a in a s th e E A
d e g e n e ra te s in to ra n d o m s e a rc h . A ls o s h o w n is th e m e a n fitn e s s o f th e b e s t s o lu tio n o f
e a c h o f th e 5 0 ru n s a t e a c h m u ta tio n ra te . T h e p la te a u re p re s e n tin g ru n s fin d in g th e
g lo b a l o p tim u m 5 0 tim e s o u t o f 5 0 is s e e n to c o in c id e w ith th e tro u g h in e v a lu a tio n s
in d ic a tin g a n id e a lly tu n e d E A .
F ig u re 8 s h o w s a 2 d im e n s io n a l p e rfo rm a n c e p ro file a t 1 m illio n e v a lu a tio n s o n th e H IF F 6 4 p ro b le m w ith a p o p u la tio n s iz e o f 2 0 . H e re 3 d is tin c t p e a k s a n d tro u g h s c a n b e
s e e n b e fo re d e te rio ra tio n in to ra n d o m s e a rc h , w ith s p e c ific p e a k s a t a m u ta tio n ra te o f
1 E -6 , a n o th e r a t 1 .6 E -3 a n d a th ir d a t 4 .4 E -2 . T h is is a n in c r e a s e in th e n u m b e r o f
fe a tu r e s s e e n a t 2 0 ,0 0 0 e v a lu a tio n s ( F ig u r e 4 ) , a n d a g a in it is n o te d th a t th e fe a tu r e s
o c c u r a t lo w e r m u ta tio n ra te s g iv e n th e h ig h e r e v a lu a tio n lim it. It c a n a ls o b e c le a rly
s e e n o n th e le ft m o s t tw o tro u g h s th a t th e s ta n d a rd d e v ia tio n in th e tro u g h s is a t a
d is tin c t m in im u m o n th e le ft h a n d s id e o f e a c h tro u g h in e v a lu a tio n s . A ls o , th e a v e ra g e
2 3 2
M . O a te s , D . C o rn e , a n d R . L o a d e r
fitn e s s p lo t c le a rly s h o w s m a rk e d in c re a s e s b e tw e e n th e tro u g h s s h o w in g th a t th e E A
is a b le to fin d s te p -lik e im p ro v e m e n t a s m u ta tio n ra te s a re in c re a s e d .
F in a lly , F ig u re 9 s h o w s th e 2 d im e n s io n a l p e rfo rm a n c e p ro file fo r o u r A D D M P
s c e n a rio a t 1 m illio n e v a lu a tio n s w ith a p o p u la tio n s iz e o f 2 0 . W h ils t th e firs t p e a k a n d
tr o u g h a r e c le a r ( a t m u ta tio n r a te s o f 1 .3 E - 5 a n d 8 .2 E - 4 r e s p e c tiv e ly , th e s e c o n d a n d
th ird p e a k a re le s s d is tin c t, a lth o u g h c le a rly e x is t a s fe a tu re s in F ig u re 6 . T h e re a re
c le a r ly r e g io n s o f lo w s ta n d a r d d e v ia tio n c o r r e s p o n d in g to tr o u g h s a t 8 .2 E - 4 , 2 .5 E - 3
a n d 0 .2 , a n d a r e g io n o f h ig h e r m e a n a n d m u c h h ig h e r d e v ia tio n b e tw e e n m u ta tio n
r a te s o f 2 .5 E -3 a n d 0 .1 6 . T h is r e la tiv e d is o r d e r is to b e e x p e c te d a s a n y s tr u c tu r e in
th e s e a rc h s p a c e is lik e ly to b e fa r le s s re g u la r in th e A D D M P s c e n a rio th a n in th e
c a s e o f H -IF F . N o n e th e le s s , F ig u re s 6 a n d 9 c le a rly s h o w th e re to b e s o m e fe a tu re s fo r
w h ic h k e y ra te s o f m u ta tio n a re e ith e r h ig h ly o p tim a l o r h ig h ly s u b -o p tim a l. T h e p lo t
o f m e a n fitn e s s d o e s n o t s h o w s u c h c le a rly d e fin e d ‘s te p -lik e ’ im p ro v e m e n t a s in
F ig u re 8 , h o w e v e r o n c e a g a in it c a n b e s e e n th a t th e re is g e n e ra l im p ro v e m e n t to w a rd s
h ig h e r m u ta tio n ra te s . In te re s tin g ly , a n d in c o n tra s t to o th e r re s u lts s h o w n h e re , fitn e s s
d e te rio ra tio n a t e x c e e d in g ly h ig h ra te s o f m u ta tio n is n o t a s m a rk e d , a g a in s u g g e s tin g
a la c k o f s tru c tu re in th e s e a rc h s p a c e , fa v o u rin g a ‘ra n d o m s e a rc h ’ lik e p ro c e s s . O f
c o u rs e th e m a rk e d in c re a s e in b o th m e a n a n d s ta n d a rd d e v ia tio n o f th e n u m b e r o f
e v a lu a tio n s ta k e n to fin d th e s e g o o d s o lu tio n s s h o w s th e re la tiv e in e ffic ie n c y o f
ra n d o m s e a rc h w ith re s p e c t to a w e ll tu n e d E A o n th is p ro b le m . T h e id e a l m u ta tio n
r a te h e r e b e in g 0 .3 1 ( N R A ) in d u c in g h ig h f itn e s s s o lu tio n s , in a m in im u m o f
e v a lu a tio n s w ith a m in im u m o f p ro c e s s v a ria tio n .
4 C o n c lu s io n s
F o r th e s im p le u
m o d e ra te ra te o f
b o th in te rm s o
e v a lu a tio n s ta k e n
ra n g e o f s u ita b le
n im o d a l s e a rc h s p a c e
m u ta tio n (u s u a lly c lo s
f c o n s is te n t q u a lity o
to fin d th e m . W h e re a
m u ta tio n ra te s e x te n d s
e x p lo re d h
e to 1 / L )
f s o lu tio n
h ig h e r n u m
to w a rd s e v e
e re ,
c a n
fo u n
b e r
r lo w
it
p r
d
o f
e
c a n c le a rly
o d u c e o p tim
a n d m in im
e v a lu a tio n s
r v a lu e s .
F o r th e s tru c tu re d m u lti-m o d a l s e a rc h s p a c e , th e re a p p e a r to b e a
m u ta tio n ra te s , c a p a b le o f e x p lo itin g c e rta in fe a tu re s o f th e s e a rc h
o f H -IF F th is is lik e ly to b e a lig n e d b lo c k s o f c o n s ta n t le n g th e a c h
H a m m in g d is ta n c e . V a lu e s a ro u n d th e s e ‘o p tim a l’ ra te s a re s e e n
o p tim a l in te rm s o f p ro c e s s re p e a ta b ility (in c re a s e d m e a n a n d s ta n
‘g lo b a lly o p tim u m ’ m u ta tio n ra te a g a in s e e m s to e x is t c lo s e to
o p tim u m fitn e s s s o lu tio n s , b u t in th is c a s e , n o t in a n o p tim a lly m
e v a lu a tio n s (h o w e v e r it is p o s s ib le th a t fu rth e r e x p e rim e n ta tio n w
m u ta tio n ra te ra n g e m a y s h o w th is n o t to b e th e c a s e ).
b e se e n
a l p e rfo
u m n u m
is a llo w
th a t a
rm a n c e
b e r o f
e d , th is
s e rie s o f ‘o p tim a l’
s p a c e . In th e c a s e
s e p a ra te d b y e q u a l
to b e c le a rly s u b d a rd d e v ia tio n ). A
1 / L w h ic h g iv e s
in im u m n u m b e r o f
ith a fin e r g ra in e d
T h e p e rfo rm a n c e p ro file o f th e A D D M P h a s b e e n s h o w n to e x h ib it s o m e fe a tu re s in
c o m m o n w ith b o th s tru c tu re d a n d u n s tru c tu re d m u lti-m o d a l s e a rc h s p a c e s . It is c le a rly
m u ltim o d a l o v e r a ra n g e o f p o p u la tio n s iz e s w h e re s u ffic ie n t e v a lu a tio n s a re
M u ltim o d a l P e rfo rm a n c e P ro file s
p e rm itte d . T h e e x is te n c e o f k e
n u m b e r o f e v a lu a tio n s is h ig h ly
th is p ro b le m d o m a in a llo w in g
T h e fa c t th a t o n c e a g a in a m u
s o lu tio n s in a lo w n u m b e r o f e
a llo w e d ra n g e o f e v a lu a tio n s w
re a l tim e a p p lic a tio n o f th e s e te
2 3 3
y m u ta tio n ra te s in d u c in g a lo w s ta n d a rd d e v ia tio n in
s ig n ific a n t w h e n c o n s id e rin g th e a p p lic a tio n o f E A s to
a h ig h d e g re e o f c o n fid e n c e in p ro c e s s re p e a ta b ility .
ta tio n ra te c lo s e to 1 / L is s e e n to g iv e h ig h fitn e s s
v a lu a tio n s w ith lo w s ta n d a rd d e v ia tio n , e v e n w h e n th e
a s c o n s id e ra b ly h ig h e r, is o f c ritic a l im p o rta n c e to th e
c h n iq u e s to a re a l w o rld p ro b le m d o m a in .
5 A c k n o w le d g e m e n ts
T h e a u th o rs a re g ra te fu l to B ritis h T e le c o m m u n ic a tio n s P lc fo r o n g o in g s u p p o rt fo r
th is re s e a rc h .
R e fe r e n c e s
1 . T B ä c k , E v o lu tio n a r y A lg o r ith m s in T h e o r y a n d P r a c tic e , O x fo rd U n iv e rs ity P re s s , 1 9 9 6
2 . K
D e b a n d S A g ra w a l : U n d e r s ta n d in g In te r a c tio n s a m o n g G e n e tic A lg o r ith m
in F o u n d a tio n s o f G e n e tic A lg o rith m s 1 9 9 8 , M o rg a n K a u fm a n n .
P a r a m e te r s .
3 . D
G o ld b e rg (1 9 8 9 ), G e n e tic A lg o r ith m s in S e a r c h O p tim is a tio n a n d M a c h in e L e a r n in g ,
A d d is o n W e s le y .
4 . J H o lla n d , A d a p ta tio n in N a tu r a l a n d A r tific ia l S y s te m s , M IT p re s s , C a m b rid g e , M A , 1 9 9 3
5 . K a u ffm a n , S .A ., T h e O r ig in g s o f O r d e r : S e lf- O r g a n iz a tio n
O x fo rd U n iv e rs ity P re s s , 1 9 9 3
a n d
S e le c tio n
in
E v o lu tio n ,
6 . Z
M ic h a le w ic z , G e n e tic A lg o r ith m s +
1 9 9 6 .
7 . H
M ü h le n b e in a n d D S c h lie rk a m p -V o o s e n (1 9 9 4 ), T h e S c ie n c e o f B r e e d in g a n d its
a p p lic a tio n to th e B r e e d e r G e n e tic A lg o r ith m , E v o lu tio n a ry C o m p u ta tio n 1 , p p . 3 3 5 -3 6 0 .
8 . H
M ü h le n b e in , H o w g e n e tic a lg o r ith m s r e a lly w o r k : I. M u ta tio n a n d h illc lim b in g , in
R . M a n n e r , B . M a n d e r i c k ( e d s ) , P r o c . o f 2 n d I n t ’l C o n f e r e n c e o n P a r a l l e l P r o b l e m S o l v i n g
fro m N a tu re , E ls e v ie r, p p 1 5 -2 5 .
9 . E v a n N im w e g e n a n d J
S iz e In d e p e n d e n t T h e o
C o m p u te r M e th o d s in
a n d G e n e tic A lg o rith m
D e b , e d ito rs , 1 9 9 8 .
D a ta S tr u c tu r e s =
C ru tc h fie ld : O p tm iz in g
r y , S a n ta F e In s titu te W
A p p lie d M e c h a n ic s a n d
s in C o m p u ta tio n a l M e c
E p o c h a l E
o rk in g P a p
E n g in e e rin
h a n ic s a n d
E v o lu tio n P r o g r a m s , S p rin g e r,
v o lu tio n a r y
e r 9 8 -0 6 -0 4
g , s p e c ia l is
E n g in e e rin g
S e a rc h
6 , a ls o
su e o n
, D G o
: P
su
E v
ld b
o p u la tio
b m itte d
o lu tio n a
e rg a n d
n to
ry
K
1 0 . E v a n N im w e g e n a n d J C ru tc h fie ld : O p tm iz in g E p o c h a l E v o lu tio n a r y S e a r c h : P o p u la tio n S iz e D e p e n d e n t T h e o r y , S a n ta F e In s titu te W o rk in g P a p e r 9 8 -1 0 -0 9 0 , a ls o s u b m itte d to
M a c h in e L e a rn in g , 1 9 9 8 .
2 3 4
1 1 . M
M . O a te s , D . C o rn e , a n d R . L o a d e r
O a te s , D C o rn e a n d R L o a d e r, In v e s tig a tin g E v o lu tio n a r y A p p r o a c h e s fo r S e lf-A d a p tio n
in L a r g e D is tr ib u te d D a ta b a s e s , in P ro c e e d in g s o f th e 1 9 9 8 IE E E IC E C , p p . 4 5 2 -4 5 7 .
1 2 . M O a te s a n d D C o rn e , Q o S b a s e d G A P a r a m e te r S e le c tio n fo r A u to n o m o u s ly M a n a g e d
D is tr ib u te d In fo r m a tio n S y s te m s , in P ro c s o f E C A I 9 8 , th e 1 9 9 8 E u ro p e a n C o n fe re n c e o n
A rtific ia l In te llig e n c e , p p . 6 7 0 -6 7 4 .
1 3 . M O a te s a n d D C o rn e , In v e s tig a tin g E v o lu tio n a r y A p p r o a c h e s to A d a p tiv e D a ta b a s e
M a n a g e m e n t a g a i n s t v a r i o u s Q u a l i t y o f S e r v i c e M e t r i c s , L N C S , P r o c s o f 5 th I n t l C o n f o n
P a ra lle l P ro b le m S o lv in g fro m N a tu re , P P S N -V (1 9 9 8 ), p p . 7 7 5 -7 8 4 .
1 4 . M
O a te s , A u to n o m o u s M a n a g e m e n t o f D is tr ib u te d In fo r m a tio n S y s te m s u s in g E v o lu tio n a r y
C o m p u tin g T e c h n iq u e s , C o m p u tin g A n tic ip a to ry S y s te m s , A IP C o n f P ro c s 4 6 5 , 1 9 9 8 , p p .
2 6 9 -2 8 1 .
1 5 . M O a te s , D C o rn e a n d R L o a d e r, S k e w e d C r o s s o v e r a n d th e D y n a m ic D is tr ib u te d
D a ta b a s e P r o b le m , A rtific ia l N e u ra l N e tw o rk s a n d G e n e tic A lg o rith m s 1 9 9 9 , D o b n ik a r e t
a l (e d s ), S p rin g e r p p 2 8 0 -2 8 7 .
1 6 . M
O a te s , D C o rn e a n d R L o a d e r , In v e s tig a tio n o f a C h a r a c te r is tic B im o d a l C o n v e r g e n c e tim e /M u ta tio n -r a te F e a tu r e in E v o lu tio n a r y S e a r c h , in P ro c s o f C o n g re s s o n E v o lu tio n a ry
C o m p u ta tio n 9 9 V o l 3 , IE E E , p p . 2 1 7 5 -2 1 8 2
1 7 . O a te s M , C o rn e D a n d L o a d e r R , V a r ia tio n in E v o lu tio n a r y A lg o r ith m P e r fo r m a n c e
C h a r a c te r is tic s o n th e A d a p tiv e D is tr ib u te d D a ta b a s e M a n a g e m e n t P r o b le m , in P ro c s o f
G e n e tic a n d E v o lu tio n a r y C o m p u ta tio n C o n fe r e n c e 9 9 , M o r g a n K a u fm a n n , p p .4 8 0 - 4 8 7
1 8 . M . O a te s , J . S m e d le y , D . C o rn e , R . L o a d e r, B im o d a l P e r fo r m a n c e P r o file o f E v o lu tio n a r y
S e a r c h a n d th e E ffe c ts o f C r o s s o v e r , in P ro c s o f 1 9 9 9 E v o n e t S u m m e r S c h o o l o n
T h e o re tic a l a s p e c ts o f E v o lu tio n a ry C o m p u ta tio n .
1 9 . G S y s w e rd a (1 9 8 9 ), U n ifo r m C r o s s o v e r in G e n e tic A lg o r ith m s , in S c h a ffe r J . (e d ), P ro c s o f
th e T h ird In t. C o n f. o n G e n e tic A lg o rith m s . M o rg a n K a u fm a n n , p p . 2 – 9
2 0 . W a ts o n R A , H o rn b y G S , a n d P o lla c k J B , M o d e llin g B u ild in g -B lo c k In te r d e p e n d e n c y ,
L N C S , P r o c s o f 5 th I n t l C o n f o n P a r a l l e l P r o b l e m S o l v i n g f r o m N a t u r e , P P S N - V ( 1 9 9 8 ) ,
p p . 9 7 -1 0 6 .
2 1 . W a ts o n R A , P o lla c k J B , H ie r a r c h ic a lly C o n s is te n t T e s t P r o b le m s fo r G e n e tic A lg o r ith m s,
in P ro c s o f C o n g re s s o n E v o lu tio n a ry C o m p u ta tio n 9 9 V o l 2 , IE E E , p p . 1 4 0 6 -1 4 1 3
r
c l
ic
ic o l
as
s r c i Usi
rc
c iq s
a rpls & Ia
ak
a
ni rsit f uss
a l r, right n,
H
nichs@cogs.susx.ac.uk & ianw@cogs.susx.ac.uk
c
h c nstructin ftra nsp rtpr t c l
s hich ff r r l
ia bl
c
unica tin is a c pl
ica t d ta sk. h c
unica ting a g nt ust
quickl a da ptt cha ng s in th n t rk in rd r t a inta in pti a l
p rf r a nc . his a da pti l ntis a n tr
ldifficul
tc p n nt
t c nstructa s th highl d na ic n ir n ntin hich th pr t c l
p ra t s isdifficul
tt pr dict. ur rka tt ptst a ut a t th d sign
pr c ss b c n rtingitfr a d sign pr bl t n f pti isa tin,in
hichg n tic a l
g rith s a r us d t s a rchth spa c fp ssiblpr t c l
d signs in a n a tt ptt find th pti a ls l
utin.
pr s ntr sul
ts
fr
p ri nts in hich
ha
l d al
t rna tingbitpr t c l
s a nd
al
s ind
d fl c ntr l
pr t c l
s, hichha highcha nn l
util
isa tin.
r
ci
is o rk pl
o r s t po ssibil
it o f a ppli g o ltio a r s a rc stra t gis
to t s t sis o f t o rk co
ica tio s pro to co l
s.
pri cipls
a
a o pt
r o rigi a l
lpio
r it
lo f o ltio a r o bo tics[3]a [2]
ic s s g tic a l
go rit s to s a rc t
sig spa c o fro bo tco tro la rc it ct r s fo r s st s ic
ibita sir b a io r.
l
ti a t go a lofo r
r s a rc is to
l
o p a r sa bl t o o l
o g o r co c pt a lfra
o rk fo r t
co str ctio o ft a a pti a l
go rit s a
pro to co l
s
pl
o
b tra spo rt
l
a r pro to co l
s fo r pa ck ts itc
t o rks. p ci ca l
l
a r fo c si go t
l
op
to fco g stio co tro la
a o ia c al
go rit s s c a s t o s
pl
o
b rl
ia blpro to co l
ss c a st
ra spo rt o tro l ro to co l
(
). i c
is so
t si l s
itis cla r t a ta
s al
li pro
ti its a bil
it to til
is r so rc s a
a a ptto c a g il
lgr a tli pro gl
o ba l t o rk
p rfo r a c . H o
r a - sig o ft s co c rr tpro c ss s
ic
st
co -o p ra t o r a
rl
ia bla
co ti a l
lc a gi gco
ica tio
i
is a p rpl i ga
fo r i a blta sk. sig fla s a r iffic l
tto pr ict a
fo r
t
o stpa rt t ct a s pro bl s i t
sig il
lo lb co
a ppa r t
t al
go rit is i pl
t at ac o
o ft
t o rk.
ta p r lso ft a r l lt I t r tisco po s o f a s a l
lit ra cti g
pro c ss s a c o f ic a s a go a lt a tis cla rl
a tt l lo ft a t
pro c ss: t a t o fco
ica ti g a ta to a r c ii g p r pro c ss. t a ig r
l lt co l
lcti it ra ctio o ft s pro c ss s is a ir ctr s l
to ft r ls
S . C a g n o n i e t a l. ( E d s .) : E v o W o r k s h o p s 2 0 0 0 , L N C S 1 8 0 3 , p p . 2 3 5 − 2 4 6 , 2 0 0 0 .
S p rin g e r-V e rla g B e rlin H e id e lb e rg 2 0 0 0
2 3 6
N . S h a rp le s a n d I. W a k e m a n
t a tgo r a si glco
ica tio . H o
r
lss
a r ca r f la tt l
o r
l lt
rg t
a ic a tt a t ig r l l il
lb
pr icta bl la i g
to
t o rk co g stio a
al
l co l
l
a ps . o b ca r f l
l
op
ta tt is
l
o r l l ca co tro lt
ig r-l l rg t
a ic.
l
o pi g a a pti tra spo rt co tro lpro to co l
s fo r co
ica tio s to rks a s l
a rg a
ig l
a ic a s t I t r t a s i r ta
co flicti g
pro bl s.
pri a r f ctio o fr l
ia blco
ica ti g a ta i o l s t co o r i a tio o fr o t pro c ss s. is is a iffic l
tpro bl
ic is co pl
ica t
f rt r
t co
ica tio
i
s
ibits pr icta bll
o ss c a ra ct ristics.
i g to t is iffic l
t co
ica tio
stb p rfo r
i a
ffici t a
r. pro to co l st s t
t o rk r so rc s i a fri l a r it r sp ctto p r pro to co l
s
ic is i co flict it t a i o f ffici c .
It stb a blto
t ctc a g s i t
t o rk
iro
ta
co s
it r o r o r lss o fits r so rc s p
i go t circ sta c s a ro
it.
pro to co l stsa tisf a l
lo ft s
a s it o tgl
o ba li fo r a tio a bo tt
sta t o ft
t o rk a
f rt r stil
l it o tgl
o ba lco tro lo r t pa t a
pa ck t a ta k .
f a
ta lgo a loft is r s a rc is to
l
op a
to ol
og
ic il
l
al
l
o t s t sis o fa a pti pro to co l
sa
pro i i sig ts ito t
sig spa c o fs c a l
go rit s. a t rt a a pro to co l sig r a -cra fti gt
it
sta t a c i a tt
a rto ft pro to co l / s
il
lsp cif t so ltio spa c
a
a t ss f ctio
ic
il
la ss ss so ltio s a ga i st t pro to co l
r q ir
ts.
sig r il
lt
s
tic a rc t c iq s to l
o ca t
so ltio s
ic
t s t
t ss f ctio a
so
t t r q ir
ts.
I
ctio 2
so
o t
a rio s co po
ts o fo r
to ol
o g o rk
to g t r to o l o rki gpro to co l
s.
o ft
o sti po rta tf ctio s o fa co
ica ti gs st is t a bil
it to co
ica t a ta r l
ia bl si g a
rl
ia bl
i . Ifa s st ca
b
l
op i
ic r l
ia blpro to co l
s ca b s t sis t
it il
lgo so
a to pro i gt a t ig r-l lb a io rs a r a l
so a tta i a bl si g t sa
t c iq . I o r rst p ri t scrib i s ctio 3.
s o t a tr l
ia bl
pro to co l
s si g iff r tt c iq s s c a s a ck o l g
ts a
r
a t
tra s issio s a r o l
r iff r tsi l
a tio co itio s.
fa cto r
c a
lca pa cit ito o r pro bl spa c c a
l til
isa tio b co s a i po rta t a s r o ft
t ss o ft so ltio . I t s co
p ri t scrib
i
ctio 3.2
s o t a t i o i g pro to co l
s ar
ol
ic a t
a bil
it to r t c a
la t
% til
isa tio .
co cl
it a isc ssio
a
gi po it rs a s to o pro to co l
s ic a a ptto co g stio ca b
ol .
l
t o rkca b i
a s a irt a l iro
t. pro to co lis a co
ica ti g tit o ft a t iro
t o s a ctio s a r icta t b its
rli g it
sta t a c i . i c i t is r sp cta pro to co lis sta co tro la rc it ct r itis
cla r t a ti o r r to
o l a a pti co
ica tio pro to co l
s
ca
s t
P ro to c o l C o n s tru c tio n U s in g G e n e tic S e a rc h T e c h n iq u e s
sa
o ltio a r t c iq s a s t o s s
a rc it ct r s.
I o r r to
ol
t o rk pro to co l
s
fo l
l
o i gco po
ts:
i t
pro
r q ir t
2 3 7
ctio o fro bo tco tro l
i pl
ta tio o ft
pr s ta tio o fpro to co l
a rc stra t g
co i gsc
p ra to rs to a ip l
a t co i g
t o rksi l
a tio
it ss f ctio
.
pr s t t
fpr t c
r pr s ta tio
a c o s to s is a
t sio o f t
it sta t
ac i (
) ca l
l t co
ica ti g it sta t a c i (
).
a
c o s to s t
a so rr pr s ta tio fo rt o i po rta tr a so s. irstl
pro to co l
s a r a sil r pr s t
si gt
a
t is a s b
a o pt b
t pro to co lr s a rc co
it a s a sta a r r pr s ta tio [6 ] [4]. co
t
as b
s i co p t r sci c a
a rti cia lit l
l
ig c fo r a
a rs a
its c a ra ct ristics a r
l
lo c
t a
rsto o . s a rc rs
a
s t
a s a r pr s ta tio fo r o pti isa tio pro bl s fo r so
ti
si gbo t tra itio a la
o ltio a r s a rc stra t gis [ ].
iff rs fro t
i t a t it ca a ff ct a s l
la s r a ct to
t
iro
ti
ic ito p ra t s.
pa ss s ssa g s o r sig a l
s ia
bo
I
q
s. a c
ac i
as a ip ta
a o tp t q
a
c t s a t o -st p a l
go rit . ri gt
rstst p i p tsig a l
s a r a la t
a
a tra sitio r ls lct . I t s co
st p t o tp tsig a lis p a t
a
t
a c i c a g s sta t . If o
o ft r ls fo r a sta t ca b
c t
t
a c i r a i s i its c rr tsta t .
.
rc
tr t g
a
ta lto t is o rk is t a bil
it to i t pro to co la s a co tro la rc it ct r
ic ca b r pr s t i so
s bo l
ic fo r . is r pr s ta tio
stb o
t a tca b
ta t o r s cc ssi it ra tio s t ro g t s a rc
spa c o fpo ssiblso ltio s. is tra sfo r s t pro bl fro o
o fco tro l
lr
l
op
t ito o
o fco tro l
lr t s t .
ta sk o fg ra ti g co pl co tro la rc it ct r s si g s a rc a s b
a cco pl
is
i o t r ls
si g
tic a l
go rit s( s).
il
la tt pt to co str ctco
ica tio
pro to co l
s si gsi il
a r pri cipls.
.
c
g c
co i g sc
s to r pr s t t
cl
o s lti to t s a rc stra t g
pl
o
.
so ltio i t pro bl spa c is
o b io sc o ic o fr pr s ta tio
2 3 8
N . S h a rp le s a n d I. W a k e m a n
si g g tic
r pr s t b a s q
b t sta a r g
to a ria bll gt
g o
a
r pr s
g o
is tra sfo r
sig spa c .
.4
al
go rit s is a bi a r stri g i
ic a c tra sitio is
c o fbit a l s.I a itio to a l
l
o i g a s a ip l
a tio
tic o p ra to rs t is r pr s ta tio
il
la l
so l its l
f a sil
co i g.
bi a r
co i g o ft
is k o a s t
ts a so ltio i t s a rc spa c . rio r to
a la tio t
ito t p
o
a i sta tia tio o f t so ltio i
p r t rs t
pu t
c
g
g tic a l
go rit
pl
o
ri gt is r s a rc s s t sta a r g tic
o p ra to rs i. .
ta tio
cro sso r a itio a
s btra ctio . o r t sa k o f
cl
a rit t fo l
l
o i gs ctio s ta ilt a ct a li pl
ta tio
s .
ut t .
g r issp ci fo r a c r r l
a ti gto t
b ro f ta tio s
p r g o . ri g a ti g a c biti t stri gis gi a i ii a lc a c
o f ta tio
a c c a c b i ga q a lpro po rtio o ft
ta tio ra t fo r t is
g o . a c biti t stri g is t
c ck a ga i stt r ca l
c l
at
ta tio
ra t a
a
ta tio flips t bit a l .
r ss r. a fo r
bitba sis. o r a l
lt
cro sso r a s pro
a
a p rg
ba sis
r
s o fcro sso r ista
ca o p ra t o a p r g o r p r
p ri ts ca rri o t ri g t is r s a rc si glpo it
q at. o rt
o stpa rtt is a s b
i pl
t o
t cro sso r po itis r strict to g
bo
a ris.
t
u tr ct .
a itio a
s btra ctio o p ra to rs a l
l
o
fo r a t o ro g
pl
o ra tio o ft s a rc spa c a i g o r r o i gg s fro
t g o
ra o l.
itio a s a
ra o l co str ct g
to t
g o . l
t o g o s lctio pr ss r fo r si is itro c b t
t ss
f ctio t r
a c il
o g r g o s cr a t s ro b stso ltio s it r ga r
ta tio .
is is b ca s t
o rki g g s o ft so ltio (t pa rts o ft
it sta t a c i t a tco trib t to t ss) ca
i t
s l s a o gstt
r
a tg s o ft g o .
.
t r
u t
orl
i a r o pti isa tio pro bl s a c ca i a t so ltio is a la t a ga i st
a sta tic t ss f ctio
ic pro i s a
a s r o f its a bil
it to so l t
pro bl . o r o -l
i a r pro bl s s c a s t o s fa c
r
st a la t
a pro to co l
’
s f ctio a l
it i its o rki g
iro
t. o r a
a ppl
ica tio s
t is is i pra ctica l
to t a o to fti r q ir fo r a
a la tio a
a
co p t r o l stb s to si l
a t t a g t’
s
iro
ta
its it ra ctio
it it. is is o tt ca s fo r t a ppl
ica tio
pro po s ; a la tio o f
a pro to co l
sp rfo r a c co lta k pl
a c i a r a l t o rk it o t r a so a bl
P ro to c o l C o n s tru c tio n U s in g G e n e tic S e a rc h T e c h n iq u e s
2 3 9
ti o r a a l
t o g a tt is sta g r a l o rl a la tio
o la
cssa r co pl it.
a tis r q ir t
is a t stb i
ic to a la t t
p rfo r a c o fa pro to co l it i its o rki g
iro
t o r a s cl
o s to t a t
iro
ta s is po ssibl.
fo l
l
o i g p ri ts s a si pls
r-r c i r t o rk si l
a tio
a
a plo f ic ca b s
i g r . o i itia t a si l
a tio t o p o s a r co str ct : a s
ra
a r c i r.
co po
to f a c
p
o
isg ra t fro t g o
to b
a la t . o t s
ra r c i r
a a a sso cia t co
ica tio c a
lt c a ra ct ristics o f ic a r
p
to t
p ri tb i g p rfo r
. pica l
l ac p
o
al
so a s
a tra s issio b ff r a
a r c i b ff r.
t s i g a c i t tra sissio b ff r co ta i s t
ssa g to tra s it. I a itio
ac
ac i al
so
a s so
fo r o fc rr t
or i
ic a ta it s ca b
l t po ra ril
b fo r b i gpl
a c ito b ff rs o r ssa g s.
si l
a tio
c t s fo r a sp ci c
b r o f ti st ps.
ri g a c
st p t s
ra
t
t r c i rp
o s g ta c a c to
c t . ri g
c tio
ac
a c i it ra t st ro g t tra sitio so fitsc rr tsta t fro
rstto l
a st.
a ctio a sso cia t
it t tra sitio is a la t ift a ctio
ca b p rfo r
t tra sitio
rs. I ff ct a c tra sitio fo r s a bo o la
pr ssio . Ift
pr ssio
o ls t
t
ac i
a k s t tra sitio to t
tsta t . t a c ti st p i t si l
a tio
r po ssibltra sitio o ft
c rr tsta t is a tt pt
tila
c tio is ca pa blo f ri g. tt is po it
t
a c i c a g s sta t to t sta t i ica t b t
t-sta t l o ft
tra sitio .
t r o rk il
li o l l
a rg rsi l
a tio s i
ic
a si l
at
t o rk
o s il
lco
ica t a cro ss a si gl s a r c a
l it a c o
pl
o i gt sa
pro to co la s i t r a l o rl.
a la tio
il
lb a
as r o f
t gl
o ba ltra ffic t ro g -p ta cro ss t c a
l
. o ti
o rk a
s t
c rr tr a l
- o rl I t r ti pl
a c o ft si l
a tio .
.6
t ss u ct
p rpo s o ft t ssf ctio isto a la t t p rfo r a c o ft pro to co l
ri g its
c tio i t
t o rk si l
a tio a
pro i so
q a ti a bl
a s r o fitsf ctio a l
it.
a s r its l
fs o li cl a sf co po
ts
a s po ssibl si c t gr a t r t
b r o fco po
ts t lss pr icta blt
po p l
a tio ’
s co
rg c o t
sir po iti so ltio spa c . is
as r
ca b ca l
c l
a t i t o a s:t
o st iffic l
tis to ga t r i fo r a tio
ri g
t si l
a tio a
ca l
c l
at t
t ss fro t
a ta co l
lct ;a a sir o ptio
is to
as r t
t ss o ft sta t o ft
t o rk o s a ft r t si l
a tio
a s is .
st b a r i i t a t a ss ssi g t
a l o fa so ltio
si g o l a
si gl si l
a tio
o s o tpro c a a cc ra t
a s r o fits a bil
it si c it
is po ssiblt a tt pro to co lis pl
o iti gso
a sp cto ft t st a ta s . I
o r r to ga i a fa ir a s r o ft
t ss o fa so ltio
ac
stb t st a
b ro fti s si g iff r tt st a ta .
stb ca r f l
s lcti gt
2 4 0
N . S h a rp le s a n d I. W a k e m a n
b r o fr -t sts to p rfo r
gra a
if
s to o f
la s s so
c o ic a s to
so
fo r o f a .
ro g
a l g ra t
al
i ac i
t is is t a l
a
s
3
p ri
if
s to o a r -t sts t p rfo r a c il
l
t so ltio
il
l o tb a q a t l t st . is
ic a l s o l b c o s :t
ig st l
o sto r
p ri ta tio
fo
t a t si g t
a
s i t la st
b r o fg ra tio s.
r fo r
i t fo l
l
o i g p ri ts.
s
fo l
l
o i g s ctio s ta ilt o o f t
p ri ts
rta k to a t . I
a c ca s t
p ri t a s p rfo r
si ga istrib t po p l
a tio o ffo r
r i ii a l
s o l fo r o t o sa
g ra tio s.
ta sk a s to co ica t t
a ta it s fro t s
r to t r c i r.
.
t
I o r i itia l p ri t
s to tto s t sis a f l
lr l
ia bl co
ica tio pro to co la
i o i g so pro t a tt co tro la rc it ct r
rli g a
co
ica tio pro to co lca b g ra t
si g o ltio a r s a rc . o pro i
a
a s r o ft q a l
it o ft s t sis pro to co l a c ca i a t so ltio
a s a la t
si g a t o rk si l
a tio
ic
l
ats t
iro
tfo r
ic t a l
t r a ti gbitpro to co lis a o pti a lso ltio . It a s o p t a tt
s st
o l pro c a a c i
it si il
a r sta t tra sitio s a
t r a lb a io r.
al
t r a ti gbitpro to co lis a si plpro to co lt a tpro i s r l
ia bl
co
ica tio
si g a a l
t r a ti g a l
i a tio bit iti ica t s
t ror ot
a
ssa g a s l
o st ri g tra s issio . Itis l
lo c
t i t l
it ra t r
a
a s rstpro po s i [ ].
t r
u t . o g ra t a s
r si il
a r to t a to fo r b c a rk
a c i t si l
a tio
stcl
o s l o lt o p ra ti g
iro
to ft
o rigi a lpro to co l
. o r t is t c a
l st a t ca pa cit to o l o l a
si gl ssa g . I a cco r a c it t r fr c pro to co l
;o l ssa g s ca b
l
o sta
ifa
ssa g is l
o st t c a
l il
lflip t biti t
ssa g
a r.
o r a sc
a tic o r i o ft si l
a tio s
gr .
i
ct s.
fo l
l
o i ga ctio s a b
tra po l
a t fro
i t o rigi a la l
t r a ti gbitpro to co l
.
t
a c i sp c-
:
a
ssa g o t co
ica tio s c a
lta ki g a ta fro
o r ifr q ir .
ci : ci a
ssa g fro t co
ica tio s c a
la
pl
ac t
a ta ito
or.
q
: l
a c t it i
o r ito t r c i b ff r.
q
: l
ac t
rst l
to ft tra s itb ff r ito
or.
P ro to c o l C o n s tru c tio n U s in g G e n e tic S e a rc h T e c h n iq u e s
Transmition Buffer
Transmition Buffer
0:4
0:3
0:2
0:1
2 4 1
Channel
0:0
Data
FSM
FSM
Channel
Receive Buffer
Receive Buffer
Data
bit : seq
Memory
t rksi ul
a tin f r
l
l:
Memory
l
utin fr l
ia blc
unica tingfinit sta t
a chin .
l
la ctio a la s o rks
a c o ft s a ctio s r t r s a tr o r fa l
s bo o la a l
p
i go its
a bil
it to co plt t sp ci
ta sk.
l fo r t s
a ctio is t is o tt
ca s it il
lr t r tr
r al
lco itio s to pr
t it tio a lf ba ck
fro t r c ii g a c i . I t ca s
r t c a
la l
ra
co ta i s a
ssa g t
ssa g is isca r .
t ss u ct .
t ss fo r a si glpro to co lis ba s o t pro po rtio
o fa
ssa g r c i a ft r a
b r o fti st ps. o
s r a rl
ia bl
pro to co lis pro c o lt i -o r r po rtio o ft r c i
ssa g is co si r . g o
il
lsco r ro
ifa l
lb tt
rst ssa g a s s cc ssf l
l
r c i . Ift is a s o tt ca s a
t
tir pro po rtio o ft
ssa g a s
co si r
t
t po p l
a tio
o l pr a t r l co
rg o a
rl
ia bl
so ltio a
b co
i ca pa bl o f ba ck-tra cki g t ro g t s a rc spa c to
rl
ia blso ltio s.
su ts.
so
i g r 3 asb
g ra t
ri ga t pica lr
o ft s st . s it a t pica lsta t tra sitio
ia gra
ac o
r pr s ts
a sta t . a c tra sitio is r pr s t a s a a rc fro t sta t i
ic itis
c t to t t
tsta t . tra sitio is a rk
it t a ctio to b ta k
a t pa ra t rsfo rt a ta ctio . o
a ctio sr q ir o pa ra t rs i
ic
ca s t is r
a t co i g i t tra sitio a s
tra l
it to t g o :
i cr a si g its ro b st ss it r ga r to
ta tio .
tra sitio s a r a rk
it t o r r i
ic t
a r to b a tt pt s o
i sq a r bra ck ts.
s ca b s
i t
ia gra
t s
r co ti a l
ls s a
ssa g it
i ti r “ ”. la ft r r c ii ga a ck o l g
tfro t r c i r o s it
q
t
t a ta it . is is a si plsto p-go stra t g a
is t cl
o s st
a co
to g ra ti g t a l
t r a ti g bitpro to co l
. o t : a c tra sitio
a s t sa
b r o fpa ra t rs a l
t o g fo r q
q
a
l
l
t s a r s p rfl o s a
isca r
ri g
c tio .
2 4 2
N . S h a rp le s a n d I. W a k e m a n
1
Fitness Value
0.8
0.6
0.4
0.2
0
0
100
200
300
400
500
600
700
800
900
1000
Generations
itn ss a l
u
bta in d b b stindiidua l
during
0
N u ll a c k 0 [2 ]
D e q u e u e a c k 1 [0 ]
1
R e c e iv e m e s g 1 [1 ]
R e c e iv e a c k 0 [0 ]
S e n d m e sg 1 [0 ]
6
pt
S e n d m e sg 1 [1 ]
u
7
E n q u e u e m e sg 0 [0 ]
S e n d a c k 0 [0 ]
h t pica lr sul
t fth r l
ia blc
.
l
utin fa r l
ia blpr t c l
.
2
unica tin
p ri nt,a st p g stra t g .
c t
I t s co
s to f p ri ts
s to tto
l
o p pro to co l
s ic a tt pt
to
a k o pti a l s o f t ba
i t a a il
a bl.
r fo r a go o so ltio
o lb o
ic tra s itt a
ssa g
ri g a c ti st p. I a l
l
o ll
ik to g ra t a a c i
ic
ibits a i o i gb a io r.
t r
u t . fcritica li po rta c fo rt is p ri tist l
at c
co t r
ri gco
ica tio . o si l
a t t is t co
ica tio c a l
s o ft p
o
a b
co
rt fro t sta a r q
str ct r
s i t sta a r
ito a rr o fsi q a lto t l
a t c r q ir .
ri g a c si l
a tio ti st p
ssa g s o t c a
la r o
o
pl
ac
al
o gt a rra . s it t r l
ia bil
it p ri ts l
o ss is ca l
c l
at
ri gt
r c i p a s o fco
ica tio . Ifa
ssa g is l
o st ri gco
ica tio itis
o tpa ss to t r c ii gp
o . I t pr io s p ri t a c
ac i
a s p r itt o
a ctio p r ti st p. H o
r i itia l p ri ts r r a l t a tso ltio s g ra t
si gt is ti i gsc
o rco
rl
ia bil
it b
P ro to c o l C o n s tru c tio n U s in g G e n e tic S e a rc h T e c h n iq u e s
2 4 3
s ig l
tiplpa ck ts.
ti sp tsto ri ga pa ck ti t r tra s issio
b ff r a s s to s
a s co
pa ck t s cc ssi pa ck ts a i gfa rlssc a c
o fl
o ss t a t pr io s. o fo rc t
l
op
to fa r tra s issio sc
a c tra sitio
asb
a ssig
a pa ra t r ic sp ci s t l gt o fti
it a ta k to
c t t a ctio .
Transmission
Buffer
0:4
0:3
0:2
0:1
Transmission
Buffer
0:0
Receive Buffer
FSM
Re-transmission
Buffer
Memory
FSM
Latency
Variable B
Variable B
t rksi ul
a tin f r
l
utin fr l
ia blc
hich a l
s c
unica t s da ta in pti a lti .
ct
Re-transmission
Buffer
Latency
Variable A
Receive Buffer
Variable A
Memory
unica tingfinit sta t
a chin ,
s.
:
a
ssa g o t co
ica tio s c a
lta ki g a ta fro
o r ifr q ir .
ci : ci a
ssa g fro t co
ica tio s c a
la
pl
ac t
a ta ito
or.
q
: l
a c t it i
o r ito t r c i b ff r.
q
: l
ac t
rst l
to ft tra s itb ff r ito
or.
q
tra s it: l
a c t it i
o r ito t r tra s itb ff r.
q
tra s it: l
ac t
rst l
t o ft r tra s itb ff r ito
or.
o
ro
tra s it:
o t c rr tit o f a ta fro t r tra sissio b ff r.
t a ria bl : t a ria bl
it t c rr tit i
or.
t a ria bl : tt
o r to t a l o f a ria bl .
t a ria bl : t a ria bl
it t c rr tit i
or.
t a ria bl : tt
o r to t a l o f a ria bl .
l
l: l
la ctio a la s o rks
.I
t p
o s
al
so q ipp
to a ip l
at
so ltio s to o p
a bls a blt
fl
l
.
a itio
s pr
it a
a ta .
ra t o
po ra r
to t sta a r tra s it r c i a
o r b ff rs o f
io sl t p
o
s fo r t is s to f p ri ts is
it r -tra s issio b ff r a
t o a ria bls it
ic
r -tra s issio b ff r a s b
a
to
co ra g
s
ts o ft o ra l
l ssa g .
tra
o r a risto ra g o f a ta
t r -tra s issio b ff r b co s
2 4 4
N . S h a rp le s a n d I. W a k e m a n
t ss u ct
o r t is p ri t t
t ss f ctio
st a ss ss t
so ltio ’
s a bil
it to r l
ia bl co
ica t t
a ta . I a itio
it st a l
so
pro i so
a s r o ft ti ta k fo r t co
ica tio to ta k pl
ac ;
so ltio s ic co
ica t o stq ickl sco r a ig r t ss a l . o
t r i t is a l
a co bi t o rigi a lf ctio
s to
l
op t
rl
ia blpro to co l it a a itio a la l
ic i ica t s t pro po rtio o ft
si l
a tio -ti ta k to co plt t
a ta c a g .
m
t
2
− m
m t
+
tm
omm
t
t
h
su ts.
gra p s pr s t i g r
s o t a ra g a o to fsi l
a tio ti ta k to s
t co plt
ssa g . Ift
tir
ssa g a s o t
r c i b t irt ti st ps t si l
a tio
as tr iat . l
so s o
is t
a ra g
b r o f a ta it s s cc ssf l
lr c i
ri g a
a la tio .
gra p i g r 7 s o s t co rr spo i g t ss a l o ft
ig stsco ri g
i ii a la t a c g ra tio .
expB
100
80
Time
60
40
20
0
0
100
200
300
400
500
600
700
800
900
1000
Generations
i
4
isc ssi
ak n t
r
unica t
plt
ssa g
r
is pa p r a s s o
t a tt s a rc spa c o f it sta t a c i s fo r co
ica tio pro to co l
s ca b
pl
or
si gg tic s a rc t c iq s.
I itia l
l
a tt pt to co str ct l
lk o
pro to co l
s s c as t al
t ra ti gbitpro to co l
. tt pti gto fo rc t s a rc pro c ss to g ra t sp ci c
pro to co l
s pro
iffic l
t it t
sir pro to co l
r b ig o l . H o r al
i pro to co l
s ic p rfo r
t q ia l tta sk r pro c
t s
P ro to c o l C o n s tru c tio n U s in g G e n e tic S e a rc h T e c h n iq u e s
2 4 5
expB
10
Data Items
8
6
4
2
0
0
100
200
300
400
500
600
700
800
900
1000
Generations
u b r f a ta It s
6
unica t d
1
Fitness Value
0.8
0.6
0.4
0.2
0
0
100
200
300
400
500
600
700
800
900
1000
Generations
7
itn ss a l
u
bta in d b b stindiidua l
during
l
utin fa pti a l
pr t c l
.
0
[2 ] D e q u e u e R e tra n s m it
6
[3 ] R e c e iv e
1 4
[1 ] D e q u e u e
[1 ] G e t V a ria b le B
[0 ] E n q u e u e R e tra n s m it
3
[0 ] D e q u e u e R e tra n s m it
[0 ] D e q u e u e
[1 ] E n q u e u e R e tra n s m it
[0 ] S e n d
[0 ] E n q u e u e
[2 ] N u ll
[2 ] E n q u e u e
[1 ] G e t V a ria b le B
[2 ] E n q u e u e
[3 ] R e c e iv e
8
1 0
[2 ] R e c e iv e
[1 ] R e m o v e V a ria b le fro m
1 1
R e tra n s m it b u ffe r
1 5
h t pica lr sul
t fth
pro i g t a t pro to co l
s
t is t c iq .
ibiti g a
pti a lc
sir
b
unica tin
a io r ca
p ri nt.
b g
ra t
si g
2 4 6
N . S h a rp le s a n d I. W a k e m a n
o rk
rta k
r as s o
t a tt t pro c ss o fco str cti g
al
i t o rksi l
a tio s is a iffic l
tpro bl . H o
r fo r co pl pro to co l
s
ic o p ra t i ig l
a ic
iro
ts t iffic l
t i co str cti gt
si l
a tio
il
lb
ga t b t
iffic l
tis o ftr i g to pr ictt
rg t
a ic.
a a ta g o f si g t is a ppro a c is its a bil
it to gi s ir ct
f ba cko t pro to co l
’
sp rfo r a c i t si l
at
iro
ti o r rto
ri t
sig pro c ss si c t
t ss f ctio
s s a si l
a tio o ft
to rk
iro
tto
t r i t p rfo r a c o ft so ltio s.
t ig
t si l
a tio spa c fro a si glpa ir o fpro to co l titis to
l
tiplco
ica ti g titis
ca a ss ss t i pa cto fa si glco
ica tio stra t g o
t gl
o ba l t o rk
a ic.
f i g-ba ck si l
a tio p rfo r a c ito t
sig pro c ss
ca
s r t a t t pro to co lpro c is bo t ro b sta
o pti a l it r sp ctto t gl
o ba l t o rk
a ic.
r
t p ri ts il
la
l
tipl titis co
ica ti go ra si gl
c a
l si g t sa
pro to co l
.
t ss f ctio
il
l s so
o tio o f
fa ir ss a s l
la s r l
ia bil
it a
til
isa tio to
s r t a tt
o l pro to co l
r a cts to co g stio .
fr c s
. ca ntlbur . . a rtlttK. . a nd . . il
kins n. n t n r l
ia blful
l
-dupl
tra ns issin
r ha l
f-dupl l
in s.
t
t
, (5): 6 – 6 5,
6 .
. . l
iff, I. H a r , a nd .H usba nds. Incr
nta l l
utin fn ura ln t rk a rchit ctur s f r a da pti b ha iur.
chnica l p rt gniti
cinc
s a rch
ap r
56 , ch l f gniti a nd
puting cinc s, ni rsit f uss , right n
H , ngl
a nd, K,
.
3. . l
iff, . H usba nds, a nd I. H a r .
ling isua l
l guid d r b ts.
chnica l p rt gniti
cinc
s a rch a p r
, ch l f gniti a nd
puting cinc s, ni rsit f uss , right n
H , ngl
a nd, K,
.
4. ra nd . a nd Z a fir pul . n c
unica ting finit sta t a chin s.
r
t
, 3 ( ):3 3–34,
3.
5. a r nc . g l
. rt
t g
tr g s
t
t . hn il
a nd ns Inc, 6 6 .
6 . . . ch a nn. init sta t d scriptins fc
unica tin pr t c l
s.
tr
t rs, (4/ 5):36 –37 , 7 .
P r e d ic tio n o f P o w e r R e q u ir e m e n ts
fo r H ig h -S p e e d C ir c u its
F . C O R N O , M . R E B A U D E N G O , M . S O N Z A R E O R D A , M . V IO L A N T E
D ip . A u to m a tic a e In fo rm a tic a
P o lite c n ic o d i T o rin o
h t t p : / / w w w . c a d . p o l i t o . i t
A b s tr a c t. M o d e rn V L S I d e s ig n m e th o d o lo g ie s a n d m a n u fa c tu rin g te c h n o lo g ie s
a re m a k in g c irc u its in c re a s in g ly fa s t. T h e q u e s t fo r h ig h e r c irc u it p e rfo rm a n c e
a n d in te g ra tio n d e n s ity s te m s fro m fie ld s s u c h a s th e te le c o m m u n ic a tio n o n e
w h e re h ig h s p e e d a n d c a p a b ility o f d e a lin g w ith la rg e d a ta s e ts is m a n d a to ry .
T h e d e s ig n o f h ig h -s p e e d c irc u its is a c h a lle n g in g ta s k , a n d c a n b e c a rrie d o u t
o n ly if d e s ig n e rs c a n e x p lo it s u ita b le C A D to o ls . A m o n g th e s e v e ra l a s p e c ts o f
h ig h -s p e e d c irc u it d e s ig n , c o n tro llin g p o w e r c o n s u m p tio n is to d a y a m a jo r is s u e
fo r e n s u rin g th a t c irc u its c a n o p e ra te a t fu ll s p e e d w ith o u t d a m a g e s . In
p a rtic u la r, to o ls fo r fa s t a n d a c c u ra te e s tim a tio n o f p o w e r c o n s u m p tio n o f h ig h s p e e d c irc u its a re re q u ire d . In th is p a p e r w e fo c u s o n th e p ro b le m o f p re d ic tin g
th e m a x im u m p o w e r c o n s u m p tio n o f s e q u e n tia l c irc u its . W e fo rm u la te th e
p ro b le m a s a c o n s tra in e d o p tim iz a tio n p ro b le m , a n d s o lv e it re s o rtin g to a n
e v o lu tio n a ry a lg o rith m . M o re o v e r, w e e m p iric a lly a s s e s s th e e ffe c tiv e n e s s o f
o u r p ro b le m
fo rm u la tio n w ith re s p e c t to th e c la s s ic a l u n c o n s tra in e d
fo rm u la tio n . F in a lly , w e re p o rt e x p e rim e n ta l re s u lts a s s e s s in g th e e ffe c tiv e n e s s
o f th e p ro to ty p ic a l to o l w e im p le m e n te d .
1 . I n t r o d u c t i o n
M o d e rn te le c o m m u n ic a tio n s y s te m s m u s t e n s u re h ig h b a n d w id th a n d h ig h
p e rfo rm a n c e to c o p e w ith th e in c re a s in g a m o u n t o f d e liv e re d d a ta . D iffe re n t s o lu tio n s
a t d iffe re n t le v e ls o f a b s tra c tio n c a n b e e x p lo ite d : im p ro v e d c o m m u n ic a tio n c h a n n e ls ,
b e tte r c o m m u n ic a tio n p ro to c o l a n d h ig h -s p e e d c o m m u n ic a tio n e q u ip m e n t. M o d e rn
V L S I d e s ig n m e th o d o lo g ie s a n d m a n u fa c tu rin g te c h n o lo g ie s a re m a k in g c irc u its
e v e ry d a y fa s te r. T o c o p e w ith th e c o m p le x ity o f d e s ig n in g h ig h -s p e e d c irc u its
d e s ig n e rs m u s t e x p lo it s u ita b le C A D to o ls . A m o n g th e s e v e ra l a s p e c ts o f h ig h -s p e e d
c irc u it d e s ig n , p o w e r c o n s u m p tio n is to d a y a m a jo r is s u e . In th e la s t y e a rs , d e s ig n fo r
lo w -p o w e r c o n s u m p tio n h a s b e c a m e a w id e s p re a d d e s ig n p a ra d ig m . D e s ig n
te c h n iq u e s to c o n tro l p o w e r c o n s u m p tio n a re m a n d a to ry b e c a u s e e x c e s s iv e p o w e r
d is s ip a tio n c a n c a u s e p e rfo rm a n c e d e g ra d a tio n , ru n -tim e e rro rs , o r d e v ic e d e s tru c tio n
d u e to o v e rh e a tin g . L a rg e in s ta n ta n e o u s p o w e r d is s ip a tio n m a y a ls o c a u s e lo c a l h o t
s p o ts th a t h a v e n e g a tiv e im p a c t o n c irc u it re lia b ility . W ith in c re a s in g d e m a n d s fo r
re lia b le h ig h -s p e e d c irc u its a c c u ra te e s tim a tio n o f p e a k -p o w e r d is s ip a tio n d u rin g
d e s ig n p ro c e s s is b e c o m in g e s s e n tia l.
S . C a g n o n i e t a l. ( E d s .) : E v o W o r k s h o p s 2 0 0 0 , L N C S 1 8 0 3 , p p . 2 4 7 − 2 5 4 , 2 0 0 0 .
S p rin g e r-V e rla g B e rlin H e id e lb e rg 2 0 0 0
2 4 8
F . C o rn o e t a l.
In th is p a p e r w e a d d re s s th e p e a k -p o w e r e s tim a tio n p ro b le m fo r h ig h -s p e e d
s e q u e n tia l c irc u its . P o w e r c o n s u m p tio n in s e q u e n tia l c irc u its is a n o n -lin e a r fu n c tio n ,
3 , o f th r e e v a r ia b le s : th e c ir c u it in itia l s ta te , S , a n d tw o v e c to r s ( 9− 1 , 9 ) th a t a r e
I
W
a p p lie d to th e c irc u it in p u ts a t tim e t-1 a n d t. W e re p re s e n t th e m
W
a s a 3 - tu p le ( S , 9− 1 ,
W
9 ). T h e p ro b le m o f c o m p u tin g th e p e a k -p o w e r c o n s u m p tio n o f s e q u e n tia l c irc u its
c a n b e f o r m u la te d a s th e p r o b le m o f f in d in g th e 3 - tu p le ( S , 9− 1 , 9 ) th a t m a x im iz e s
W
W
W
th e fu n c tio n 3 .
I
S e v e ra l te c h n iq u e s h a v e b e e n p ro p o s e d to s o lv e th is p ro b le m re s o rtin g to e ith e r
e x a c t a p p ro a c h e s [1 ] o r te c h n iq u e s b a s e d o n A u to m a tic T e s t P a tte rn G e n e ra to rs [2 ].
B u t o n ly in [3 ] a n d [4 ] a n a lg o rith m is p ro p o s e d a b le to c o p e w ith la rg e s e q u e n tia l
c irc u its . In p a rtic u la r th e p e a k -p o w e r p ro b le m is fo rm u la te d a s a n u n c o n s tra in e d
o p tim iz a tio n p ro b le m a n d a te c h n iq u e b a s e d o n G e n e tic A lg o rith m s is p ro p o s e d . T h is
te c h n iq u e is s tro n g ly a p p ro x im a te d s in c e th e re a c h a b ility p ro p e rty o f th e s ta te S is
n e g le c te d d u rin g th e o p tim iz a tio n p ro c e s s . A s a c o n s e q u e n c e , th is a p p ro a c h c o u ld
le a d to s o lu tio n s w h e r e th e 3 - tu p le ( S , 9− 1 , 9 ) e m b e d s a s ta te S th a t th e s e q u e n tia l
c irc u it n e v e r re a c h e s . T o m a k e th e 3 -tu p le a fe a s ib le s o lu tio n , th e a u th o rs o f [3 ] a n d
[4 ] p ro p o s e to re p la c e a t th e e n d o f th e o p tim iz a tio n a lg o rith m th e s ta te S w ith a
re a c h a b le o n e a s c lo s e a s p o s s ib le to S .
In th e fo llo w in g , w e w ill a n a ly z e th e b e h a v io r o f th is a lg o rith m , a n d
e x p e rim e n ta lly s h o w th a t it c a n u n d e re s tim a te p e a k -p o w e r w h e n th e n u m b e r o f
re a c h a b le s ta te s is a s m a ll fra c tio n o f a ll th e p o s s ib le s ta te s . W e w ill th e n p ro p o s e a n
a p p ro a c h th a t o v e rc o m e s th is d ra w b a c k b y e x p lo itin g th e k n o w le d g e a b o u t th e
re a c h a b le s ta te s e t d u rin g th e o p tim iz a tio n p ro c e s s . In th is p a p e r w e p ro p o s e a n d
e v a lu a te a n im p le m e n ta tio n o f o u r p e a k -p o w e r e s tim a tio n a lg o rith m th a t is b a s e d o n
th e S e lfis h G e n e a lg o rith m [5 ], a n e v o lu tio n a ry o p tim iz a tio n a lg o rith m w e d e v e lo p e d .
W e s u p p o rt o u r a p p ro a c h b y re p o rtin g e x p e rim e n ta l re s u lts a s s e s s in g th e e ffe c tiv e n e s s
o f o u r a p p ro a c h .
T h e a lg o rith m p re s e n te d h e re is b a s e d o n a p re v io u s w o rk [8 ] w h e re G e n e tic
A lg o rith m s w e re u s e d to s o lv e th e p e a k -p o w e r p ro b le m . In [8 ] a G e n e tic A lg o rith m
e v o lv e s a p o p u la tio n w h e re e a c h in d iv id u a l is a 3 -tu p le c o m p o s e d o f a re a c h a b le s ta te
a n d tw o v e c to rs . T o g u a ra n te e th a t th e p o p u la tio n c o n ta in s o n ly v a lid in d iv id u a ls ,
e a c h tim e c ro s s o v e r a n d m u ta tio n o p e ra to rs a re a p p lie d th e s ta te in th e n e w in d iv id u a l
is a n a ly z e d , a n d re p la c e d w ith a re a c h a b le o n e w h e n re q u ire d . T h is a p p ro a c h c a n lim it
th e s e a rc h c a p a b ility o f G e n e tic A lg o rith m s a n d th u s th e e ffe c tiv e n e s s o f o u r p e a k p o w e r e s tim a tio n a lg o rith m . B y re s o rtin g to th e p e c u lia ritie s o f S e lfis h G e n e , s ta te
m o d ific a tio n is n o lo n g e r re q u ire d , a n d th u s th e e ffe c tiv e n e s s o f o u r a lg o rith m c a n b e
im p ro v e d .
T h e re m a in d e r o f th e p a p e r is o rg a n iz e d a s fo llo w s . S e c tio n 2 p re s e n ts a fo rm a l
d e s c rip tio n o f th e p ro b le m . S e c tio n 3 d e s c rib e s th e e v o lu tio n a ry a lg o rith m w e e x p lo it
in o u r to o l, a n d S e c tio n 4 p re s e n ts th e p e a k -p o w e r e s tim a tio n a lg o rith m w e
d e v e lo p e d . S e c tio n 5 re p o rts e x p e rim e n ta l re s u lts o n s o m e b e n c h m a rk c irc u its .
F in a lly , S e c tio n 6 d ra w s s o m e c o n c lu s io n s .
W
W
Prediction of Power Requirements for High-Speed Circuits
249
2 . Pro b l e m f o rmul at io n
We assume that the sequential circuits we are analyzing are manufactured with
CMOS technology. Power consumption in CMOS circuits is mainly due to the
number of times the output of circuit gates switches from 0 to 1 and vice-versa. Peakpower consumption is defined as the maximum power consumption that takes place
when a couple of input vectors are applied to the circuit inputs, starting from a given
initial state. The procedure to measure peak-power can be outlined as follows:
1. the circuit memory elements are initialized to a given state S;
2. an input vector 9 1 is applied to the circuit, and a clock pulse is applied;
W
3. an input vector 9 is applied to the circuit, a clock pulse is applied and the power
consumption is measured during the transient period.
The first two steps lead the circuit to a configuration where the circuit nodes hold
known values (either 0 or 1). We refer to the power measured during the last step as
3 , and use to notation 3 ( 6 , 9 1 , 9 ) to indicate that power is function of the 3-tuple
W
I
I
W
W
(S, 9 1 , 9 ). The classical peak-power problem formulation that can be found in [3]
and [4] is the following:
W
W
33
0$; ( 3I ( 6 , 9W 1 , 9W )), 6 % Q , 9 % P
(1)
where PP is the peak-power consumption, % {0,1} , n is the number of memory
elements in the circuit and m is the number of circuit inputs. This formulation neglects
the fact that in sequential circuits the number of reachable states is often less than 2 Q .
Therefore, algorithms that solve equation (1) could compute 3-tuples that correspond
to PP values that cannot be obtained, since the memory elements can never reach the
configuration S during circuit operations. Power figures attained by solving equation
(1) are overestimation of actual peak-power consumption, and if adopted during
design could lead to unnecessary expensive design solutions. The authors of [3] and
[4] suggest replacing an unreachable state in the solution of (1) with the closest
reachable one. This approach weakens the correlation between S and (S, 9 1 , 9 ) and
usually lead to peak-power figures that underestimate the actual power consumption.
As a consequence, the obtained peak-power prediction could lead to wrong design
solutions.
The problem formulation we propose is the following:
W
33
0$; ( 3I ( 6 ,9W 1 ,9W )), 6 6,9 % P
W
(2)
where 6 % Q is the set of states that the sequential circuit under analysis can reach
starting from the reset state, i.e., the state where all the memory elements are set to 0.
By solving equation (2) and therefore by considering reachability of the state S as a
dimension of the search space, we compute peak-power figures more accurate than
what equation (1) provides.
2 5 0
F . C o rn o e t a l.
3 . T h e S e l f i s h G e n e a l g o r i t h m
T h e S e lfis h G e n e a lg o r ith m (S G ) is a n e v o lu tio n a ry o p tim iz a tio n a lg o rith m b a s e d
o n a re c e n t in te rp re ta tio n o f th e D a rw in ia n th e o ry . It e v o lv e s a p o p u la tio n o f
in d iv id u a ls s e e k in g fo r th e fitte s t o n e . In th e s e lfis h g e n e b io lo g ic a l th e o ry , p o p u la tio n
its e lf c a n b e s im p ly s e e n a s a p o o l o f g e n e s w h e re th e n u m b e r o f in d iv id u a ls , a n d th e ir
s p e c ific id e n tity , a re n o t o f in te re s t. T h e re fo re , d iffe re n tly fro m o th e r e v o lu tio n a ry
a lg o rith m s , th e S G re s o rts to a s ta tis tic a l c h a ra c te riz a tio n o f th e p o p u la tio n , b y
re p re s e n tin g a n d e v o lv in g s o m e s ta tis tic a l p a ra m e te rs o n ly . E v o lu tio n p ro c e e d s in
d is c re te s te p s : in d iv id u a ls a re e x tra c te d fro m th e p o p u la tio n , c o lla te d in to u rn a m e n ts
a n d w in n e r o ffs p rin g is a llo w e d to s p re a d b a c k in to th e p o p u la tio n .
A n in d iv id u a l is id e n tifie d b y th e lis t o f its g e n e s . T h e w h o le lis t o f g e n e s is c a lle d
g e n o m e a n d a p o s itio n in th e g e n o m e is te rm e d lo c u s . E a c h lo c u s c a n b e o c c u p ie d b y
d iffe re n t g e n e s . A ll th e s e c a n d id a te s a re c a lle d th e g e n e a lle le s . In th e c o n te x t o f a n
o p tim iz a tio n p ro b le m , lo o k in g fo r th e fitte s t in d iv id u a l c o rre s p o n d s to d e te rm in e th e
b e s t s e t o f g e n e s a c c o rd in g to th e fu n c tio n to b e o p tim iz e d .
S in c e th e S G “ v irtu a l” p o p u la tio n is u n lim ite d , in d iv id u a ls c a n b e c o n s id e re d to b e
u n iq u e , b u t s o m e g e n e s w o u ld c e rta in ly b e m o re fre q u e n t th a n o th e rs m ig h t. A t th e
e n d o f th e e v o lu tio n p ro c e s s , th e fr e q u e n c y o f a g e n e m e a s u re s its s u c c e s s a g a in s t its
a lle le s . H o w e v e r, a t th e b e g in n in g o f th e e v o lu tio n p ro c e s s , th e fre q u e n c y c a n b e
re g a rd e d a s th e g e n e d e s ir a b ility . W h e n th e m a jo rity o f a p o p u la tio n is c h a ra c te riz e d
b y th e p re s e n c e o f a c e rta in c h a ra c te ris tic , n e w tra its m u s t h a rm o n iz e w ith it in o rd e r
to s p re a d .
JHQRPH S G ( 9LUWXDO3RSXODWLRQ3)
{
JHQRPH%(67, *, *, ZLQQHU, ORVHU;
it e r = 0 ;
%(67 = se l e c t _ individual ( 3) ;
EHVWVRIDU
do {
++it e r;
* = se l e c t _ individual ( 3) ;
* = se l e c t _ individual ( 3) ;
t o urname nt ( *, *) ;LGHQWLI\ZLQQHUDQGORVHU
inc re ase _ al l e l e _ f re que nc ie s( 3, ZLQQHU) ;
de c re ase _ al l e l e _ f re que nc ie s( 3, ORVHU) ;
if ( ZLQQHU is pre f e rabl e t o %(67)
%(67 = ZLQQHU;
} wh il e ( st e ady_ st at e ( ) ==F AL SE && it e r<max_ it e r) ;
re t urn %(67;
}
F ig . 1 : S e lfis h G e n e a lg o rith m
T h
im p le
m o tiv
d e a l w
e
p
m e n
a tio
ith
s e u d o -c o d e o f th e S
ta tio n d e ta ils a b o u t th
n s a re b e tte r a n a ly z e d
m o re c o m p le x fitn e s s
G
e S
in
la n
a lg o rith
a lg o r
[1 0 ]. A
d sc a p e s
G
m
ith
n
is
p se u d o -c o d e .
c o re is
a re a v
e x te n s io n
d e s c rib e d
m
re p o rte d in F ig . 1 . F u rth e r
a ila b le in [9 ], w h ile b io lo g ic a l
to th e b a s ic S G a lg o rith m to
in [1 1 ].
Prediction of Power Requirements for High-Speed Circuits
251
4. Po w e r e st imat io n al go rit h m
The peak-power estimation algorithm we developed is composed of the following
steps:
1. the set of the reachable states 6 % Q is computed, which cardinality is 6 . It can
be computed either resorting to exact symbolic calculation techniques [7] or
through logic simulation;
2. peak-power estimation is performed. The SG is run to solve equation (2) where the
genome represents the 3-tuple (S, 9 1 , 9 ) and is composed of 1 2 P loci. The
W
W
first locus is an index ranging from 0 to 6 1 representing a reachable state
6 6 . The remaining loci are binary values, coding the couple of vectors ( 9 1 ,
9 ).
W
W
During step 2, the power consumption 3 ( 6 , 9 1 , 9 ) is computed resorting to a
I
W
W
unit-delay logic simulator. The adoption of a logic simulator is a well-known effective
approach to measure power consumption in CMOS circuits [6], since it conjugates
simulation speed with accuracy.
5 . Expe rime nt al re sul t s
A prototypical version of our algorithm named SG-ALPS, Selfish Gene-based
AnaLyzer of Power in Sequential circuits, has been written, which implements the
above-introduced procedures. The tool consists of about 500 lines of ANSI C code
and exploits the SG and logic simulation packages developed at our institution.
Reachability analysis has been performed resorting to exact calculation techniques
exploiting the BDD [7] package developed at our institution.
The subset of ISCAS’89 sequential circuits tractable with symbolic calculation
techniques have been used to evaluate the performance of SG-ALPS: all the
experiments have been performed on a Sun UltraSparc 5/333 with 256 MB RAM. To
compare SG-ALPS with a state-of-the-art tool, we have re-implemented the algorithm
proposed in [4]. Experimental results of our re-implementation of [4] are the same of
[4]; we can therefore perform a fair comparison between the two tools.
Two sets of experiments have been performed. The first one aims at empirically
showing that solving equation (1) as done by [4] could greatly underestimate peakpower consumption. Conversely, the second set of experiments aims at assessing the
effectiveness of the approach we propose.
Table 1 reports results we gathered with our implementation of [4]. The first
column reports the benchmark name, PPU and PPR report respectively the peakpower consumption obtained by solving equation (1) (which takes into account also
unreachable states) and the power obtained by replacing the state in the solution with
the closest reachable one. Column ' reports the difference between these power
figures, while column CPU reports the time requirements.
By observing the column ', one can observe that several circuits exist where peak-
252
F. Corno et al.
power strongly depends on the reachability of the initial state S and on the correlation
between S and the vectors ( 9 1 , 9 ). Where a significant loss in peak-power is
found, a large portion of the state S has been modified to make it reachable, thus a
high difference exists between the ideal initial state and the selected reachable one.
W
Circ
s208
s298
s344
s349
s382
s386
s400
s420
s444
s499
s510
s526
s641
s713
s820
s832
s1196
s1238
s1488
s1494
Avg.
PPU
0.900
1.007
1.573
1.020
0.961
0.855
1.000
0.893
1.077
0.595
0.860
0.915
2.753
2.793
0.951
0.927
1.035
1.015
1.235
1.234
W
PPR
0.900
0.833
1.552
0.869
0.790
0.855
0.435
0.893
0.551
0.230
0.860
0.631
2.739
0.837
0.951
0.927
1.020
1.000
1.156
1.155
'
[%]
0.00
-17.28
-1.34
-14.80
-17.79
0.00
-56.50
0.00
-48.84
-61.34
0.00
-31.04
-0.50
-70.02
0.00
0.00
-1.45
-1.47
-6.40
-6.39
-15.96
CPU
[s]
3.2
3.5
4.7
4.6
4.9
4.6
5.2
5.3
6.3
4.9
6.7
7.0
17.0
16.5
9.9
9.9
14.7
15.3
22.7
22.6
10.4
T ab l e 1 . Analysis of algorithm [4]
Table 2 reports results obtained by running SG-ALPS when max_iter (Fig 1) is
set to 1000. Column SG-ALPS reports the peak-power figures predicted by our
algorithm, while the third column reports the best results attained by the algorithm
proposed in [4]. We compare the two algorithms in column '. Finally, the CPU time
requirements are reported.
As far as peak-power estimation accuracy is concerned, we can conclude that our
approach is superior to [4]. Even if on the average SG-ALPS attains PP figures 21%
higher than [4], several circuits exist where SG-ALPS computes power figures 50%
higher than [4], thus showing the importance of considering reachability during the
optimization phase.
As far as CPU time is concerned, the algorithm proposed in [4] is far more
effective than SG-ALPS. This is mainly due to two factors: SG requires more time to
converge than Genetic Algorithms, and guaranteeing state reachability during the
optimization process is a time consuming operation.
P re d ic tio n o f P o w e r R e q u ire m e n ts fo r H ig h -S p e e d C irc u its
2 5 3
W e d o p o in t o u t th a t th e C P U re q u ire m e n ts o f S G -A L P S , a lth o u g h h ig h e r th a n
th o s e o f p re v io u s ly p ro p o s e d a p p ro a c h e s , s till re m a in in th e o rd e r o f m in u te s , w h ic h is
n e g lig ib le w ith re s p e c t to th e tim e re q u ire d b y m o s t s te p s in th e c u rre n t c irc u it d e s ig n
flo w . T h e in c re a s e in th e C P U tim e is w o rth p a id b y th e im p ro v e m e n t in th e a tta in e d
e s tim a tio n q u a lity .
W h e n c o m p a re d w ith th e m e th o d p ro p o s e d in [8 ], im p le m e n tin g a s im ila r a p p ro a c h
u s in g G e n e tic A lg o rith m , w e o b s e rv e d th a t th e a lg o rith m p ro p o s e d h e re p ro v id e s
c o m p a ra b le re s u lts . H o w e v e r, th e a d o p tio n o f th e S G a lg o rith m a llo w s to m o re e a s ily
d e a l w ith o p tim iz a tio n p ro b le m s fo r w h ic h a h e te ro g e n e o u s s o lu tio n e n c o d in g is m o re
s u ita b le .
C irc
s2 0
s2 9
s3 4
s3 4
s3 8
s3 8
s4 0
s4 2
s4 4
s4 9
s5 1
s5 2
s6 4
s7 1
s8 2
s8 3
s1 1
s1 2
s1 4
s1 4
A v
8
8
4
9
2
6
0
0
4
9
0
6
1
3
0
2
9 6
3 8
8 8
9 4
g .
S G -A L P S
[4 ]
P P
P P
0 .9
0 .8
1 .5
0 .8
0 .7
0 .8
0 .4
0 .8
0 .5
0 .2
0 .8
0 .6
2 .7
0 .8
0 .9
0 .9
1 .0
1 .0
1 .1
1 .1
0 .9
0 .8
1 .5
1 .5
0 .8
0 .8
0 .8
0 .8
0 .9
0 .2
0 .8
0 .7
2 .7
2 .1
0 .9
0 .9
1 .0
1 .0
1 .1
1 .1
0 0
3 3
5 3
3 5
5 9
7 2
5 0
9 3
4 2
3 0
6 9
3 7
3 9
5 5
6 2
6 7
4 4
5 0
5 7
5 6
∆
[% ]
0 0
3 3
5 2
6 9
7
9 0
5 5
3 5
9
9 3
5 1
7
3 0
6 0
3 1
1
3 9
3 7
5 1
2 7
2 0
0 0
5 6
5 5
1 5
C P U
[s]
0 .0 0
0 .0 0
0 .0 3
6 .6 5
8 .7 2
2 .0 1
5 .4 3
0 .0 0
0 .9 6
0 .0 0
1 .0 5
6 .8 0
0 .0 0
7 .3 6
1 .1 6
4 .3 5
2 .3 7
5 .0 0
0 .0 4
0 .0 8
2 1 .0
2
2
1
1
1
3
1
5
1
1
1
2 8
5 6
1 8
3 1
9 8
3 7
1 0
2 7
0 3
1 5
4 5
9 9
5 8
9 8
5 0
7 1
6 8
9 5
1 7
1 6
3 0
.1
.6
.0
.9
.5
.5
.6
.1
.7
.5
.0
.7
.3
.2
.5
.4
.8
.6
.8
.4
.9
T a b le 2 . S G -A L P S re s u lts
6 . C o n c l u s i o n s
H ig h
re q u e st
c o m m u
c o n su m
-s p e e d te le c o m m u n ic a tio n e q u ip m e n
fo r h ig h -s p e e d d a ta d e liv e rin g . W
n ic a tio n e q u ip m e n t, p a rtic u la r c a re m
p tio n . In p a rtic u la r, p e a k -p o w e r e s tim
t is
h e n
u st
a tio
re q u ire d
d e s ig n in
b e p o se
n is m a n
to
c o p e
h ig h -s
d to th e
d a to ry fo
g
w ith th e
p e e d c irc u
c o n tro l o f
r d e s ig n in g
m a
its
p o
to
rk e t
fo r
w e r
d a y
2 5 4
F . C o rn o e t a l.
h ig h -s p e e d c irc u its . In th is p a p e r th e p e a k -p o w e r p re d ic tio n p ro b le m h a s b e e n
fo rm u la te d a s a c o n s tra in e d o p tim iz a tio n p ro b le m , a n d a n a lg o rith m b a s e d o n a n e w
e v o lu tio n a ry p a ra d ig m h a s b e e n p ro p o s e d . T h e a lg o rith m is d e s ig n e d fo r a d d re s s in g
s e q u e n tia l c irc u its a n d th a n k s to its a b ility o f g u a ra n te e in g fe a s ib ility o f s o lu tio n s
d u rin g th e o p tim iz a tio n p ro c e s s it is m o re e ffe c tiv e th a n o th e r a p p ro a c h e s .
M o re o v e r, in th is w o rk w e e x p e rim e n te d th a t S G is b e tte r s u ite d to a d d re s s
o p tim iz a tio n p r o b le m s h a v in g h e te r o g e n e o u s s o lu tio n s , i.e ., s o lu tio n s m ix in g
c o m p o n e n ts d e fin e d o v e r d iffe re n t d o m a in s . A s a n e x a m p le , in p e a k -p o w e r e s tim a tio n
th e s o lu tio n e m b e d s a s ta te th a t is n a tu ra lly e x p re s s e d a s a n in d e x in a s ta te ta b le , a n d
a c o u p le o f v e c to rs , re p re s e n te d a s a c o lle c tio n o f b in a ry v a lu e s .
W e a re c u rre n tly w o rk in g to w a rd a n im p le m e n ta tio n o f S G -A L P S w h e re
a p p ro x im a te d re a c h a b ility a n a ly s is is p e rfo rm e d th ro u g h lo g ic s im u la tio n . T h a n k s to
th is im p ro v e m e n t, w e w ill b e a b le to a d d re s s la rg e s e q u e n tia l c irc u it c u rre n tly n o t
tra c ta b le b y s y m b o lic c a lc u la tio n te c h n iq u e s .
7 . R e f e r e n c e s
[1 ]
[2 ]
[3 ]
[4 ]
[5 ]
[6 ]
[7 ]
[8 ]
[9 ]
[1 0 ]
[1 1 ]
S . M a n n e , A . P a rd o , R . I. B a h a r, G . D . H a c h te l, F . S o m e n z i, E . M a c ii, M . P o n c in o ,
“ C o m p u tin g th e m a x im u m p o w e r c y c le s o f a s e q u e n tia l c irc u it” , P ro c . o f IE E E /A C M
D A C , 1 9 9 5 , p p . 2 3 -2 8 7
C .- Y . W a n g , K . R o y , “ M a x im u m C u r r e n t E s tim a tio n in C M O S C ir c u its U s in g
D e te rm in is tic a n d S ta tis tic a l T e c h n iq u e s ” , IE E E T ra n s . o n V L S I S y s te m s , M a rc h 1 9 9 8 ,
p p . 1 3 4 -1 4 0
M .S . H s ia o , E .M . R u d n ic k , J . P a te l, “ K 2 : A n E s tim a to r f o r P e a k S u s ta in a b le P o w e r o f
V L S I C irc u its ” , P ro c . o f In t. S y m p . o n L o w P o w e r E le c tro n ic s a n d D e s ig n , 1 9 9 7 , p p .
1 7 8 -1 8 3
M .S . H s ia o , E .M . R u d n ic k , J . P a te l, “ E f f e c ts o f D e la y M o d e ls o n P e a k P o w e r E s tim a tio n
o f V L S I S e q u e n tia l C irc u its ” , P ro c . o f IE E E /A C M IC C A D , 1 9 9 7 , p p . 4 5 -5 1
F . C o rn o , M . S o n z a R e o rd a , G . S q u ille ro , “ O p tim iz in g D e c e p tiv e F u n c tio n s w ith th e S G C l a n s A l g o r i t h m ” , C E C '9 9 : 1 9 9 9 C o n g r e s s o n E v o l u t i o n a r y C o m p u t a t i o n , W a s h i n g t o n
D C (U S A ), J u ly 1 9 9 9 , p p . 2 1 9 0 -2 1 9 5
A . G h o s h , S . D e v a d a s , K . K u e tz e r, J . W h ite , “ E s tim a tio n o f a v e ra g e s w itc h in g a c tiv ity in
c o m b in a tio n a l a n d s e q u e n tia l c irc u its ” , P ro c . o f IE E E /A C M D A C , 1 9 9 2 , p p . 2 5 3 -2 5 9
R . E . B ry a n t, “ S y m b o lic B o o le a n M a n ip u la tio n w ith O rd e re d B in a ry D e c is io n
D ia g r a m s ,” A C M C o m p u tin g S u r v e y s , V o l. 2 4 , N o . 3 , 1 9 9 2 , p p . 2 9 3 - 3 1 8
F . C o rn o , M . R e b a u d e n g o , M . S o n z a R e o rd a , M . V io la n te , “ A L P S : A P e a k -P o w e r
E s tim a tio n A lg o rith m
f o r S e q u e n t i a l C i r c u i t s ” , G L S - V L S I '9 9 : 8 t h G r e a t L a k e s
S y m p o s iu m o n V L S I, 1 9 9 9 , p p . 3 5 0 -3 5 3
F . C o rn o , M . S o n z a R e o rd a , G . S q u ille ro , “ T h e S e lfis h G e n e A lg o rith m : a N e w
E v o l u t i o n a r y O p t i m i z a t i o n S t r a t e g y ” , S A C '9 8 : 1 3 t h A n n u a l A C M
S y m p o s iu m o n
A p p lie d C o m p u tin g , 1 9 9 8 , p p . 3 4 9 -3 5 5
F . C o rn o , M . S o n z a R e o rd a , G . S q u ille ro , “ A N e w E v o lu tio n a ry A lg o rith m In s p ire d b y
t h e S e l f i s h G e n e T h e o r y ” , I C E C '9 8 : I E E E I n t e r n a t i o n a l C o n f e r e n c e o n E v o l u t i o n a r y
C o m p u ta tio n , 1 9 9 8 , p p . 5 7 5 -5 8 0
F . C o rn o , M . S o n z a R e o rd a , G . S q u ille ro , “ O p tim iz in g D e c e p tiv e F u n c tio n s w ith th e S G C l a n s A l g o r i t h m ” , C E C '9 9 : 1 9 9 9 C o n g r e s s o n E v o l u t i o n a r y C o m p u t a t i o n , 1 9 9 9 , p p .
2 1 9 0 -2 1 9 5
A Communication Architecture for Multi-Agent
Learning Systems
N. Ireson, Y. J. Cao, L. Bull and R. Miles
Intelligent Computer Systems Centre
Faculty of Computer Studies and Mathematics
University of the West of England, Bristol, BS16 1QY, UK
Bristol, BS16 1QY, UK
Abstract. This paper presents a simple communication architecture
for Multi-Agent Learning Systems. The service provided by the communication architecture allows each agent to connect to the user interface,
the application and the other agents. The communication architecture is
implemented using TCPIP. An application example in a simplied trafc environment shows that the communication architecture can provide
reliable and e cient communication services for Multi-Agent Learning
Systems.
1 Introduction
Many researchers in the eld of Distributed Articial Intelligence are beginning
to build agents that can work in a complex, dynamic multi-agent domains 1.
Such domains include virtual theater 2, realistic virtual training environments
1, RoboCup robotic and virtual soccer 3 and robotic collaboration by observation 4. This is because that there is a realisation of the benets of using
problem-solving models based upon an interacting group of agents rather than
a single agent and multi-agent systems can benet from the inherent properties
of distributed systems, i.e. parallelism, robustness, scalability.
Learning in multi-agent systems has been seen as important both in removing the need to hard code" the agent behaviour, as for certain problems
the appropriate behaviour is unknown 5, 6. It is motivated by the insight that
it is impossible to determine a-priori the complete knowledge that must exist
within each component of a distributed, heterogeneous system in order to allow
satisfactory performance of that system. Especially if we want to exploit the
potential of modularity, such that it is possible for individual agents to join and
leave the multi-agent system, there is a constant need for the acquisition of new
and the adaptation of already existing knowledge, i.e., for learning.
Within this setting, dierent kinds of learning tasks must be investigated,
such as `traditional' single agent learning tasks, learning in teams, learning to
act within a team, and learning to cooperate with other agents. To solve any
of these tasks, communication, i.e., the existence of appropriate information
that can be communicated to the learning agents is of primary importance.
S . C a g n o n i e t a l. ( E d s .) : E v o W o r k s h o p s 2 0 0 0 , L N C S 1 8 0 3 , p p . 2 5 5 − 2 6 6 , 2 0 0 0 .
S p rin g e r-V e rla g B e rlin H e id e lb e rg 2 0 0 0
2 5 6
N . Ire s o n e t a l.
In this work, we consider the communication issues in the cooperative
multi-agent learning systems. In a cooperative multi-agent system, the agents
can cooperate either implicitly or explicitly. With implicit cooperation agents
act selshly to satisfy their individual goals but their actions can have benecial
eects upon other agents. With explicit cooperation, agents share information
and rewards, thus perform actions which provide mutual benet, involving some
form of direct communication.
Sen et al showed that agents attempting to optimise the use of limited
resources converge to optimal states faster with a limited view 7 . They also
investigated the formation of agent coalitions, where agents within a coalition
share information about their resource utilisation intentions. The use of such
coalitions further improves the system's convergence to the optima. Prasad
and Lesser examined an explicitly cooperative directly communicating system
in which agents learning to communicate only relevant information dependant
on situation specics 8 . The learning algorithm uses the Instance-Based learning paradigm. At the end of a particular problem solving run, agents assess their
coordination strategy according to four performance measures. Agents then derive new strategies and broadcast these to the other agents. Seredynski et al
developed a coevolutionary multi-agent system showing that global behaviour
evolves via only local cooperation between agents acting without global information about the system. Seredynski introduces an exchange process into the
game, which redistributes the payo 9, 10 . The players are placed in a ring
and play a set number of games with a selected number of neighbours, i.e. local
interactions. Three exchange schemes were tested:
1. no payo exchange, no cooperation
2. payo is exchanged amongst interacting players, local cooperation sharing
3. payo is exchanged amongst all players, global cooperative sharing.
They considered two simple coevolutionary schemes: loosely coupled GA
and loosely coupled classier systems. The term loosely coupled" refers to the
individuals in a population being evaluated only on their local rather than a
global tness function. The experiments showed that with no payo exchange
the players evolve to the Nash equilibrium both defecting, whilst with both the
local and global payo exchange players evolved to cooperate. In fact the payo
exchange alters the game's payo matrix and encourage cooperation. Seredynski
applied the technique of payo exchange to a dynamic mapping and scheduling
problem 9 . Bull et al applied the learning classier system design to a multiagent environment in which each of a number of classier systems represents a
co-operating communicating agent 11 . It is demonstrated that the evolution
of multiple co-operating agents can give improved performance over an equivalent single agent model. The performances of the varying components, such as
reinforcements, discovery system, in a multi-agent environment, were examined
in detail in 12 .
In this work, we extend the framework of the multi-agent systems in 10, 12
to a general version and aim at developing a communication architecture for
these multi-agent learning systems. In these environments, each agent is dened
by a classier system which evolves a set of control rules and is concerned with
agents as software rather than theoretical constructs. In this communication
architecture, each agent is allowed to connect to user interface, the application
A C o m m u n ic a tio n A rc h ite c tu re fo r M u lti-a g e n t L e a rn in g S y s te m s
2 5 7
A g e n t
C la s s ifie r S y s te m
C o m m u n ic a tio n
S e rv e r
Figure 1: Structure of the agent
and the other agents. An application example is given to show that the communication architecture can provide reliable and ecient communication services.
2 The Framework of the Multi-Agent Learning
System
Multi-agent learning systems usually consist of a number of collaborative agents
and consider how these agents can interact to eectively cooperate in problem
solving tasks. They have many applications in process control, network management, scheduling, etc. 9, 13, 17 .
As with a number of contemporary elds on computer science, such as AI
and Articial Life ALife the denition of an agent ranges from a strong to a
weak notion. On the stronger end of the scale an agent is deemed to possess
properties akin to those found in humans, such as knowledge, belief, intention,
etc. The weaker notion of an agent tends to be more pragmatic and associated
with agent software engineering. Such agents can be dened as possessing less
anthropomorphic properties, for example 18 :
autonomy: operate without the direct intervention of others, and have
some control over actions and internal states
communication: interactions with other agents
reactivity: perception of their environment and timely response to changes
that occur in it.
Pro-activeness: not merely reacting to events but exhibit pre-active goaldirected behaviours.
The multi-agent learning system considered in this paper contains a number of distributed, communicating agents, where each agent, as shown in Figure
1, has a learning classier system providing the rule base and control actions
and a communication server which is used to connect the agent to the user interface, the application and to other agents. These two elements of the agent are
separate since as messages are passed around the agent network, the communication server acts independently of the classier system to route the message
to its neighbours. Another reason for keeping the communication server distinct
2 5 8
N . Ire s o n e t a l.
S e n s o rs /
A c tu a to rs
U s e r I n te r f a c e
M
o n ito r
In te rf a c e s
C o n tro l
T C P /I P
A g e n t
S ta te D e te c to r
R M
S h a r e d
I n fo r m a t io n
I
A c t io n
E ffe c to r
S h a r e d
R e w a r d
P e r fo r m a n c e
E v a lu a t io n
A g e n t
A p p lic a tio n
Figure 2: Interaction structure of the agent
from the classier system is that communication is likely to be implementation
specic even in the test applications. Thus it is necessary when specifying the
communication server to consider the general requirements of setting up and
maintaining the communication in a multi-agent learning system rather than
those in a specic software and hardware implementation.
The communication server must provide the following services:
Interaction between the application and the agents: The interface between
the agents and the application is dened by the agents' detectors and
e ectors. The application must package the system state into a message
which is in the representation expected by the agents. The application
receives an action from the detectors which is interpreted to e ect the
system.
Interaction between the User Interface and the agents: The amount of
performance information reported by the agents is parameterised. During
the development of an application it is likely that the user will wish to
monitor performance more closely to ensure the system has been correctly
congured. Another consideration is whether information performance
statistics, warnings, error, etc. is logged to a le or sent to the user
interface directly.
Inter-agent communication: The communication strategy between the agents
is e ected by the application and the learning strategy to be employed.
Each agent can communicate with a specic other agent or group of agents,
known as an agents neighbourhood, the user species the agents contained
in each agents neighbourhood.
The Figure 2 shows a representation of the interactions between an agent and
A C o m m u n ic a tio n A rc h ite c tu re fo r M u lti-a g e n t L e a rn in g S y s te m s
2 5 9
A g e n t N e tw o rk
2
1
3
6
5
7
1 0
9
1 3
4
8
1 1
1 4
1 2
1 5
1 6
Figure 3: Structure of the agent network
its environment which contains three types of actors: the user, the application,
and other agents.
While it is important to examine the research into agent languages and
implementations such as KQML, such work provides functionality beyond the
needs of these systems. The communication in such learning systems will have a
limited syntax and have a well dened content, thus the language required will
be fairly simplistic. We have specied the nature of the communication, which
includes: utility measure enabling tness" sharing environmental state actions performed or to perform intentions shared classiers.
3 The Proposed Communication Architecture
3.1 The Whole Agent Network
An agent network is dened by the characteristics of the application, which determines the appropriate distribution and connections between agents. Each
agent is allocated a specic address, and a port for each channel of communication, this includes communication with the user interface, application and other
agents.
There are two basic protocols of communication. The simplest is to broadcast the message which is received by all the agents, the message can include a
tag to identify the sender and nature of the message. It is the responsibility of
the receiving agent to determine whether to utilise or ignore the information in
the message. The second means of communication is to send messages to the
agents in the local neighbourhood. In this approach each agent is the centre of
a neighbourhood, those agents contained in the neighbourhood from none, i.e.
no communication, to all other agents receive any messages sent by the central
agent. As neighbourhoods might overlap an agent can be a member of a number of neighbourhoods, for example, for the agent network, shown in Figure 3,
neighbourhoods are constructed from an agents nearest neighbours, thus sixteen
2 6 0
N . Ire s o n e t a l.
groups, each centred on one agent, are formed:
Group 1: 1, 2, 5
Group 2: 2, 1, 3, 6
Group 3: 3, 2, 4, 7
...
Group 6: 6, 2, 5, 7, 10
Group 7: 7, 3, 6, 8, 11
...
Group 16: 16, 12, 15.
This approach is general known as multicast or one-to-many communication. There are a number of possible approaches to this form of communication
in distributed systems from ooding where each node sends a copy of the incoming message to all the connected nodes except for the message source node to
routing where messages follow a pre-specied path to their destination node.
The principal problem with ooding is that it causes a great deal of redundant communication, also as each node must check if it has already received
each incoming message, with frequent communication this can lead to a communication overload. Routing requires a more complex initialisation process but
minimises the communication trac. Each node when it receives a message
refers to a lookup table giving the nodes to which copies of the message are sent.
Unlike ood broadcast there is no redundant communication, thus the severing
of a communication link will cause at least one node to fail to receive messages.
In practice the choice of method should reect the constraints of the system, i.e.
the trade-o between fault tolerance and communication load.
3.2 Initialisation of the Agent Network
The initialisation process rstly involves the setup and opening of communication
channels from agent to user interface, agent to application and agent to other
agents. Once this has been successfully completed, the classier system can be
initialised, the agent then waits for the rst message from the application to
begin its control process.
The initialisation of communication involve each agent connecting to the
user interface, application and other agents. All these channels might involve
two-way communication. During the initialisation the agents open a communication channel and await a connection message. The channel is tested to ensure
the communication is setup correctly as although the conguration parameters
have been previously checked for consistency, the parameters may be inconsistent
with the physical communication process, also this process might be faulty.
Although the term socket" is used in the specication as the medium
to connect communication channels in implementation other methods can be
used, such as calls to remote objects, when using RMI or DCOM. The basis of
the communication initialisation and run-time processes are not aected. The
creation of the communication object and binding in a remote registry on a
given hostname and port replaces the creation of a server socket and calls to
the remote object replace read and write calls to the sockets.
Note that it is possible for the communication server to create separate
processes to listen on the communication channel for messages, this allows the
A C o m m u n ic a tio n A rc h ite c tu re fo r M u lti-a g e n t L e a rn in g S y s te m s
C o m m u n ic a tio n
C o n fig u ra tio n
2 6 1
P a ra m e te rs
C o m m u n ic a tio n S e rv e r
2 . C re a te S e rv e r
3 . C o n n e c t
(C o m m u n ic a tio n S e rv e r)
C o n n e c to r
4 . R e q u e st
C o n n e c tio n
5 . R e q u e st
C o n n e c tio n
U s e r In te rfa c e ,
A p p lic a tio n o r
A g e n t N e ig h b o u r
1 . C re a te M o n ito r
6 . A d d C o n n e c tio n
C o n n e c to r M o n ito r
7 . A c c e p t C o n n e c tio n
C o m m u n ic a tio n S e rv ic e
8 . In itia lis e S tre a m s
C o n n e c tio n
Figure 4: Structure of communication server
agent to be reactive to external messages. The communication with neighbours
requires a single channel for incoming messages, and separate channels from
sending to each neighbour except if the messages are broadcast on sent via a
proxy. The initialisation of the communication server, as shown in Figure 4,
involves the following steps:
1. The Communication Server object creates the speci c Communication Services Application, User Interface or Neighbourhood as speci ed by the
con guration.
2. The Communication Server object create a monitor which maintains the
list of current connections.
3. The Communication Server object passes the Communication Service object and connection con guration information to the Connector object
which, for connection with the User Interface and Application and incoming channel from the neighbouring agents, opens a Server Socket on the
speci ed port and waits for a request to connect. For the outgoing channel
to the neighbouring agents the Connector object intermittently requests a
connection to the neighbours speci ed port.
4. The User Interface, Application or Neighbouring Agent sends a request to
connect.
5. The request to connect is accepted by the neighbour's server socket.
6. The Connector sends the Communication Service object and open socket
to the Communication Monitor.
7. The Communication Monitor object tests the communication channel, if
the test succeeds the Communication Service is passed to the Connection
object, otherwise the socket is closed and the failure reported.
2 6 2
N . Ire s o n e t a l.
Figure 5: Class diagram of the communication
8. The Connection object starts the thread to handle the connection and
passes the input and output streams to the Communication Service object.
3.3 Logical View of the Communication
A class diagram of the communication is given in Figure 5. The function of each
class object in the communication is described briey as follows:
Communication Server: The Communication server supports a number of
multi-threaded channels. It creates or opens each channel on a speci ed
port allowing the Agent to communicate with the User Interface, Application and other Agents. It provides the ability to send messages whilst
listening on the port for incoming messages which are passed to the appropriate objects.
Connection Monitor: The Connection Monitor object maintains a list of
the current connections. The thread waits to be noti ed if a connection
terminates and updates the list.
Connector: The Connector class either listens for a connection on a speci ed port using a server socket or connects to another agent's server
socket. Once accepted the socket is sent to the Connection Monitor.
Connection: Connection objects are created by the Connector thread using
the Communication Monitor method addConnection. It simply creates
a thread to handle the connection.
A C o m m u n ic a tio n A rc h ite c tu re fo r M u lti-a g e n t L e a rn in g S y s te m s
I
II
III
IV
2 6 3
Figure 6: The simulated trac environment
Communication Service: A general class for each of the types of communication required.
Listener: An object that listens on a communication channel for an incoming message.
User Interface: The communication service require by the User Interface.
This object handles messages received from or sent to the User Interface.
Incoming messages are interpreted and, if necessary, call the appropriate
function. Outgoing messages are packaged and sent to the User Interface.
Application: The communication service require by the Application. This
object handles messages received from or sent to the Application. Incoming messages generally system state or rewards are interpreted and, if
necessary, sent to the appropriate objects such as classi ers in the classi er system. Outgoing messages general actions are packaged and sent
to the Application.
Neighbourhood: The communication service require by the Agent Neighbourhood. This object handles messages received from or sent to the Agent
Neighbourhood. Incoming messages are interpreted and, if necessary, sent
to the appropriate functions in the classi er system. Outgoing messages
are packaged and sent to the neighbouring agents.
4 Application Example
Optimization of a group of trac signals over an area is typical multi-agent type
real-time planning problem without precise reference model given. To do this
planning, each signal should learn not only to acquire its control plans individually through reinforcement learning but also to cooperate with each other. This
requires communication between the agents. In this example, we developed a
multi-agent learning system, which is aimed at learning the ecient control rules
for the dynamic trac environment and with the communication provided by
2 6 4
N . Ire s o n e t a l.
5.5
5
4.5
Traffic Speed
4
3.5
3
2.5
2
1.5
0
1000
2000
3000
4000
5000
6000
Time Steps
7000
8000
9000
10000
Figure 7: Performance comparison of dierent control strategies
the developed communication architecture. Each agent has a classier system
providing the control strategy and a communication server which is used to
connect the agent to the user interface, the application and to other agents.
To control a tra c network, we associate an agent to each junction of the
tra c network. The agents are initialised according to the tra c network conguration and user-specied parameters. For the simulated 2 2 tra c network,
as shown in Figure 6, four agents, i.e., agents I, II, III, and IV, associating with
junctions I, II, III, and IV, are need to provide comprehensive control of the
network. Agent I has the neighbouring agents II and III, and agent II has the
neighbouring agents I and IV, etc. The communication server in each agent
provides the control actions of its neighbouring agents, and these information is
used to construct control rules for its junction.
The classier system employed is a version of Wilson's zeroth-level" system
ZCS 19, with some changes on the classier representation 20. The condition
part of each classier consists of six bits, which reects the scalar level of queue
length from each direction and the previous actions of the neighbouring agents.
In this application, the scalar level of the queue length is set to 4, which ranges
from 0 to 3, corresponding to the four linguistic variables, fzero small medium
large g. The action part indicates the required state of the signal. For instance,
for junction I, the rule 130201:1 says that if the queue from directions east and
west are small 1 and zero 0, but the queue from directions south and north are
large 3 and medium 2, and the previous neighbourhood junction controllers'
actions are vertically red 0 junction II and green 1 junction III, then the
tra c light stays green vertically 1 for a xed period of time. The performance
evaluation, reinforcement learning strategy, genetic algorithm and the simulated
tra c environment are all similar to those used in 20.
For comparison purpose, two types of control strategies are employed: random control strategy and the developed multi-agent learning system MALS
strategy. The random control strategy determines the tra c light's state 0
or 1 randomly at 50 of probability whilst MALS strategy determines the
tra c light's state according to the action of the winning classier of the agent.
A C o m m u n ic a tio n A rc h ite c tu re fo r M u lti-a g e n t L e a rn in g S y s te m s
2 6 5
Experiments were carried out for three dierent types of trac conditions. In
these simulations, the mean arrival rates for the cars are the same but the number of cars in the area is limited to 30, 60, and 90, corresponding to a sparse,
medium, and crowded trac condition. In all cases, the MALS strategy is found
to learn how to reduce the average queue length and improve the trac speed
in the network. For example, Figure 7 shows the average performances of the
random control strategy and MALS strategy respectively over 10 runs in the
crowded case, where the solid line represents MALS strategy and the dotted
line represents random control strategy. It can be seen that the MALS strategy
consistently learns and improves the trac speed over 10,000 iterations.
5 Conclusion and Future Work
We have extended the framework of the multi-agent learning systems in 10, 12
to a general case and developed a simple communication architecture for these
systems. The service provided by the communication architecture allows each
agent to connect to the user interface, the application and the other agents.
An application example shows that the communication architecture is reliable
and ecient. Although the communication architecture is implemented using
TCPIP, it can also be implemented using RMI or DCOM, via binding the
objects in the remote registry and making calls to the remote objects.
6 Acknowledgment
This work was carried out as part of the ESPRIT Framework V Vintage project
ESPRIT 25.569.
References
1 Tamble, M., Rosenbloom, P. S.: RESC: An approach for real-time, dynamic agent
tracking. In Proc. of the International Joint Conference on Articial Intelligence,
Montreal, Canada, 1995
2 Hayes-Roth, B., Brownston, L., Gen, R. V.: Multiagent collaboration in directed
improvisation. In Proc. of International Conference on Multi-Agent Systems. USA
1995
3 Kitano, H., Asada, M. Kuniyoshi, Y., Noda, I., Osawa, E.: The robot world cup
initiative. In Proc. IJCAI-95 Workshop on Entertainment and AIAlife, Montreal,
Canada 1995
4 Kuniyoshi, Y., Rougeaux, S., Ishii, M., Kita, N., Sakane, S., Kakikura, M.: Cooperation by observation: the framework and the basic task pattern. In Proc. IEEE
International Conference on Robotics and Automation. 1994
5 Weiss, G. and Sen, S. eds : Adaptation and Learning in Multi-Agent Systems.
Springer-Verlag, Berlin, Heidelberg, New York, 1995
6 Sen, S. ed : AAAI Spring Symposium on Adaptation, Coevolution and Learning
in Multiagent Systems. AAAI Press, 1996
7 Sen, S. Sekaran, M. and Hale: Learning To Coordinate without Sharing Information. In Proceedings of the Twelfth National Conference on Articial Intelligence,
1994 426-431.
2 6 6
N . Ire s o n e t a l.
8 Prasad, M.V.N. and Lesser, V.R.: Learning Problem Solving Control in Cooperative Multi-Agent Systems. Workshop on Multi-Agent Learning AAAI-97, 1997
9 Seredynski, F.: Coevolutionary Game-Theoretic Multi-Agent Systems: the Application to Mapping and Scheduling Problems Technical Report TR-96-045 Institute
of Computer Science, Polish Academy of Sciences, Warsaw, Poland. 1996
10 Seredynski, F., Cichosz, P. and Klebus, G. P: Learning classier systems in MultiAgent Environments, In Proc. First IEEIEEE International Conference on Genetic Algorithms in Engineering: Innovations and Applications, 1995 287292
11 Bull, L., Fogarty, T. C., and Snaith, M.: Evolution in Multi-Agent Systems:
Evolving Communicating Classier Systems for Gait in a Quadrupedal Robot.
In Eshelman, L. J. ed : Proceedings of the Sixth International Conference on
Genetic Algorithms, Morgan Kaufmann, 1995 382388
12 Bull, L: On ZCS in Multi-Agent Environments. Parallel Problem Solving From
Nature - PPSN V, Springer Verlag 1998 471480
13 Fleury, G., Goujon, J., Gourgand, M. and Lacomme, P., Multi-agent approach
and stochastic optimization: random events in manufacturing systems. Journal of
Intelligent Manufacturing, 10, 1, 1999 81102
14 Cao, Y. J. and Wu, Q. H.: A mixed-variable evolutionary programming for optimisation of mechanical design. International Journal of Engineering Intelligent
Systems, 7, 2, 1999 7782
15 Cao, Y. J. and Wu, Q. H.: An improved evolutionary programming approach to
economic dispatch. International Journal of Engineering Intelligent Systems, 6,
2, 1998 187194
16 Cao, Y. J. and Wu, Q. H.: Optimisation of control parameters in genetic algorithms: a stochastic approach. International Journal of Systems Science, 20, 2,
1999 551559
17 Kouiss, K., Pierreval, H. and Mebarki, N., Using multi-agent architecture in FMS
for dynamic scheduling. Journal of Intelligent Manufacturing, 8, 1, 1997 4148
18 Wooldridge, M. and Jennings, N.R.: Intelligent agents: theory and practice. In
The Knowledge Engineering Review, 10 2, 1995 115-152.
19 Wilson, S. W.: ZCS: A zeroth level classier system. Evolutionary Computation,
2, 1994 118
20 Cao, Y. J., Ireson, N. I., Bull, L. and Miles, R.: Design of Trac Junction Controller Using a Classier System and Fuzzy Logic. In Computational Intelligence:
Theory and Applications, Reusch, B. ed, Lecture Notes in Computer Sciences,
1625, Springer Verlag, 1999 342353
A n A m b u la n c e C r e w R o s te r in g S y s te m
P . V . G . B r a d b e e r † , C . F in d la y ‡ a n d T .C
F o g a r t y .¶
†
F if e C o lle g e o f F u r th e r a n d H ig h e r E d u c a tio n ., p v g b @ c it.f if e .a c .u k
F ife A m b u la n c e S e rv ic e .
¶
N a p ie r U n iv e r s ity ., t.f o g a r ty @ d c s .n a p ie r .a c .u k
‡
A b str a c t
T h e p r o d u c tio n o f a r o s te r fo r
p r a c tic a l, m a n a g e r ia l a n d s o c
o f a n in v e s tig a tio n in to th e c h a
r e p o r ts o n th e s u c c e s s to d a te o
c e p ta b le s o lu tio n to th e p r o b le
a q u ic k e r w a y o f te s tin g th e a
fo r m m u ltip le tim e c o n s u m in g
th e d u tie s o f a m b u la n c e c r
ia l c o n s tr a in ts . T h is d o c u m
r a c te r is tic s o f th e s e a r c h s p
f a n e v o lu tio n a r y a lg o r ith m
m . T h e v is u a lis a tio n m e th o
p p r o p r ia te n e s s o f r e p r e s e n
e x p e r im e n ts .
e w is s u b je c t to a v a r ie ty o f
e n t d e s c r ib e s th e fir s t s te p s
a c e fo r s u c h a p r o b le m a n d
a p p r o a c h in fin d in g a n a c d d e s c r ib e d is s u g g e s te d a s
ta tio n s th a n h a v in g to p e r -
S e c tio n 1 - I n tr o d u c tio n
R e c e n tly , th e d e c is io n w a s m a d e to re -e s ta b lis h th e a m b u la n c e s e rv ic e s u b -s ta tio n
a tta c h e d to th e V ic to ria H o s p ita l in K irk c a ld y . T h e p re v io u s a rra n g e m e n t w a s to p ro v id e a ll
a m b u la n c e s e rv ic e s fro m a c e n tra l b a s e in G le n ro th e s . ‘F ro n t L in e ’ p e rs o n n e l fo r th e n e w
s e rv ic e c o m p ris e s a g ro u p o f e ig h t s ta ff, fo u r o f w h o m a re d e s ig n a te d p a ra m e d ic s a n d h a v e
re c e iv e d s p e c ia lis e d tra in in g fo r th e ir d u tie s , a n d fo u r s ta ff w h o a re d e s ig n a te d m e d ic a l
te c h n ic ia n s , a ls o a fte r a p p ro p ria te tra in in g . T h is g ro u p o f s ta ff a re re q u ire d to p ro v id e tw e n ty fo u r h o u r a d a y , s e v e n d a y a w e e k c o v e ra g e . E a c h a m b u la n c e is c re w e d a t a ll tim e s b y a
te a m c o n s is tin g o f o n e p a ra m e d ic a n d o n e te c h n ic ia n , w o rk in g a tw e lv e h o u r s h ift. C le a rly
s u c h a c o v e ra g e s c h e m e re q u ire s c a re fu l a rra n g e m e n t to e n s u re th a t s e n s ib le ro s te rs a re p ro d u c e d fo r a ll o f th e in d iv id u a ls c o n c e rn e d . T h e re a re a n u m b e r o f c o n s tra in ts o n w h e n in d iv id u a ls a re a v a ila b le to c re w a v e h ic le .
G iv e n th e n a tu re o f th e d u tie s th a t c re w m e m b e rs u n d e rta k e it is n o t a p ra c tic a b le
o p tio n to a llo w d o u b le s h ift w o rk in g , fo r b o th th e s a fe ty o f u s e rs o f th e s e rv ic e a n d th e c re w s
th e m s e lv e s . T h is c o m p le m e n t o f s ta ff im p lie s th a t a c re w m e m b e r is re q u ire d to p e rfo rm a n
a v e ra g e o f th re e a n d a h a lf s h ifts p e r w e e k . In p ra c tic e th is tra n s la te s to e ith e r th re e o r fo u r
s h ifts in o n e s e v e n d a y p e rio d . E x p e rie n c e h a s s h o w n th a t re g u la rly w o rk in g e x c e s s s h ifts
in a w e e k to b e b o th p h y s ic a lly a n d e m o tio n a lly d ra in in g , a n d th is is to b e a v o id e d a s fa r a s
p o s s ib le .
In a d d itio n to th e p ra c tic a l c o n s tra in ts d e s c rib e d , m a n a g e m e n t h a v e d e c id e d a g a in s t
fo rm in g p e rm a n e n t p a irin g s . T h is m e a n s th a t e a c h p o s s ib le c o m b in a tio n o f p a ra m e d ic a n d
te c h n ic ia n m u s t w o rk to g e th e r a t s o m e s ta g e in th e s c h e d u le . T h is is s e e n to b rin g w ith it
S . C a g n o n i e t a l. ( E d s .) : E v o W o r k s h o p s 2 0 0 0 , L N C S 1 8 0 3 , p p . 2 6 7 − 2 7 9 , 2 0 0 0 .
S p rin g e r-V e rla g B e rlin H e id e lb e rg 2 0 0 0
2 6 8
P .V .G . B r a d b e e r , C . F in d la y , a n d T .C . F o g a r ty
th e b e n e fits o f p ro m o tin g ‘s k ill p u ll-th ro u g h ’ a n d e n s u rin g th a t th e re la tio n s h ip s b e tw e e n
p a ra m e d ic s a n d te c h n ic ia n s re m a in p ro d u c tiv e . In th e in te re s ts o f ‘fa irn e s s ’ it is d e s ira b le
th a t e a c h o f th e s ix te e n (fo u r p a ra m e d ic s a n d fo u r te c h n ic ia n s ) p o s s ib le te a m s a re a c tiv e a n
e q u a l n u m b e r o f tim e s .
A fu rth e r c o m p lic a tio n a ris e s fro m th e a m b u la n c e s ta ff th e m s e lv e s . M a n y o f th e m
h a v e fa m ilie s , a n d a re d e s iro u s o f s p e n d in g ‘q u a lity tim e ’ w ith th e m . T h is n o t o n ly p ro m o te s d e v e lo p m e n t o f th e fa m ily u n it, b u t a ls o a llo w s th e s ta ff to u n w in d a fte r s o m e o f th e
m o re h a rro w in g in c id e n ts th a t th e y a re re q u ire d to a tte n d . P o s s ib ly th e m o s t im p o rta n t o p p o rtu n ity is to s p e n d tim e w ith a fa m ily is d u rin g th e w e e k e n d , w h e n c h ild re n a re n o t a t
s c h o o l.
It is c le a rly n o t p o s s ib le to a v o id w e e k e n d w o rk , b u t th e s ta ff h a v e a s ta te d p re fe re n c e to e ith e r w o rk a c o m p le te w e e k e n d , o r to h a v e n o s h ifts in a w e e k e n d a t a ll. T h e in te rm e d ia te s itu a tio n w h e re o n ly o n e s h ift is w o rk e d in a w e e k e n d is re fe rre d to in h o u s e a s a
‘ru in e d w e e k e n d ’. T h e re is th u s a re q u e s t th a t s in g le s h ift w e e k e n d s a re a v o id e d b y a n y s y s te m u s e d to g e n e ra te a m b u la n c e s ta ffin g ro s te rs .
T h e re is a ls o th e n o tio n o f p e rc e iv e d ‘fa irn e s s ’ w ith in a ro s te r. T h is is b e s t v ie w e d
a s a d e s ire to h a v e a ll m e m b e rs o f s ta ff w o rk in g a n e q u a l n u m b e r o f d a y /n ig h t s h ifts a n d th e
s a m e n u m b e r o f w e e k e n d /w e e k d a y s h ifts . T h is p re v e n ts a n y a c c u s a tio n s o f b ia s w h e n s lig h t
im b a la n c e s a re n o tic e d in ro s te rs .
S ta ff le a v e re q u ire s n o s p e c ia l tre a tm e n t, a s s ta ff a re s e c o n d e d fro m th e c e n tra l a m b u la n c e b a s e in G le n ro th e s to fill in fo r s c h e d u le d h o lid a y s . T h is m a y c a u s e s lig h t im b a la n c e s in th e fre q u e n c y w ith w h ic h in d iv id u a ls a re p a rtn e re d , b u t th is is n o t re g a rd e d a s
im p o rta n t.
S e c tio n 2 - A p p r o a c h 1
A tte m p ts to p ro d u c e a ro s te r b y h a n d p ro v e d th a t th e p ro b le m w a s q u ite d iffic u lt, a n d
a c o m p u te ris e d s o lu tio n w a s s o u g h t. A t firs t g la n c e th e d e v e lo p m e n t o f a s o ftw a re s y s te m
to a u to m a tic a lly g e n e ra te u s e a b le a m b u la n c e ro s te rs w ith in th e c o n s tra in ts d e s c rib e d in s e c tio n 1 lo o k s a s th o u g h a s im p le d e p th firs t re c u rs iv e tre e s e a rc h w o u ld b e a fe a s ib le a p p ro a c h , a s th e c h o ic e o f th e ‘n e x t s h ift’ is s o h ig h ly c o n s tra in e d th a t c o n s id e ra b le p ru n in g
o f th e s e a rc h tre e w o u ld b e p o s s ib le . U n fo rtu n a te ly th e e ffe c t o f s o m e o f th e c o n s tra in ts c a n
o n ly b e a s s e s s e d w h e n th e tre e h a s b e e n g ro w n to s o m e d e p th . T h is h a s th e e ffe c t o f m a k in g
e x h a u s tiv e s e a rc h a n n o n -v ia b le o p tio n e v e n if v e ry fa s t c o m p u ta tio n a l m a c h in e ry w e re
a v a ila b le .
T h is le a d s u s in to th e a re a w h e re a s e a rc h b a s e d o n e v o lu tio n a ry te c h n iq u e s s u g g e s ts
its e lf a s a p o s s ib ility . E v e n fro m th e o u ts e t, th e c o n s tra in ts p la c e d o n th e s y s te m s u g g e s t th a t
fin d in g a s o lu tio n w ill p re s e n t a c o n s id e ra b le c h a lle n g e , b u t it w a s fe lt th a t g iv e n th e p a s t
s u c c e s s e s in s o lv in g s c h e d u le b a s e d p ro b le m s w ith g e n e tic a lg o rith m s (G A s ), in c lu d in g
th o s e d e s c rib e d in L a n g d o n [L a n g d o n 1 9 9 5 ] W re n a n d W re n [W re n 1 9 9 5 ] a n d F a n g e t. a l.
[F a n g 1 9 9 3 ], th e re w a s a re a s o n a b le c h a n c e o f s u c c e s s .
A n A m b u la n c e C re w R o s te rin g S y s te m
2 6 9
T h e a lp h a b e t in itia lly c h o s e n to re p re s e n t th e g e n e tic m a te ria l in th e c a n d id a te s o lu tio n s h a s s ix te e n c h a ra c te rs , o n e fo r e a c h o f th e d iffe re n t te c h n ic ia n p a ra m e d ic p a irin g s .
T h is w a s p a rtly to a llo w e a s y c h e c k in g to s e e if th e p a rtn e rs h ip c o n s tra in ts w e re b e in g m e t.
T h e s e le c tio n o f th e le n g th o f th e g e n e tic m a te ria l (c h ro m o s o m e le n g th ) is a ls o a n is s u e , a s
th e le n g th o f th e ro s te r is n o t s p e c ifie d . U n d e r s u c h c irc u m s ta n c e s a n a p p lic a tio n o f m e s s y
G A s [G o ld b e rg 1 9 9 0 ] m a y p ro v e p a rtic u la rly a p p ro p ria te , b u t fo r a n in itia l in v e s tig a tio n a
fix e d le n g th re p re s e n ta tio n w a s c h o s e n .
A s th e re a re s ix te e n te a m s to a c c o m m o d a te th e fix e d le n g th w a s s e t a t s ix te e n w e e k s ,
g iv in g a p a tte rn th a t re p e a ts th re e tim e s a y e a r. E a c h o f th e s e s ix te e n w e e k c y c le s w o u ld
id e a lly h a v e e a c h p a rtn e rs h ip a c tiv e fo u rte e n tim e s (s e v e n o n d a y s h ift a n d s e v e n o n n ig h ts ).
A y e a r c o u ld th u s c o m p ris e th re e o f th e s e c y c le s , le a v in g a g a p o f a b o u t fo u r w e e k s . T h is
is c o n v e n ie n t fo r th e d is ru p tio n in p a tte rn re q u ire d fo r th e C h ris tm a s /N e w Y e a r p e rio d w h e n
th e w o rk in g re q u ire m e n ts a re s u b je c t to c h a n g e . T h e fe s tiv e p e rio d is a b u s y tim e fo r th e
e m e rg e n c y s e rv ic e s .
N o te th a t if w o rk in g to th e s e s ix te e n w e e k ro s te rs th a t th e ‘w ra p a ro u n d ’ e ffe c t fro m
th e e n d o f o n e c y c le to th e b e g in n in g o f th e n e x t m u s t b e ta k e n in to a c c o u n t. F a ilu re to d o
s o m a y re s u lt in d o u b le s h ift w o rk in g b e tw e e n th e e n d o f a c y c le , a n d th e b e g in n in g o f th e
n e x t.
T h e firs t in c a rn a tio n o f g e n e tic s e a rc h w a s la rg e ly to g a in in s ig h t in to th e c h a ra c te ris tic s o f th e s p a c e , a n d u s e d a fa irly tra d itio n a l a p p ro a c h , in th a t e a c h lo c u s o n th e c h ro m o s o m e w a s a llo w e d to ta k e a n y o f th e a llo w a b le a lle le s , a lo w m u ta tio n ra te , u n ifo rm
c ro s s o v e r [S y s w e rd a 1 9 8 9 ] a n d b in a ry to u rn a m e n t s e le c tio n [B rin d le 1 9 8 1 ] (w ith 1 0 0 %
c h a n c e o f th e b e tte r o f th e tw o c a n d id a te s p ro g re s s in g ).
F o llo w in g th e s u c c e s s re p o rte d b y M ille r e t. a l. [M ille r1 9 9 5 ], J o n e s a n d B ra d b e e r
[J o n e s 1 9 9 4 ] a n d C h is h o lm a n d B ra d b e e r [C h is h o lm 1 9 9 7 ] u s in g s m a ll b re e d in g p o o l s iz e s ,
a re la tiv e ly s m a ll b re e d in g p o o l o f th irty w a s in itia lly a d o p te d . A ra n d o m s a m p le o f fiv e
th o u s a n d in d iv id u a ls g e n e ra te d th e d is trib u tio n s h o w n a s fig u re 1 . T h is s u g g e s ts th a t a la rg e
p o rtio n o f th e p o p u la tio n h a v e p o o r fitn e s s , a n d le a v e s u s h o p in g th a t th e re a re v e ry lo n g (if
th in ) ta ils to th e d is trib u tio n . T h is ty p e o f d is trib u tio n , a llie d w ith th e fa c t th a t th e re a re
m a n y c o n s tra in ts o n th e s e a rc h s p a c e re in fo rc e s th e s u s p ic io n th a t th is w ill b e a d iffic u lt
p ro b le m fo r th is b ra n d o f G A . It is w o rth c o m m e n tin g th a t th e re w a s little c o d e p ro d u c e d
fo r th is in itia l s y s te m th a t w a s n o t to b e o f u s e in s u b s e q u e n t im p le m e n ta tio n s .
T h e fitn e s s fu n c tio n u s e d fo r e v a lu a tio n p u rp o s e s is b a s e d o n th e a c c ru a l o f p e n a ltie s , w ith d iffe re n t w e ig h ts b e in g a s s ig n e d to b re a c h e s o f d iffe re n t c o n s tra in ts . E a c h tim e a
c o n s tra in t is v io la te d th e in te g e r re p re s e n tin g th e fitn e s s o f th e c a n d id a te s o lu tio n is in c re a s e d . T h is m e a n s th a t in th e re p o rtin g o f re s u lts a lo w fitn e s s n u m b e r in d ic a te s a a g o o d
s o lu tio n .
2 7 0
P .V .G . B r a d b e e r , C . F in d la y , a n d T .C . F o g a r ty
Initially the following penalties were allocated
C o n s tra in t b re a c h
P e n a lty le v e l
D a y /n ig h t im b a la n c e
4 8 h o u r+ p e r se v e n d a y s
1
D o u b le S h ift
1
1
R u in e d W e e k e n d s
1
Table 1: Initial Penalty Scheme
T h is a p p o rtio n m e n t o f p e n a ltie s im p lie s th a t a ll b re a c h e s o f c o n s tra in ts a re e q u a l,
w h ic h is c le a rly a n o v e rs im p lific a tio n o f th e p ro b le m . In re a lity , a s ix te e n w e e k ro s te r th a t
p ro d u c e d o n e o r tw o b re a c h e s o f th e 4 8 h o u r ru le w o u ld p ro b a b ly b e a c c e p ta b le , w h e re a s a
s c h e m e w ith th e s a m e n u m b e r o f d o u b le s h ifts w o u ld n o t b e a c c e p ta b le . T h o s e c o n s tra in ts
th a t m u s t n o t b e b re a c h e d a re u s u a lly re fe rre d to a s h a r d c o n s tr a in ts , w h ile th o s e fo r
w h ic h m in o r o r lim ite d b re a c h e s c o u ld b e to le ra te d a re re fe rre d to a s s o ft c o n s tr a in ts .
U s u a lly a h a rd c o n s tra in t v io la tio n w o u ld a ttra c t a h ig h e r p e n a lty v a lu e th a n a s o ft p e n a lty .
M ic h c h a le w ic z [M ic h c h a le w ic z 1 9 9 4 ] p ro v id e s a d is c u s s io n o f s o m e o f th e s o ft c o n s tra in ts th a t c a n a p p e a r in tim e ta b lin g p ro b le m s , a n d th e c o n tra s tin g h a rd c o n s tra in ts .
T h e a p p o rtio n m e n t o f re la tiv e p e n a lty le v e ls is a n a re a o f in te re s t in its o w n rig h t,
a n d c a n e ith e r b e s ta tic , o r d y n a m ic a s c a n b e s e e n fo r e x a m p le in E ib e n e t. a l. [E ib e n 1 9 9 8 ]
w h e re d iffe re n t a p p ro a c h e s a re c o m p a re d .
A s th is is
o f th e s y s te m , a n
p e n a lty re s u lt.
F o r m o re
m o re a tte n tio n ,
[P a e c h te r1 9 9 8 ] w
a n in itia l s tu d y h o w e v e r w e a re m o re in te re s te d in th e g ro s s p e rfo rm a n c e
d m e re ly n o te th a t a n id e a l s o lu tio n to th is p ro b le m w o u ld p ro d u c e a z e ro
d e ta ile d s tu d y , th e e x a c t w e ig h t g iv e n to e a c h c la s s o f b re a c h w o u ld n e e d
p o s s ib ly u s in g a n a p p ro a c h s u c h a s th a t d e ta ile d b y P a e c h te r e t. a l.
h e re a ‘fro n t p a n e l’ is a tta c h e d to th e s y s te m a llo w in g d y n a m ic a lte ra tio n
A n A m b u la n c e C re w R o s te rin g S y s te m
2 7 1
o f th e w e ig h ts . It is c e rta in ly p o s s ib le th a t d iffe re n t p e n a lty le v e ls c o u ld c h a n g e th e tra je c to ry o f th e s e a rc h .
Fitness Distribution (30000 samples)
4000
3500
Frequency
3000
2500
2000
1500
1000
500
580
560
540
520
500
480
460
440
420
400
380
360
340
320
300
280
0
Fitness
F ig u re 1 : D is trib u tio n o f fitn e s s fo r ra n d o m
T h is s a
a b o u t fo u r h u n
th is v ic in ity . It
tw e n ty c o n s tra
m p le in
d re d a n
is d iffic
in ts , s o
d ic a te s th a t th e
d fo rty c o n s tra in
u lt to b e lie v e th a
th a t le a v e s a c o n
s a m p le u s in g ‘u n c o n s tra in e d ’ re p re s e n ta tio n .
a v e ra g e
t re q u ire m
t a n a c c e p
s id e ra b le
ra n d o m ly
e n ts , a n d
ta b le s o lu
c h a lle n g e
g
th
tio
fo
e n e
e b
n w
r th
ra te
u lk
ill b
e se
d in d iv id
o f th e p o
re a c h m o
a rc h m e c
u a
p u
re
h a
l b re a c h e s
la tio n is in
th a n a b o u t
n is m .
In o rd e r to p ro g re s s th e s e a rc h , s o m e s o rt o f m o v e o p e ra to r m u s t b e e m p lo y e d . It is
u s e fu l to h a v e a n in d ic a tio n o f th e lik e ly e ffe c tiv e n e s s o f o p e ra to rs . R a n d o m m u ta tio n is o n e
o f th e o p e ra to rs tra d itio n a lly u s e d , a n d th e e ffe c t o f th is o p e ra to r in th is re p re s e n ta tio n in
th is p ro b le m is s u m m a ris e d in fig u re 2 . M u ta tio n in th is c a s e in v o lv e s th e s e le c tio n o f a s in g le lo c u s o n th e c h ro m o s o m e , a n d re p la c in g it w ith a s in g le ra n d o m ly g e n e ra te d le g a l v a lu e .
T h e fig u re is g e n e ra te d fro m a s a m p le o f 3 0 0 0 0 in d iv id u a ls , e a c h o f w h ic h re c e iv e s o n e m u ta tio n . T h e re s u ltin g c h a n g e in fitn e s s d u e to th e m u ta tio n is n o te d a g a in s t th e fitn e s s o f th e
o rig in a l. R a th e r th a n k e e p re c o rd s o f in d iv id u a l fitn e s s n u m b e rs , th e o u tc o m e s a re g ro u p e d
in to ra n g e s , a n d a v e ra g e d .
T h e g ra p h s h o w s th e p h e n o m e n o n o f a b o v e a v e ra g e fitn e s s in d iv id u a ls g e n e ra lly b e in g a d v e rs e ly a ffe c te d b y ra n d o m m u ta tio n s , w h e re a s b e lo w a v e ra g e fitn e s s in d iv id u a ls a re
o n a v e ra g e im p ro v e d . T h e k in k s a t th e e x tre m itie s c a n b e e x p la in e d a s d u e to s m a ll s a m p le
s iz e s a t th e ta ils o f th e d is trib u tio n .
2 7 2
P .V .G . B r a d b e e r , C . F in d la y , a n d T .C . F o g a r ty
N o s te p s w e re ta k e n to p re v e n t th e m u ta tio n re p lic a tin g th e o rig in a l v a lu e . A s a n a lp h a b e t o f s ix te e n w a s u s e d , a p p ro x im a te ly o n e in s ix te e n m u ta tio n s re s u lt in n o c h a n g e .
T h is m a y b e v ie w e d a s re d u c in g th e s lo p e o f th e g ra p h s lig h tly .
A g r
in d u c e s s e v
A se
ly p o o r re s u
a p h
e re
rie s
lts
w ith a g ra
e p is ta tic e
o f ru n s (w
to c o n firm
d ie
ffe
ith
th
n ts lo w a
c ts , o r th
a ra n g e
a t a b e tte
s th is
a t th e
o f p a r
r re p r
c a n
p ro
a m
e s e
b e ta
b le m
e te r s
n ta tio
k e n
sp a
e ts )
n sh
a s in d ic
c e h a s n
u s in g th
o u ld b e
a tin g th a t th e re p re s e n ta tio n
o s tru c tu re .
is a p p ro a c h g a v e s u ffic ie n ts o u g h t.
Average Change due to Mutation
7
6
5
Change
4
3
2
1
560
540
520
500
480
460
440
420
400
380
360
340
-1
320
0
-2
Original fitness
F ig u r e 2 : A v e r a g e c h a n g e in fitn e s s u n d e r m u ta tio n .( 3 0 0 0 0 s a m p le s )
S e c tio n 3 - A p p r o a c h 2
In s tru m e n ta tio n o f th e
s o u rc e o f p e n a lty p o in ts w a s d
n u m b e r o f tim e s e a c h te a m w a
s e n ta tio n . A s a w a y o f re d u c in
e m p lo y e d , w ith e a c h te a m a p p
T h is c h a n g e la rg e ly fitte d in to
c o d e c h a n g e s. A s b e fo re , a sa m
c e rta in if a n y b e n e fit a c c ru e d .
ru
u
s
g
n s o f th e firs t a p p ro a c h s e e m e d to in d ic a te th a t th e
e to d iffic u lty in m a in ta in in g th e e v e n s p re a d b e tw e
ro s te re d . C le a rly , th e n e x t s te p w a s to re c o n s id e r th e
th e im b a la n c e p e n a ltie s a p e rm u ta tio n re p re s e n ta tio
e a rin g a g iv e n n u m b e r o f tim e s w ith in a s ix te e n w e e k
th e fra m e w o rk o f th e p re v io u s s y s te m , w ith re la tiv e ly
p le o f ra n d o m ly g e n e ra te d in d iv id u a ls w a s e v a lu a te d
T h e re s u lts a re n o te d in fig u re 3 .
m a jo r
e n th e
re p re n w a s
ro s te r.
s m a ll
to a s -
A n A m b u la n c e C re w R o s te rin g S y s te m
Permutation vs
Unconstrained representation s
2 7 3
Permutation
Unconstrained
4000
3500
Frequency
3000
2500
2000
1500
1000
500
570
550
530
510
490
470
450
430
410
390
370
350
330
310
290
270
0
Fitness
F ig u re 3 : D is trib u tio n o f fitn e s s fo r ra n d o m s a m p le u s in g s in g le p e rm u ta tio n c h o m o s o m e
c o m p a re d to u n c o n s tra in e d c h ro m o s o m e (3 0 0 0 0 s a m p le s ).
C o m p a rin g fig u re 3 w ith fig u re 1 w e s e e th a t th e d is trib u tio n h a s m o v e d fa r e n o u g h
to p ro v id e e n c o u ra g e m e n t to c o d e th e re s t o f th e s y s te m to p e rfo rm fu rth e r te s t ru n s .
Effect of mutation on permutation representation
6
4
2
-4
-6
-8
-10
-12
-14
Fitness
570
550
530
510
490
470
450
430
410
390
370
350
330
310
Change
0
-2
2 7 4
P .V .G . B r a d b e e r , C . F in d la y , a n d T .C . F o g a r ty
F ig u re 4 : A v e ra g e c h a n g e in fitn e s s to p e rm u ta tio n re p re s e n ta tio n u n d e r m u ta tio n .(3 0 0 0 0
s a m p le s )
O n c e a g a in p e rfo rm in g th e c h a n g e u n d e r m
th e re s p o n s e is v e ry lo w , a g a in le a d in g u s to s u s p e c
tio n le ft in th e re p re s e n ta tio n .
U s in g P M X [G o ld b e rg 1 9 8 5 ] a s th e c ro s s o v
p e rfo rm e d . V a rio u s p a ra m e te r s e ts fa ile d to p ro d u c e
re p re s e n ta tio n o f th e c a n d id a te s o lu tio n s p e rm itte d
It is b e lie v e d th a t th is in tu rn le d to a h ig h ly m o d a l
th e s e a rc h a lg o rith m .
u ta tio n te s t, w e s e e th a t th e g ra d ie n t o f
t th a t th e re is a la rg e a m o u n t o f in te ra c e r
e v
fa r
se a
m e th o
e n n e a
to o m
rc h sp
d , a
r a c
a n y
a c e
n u m b
c e p ta b
c o n s tr
th a t c a
e r o f
le re s
a in ts
u se d
te s t ru
u lts . A
to b e v
d iffic u
n s w e r
g a in th
io la te d
ltie s fo
e
e
.
r
S e c tio n 4 - A p p r o a c h 3
T h e re p re s e n ta tio n d e s c rib e d in th e p re v io u s s e c tio n h a d th e e ffe c t o f fo rc in g a ll c a n d id a te s o lu tio n s to o b e y th e c o n s tra in t re q u irin g a ll te a m p a irin g s to b e e q u a lly re p re s e n te d ,
b u t th e ‘e q u ita b ility ’ re q u e s t is le ft to th e e v o lu tio n a ry m e c h a n is m . In o rd e r to re d u c e th e
n u m b e r o f u n d e s ira b le c o m b in a tio n s fu rth e r, a th ird e n c o d in g w a s d e v is e d . T h is w a s m a d e
u p fro m fo u r s e p a ra te c o m p o n e n ts (c h ro m o s o m e s ), e a c h e n c o d in g a p e rm u ta tio n o f te a m s .
T h e firs t tw o , e a c h o f le n g th s ix te e n , d e te rm in e w h ic h w e e k e n d th e te a m w ill w o rk a d a y
s h ift a n d w h ic h w e e k e n d th e y w o rk a n ig h t s h ift. If a te a m is s c h e d u le d to w o rk b o th d a y
a n d n ig h t s h ift th e n th is w ill b e d e te c te d a n d a c c ru e th e a s s o c ia te d p e n a lty . T h is a p p ro a c h
g u a ra n te e s th a t n o t o n ly d o e s e a c h te a m w o rk a n e q u a l n u m b e r o f w e e k e n d d a y a n d n ig h t
s h ifts , b u t a ls o th a t th e re q u ire m e n t to a v o id ru in e d w e e k e n d s is a v o id e d . T h is re n d e rs h a lf
o f th e fitn e s s fu n c tio n re d u n d a n t, th u s s p e e d in g th e e v a lu a tio n p o rtio n o f th e s y s te m b y
a b o u t 2 5 % . S im ila rly th e th ird a n d fo u rth c h ro m o s o m e s , e a c h o f le n g th e ig h ty e n c o d e fiv e
w e e k d a y s h ift a n d fiv e w e e k n ig h t s h ift a p p e a ra n c e s fo r e a c h te a m .
F ig u re 5 c o m p
o f th is a d m itte d ly s m
th e m u ta tio n o p e ra tio
s u lts fro m th e p re v io u
a re s
a ll s
n . E
s tw
th e fitn e s
a m p le s e e
x a m in a tio
o re p re se n
s d is
m s o
n o f
ta tio
trib u tio n o f th is re p re s e n ta tio n
n ly a little b e tte r. F ig u re 6 e x
th e s lo p e o f th is g ra p h in c o m
n s re v e a ls th a t it is m o re p ro n o
, w
a m
p a
u n
h ic
in e
ris o
c e d
h o n th e b
s th e e ffe c
n w ith th e
. T h is is ta
a s is
t o f
re k e n
A n A m b u la n c e C re w R o s te rin g S y s te m
2 7 5
4000
3500
3000
2500
2000
1500
1000
500
0
Permutation
Unconstrained
570
540
510
480
450
420
390
360
330
300
4 Chromosome
270
Frequency
a s in d ic a tio n th a t th e re is le s s e p is ta s is e v id e n t in th e re p re s e n ta tio n . E v e n s o th e s lo p e is
s till re la tiv e ly s h a llo w , in d ic a tin g th a t th e p ro b le m is s till ‘h a rd ’.
Fitness
F ig u re 5 : D is trib u tio n o f fitn e s s fo r ra n d o m s a m p le u s in g 4 c h ro m o s o m e a rra n g e m e n t, c o m p a re d to s in g le p e rm u ta tio n c h o m o s o m e a n d u n c o n s tra in e d c h ro m o s o m e (3 0 0 0 0 s a m p le s ).
Effect of mutation on 4 Chomosome permutation representation
10
530
510
490
470
450
430
410
390
370
350
330
310
-10
290
0
270
Average change
20
-20
-30
Fitness
F ig u re 6 : A v e ra g e c h a n g e in fitn e s s to 4 C h ro m o s o m e p e rm u ta tio n re p re s e n ta tio n u n d e r
m u ta tio n .( 3 0 0 0 0 s a m p le s )
P re lim in a ry ru n s w ith th
o rd e r o f te n c o n s tra in t b re a c h e
v e a le d th a t a b o u t th re e o f th e s e
th u s n o t a c c e p ta b le . R a th e r th a n
o c c u rre n c e o f d o u b le s h ifts , it w
th e v a rio u s b re a c h e s . T h e v a lu e
th a t th is w o u ld d riv e th e s e a rc h
n e s s d is trib u tio n v a lu e s a lre a d y
is re p re s e n ta tio n g a v e a s e rie s o f re s u lts w ith fitn e s s in th e
s . C lo s e r e x a m in a tio n o f th e re s u ltin g ro s te r h o w e v e r re w e re b re a c h e s o f th e ‘n o d o u b le s h ift’ c o n s tra in t, a n d w e re
try to re c o d e th e re p re s e n ta tio n fu rth e r to to ta lly a v o id th e
a s d e c id e d to re b a la n c e th e w e ig h t o f p e n a lty a c c ru e d b y
fo r d o u b le s h ift v io la tio n s w a s in c re a s e d to 2 5 , in th e h o p e
in to e lim in a tin g d o u b le s h ifts . T h is o f c o u rs e a lte rs th e fitp ro d u c e d , b u t a s w e w e re n o w n e a rin g th e id e a l s o lu tio n s
2 7 6
P .V .G . B r a d b e e r , C . F in d la y , a n d T .C . F o g a r ty
th is w a s fe lt to b e a s m a ll p ric e to p a y . It is a ls o p o s s ib le o r e v e n lik e ly th a t th e s e c h a n g e s
w ill c h a n g e th e s e a rc h s p a c e s lig h tly , w ith a re s u ltin g c h a n g e in p ro g re s s
S e c tio n 5 - R e s u lts
A fte r m a k in g th
tio n , th e s y s te m w a s ru
2 0 , P M X , b in a ry to u rn
o n ly 5 m in o r b re a c h e s
tio n to th e p ro b le m , a n
w ith p a ra m e te rs , s o fa r
fitn e s s v a lu e .
e c h a n g e
n w ith v a
a m e n t se
o f th e 4 8
d is s h o w
n o o th e r
d e s c rib e d to p e n a lty w
rio u s p a ra m e te rs . O n e
le c tio n a n d a 1 0 % m u
h o u rs p e r w e e k ru le . T
n in fig u re 7 . D e s p ite
s o lu tio n s h a v e b e e n g e
e ig h tin g s g iv e n in th
o f th e firs t ru n s , u s in
ta tio n ra te p ro d u c e d
h is w a s fe lt to b e a p
a c e rta in a m o u n t o f e
n e ra te d th a t h a v e s u c
e p re v io u s s e c g a p o o l s iz e o f
a s o lu tio n w ith
ra c tic a b le s o lu x p e rim e n ta tio n
h a n a c c e p ta b le
F ig u re 7 : S c re e n d u m p o f b e s t re s u lt s o fa r.
S e c t io n 6 - C o n c lu s io n s
T h e s y s te m , d e v e lo p e d a g a in s t a p ra c tic a l n e e d a n d re la tiv e ly u n d e v e lo p e d a s it is
h a s p ro d u c e d a w o rk a b le ro s te r fo r a h ig h ly c o n s tra in e d p ro b le m .
A n A m b u la n c e C re w R o s te rin g S y s te m
2 7 7
A s th e re a re a n u m b e r o f d iffe re n t lo c a l o p tim a b e in g fo u n d , it is p o s s ib le to g iv e th e
‘c lie n t’ a n u m b e r o f d iffe re n t c a n d id a te s o lu tio n s . T h is w ill a llo w c o n s id e ra tio n o f o th e r
‘s o c ia l’ fa c to rs n o t e x p lic itly m e n tio n e d in th e in itia l o u tlin e o f th e p ro b le m .
T h is a p p ro a c h o f lo o k in g a t th e b e h a v io u r o f th e s e a rc h s p a c e h a s s a v e d a c o n s id e ra b le a m o u n t o f tim e in ru n n in g e x p e rim e n ts a n d a n a ly s in g th e re s u lts fro m th e m . W h ile th e
re la tiv e s h o rtc o m in g s o f th e firs t tw o re p re s e n ta tio n s a re to a n e x te n t p re d ic ta b le , g iv e n th e
n a tu re o f th e c o n s tra in ts , it is u s e fu l to b e a b le to c o n firm th is w ith o u t e x p e n d in g la rg e
a m o u n ts o f c o m p u te c y c le s .
S e c tio n 7 - F u r th e r W o r k
T h is d o c u m e n t d e s c rib e s th e firs t s te p s to w a rd s s o lv in g th is p ro b le m . T h e
a re a w a re th a t m a n y v a ria n ts a re p o s s ib le , a n d e v e n d e s ira b le . A s m e n tio n e d in th e
th e te x t, th e le n g th o f th e c y c le is n o t c le a r, s o a n a p p ro a c h a llo w in g a v a ria b le le n g
is o n e p o s s ib le d ire c tio n fo r fu rth e r s tu d y .
It w o u ld a ls o b e o f in te re s t to d is c o v e r m o re a b o u t th e m o d a lity o f th e s e a rc
a n d it is p la n n e d to u s e a re v e rs e h ill-c lim b in g te c h n iq u e , s u c h a s d e s c rib e d b
[Jo n e s1 9 9 5 ].
P ro d u c in g m o re in fo rm a tio n o n th e s e n s itiv ity o f th e p ro b le m to d iffe re n t
te rs , s u c h a s p o o l s iz e a n d c ro s s o v e r m e c h a n is m re m a in s a n ite m o n th e a g e n d a .
It h a s b e e n n o te d th a t th e a d d itio n o f h e u ris tic s c o u ld b e u s e d to im p ro v e p
a n c e , p o s s ib ly in a s im ila r fa s h io n to th a t re p o rte d b y H a rt e t. a l. [H a rt1 9 9 8 ].
a u th o rs
b o d y o f
th c y c le
h sp a c e ,
y J o n e s
p a ra m e e rfo rm -
A c k n o w le d g e m e n ts
T h e c o n s tru c tiv e c o m m e n ts o f th e a n o n y m o u s re fe re e s a re n o te d a n d a p p re c ia te d .
T h a n k s a ls o to C o lin W ils o n fo r h is h e lp .
B ib lio g r a p h y
[B rin d le 1 9 8 1 ]. B rin d le , A , “ G e n e tic a lg o rith m s fo r fu n c tio n o p tim iz a tio n ” , D o c to ra l D is s e rta tio n a n d T e c h n ic a l R e p o rt T R 8 1 -2 , D e p a rtm e n t o f C o m p u te r S c ie n c e , U n iv e rs ity o f
A lb e rta , E d m o n to n , 1 9 8 1 .
[ C h is h o lm 1 9 9 7 ] . C h is h o lm K .J . a n d B r a d b e e r P .V .G . “ U s in g a G e n e tic A lg o r ith m to O p tim is e a D ra u g h ts P ro g ra m B o a rd E v a lu a tio n F u n c tio n ” , P ro c e e d in g s o f IE E E IC E C ’9 7 , In d ia n a p o lis , 1 9 9 7 .
[E ib e n 1 9 9 8 a ] E ib e n A .E ., B a c k T ,. S c h o e n a u e r M , a n d S c h w e fe l (e d s .) “ P ro c e e d in g s o f
P a ra lle l P ro b le m S o lv in g F ro m N a tu re - P P S N V ” , L N C S 1 4 9 8 , S p rin g e r V e rla g
2 7 8
P .V .G . B r a d b e e r , C . F in d la y , a n d T .C . F o g a r ty
[ E ib e n 1 9 9 8 b ] E ib e n A .E ., v a n H e m e r t J .I ., M a r c h io r i E . a n d S te e n b e e k A .G , “ S o lv in g B in a ry C o n s tra in t S a tis fa c tio n P ro b le m s U s in g E v o lu tio n a ry A lg o rith m s w ith a n A d a p tiv e
F itn e s s F u n c tio n ” , in [E ib e n 1 9 9 8 a ]
[F a n g 1 9 9 3 ]. F a n g H -L , R o s s P . a n d C o rn e D . “ A p ro m is in g G e n e tic A lg o rith m a p p ro a c h to
jo b -s h o p S c h e d u lin g , re s c h e d u lin g a n d o p e n -s h o p s c h e d u lin g p ro b le m s ” , in [F o rre s t1 9 9 3 ].
[ F o r r e s t1 9 9 3 ] . F o r r e s t S ., ( e d .) “ P ro c e e d in g s o f th e F if th I n te rn a tio n a l C o n f e r e n c e o n G e n e tic A lg o rith m s ” , M o rg a n K a u fm a n n , S a n M a te o , 1 9 9 3 .
[ F o g a r ty 1 9 9 5 ] . F o g a r ty T .C ( e d .) “ P r o c e e d in g s o f th e A I S B W o rk s h o p o n E v o lu tio n a r y
C o m p u tin g ” , S h e ffie ld , L N C S 9 9 3 , S p rin g e r V e rla g , B e rlin 1 9 9 5 .
[ G o ld b e r g 1 9 9 0 ] . G o ld b e r g D .E . “ M e s s y G e n e tic A lg o r ith m s : M o tiv a tio n , A n a ly s is a n d
F irs t R e s u lts ” , C o m p le x S y s te m s , V o l. 3 .
[G o ld b e rg 1 9 8 5 ]. G o ld b e rg D .E . a n d L in g le , R ., “ A lle le s , lo c i, a n d th e tra v e llin g s a le s m a n
p ro b le m ” , in [G re fe n s te tte 1 9 8 5 ].
[ G r e f e n s te tte 1 9 8 5 ] . G r e f e n s te tte J .J . ( e d .) “ P r o c e e d in g s o f th e F ir s t I n te r n a tio n a l C o n f e r e n c e o n G e n e tic A lg o rith m s ” , L a w re n c e E a rlb a u m , H ills d a le , 1 9 8 5
[ H a r t1 9 9 8 ] H a r t E ., N e ls o n J . a n d R o s s P ., “ S o lv in g a R e a l- W o r ld P r o b le m U s in g a n E v o lv in g H e u ris tic a lly D riv e n S c h e d u le B u ild e r” , E v o lu tio n a ry C o m p u ta tio n V 6 (1 ):6 1 -8 0 .
[J o n e s 1 9 9 5 ]. J o n e s T ., “ E v o lu tio n a ry A lg o rith m s , F itn e s s L a n d s c a p e s a n d S e a rc h ” , P h D
D is s e rta tio n , U n iv e rs ity o f N e w M e x ic o , 1 9 9 5 .
[ J o n e s 1 9 9 4 ] . J o n e s P .A . a n d B r a d b e e r P .V .G ., “ D is c o v e r y o f o p tim a l w e ig h ts in a c o n c e p t
s e le c tio n s y s te m ” , in [L e o n 1 9 9 4 ].
[ L a n g d o n 1 9 9 5 ] . L a n g d o n W .B . “ S c h e d u lin g P la n n e d M a in te n a n c e o f th e N a tio n a l G r id ” ,
in [F o g a rty 1 9 9 5 ].
[ L e o n 1 9 9 4 ] . L e o n R ., ( e d .) “ P r o c e e d in g s o f th e 1 6 th R e s e a rc h C o llo q u iu m
fo rm a tio n R e trie v a l S p e c ia lis t G ro u p ” , T a y lo r G ra h a m , 1 9 9 4
o f th e B C S In -
[M ic h c h a le w ic z 1 9 9 4 ] M ic h c h a le w ic z Z ., “ G e n e tic A lg o rith m s + D a ta S tru c tu re s = E v o lu tio n P ro g ra m s ” , (2 n d E d itio n ) p 2 5 6 , S p rin g e r V e rla g , 1 9 9 4 .
A n A m b u la n c e C re w R o s te rin g S y s te m
2 7 9
[ M ille r 1 9 9 5 ] . M ille r J .F ., T h o m s o n P . a n d B r a d b e e r P .V .G . “ T e rn a r y D e c is io n D ia g r a m
O p tim iz a tio n o f R e e d -M u lle r L o g ic F u n c tio n s u s in g a G e n e tic A lg o rith m fo r V a ria b le a n d
S im p lific a tio n R u le O rd e rin g ” , in [F o g a rty 1 9 9 5 ].
[ P a e c h te r 1 9 9 8 ]. P a e c h te r B ., R a n k in R .C ., C u m m in g A . a n d F o g a r ty T .C ., “ T im e ta b lin g th e
C la s s e s o f a n E n tire U n iv e rs ity w ith a n E v o lu tio n a ry A lg o rith m .” in [E ib e n 1 9 9 8 a ]
[ S c h a f f e r 1 9 9 3 ] . S c h a f f e r J .D . “ P r o c e e d in g s o f th e T h ir d I n te r n a tio n a l C o n f e r e n c e o n G e n e tic A lg o rith m s ” , M o rg a n K a u fm a n n , S a n M a te o , 1 9 9 3 .
[S y sw e rd a 1 9 8 9 ].
[S c h a ffe r1 9 9 3 ]
S y sw e rd a
G .
“ U n ifo rm
C ro s so v e r
in
[ W r e n 1 9 9 5 ] . W r e n A . a n d W r e n D . O ., “ A G e n e tic a lg o r ith m
s c h e d u lin g ” , C o m p u te rs in O p e ra tio n s R e s e a rc h , 2 2 (1 ), 1 9 9 5 .
G e n e tic
A lg o rith m s ” ,
in
fo r p u b lic tra n s p o rt d riv e r
A Systematic Investigation of GA Performance
on Jobshop Scheduling Problems
Emma Hart, Peter Ross
Division of Informatics, University of Edinburgh,
Edinburgh EH1 2QL, Scotland
femmah,peterg@dai.ed.ac.uk
Abstract. Although there has been a wealth of work reported in the
literature on the application of genetic algorithms GAs to jobshop
scheduling problems, much of it contains some gross over-generalisations,
i.e that the observed performance of a GA on a small set of problems
can be extrapolated to whole classes of other problems. In this work we
present part of an ongoing investigation that aims to explore in depth
the performance of one GA across a whole range of classes of jobshop
scheduling problems, in order to try and characterise the strengths and
weaknesses of the GA approach. To do this, we have designed a con gurable problem generator which can generate problems of tunable di culty,
with a number of di erent features. We conclude that the GA tested is
relatively robust over wide range of problems, in that it nds a reasonable
solution to most of the problems most of the time, and is capable of nding the optimum solutions when run 3 or 4 times. This is promising for
many real world scheduling applications, in which a reasonable solution
that can be quickly produced is all that is required. The investigation
also throws up some interesting trends in problem di culty, worthy of
further investigation.
1 Introduction
Since the rst applications of GAs to scheduling problems, 4, there have been
many reported applications of GAs to scheduling problems in general, with the
jobshop problem receiving a great deal of attention, for example 2, 8, 9, and
more recently 3, 13. A large variety of representations and operators have been
reported, each showing impressive performance of a GA on some small subset of
benchmark problems. Often this performance is then extrapolated to claim that
the GA in question is a good algorithm for solving job-shop scheduling problems.
However, in order to properly evaluate the quality of a method, it is important
to show that it works over a wide range of problems. This requires testing the
method over an extremely large number of problems the problems should be
chosen such that they exhibit a variety of features, and should vary in diculty in
some congurable way. Rather than use benchmarks problems, it is more useful
to use a parameterised problem generator which can generate problem instances
at random, and in some tunable manner, so that many dierent instances of
problem classes can be generated.
S . C a g n o n i e t a l. ( E d s .) : E v o W o r k s h o p s 2 0 0 0 , L N C S 1 8 0 3 , p p . 2 7 7 − 2 8 6 , 2 0 0 0 .
S p rin g e r-V e rla g B e rlin H e id e lb e rg 2 0 0 0
2 7 8
E . H a rt a n d P . R o ss
Therefore, in this article we describe a problem generator which we have used
to generate instances of jobshop scheduling problems falling into 280 dierent
problem classes. We present the preliminary results from an ongoing investigation, which allows us to come to some general conclusions about the robustness
of a GA as a technique for solving scheduling problems. We evaluate our results
not only in terms of the capability of the GA to nd the optimum solution to
the problems we generate which all have a known optimum value , but also
in terms of its accuracy at producing 'reasonable' quality solutions on a reliable basis. For real-world problems, it is seldom necessary to produce optimum
solutions to problems | good enough fast enough" will generally suce. Therefore, it is important to take this into account when making statements about
the performance of the GA.
We rst describe the features of the problem generator, then present a summary of the results from 28000 experiments on a range of scheduling problems.
Some general conclusions about the performance of the GA are then drawn.
2 The Problem Generator
We designed a parameterised, tunable problem generator which generates a
solution to a jobshop scheduling problem, i.e. a gantt chart, characterised by
P O M D I S , where O is the number of operations to be scheduled, M is
the number of machines, D de nes the distribution that the operation sizes are
drawn from, either Gaussian or uniform , I is the total amount of idle time per
machine, and S is the amount of slack in the arrival and due dates of each job.
The parameters are described in more detail below:
OperationsJobsMachines The operations, O, are divided equally between the
number of machines, m. A schedule is generated at random, by placing operations
into the gantt chart. The operations are then assigned to jobs, j , such that each
job is processed only once on each machine, and an operation of a job on a
machine cannot begin until its operation on the previous machine has nished.
Therefore, although the minimum number of resulting jobs is O=M , the actual
number depends on the exact manner in which the tasks are allocated to jobs.
Sizes of Operations The size of each operation in O is drawn from either a
uniform distribution between min and max, U min max , or a Gaussian distribution of speci ed mean and deviation, G m d .
Idle Time The amount of idle time on each machine is speci ed as a percentage
of the total processing time of operations on that machine. The idle time is
randomly and uniformally distributed between the actual operations.
Slack The maximum amount of slack in the arrival A and due-dates D of each
job is also speci ed as a percentage of the total processing time pt of the job.
If there is no slack, then the arrival date of the job is set to be equal to the
A S y s te m a tic In v e s tig a tio n o f G A P e rfo rm a n c e o n J o b s h o p S c h e d u lin g P ro b le m s
2 7 9
time slot at which the job is rst processed in the generated schedule, and the
due-date is set to the time-slot in which the job nishes. If slack is specied,
then the arrival and due dates of each job are altered randomly such that
A ! A , random0 S pt
D ! D + random0 S pt
Therefore, no job is ever tardy, and the maximum tardiness objective, Tmax
for each generated problem has an optimum value of 0. For each solution, there
is also a known upper bound on the makespan of each problem | if the problem
has no idle time, then the optimum makespan is known exactly.
3 Experimental Parameters
3.1 The Genetic Algorithm
The GA used is HGA which was described by the authors in 6 . This GA
outperformed other heuristic combination methods, and compared well to the
most recently published results on a number of benchmark problems. It uses an
indirect representation, in which each gene on the chromosome encodes a pair
Method Heuristic. The method gene denotes the methodology that should
be used to calculate a conicting set of schedulable operations at each iteration
of the algorithm. An operation is then chosen using the heuristic denoted by the
heuristic gene. The method applied is either the Gier and Thompson algorithm,
5 , which produces an active schedule, or a modied G&T algorithm which
produces a non-delay schedule.
The GA used a population size of 100, uniform crossover and a swap mutation
operator. Each experiment is run for a maximum of 500 generations, or until it
nds the optimum solution. The maximum tardiness of the best solution is noted,
with the number of generations required to nd it if the optimum was discovered.
The best solutions for each of the 100 trials in each problem class are averaged,
and we also record the minimum and maximum solution quality in each class.
3.2 The Problem Classes
In the experiments presented in this article we x the number of operations,
and hence the chromosome length, at 60. This facilitates a thorough and fair
investigation of the eects of the other four parameters on problem diculty.
Furthermore, as each GA experiment uses a chromosome population of identical
size and length, we can x the GA" parameters such as the number of generations to be the same in all experiments.
The other four parameters are varied as shown in table 1, resulting in a
total of 280 problem classes, each dened by a tuple P O M D I S . For all
experiments in which S 0, then I is xed at 0, and vice versa. For each of the
280 problem classes, 10 problems are generated from dierent random number
2 8 0
E . H a rt a n d P . R o ss
seeds. We then run a GA 10 times on each problem instance, and using max
as the objective function which we know to have an optimum solution of 0,
average the results over the 100 experiments in that class.
In the remainder of this article, the term problem class refers to the general
class of problems dened by a tuple
. The term problem instance
refers to one of the 10 problems generated for a problem class.
T
P O M D I S
Parameter
Value
Total Number of Operations
60
Number of Machines
2,3,4,5,6
Task Size Distribution U 0,10 , U 0,30 , U 0,50 , U 0,100
G 50,1 , G 50,10 , G 50,20 , G 50,40
Idle Time
5 , 10 , 25
Slack
10 , 25 , 50
Table 1. Experimental Parameters
4 Results
As space limitations do not allow us to present the results of all 280 experiments
here, we attempt to present summaries of the ndings and general trends observed, and discuss some individual problem classes in more detail. The complete
set of results can be found at 1 .
4.1 Overall Performance
Firstly, we note that in every problem class, the GA is able to nd the optimum
solution for at least 1 instance of the problems in that class. On the whole, better
performance is observed in the 140 experiments in which the operation sizes were
distributed uniformally, regardless of distribution size. For example, in 81 of
problem classes that had a uniform distribution of task size, then running the GA
on problem instances of that class resulted in the optimum solution being found
in at least half of all experiments. This compares to a gure of 59 for problem
classes with a Gaussian distribution of task size. These statements should be
treated with some caution however as they may simply be an artifact of the
parameters chosen to dene the uniform and Gaussian distributions, and the
two series of experiments cannot be directly compared.
Turning our attention to those problem classes which are solved to reasonable" accuracy, we note that 55 of the problem classes with Gaussian task
size distribution are solved to within 10 time units of the optimum solution. This
increases to 96 for problem classes with uniformally distributed operation sizes.
Tables 2 and 3 show the problem classes in which all generated instances
where solved with 100 accuracy. There is no obvious pattern, except that it
A S y s te m a tic In v e s tig a tio n o f G A P e rfo rm a n c e o n J o b s h o p S c h e d u lin g P ro b le m s
Idle Time Slack Std.Dev Machines
10
0
10
2
25
0
1
3
25
0
1
2
25
0
10
2
25
0
20
2
25
0
40
2
25
0
40
3
0
50 40
6
Table 2. Problems with Gaussian
Operation Size Distribution which
are solved with 100 accuracy
2 8 1
Idle Time Slack Max Size Machines
25
0
10
2
25
0
30
2
25
0
50
2
25
0
100
2
25
0
30
4
0
25 10
5
0
25 10
6
0
50 30
6
0
50 50
6
Table 3. Problems with Uniform
Operation Size Distribution which
are solved with 100 accuracy
appears that for perfect performance, either a large value of idle time or large
slack is required in all cases, and that for Gaussian distributions of operation
size, a large value of standard deviation helps.
For problems with a Gaussian distribution of operation size, the worst performance is observed in 2 problem classes which both result in the optimum
solution only being in 1 run of 1 of the 10 problem instances. These problems
classes are P 60 5 G50 10 0 0 and P 60 6 G50 10 0 0 . In the uniform
case, the worst performance, again with only 1 optimum solutions is for a problem class P 60 6 U 50 0 0 .
4.2 Number of Machines
Intuitively, it would be expected that as the number of machines increases for
a xed number of operations, O, then the problems would become easier to
solve. This is because the number of schedulable operations in the con ict set
produced as a result of applying the G&T or non-delay algorithm must reduce
as the number of machines increases. Table 4 shows the percentage of problem
classes with m 2 2 3 4 5 6 in which the optimum solution was found. For a
uniform distribution of jobs, the results are as expected. However, for the Gaussian distributions, we see exactly the opposite | i.e the percentage of optimum
solutions decreases as the number of machines is increased.
Number of Machines
2 3 4 5
6
Gaussian 63 56 55 55 58
Uniform 62 65 74 75 73
Table 4. of problem classes resulting in an optimum solution vs
no. machines
Slack Parameter
0.0 0.1 0.25 0.5
Gaussian 33 32 60 82
Uniform 46 50 76 80
Table 5. of problem classes resulting in an optimum solution vs
slack
2 8 2
E . H a rt a n d P . R o ss
4.3 Slack
Table 5 shows the percentage of problem classes with slack 2 0:0 0:1:0:25:0:5
in which the optimum solution was found. For problems with Gaussian distributions of task size, we observe that increasing the slack in due and arrival dates
to 10 does not increase performance, as would be expected, and in fact results
in a slight decrease. Increasing the slack to larger values however increases performance. This trend is not observed with uniform distributions, which tend to
become easier to solve as the slack parameter is increased.
Examining the e ect of the slack parameter in more detail in the Gaussian
experiments, we notice a correlation between the number of machines parameter
m and the slack S in determining solution quality. Table 6 shows the minimum
value of the slack parameter S that was required for each set of problem classes
with machine m before an improvement was observed in solution quality compared to the equivalent experiment with S = 0:0. Where no improvement was
observed, adding slack had a detrimental e ect on solution quality.
Number of Machines Minimum Percentage Standard Deviation
Machines
Slack Required
2
0.5
20,40
0.25
all values of std. dev.
3
4
0.1
all values of std. ded.
5
0.05
all values of std. dev.
0.05
all values of std. dev.
6
Table 6. Minimum Value of Slack Parameter Required to Improve Solution Quality
In problem classes where operation size was uniformally distributed, then
adding slack generally increases performance. The only exceptions where a signicant di erence occurs are for the problem classes P 60 2 U 0max 0:0 0:1,
where max 2 10 30 50 100.
4.4 Idle Time
For problems with both uniform and Gaussian task distributions, adding idle
time generally improves performance, as is seen in table 7, which shows the
percentage of problem classes in which the optimum solution was found for
various values of I . This is as expected | inserting idle time into the schedule
allows some exibility in the exact placement of operations in the schedule,
without necessarily decreasing schedule quality.
5 A Phase Transition in Problem Classes with Gaussian
Distributions
In the majority of problems with a Gaussian distribution of operation sizes, we
notice an interesting transition in the diculty of the problems as the standard
A S y s te m a tic In v e s tig a tio n o f G A P e rfo rm a n c e o n J o b s h o p S c h e d u lin g P ro b le m s
Table 7.
time
2 8 3
Idle Time
0.0 0.05 0.1 0.25
Gaussian 33 48 55 85
Uniform 46 65 75 89
of problem classes in which the optimum solution was achieved vs idle
deviation of the distribution is varied.The GA performs best on those problems
with large standard deviations. For very small standard deviations, i.e when the
all tasks have very similar sizes, then the GA also performs reasonably well.
However, for a range of values of standard deviation in the middle, performance decreases considerably. For example, gure 1 shows an expanded graph for
the problem class 60 5 50
0 0 0 0 , in which some extra experimental
points have been added. A clear peak in di culty is seen, centered around a
standard deviation of 7.
P
G
sd
:
:
A v e ra g e T a rd in e s s o f S o lu tio n s
5 0
4 5
4 0
3 5
3 0
2 5
2 0
1 5
1 0
0
5
0
5
1 0
1 5
2 0
2 5
3 0
3 5
4 0
S ta n d a rd D e v ia tio n o f D is trib u tio n
Fig. 1. Solution Quality vs Standard Deviation of
Task Size Distribution
This appearance of this phenomenum shows remarkably similar properties to
earlier work performed by the authors in the timetabling domain. Work described
in 11 showed that there was a clear phase transition in the performance of a
GA on a sequence of solvable timetabling problems designed to be of increasing
di culty the GA tested could solve very lightly constrained, and also very highly
constrained problems, however, for moderately constrained problems, the GA
would often fail to nd a solution. The appearance of similar phase-transition
regions has also been reported in other classes of constraint satisfaction problems,
for example see 7, 10, 12.
In the timetabling case, it was noted that other non-evolutionary algorithms
also failed on the same subset of problems, suggesting that it was not the GAs
fault", but that the problems were intrinsically di cult. We have not yet investigated the performance of other non-evolutionary methods on these problems,
but expect to see a similar pattern in performance.
2 8 4
E . H a rt a n d P . R o ss
1. Calculate the set C of all operations that can be
scheduled next
2. Calculate the completion time of all operations
in C , and let m equal the machine on which the
minimum completion time t is achieved.
3. Let G denote the conict set of operations on
machine m - this is the set of operations in C which
take place on m , and whose start time is less than t.
4. Select an operation from G to schedule
5. Delete the chosen operation from C and return
to step 1.
Fig. 2.
Gi er and Thompson Algorithm
A possible reason, still to be looked at in more detail, is the use of the
Gier and Thompson G&T algorithm or the modied non-delay version in
constructing the con ict sets of operations at each iteration. The G&T algorithm
is shown in gure 2. For problems in which there is a large deviation in operation
sizes, then it is possible that the con ict set is generally smaller at each iteration,
and therefore it is more straightforward to choose the 'correct' operation. When
all the operations are of similar size, then it seems likely that the size of the
con ict set is non-trivial, and hence it is more di cult to choose the 'correct'
operation. Figure 3 shows the size of the con ict set at each iteration, averaged
over 100 runs of experiments in which the standard deviation of operations
sizes was set to 5, and then to 40, using a constant mean of 50. Early in the
scheduling process, there is a small region, highlighted on the gure, in which
there a fewer items in the con ict set for the case when sd = 40. As we know that
the placement of operations early in the schedule is crucial to the success of the
algorithm, this may provide a clue, however the matter needs further attention.
6 Problem Classes with Uniform Distributions
As noted earlier, better results appear to be obtained when the operation sizes
are uniformally distributed, compared to those problems with gaussian distribution of operation sizes. Solution quality tends to decrease as the range of the
distribution increases. This information is summarised in table 8 which reports
the percentage of optimum solutions obtained for instances of problems tested
with each dierent range value.
A S y s te m a tic In v e s tig a tio n o f G A P e rfo rm a n c e o n J o b s h o p S c h e d u lin g P ro b le m s
2 8 5
4
S iz e o f c o n flc it s e t
3 .5
S ta n d a rd d e v ia tio n 4 0
3
S ta n d a rd d e v ia tio n 5
c r u c ia l r e g io n ?
2 .5
2
1 .5
1
0
1 0
2 0
3 0
4 0
5 0
6 0
Ite ra tio n s o f a lg o rith m
Fig. 3.
Size of Conict Set For Dierent Distributions of Operation Sizes
Maximum
Task Size
10 30 50 100
optimum solutions 80 76 71 68
Table 8. Percentage of problem instances resulting in an optimum solution vs maximum task size
7 Conclusion
This article has presented some initial observations made whilst attempting to
perform a systematic investigation on the performance of a genetic algorithm on
a range of job-shop scheduling problems. The investigation involved a total of
280 dierent problem classes, each containing 10 randomly generated problems.
Initial ndings suggest that although the GA does not nd the optimum
solution for all problem classes, for most instances of all problem classes it is
capable of nding the optimum in at least 1 in 10 trials, and that the quality of
solution is generally satisfactory. This suggests the GA appears to be a relatively
robust method of tackling such problems, and that since runs are reasonably
fast, running the GA several times to nd a solution is a viable strategy. Some
interesting trends have been observed | we now intend to try and understand
and explain these trends, and to determine whether they are a feature of the
problems themselves or due to the genetic algorithm itself.
This study concentrated on problem classes which all contained a xed number of operations. Therefore, in order to complete the study, we will also use
the generator to test GA performance on a sequence of much larger problems.
Although we expect performance to degrade as the problems get bigger, it is
fruitful to determine at what point this happens.
Finally, further work is planned to compare the performance of other nonevolutionary algorithms on the same set of problem classes, to see if similar
trends are observed, and to compare overall performance. These other methods
will include constraint satisfaction techniques, and simple scheduling rules.
2 8 6
E . H a rt a n d P . R o ss
Acknowledgements
Emma Hart is supported by EPSRC grant GRL22232.
References
1. http:www.dai.ed.ac.uk emmahjobshop-expts.html.
2. Sugato Bagchi, Serdar Uckun, Yutaka Miyabe, and Kazuhiko Kawamura. Exploring problem-specic recombination operators for job shop scheduling. In R.K.
Belew and L.B. Booker, editors, Proceedings of the Fourth International Conference
on Genetic Algorithms, pages 10 17. San Mateo: Morgan Kaufmann, 1991.
3. Brizuela. C.A. and N. Sannomiya. A diversity study in genetic algorithms for
jobshop. In Proceedings of the Genetic and Evolutionary Computation Conference,
pages 75 83, 1999.
4. L. Davis. Job shop scheduling with genetic algorithms. In J. J. Grefenstette,
editor, Proceedings of the International Conference on Genetic Algorithms and
their Applications, pages 136 140. San Mateo: Morgan Kaufmann, 1985.
5. B. Gier and G.L. Thompson. Algorithm for solving production scheduling problems. Operations Research, 84:487 503, 1960.
6. E. Hart and P. Ross. A heuristic combination method for jobshop scheduling
problems. In Parallel Problem Solving from Nature, PPSN-V, pages 845 854, 1998.
7. T. Hogg, A. Huberman, and C.P. Williams. Phase transitions and the search
problem. Articial Intelligence, 811-2:1 15, 1996.
8. S-C. Lin, E.D. Goodman, and W.F. Punch. A genetic algorithm approach to
dynamic job-shop scheduling problems. In Thomas Back, editor, Proceedings of the
Seventh International Conference on Genetic Algorithms, pages 481 489. MorganKaufmann, 1997.
9. R. Nakano and T. Yamada. Conventional genetic algorithms for job shop problems.
In R.K. Belew and L.B. Booker, editors, Proceedings of the Fourth International
Conference on Genetic Algorithms, pages 474 479. San Mateo: Morgan Kaufmann,
1991.
10. P. Prosser. An empirical study of phase transitions in binary constraint satisfaction
problems. Articial Intelligence, 811-2:81 109, 1996.
11. P Ross, E Hart, and D Corne. Some observations about ga-based exam timetabling.
In Practice and Theory of Automated Timetabling, pages 115 130, 1997.
12. B.M. Smith and M.E. Dyer. Locating the phase transitions in binary constraint
satisfaction problems. Articial Intelligence, 811-2:155 181, 1996.
13. P. Van Bael, D. Devogelaere, and M. Rijckaert. The job shop problem solved
with simple, basis evolutionary search elements. In Proceedings of the Genetic and
Evolutionary Computation Conference, pages 665 670, 1999.
An Ant Algorithm with a New Pheromone
Evaluation Rule for Total Tardiness Problems
Daniel Merkle1 and Martin Middendorf2
Institute for Applied Computer Science and Formal Description Methods,
University of Karlsruhe, Germany
f1 merkle,2middendorfg@aifb.uni-karlsruhe.de
Abstract. Ant Colony Optimization is an evolutionary method that
has recently been applied to scheduling problems. We propose an ACO
algorithm for the Single Machine Total Weighted Tardiness Problem.
Compared to an existing ACO algorithm for the unweighted Total Tardiness Problem our algorithm has several improvements. The main novelty is that in our algorithm the ants are guided on their way to good
solutions by sums of pheromone values. This allows the ants to take into
account pheromone values that have already been used for making earlier
decisions.
1
Introduction
Ant Colony Optimization ACO is an evolutionary metaheuristic to solve combinatorial optimization problems by using principles of communicative behaviour
found in real ant colonies for an introduction and overview see 5. Recently
the ACO approach has been applied to scheduling problems, like Job-Shop 2, 7,
Flow-Shop 13, and the Single Machine Total Tardiness problem 1. Bullnheimer
et al. 1 have compared an ACO algorithm with several other heuristics to solve
the Single Machine Total Tardiness problem e.g. decomposition heuristics, interchange heuristics and simulated annealing. They have shown that the ACO
algorithm found the optimal solution of 125 benchmark problems more often
than the other heuristics these benchmark problems where generated with the
same method from 12 as the benchmarks problems used in this paper .
In this paper we propose alternative and improved ways to solve the Single
Machine Total Tardiness problem by ACO. Moreover, we also study the weighted
version of the total tardiness problem.
In ACO algorithms several generations of articial ants search for good solutions. Every ant of a generation builds up a solution step by step going through
several probabilistic decisions until a solution is found. In general, ants that
found a good solution mark their paths through the decision space by putting
some amount of pheromone on the edges of the path. The following ants of the
next generation are attracted by the pheromone so that they will search in the
solution space near good solutions. In addition to the pheromone values the
ants will usually be guided by some problem specic heuristic for evaluating the
possible decisions.
S . C a g n o n i e t a l. ( E d s .) : E v o W o r k s h o p s 2 0 0 0 , L N C S 1 8 0 3 , p p . 2 8 7 − 2 9 6 , 2 0 0 0 .
S p rin g e r-V e rla g B e rlin H e id e lb e rg 2 0 0 0
2 8 8
D . M e rk le a n d M . M id d e n d o rf
The approach used in 1 and 13 to solve scheduling problems with ACO
algorithms is to use a pheromone matrix = f g where pheromone is added to
an element of the pheromone matrix when a good solution was found where
job is the th job on the machine. The following ants of the next generation
then directly use the value of to estimate the desirability of placing job as
the th job on the machine when computing a new solution.
Here we propose
P a dierent approach. Instead of using only the value of
the ants use =1 to compute the probability of placing job as the th
on the machine. A problem with using only can occur when the ant does
not chose job as the th job in the schedule. Because, if the +1 , +2
values are small then job might be scheduled much later than at the th place
and possibly
long after its due date . It is likely that this will not happen
P
when using =1 . Note, that this approach diers from nearly all other ant
algorithms proposed so far, in that we base one possible decision of an ant on
several pheromone values. The only other work that uses several pheromone
values to estimate the quality of one possible decision is 11. Moreover, we let
the ants make optimal decisions when this is possible and use a heuristic that is
a modi cation of the heuristic used in 1.
This paper is organized as follows. The Single Machine Total Weighted Tardiness Problem is de ned in Section 2. In Section 3 we describe an ACO algorithm
for the unweighted problem. The pheromone summation rule is introduced in
Section 4. Section 5 contains further variants and improvements. The choice of
the parameter values of our algorithms used in the test runs and the test instances and are described in Section 6. The results are reported in Section 7. A
conclusion is given in Section 8.
T
Tij
Tij
j
i
Tij
j
i
Tij
i
Tkj
k
j
i
Tij
j
i
Ti
j
Ti
j
i
k
j : : :
i
Tkj
2 The Single Machine Total Weighted Tardiness Problem
The Single Machine Total Weighted Tardiness Problem SMTWTP is to nd
for jobs, where job , 1 has a processing time , a due date ,
, a non-preemptive one machine schedule that minimizes =
Pand=1a weight
maxf0 , g where is the completion time of job . is called
the total weighted tardiness of the schedule. The unweighted case, i.e. = 1
for all 2 f1
g, is the Single Machine Total Tardiness Problem SMTTP .
It is known that SMTTP is NP-hard in the weak sense 8 and SMTWTP
is NP-hard in the strong sense 10. A pseudopolynomial time algorithm for
SMTWTP in case that the weights agree with the processing times i.e.
implies
was given in 10. Observe, that the last result implies
that SMTTP is pseudopolynomial time solvable. For an overview over dierent
heuristics for SMTWTP see 4.
n
n
j
j
j
n
pj
dj
wj
wj
T
Cj
dj
Cj
j
T
wj
j
: : : n
pj
ph
wj
wh
3 ACO Algorithm for SMTTP
The ACO algorithm of Bullnheimer et al. 1 is described in this section. The
general idea was to adapt an ACO algorithm called ACS-TSP for the traveling
A n A n t A lg o rith m
w ith a N e w P h e ro m o n e E v a lu a tio n R u le
2 8 9
salesperson problem of Dorigo et al. 6 for the SMTTP. In every generation
each of m ants constructs one solution. An ant selects the jobs in the order in
which they will appear in the schedule. For the selection of a job the ant uses
heuristic information as well as pheromone information. The heuristic information, denoted by ij , and the pheromone information, denoted by ij , are an
indicator of how good it seems to have job j at place i of the schedule. The
heuristic value is generated by some problem dependent heuristic whereas the
pheromone information stems from former ants that have found good solutions.
With probability q0, where 0 q0 1 is a parameter of the algorithm, the ant
chooses a job j from the set S of jobs that have not been scheduled so far which
maximizes
ij ij
where and are constants that determine the relative in uence of the
pheromone values and the heuristic values on the decision of the ant. With
probability 1 , q0 the next job is chosen according to the probability distribution
over S determined by
pij = P ij ij
h2S ih ih
The heuristic values ij are computed according the Modi ed Due Date rule
MDD, i.e.,
1
ij = maxfT 1+ p d g
j j
where T is the total processing time of all jobs already scheduled.
After an ant has selected the next job j, a local pheromone update is performed at element i j of the pheromone matrix according to
ij = 1 , ij + 0
for some constant , 0 1 and where
0 = m T1
EDD
and TEDD is the total tardiness of the schedule that is obtained when the
jobs are ordered according to the Earliest Due Date heuristic EDD, i.e., with
falling values of 1=dj . The value 0 is also used to initialize the elements of the
pheromone matrix.
After all m ants have constructed a solution the best of these solutions is
further improved with a 2-opt strategy. The 2-opt strategy considers swaps between all pairs of jobs in the sequence. Then it is checked whether the so derived
schedule is the new best solution found so far.
The best solution found so far is then used to update the pheromone matrix.
But before that some of the old pheromone is evaporated according to
2 9 0
D . M e rk le a n d M . M id d e n d o rf
= 1 , ij
The reason for this is that old pheromone should not have a too strong
inuence on the future. Then, for every job j in the schedule of the best solution found so far some amount of pheromone is added to element ij of the
pheromone matrix where i is the place of job j in the schedule. The amount
of pheromone added is =T where T is the total tardiness of the best found
schedule, i.e.,
ij
ij
=
1
ij + T
The algorithm stops when some stopping criterion is met, e.g. a certain number of generations has been done or the best found solution has not changed for
several generations.
4 The Pheromone Summation Rule
In this section we describe a new approach of using the pheromone values which
is used in our ACO algorithm for SMTTP. In general, a high pheromone value ij
means that it is advantageous to put job j at place i in the schedule. Assume now
that by chance an ant chooses to put some job h at place i of the schedule that
has a low pheromone value ih instead of a job j that has a high pheromone
value ij . Then in order to have a high chance to still end up with a good
solution it will likely be necessary for the ant to place job j not too late in the
schedule when j has a small due date. To some extend the heuristic values lj
for l i will then force the ant to choose j soon. But a problem occurs when
the values lj are small because no good solutions have been found before that
have job j at some place l i. Then the product lj lj is small and it is
likely that the ant will not choose j soon. In this case the ant will end up with
a useless solution having a high total tardiness value.
To handle this problem we propose to let a pheromone value ij also inuence
later decisions when choosing a job for some place l i. A simple way to guaranty
this inuence is to use the sum of all pheromone values for every job from the
rst row of the matrix up to row i when deciding about the job for place i. When
using this pheromone summation rule we have the following modi ed decision
formulas. An ant chooses as next job for place i in the schedule with probability
q0 the job j 2 S that maximizes
X
i
k=1
kj
kj
2
and with probability 1 , q0 job j 2 S is chosen according to the probability
distribution over S determined by
A n A n t A lg o rith m
pij
w ith a N e w P h e ro m o n e E v a lu a tio n R u le
2 9 1
Pi kj ij
kP
=1
i kh ih
=P
h2S
3
k=1
5 Further Variations and Improvements
In this section we describe further variations and improvements that we used in
our ACO algorithm.
5.1 Modied Heuristic
A problem when using the heuristic values according to formula 1 is that the
values of max + pj dj become much larger | due to | when deciding
about jobs to place at the end of the schedule than they are when placing jobs
at the start of the schedule. As a consequence the heuristic di erences between
the jobs are, in general, small at the end of the schedule. To avoid this e ect we
used the following modi ed values
1
4
ij =
max + pj dj
For the weighted problem SMTWTP we multiplied every value on the right
side of equation 4 with the weight wj of job j . Note that jobs with a small
weighted processing time pj =wj have a high heuristic value when + pj dj .
fT
g
T
fT
g, T
T
5.2 Deterministic Scheduling Between Due Dates
Consider the construction of a schedule for the unweighted problem SMTTP.
Assume that some jobs have already been scheduled. Assume further that the
sum of the processing times of all jobs scheduled so far lies between some due
date dj and a due date dh dj and every other due date is smaller than dj or
larger than dh . For this case it is easy to show that it is optimal to schedule all
jobs with a due date dj before scheduling a job with a due date dh as long
as the sum of the processing times of the scheduled jobs is at most dh . Moreover
when there are several jobs with due date dj it is optimal to schedule these
jobs ordered by increasing processing times. If the ants apply this deterministic
rule whenever possible we say that the ants work locally deterministic. Then the
ants will switch between probabilistic and deterministic behaviour.
T
6 Test Instances and Parameters
We tested the di erent variants of ACO algorithms on 125 benchmark instances
for SMTWTP of size 100 jobs that are included in the OR-Library 14. These
benchmark instances were generated as follows: for each job j 1 : 125 an
integer processing time pj is taken randomly from the interval 1 : 100, an
2
2 9 2
D . M e rk le a n d M . M id d e n d o rf
integer weight wj is taken randomly from the interval 1 : 10 and an integer due
date dj is taken randomly from the interval
3
2
125
125
X
X
4 pj 1 , TF , RDD pj 1 , TF + RDD 5
j =1
2
j =1
2
The value RDD relative range of due dates determines the length of the interval from which the due dates were taken. TF tardiness factor
P determines the
relative position of the centre of this interval between 0 and 125
j =1 pj . The values
for TF and RDD are chosen from the set 0:2 0:4 0:6 0:81:0 . The benchmark
set contains ve instances for each combination of TF and RDD values. For the
unweighted problem SMTTP we used the same benchmark instances but ignored
the dierent weights.
Our results for SMTWTP were compared to the best known results for the
benchmark instances that are from 3 and can be found in 14.
The parameters used for the test runs are: = 1, = 1, = 0:1, q0 0 0:9 .
The number of ants in every generation was m = 20. Every test was performed
with 4 runs on every instance. Every run was stopped after 500 generations.
We used a 2-opt strategy to improve the best solution that was found in every
generation which diers slightly from the 2-opt strategy used in 1. For every
pair of jobs it was checked exactly once whether a swap of these jobs improves
the schedule. A swap that improves the schedule was xed immediately. Thus
we tried exactly 4950 swaps per generation.
In the following ACS-SMTTP or short ACS denotes the algorithm of 1
as described in Section 3 but with the new 2-opt strategy described in the last
paragraph. Our algorithm ACS-SMTWTP- is similar to ACS but uses the
pheromone summation rule as described in Section 4. Algorithm ACS-SMTWTPH is similar to ACS but uses the new heuristic from Section 5.1. Algorithm ACSSMTWTP-D is similar to ACS but additionally uses the deterministic strategy
from Section 5.2 for scheduling between due dates. Algorithms that use combinations of new features are denoted by ACS-SMTWTP-XYZ where X,Y,Z ,
H, D e.g. ACS-SMTWTP-H uses the new heuristic and the pheromone summation rule . For shortness we write ACS-XYZ for ACS-SMTWTP-XYZ.
f
g
2 f
g
2 f
g
7 Experimental Results
The inuence of the pheromone summation rule called -rule in the following
and the modied heuristic was tested on weighted and unweighted problem instances. Since the parameter q0 has some inuence on the results we performed
tests with q0 = 0 and q0 = 0:9.
Table 1 shows the results for SMTWTP. The average total tardiness values
found by the ACO algorithms for SMTWTP were compared to the average total tardiness of the best known solutions that are from 3. The average total
tardiness per instance of the best solutions from 3 is 217851:34. Table 1 shows
that ACS-H performed better than ACS-H and also that ACS- performed
A n A n t A lg o rith m
w ith a N e w P h e ro m o n e E v a lu a tio n R u le
2 9 3
better than ACS this holds for both cases 0 = 0 and 0 = 0 9. In all cases the
dierence of the total tardiness values compared to the best known solutions are
at least 61 1 lower for the ACO algorithm with -rule 79 5 for ACS- H compared to 204 5 for ACS-H with 0 = 0 9. Moreover, the ACO algorithms with
-rule found for more instances a better total tardiness than their counterparts
without -rule at least 5 3 times as often. The dierences of the total tardiness values compared to the best known values over the rst 200 generations are
shown in Figure 1. The best solution of ACS- H was found after an average of
80 generations, which was after less than 3.5 seconds on a 450 MHz Pentium-II
processor.
Table 1 also shows that the ACO algorithms with modied heuristic performed in all cases better than their counterparts using the heuristic from 1.
For 0 = 0 9 the advantage of the modied heuristic is smaller than for 0 = 0
e.g. for 0 = 0 9 ACS- H has a 60 2 smaller dierence to optimal total tardiness than ACS- compared to a 92 9 smaller dierence for 0 = 0.
q
q
:
:
q
:
:
:
:
q
:
q
q
:
:
:
q
Table 1. Inuence of pheromone summation rule and new heuristic on solution quality
for SMTWTP. Total Tardiness: average dierence to total tardiness of best found solutions from 3 average over 500 test runs, 125 instances and 4 runs for each instance
Better: comparisons between ACS- H and ACS-H respectively ACS- and ACS,
number of instances with smaller average total tardiness average over 125 instances
and 4 runs for each instance.
weighted
ACS- H ACS-H ACS- ACS
Total
q0 = 0
191.8 3024.7 946.1 9914.7
Tardiness q0 = 0:9
79.5 204.5 200.0 1198.6
q
97
2
106
0
0 =0
Better q = 0:9
86
16
97
3
0
Table 2 shows the results for the unweighted problem SMTTP. The results
are compared with the average of the best total tardiness values we found for
the unweighted instances, i.e. 54309 5. Similarly as for the weighted problem in
all cases the ACO algorithms with -rule are better than their counterparts
without -rule. Also the modied heuristic performed better in all cases than
the heuristic from 1.
Since the 2-opt strategy signicantly inuences of the quality of the solutions
we also compared the ACS- H with ACS-H when using no 2-opt strategy. The
results can be found in Table 3 for SMTWTP and in Table 4 for SMTTP. The
only case where ACS- H performed not signicantly better than ACS-H is the
unweighted case with 0 = 0 9. In this case ACS-H found a slightly better average
total tardiness ACS- H dierence is 331 5 for ACS-H and 332 3 for ACS- H.
On the other hand ACS- H found for more instances better solutions than
ACS-H For 65 instances ACS- H found better solutions than ACS-H whereas
ACS-H performed better than ACS- H for 33 instances.
:
q
:
:
:
2 9 4
D . M e rk le a n d M . M id d e n d o rf
Fig. 1. SMTWTP: Average dierence to total tardiness of best found solutions from
3 over the rst 200 generations.
ACS- H
ACS-H
ACSACS
2000
1500
1000
500
0
0
50
100
150
200
Inuence of pheromone summation rule and new heuristic on solution quality
for SMTTP. Total Tardiness: average dierence to total tardiness of best found solutions average over 500 test runs, 125 instances and 4 runs for each instance Better"
as in Table 1.
unweighted
ACS- H ACS-H ACS- ACS
Total
q0 = 0
47.9 48.5 112.9 256.4
Tardiness q0 = 0:9
7.0 19.0
8.7 26.3
q
53
32
82 17
0 =0
Better q = 0:9
53
22
67 14
Table 2.
0
Inuence of pheromone summation rule and new heuristic on solution quality
for SMTWTP when using no 2-opt. Total Tardiness" as in Table 1 Better" as in
Table 1 but comparison between ACS- H and ACS-H.
Table 3.
no 2-0pt, weighted ACS- H ACS-H
Total
q0 = 0 11894.4 22046.8
Tardiness q0 = 0:9 1733.2 1793.5
0
76
48
Better qq0 =
67
42
0 = 0:9
A n A n t A lg o rith m
w ith a N e w P h e ro m o n e E v a lu a tio n R u le
2 9 5
Table 4. Inuence of pheromone summation rule and new heuristic on solution quality
for SMTTP when using no 2-opt. Total Tardiness" as in Table 2 Better" as in Table
1 but comparison between ACS- H and ACS-H.
no 2-0pt, unweighted ACS- H ACS-H
Total
q0 = 0
3943.7 4515.8
Tardiness q0 = 0:9
332.3 331.5
q
59
50
0 =0
Better q = 0:9
65
33
0
The inuence of the deterministic strategy for scheduling between due dates
for SMTTP has only a minor inuence on the results for the unweighted benchmark instances from the OR-Library. The reason is that these instances have
small gaps between the due dates. Thereby, the deterministic strategy does
come into play only rarely. Hence, we created new test instances which have
two neighboured due dates that have a large gap in between. We changed each
of the problem instances from the OR-Library as follows. The jobs were ordered
by their due dates and the due dates of jobs 41 to 59 were set to the same due
date that job 40 has. The average of the best total tardiness values we found for
these modi ed instances was 56416 3. Table 5 shows for 0 = 0, that ACS- HD
performed much better than ACS- H and also that ACS-HD performed much
better than ACS-H. For 0 = 0 9 the ACS- HD algorithm could not pro t from
the deterministic scheduling between due dates.
:
q
q
:
Table 5. Inuence of deterministic
strategy between due dates on solution quality for
SMTTP and problem instances with modied due dates. Total Tardiness as in Table
2 Better" as in Table 1 but comparison between ACS- H and ACS- HD.
unweighted
ACS- H ACSTotal
q0 = 0
101.4
Tardiness q0 = 0:9
2.9
q
8
0 =0
Better q = 0:9
36
0
HD ACS-H ACS-HD
45.7 120.1
3.8
8.7 11.1
9.2
78
1
92
14
29
36
8 Conclusion
We have introduced a new method to use the pheromone values in an Ant Colony
Optimization ACO algorithm for the Single Machine Total Weighted Tardiness
problem. An ACO algorithm using this pheromone summation rule gives better
solutions for 125 benchmark than its counterpart that does not use the summation rule. This holds also for the unweighted total tardiness problem. Moreover,
2 9 6
D . M e rk le a n d M . M id d e n d o rf
we proposed a new heuristic that can be used by the ants when searching for a
solution. For the unweighted problem we have shown that the ACO algorithm
can prot from ants that switch between a deterministic behaviour in case that
optimal decisions can be made and the standard" probabilistic behaviour.
References
1. A. Bauer, B. Bullnheimer, R.F. Hartl, C. Strauss: An Ant Colony Optimization
Approach for the Single Machine Total Tardiness Problem in: Proceedings of the
1999 Congress on Evolutionary Computation CEC99, 6-9 July Washington D.C.,
USA, 1445-1450, 1999.
2. A. Colorni, M. Dorigo, V. Maniezzo, M. Trubian: Ant System for Job-Shop Scheduling JORBEL - Belgian Journal of Operations Research, Statistics and Computer
Science, 34: 3953 1994.
3. R.K. Congram, C.N. Potts, S. L. van de Velde: An iterated dynasearch algorithm
for the single-machine total weighted tardiness scheduling problem submitted to
INFORMS Journal on Computing.
4. H.A.J. Crauwels, C.N. Potts, L.N. Van Wassenhove: Local Search Heuristics for the
Single Machine Total Weighted Tardiness Scheduling Problem INFORMS Journal
on Computing, 10: 341359 1998.
5. M. Dorigo, G. Di Caro: The ant colony optimization meta-heuristic in: D. Corne,
M. Dorigo, F. Glover Eds., New Ideas in Optimization, McGraw-Hill, 1999, 11-32.
6. M. Dorigo, L. M. Gambardella: Ant colony system: A cooperative learning approach
to the travelling salesman problem IEEE Trans. on Evolutionary Comp., 1: 53-66
1997.
7. M. Dorigo, V. Maniezzo, A. Colorni: The Ant System: Optimization by a Colony
of Cooperating Agents IEEE Trans. Systems, Man, and Cybernetics Part B, 26:
29-41 1996.
8. J. Du, J.Y.-T. Leung: Minimizing the Total Tardiness on One Machine is NP-hard
Mathematics of Operations Research, 15: 483496 1990.
9. P. Forsyth, A. Wren: An Ant System for Bus Driver Scheduling Report 97.25,
University of Leeds - School of Computer Studies, 1997.
10. E.L. Lawler: A `pseudopolynomial' algorithm for sequencing jobs to minimize total
tardiness Annals of Discrete Mathematics, i: 331342 1977.
11. R. Michels, M. Middendorf: An Ant System for the Shortest Common Supersequence Problem in: D. Corne, M. Dorigo, F. Glover Eds., New Ideas in Optimization, McGraw-Hill, 1999 692-701.
12. C.N. Potts, L.N. Van Wessenhove: Single machine tardiness sequencing heuristics
IEE Transactions, 23: 346-354 1991
13. T. Stutzle: An ant approach for the ow shop problem in Proc. of the 6th European
Congress on Intelligent Techniques & Soft Computing EUFIT '98 , Vol. 3, Verlag
Mainz, Aachen, 1560-1564, 1998.
14. http:mscmga.ms.ic.ac.ukjeborlibwtinfo.html.
$1HZ*HQHWLF5HSUHVHQWDWLRQDQG
&RPPRQ&OXVWHU&URVVRYHU
IRU-RE6KRS6FKHGXOLQJ3UREOHPV
7(=8.$ 0DVDUX +,-, 0DVDKLUR
0,<$%$<$6+, .D]XQRUL DQG 2.8085$ .HLJR
+LWDFKL 7RKRNX 6RIWZDUH /WG
+RQFKR $REDNX 6HQGDL 0L\DJL -DSDQ
^WH]XND KLML PL\DED\ RNXPXUD`#KLWDFKLWRFRMS
-897,.9 7KLV SDSHU GHVFULEHV D JHQHWLF DOJRULWKP DSSURDFK IRU VHTXHQFLQJ
SUREOHPV HVSHFLDOO\ IRU MRE VKRS VFKHGXOLQJ ,Q DFWXDO SUREOHPV VHWXS WLPH
VKRXOG EH RSWLPL]HG ,Q RUGHU WR UHGXFH VHWXS WLPH ZH GHYHORSHG D QHZ JHQHWLF
UHSUHVHQWDWLRQ DQG DQ HIILFLHQW FURVVRYHU RSHUDWRU FDOOHG &RPPRQ &OXVWHU
&URVVRYHU &&; ,Q RXU UHSUHVHQWDWLRQ FKURPRVRPHV UHSUHVHQW WKH VKLIW RI WKH
RUGHU LQ D VHTXHQFH 7R SUHVHUYH VXEVHTXHQFHV LQ FURVVRYHU RSHUDWLRQV ZH LP
SOHPHQWHG WKH SURFHVV WR LGHQWLI\ WKH FOXVWHU RI WKH VXEVHTXHQFHV DQG DSSOLHG
&&; WKDW H[FKDQJHV FRPPRQ FOXVWHUV EHWZHHQ WZR SDUHQWV 7KH DSSURDFK ZDV
WHVWHG RQ WZR VWDQGDUG EHQFKPDUNV DQG DSSOLHG WR DQ DXGLR SDUWV PDQXIDFWXUHU
&&; DFKLHYHG UHPDUNDEOH UHVXOWV RQ WKH DFWXDO MRE VKRS VFKHGXOLQJ SUREOHP
,QWURGXFWLRQ
-RE 6KRS 6FKHGXOLQJ 3UREOHPV -663V DUH VROYHG E\ DVVLJQLQJ MREV WR UHVRXUFHV
7KRVH DVVLJQPHQWV DUH FDOOHG RSHUDWLRQV -REV DUH EDVLF WDVNV 7KURXJK D FKDLQ RI
RSHUDWLRQV PDWHULDOV DUH SURFHVVHG LQWR LQWHUPHGLDWH SDUWV DQG WKH SDUWV DUH SURF
HVVHG LQWR SURGXFWV &RQVHTXHQWO\ HDFK MRE FRPSOHWHV 7KH VHTXHQFH DPRQJ RSHUD
WLRQV RI HDFK MRE PXVW EH UHVWULFWHG $W DQ\ WLPH RQH UHVRXUFH FDQ SURFHVV RQO\ RQH
RSHUDWLRQ
-663V DUH NQRZQ DV WKH PRVW GLIILFXOW SUREOHPV DQG WKH\ FDQQRW EH VROYHG LQ D
SUDFWLFDOO\ WLPHO\ PDQQHU EHFDXVH WKH\ DUH 13KDUG 7KH JHQHWLF DOJRULWKP *$
ZKLFK LV UHFRJQL]HG DV WKH HIIHFWLYH WRRO WR VROYH FRPELQDWRULDO SUREOHPV KDV EHHQ
DSSOLHG WR -663V LQ YDULRXV DSSURDFKHV >@>@>@>@
S . C a g n o n i e t a l. ( E d s .) : E v o W o r k s h o p s 2 0 0 0 , L N C S 1 8 0 3 , p p . 2 9 7 − 3 0 6 , 2 0 0 0 .
S p rin g e r-V e rla g B e rlin H e id e lb e rg 2 0 0 0
2 9 8
T . M a s a ru e t a l.
,Q WKLV SDSHU ZH SURSRVH D QHZ JHQHWLF VHTXHQFLQJ DSSURDFK DQG D QHZ FURVVRYHU
RSHUDWRU WKDW ZRUNV HIILFLHQWO\ WR PLQLPL]H VHWXS WLPH RQ -663V :KHQ *$V DUH XVHG
LQ VHTXHQFLQJ DSSURDFK VHYHUDO VSHFLDOO\ GHVLJQHG FURVVRYHU RSHUDWRUV VXFK DV WKH
SDUWLDOO\ PDSSHG FURVVRYHU
30; >@ WKH RUGHU EDVHG FURVVRYHU >@ DQG & >@ DUH
XVHG WR HOLPLQDWH LQIHDVLEOH VHTXHQFHV 7KRVH RSHUDWRUV SUHVHUYH VRPH RI VXE
VHTXHQFHV EXW IUHTXHQWO\ GHVWUR\ RWKHUV )RU H[DPSOH WKH RUGHU EDVHG DQG & SUH
VHUYH WKH FRPSOHWH VXEVHTXHQFHV LQKHULWHG IURP RQH SDUHQW EXW MXVW D UHODWLYH
SRVLWLRQ IURP WKH RWKHU
,Q PRVW SUDFWLFDO -663V VXEVHTXHQFHV DUH YHU\ LPSRUWDQW EHFDXVH VHWXS WLPH LV
UHTXLUHG IRU UHVRXUFHV WR FKDQJH VHWXSV IURP WKH RQH IRU D FHUWDLQ LWHP WR WKH RQH IRU D
QHZ LWHP 6XFFHVVLYH SURGXFWLRQ RI WKH VDPH LWHPV UHGXFHV WKH VHWXS WLPH 7KXV ZH
FRQVLGHU WKDW SUHVHUYLQJ JRRG VXEVHTXHQFHV WKDW UHGXFHV WKH VHWXS WLPH LV LPSRUWDQW
WR UHDFK WKH JRRG VROXWLRQ TXLFNO\
,Q RUGHU WR DOORZ RIIVSULQJ WR LQKHULW VXEVHTXHQFHV IURP WKH ERWK SDUHQWV ZH
SURSRVH D QHZ JHQHWLF VHTXHQFLQJ UHSUHVHQWDWLRQ DQG WKH &RPPRQ &OXVWHU &URVVRYHU
&&; ,Q RXU IUDPHZRUN FKURPRVRPHV UHSUHVHQW WKH VKLIW RI WKH RUGHU LQ D VHTXHQFH
:H DGGHG WKH SURFHVV WR LGHQWLI\ WKH FOXVWHU RI WKH VXEVHTXHQFHV DQG FURVVRYHU
RSHUDWLRQ LV DSSOLHG WR WKH FRPPRQ FOXVWHUV EHWZHHQ WZR SDUHQWV
,Q 6HFWLRQ ZH GHVFULEH WKH GHWDLOV RI WKH UHSUHVHQWDWLRQ DQG &&; ,Q 6HFWLRQ
ZH UHSRUW WKH UHVXOWV RI WZR FRPSXWDWLRQDO H[SHULPHQWV RQH LV IRU )7[ DQG
)7[ DQG WKH RWKHU LV IRU DQ DXGLR SDUWV PDQXIDFWXUHU V VFKHGXOLQJ SUREOHP 7KHQ
ZH FRQFOXGH ZLWK WKH UHVXOWV LQ 6HFWLRQ
3UHVHUYLQJ6XEVHTXHQFHV
*HQHWLF 5HSUHVHQWDWLRQ
,Q RXU DSSURDFK FKURPRVRPHV UHSUHVHQW WKH VKLIWLQJ GLVWDQFH RI WKH RUGHU LQ D
VHTXHQFH 2QH FKURPRVRPH FRUUHVSRQGV WR RQH UHVRXUFH DQG RQH LQGLYLGXDO FRQVLVWV
RI D JURXS RI FKURPRVRPHV 7KH QXPEHU RI FKURPRVRPHV LQ RQH LQGLYLGXDO HTXDOV WR
WKH QXPEHU RI UHVRXUFHV LQ D VFKHGXOH
(DFK FKURPRVRPH LV GHFRGHG WR WKH VHTXHQFH RI WKH MREV RQ WKH UHVRXUFH /HW XV
FRQVLGHU WKH FDVH LQ ZKLFK M MREV DUH SURFHVVHG E\ D FHUWDLQ UHVRXUFH 7KH OHQJWK RI WKH
FKURPRVRPH FRUUHVSRQGLQJ WR WKH UHVRXUFH LV M 7KH LWK DOOHOH RI WKH FKURPRVRPH
« M M M
« M M
« ML
« M
&KURPRVRPH
JHQH
JHQH
JHQH
«
)LJ 7KH DOOHOHV
JHQH L
«
JHQH M
A N e w G e n e tic R e p re s e n ta tio n a n d C o m m o n C lu s te r C ro s s o v e r
2 9 9
FDQ WDNH DQ LQWHJHU > ML@ DV VKRZQ LQ )LJ
7KH GHFRGLQJ SURFHVV RI WKH FKURPRVRPH WR WKH VHTXHQFH LV VKRZQ LQ )LJ ,Q WKH
ILJXUH WKH FKURPRVRPH LV ^` $W ILUVW WKH VHTXHQFH LV LQLWLDOL]HG DV
^` 7KH YDOXH RI WKH ILUVW JHQH LV ILYH DQG WKLV PHDQV WKDW WKH ILUVW V\PERO
RI WKH VHTXHQFH LV WR EH H[FKDQJHG IRU WKH V\PERO RQ ILYH VWHSV ULJKW 7KXV V\PERO
µ¶ DQG µ¶ DUH H[FKDQJHG DQG WKH VHTXHQFH EHFRPHV ^` 7KH YDOXH RI WKH
VHFRQG JHQH LV IRXU DQG WKLV PHDQV WKDW WKH VHFRQG V\PERO RI WKH VHTXHQFH LV WR EH
H[FKDQJHG IRU WKH V\PERO RQ IRXU VWHSV ULJKW 7KXV V\PERO µ¶ DQG µ¶ DUH H[FKDQJHG
DQG WKH VHTXHQFH EHFRPHV ^` 5HSHWLWLRQ RI WKLV SURFHGXUH IRU WKH ZKROH
JHQHV OHDGV WR WKH VHTXHQFH ^`
$ FKURPRVRPH WR EH GHFRGHG
&KURPRVRPH
6HTXHQFH
&KURPRVRPH
6HTXHQFH
&KURPRVRPH
,QLWLDOL]H WKH VHTXHQFH
6HTXHQFH
&KURPRVRPH
6HTXHQFH
&KURPRVRPH
6HTXHQFH
6KLIW WKH RUGHUV
7KH VHTXHQFH GHFRGHG IURP WKH FKURPRVRPH
6HTXHQFH
)LJ 'HFRGLQJ SURFHVV RI WKH FKURPRVRPH WR WKH VHTXHQFH
3 0 0
T . M a s a ru e t a l.
&RPPRQ &OXVWHU &URVVRYHU &&;
7KH UDQGRP VHOHFWLRQ RI WKH FURVVRYHU SRLQW LV W\SLFDOO\ XVHG LQ WKH FRQYHQWLRQDO *$
DSSURDFKHV (YHQ LI WKH QHZ JHQHWLF UHSUHVHQWDWLRQ ZH PHQWLRQHG DERYH LV DGRSWHG
WKH UDQGRP VHOHFWLRQ FDQ GHVWUR\ WKH VXEVHTXHQFH ,Q RUGHU WR VHOHFW WKH HIIHFWLYH
FURVVRYHU SRLQW ZH SURSRVH WKH &RPPRQ &OXVWHU &URVVRYHU &&; &&; LGHQWLILHV
WKH FOXVWHUV RI VXEVHTXHQFHV DQG H[FKDQJHV WKH FRPPRQ FOXVWHUV EHWZHHQ WZR
SDUHQWV
$ FOXVWHU FRQVLVWV RI WKH JHQHV OLQNHG WRJHWKHU )LJ VKRZV WKH H[DPSOH RI FOXVWHUV
$V PHQWLRQHG LQ 6HFWLRQ HDFK YDOXH RI D JHQH LV WKH VKLIWLQJ GLVWDQFH 7KH DUURZV
LQ WKH ILJXUH LQGLFDWH WKH OLQN RI WKH VKLIWV ,Q WKLV FDVH WKH FKURPRVRPH KDV WZR
FOXVWHUV $V VKRZQ LQ WKH ILJXUH D FOXVWHU FDQ FRQVLVW RI VHYHUDO VHSDUDWH VHJPHQWV
&&; H[FKDQJHV WKH FRPPRQ FOXVWHUV RI WZR SDUHQWV )LJ GHVFULEHV &&; SURFHVV
,Q WKLV H[DPSOH 3DUHQW KDV WZR FOXVWHUV $ DQG % 3DUHQW KDV IRXU FOXVWHUV & ' (
DQG ) 7ZR GRWWHG DUURZV LQGLFDWH WKH FRPPRQ ERXQGDULHV RI FOXVWHUV &&;
H[FKDQJHV WKH FRPPRQ FOXVWHUV WKDW DUH VHSDUDWHG E\ WKH FRPPRQ ERXQGDULHV DQG
FUHDWH WZR RIIVSULQJ
)LJ VKRZV WKH H[DPSOH RI &&; 2QH RIIVSULQJ LV UHSURGXFHG IURP 3DUHQW 3
DQG 3DUHQW 3 E\ &&; 3 KDV WZR FOXVWHUV $ DQG % 3 KDV IRXU FOXVWHUV & ' (
DQG ) 6LQFH WZR ERXQGDULHV PDWFK EHWZHHQ 3 DQG 3 &&; H[FKDQJHV FOXVWHUV DW
&KURPRVRPH
&OXVWHU
&OXVWHU
)LJ 7KH FOXVWHU RI VXEVHTXHQFHV
3DUHQW
$
%
$
3DUHQW
&
'
(
2IIVSULQJ
$
'
(
2IIVSULQJ
&
%
)LJ &RPPRQ &OXVWHU &URVVRYHU &&;
)
&
$
)
&
A N e w G e n e tic R e p re s e n ta tio n a n d C o m m o n C lu s te r C ro s s o v e r
3 0 1
3DUHQW
&KURPV
$
6HTXHQFH
$
%
3DUHQW
&KURPV
&
6HTXHQFH
(
'
'
•
)
2IIVSULQJ
&KURPV
6HTXHQFH
)URP 3DUHQW
)URP 3DUHQW
)URP 3DUHQW
)LJ &RPPRQ &OXVWHU &URVVRYHU
WKHVH SRLQWV
, ZRXOG OLNH WR HPSKDVL]H WKDW WKH RIIVSULQJ LQKHULWV WKH VXEVHTXHQFHV IURP ERWK
3 DQG 3 ,Q WKLV FDVH WKH VXEVHTXHQFH
IURP 3 DQG
ZDV LQKHULWHG
IURP 3 :H EHOLHYH WKDW WKLV IHDWXUH RI
&&; LV YHU\ HIIHFWLYH WR RSWLPL]H DQG UHGXFH WKH VHWXS WLPH LQ -663V
7KHUH DUH WZR FDVHV LQ ZKLFK &&; FDQQRW EH DSSOLHG 7KH ILUVW FDVH RFFXUV ZKHQ
QR SRVLWLRQ RI WKH ERXQGDULHV PDWFKHV EHWZHHQ WZR SDUHQWV 7KH VHFRQG FDVH RFFXUV
ZKHQ ERWK SDUHQWV FRQVLVW RI RQH FOXVWHU ,Q RWKHU ZRUGV WKHUH LV QR ERXQGDU\ ,Q WKH
ILUVW FDVH &&; LV DSSOLHG DFFRUGLQJ WR WKH FOXVWHUV RI RQH SDUHQW DQG WKRVH RI WKH
RWKHU SDUHQW DUH LJQRUHG :H FDOO WKLV FURVVRYHU TXDVL&&; ,Q WKH VHFRQG FDVH
FRQYHQWLRQDO RQHSRLQW FURVVRYHU LV DSSOLHG 1DWXUDOO\ SUHVHUYDWLRQ RI VXEVHTXHQFHV
LV QRW JXDUDQWHHG LQ WKHVH FDVHV
&RPSXWDWLRQDO5HVXOWV
6WDQGDUG -RE 6KRS 6FKHGXOLQJ 3UREOHP
7KH *$ EDVHG RQ RXU UHSUHVHQWDWLRQ DQG &&; KDV EHHQ LPSOHPHQWHG LQ -DYD
-'. RQ WKH 3HQWLXP 0+] 3& ZLWK :LQGRZV
3 0 2
T . M a s a ru e t a l.
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
ɼÈ̼źÐ
7KH *$ ZDV WHVWHG RQ WZR VWDQGDUG MRE VKRS SUREOHPV )LVKHU DQG 7KRPSVRQ¶V
)7[ DQG )7[ >@ DV D EHQFKPDUN 7KHVH WZR SUREOHPV KDYH EHHQ ZLGHO\
XVHG E\ PDQ\ UHVHDUFKHUV DV EHQFKPDUNV 7KH REMHFWLYH RI WKH EHQFKPDUN SUREOHPV LV
WR PLQLPL]H PDNHVSDQ 7KH RSWLPDO VROXWLRQV DUH RQ )7[ DQG RQ
)7[ ,Q WKH ILHOG RI *$V RQ )7[ <DPDGD>@ ILUVW IRXQG WKH RSWLPXP
IRXU WLPHV DPRQJ WULDOV LQ .RED\DVKL>@ IRXQGV ILIW\RQH WLPHV
DPRQJ WULDOV LQ 2Q )7[ 0DWWIHOG>@ DQG /LQ>@ IRXQG RSWLPXP
LQ UHVSHFWLYHO\
,W LV NQRZQ WKDW WKH RSWLPDO VFKHGXOHV RI PDNHVSDQ DUH DPRQJ DFWLYH VFKHGXOHV
7KH PHWKRG WR JHQHUDWH DFWLYH VFKHGXOHV *7 PHWKRG ZDV SURSRVHG E\ *LIIOHU DQG
7KRPSVRQ >@ 6RPH RWKHU DSSURDFKHV XVH WKH *7 PHWKRG >@>@>@ :H DOVR XVHG
WKH *7 PHWKRG WR LQWHUSUHW WKH MRE VHTXHQFH LQWR DQ DFWLYH VFKHGXOH
:H DGRSWHG WKH UDQNLQJ VHOHFWLRQ >@ DQG WKH HOLWLVW SROLF\ >@ 0XWDWLRQ FKDQJHV
WKH YDOXH RI D JHQH LQWR D UDQGRP LQWHJHU IROORZLQJ XQLIRUP GLVWULEXWLRQ > ML@
ZKHUH M LV WKH QXPEHU RI MREV DQG L LV WKH ORFL RI WKH JHQH
,Q RXU H[SHULPHQWV WKH SRSXODWLRQ VL]H ZDV VHW WR DQG WHUPLQDWLRQ ZDV VHW WR
JHQHUDWLRQV &URVVRYHU UDWH ZDV VHW WR DQG PXWDWLRQ UDWH ZDV VHW WR ,W
WRRN DERXW VHFRQGV WR UXQ RQH WULDO RQ )7[ DQG VHFRQGV RQ )7[
)LJ DQG )LJ VKRZ WKH IUHTXHQF\ GLVWULEXWLRQ RI VROXWLRQV PDNHVSDQ LQ
WULDOV RQ )7[ DQG )7[ UHVSHFWLYHO\ 7KH DSSURDFK ZH SURSRVHG FRXOG QRW
REWDLQ RSWLPD EXW SURGXFHG KLJKO\ HIILFLHQW VROXWLRQV , PXVW SRLQW RXW WKDW LQ
SUDFWLFDO XVH LW LV QHFHVVDU\ WR FUHDWH JRRG HQRXJK VFKHGXOHV UHOLDEO\ DQG IDVW UDWKHU
WKDQ ILQG RSWLPDO VROXWLRQV
$V PHQWLRQHG LQ 6HFWLRQ RXU FURVVRYHU RSHUDWLRQ FRQVLVWV RI WKUHH NLQGV RI
FURVVRYHU &&; LV DSSOLHG ZKHQ WKHUH DUH FRPPRQ FOXVWHUV EHWZHHQ WZR SDUHQWV
4XDVL&&; LV DSSOLHG ZKHQ QR ERXQGDU\ PDWFKHV EHWZHHQ WZR SDUHQWV 2QHSRLQW
FURVVRYHU LV DSSOLHG ZKHQ ERWK SDUHQWV FRQVLVW RI RQH FOXVWHU )LJ VKRZV D WHPSRUDO
¤¸Â¼ÊǸÅ
)LJ )UHTXHQF\ GLVWULEXWLRQ RI VROXWLRQV RQ )7[
3 0 3
>
>
>
>
>
>
>
>
>
>
>
>
ɼÈ̼źÐ
A N e w G e n e tic R e p re s e n ta tio n a n d C o m m o n C lu s te r C ro s s o v e r
¤¸Â¼ÊǸÅ
«¿¼ ÅÌĹ¼É ƽ ºÉÆÊÊÆͼÉ
)LJ )UHTXHQF\ GLVWULEXWLRQ RI VROXWLRQV RQ )7[
&&;
4XDVL&&;
2QHSRLQW
¼Å¼É¸ËÀÆÅ
)LJ $ W\SLFDO WHPSRUDO FKDQJH RI WKH QXPEHU RI WKUHH NLQGV RI FURVVRYHU
FKDQJH RI WKH QXPEHUV RI WKUHH NLQGV RI FURVVRYHU LQ D FHUWDLQ WULDO RQ )7[ ,Q WKH
ILUVW JHQHUDWLRQ DERXW RI FURVVRYHU RSHUDWLRQV ZDV &&; ZDV TXDVL&&;
DQG ZDV RQHSRLQW FURVVRYHU 7KH SURSRUWLRQ RI &&; LQFUHDVHG JUDGXDOO\ DQG LQ
WKH HQG RI LWHUDWLRQ &&; LQFUHDVHG WR TXDVL&&; GHFUHDVHG WR DQG RQH
SRLQW FURVVRYHU GHFUHDVHG WR
:H VXUPLVH WKDW WKH SURSRUWLRQ RI &&; LQFUHDVHG VLQFH JRRG VXEVHTXHQFHV DV
EXLOGLQJ EORFNV ZHUH FUHDWHG PDLQWDLQHG DQG FRPELQHG GXULQJ LWHUDWLRQ
$SSOLFDWLRQ WR DQ $XGLR 3DUWV 0DQXIDFWXUHU
7KH *$ DSSURDFK ZH SURSRVHG ZDV DSSOLHG WR WKH VFKHGXOLQJ RI DQ DXGLR SDUWV PDQX
IDFWXUHU 6LQFH WKH VFKHGXOLQJ SUREOHP LQYROYHV VHWXS WLPH UHGXFWLRQ ZH FDQ WHVW WKH
RSWLPL]DWLRQ DELOLW\ RI &&; E\ PHDVXULQJ WKH UHGXFWLRQ LQ VHWXS WLPH
3 0 4
T . M a s a ru e t a l.
7KH PDQXIDFWXUHU LV D VXEFRQWUDFWRU RI DQ HOHFWULFDO DSSOLDQFH PDQXIDFWXUHU DQG
UHFHLYHV SURGXFWLRQ RUGHUV RQFH D PRQWK $FFRUGLQJ WR WKH SURGXFWLRQ RUGHUV LW
PDNHV LWV SURGXFWLRQ VFKHGXOH RQFH D PRQWK
,W KDV ILIWHHQ ZRUN FHQWHUV DV UHVRXUFHV 2QH KXQGUHG ZRUNHUV DUH LQYROYHG LQ WKH
SURGXFWLRQ ,WV PDLQ SURGXFWV DUH DXGLR VSHDNHUV DQG KHDGSKRQHV ,W SURGXFHV DERXW
SURGXFWV SHU PRQWK /RW VL]HV DUH FDOFXODWHG E\ D 053 V\VWHP ,WV PRQWKO\
SURGXFWLRQ VFKHGXOH W\SLFDOO\ FRQVLVWV RI RSHUDWLRQV ,Q WKH SURGXFWLRQ VHWXS WLPH
LV UHTXLUHG IRU UHVRXUFHV WR FRQYHUW IURP WKH SURGXFWLRQ RI D FHUWDLQ LWHP WR DQRWKHU
LWHP
)RUPHUO\ WKH PDQXIDFWXUHU KDG XVHG D GLVSDWFKLQJEDVHG VFKHGXOHU ,Q RUGHU WR
UHGXFH VHWXS WLPH WKH ZRUNHU KDG WR PRGLI\ WKH VFKHGXOH FUHDWHG E\ WKH VFKHGXOHU ,W
WRRN DERXW WKUHH WR IRXU KRXUV WR PRGLI\ ZLWK D *DQWW FKDUW HGLWRU 7R UHILQH WKH
VFKHGXOLQJ SURFHVV RXU *$EDVHG VFKHGXOLQJ V\VWHP ZDV LQWURGXFHG 7KH REMHFWLYH
RI WKH PDQXIDFWXUH LV WR PLQLPL]H WRWDO GHOLYHU\ GHOD\ WLPH WKH QXPEHU RI GHOD\LQJ
MREV DQG WRWDO VHWXS WLPH
)LUVW ZH XVHG WRWDO VHWXS WLPH VROHO\ DV WKH ILWWLQJ IXQFWLRQ LQ RUGHU WR WHVW WKH DELO
LW\ RI &&; 7KH WHVW GDWD FRQVLVW RI MREV DQG RSHUDWLRQV )RU FRPSDULVRQ ZH
DOVR DSSOLHG RUGHU EDVHG RSHUDWRU >@ DV WKH WUDGLWLRQDO RSHUDWRU WR WKH VDPH SUREOHP
3RSXODWLRQ VL]H ZDV VHW WR DQG WHUPLQDWLRQ ZDV VHW WR 7R FUHDWH VFKHGXOHV LQ D
UHDVRQDEO\ VKRUW WLPH :H VHW WKH YDOXHV VPDOOHU WKDQ RQ WKH EHQFKPDUNV ,W WRRN
DERXW VHFRQGV WR UXQ RQH WULDO ZLWK ERWK &&; DQG RUGHU EDVHG )LJ VKRZV WKH
IUHTXHQF\ GLVWULEXWLRQ RI VROXWLRQV VHWXS WLPH LQ WULDOV 2EYLRXVO\ &&; LV VXSH
ULRU WR WKH WUDGLWLRQDO FURVVRYHU RSHUDWRU LQ PLQLPL]LQJ VHWXS WLPH
1H[W ZH WHVWHG ZHLJKWHG VXP RI WRWDO GHOLYHU\ GHOD\ WLPH WKH QXPEHU RI GHOD\LQJ
MREV DQG WRWDO VHWXS WLPH DV WKH ILWWLQJ IXQFWLRQ RQ WKH VDPH GDWD 3RSXODWLRQ VL]H DQG
WHUPLQDWLRQ ZHUH DOVR VHW WR DQG UHVSHFWLYHO\ 7DEOH VKRZV WKH EHVW DQG WKH
DYHUDJH RI &&; DQG RUGHUEDVHG LQ WULDOV 7KH UHVXOW RI D OHJDF\ GLVSDWFKLQJ
¯
¦É»¼É¹¸Ê¼»
>
>
>
>
>
>
>
>
>
>
>
>
ɼÈ̼źÐ
«Æ˸à ʼËÌÇ ËÀļ ÄÀÅÌ˼
)LJ )UHTXHQF\ GLVWULEXWLRQ RI VROXWLRQV RQ VHWXS WLPH PLQLPL]LQJ
A N e w G e n e tic R e p re s e n ta tio n a n d C o m m o n C lu s te r C ro s s o v e r
3 0 5
7DEOH $SSOLFDWLRQ WR DQ DXGLR SDUWV PDQXIDFWXUHU
EHVW
DYHUDJH
EHVW
2UGHU
EDVHG
DYHUDJH
/HJDF\ VFKHGXOHU
&&;
7RWDO VHWXS
WLPH PLQXWHV
7RWDO GHOLYHU\
GHOD\ WLPH GD\V
7KH QXPEHU RI
GHOD\LQJ MREV
EDVHG VFKHGXOHU LV DOVR VKRZQ IRU FRPSDULVRQ &&; DSSURDFK VXFFHVVIXOO\ RSWLPL]HG
WKH ZKROH LQGLFHV HVSHFLDOO\ VHWXS WLPH :KDW QHHGV WR EH HPSKDVL]HG KHUH LV WKDW LQ
RUGHU WR PLQLPL]H WRWDO VHWXS WLPH VXEVHTXHQFHV DUH TXLWH LPSRUWDQW EHFDXVH
VXFFHVVLYH SURGXFWLRQ RI WKH VDPH NLQG RI LWHPV UHGXFHV WKH WRWDO VHWXS WLPH
UHPDUNDEO\
&RQFOXVLRQV
,Q WKLV SDSHU ZH SURSRVHG D QHZ JHQHWLF UHSUHVHQWDWLRQ DQG &RPPRQ &OXVWHU
&URVVRYHU &&; &KURPRVRPHV UHSUHVHQW WKH VKLIWLQJ GLVWDQFH RI WKH RUGHU LQ D
VHTXHQFH DQG DUH GHFRGHG WR WKH VHTXHQFH RI WKH MREV &&; H[FKDQJHV WKH FRPPRQ
FOXVWHUV RI WZR SDUHQWV ,QWHJUDWHG XVH RI RXU UHSUHVHQWDWLRQ DQG &&; DOORZV RIIVSULQJ
WR SUHVHUYH VXEVHTXHQFHV 2XU GLVFXVVLRQ VKRZV WKDW WKLV IHDWXUH LV KLJKO\ HIIHFWLYH
LQ VROYLQJ -663V HVSHFLDOO\ SUREOHPV WR RSWLPL]H VHWXS WLPH
:H WHVWHG RXU DSSURDFK RQ WKH VWDQGDUG -663V )7[ DQG )7[ DQG IRXQG WKH
DSSURDFK ZDV HIIHFWLYH
6LQFH ZH GHYHORSHG &&; LQ RUGHU WR RSWLPL]H VHWXS WLPH LQ SDUWLFXODU ZH DOVR
DSSOLHG RXU QHZ DSSURDFK WR DQ DXGLR SDUWV PDQXIDFWXUHU 6HWXS WLPH VKRXOG EH GHDOW
ZLWK RQ WKH SUREOHP )LUVW RQO\ VHWXS WLPH ZDV XVHG DV WKH ILWWLQJ IXQFWLRQ WR WHVW WKH
DELOLW\ RI &&; DQG &&; VXFFHVVIXOO\ PLQLPL]HG VHWXS WLPH 1H[W WKH ILWWLQJ
IXQFWLRQ ZDV VHW WR PLQLPL]H WKH ZHLJKWHG VXP RI WKH WRWDO GHOLYHU\ GHOD\ WLPH WKH
QXPEHU RI GHOD\LQJ MREV DQG WKH WRWDO VHWXS WLPH 7KH SURSRVHG DSSURDFK
VXFFHVVIXOO\ RSWLPL]HG WKH -663 EDVHG RQ WKH DFWXDO GDWD
5HIHUHQFHV
>@ /DZUHQFH 'DYLV -RE 6KRS 6FKHGXOLQJ ZLWK *HQHWLF $OJRULWKPV 3URFHHGLQJV RI
WKH )LUVW ,QWO &RQIHUHQFH RQ *HQHWLF $OJRULWKPV DQG WKHLU $SSOLFDWLRQV
>@ 6XJDWR %DJFKL 6HUGDU 8FNXQ <XWDND 0L\DEH DQG .D]XKLNR .DZDPXUD
([SORULQJ 3UREOHP6SHFLILF 5HFRPELQDWLRQ 2SHUDWRUV IRU -RE 6KRS 6FKHGXOLQJ
3URFHHGLQJV RI WKH )RUWK ,QWO &RQIHUHQFH RQ *HQHWLF $OJRULWKPV
3 0 6
T . M a s a ru e t a l.
>@ 5\RKHL 1DNDQR DQG 7DNHVKL <DPDGD &RQYHQWLRQDO *HQHWLF $OJRULWKP IRU -RE
6KRS 3UREOHPV 3URFHHGLQJV RI WKH )RUWK ,QWO &RQIHUHQFH RQ *HQHWLF $OJRULWKPV
>@ &KULVWLDQ %LHUZLUWK $ JHQHUDOL]HG SHUPXWDWLRQ DSSURDFK WR MRE VKRS VFKHGXOLQJ
ZLWK JHQHWLF DOJRULWKPV 25 6SHNWUXP 9RO
>@ 'DYLG ( *ORGEHUJ DQG 5REHUW /LQJOH -U $OOHOHV /RFL DQG 7UDYHOLQJ 6DOHVPDQ
3UREOHP 3URFHHGLQJV RI WKH )LUVW ,QWO &RQIHUHQFH RQ *HQHWLF $OJRULWKPV DQG WKHLU
$SSOLFDWLRQV
>@
*LOEHUW
6\VZHUGD
6FKHGXOH
2SWLPL]DWLRQ
8VLQJ
*HQHWLF
$OJRULWKPV
,Q
+DQGERRN RI *HQHWLF $OJRULWKPV /DZUHQFH 'DYLV HG 9DQ 1RVWUDQG 5HLQKROG 1HZ
<RUN
>@ &ROLQ 5 5HHYHV $ *HQHWLF $OJRULWKP IRU )ORZVKRS 6HTXHQFLQJ &RPSXWHUV DQG
2SHUDWLRQV 5HVHDUFK 9RO 1R
>@ + )LVKHU * / 7KRPSVRQ 3UREDELOLVWLF OHDUQLQJ FRPELQDWLRQV RI ORFDO MREVKRS
VFKHGXOLQJ UXOHV ,Q ,QGXVWULDO 6FKHGXOLQJ
-) 0XWK * / 7KRPSVRQ HG 3UHQWLFH
+DOO (QJOHZRRG &OLIIV 1HZ -HUVH\
>@ 7DNHVKL <DPDGD DQG 5\RKHL 1DNDQR $ *HQHWLF $OJRULWKP $SSOLFDEOH WR /DUJH
6FDOH -RE 6KRS 3UREOHPV 3DUDOOHO 3UREOHP 6ROYLQJ IURP 1DWXUH
>@ 6KLJHQREX .RED\DVKL ,VDR 2QR DQG 0DVD\XNL <DPDPXUD $Q (IILFLHQW *HQHWLF
$OJRULWKP
IRU
-RE
6KRS
6FKHGXOLQJ
3UREOHPV
3URFHHGLQJV
2I
WKH
WK
,QWO
&RQIHUHQFH 2Q *HQHWLF $OJRULWKPV
>@ ' & 0DWWIHOG + .RSIHU DQG & %LHUZLUWK &RQWURO RI 3DUDOOHO 3RSXODWLRQ
'\QDPLFV E\ 6RFLDOOLNH %HKDYLRU RI *$,QGLYLGXDOV 3DUDOOHO 3UREOHP 6ROYLQJ IURP
1DWXUH
>@ 6K\K&KDQJ /LQ (ULN ' *RRGPDQ DQG :LOOLDP ) 3XQFK ,,, ,QYHVWLJDWLQJ
3DUDOOHO
*HQHWLF
$OJRULWKPV
RQ
-RE
6KRS
6FKHGXOLQJ
3UREOHPV
(YROXWLRQDU\
3URJUDPPLQJ 9, 3URFHHGLQJV RI 6L[WK ,QWHUQDWLRQDO &RQIHUHQFH (3
>@ % *LIIOHU DQG * / 7KRPSVRQ $OJRULWKP IRU 6ROYLQJ 3URGXFWLRQ 6FKHGXOH
3UREOHPV 2SHUDWLRQV 5HVHDUFK 9RO
>@ 5REHUW + 6WRUHU 6 'DYLG :X DQG 5HQ]R 9DFFDUL 1HZ 6HDUFK 6SDFHV IRU
6HTXHQFLQJ 3UREOHPV ZLWK $SSOLFDWLRQ WR -RE 6KRS 6FKHGXOLQJ 0DQDJHPHQW 6FLHQFH
9RO
>@
-DPHV
(
%DNHU
3URFHHGLQJV
RI
WKH
$GDSWLYH
)LUVW
,QWO
6HOHFWLRQ
&RQIHUHQFH
0HWKRGV
RQ
IRU
*HQHWLF
*HQHWLF
$OJRULWKPV
$OJRULWKPV
DQG
WKHLU
$SSOLFDWLRQV
>@ *XQDU ( /LHSLQV DQG : ' 3RWWHU $ *HQHWLF $OJRULWKP $SSURDFK WR 0XOWLSOH
)DXOW 'LDJQRVLV ,Q +DQGERRN RI *HQHWLF $OJRULWKPV /DZUHQFH 'DYLV HG 9DQ
1RVWUDQG 5HLQKROG 1HZ <RUN
O p tim is in g a n E v o lu tio n a r y A lg o r ith m
N e il U rq u h a rt, K e n C h is h o lm
S c h o o l O f C o m
N a p ie r U n iv
2 1 9 C o lin to n
E d in b u rg
E H 1 4 1 D
{n e i l u , k e n , b e n P }@ d c
fo r S c h e d u lin g
a n d B e n P a e c h te r
p u tin g
e rs ity
R o a d
h
J
s . n a p i e r . a c . u k
A b s t r a c t . T h is p a p e r e x a m in e s tw o te c h n iq u e s fo r s e ttin g th e p a ra m e te rs o f a n
e v o lu tio n a r y A lg o r ith m ( E A ) . T h e e x a m p le E A u s e d f o r te s t p u r p o s e s
u n d e r ta k e s a s im p le s c h e d u lin g p r o b le m . A n in itia l v e rs io n o f th e E A w a s
te s te d u tilis in g a s e t o f p a ra m e te rs th a t w e r e d e c id e d b y b a s ic e x p e rim e n ta tio n .
T w o s u b s e q u e n t v e r s io n s w e re c o m p a r e d w ith th e in itia l v e rs io n , th e firs t o f
th e s e a d ju s te d th e p a ra m e te rs a t ru n tim e , th e s e c o n d u s e d a s e t o f p a r a m e te r s
d e c id e d o n b y r u n n in g a m e ta - E A . T h e a u th o r s h a v e b e e n a b le to c o n c lu d e th a t
th e u s a g e o f a m e ta - E A a llo w s a n e f fic ie n t s e t o f p a r a m e te r s to b e d e r iv e d fo r
th e p ro b le m E A .
1 .
A D e s c r ip tio n o f th e P r o b le m
T h e u s e o f E v o lu tio n a ry A lg o rith m s (E A s ) fo r s o lv in g tim e ta b lin g a n d s c h e d u lin g
p ro b le m s h a s b e c o m e c o m m o n p la c e in re c e n t y e a rs [1 ], [9 ]. T h is p a p e r e x a m in e s tw o
m e th o d s fo r o p tim is in g th e p a ra m e te rs u s e d b y th e a lg o rith m . A lth o u g h th e c o n c e p ts
a n d m e th o d s u s e d fo r d e v e lo p in g a n d o p tim is in g th e e v o lu tio n a ry a lg o rith m a re
in te n d e d to b e g e n e ra l, fo r th e p u rp o s e s o f p re s e n ta tio n w ith in th is p a p e r th e y w ill b e
a p p lie d to a s im p le s c h e d u lin g p ro b le m . T h e s c h e d u lin g p ro b le m u n d e r c o n s id e ra tio n
re q u ire s a n u m b e r o f jo b s to b e p ro c e s s e d th ro u g h a fa c to ry . E a c h jo b h a s s ta rt a n d
e n d tim e s th a t fo rm its tim e w in d o w . W ith in th is w in d o w th e jo b m u s t “ v is it” a ll th e
re s o u rc e s re q u ire d fo r th is jo b , w ith in a s p e c ifie d o rd e r. E a c h re s o u rc e is m u tu a lly
e x c lu s iv e (ie o n ly o n e jo b m a y m a k e u s e o f a re s o u rc e a t a n y o n e tim e ).
T h e v a rio u s c o n flic tin g c o n s tra in ts fo r th e s c h e d u lin g ta s k c a n b e d iv id e d in to tw o
c a te g o rie s h a rd c o n s tra in ts a n d s o ft c o n s tra in ts , a s s h o w n in ta b le 1 . A h a rd c o n s tra in t
is o n e th a t m u s t b e s a tis fie d in o rd e r to p ro d u c e a fe a s ib le s c h e d u le , s o ft c o n s tra in ts
a lte r th e q u a lity o f th e s c h e d u le .
2 .
A D e s c r ip tio n o f th e E A
T h e g e n e ra l c o n c e p ts o f e v o lu tio n a ry a lg o rith m s a re w e ll u n d e rs to o d . E a c h in d iv id u a l
re q u ire s a re p re s e n ta tio n (a g e n o ty p e ) th a t m a y b e d e c o d e d in to a s p e c ific s c h e d u le (a
S . C a g n o n i e t a l. ( E d s .) : E v o W o r k s h o p s 2 0 0 0 , L N C S 1 8 0 3 , p p . 3 0 7 − 3 1 8 , 2 0 0 0 .
S p rin g e r-V e rla g B e rlin H e id e lb e rg 2 0 0 0
3 0 8
N . U rq u h a rt, K . C h is h o lm , a n d B . P a e c h te r
p h e n o ty p e ). E a c h
(s to re d a s a re a l
c h ro m o s o m e , th e
re s o u rc e a re s to re
c o n s id e re d a d ire c
d ire c t re p re s e n ta tio
a c t d ire c tly o n th e
w o u ld re p re s e n t a
p ro c e s s in g w o u ld b
g e n e w ith in th e g e n o ty p e re p re s e n ts th e s ta rt tim e o f th e e v e n t
n u m b e r). W ith in e a c h g e n o ty p e th e e v e n t g e n e s m a k e u p a
e v e n t g e n e s a re g ro u p e d s o th a t a ll th e g e n e s fo r a p a rtic u la r
d to g e th e r in o rd e r o f s ta rt tim e . T h is re p re s e n ta tio n m a y b e
t re p re s e n ta tio n , in th a t th e g e n e s w ith in th e c h ro m o s o m e a re a
n o f th e fin is h e d s c h e d u le . T h e m u ta tio n a n d c ro s s o v e r o p e ra to rs
ite m s in th e s c h e d u le . W ith a n in d ire c t re p re s e n ta tio n th e g e n e s
s e t o f in s tru c tio n s fo r b u ild in g th e s c h e d u le [1 0 ], th u s fu rth e r
e re q u ire d to c o n v e rt th e g e n o ty p e in to a s c h e d u le .
T a b le 1 . H a rd a n d s o f t c o n s tra in ts
C o
E v
E v
T h
T h
T h
T h
n s tr a in t
e n t m u s t s ta r t a f te r th e
e n t m u s t s ta r t a f te r th e
e f ir s t e v e n t m u s t s ta r t
e la s t e v e n t m u s t fin ish
e e v e n t s h o u ld ta k e p la
e e v e n t s h o u ld ta k e p la
In th e c o u rs e
s u m m a rise d a s :
o f th e se
p re v io u s
p re v io u s
a f te r th e
a fte r th e
c e in n o r
c e o n w e
e v
e v
jo b
jo
m a
e k
e n t o n
e n t in
s “ a v a
b s “ d u
l w o rk
d a y s
e x p e rim e n ts 3
th is r e s
th is jo b
ila b le ”
e ” tim e
in g h o u
d a ta
T y
H a
H a
H a
S o
S o
S o
o u r c e h a s fin is h e d .
h a s fin ish e d .
tim e
rs
se ts w e re
u tilis e d , th e y
p e
rd
rd
rd
f t1
ft
ft
m a y
b e
D a ta s e t 1 : 3 5 E v e n ts , w ith in 1 6 J o b s u s in g 4 R e s o u rc e s , o v e r a 9 d a y p e rio d
D a ta s e t 2 : 4 1 E v e n ts , w ith in 1 5 J o b s u s in g 8 R e s o u rc e s , o v e r a 4 d a y p e rio d
D a ta s e t 3 : 5 0 E v e n ts , w ith in 1 8 J o b s u s in g 8 R e s o u rc e s , o v e r a 4 d a y p e rio d
E a c h re so u rc e h a s a ‘q
m in u te . T h u s fo r a n e v e
p e rio d th e re a re 1 4 4 0 * 4 p
d a y ). T h u s if th e re w e re 1
n u m b e r o f p o te n tia l q u e
d iffe re n t q u e u e s . T h e to ta
th e p ro d u c t o f th e s iz e o f
th e s e s c h e d u le s a re in fe a
p ro d u c tio n .
2 .1
u e u e ’ o f e v e n ts , e a c h e v e n t h a s a s ta rt tim e s e t to th e n e a re s t
n t b e in g p la c e d w ith in a s c h e d u le b e in g b u ilt fo r a 4 d a y
o s s ib le v a lu e s fo r th e s ta rt tim e (th e re a re 1 4 4 0 m in u te s in a
0 e v e n ts s c h e d u le d to u s e a re s o u rc e o v e r a 2 d a y p e rio d th e
u e s fo r th a t re s o u rc e w o u ld b e (4 8 * 1 4 4 0 )* 1 0 = 6 1 9 2 0 0
l n u m b e r o f p o te n tia l s o lu tio n s to th e s c h e d u lin g p ro b le m is
th e p o te n tia l q u e u e s fo r e a c h re s o u rc e . T h e v a s t m a jo rity o f
s ib le , a n d th e re s tric te d m u ta tio n o p e ra to r d is c o u ra g e s th e ir
B a s ic O p e r a tio n s
W ith in th e s y s te m tw o -p o in t c ro s s o v e r, b a s e d o n th e D a v is s y s te m is u s e d to c re a te
c h ild re n , b a s e d o n e x c h a n g in g s ta rt tim e g e n e s . T h e c h ild re p la c e s o n e o f th e
ra n d o m ly s e le c te d p a re n ts . T h e E A u s e s s te a d y s ta te p o p u la tio n to re d u c e c o m p le x ity .
1
A jo b w h ic h is n o t c o m p le te d u n til a fte r th e e n d o f its tim e w in d o w
d e s ira b le , b u t w ill s till re s u lt in a s c h e d u le th a t is fe a s ib le .
m a y n o t b e
O p tim is in g a n E v o lu tio n a ry A lg o rith m
U s e is m a d e o f e litis m , th is e n s u re s th a t th e b e s t in d iv id u a ls o f
p re s e rv e d fo r in c lu sio n in th e n e x t g e n e ra tio n .
T h e m u ta tio n o p e ra to r c o n s id e rs e a c h m e m b e r o f th e p o p
m u ta tio n , th e c h a n c e o f a s p e c ific m e m b e r b e in g s e le c te d is b a
ra te w h ic h is e x p re s s e d a s a p e rc e n ta g e . F o r in s ta n c e w ith a m
th e re is a 3 5 % c h a n c e o f a n y in d iv id u a l b e in g s e le c te d . W h
s e le c te d , o n e g e n e is s e le c te d fo r m u ta tio n w ith a s e le c tio n p re s s u
th e le n g th o f th e c h ro m o s o m e ).
T h e fitn e s s fu n c tio n is b a s e d o n th e c o n s tra in ts o u tlin e d in
s a tis fa c tio n o f h a rd c o n s tra in ts b e in g g iv e n g re a te r re w a rd s th a
s o ft c o n s tra in ts . T h e fitn e s s fo r e a c h in d iv id u a l e v e n t is c a lc u la
s h o w n in ta b le 2 .
fo r S c h e d u lin g
3 0 9
e a c h g e n e ra tio n a re
u la tio n fo r p o s s ib le
s e d o n th e m u ta tio n
u ta tio n ra te o f 3 5 %
e n a n in d iv id u a l is
re o f 1 /L (w h e re L is
s e c tio n 1 , w ith th e
n th e s a tis fa c tio n o f
te d u s in g th e v a lu e s
T a b le 2 . I n itia l w e ig h t v a lu e s u s e d in th e f itn e s s fu n c tio n
C o n s tra in t
C u rre n t e v
A ll e v e n ts
C u rre n t e v
C u rre n t e v
E v e n t ta k e
E v e n t ta k e
E v e n t ta k e
e n t is
th a t m
e n t ta
e n t ta
s p la c
s p la c
s p la c
“ p la c e d ” (d o e s n o t c o n flic t w ith a n y o th e r e v e n t)
a k e u p th e c u rre n t jo b p la c e d
k e s p la c e b e fo re th e jo b is d u e fo r c o m p le tio n
k e s p la c e a fte r th e jo b is “ a v a ila b le ”
e b e tw e e n M o n d a y a n d F rid a y
e b e fo re “ n ig h t tim e ”
e a fte r “ d a y tim e ”
R e w a rd
1 2 0
1 0 0
5 0
5 0
2 5
4
4
T h e E A w a s a llo w e d to ru n u n til a ll th e e v e n ts w e re p la c e d (ie a ll th e h a rd
c o n s tra in ts a re s a tis fie d ). T h e re m a y a ris e s itu a tio n s w h e re th e re is n o fe a s ib le
s c h e d u le fo r a g iv e n d a ta s e t (ie if 2 5 h o u rs o f w o rk a re s c h e d u le d to ta k e p la c e in o n e
d a y ) th e re fo re a lim it is s e t o n th e to ta l n u m b e r o f g e n e ra tio n s th a t m a y b e c a lc u la te d .
T h is lim it is c a lc u la te d a s n o o f e v e n ts * 1 5 0 . T h ro u g h in itia l e x p e rim e n ta tio n it w a s
fo u n d th a t a llo w in g th e E A to c o n tin u e ru n n in g fo r a s h o rt p e rio d a fte r a fe a s ib le
s o lu tio n h a d e v o lv e d a llo w e d fu rth e r im p ro v e m e n t w ith re g a rd s to th e s o ft
c o n s tra in ts . T h e n u m b e r o f g e n e ra tio n s u s e d fo r th is ru n -o n p e rio d w a s c a lc u la te d a s
to ta l n u m b e r o f e v e n ts * 5 .
2 .2
I m p r o v in g th e Q u a lity o f S c h e d u le s u s in g L a m a r c k is m
T o im p ro v e th e q u a lity o f s c h e d u le s c o n s tru c te d b y th e s y s te m , th e a u th o rs
im p le m e n te d L a m a rc k ia n w rite b a c k [1 1 ],[5 ]. L a m a rc k ia n th e o ry s u g g e s ts th a t
k n o w le d g e a c q u ire d b y in d iv id u a ls d u rin g th e ir life tim e m a y b e p a s s e d g e n e tic a lly
o n to th e ir d e s c e n d a n ts . T h is th e o ry h a s s in c e b e e n d is c re d ite d b y b io lo g is ts w h o
m a in ta in th a t c ro s s o v e r a n d ra n d o m m u ta tio n a re th e m a in in flu e n c e s o n e v o lu tio n a ry
d e v e lo p m e n t. T h e ty p e o f im p ro v e m e n t m a d e to a s c h e d u le is illu s tra te d in fig u re 1 , it
m a y b e a c c e p te d th a t a n y s o lu tio n w h e re e v e n ts w ith in a jo b “ o v e rla p ” is n o t fe a s ib le
a n d th u s o f n o p ra c tic a l v a lu e . T h e w o rk o f [1 0 ] o u tlin e s th e u s e o f L a m a rc k ia n is m
w ith in a n E A u s e d fo r s o lv in g th e N a p ie r U n iv e rs ity tim e ta b lin g p ro b le m .
3 1 0
2 .3
N . U rq u h a rt, K . C h is h o lm , a n d B . P a e c h te r
T a r g e te d M u ta tio n
In [1 0 ] u s e is m a d e o f w h a t is k n o w n a s “ T a rg e te d m u ta tio n ” . T h is m o d ifie s th e
m u ta tio n o p e ra to r to in c re a s e th e c h a n c e s o f g e n e s w ith s p e c ific c h a ra c te ris tic s b e in g
s e le c te d fo r m u ta tio n . T h e a c tu a l c h a ra c te ris tic s th a t a re u s e d to s e le c t a g e n e fo r a n
in c re a s e d c h a n c e o f m u ta tio n w ill v a ry d e p e n d in g o n th e a p p lic a tio n . In m o s t c a s e s
th e g e n e s th a t w ill h a v e th e e x tra b ia s in fa v o r o f m u ta tio n w ill b e th o s e th a t e x h ib it
s o m e d e fe c t re q u irin g im p ro v e m e n t. A n e x tra w e ig h t v a lu e w ill b e a d d e d to th e
c h a n c e s o f g e n e s e x h ib itin g th a t d e fe c t. In th e c a s e o f th e p ro b le m u n d e r
c o n s id e ra tio n th e a u th o rs d e c id e d to im p le m e n t th e b ia s s o th a t g e n e s re la tin g to
e v e n ts b re a k in g a h a rd c o n s tra in t h a v e a h ig h e r c h a n c e o f b e in g s e le c te d fo r m u ta tio n .
T h e m u ta tio n fu n c tio n w a s b e e n m o d ifie d s o th a t if a g e n e s e le c te d fo r m u ta tio n
re p re s e n ts a n e v e n t th a t d o e s n o t b re a k a h a rd c o n s tra in t th e re is a c h a n c e (e x p re s s e d
a s a p e rc e n ta g e ) th a t a n o th e r g e n e m a y b e s e le c te d . T a b le 3 s h o w s th e re s u lts o f u s in g
a 4 0 % , 5 0 % , a n d 6 0 % b ia s a g a in s t th e m u ta tio n o f p la c e d g e n e s .
F ig . 1 . A n e x a m p le o f p a r t o f a s c h e d u le b e f o r e w rite b a c k ( le ft) a n d a fte r (r ig h t)
w r ite b a c k
2 .4
P r e m a tu r e C o n v e r g e n c e a n d R e g e n e r a tio n
W ith in th e E A it is d e s ir a b le to m a in ta in a d iv e r s e p o p u la tio n . W h e n a g o o d s o lu tio n
is fo u n d , th is in d iv id u a l w ill b e g in to d o m in a te a n d th u s r e d u c e th e b io -d iv e r s ity
w ith in th e p o p u la tio n . T h e c o n v e r g e n c e o f m a n y p o p u la tio n m e m b e r s to a s im ila r
p o in t w ith in th e s e a r c h s p a c e is fe a tu r e o f a ll E A s . It is d e s ir a b le th o u g h , to tr y to
p r e v e n t c o n v e r g e n c e u n til th e E A b e in g s to c o n v e r g e o n th e o p tim u m s o lu tio n . O n e
m e th o d o f e lim in a tin g th e p r e m a tu r e c o n v e r g e n c e e ffe c t is to im p le m e n t a
r e g e n e ra tio n o p e ra to r . W ith in th e r e g e n e ra tio n o p e r a to r a n u m b e r o f in d iv id u a ls h a v e
th e ir g e n e s ra n d o m ly r e s e t. T h u s in th e c a s e o f th e s c h e d u lin g a lg o r ith m u n d e r
c o n s id e r a tio n a n u m b e r o f m e m b e r s o f th e p o p u la tio n w ill h a v e th e tim e s o f th e ir a ll
e v e n ts ra n d o m ly m u ta te d . C o n s id e r a tio n m u s t b e g iv e n to th e fr e q u e n c y o f th e u s a g e
o f th e o p e r a to r a n d th e p e r c e n ta g e o f th e p o p u la tio n in v o lv e d . If th e r e g e n e r a tio n
O p tim is in g a n E v o lu tio n a ry A lg o rith m
o p e r a to r is u tilis e d to o fr e q u e n tly th e n th
it a ffe c ts to o m a n y in d iv id u a ls w ith a
g e n e tic m a te ria l.
T h e o p e ra to r a s im p le m e n te d in itia lly ,
ra n d o m ly e v e ry 5 0 g e n e ra tio n s . T h e re
m u ta tio n o p e ra to r to a ll o f th e g e n e s in
in d iv id u a l. T h e tim e s ta k e n to p ro d u c e
s h o w n in ta b le 3 . It m a y b e s e e n th e re fo r
(a p p ro x im a te ly 4 0 % ) s a v in g in tim e .
3 .
fo r S c h e d u lin g
3 1 1
e p o p u la tio n w ill n o t h a v e tim e to e v o lv e , if
h ig h fitn e s s th e n th e E A m a y lo o s e u s e fu l
re in itia lis e s
g e n e ra tio n
th e s c h e d u
a fe a s ib le
e th a t u s in g
th e
o p e
le ,
sc h
re g
lo w e r 5 0 % o f th
ra to r in p ra c tic e
th u s c re a tin g a n
e d u le u s in g re g e
e n e ra tio n o ffe rs
e p o p u la tio n
a p p lie s th e
e w , ra n d o m
n e ra tio n a re
a s u b s ta n tia l
S e ttin g th e E A P a r a m e te r s
T h e p a ra m e te rs
m u ta tio n ra te a n d
E A . C o n v e n tio n a
w ith re fe re n c e to
a lte rn a tiv e m e th o
in itia l p a ra m e te rs
u s
re
lly
p
d s
w
e d w
g e n e
th e
re v io
o f a
a s c o
ith in th e E A to s p e c if
ra tio n h a v e th e b ig g e s t b e
E A d e s ig n e r u s u a lly s e ts
u s e x p e rim e n ta tio n . T h e
rriv in g a t a se t o f p a ra m e
n s tru c te d a n d w ill b e re fe
y
p o p u la tio n s iz e ,
a rin g o n th e s u c c e s s
th e s e p a ra m e te rs to
a u th o rs d e c id e d to
te rs . A v e rs io n o f th
rre d to a s v e rs io n C P
c ro
o r
v a
in v
e E
.
sso v
fa ilu
lu e s
e s tig
A u
e r ra te ,
re o f a n
d e c id e d
a te tw o
s in g th e
T a b le 3 . T im e in s e c o n d s to p r o d u c e a fe a s ib le s c h e d u le u s in g d a ta s e t 2
5
4
6
R e g e
a n d 5
3 .1
N o
0 %
0 %
0 %
n e r
0 %
R u
B ia
B ia
B ia
B ia
a tio
b ia
n
1
s
s
s
s
n
5 7
1 1
2 3
5
0
0
8
1
5 7
2
1 1
7 8
3 5
1 6 1
1 8
8
0
1 6
3
5
9
1
1
4
3
1 9
1 7
3 4
3 7
7 2
4
5 0
2 8
1 6 1
9 3
2 2
2
0
4
4
8
1 4
1 5
1
1
1 6
5
5 7
9 9
6 3
0 8
3 1
A v e
7
5
7
6
3
ra
6 8
6 3
1 6
4 6
4 3
g e
.5
.3
.3
.9
.6
s
M o d ify in g th e P a r a m e te r s a t R u n T im e
T h e firs t m e th o d in v e s tig a te d is b a s e d o n a s im p lifie d v e rs io n o f th e s y s te m d e s c rib e d
in [1 3 ] a n d [5 ]. E a c h o f th e s e p a ra m e te rs s ta rts o ff s e t to a lo w v a lu e , w h e n th e
in d iv id u a l a t th e h e a d o f th e p o p u la tio n (ie th e m o s t fit in d iv id u a l) c h a n g e s th e
p a ra m e te r th a t in v o k e s th e o p e ra to r re s p o n s ib le fo r c re a tin g th a t in d iv id u a l is
in c re a s e d s lig h tly . T h is is s im p lifie d fro m th e o rig in a l s y s te m p ro p o s e d b y D a v is [1 3 ]
w h ic h n o t o n ly to o k in to a c c o u n t th e m o s t fit in d iv id u a l, b u t a ls o th e o p e ra to rs u s e d to
c re a te th e a n c e s to rs o f th a t in d iv id u a l fo r s e v e ra l g e n e ra tio n s p re v io u s to th e c u rre n t
g e n e ra tio n .
A n e w v e rs io n o f th e E A in c o rp o ra tin g th e a b ility to m o d ify its p a ra m e te rs a t ru n
tim e w a s c o n s tru c te d . W h e n e v e r th e m o s t fit in d iv id u a l c h a n g e s th e p a ra m e te r
c o n tro llin g th e c re a tin g o p e ra tio n is m o d ifie d to in c re a s e th e u s a g e o f th e o p e ra to r th a t
c a u s e d th e im p ro v e m e n t. F o r in s ta n c e if a n e w m o s t-fit in d iv id u a l is c re a te d b y th e
m u ta tio n o p e ra to r th e n th e m u ta tio n ra te is in c re a s e d b y 0 .1 % , in th e e v e n t o f th e
c ro s s o v e r o p e ra tio n c re a tin g a n e w m o s t-fit in d iv id u a l th e n th e c ro s s -o v e r ra te is
3 1 2
N . U rq u h a rt, K . C h is h o lm , a n d B . P a e c h te r
in c re a s e d b y 0 .1 % . In th e e v e n t o f th e re g e n e ra tio n o p e ra to r b e in g re s p o n s ib le th e n
th e fre q u e n c y o f th e re g e n e ra tio n o p e ra to r is in c re a s e d b y 1 g e n e ra tio n . A n e w
v e rs io n o f th e E A im p le m e n tin g th is o p e ra to r w a s c o n s tru c te d , it w ill b e re fe rre d to a s
V P 1 . In itia l v a lu e s fo r th e m u ta tio n ra te a n d c ro s s o v e r ra te a re 7 5 % a n d 5 0 % w ith th e
re g e n e ra tio n in te rv a l b e in g s e t a t 5 0 g e n e ra tio n s , a s u s e d in C P . T h e re s u lts o b ta in e d
w h e n u tilis in g d a ta s e t2 m a y b e s e e in ta b le 4 .
In o rd e r to a tte m p t to d e c re a s e th e tim e ta k e n to b u ild a s c h e d u le , it w a s d e c id e d to
m o d ify th e s ta rtin g v a lu e s , th u s a llo w in g th e E A its e lf to m a k e o n e o r th e o th e r
b e c o m e th e d o m in a n t o p e ra to r. A s e c o n d v e rs io n (V P 2 ) w ith s ta rtin g m u ta tio n a n d
c ro s s o v e r ra te s w e re s e t to 1 0 % a n d th e s ta rtin g re g e n e ra tio n in te rv a l w a s s e t to 1 0 0 0
g e n e ra tio n s , th u s n o o p e ra to r is a llo w e d to d o m in a te a t th e s ta rt.
T a b le 4 . F itn e s s o f s c h e d u le s g e n e r a te d u s in g d a ta s e t 2 . ( C P – C o n s ta n t p a r a m e te r s ; V P 1 –
V a r ia b le p a r a m e te r s , s ta rts w ith 7 5 % c r o s s o v e r a n d 5 0 % m u ta tio n r a te s ; V P 2 – V a ria b le
p a r a m te r s s ta r ts w ith 1 0 % c r o s s o v e r a n d m u ta tio n ra te s )
R u
C P
V P
V P
n
.
2
1
1
3 4 1 6 2 .9
3 4 5 9 8 .1
3 4 2 1 4 .1
2
3 4 0 1 9 .4
3 4 4 8 7 .4
3 4 3 8 2 .8
3
3 4 1 9 3 .2
3 4 4 9 7 .6
3 4 4 2 4 .7
4
3 4 0 6 2 .4
3 4 3 5 2 .3
3 4 3 7 2 .8
F ro m th e fig u re s in ta b le 4 it m a y b e s e e n th a t v e rs io n V P 1 g iv e s th e h ig h e s t
e ra g e fitn e s s w h e n b u ild in g s c h e d u le s b a s e d o n d a ta s e t 2 . T h e re s u lts w o u ld
g g e s t th a t in o rd e r to b e e ffe c tiv e th e ru n tim e p a ra m e te rs m u s t h a v e s ta rtin g v a lu e s
a t a re p ro v e n to p ro v id e a n e ffic ie n t re s u lt w h e n u s e d w ith o u t m o d ific a tio n . It w a s
u n d th a t w h e n s ta rtin g w ith lo w in itia l v a lu e s th e E A is u n a b le to p ro g re s s d u e to
e la c k o r c ro s s o v e r, m u ta tio n o r re g e n e ra tio n .
B a s e d o n th e re s u lts o b ta in e d fro m th e te s tin g o f v e rs io n s V P 1 a n d V P 2 ,
p tim is in g th e p a ra m e te rs a t ru n -tim e d o e s n o t a p p e a r to g iv e a n y a d v a n ta g e s o v e r
s in g p a ra m e te rs th a t th e u s e r h a s a rriv e d a t u s in g “ tria l a n d e rro r” m e th o d s . T h e
e a s o n s fo r fa ilu re m a y b e c o n n e c te d w ith th e in itia l p a ra m e te r v a lu e s . If th e
a ra m e te rs a re in itia lly s e t to s m a ll v a lu e s (a s in v e rs io n V P 2 ), th e re is a la c k o f
c tiv ity w ith in th e G A , a n d th e p a ra m e te rs c a n n o t in c re a s e , a n d th u s in c re a s e th e
c tiv ity w ith in th e a lg o rith m u n til n e w m o s t-fit in d iv id u a ls a re b e in g c re a te d .
a v
su
th
fo
th
o
u
r
p
a
a
3 .2
U s in g a M e ta -L e v e l E A to O p tim ise th e P a r a m e te r s
T h e p a ra m e te rs o f a n E A m a y th e m s e lv e s b e o p tim is e d b y u s in g a n o th e r E A [1 3 ], [4 ],
k n o w n a s a m e ta -E A , to a tte m p t to e v o lv e a s e t o f e ffic ie n t p a ra m e te rs . W ith in th is
m e ta -le v e l E A th e g e n e s re p re s e n t th e p a ra m e te rs o f th e E A th a t w e w is h to o p tim is e
(h e re in a fte r re fe rre d to a s th e p ro b le m E A ). T h e fitn e s s fu n c tio n o f th e m e ta -E A is
b a s e d o n th e re s u lts o f ru n n in g th e p ro b le m E A w ith th e p a ra m e te rs s to re d in e a c h
in d iv id u a l w ith in th e m e ta -E A p o p u la tio n .
T h e p a ra m e te rs o f th e p ro b le m E A to b e o p tim is e d a re lis te d in ta b le 5 . T h e fitn e s s
o f e a c h in d iv id u a l w ith in th e m e ta -E A w a s c a lc u la te d b y ru n n in g th e s c h e d u lin g -E A
u s in g th e p a ra m e te rs e n c o d e d w ith in th a t in d iv id u a l, th e fitn e s s w a s c a lc u la te d a s th e
fitn e s s o f th e fin a l s c h e d u le le s s a tim e v a lu e .
O p tim is in g a n E v o lu tio n a ry A lg o rith m
fo r S c h e d u lin g
3 1 3
T a b le 5 . P a ra m e te rs u se d fo r th e m e ta -E A
P a r a m e te
P o p u la tio
E lite s iz e
M u ta tio n
C ro sso v e
r
V a lu e
1 0
1
4 0 %
5 0 %
n s iz e
D u rin g in itia lis a tio n th e firs t 2 m
p a ra m e te rs u s e d w ith v e rs io n C P
ra n d o m ly .
In o rd e r to a s c e rta in th e a b ility
p a ra m e te rs th e fo llo w in g e x p e rim e
w o rk s ta tio n s w e re a llo w e d to ru n th
th e E A s w a s ru n n in g c o m p le te ly in d
w e re e v a lu a te d , th is w o u ld m e a n
c o n fig u ra tio n s , w h ic h w o u ld b e 1 2 .5
R a te
r R a te
e m b e rs o f th e p o p u la tio n w e re s e t to e q u a l th e
a n d th e re m a in in g m e m b e rs w e re in itia lis e d
o f th e
n t w a s
e m e ta -E
e p e n d e n
e a c h P
m in s p e
m
c a
A
tly
e ta -E A to
rrie d o u t
fo r a p p ro
. In to ta l 3
C e v a lu a te d
r e v a lu a tio n .
e v o
o v e r
x im a
9 6 1
o n
lv e a n e ffic ie n t s e t o
6 d a y s . S ix id e n tic a
te ly 1 3 6 h o u rs , e a c h o
d iffe re n t c o n fig u ra tio n
a v e ra g e 6 6 0 d iffe re n
f
l
f
s
T a b le 6 . P a ra m e te rs r e q u irin g o p tim is a tio n b y th e m e ta -E A
C h ro m o so m e
P o p s iz e
E lite s iz e
M u ta tio n r a te
C ro ss o v e r ra te
B
fa v o
R
R
ia s
u r o
e g e
e g e
m u ta tio n in
f f itn e s s
n e r a te tim e
n e ra te P C
M in
1
0
0
0
F a lse
0
0
M a
2 0
1 0
1 0
1 0
x
N o te
0
0
0
T ru e
1 0 0
1 0 0
I f 0 th e n e litis m is s w itc h e d o f f .
c h a n c e o f a n y o n e in d iv id u a l b e in g m
% c h a n c e o f a n y in d iv id u a l b e in g s e le c
c ro sso v e r.
I f tr u e p r o v id e in c r e a s e c h a n c e s o f b e in
m u ta te d to th e m o s t fit e x a m p le s
N o o f g e n e r a tio n s b e tw e e n re - g e n e r a tio
P e r c e n ta g e o f p o p u la tio n to r e - g e n e ra te
%
u ta te d .
te d f o r
g
n
.
In a ll o f th e ru n s th e m e ta E A fitn e s s in itia lly in c re a s e d ra p id ly , b e fo re s lo w in g . It
s h o u ld b e re m e m b e re d th a t th e g ra p h c o v e rs a p e rio d o f a p p ro x im a te ly 6 d a y s . If it
h a d b e e n p ra c tic a l to h a v e a llo w e d th e m e ta -E A e x p e rim e n t to c o n tin u e th e g ra p h
s u g g e s ts th a t s o m e , if n o t a ll o f th e fitn e s s ra te s w o u ld h a v e c o n tin u e d to g ro w .
A fte r th e m e ta E A e x p e rim e n t h a d fin is h e d th e p a ra m e te r s e t w a s e x tra c te d fro m
th e in d iv id u a l w ith th e h ig h e s t fitn e s s . T h is is s h o w n in ta b le 7 . T h e p e rc e n ta g e o f th e
p o p u la tio n to b e re g e n e ra te d h a s b e e n d e c re a s e d a n d th e in te rv a l b e tw e e n
re g e n e ra tio n in c re a s e d fro m th e s e ttin g s u s e d in v e rs io n C P o f th e G A . T h e
p o p u la tio n a n d th e e lite s iz e s h a v e b o th b e e n in c re a s e d . T h e s e p a ra m e te rs h a v e b e e n
in c o rp o ra te d in a n e w v e rs io n o f th e E A k n o w n a s M P . It is n o w p o s s ib le to c o m p a re
v e rs io n M P w ith v e rs io n V P 1 a n d v e rs io n C P , th e re s u lts a re s h o w n in ta b le 7 .
t
3 1 4
N . U rq u h a rt, K . C h is h o lm , a n d B . P a e c h te r
T a b le 7 . :T h e p a r a m e te r s e v o lv e d u s in g th e m e ta - G A
P a ra m e te r
C ro s s o v e r R a te
M u ta tio n R a te
R e g e n e ra tio n In te rv a l
R e g e n e ra tio n % o f p o p u la tio n
P o p u la tio n s iz e
E lite s iz e
B ia s m u ta tio n in fa v o u r o f fitn e s s
V a lu e
6 8
3 0
7 4
2 9
9 5
4
Y e s
T a b le 8 . A c o m p a r is o n o f th e fitn e s s v a lu e o f s c h e d u le s
R u n
C P
V P 1
M P
1
3 4 5 0 0 .0 8
3 4 4 8 5 .0 8
3 4 5 4 5 .0 5
2
3 4 7 1 5 .8
3 4 5 9 8 .1
3 4 4 5 4 .8
4 .
C o n c lu s io n s a n d F u tu r e W
4 .1
S o m e G e n e r a l C o n c lu s io n s o n th e E A
3
3 4 1 6 2
3 4 4 8 7 .4
3 4 4 3 5 .8
4
3 2 8 9 1
3 4 5 1 8
3 4 6 1 1
o r k
T h e re p re s e n ta tio n u s e d is a b a s ic d ire c t re p re s e n ta tio n . A lth o u g h m a n y E A
s c h e d u lin g te c h n iq u e s , m a k e e x te n s iv e u s e o f m e m e tic a lg o rith m s , th e b a s ic
s c h e d u lin g p ro b le m o u tlin e d in th is re p o rt w a s s o lv e d w ith o u t re c o u rs e to a p u re ly
m e m e tic a lg o rith m . A n u m b e r o f c h a n g e s fro m th e b io lo g ic a l m o d e l w e re n e c e s s a ry
to a llo w th e s y s te m to p ro d u c e s c h e d u le s w ith a h ig h fitn e s s in a re a s o n a b le tim e .
T h e s e c h a n g e s w e re th e u s e o f a s te a d y -s ta te p o p u la tio n , th e u s e o f re s tric te d m u ta tio n
a n d th e u s e o f L a m a rk ia n w rite b a c k .
T h e m u ta tio n o p e ra to r w a s e x te n s iv e ly m o d ifie d a n d it w a s re s tric te d to o n ly
m u ta tin g e v e n t s ta rt tim e s to w ith in th e s ta rt a n d e n d tim e fo r th a t jo b , ra th e r th a n
a n y w h e re o n th e s c h e d u le . T h e m u ta tio n o p e ra to r w a s a ls o g iv e n a b ia s to d is c o u ra g e
th e m u ta tio n o f g e n e s th a t re p re s e n te d a lre a d y p la c e d e v e n ts . It m a y b e c o n c lu d e d th a t
th e m u ta tio n o p e ra to r w o rk s m o s t e ffe c tiv e ly w h e n it is re s tric te d fro m p ro d u c in g
s c h e d u le s th a t a re o b v io u s ly in fe a s ib le . T h e m u ta tio n o p e ra to r’s p o w e r s te m s fro m its
a b ility to c re a te c h a n g e s th a t a re n o t in flu e n c e d b y e x is tin g g e n e tic m a te ria l w ith in th e
p o p u la tio n . T h is a b ility to m a k e “ ra d ic a l” c h a n g e s c a n c a u s e a b ig g e r d e c re a s e th a n
in c re a s e in fitn e s s , if th e c h a n g e s a re d e trim e n ta l to th e o v e ra ll p ro b le m th a t th e E A is
a tte m p tin g to s o lv e .
T h e d ire c te d m u ta tio n o p e ra to r s h o w e d a s ig n ific a n t im p ro v e m e n t, th e a v e ra g e
tim e s to b u ild a fe a s ib le s c h e d u le d e c re a s e d fro m 7 6 8 .5 to 5 6 3 .3 s e c o n d s fo r v e rs io n s
w ith n o b ia s in fa v o u r o f b a d ly p la c e d e v e n ts , a n d 5 0 % b ia s re s p e c tiv e ly . T h e
s ta n d a rd d e v ia tio n in tim e ta k e n w ith o u t b ia s w a s 6 3 2 .3 s e c o n d s , b u t a d d in g a 5 0 %
b ia s , th is d e c re a s e s to 5 8 4 s e c o n d s . T o fu rth e r e x a m in e th e e ffe c ts o f ta rg e te d
O p tim is in g a n E v o lu tio n a ry A lg o rith m
fo r S c h e d u lin g
3 1 5
m u ta tio n tw o m o re v e rs io n s o f th e E A w e re c o n s tru c te d w ith 4 0 % a n d 6 0 % b ia s
a g a in s t th e m u ta tio n o f p la c e d e v e n ts . A m o re d e ta ile d c o m p a ris o n o f ty p ic a l ru n s is
s h o w n in fig u re 2 .
T h e fitn e s s fu n c tio n is o f p a ra m o u n t im p o rta n c e to th e E A , a n d a s u b tle a lte ra tio n
to th e fitn e s s fu n c tio n w ill ra d ic a lly a ffe c t th e o u tp u t o b ta in e d fro m th e E A . T h e e x a c t
v a lu e s u s e d a s w e ig h ts w ith in th e fitn e s s fu n c tio n w e re e s ta b lis h e d b y “ tria l a n d
e rro r” . It w a s n o t p o s s ib le to a llo w th e m e ta -E A to o p tim is e th e fitn e s s fu n c tio n
v a lu e s d u rin g th e m e ta -E A e x p e rim e n t. If th is h a d ta k e n p la c e it is lik e ly th a t th e
m e ta -E A w o u ld h a v e a lte re d th e fitn e s s fu n c tio n v a lu e s to g iv e a h ig h fitn e s s
re g a rd le s s o f th e a c tu a l fitn e s s o f th e s c h e d u le . It is a c k n o w le d g e d th a t a m o re
e ffic ie n t s e t o f w e ig h ts m a y e x is t. S o m e re s e a rc h in to e s ta b lis h in g a n o p tim u m s e t o f
w e ig h ts fo r a fitn e s s fu n c tio n w ith in a d ra fts p ro g ra m [4 ] h a s b e e n c a rrie d o u t w ith
c o n s id e ra b le s u c c e s s
T h e u s e o f th e re -g e n e ra tio n o p e ra to r p ro v id e d a m a rk e d im p ro v e m e n t. In itia lly
th is o p e ra to r a p p e a re d to b e v e ry d is ru p tiv e a n d c o u ld p o s s ib ly c a u s e th e lo s s o f
g e n e tic m a te ria l. T h e re -g e n e ra tio n o p e ra to r re q u ire s s k ill in its u s a g e , if to o m a n y
in d iv id u a ls a re re -in itia lis e d o r th e o p e ra to r in v o k e d to o fre q u e n tly th e n u s e fu l g e n e tic
m a te ria l w ill b e d e s tro y e d . F u rth e r re s e a rc h to fin d if th e re is a n o p tim u m p e rc e n ta g e
o f p o p u la tio n to re -in itia lis e a n d a n o p tim u m fre q u e n c y o f u s e w o u ld b e d e s ira b le .
T h e m e ta -E A lo w e re d th e p e rc e n ta g e o f th e p o p u la tio n to b e re -in itia lis e d to 2 9 % .
T h e re -g e n e ra tio n in te rv a l w a s in c re a s e d to 7 4 g e n e ra tio n s .
4 .2
F ig
M P
d e c
v e r
re p
S o m e C o n c lu s io n s o n P a r a m e te r O p tim is a tio n
u re 3 s h o w s th e fitn e s s g ro w th ra te s fo r ty p ic a l ru n s u s in g v e rs io n s C P , V P 1 a n d
, w h e n ru n n in g d a ta s e t 2 . V e rs io n C P re p re s e n ts th e E A ru n n in g w ith p a ra m e te rs
id e d b y th e u s e r u s in g “ tria l a n d e rro r” , v e rs io n V P 1 re p re s e n ts th e b e s t o f th e
s io n s th a t a llo w e d th e p a ra m e te rs to b e o p tim is e d a t ru n tim e , fin a lly v e rs io n M P
re s e n ts th e re s u lt o f th e m e ta -E A .
W h e n try in g to p ic k a “ b e s t” v e rs io n o f th e E A it is im p o rta n t to n o tic e th a t
d iffe re n t v e rs io n s d is p la y d iffe re n t q u a litie s . F o r in s ta n c e v e rs io n C P p ro d u c e d
s c h e d u le s w ith a s lig h tly h ig h e r a v e ra g e fitn e s s (fo r d a ta s e t 3 , s e e ta b le 1 0 ) th a n M G ,
b u t th e tim e re q u ire d to p ro d u c e th e s c h e d u le s is s ig n ific a n tly le s s (s e e ta b le 1 1 ). It is
th e c o n c lu s io n o f th e a u th o rs th a t v e rs io n M P a p p e a rs to re p re s e n t th e b e s t “ tra d e -o ff”
in te rm s o f a c h ie v in g a n a c c e p ta b le le v e l o f fe a s ib le s c h e d u le s , a c h ie v in g a h ig h e r
a v e ra g e fitn e s s th a n th e v e rs io n s a n d c o m p le tin g th e ru n s in s ig n ific a n tly le s s tim e .
V e rs io n s V P 1 a n d V P 2 w e re u n a b le to p ro d u c e s c h e d u le s th a t w e re e ffic ie n t w h e n
c o m p a re d w ith th e re s u lts o b ta in e d u s in g C P a n d M P (s e e ta b le 9 ). T h e m a in re a s o n
fo r th is a p p e a rs to b e d u e to th e s u b tle n a tu re o f th e lin k s b e tw e e n th e p a ra m e te rs .
A d ju s tin g th e p a ra m e te rs b a s e d o n ly o n th e m o s t fit in d iv id u a l in th e p o p u la tio n is n o t
n e c e s s a rily p ro d u c tiv e . T h e e n tire p o p u la tio n c o n trib u te s to e a c h o th e r fitn e s s th ro u g h
c ro s s o v e r th e re fo re a lth o u g h o n e o p e ra to r m a y a p p e a r to b e d o m in a n t in te rm s o f
c re a tin g th e in d iv id u a l w ith th e h ig h e s t fitn e s s a n d th is o p e ra to r m a y n o t b e d o m in a n t
th ro u g h o u t th e p o p u la tio n .
3 1 6
N . U rq u h a rt, K . C h is h o lm , a n d B . P a e c h te r
35000
Fitnes
34500
34000
33500
33000
2704
2545
2386
2227
2068
1909
1750
1591
1432
1273
1114
955
796
637
478
319
1
32000
160
32500
Generation
No bias.
50 % bias
40 % bias
F ig . 2 . F itn e s s G r o w th w ith 0 ,4 0 ,5 0 a n d 6 0 %
60 % bias
m u ta tio n b ia s u s in g d a ta s e t 2
A m a jo r c o n c lu s io n th a t m a y b e d ra w n fro m th e w o rk is th a t th e o p tim is a tio n a n d
s e le c tio n o f th e E A p a ra m e te rs m a y b e b e s t c a rrie d o u t b y a m e ta -E A . In th is c a s e th e
m e ta -E A p a ra m e te rs re c o rd e d a s m a ll b u t s ig n ific a n t in c re a s e in p e rfo rm a n c e fro m
th o s e o b ta in e d u s in g th e m o re tra d itio n a l “ tria l a n d e rro r” m e th o d s . W ith in v e rs io n
C P , th e v e rs io n th a t u s e d th e “ tria l a n d e rro r” p a ra m e te rs 8 6 % o f a ll th e E A ru n s
re s u lte d in fe a s ib le s c h e d u le s , u s in g v e rs io n M P th a t in c re a s e d to 9 3 % .
It m a y b e s e e n th a t u s in g p a ra m e te rs d e riv e d b y th e m e ta -E A it is p o s s ib le to
a c h ie v e a fa s te r a n d h ig h e r g ro w th ra te . O p tim is in g w ith th e m e ta -E A g iv e s s u p e rio r
re s u lts to th o s e a c h ie v e d u s in g p a ra m e te rs d e c id e d b y th e u s e r (v e rs io n C P ) o r b y
m o d ify in g th e p a ra m e te rs a t ru n tim e (V P 1 a n d V P 2 , s e e fig u re 4 )
35000
Fitnes
34500
CP
34000
VP1
MP
33500
33000
1111
963
889
815
741
667
593
519
445
371
297
223
149
1
75
32000
1037
32500
Generation
F ig . 3 . F itn e s s g ro w th c o m p a ris o n fo r ty p ic a l ru n s o f v e rs io n s C P , V P 1 a n d M P
O p tim is in g a n E v o lu tio n a ry A lg o rith m
fo r S c h e d u lin g
3 1 7
T a b le 9 . T o ta l jo b s p la c e d (a v e ra g e d o v e r 1 0 ru n s )
D a ta se t 1 (1 6 jo b s )
1 6
1 5 .8
1 5 .9
1 6
C P
V P 1
V P 2
M P
D a ta
1 4
1 4
1 2
1 5
se t 2 (1 5 Jo b s)
.7
.4
.1
D a ta
1 7
1 7
1 3
1 7
se t 3 (1 8 jo b s )
.7
.6
.8
T a b le 1 0 . S c h e d u le fitn e s s (A v e ra g e d o v e r 1 0 ru n s )
C P
V P 1
V P 2
M P
D a ta se t 1
7 4 6 2 6 .5
7 4 5 0 0 .1 3
7 4 5 9 2 .8 1
7 4 6 0 0 .8 1
D a ta
3 4 4 3
3 4 2 6
3 3 9 5
3 4 4 7
se t 2
3 .3 8 3
0 .0 4 1
0 .8 6
8 .6 3 4
D a ta
5 3 6 9
5 3 6 2
5 3 0 0
5 3 6 8
se t 3
8 .6 0
8 .3 3
9 .9 2
1 .2 6
T a b le 1 1 . T im e ta k e n f o r r u n ( s e c o n d s a v e r a g e d o v e r 1 0 r u n s )
C P
V P 1
V P 2
M P
D a ta se t 1
2 0 7 .2
2 1 4 .5
4 1 2 .1
1 5 4 .3
D a ta se t 2
8 0 5
6 7 8 .4
1 6 5 7 .6
3 6 7 .5
D a ta se t 3
1 3 8 8 .3
1 2 6 0 .9
2 4 3 1
3 3 8
400 214.5
350 412.1
300 154.3
250
CP
200
VP1
150
VP2
100
VP3
50
MP
0
CP
VP1
VP2
VP3
MP
F ig . 4 . A v e r a g e tim e ( s e c o n d s ) to b u ild a f e a s ib le s c h e d u le
4 .3 F u tu r e W o r k
T h e a u th o rs p la n to a p p ly m e ta -E A p rin
s y s te m u s e d to s o lv e a ro u tin g p ro b le m . R
ra is e d th e p o s s ib ility o f u s in g d iffe re n tia l
u s e d w ith in th e fitn e s s fu n c tio n [2 ]. It
e v o lu tio n to o p tim is e th e p a ra m e te rs fo r
m e ta -E A is th e p ro c e s s in g tim e re q u ire d ,
re s u lts to th e m e ta -E A , b u t w ith fe w e r C P U
c ip le s to th e o p tim is in g o f a n E A b a s e d
e c e n t w o rk in to d iffe re n tia l e v o lu tio n h a s
e v o lu tio n to o p tim is e th e w e ig h tin g v a lu e s
m a y a ls o b e p o s s ib le to u s e d iffe re n tia l
th e w h o le E A . A m a jo r p ro b le m w ith th e
d iffe re n tia l e v o lu tio n m a y a c h ie v e s im ila r
c y c le s b e in g re q u ire d .
3 1 8
N . U rq u h a rt, K . C h is h o lm , a n d B . P a e c h te r
R e fe r e n c e s
[ 1 ] P r o d u c tio n S c h e d u lin g a n d R e s c h e d u lin g w ith g e n e tic a lg o rith m s . B ie r w ir th C , M a ttf e ld D .
E v o lu tio n a r y C o m p u ta tio n v o lu m e 7 , N o 1 . M I T P re s s 1 9 9 9 .
[ 2 ] C o - E v o lv in g D ra u g h ts S tra te g ie s w ith D if fe r e n tia l E v o lu tio n , C h a p 9 p p 1 4 7 - 1 5 8 in
N e w I d e a s in O p tim is z a tio n , C o r n e D , D o r ig o M , G lo v e r F E d s . M c G r a w - H ill 1 9 9 9 .
[ 3 ] B u ild in g a n d O p tim is in g a S c h e d u lin g G A . B S c H o n o u rs d is s e r ta tio n . U r q u h a r t, N
C h is h o lm , K .( s u p e r v is o r ) . N a p ie r U n iv e r s ity , E d in b u r g h 1 9 9 8 .
[ 4 ] M a c h in e L e a r n in g U s in g a G e n e tic A lg o rith m to O p tim is e a D ra u g h ts P r o g r a m B o a r d
E v a lu a ti o n F u n c t io n . C h i s h o lm K .J , B r a d b e e r P .V .G .. P r o c e e d i n g s o f I E E E I C E C ’9 7 ,
I n d ia n a p o lis , U S A , 1 9 9 7 .
[ 5 ] A n in tr o d u c tio n to G e n e tic A lg o rith m s , M itc h e ll, M . M I T p r e s s 1 9 9 6 .
[ 6 ] . E x te n s io n s to a M e m e tic T im e ta b lin g S y s te m . P a e c h te r B , N o r m a n M , L u c h ia n H . P r a c tic e
a n d th e o r y o f A u to m a te d T im e ta b lin g , B u r k e a n d R o s s E d s . S p rin g e r V e rla g 1 9 9 6 .
[ 7 ] E v o lu tio n a r y C o m p u ta tio n , F o g e l D B . I E E E P r e s s 1 9 9 5 .
[ 8 ] S p e c ia lis e d R e c o m b in a tiv e O p e ra to rs f o r T im e ta b lin g P r o b le m s , B u r k e E , E llim a n D ,
W e a r e R . P r o c e e d in g o f E v o lu tio n a r y C o m p u tin g A I S B W o r k s h o p S h e ffie ld U K A p ril 1 9 9 5 e d
F o g a r ty , T . S p r in g e r - V e rla g 1 9 9 5 .
[ 9 ] O p tim is in g a P re s e n ta tio n T im e ta b le U s in g E v o lu tio n a r y A lg o rith m s . P a e c h te r B . L e c tu r e
N o te s I n C o m p u te r S c ie n c e N o 8 6 4 , S p rin g e r- V e rla g 1 9 9 4 .
[ 1 0 ] T w o s o lu tio n s to th e G e n e r a l T im e ta b le P r o b le m U s in g E v o lu tio n a r y M e th o d s , P a e c h te r
B , C u m m in g A , L u c h ia n H , P e tr iu c , M . P r o c e e d in g s o f th e I E E E W o r ld C o n g r e s s o n
C o m p u ta tio n a l I n te llig e n c e , J u n e 1 9 9 4 .
[ 1 1 ] A C a s e f o r L a m a r c k ia n E v o lu tio n , A c k le y D H , L ittm a n M L . A r tific ia l L ife I II, L a n g to n
C e d . A d d is o n -W e sle y 1 9 9 4 .
[ 1 2 ] G e n e tic A lg o r ith m s in S e a r c h O p tim is a tio n a n d M a c h in e L e a r n in g , G o ld b e r g D . A d d is o n W e sle y 1 9 8 9 .
[ 1 3 ] A d a p tin g O p e ra to r P ro b a b ilitie s in G e n e tic A lg o r ith m s , D a v is L . P r o c e e d in g s o f th e th ir d
I n te r n a tio n a l C o n fe r e n c e o n G e n e tic A lg o r ith m s . S c h a f f e r J . e d . M o r g a n K a u fm a n n 1 9 8 9 .
[ 1 4 ] O p tim is a tio n & C o n tr o l P a r a m e te r s f o r G e n e tic A lg o rith m s , G r e fe n s te tte J . I E E E
T r a n s a c tio n s o n S y s te m s , M a n a n d C y b e r n e tic s 1 9 8 6 .
[ 1 5 ] F a s t P r a c tic a l E v o lu tio n a r y T im e ta b lin g .. C o r n e D , R o s s P , F a n g H .. L e c tu r e N o te s I n
C o m p u te r S c ie n c e N o 8 6 4 , S p r in g e r - V e r la g 1 9 9 4 .
[ 1 6 ] A d a p tin g o p e r a to r s e ttin g s in G e n e tic A lg o rith m s . T u s o n A , R o s s P . E v o lu tio n a r y
C o m p u ta tio n V o l 6 N o 2 . M a s s a c h u s e tts I n s titu te o f T e c h n o lo g y .
O n - l i n e E v o l u t i o n o f C o n t r o l f o r a Fo u r - Le g g e d R o b o t
Us i n g G e n e t i c P r o g r a m m i n g
B jö rn A n d e rs s o n , P e r S v e n s s o n , M a ts N o rd a h l a n d P e te r N o rd in
C o m p le x S y s te m s , C h a lm e rs U n iv e rs ity o f T e c h n o lo g y ,
S E -4 1 2 9 6 G o th e n b u rg , S w e d e n
{t f e m n , n o r d i n }@ f y . c h a l m e r s . s e
A b s
to w
G P
R /C
tio n
1
tr a c t. W
a lk d y n
s y s te m .
se rv o s.
a n d th e
e e v o lv e a ro b o tic c o n tro lle r fo r a fo u ra m ic a lly . E v o lu tio n is p e rfo rm e d o n -lin
T h e ro b o t h a s e ig h t d e g re e s o f fre e d o m
D iffe re n t w a lk in g s tra te g ie s a re s h o w n
e v o lv in g s y s te m is ro b u s t a g a in s t m e c h a
le g g e d re a l ro b o
e b y a lin e a r m
a n d is b u ilt fr
b y th e ro b o t d
n ic a l fa ilu re s .
t e n
a c h
o m
u rin
a b lin g it
in e c o d e
s ta n d a rd
g e v o lu -
I n tr o d u c tio n
R o b o ts o n le g s c o n s titu te b o th o n e o f th e la rg e s t p o te n tia ls a n d o n e o f th e la rg e s t c h a lle n g e s fo r in te llig e n t ro b o tic c o n tro l. A p p lic a tio n s a re n u m e ro u s in a ll e n v iro n m e n ts
a c c e s s ib le to h u m a n s a n d a n im a ls b u t in a c c e s s ib le to w h e e le d a u to n o m o u s a g e n ts . In
g e n e ra l th e fle x ib ility o f th e le g g e d ro b o t in c re a s e s w ith d e c re a s in g n u m b e r o f le g s ,
u n fo rtu n a te ly s o d o e s a ls o c o n tro l c o m p le x ity . It is le s s c o m p lic a te d to c o n tro l a ro b o t
w ith s ix le g s o r m o re , s in c e th e ro b o t c a n h a v e fo u r le g s o n th e g ro u n d a ll th e tim e
p r o v i d i n g s t a b l e s t a t i c b a l a n c e [ 2 , 4 , 9 ] . A l r e a d y t h e f o u r -l e g g e d c a s e b e c o m e s m o r e
d i f f i c u l t . A f o u r -l e g g e d c r e a t u r e c a n c r a w l l i k e a t u r t l e w i t h p a r t i a l s u p p o r t b y i t s b o d y
o r b y a ta il. In th is w a y it is p o s s ib le to w a lk o n fo u r le g s w ith s ta tic b a la n c e . S ta tic
b a la n c e m e a n s th a t th e a g e n t is in b a la n c e a t a ll m o m e n ts s c a rify in g s o m e e ffic ie n c y
a n d fle x ib ility b y d ra g g in g th e ir b o d y o v e r th e g ro u n d . H o w e v e r, it is m o re a d v a n ta g e o u s in g e n e ra l to u s e a d y n a m ic w a lk w h e re th e a g e n t w ill fa ll if in te rru p te d in th e
m id d le o f a m o v e m e n t.
W e h a v e e v o l v e d t h e f i r s t c o n t r o l l e r f o r a f o u r -l e g g e d r o b o t , l e a r n u s i n g a g e n e t i c
p ro g ra m m in g s y s te m a n d a re a l ro b o t. T h e e v o lu tio n o f b e h a v io r p a s s e s th ro u g h s e v e ra l s ta g e s s ta rtin g w ith s im p le p a d d lin g b e h a v io r, c o n tin u in g th ro u g h c ra w lin g ,
" c a m e l w a lk " , a n d fin a lly g a llo p in g w ith d y n a m ic e le m e n ts in th e w a lk .
2
E x p e r im e n ts
S in c e w e a re u s in g a re a l ro b o t th e e x p e rim e n t c o n s is ts o f o n e h a rd w a re p a rt, th e
ro b o t, a s w e ll a s a s o ftw a re p a rt, th e G P s y s te m . T h e re is a ls o a c o m m u n ic a tio n p a rt o f
S . C a g n o n i e t a l. ( E d s .) : E v o W o r k s h o p s 2 0 0 0 , L N C S 1 8 0 3 , p p . 3 1 9 − 3 2 6 , 2 0 0 0 .
S p rin g e r-V e rla g B e rlin H e id e lb e rg 2 0 0 0
3 2 0
th
se
re
th
B . A n d e rs s o n e t a l.
e s y s te m
n so rs. M
a so n w e
e ro b o t b
, se n d
o st e x
n e e d e
a c k if
in g
p e r
d a
it h
c
im
s
a
o n tro
e n ts
p e c ia
d a d v
l c
w
l a
a n
o m
e re
u to
c e d
m a n d
ru n o
m a tic
to o f
to th e a c tu a to rs a n d re c e iv in g fe e d b a c k fro m th e
n a lo n g tim e s c a le o f a b o u t 2 0 h o u rs a n d fo r th is
s y s te m th a t m o n ito rs th e e x p e rim e n t a n d p u lle d
a r.
2 .1 T h e f o u r le g g e d r o b o t
T h e
1 . T
N c m
w ith
n e e d
fo u r-le g
h e se rv o
. T h e se
a s e ria l
e d fo r th
g e d ro b o t is b u ilt w ith e ig h t s ta n
s a re o f o n e o f th e s m a lle s t a v a
rv o s a re c o n tro lle d b y a s ta n d a rd
p o rt to c o n tro l u p to e ig h t s e rv o
is e x p e rim e n t.
d a
ila
se
s.
rd R
b le
rv o
N o
/C s e rv o
b u t s till
c o n tro lle
in p u t s e n
s a s a c tu a to
h a v e a m o m
r c a rd , w h ic
s o rs o n th e
rs, se e
e n tu m
h a llo w
d o g its
F ig
o f
s a
e lf
u re
1 1
P C
a re
F ig u r e 1 : T h e fo u r-le g g e d ro b o t d o g .
F itn e s s is g e n e ra te d th ro u g h a c o m p u te r m o u s e , w h ic h th e ro b o t d ra g s b e h in d its e lf.
P o s itiv e re in fo rc e m e n t is g iv e n b y d ra g g in g th e m o u s e in th e fo rw a rd d ire c tio n . D ra g g in g b a c k w a rd s g iv e s a n e g a tiv e fitn e s s s ig n a l. J u s t m o v in g th e ro b o t g iv e s a s m a ll
p o s itiv e c o n trib u tio n to fitn e s s . T h e e x p e rim e n t s e t-u p c o n s is ts o f a re c ta n g u la r b o x
( 0 .3 m X 1 .2 m ) w h e r e th e a g e n t c a n le a r n f o r w a r d w a lk in g . I f it r e a c h e s th e e n d o f th e
b o x th e n it is a u to m a tic a lly p u lle d b a c k w a rd s to th e o th e r e n d b y a n e le c tric m o to r.
S in g le e x p e rim e n ts ra n a u to m a tic a lly w ith th is s e t-u p fo r m o re th a n 1 8 h o u rs .
2 .2 C o n t r o l s y s t e m
T h
b in
e ra
P C
e c o
a ry
tin g
a n d
n tro l
m a c h
o n a
n o rm
s y s te m is b u ilt
in e c o d e [ 1 ,5 ] . T
s e t o f re g is te rs o
a l m o u se c o m m
a r
h e
f a
u n
o u n d a
e v o lv e
re g is te
ic a tio n
lin e a r g e n
d s tru c tu re
r m a c h in e .
is u s e d fo r
e tic
is a
T h e
fe e
p ro g ra m
lin e a r s
c o n tro l
d b a c k w
m in g
trin g o
s y s te m
h ile th
s y s te m e v
f in s tru c tio
is e x e c u te
e se rv o s a r
o lv
n s
d o
e c
in
o p
n
o n
g
a
-
O n -l i n e E v o l u t i o n o f C o n t r o l f o r a F o u r - L e g g e d R o b o t
tro lle d th ro u g h s e ria l c o m m u
e v o lv e d p ro g ra m s a re p la c e d
w h ile th e o u tp u t v e c to r is w h
p ro g ra m . E a c h in d iv id u a l th a t
e ig h t o u tp u ts a re s e n t a s a n g le
in p u t to th e n e x t ite ra tio n e n a b
th e e ig h th ite ra tio n s (a n d b a s ic
m o v e d fo rw a rd ). S o m e in d iv
w h ic h d o n o t m o v e th e s e rv o s
3 2 1
n ic a tio n w ith a s e rv o c o n tro l c a rd . T h e in p u ts to th e
in to th e re g is te rs o f th e re g is te r m a c h in e (p ro c e s s o r)
a te v e r re m a in s in th e re g is te rs u p o n te rm in a tio n o f th e
is e v a lu a te d is ite ra te d e ig h t tim e s . In e a c h ite ra tio n th e
s to th e s e rv o s a n d th e p re v io u s o u tp u t v e c to r is s e n t a s
lin g a s e q u e n c e to b e e v o lv e d . F itn e s s is m e a s u re d a fte r
a lly c o n s is ts o f th e n u m b e r o f " tic k s " th a t th e m o u s e h a s
id u a ls a re re m o v e d b e fo re e v a lu a tio n , s u c h a s th o s e ,
a t a ll.
F ig u r e 2 : F e e d b a c k th ro u g h c o m p u te r m o u s e
F ig u r e 3 : C a b le s to P C a n d p o w e r s u p p ly
T h e p o p u la tio n c o n s is te d o f 1 0 0 in d iv id u a ls a n d th e g e n e tic o p e ra to rs w h e re c ro s s o v e r a n d m u ta tio n . C ro s s o v e r w a s p e rfo rm e d b o th a s tw o -p o in t s trin g c ro s s o v e r a n d a s
h o m o lo g o u s lin e a r c r o s s o v e r [ 3 ,7 ] . T h e s y s te m u s e d a s m a ll to u r n a m e n t f o r s e le c tio n
3 2 2
B . A n d e rs s o n e t a l.
a n d o p e ra te s u n d e r s te a d y s ta te , s e e F ig u re 4 . T h e fu n c tio n s e t c o n s is te d o f a rith m e tic
o p e r a to r s a n d s q u a r e r o o t. A s in g le c o n s ta n t o f 0 .5 w a s u s e d in th e te r m in a l s e t.
F ig u r e 4 : O n -lin e G P s y s te m
O u r m e th o d fo r u s in g G P w ith a re a l-tim e a p p lic a tio n is b a s e d o n a p ro b a b ilis tic s a m p lin g o f th e e n v iro n m e n t [6 ]. D iffe re n t s o lu tio n c a n d id a te s (p ro g ra m s ) a re e v a lu a te d in
d iffe re n t s itu a tio n s . T h is is u n fa ir b e c a u s e a g o o d in d iv id u a l d e a lin g w ith a h a rd s itu a tio n c a n b e re je c te d in fa v o r o f a b a d in d iv id u a l d e a lin g w ith a v e ry e a s y s itu a tio n . F o r
in s ta n c e , a n in d iv id u a l th a t g e ts s tu c k n e a r th e w a ll o f th e b o x a n d d o e s a g o o d jo b o f
m o v in g a w a y fro m th e w a ll b u t a d v a n c e s little in th e fo rw a rd d ire c tio n m ig h t g e t a
lo w s c o re w h ile a p o o r in d iv id u a l in th e m id d le o f th e b o x m ig h t p e rfo rm b e tte r. O u r
e x p e rie n c e is , h o w e v e r, th a t a g o o d o v e ra ll in d iv id u a l te n d s to s u rv iv e a n d re p ro d u c e
in th e lo n g te rm . T h e s o m e w h a t p a ra d o x ic a l fa c t is th a t s p a rs e tra in in g d a ta s e ts o r
p ro b a b ilis tic s a m p lin g in e v o lu tio n a ry a lg o rith m s o fte n b o th in c re a s e s p e e d to w a rd th e
g o a l a n d k e e p th e d iv e rs ity h ig h e n o u g h to e s c a p e lo c a l o p tim a d u rin g s e a rc h .
3
E v o
d o m
m o u
k in d
b a re
w h e
R e s u lts
lv in g w
a in k n o
se o n a
o f c h a
ly v is ib
re th e r
a lk in
w le d
ro d ,
o tic
le . A
o b o t
g b e h a v io r, th ro u g h c o o rd in a tio n o f e ig h t s e rv o s a n d
g e o r g u id a n c e a n d w h e re th e o n ly fe e d b a c k is fro m
is n o t a triv ia l p ro b le m . T h e firs t e m e rg in g b e h a v io r is
p a d d lin g , w h ic h s lo w ly m o v e s th e ro b o t fo rw a rd e v e n
c o m m o n lo c a l o p tim u m a fte r th e p a d d lin g is a s tra
s ta n d s u p rig h t o n a ll le g s c a re fu lly b a la n c in g a n d m o
w ith o u t a n y
a c o m p u te r
u s u a lly s o m e
th o u g h it is
n g e s tra te g y ,
v in g q u ic k ly
O n -l i n e E v o l u t i o n o f C o n t r o l f o r a F o u r - L e g g e d R o b o t
3 2 3
b a c k a n d fo rth w ith o u t re a lly a d v a n c in g th e ro b o t. It is u n c le a r h o w th is s tra te g y
e m e rg e s b u t it c o u ld b e th e re s u lt o f s o m e la c k o f s y m m e try in th e h a rd w a re a n d
m e a s u re m e n t s y s te m . T h e n e x t s tra te g y is o fte n s o m e k in d o f c ra w lin g s im ila r to th a t
o f tu rtle . O th e r o b s e rv a b le b e h a v io rs a re a " c a m e l w a lk " w h e re th e le g s m o v e in p a ra lle l a n d p a ir w is e o n e a c h s id e . T h e m o s t e ffic ie n t s tra te g y fo r w a lk in g th a t h a s
e v o lv e d is a ls o th e m o s t d iffic u lt to le a rn a n d g a llo p in g o n ly a p p e a rs a fte r m a n y h o u rs
o f tra in in g . H e re b o th th e fro n t le g s a n d b a c k le g s a re p a ra lle l a n d th e ro b o t o fte n
s ta n d s u p o n its b a c k le g s b e fo re p u s h in g fo rw a rd . It u s e s th e d y n a m ic in te ra c tio n w ith
th e h e a v y m o u s e lo d e to a c h ie v e m a x im u m s p e e d fo rw a rd .
E m e rg e n c e o f b e h a v io r d o e s n o t a lw a y s p a s s th ro u g h a ll o f th e s e b e h a v io rs b u t u s u a lly
m o s t o f th e m a p p e a r in o rd e r.
S u m m a r
1
2
3
4
y o
.
.
.
.
f c o m
c h a o
c ra w
" c a m
g a llo
m o n
tic p a
lin g
e l w
p in g
e v o lu tio n o f b e h a v io r o v e r tim e c o u ld lo o k lik e :
d d lin g b e h a v io r
a lk "
w ith d y n a m ic e le m e n ts in th e w a lk
L ik e m a n y e v o lu tio n a ry c o n tro l s
c h a n ic a l fa ilu re . F ig u re 5 s h o w s th
h o u rs . D u rin g e v o lu tio n a tre e s e rv
m a n a g e d to in c re a s e th e s p e e d a fte
c u rre d in g e n e ra tio n 2 0 , 7 5 a n d 1 0
d e c re a s e s ra p id ly b u t th e s y s te m m
a g e d ) h a rd w a re c o n fig u ra tio n .
y s te m s
e e v o lu
o s b ro k
r th e fa
0 . T h e
a n a g e s
w e o b s e rv e s o m e ro b u s tn e s s a g a in s t m e tio n o f s p e e d o v e r 1 1 1 g e n e ra tio n s a n d 1 5
e d o w n b u t th e c o n tro l s y s te m a d a p te d a n d
ilu re s , s e e F ig u re 5 . S e rv o b re a k d o w n s o c fig u re c le a rly in d ic a te s h o w fitn e s s a t firs t
to re le a rn a n d a d a p t to th e c h a n g e d (d a m -
F ig u r e 5 : T h e v e lo c ity (in m /h ) o f th e d o g w h e n c o n tro lle d b y th e (c u rre n tly ) b e s t
in d iv id u a l d u rin g 1 1 7 g e n e ra tio n s . S e rv o fa ilu re o c c u rre d in g e n e ra tio n 2 0 , 7 5 a n d
1 0 0 .
3 2 4
B . A n d e rs s o n e t a l.
4
F u tu r e W
O u
m a
fo u
o n
r in te n
n o id p
r le g g
th e E L
tio n
ro je
e d w
V IS
is
c t,
a lk
h u
o r k
to
se
in
m
u s e th
e F ig u
g a n d
a n o id
e re s u lts
re 6 . H e r
d e m a n d s
c o n firm s
o b ta in e d
e th e a im
m u c h m o
th e fe a s ib
w ith th
is b i-p
re e m p
ility o f
e fo
e d a
h a s
th e
u r-le g
l w a lk
is o n b
a p p ro
g e d
in g
a la
a c h
ro b o t fo r a la rg e r h u , w h ic h is h a rd e r th a n
n c e . P re lim in a ry w o rk
[8 ].
F ig u r e 6 : E L V IS h u m a n o id a b ip e d a l w a lk in g ro b o t
M a k in g th e ro b o t fu lly a u to n o m o u s is a n o th e r a m b itio n fo r th e fu tu re . T h a t
m e a n th a t th e e x p e rim e n ts w ill n o t b e h in d e re d b y a ll th e c a b le s a n d to ta l a u to
w o u ld a ls o b e in te re s tin g fro m a m o re p h ilo s o p h ic a l s ta n d p o in t. F ig u re 7 s h
c h a o tic ro b o t m a d e o f " g a rb a g e " w h ic h h a s le a rn e d to m o v e b y th e s a m e m e c h
a s th e fo u r-le g g e d ro b o t b u t w h ic h is fu lly a u to n o m o u s . T h e o n -b o a rd G P s y s te m
o n a s m a ll, e m b e d d e d P IC -c h ip a n d e v o lv e s b in a ry c o d e fo r th is tin y p ro c e s s o
p la n is to m o v e th is s y s te m to th e w a lk in g e x p e rim e n ts .
w o
n o
o w
a n
u ld
m y
s a
is m
ru n s
r. O u r
O n -l i n e E v o l u t i o n o f C o n t r o l f o r a F o u r - L e g g e d R o b o t
F ig u r e 7 : A n a u to n o m o u s c h a o tic ro b o t m a d e fro m
te m in a n e m b e d d e d c h ip .
5
3 2 5
" tra s h " w ith a n o n -b o a rd G P s y s -
S u m m a r y a n d C o n c lu s io n s
W e h a v e e v o lv e d th e firs t c o n tro lle r fo r a fo u r-le g g e d ro b o t, w h ic h le a rn o n -lin e u s in g
a g e n e tic p ro g ra m m in g s y s te m a n d a re a l ro b o t. T h e e v o lu tio n is o f b e h a v io r p a s s e s
th ro u g h s e v e ra l s ta g e s s ta rtin g w ith s im p le p a d d lin g b e h a v io r, c o n tin u in g th ro u g h
c ra w lin g , " c a m e l w a lk " , a n d fin a lly g a llo p in g w ith d y n a m ic e le m e n ts in th e w a lk . A
h ig h d e g re e o f ro b u s tn e s s is s e e n fo r m e c h a n ic a l fa ilu re - th e s y s te m is o b s e rv e d to
a d a p t to c h a n g e s in th e m e c h a n ic a l c o n fig u ra tio n d u e to c o m p o n e n t fa ilu re .
A c k n o w le d g e m e n t
P e te r N o rd in g ra te fu lly a c k n o w le d g e s s u p p o rt fro m
T F R a n d N U T E K .
R e fe r e n c e s
1 .
B a n z h a f , W ., N o r d in , P . K e lle r , R . E ., a n d F r a n c o n e , F . D .( 1 9 9 8 ) G e n e tic
P r o g r a m m in g : A n In tr o d u c tio n o n th e A u to m a tic E v o lu tio n o f C o m p u te r
P r o g r a m s a n d Its A p p lic a tio n s . M o rg a n K a u fm a n n , G e rm a n y .
2 .
T .
d im
g ra
c h a
B r
e n
m m
p te
o u
s io
in
r 1
g h
n a
g .
4 .
to n
l d
In
A c
, P .
e s ig n
P e te r
a d e m
S . C o a te s , a n d H
w o rld s u s in g L in
B e n tle y , e d ito r, E v
ic p re s s , L o n d o n , U
. J a c k s o n . (1 9 9 9 ) E x p lo rin g th re e d e n m e y e r s y s te m s a n d g e n e tic p ro o lu tio n a ry D e s ig n U s in g C o m p u te rs ,
K ,.
3 2 6
B . A n d e rs s o n e t a l.
3 .
F r a n c o n e F .D ., C o n r a d s M ., N o r d in J .P .a n d B a n z h a f W .( 1 9 9 9 ) H o m o lo g o u s
C ro s s o v e r in G e n e tic P ro g ra m m in g , In P ro c e e d in g s o f: G e n e tic a n d E v o lu tio n a ry C o m p u ta tio n C o n fe re n c e (G E C C O 9 9 ) M o rg a n -K a u fm a n n
4 .
M . A n th o n y
g ra m m in g a p
in g ro b o t. In
b o tic s a n d A
tro n ic a B k s .
L e w is , A n d re
p ro a c h to th e
P ro c e e d in g s
u to m a tio n ,
w H . F
c o n s tr
o f th e
p a g e s
a g g , a n d A la n S o lid u m . (1
u c tio n o f a n e u ra l n e tw o rk
1 9 9 2 IE E E In te rn a tio n a lC
2 6 1 8 -2 6 2 3 , N ic e , F ra n c e ,
9 9
c o
o n
M
2 ) G e n e tic
n tro l o f a w
fe re n c e o n
a y 1 9 9 2 . E
p ro
a lk
R o
le c
-
5 .
N o r d in , J .P . ( 1 9 9 7 ) E v o lu tio n a r y P r o g r a m
In d u c tio n o f B in a ry M a c h in e
C o d e a n d its A p p lic a tio n s . K re h l V e rla g , M u e n s te r, G e rm a n y
6 .
N
to
In
U
7 .
N o r
c h in
T o
O ’R
8 .
N o r d in J . P ., N o r d a h l M . ( 1 9 9 9 ) : A n E v o lu tio n a r y A r c h ite c tu r e F o r A H u m a n o id R o b o t, In P ro c e e d in g o f: T h e F o u rth In te rn a tio n a l S y m p o s iu m o n A rtific ia l L ife a n d R o b o tic s (A R O B 4 th 9 9 ) O ita J a p a n
9 .
G ra h a m F . S p e n c e r. A u to m a tic g e n e ra tio n o f p ro g ra m s fo r c ra w lin g a n d
w a lk in g ( 1 9 9 4 ) . I n K e n n e th E . K in n e a r , J r ., e d ito r , A d v a n c e s in G e n e tic P r o g ra m m in g , c h a p te r 1 5 , p a g e s 3 3 5 -3 5 3 . M IT P re s s , 1 9 9 4 .
o r d in , J .P ., B a n z h a f W .( 1 9 9 7 ) A n O n - lin e M e th o d to E v o lv e B e h a v io r a n d
c o n tro l a M in ia tu re R o b o t in R e a l T im e w ith G e n e tic P ro g ra m m in g : T h e
te rn a tio n a l J o u rn a l o f A d a p tiv e B e h a v io r, (5 ) p p 1 0 7 - 1 4 0 M IT P re s s ,
S A .
d in J .
e C o d
a p p e a
e illy ,
P .,
e f
r
A n
B a n z h a f
o r C IS C
in A d v a
g e lin e , S
W ., a n d
A rc h ite c
n c e s in
p e c to r, M
F ra n
tu re s
G e n
IT -P
c o n e
u s in
e tic
re ss
F . (1 9 9 9 ) E ffic ie n t E v o lu tio n o f M a g B lo c k s a n d H o m o lo g o u s C ro s s o v e r.
P ro g ra m m in g III, (E d s ) L a n g d o n ,
, U S A
O p tim iz e d C o llis io n F r e e R o b o t M o v e S ta te m e n t
G e n e r a tio n b y th e E v o lu tio n a r y S o ftw a r e G L E A M
C h ris tia n B lu m e
F a c h h o c h s c h u le K ö ln
A m S a n d b e rg 1 , D -5 1 6
T e l. + 4 9 /2 2 6 1 /8 1 9 6 -2 9 6 , -3 3
e m a i l : b l u m e
A b
e v o
g e n
tim
ro b
c o m
ro b
, A
4 3
0 o
@ g
str a c t.
T h e G L E A M
a lg o rith m
a
lu tio n a ry m e th o d a p p lic a tio n in th e fie
e ra te s c o n tro l c o d e fo r re a l in d u s tria l
e re la te d d e s c rip tio n o f th e ro b o t m o v e
o t a rm c o n fig u ra tio n s ). T h is in te rn a
m a n d s is m a p p e d to a re p re s e n ta tio n
o t la n g u a g e , w h ic h c a n b e lo a d e d a t th e
b te ilu n g
G u m m e r
r -3 3 2 , F
m . f h - k
G u m m e rsb a c h
sb a c h , G e rm a n y
a x : + 4 9 /2 2 6 1 /8 1 9 6 1 5
o e l n . d e
n d its im p le m e n ta tio n a re a n e w
ld o f ro b o tic s . T h e G L E A M s o ftw a re
ro b o ts . T h e re fo re G L E A M a llo w s a
m e n t (n o t o n ly a s ta tic d e s c rip tio n o f
l re p re s e n ta tio n o f p rim itiv e m o v e
o f m o v e s ta te m e n ts o f a n in d u s tria l
ro b o t c o n tro l a n d e x e c u te d .
I n tr o d u c tio n
T h e re a re m a n y d iffe re n t id e a s a n d p ro c e d u re s a b o u t G e n e tic A lg o rith m s a n d
E v o lu tio n a ry P ro g ra m m in g , b u t o n ly s o m e o f th e s e le a d to im p le m e n ta tio n s w h ic h
s o lv e „ re a l w o rld p ro b le m s “ fo r in d u s tria l a p p lic a tio n s . G L E A M (G e n e tic L e a r n in g
A lg o r ith m a n d M e th o d s ) is a n a lg o rith m b a s e d o n e v o lu tio n a ry s tra te g y , w h ic h w a s
u s e d to im p le m e n t a s o ftw a re to o l fo r s o lv in g u s e fu l a n d n o t o n ly a c a d e m ic p ro b le m s ,
se e [1 ] a n d [2 ].
T h e a im o f th e G L E A M a p p lic a tio n to in d u s tria l ro b o ts w a s a n im p le m e n ta tio n
w ith re s p e c t to p ra c tic a l re q u ire m e n ts a n d re s tric tio n s to p ro o f th e G L E A M m e th o d .
T h e re fo re a firs t p ro to ty p e im p le m e n ta tio n w a s d o n e a t A B B in V ä s te ra s (S w e d e n )
a n d a s e c o n d o n e w ith im p ro v e d fa c ilitie s a t D a im le r C h ry s le r in B e rlin (G e rm a n y ).
B o th im p le m e n ta tio n s in c lu d e th e o u tp u t o f ro b o t la n g u a g e c o d e fo r th e s p e c ific
in d u s tria l ro b o t (A B B IR B 2 4 0 0 a n d K U K A K R 6 ).
T h e E v o lu tio n a r y A lg o r ith m
T h e
a re
d e sc
o n e
p rin c ip le s o f G L E A M in
fo rm u la te d w ith c lo s e
rip tio n in [3 ]). T h e g e n o
s te p o f a p la n to b e e x e c
c lu d in
c o n n e
ty p e o
u te d . T
G L E A M
g its g e n e tic o
c tio n to th e
f G L E A M is
h e p la n e x e c u
p e ra to rs a
b io lo g ic a
c a lle d a n
tio n is d o
n d
l
„ a
n e
g e n o ty p e
e v o lu tio n
c tio n “ , w h
b y s im u la
S . C a g n o n i e t a l. (E d s .): E v o W o rk s h o p s 2 0 0 0 , L N C S 1 8 0 3 , p p . 3 2 7 -3 3 8 , 2 0 0 0 .
© S p rin g e r-V e rla g B e rlin H e id e lb e rg 2 0 0 0
re p re se
(se e d
ic h re p
tio n a s
n ta tio n
e ta ile d
re s e n ts
p a rt o f
3 2 8
C . B lu m e
th e e v o lu tio n a lg o rith m to g e t a v a lu e fo r th e „ fitn e s s “ o f th e
G L E A M g e n e ra te s a s e q u e n c e o f b a s ic a c tio n s o r s ta te m e n ts , w h ic
o f th e p la n , w h ic h is a m e m b e r o f th e e v o lu tio n p o p u la tio n . A p la n
th e g e n e tic in fo rm a tio n , s e e [1 ]. T h e p u rp o s e o f th e p la n is n o t
e v o lu tio n its e lf, th e re fo re th e k e rn e l o f G L E A M in c lu d in g th e e v o lu
a p p lie d to d iffe re n t p ro b le m s w ith m in o r c h a n g e s . F o r e x a m p le , th
th e b a s ic c o m m a n d s o f a s im p le ro b o t c o n tro lle r o r a llo c a tio n
m a c h in e in a p ro d u c tio n p la n .
p la n . In
h a re th e
re p re se n
o f in te re
tio n a lg o
e s e a c tio
s te p s to
p a rtic u la r
e le m e n ts
ts d ire c tly
s t fo r th e
rith m w a s
n s c a n b e
re se rv e a
T h e G L E A M m e th o d w a s im p le m e n te d fo r s e v e ra l a p p lic a tio n s :
p la n n in g c o llis io n fre e m o v e s fo r in d u s tria l ro b o ts , s e e [4 ] a n d [1 1 ]
g e n e ra tin g p ro d u c tio n p la n s s o lv in g th e jo b -s h o p -p ro b le m , s e e [5 ]
p ro c e s s s c h e d u lin g in c h e m ic a l in d u s try , s e e [6 ]
G L E A M is im p le m e n te d a s w e ll in P a s c a l a s in C o n d iffe re n t h a rd w a re p la tfo rm s ,
e .g . o n P C , w o r k s ta tio n , a n d a p a r a lle l c o m p u te r s y s te m . I t is p o r ta b le , b e c a u s e th e
im p le m e n ta tio n is s tru c tu re d m o d u la r a n d c o n s is ts o f th e fo llo w in g m a in p a rts :
B a s ic m a c h in e : E .g . in itia liz a tio n , d a ta s tr u c tu r e c o n s tr u c tio n a n d m a n a g e m e n t,
e rro r h a n d lin g
O p tim iz a tio n k e rn e l: E v o lu tio n fu n c tio n s fo r m u ta tio n , re c o m b in a tio n ,
p o p u la tio n m a n a g e m e n t
S im u la tio n a n d e v a lu a tio n : E x e c u tio n o f p la n s , c rite rio n ’s c h e c k , re s tric tio n s ,
fitn e s s c a lc u la tio n
O v e ra ll c o n tro l: M a n a g e m e n t o f p a ra m e te r in p u t, d is p la y o f re s u lts ,
in te rru p tin g s im u la tio n o r o p tim iz a tio n a n d o th e rs
O v e r a ll c o n tr o l
U s e r
S im u la tio n
& e v a lu a tio n
I n te r f a c e
O p tim iz a tio n
k e r n e l
B a s ic m a c h in e
G L E A M
im p le m e n ta tio n
s tr u c tu r e
Im p le m e n ta tio n e ffo r t to a p p ly
G L E A M to a n o th e r a p p lic a tio n
R o b o t s p e c ific
m o d u le :
G e n e
o f r
c o n
s ta te
r a tio n
o b o t
tr o l
m e n ts
O p tim iz e d C o llis io n F re e R o b o t M o v e S ta te m e n t G e n e ra tio n
3 2 9
U s e r in te rfa c e : C rite rio n ’s p rio rity a n d fitn e s s d e fin itio n , ta rg e t d e fin itio n ,
s im u la tio n v is u a liz a tio n , p la n d e s c rip tio n a n d o th e rs
R o b o t c o d e g e n e ra tio n m o d u le , g e n e ra te s a ro b o t p ro g ra m p a rt in c lu d in g th e
m o v e s ta te m e n ts
S im u la tio n a n d e v a lu a tio n n e e d th e m a in a m o u n t o f im p le m e n ta tio n e ffo rt to a p p ly th e
G L E A M im p le m e n ta tio n to o th e r a p p lic a tio n s . It m e a n s o n th e o th e r s id e , th a t d u e to
th e g e n e ra lity o f th e m e th o d a n e w a p p lic a tio n re q u ire s le s s im p le m e n ta tio n c o s ts
b e c a u s e th e b a s ic m a c h in e , o p tim iz a tio n k e rn e l a n d o v e ra ll c o n tro l a re n o t m u c h
e ffe c te d .
A p p lic a tio n o f G L E A M
to R o b o tic s
T h e firs t a p p lic a tio n o f G L E A M to ro b o tic s w a s p e rfo rm e d b e c a u s e o f tw o re a s o n s :
1 . T h e a u th o r is a n e x p e rt in in d u s tria l ro b o t p ro g ra m m in g a n d c o n tro l (C o n v e n o r
o f a n IS O w o rk in g g ro u p ), th e re fo re a p ra x is o rie n te d w o rk w a s p e rfo rm e d , s e e
[7 ]
2 . T h e c a lc u la tio n fo r c o n tro llin g a n in d u s tria l ro b o t to m o v e o n a p re d ic te d
tra je c to ry w ith o u t c o llis io n is v e ry c o m p le x b u t th e re s u lt is e a s y to e v a lu a te :
e v e ry b o d y c a n s e e , if th e m o v e m e n t is c o llis io n fre e a n d ta k e s a s h o rte r tim e .
T h e a p p lic a tio n o f G L E A M w a s s o m e th in g lik e a b e n c h m a rk e s p e c ia lly to s h o w , th a t
G L E A M is a p o w e rfu l p la n n in g a n d o p tim iz a tio n to o l fo r th e c o n tro l o f d y n a m ic
p ro c e s s e s lik e th e m o v e m e n t o f in d u s tria l ro b o ts .
T h e G L E A M m e t h o d a p p l i e d t o r o b o t c o l l i s i o n f r e e m o v e s d o e s n 't p e r f o r m a n
e x p l i c i t s e a r c h o f t h e c o n f i g u r a t i o n s p a c e , b e c a u s e t h e s e a r c h p a r a m e t e r s d o n 't i n c l u d e
c o n fig u ra tio n s , lik e in [8 ] o r [9 ]. W ith G L E A M th e m o v e s ta te m e n ts fo r p e rfo rm in g
th e c o llis io n fre e m o v e a re g e n e ra te d d ire c tly . T h e G L E A M m e th o d a v o id s th e
p ro b le m o f c a lc u la tin g a s m o o th p a th b e tw e e n th e c o n fig u ra tio n s a n d o f b u ild in g u p
3 3 0
C . B lu m e
a n d s to rin g th e c o n fig u ra tio n s p a c e , s e e [1 0 ]. T h e c o llis io n fre e p a th is o p tim iz e d b y
c r ite r io n ’ s s e le c te d b y th e u s e r , e .g . th e c r ite r io n c o u ld b e a s h o r t C a r te s ia n p a th . O th e r
c rite rio n ’s lik e e n e rg y o r m o v e e x e c u tio n tim e a re in te g ra te d in to th e o p tim iz a tio n
p ro c e s s . T h e c rite rio n ’s c o u ld b e c o n tra d ic tio n a ry . T h e re fo re e v e ry c rite rio n h a s a
p rio rity a n d th e re s u lts o f th e o p tim iz a tio n re fle c t th e d iffe re n t p rio ritie s .
T h e e v o lu tio n p e rfo rm e d b y G L E A M s ta rts b y g e n e ra tin g a c tio n re s . p rim itiv e
s ta te m e n t s e q u e n c e s fo r c o n tro llin g th e ro b o t m o v e m e n t. T h e s e b a s ic a c tio n s fo r th e
ro b o t m o v e s lo o k lik e fo llo w s :
- m o v e r o b o t a x is k w ith a v e lo c ity o f r d e g r e e s p e r s e c o n d w ith a n a c c e le r a tio n o f
b d e g r e e s p e r s e c o n d 2 (E v o lu tio n p a ra m e te rs a re k , r, a n d b )
-
r o b o t a x is m
w ith a s lo w d o w n o f n d e g r e e s p e r s e c o n d 2 (E v o lu tio n
m e te rs a re m a n d n )
e s s o f a s e q u e n c e c a n b e c a lc u la te d a s a fu n c tio n o f th e fo llo w in g c rite rio n ’s :
o s itio n p r e c is io n , i.e . th e d is ta n c e o f th e e n d p o in t o f th e m o v e m e n t a n d th e
g iv e n ta rg e t p o in t
o r ie n ta tio n p r e c is io n , i.e . th e d if f e r e n c e b e tw e e n th e p la n n e d a n d r e a c h e d
o rie n ta tio n o f th e g rip p e r
o r ie n ta tio n c h a n g e s , i.e . th e o r ie n ta tio n m o v e s o f th e g r ip p e r to r e a c h th e ta r g e t
o rie n ta tio n
q u a lity o f th e m o v e tr a je c to r y , i.e . th e d if f e r e n c e b e tw e e n th e m o v e p a th a n d a
s tra ig h t lin e
le n g th o f th e m o v e tra je c to ry
f a s tn e s s o f m o v e e x e c u tio n , i.e . th e d u r a tio n o f m o v e m e n t
p r o g r a m le n g th , i.e . th e n u m b e r o f a c tio n s r e s . s ta te m e n ts r e q u ir e d to p e r f o r m
th e ro b o t m o v e
I n te r m e d ia te p o in ts , i.e . th e r o b o t m o v e p a s s e s s p e c if ie d in te r m e d ia te p o in ts
e c o n o m y , i.e . th e e n e r g y n e e d e d f o r th e m o v e e x e c u tio n
s to p
p a ra
T h e fitn
p
B e fo re th e e v a lu a tio n , th e s ta te m e n t s e q u e n c e is p ro o fe d fo r its p la u s ib ility , fo r
e x a m p le : a s to p -a x is -s ta te m e n t b e fo re a s ta rt-m o v e -a x is -s ta te m e n t is n o t m e a n in g fu l
(a n d th e re fo re c a n c e le d ). A fte r th e p la u s ib ility -c h e c k th e e v o lu tio n c o n tin u o u s b y th e
m u ta tio n o f th e e v o lu tio n p a ra m e te rs o f th e s ta te m e n ts a n d b y re c o m b in a tio n . A s th e
s e q u e n c e o f e le m e n ta ry m o v e s is e s s e n tia l fo r th e re s u ltin g o v e ra ll m o v e m e n t th e re
a re s o m e m u ta tio n s a lte rin g o n ly th e s e q u e n c e o f a c tio n s .
T h e c o m p le te s e q u e n c e o f a c tio n s c o n tro ls th e ro b o t m o v e to a p o in t n e a r th e ta rg e t
p o s itio n w ith o u t c o llis io n s . It is a s s u m e d , th a t th e ro b o t c o n tro l is b a s e d o n a fix e d
c o n tr o l c y c le tim e , e .g . o f 5 0 m s . E v e r y c o n tr o l s ta te m e n t is e x e c u te d b y th e
s im u la tio n to o l w ith re s p e c t to th is c y c le tim e , w h ic h c a n b e d e fin e d a s a c o n tro l
c h a ra c te ris tic . T h e s im u la tio n p e rfo rm s th e ro b o t m o v e a n d s to re s th e a x is v a lu e s fo r
e v e ry c o n tro l c y c le . T o m e a s u re th e d is ta n c e fro m th e ta rg e t p o s itio n a n d th e p a th
le n g th , th e g e n e ra l fo rw a rd tra n s fo rm a tio n fro m ro b o t to C a rte s ia n c o o rd in a te s is
a p p lie d .
T h e G L E A M m e th o d is a b le to d e v e lo p a s ta te m e n t s e q u e n c e fo r a n y k in d o f
ro b o t. A s o ftw a re to o l R O B M O D E F (R o b o t M o d e l D e fin itio n ) w a s im p le m e n te d
O p tim iz e d C o llis io n F re e R o b o t M o v e S ta te m e n t G e n e ra tio n
p ro v id in g a g ra p h ic s u p p
1 6 ro ta tio n a l ro b o t a x e s ,
z e ro p o s itio n ), a n d w h ic h
T h e u s e r c a n d e fin e
a v o id a n c e . T h e d e fin e d
a c tio n d e fin itio n is c a lle d
u s e r a n d s to re d fo r fu rth e
o rte d ro b o t d e fin itio n . T h e
w h ic h c a n b e d ire c te d in to
c a n ro ta te a b o u t th e x -, y o b s ta c le s in th e ro b o t m
ro b o t m o d e l, th e o b s ta c le
th e " a c tio n m o d e l" . It c a n
r e x p e rim e n ts .
u s e r c a n b u ild
th e x - o r z -c o o r
o r z -c o o rd in a te
o v e m e n t a re a
d e fin itio n s a n
e a s ily c h a n g e d
a ro b o t w
d in a te d ir
a x is .
to s h o w
d th e (p r
o r e x te n d
3 3 1
ith u p to
e c tio n (in
c o llis io n
e d e fin e d )
e d b y th e
G e n o ty p e R e p r e s e n ta tio n a n d D a ta S tr u c tu r e
T h e a c tio n o f G L E A M is o n e g e n e a n d c o n s is ts o f th e a c tio n c o d e a n d a n u m b e r o f
p a ra m e te rs . T h e p a ra m e te rs c a n b e o f in te g e r, re a l o r c h a ra c te r ty p e . T h e a c tio n c o d e
a n d p a ra m e te r d e fin itio n is s to re d in a n (re a d a b le ) file a n d c a n b e c h a n g e v e ry e a s ily .
T h e n u m b e r o f a c tio n s , w h ic h b u ild u p a m e m b e r o f th e e v o lu tio n p o p u la tio n , is n o t
lim ite d . T h e le n g th o f s u c h a n a c tio n c h a in is fle x ib le a n d a n o p tim iz a tio n c rite ria o f
th e e v o lu tio n .
G L E A M w a s d e s ig n e d fo r th e o p tim iz a tio n o f d y n a m ic p ro c e s s e s w ith re s p e c t to
th e n a tu ra l e v o lu tio n . T h e re fo re th e e v o lu tio n p ro c e s s o p e ra te s a b o u t g e n e s re s . p la n s
o f v a ria b le le n g th . D u e to th is a ttrib u te th e ta s k o f re c o m b in a tio n is m o re c o m p lic a te d
th a n in e v o lu tio n a lg o rith m s o f o th e r a u th o rs . T h e m o s t im p o rta n t im p ro v e m e n t o f
G L E A M is th e in tro d u c tio n o f th e c o n c e p t o f s o c a lle d „ s e c tio n s “ . A s e c tio n is fo rm e d
b y a ( v a r ia b le ) n u m b e r o f g e n e s r e s . a c tio n s a s a s u b s tr u c tu r e o f th e p la n , i.e . th e
a c tio n c h a in is p a rtitio n e d in to s e g m e n ts . A s e g m e n t c a n b e re g a rd e d a s a c h ro m o s o m e
o f th e b io lo g ic a l g e n e tic in fo rm a tio n . S o , o n e m e m b e r, th a t m e a n s th e a c tio n c h a in o f
th e p la n , c o n s is ts o f s e g m e n ts , a n d e a c h s e g m e n t o f a n u m b e r o f a c tio n s .
T h e re a re n e w d e fin e d e v o lu tio n o p e ra to rs (fo r m u ta tio n a n d re c o m b in a tio n ) to b e
a p p lie d to th e s e c tio n s , lik e d e le te a s e g m e n t o r m o v e a s e g m e n t to a n o th e r p la c e in
th e a c tio n c h a in . T h is c o n c e p t e n a b le s th e re c o m b in a tio n o f „ g o o d “ s u b -s tru c tu re s o f a
3 3 2
C . B lu m e
p ro b le m s o lv in g p ro c e s s to s p e e d u p th e e v o lu tio n . If tw o p la n s a re m e rg e d b y th e
re c o m b in a tio n , th e c o m b in a tio n o f a s e g m e n t o f p la n A w ith a s e g m e n t o f p la n B c a n
b e tre a te d a s a c o m b in a tio n o f tw o s u b -s o lu tio n s . If th e tw o s u b -s tru c tu re s re s .
s e g m e n ts a re g o o d in s in g le , th e y b o th to g e th e r in th e re s u ltin g p la n w ill g iv e a m u c h
h ig h e r fitn e s s v a lu e th a n th e fitn e s s o f th e b o th p a re n t p la n s . T o c o m e to s u c h a g o o d
s o lu tio n o n ly b y m u ta tio n o f th e g e n e s ta k e s p ro b a b ly m u c h lo n g e r tim e . T h e c o n c e p t
o f s e g m e n ts (th e „ c h ro m o s o m e s “ ) a n d a c tio n s (th e „ g e n e s “ ) w a s v e ry s u c c e s s fu l fo r
p la n n in g a n d o p tim iz a tio n ta s k s o f d y n a m ic p ro c e s s e s lik e ro b o t m o v e s .
S im ila r to th e „ tra d itio n a l“ a rtific ia l in te llig e n c e la n g u a g e L IS P , th e d a ta s tru c tu re
o f G L E A M is b a s e d o n d y n a m ic lis ts . T h e a c tio n c h a in s ta rts w ith a h e a d e r n o d e
s to r in g a ll n e c e s s a r y in f o r m a tio n a b o u t th e a c tio n c h a in , e .g . c h a in le n g th ( e .i. n u m b e r
o f a c tio n s a n d s e g m e n ts ), fitn e s s v a lu e , a n d o th e rs . T h e h e a d e r is fo llo w e d b y a c tio n s ,
a ll lin k e d b y p o in te rs . A n a c tio n c h a in c a n g ro w o r b e re d u c e d , it d e p e n d s o n th e
e v o lu tio n . O f c o u rs e , if a n a c tio n c h a in (th a t m e a n s a p ro b le m s o lu tio n ) g e ts a g o o d
fitn e s s v a lu e , a n d a n o th e r c h a in g e ts th e s a m e fitn e s s ra n k b u t c o n s is ts o f a lo w e r
n u m b e r o f a c tio n s , th e la s t o n e is b e tte r a n d w ill s u rv iv e .
R e p r e s e n ta tio n o f a c tio n :
A c t io n c o d e
A c t io n p a r a m e t e r s
A c tio n c h a in :
1 .A c tio n
M o v e
A x is 1
H e a d e r o f
th e a c tio n
c h a in
R e fe re n c e to th e
n e x t a c tio n
2 .A c tio n
M o v e
A x is 3
...
E x a m p le s fo r th e m u t a tio n
o p e r a to r s fo r a c tio n s :
D e le t e :
A c tio n
K
A c tio n
L
A c tio n
M
In se rt:
A c tio n
K
A c tio n
L
A c tio n
M
A c tio n
X
C h a n g e :
A c tio n
M
A c tio n
L
A c tio n
K
O p tim iz e d C o llis io n F re e R o b o t M o v e S ta te m e n t G e n e ra tio n
3 3 3
G e n o ty p e O r ie n te d C o d e a n d R o b o t C o d e G e n e r a tio n
T h e g e n o ty p e o rie n te d c o d e o f G L E A M fo r ro b o ts c o n s is ts o f p rim itiv e m o v e
m a n d s . T h e c o d e is s im u la te d d u rin g th e e v o lu tio n (w ith o u t a n y o u tp u t, o n ly th e
e s s v a lu e is c a lc u la te d d u rin g th e s im u la tio n ), a n d if th e u s e r w a n ts to s e e th e re s u lt
th e p la n n in g a n d o p tim iz a tio n p ro c e s s , th e c o d e o f s u c c e s s fu l a c tio n p la n s is
u la te d w ith g ra p h ic a l d is p la y o f th e ro b o t m o v e m e n ts .
A s m e n tio n e d a b o v e , th e ro b o t c o n tro l s im u la tio n c a lc u la te s a m o v e s te p b y s te p ,
s e e [1 1 ]. B e tw e e n e v e ry c a lc u la te d v ia p o s itio n o f th e tra je c to ry a u s e r d e fin e d c y c le
tim e h a s p a s s e d . T h e m o v e d e s c rip tio n b y p rim itiv e c o m m a n d s (u s e d in te rn a lly b y
G L E A M ) h a s to b e tra n s fe rre d to c o d e fo rm u la te d in a ro b o t p ro g ra m m in g la n g u a g e ,
b u t p ro g ra m m in g th e s a m e ro b o t m o v e . A n e w m o d u le „ ro b o t c o d e g e n e ra to r“ w a s
im p le m e n te d , w h ic h c a lc u la te s th e p a ra m e te r fo r a m o v e p ro g ra m m in g b y ro b o t m o v e
s ta te m e n ts .
A fte r e v e ry c o n tro l c y c le th e v a lu e s fo r th e ro b o t a x e s a n d th e C a rte s ia n
c o o rd in a te s fo r th e to o l c e n te r p o in t T C P a re c a lc u la te d a n d tra n s fe rre d to th e ro b o t
c o d e g e n e ra to r (if th e u s e r h a s m a rk e d „ c o d e g e n e ra tio n “ w ith th e h e lp o f m e n u in p u t).
T h e firs t s te p o f th e c o d e g e n e ra tio n is th e d e c la ra tio n o f p o s itio n d a ta a n d a n
a s s ig n m e n t o f v a lu e s . T h is p o s itio n d a ta c o n s is ts o f th e in te rm e d ia te p o in ts o f th e
m o v e a n d th e ta rg e t p o in t. T h e ta rg e t p o in t is g iv e n b y th e u s e r in C a rte s ia n
c o o rd in a te s , th e c o d e g e n e ra to r c a lc u la te s th e v a lu e s fo r th e p a ra m e te rs to d e s c rib e th e
ro b o t s p e c ific o rie n ta tio n a n d c o n fig u ra tio n d a ta . A fte r th e d a ta d e fin itio n , a s im p le
m a in fu n c tio n is g e n e ra te d c o n s is tin g o f m o v e c o m m a n d s a n d re fe rrin g to th e p o s itio n
d a ta g e n e ra te d b e fo re . T h e ro b o t c o d e c a n b e d o w n lo a d e d to th e ro b o t c o n tro l a n d
e x e c u te d .
c o m
fitn
o f
s im
E x a m p le to p r o g r a m th e A B B IR B 2 4 0 0 r o b o t:
!
!
!
C O M M E N T
C O M M E N T
C O M M E N T
M O D U L E
V A R
C O N
C O N
C O N
!
. . . . .
C O N
P o s i t i o n l i s t
w i t h t h e m o v e
( r e p r e s e n t i n g
p o s i t i o n f o r e v e r y
t h e t r a j e c t o r y ) :
d e m o 1
r o b t a r g e t t a r g e t : = [ [
[ 0 . 1 2 3 6 7 , 0 . 6 9 6 2 1 , 0 . 1 2 3
[ 9 E
S T j o i n t t a r g e t P 1 : = [ [
0 . 0 , - 4 5 . 3 ,
S T j o i n t t a r g e t P 2 : = [ [
0 . 0 , - 4 5 . 3 ,
S T j o i n t t a r g e t P 3 : = [ [
0 . 0 , - 4 4 . 6 ,
. . .
S T j o i n t t a r g e t P 7 4 : = [
1 6 . 9 , 4 . 1 , - 1
- 1 6 4
6 7 , 9 , 9 E
0
0 . 0 ]
0
0 . 0 ]
0
0 . 0 ]
7 .
0 .
9 ,
. 5
,
. 7
,
. 7
,
0 ,
6 9
9 E
,
[ 9
,
[ 9
,
[ 9
- 6
6 2
9 ,
4
E 9
4
E 9
4
E 9
0 4
1 ]
9 E
5 .
,
5 .
,
5 .
,
. 2
, [
9 ,
3 ,
. .
3 ,
. .
3 ,
. .
, 1 3 7
- 2 , 0
9 E 9 ,
4 5
. . ,
4 5
. . ,
4 5
. . ,
[ - 1 6 0 . 7 ,
4 5 . 2 ,
1 7 . 0 ] , [ 9 E 9 , . . . .
C O M M E N T
M o v e s t a t e m e n t s f o r t h e
P R O C m a i n ( )
M o v e A b s J P 1 , v 4 0 0 , z 2 0 , t o o l x ;
R A P I D
c y c l e
4 .
, 0
9 E
. 3
9 E
. 3
9 E
. 3
9 E
7 ]
, 0
9 ]
,
9 ]
,
9 ]
,
9 ]
,
] ,
] ;
] ;
] ;
] ;
0 . 6 ,
, 9 E 9 ] ] ;
p r o g r a m :
3 3 4
C . B lu m e
M o
M o
. . . .
M o
M o
E N D P R
E N D M O
v e A b s J
v e A b s J
P 2 , v 4 0 0 , z 2 0 , t o o l x ;
P 3 , v 4 0 0 , z 2 0 , t o o l x ;
v e A b s J P 7 4 , v 4 0 0 , z 2 0 , t o o l x ;
v e L t a r g e t , v 1 0 0 , z 2 0 , t o o l x ;
O C
D U L E
E x e c u tin g th e s e ro b o t s ta te m e n ts , th e A B B c o n tro l p e rfo rm s o n ly o n e ro b o t m o v e
to th e ta rg e t p o in t u s in g th e m a n y g e n e ra te d p o in ts to p e rfo rm a „ s m o o th p a th “ .
T h e re fo re th e re s u ltin g m o v e m e
a h u m a n a rm . T h e re a s o n is , th
o n ly a fe w (ty p ic a lly 3 to 1 0 )
th e m , a n d th e ro b o t m o v e lo o k s
o f th e n e w m e th o d is th e g e n e ra
d ire c tio n . T h is le a d s to a c o n tr
m o to rs a n d g e a rs . A ls o th e g e n e
s to p s .
n t lo o k s m u c h m o re „ n a tu ra
a t th e p ro g ra m m e r u s u a lly
in te rm e d ia te p o in ts w ith a
th e n m o re „ m a c h in e -lik e “ o
tio n o f m o v e s , w h ic h a v o id
o l b e h a v io r m o re a v o id in g
ra te d m o v e w a s c h e c k e d to
l“ , i.e . lik e a m o v e m e n t o f
„ d iv id e s “ a ro b o t m o v e in
g re a te r d is ta n c e b e tw e e n
r „ c o rn e re d “ . A s id e e ffe c t
su d d e n c h a n g e o f sp e e d o r
w e a r a n d te a r o f th e a x is
a v o id ru n n in g a g a in s t a x is
O p tim iz e d C o llis io n F re e R o b o t M o v e S ta te m e n t G e n e ra tio n
3 3 5
R e s u lts o f th e I m p le m e n ta tio n s fo r I n d u s tr ia l R o b o ts
T h e im p le m e n ta tio n s o f G L E A M to in d u s tria l ro b o ts h a v e b e e n d o n e to d e m o n s tra te ,
th a t G L E A M is a b le to p ro d u c e ro b o t c o d e in a n in d u s tria l e n v iro n m e n t. T h e s o ftw a re
fo r th e A B B ro b o t p ro g ra m m in g fa c ility g e n e ra te s ro b o t p ro g ra m s w ith c o llis io n fre e
ro b o t m o v e s . T h e c o llis io n fre e p a th is o p tim iz e d b y c rite rio n ’s o f d iffe re n t p rio ritie s
s e le c te d b y th e u s e r . E .g . th e c r ite r io n c o u ld b e a s h o r t C a r te s ia n p a th o f th e T C P o r a
s h o rt d is ta n c e o f o n e o re m o re ro b o t a x e s . T h e p ro c e s s o f p ro g ra m m in g o p tim iz e d
ro b o t m o v e s c a n b e d o n e b y th e s o ftw a re to o l, w h ic h re d u c e s th e n e e d e d m a n p o w e r
fo r th e u s e r.
A n o th e r re s u lt is th e d e m o n s tra tio n , th a t th e m e th o d G L E A M is a p p lic a b le to a n
in d u s tria l ro b o t c o n tro l o f th e u s e r. T h e a d v a n ta g e s a re c o n tro llin g ro b o ts w ith o u t
c o m p lic a te d m a th e m a tic s c a lc u la tio n s a n d e a s y a d a p ta tio n o f d iffe re n t e n v iro n m e n ts
a n d ro b o t m o d e ls . T h e re s u lt c a n b e u s e d fo r o ff-lin e p a th p la n n in g in c lu d in g a m u ltic rite ria o p tim iz a tio n . T h e g e n e ra te d m o v e tra je c to ry is a n o p tim iz a tio n o f a ll
c rite rio n ’s , a v o id in g p ro b le m s o f d e fin in g th e c o n fig u ra tio n s p a c e o r tra n s fe r
p o s itio n s .
T h e im p le m e n ta tio n f o r D a im le r C h r y s le r in c lu d e s s e v e r a l im p r o v e m e n ts , e .g . th e
d e fin itio n o f in te rm e d ia te p o in ts b y th e u s e r. T h e s e in te rm e d ia te p o in ts h a v e to b e
p a s s e d b y th e ro b o t tra je c to ry (m o re o r le s s ), th e u s e r c a n in flu e n c e th e ro b o t p a th a n d
h e lp th e s y s te m to fin d a c o llis io n fre e p a th in a s h o rte r tim e .
T h e im p le m e n ta tio n s h a v e b e e n p e rfo rm e d o n a P C , th e y a llo w a lo w c o s t s o lu tio n
a n d w ill b e a c c e p te d a ls o fo r s m a lle r c o m p a n ie s . T h e s o ftw a re fo r th e G L E A M
m o d u le s a re w ritte n in C . T h e c o m p a n ie s A B B a n d D a im le rC h ry s le r c a n te s t th e n e w
m e th o d a n d d e c id e , if it w ill b e in te g ra te d in to th e ir p ro g ra m m in g s o ftw a re to o ls , th e
im p le m e n ta tio n s a re th e b a s e o f fe a s ib ility s tu d ie s to a n a ly z e th e n e w m e th o d .
3 3 6
C . B lu m e
O u tlo o k
T h e G L E A M c o n c e p t w a s re a liz e d b y im p le m e n ta tio n s fo r d iffe re n t ro b o ts : th e
M its u b is h i R V -M 2 ro b o t, th e A B B IR B 2 4 0 0 in d u s tria l ro b o t , a n d th e K U K A K R 6
in d u s tria l ro b o t. T h e y g e n e ra te s ta te m e n ts fo r th e ro b o t c o n tro ls to m o v e th e ro b o t o n
a n o p tim iz e d c o llis io n fre e tra je c to ry to a g iv e n m o v e ta rg e t.
In fu tu re a n im p le m e n ta tio n fo r a p ro fe s s io n a l s im u la tio n to o l fo r ro b o ts w ill b e o f
in te r e s t, e .g . f o r th e I G R I P s y s te m o f th e D e n e b c o m p a n y . S u c h a n im p le m e n ta tio n
w ill d e m o n s tra te th e a p p lic a tio n o f th e G L E A M m e th o d to a h ig h le v e l la n g u a g e
s im u la tio n c o m m a n d la n g u a g e lik e G S L , a n d a c o m p le x m o d e le d e n v iro n m e n t o f
o b s ta c le s a n d ro b o ts . T h e fo llo w in g p ic tu re s h o w s th e in fo rm a tio n flo w b e tw e e n th e
s y s te m c o m p o n e n ts . T h e u s e r w o rk s w ith th e s im u la tio n to o l a s b e fo re a n d m o d e ls its
m a c h in e s , w o rk c e ll, o b s ta c le s , a n d o th e rs . H e c a n p ro g ra m th e ro b o t m o v e s in a
s im u la tio n la n g u a g e a n d s im u la te th e m o v e s .
U s e r In p u t to th e s im u la tio n to o l
C o m m a n d s
in c l. c o llis io n fre e
m o v e fu n c tio n
S im u la tio n
w ith g ra p h .
re p re s e n ta t.
S im u la tio n to o l
w ith m o d e llin g ,
ro b o t m o v e s,
s im u la tio n la n g u a g e
s ta te m e n t g e n e ra tio n
S im u l.s ta
fitn e s s v
c o llis io n
tio n to G
tm e n ts ,
a lu e s ,
d e te c L E A M
S im u la tio n
s ta tm e n ts to th e
s im u l. to o l
G L E A M
o p tim iz a tio n
a n d g e n e r a tio n
o f c o llis io n fr e e
ro b o t m o v e s
S im u la tio n la n g u a g e
s ta tm e n ts
P o s tp ro c e s s o r
to g e n e ra te s ta te m e n ts
fo r th e ro b o t c o n tro l
M o v e s ta tm e n ts
fo r d iffe re n t ro b o t
c o n tro ls
R o b o t
C o n tr o l
U n it
A n e w
s im u la tio n
p ro g ra m m e
ro b o t m o v e
fu n c tio n „ o p
s y s te m . T h e
d b y th e u s e r
s (s h o rte r e x e
tim iz e m
s im u la tio
to th e G
c u tio n tim
o v e s
n to
L E A
e ) w
w ith o
o l se n
M c o m
ith re s
u t c o llis io n “
d s th e s im u
p o n e n t. G L E
p e c t to c o llis
w ill b e in tro d u c
la tio n s ta te m e n ts
A M trie s to o p tim
io n a v o id a n c e . D u
e d
a lre
iz e
rin g
in to
a d y
th e
th e
O p tim iz e d C o llis io n F re e R o b o t M o v e S ta te m e n t G e n e ra tio n
3 3 7
o p tim iz a tio n p ro c e s s G L E A M s e n d s c h a n g e d s ta te m e n ts in s im u la tio n la n g u a g e c o d e
to th e s im u la tio n s y s te m , w h ic h w ill e x e c u te th e s ta te m e n ts in a „ s ile n t“ b a c k g ro u n d
m o d e w ith o u t g ra p h ic a l o u tp u t. T h e s im u la tio n to o l s e n d s b a c k to G L E A M o n ly th e
in fo rm a tio n a b o u t th e „ fitn e s s “ o f th e s im u la te d m o v e s ta te m e n ts , e s p e c ia lly if th e re
w a s a c o llis io n o r n o t. A fte r th e o p tim iz a tio n p ro c e s s G L E A M s e n d s th e re s u lt to th e
s im u la tio n to o l: th e o p tim iz e d c o llis io n fre e m o v e s ta te m e n ts in th e s im u la tio n
la n g u a g e . T h e u s e r c a n s im u la te th is re s u lt a n d te s t, if it is b e tte r th a n h is o w n
s ta te m e n ts . A t th e e n d , th e s im u la tio n to o l g e n e ra te s p ro g ra m c o d e fo r d iffe re n t ro b o t
c o n tro l u n its .
It is im p o rta n t, th a t G L E A M c a n u s e s o lu tio n s a lre a d y p ro g ra m m e d b y th e u s e r a s
a n in p u t. G L E A M trie s to fin d a b e tte r s o lu tio n (if th e re e x is ts o n e ), a n d it g e n e ra te s a
tra je c to ry w ith o u t c o llis io n s . T h e u s e r c a n v is u a liz e th e re s u lt b y s im u la te it, a n d h e
c a n u s e th e g e n e ra te d s ta te m e n ts fo r o th e r p ro g ra m s o r to p ro g ra m a n y ro b o t.
R e fe r e n c e s
1 . B lu m e , C ., J a k o b , W .: C lo s in g th e O p tim iz a tio n G a p in P r o d u c tio n b y G e n e tic
A lg o rith m s . P ro c . o f th e E u r o p e a n C o n g r e s s o n In te llig e n t T e c h n iq u e s a n d S o ft
C o m p u tin g (E U F IT 9 3 ), 1 9 9 3 , A a c h e n
2 . B lu m e , C .: I n d u s tr ie lle A n w e n d u n g e n E v o lu tio n ä r e r A lg o r ith m e n ( I n d u s tr ia l
A p p lic a tio n s o f e v o lu tio n a ry a lg o rith m s ). C o n tr ib u tio n : P la n u n g k o llis io n s fre ie r
B e w e g u n g e n fü r In d u s trie ro b o te r (P la n n in g o f c o llis io n fre e m o v e m e n ts o f
in d u s tr ia l r o b o ts ) . E d .: S . H a f n e r , R . O ld e n b o u r g V e r la g , M ü n c h e n W ie n 1 9 9 8
3 . B lu m e , C .: G L E A M - A S y s te m f o r S im u la te d „ I n tu itiv e L e a r n in g “ . P r o c e e d in g s
o f th e 1 s t In te r n a tio n a l W o r k s h o p o n P r o b le m S o lv in g fr o m N a tu r e , D o rtm u n d ,
G e rm a n y , O c to b e r 1 -3 , 1 9 9 0
4 . B lu m e , C ., J a k o b , W ., K r is c h , S .: R o b o t T r a je c to r y P la n n in g w ith C o llis io n
A v o id a n c e u s in g G e n e tic A lg o rith m s a n d S im u la tio n . P r o c . o f th e 2 5 th
I n te r n a tio n a l S y m p o s iu m o n I n d u s tr ia l R o b o ts , 2 5 .- 2 7 . A p r il 1 9 9 4 , H a n n o v e r , p p .
1 6 9 -1 7 5
5 . P la n n in g a n d O p tim iz a tio n o f S c h e d u lin g in In d u s tria l P ro d u c tio n b y G e n e tic
A lg o rith m s a n d E v o lu tio n a ry S tra te g y . P r o c . o f th e S e c o n d B ie n n ia l E u r o p e a n
J o in t C o n fe r e n c e o n E n g in e e r in g S y s te m s D e s ig n a n d A n a ly s is (E S D A ), J u ly 4 -7 ,
1 9 9 4 , L o n d o n , E n g la n d
6 . B lu m e , C ., G e r b e , M .: D e u tlic h e S e n k u n g d e r P r o d u k tio n s k o s te n d u r c h O p tim ie ru n g d e s R e s s o u rc e n e in s a tz e s . (R e d u c tio n o f p ro d u c tio n c o s ts b y o p tim iz in g
th e re s o u rc e p la n n in g ). a tp - A u to m a tis ie r u n g s te c h n is c h e P r a x is , 3 6 (1 9 9 4 ) p p . 5 9
7 . B lu m e , C ., F r ü a u f , P .: S ta n d a r d iz a tio n o f P r o g r a m m in g M e th o d s a n d la n g u a g e s f o r
M a n ip u la tin g In d u s tria l R o b o ts . 2 7 th In te r n a tio n a l S y m p o s iu m o n In d u s tr ia l
R o b o ts , O k to b e r 1 9 9 6 , M a ila n d
8 . D a i, F .: C o llis io n - F r e e M o tio n o f a n A r tic u la te d K in e m a tic C h a in in a D y n a m ic
E n v iro n m e n t, IE E E C o m p u te r G r a p h ic s & A p p lic a tio n , J a n u a ry 1 9 8 9 , p p . 7 0 -7 4
3 3 8
C . B lu m e
9 . G o ld b e r g , D ., P a r k e r , J ., K h o o g a r , A .: I n v e r s e K
u s in g G e n e tic A lg o rith m s . IE E E In te r n a tio n a l
A u to m a tio n 7 , (1 9 8 9 ). p p . 2 7 1 -2 7 6
1 0 . H e in e , R ., S c h n a r e , T .: K o llis io n s f r e ie B a h n p la n
p a th p la n n in g fo r ro b o ts ). R o b o te r s y s te m e 7 , (1 9 9 1
1 1 . G e n e ra tio n o f O p tim iz e d C o llis io n F re e R o b o t M o
A lg o rith m s . P r o c e e d in g s o f th e W o r ld A u to m a tio n
R o b o tic a n d M a n u fa c tu r in g S y s te m s , M a y 2 8 -3 0 ,
8 9 - 9 4
in e m a tic s o f R e d u n d a n t R o b o ts
C o n fe r e n c e o n R o b o tic s a n d
u n g
). p p
v e S
C o n
1 9 9
fü r R o b o te r (C o
. 1 7 -2 2
ta te m e n ts B a s e d
g re ss (W A C ‘9 6 )
6 , M o n tp e llie r , F
llis io n fre e
o n G e n e tic
V o lu m e 3 :
ra n c e . p p .
S e lf-A d a p tiv e
M u ta tio n
L a rry
In te llig
F a c u lty o f
U n iv e
B
{ L a r r y .B
e n t C
C o m
rs ity
ris to
u ll,J
B u ll &
o m
p u
o f
l B
a c o
p u te
te r S
th e
S 1 6
b 3 .H
in
Z C S
Ja c o b
H u rst
r S y s te m
tu d ie s &
W e st o f
1 Q Y , U
u rs t} @ u
s C e n
M a th
E n g la
.K .
w e .a c .u
C o n tr o lle r s
tre
e m a tic s
n d
k
A b s tr a c t. T h e u s e a n d b e n e fits o f s e lf-a d a p tiv e m u ta tio n o p e ra to rs a re w e ll-k n o w n
w ith in e v o lu tio n a ry c o m p u tin g . In th is p a p e r w e e x a m in e th e u s e o f s e lf-a d a p tiv e
m u ta tio n in M ic h ig a n -s ty le C la s s ifie r S y s te m s w ith th e a im
o f im p ro v in g th e ir
p e rfo rm a n c e a s c o n tro lle rs fo r a u to n o m o u s m o b ile ro b o ts . In itia lly , w e im p le m e n t th e
o p e ra to r in th e Z C S c la s s ifie r a n d e x a m in e its p e rfo rm a n c e in tw o
a n im a t
e n v iro n m e n ts . It is s h o w n th a t, a lth o u g h n o s ig n ific a n t in c re a s e in p e rfo rm a n c e is s e e n
o v e r re s u lts p re s e n te d in th e lite ra tu re u s in g a fix e d ra te o f m u ta tio n , th e o p e ra to r a d a p ts
t o a p p r o x i m a t e l y t h i s r a t er e g a r d l e s s o f t h e i n i t i a l r a n g e .
1
I n tr o d u c tio n
W ith in G e n e tic A lg o rith m s (G A s ) [H o lla n d 1 9 7 5 ] a n d G
1 9 9 1 ] th e m u ta tio n ra te is tra d itio n a lly a g lo b a l p a ra m e te r
H o w e v e r, in E v o lu tio n a ry S tra te g ie s [R e c h e n b e rg 1 9
E v o lu tio n a ry P ro g ra m m in g (M e ta -E P ) [F o g e l 1 9 9 2 ], th e
e v o lv in g e n tity in its e lf , i.e . it a d a p ts d u r in g th e s e a r c h
fo rm o f m u ta tio n n o t o n ly re d u c e s th e n u m b e r o f h a n
e v o lu tio n a ry a lg o rith m , it h a s a ls o b e e n s h o w n to im p
[B c k 1 9 9 2 ] fo r re s u lts w ith a s e lf-a d a p tiv e G A ). In th is
a s e lf-a d a p tiv e m u ta tio n o p e ra to r w ith in M ic h ig a n -s ty le
[H o lla n d e t a l. 1 9 8 6 ], m o re s p e c ific a lly in W ils o n ’s Z C S
T h e p e rfo rm a n c e o f th e n e w o p e ra to r w ith in Z C S
a u to n o m o u s e n tity /ro b o t - a n im a t [W ils o n 1 9 8 5 ]
e a c h o f w h ic h w e re o rig in a lly u s e d b y W ils o n to
b o th c a s e s it is fo u n d th a t n o b e n e fits in p e rfo
p re s e n te d b y W ils o n u s in g a fix e d m u ta tio n ra te .
ra n g e fo r th e a d a p tin g m u ta tio n ra te s , th e fin a l C S
s a m e m u ta tio n ra te a s th a t u s e d b y W ils o n . T h a t
a d a p ta tio n w o rk s w ith in th e C la s s ifie r S y s te m fra m
e n e tic P ro g ra m m in g [K o z a
w h ic h is c o n s ta n t o v e r tim e .
7 3 ] a n d la te r fo rm s o f
m u ta tio n ra te is a lo c a lly
p ro c e s s . T h is s e lf-a d a p tiv e
d -tu n a b le p a ra m e te rs o f th e
r o v e p e r f o r m a n c e ( e .g . s e e
p a p e r w e e x a m in e th e u s e o f
C la s s ifie r S y s te m s (C S s )
[W ils o n 1 9 9 4 ] s y s te m .
is e x a m in
- ta s k s : W
in tro d u c e
rm a n c e a re
H o w e v e r,
a n im a t c o n
is , w e s h
e w o rk .
e d u s in
o o d s 1
a n d in v
fo u n d
re g a rd le
tro lle rs
o w th e
g
tw
a n d
e s tig
o v e
ss o
h a v e
p rin
o
a
r
f
c
s im
W o o
te Z
th e
th e
ro u g
ip le
u la te d
d s 7 ,
C S . In
re s u lts
in itia l
h ly th e
o f s e lf-
T h e p a p e r is a rra n g e d a s fo llo w s : th e n e x t s e c tio n in tro d u c e s Z C S . S e c tio n 3 d e s c rib e s
h o w s e lf-a d a p tiv e m u ta tio n is im p le m e n te d a n d S e c tio n 4 d e s c rib e s th e ta s k s a n d
e x a m in e s th e e ffe c ts o f th e o p e ra to r. F in a lly , a ll re s u lts a re d is c u s s e d .
S . C a g n o n i e t a l. ( E d s .) : E v o W o r k s h o p s 2 0 0 0 , L N C S 1 8 0 3 , p p . 3 3 9 − 3 4 6 , 2 0 0 0 .
S p rin g e r-V e rla g B e rlin H e id e lb e rg 2 0 0 0
340
L. Bull and J. Hurst
2 Z C S
ZCS is a "Zeroth-level" Michigan-style Classifier System without internal memory,
where the rule-base consists of a number (N) of condition/action rules in which the
condition is a string of characters from the usual ternary alphabet {0,1,#} and the
action is represented by a binary string. Associated with each rule is a strength scalar
which acts as an indication of the perceived utility of that rule within the system. This
strength of each rule is initialised to a predetermined value termed S0.
Reinforcement in ZCS consists of redistributing strength between subsequent "action
sets", or the matched rules from the previous time step which asserted the chosen
output or "action". A fixed fraction (E) of the strength of each member of the action
set ([A]) at each time-step is placed in a "common bucket". A record is kept of the
previous action set [A]-1 and if this is not empty then the members of this action set
each receive an equal share of the contents of the current bucket, once this has been
reduced by a pre-determined discount factor (J). If a reward is received from the
environment then a fixed fraction E of this value is distributed evenly amongst the
members of [A]. Finally, a tax (W) is imposed on all matched rules that do not belong
to [A] on each time-step in order to encourage exploitation of the stronger classifiers.
Hence this is different from the traditional "Bucket-brigade" algorithm [Holland et al.
1986] and is known [Wilson 1994] to be similar to Watkin’s Q-learning [1989]
reinforcement algorithm.
ZCS employs two discovery mechanisms, a panmictic GA and a covering operator.
On each time-step there is a probability p of GA invocation. When called, the GA uses
roulette wheel selection to determine two parent rules based on strength. Two
offspring are produced via mutation (probability P) and crossover (single point with
probability F). The parents then donate half of their strengths to their offspring who
replace existing members of the rule-base. The deleted rules are chosen using roulette
wheel selection based on the reciprocal of rule strength. If on some time-step, no rules
match or all matched rules have a combined strength of less than I times the rule-base
average, then a covering operator is invoked.
The default parameters presented for ZCS, and unless otherwise stated for this paper,
are: N = 400, S0=20, E = 0.2, J = 0.71,W = 0.1, F = 0.5, P = 0.002, p = 0.25, I = 0.5
Thus ZCS represents a "basic classifier system for reinforcement learning that retains
much of Holland’s original framework while simplifying it so as to increase
understandability and performance" [Wilson 1994]. For this reason the ZCS
architecture has been chosen to examine the basic behaviour of classifier systems with
self-adaptive mutation rates. The reader is referred to [Wilson 1994] for full details of
ZCS.
Self-Adaptive Mutation in ZCS Controllers
341
3 Se l f - A dapt ive C l assif ie r Syst e m C o nt ro l l e rs
3. 1 Se l f - A dapt at io n
In this paper we use the same form of self-adaptive mutation as in Meta-EP. That is,
each rule has its own mutation rate P, stored as a real number, which is passed to its
offspring, either under recombination or directly (depending upon the satisfaction of
F). The offspring then applies its mutation rate to itself using a Gaussian distribution,
i.e. Pi’ = Pi + N(0,Pi), before mutating the rest of the rule at the resulting rate. It is
noted that this form of self-adaptation is simpler than that typically used in
Evolutionary Strategies, where a Lognormal is applied to P, however the simpler form
is shown to be adequate here and has been suggested to work better in noisy
environments [Angeline et al. 1996].
We also note that this is in contrast to the adaptive form of crossover introduced by
Wilson (1987) for CS, under which a system entropy measure was used to alter the
operator rate; Wilson showed benefits from increasing crossover as entropy dropped
using predetermined rules of change.
3. 2 C l assif ie r Syst e ms in Evo l ut io nary Ro b o t ic s
A number of investigators have examined the use of Classifier Systems in
evolutionary robotics. Dorigo, in conjunction with many others (see [Dorigo &
Colombetti 1999] for a comprehensive overview), has used multiple CSs in a
hierarchy to control an autonomous robot in a variety of environments. To our
knowledge this remains the only hardware implementation to date. A large body of
work exists on the use of CSs to control simulated robots however, e.g. [Riolo 1991],
[Cliff & Bullock 1993], [Donnart & Meyer 1994], [Stolzmann 1999], etc. The reader
is referred to [Lanzi et al. 2000] for a full CS bibliography.
The performance of a Michigan-style classifier system - ZCS - with self-adaptive
mutation in simulated evolutionary robotics tasks is now examined, with the aim of
determining ways to improve their use in real environments.
4 Re sul t s in Wo o ds 1 and Wo o ds 7
4. 1 T h e T ask s
Wilson [1994] introduced two multi-step "woods" environments with which to
examine the performance of ZCS. Woods 1 is a two dimensional rectilinear 5x5
toroidal grid. Sixteen cells are blank, eight contain rocks and one contains food. ZCS
is used to develop the controller of a robot/animat which must traverse the map in
search of food. It is positioned randomly in one of the blank cells and can move into
any one of the surrounding eight cells on each discrete time step, unless occupied by
a rock. If the animat moves into the food cell the system receives a reward from the
3 4 2
L . B u ll a n d J . H u rs t
e n v ir o n m e n t ( 1 0 0 0 ) , a n d th e ta s k is r e s e t, i.e . f o o d is r e p la c e d a n d th e a n im a t r a n d o m ly
re lo c a te d (F ig u re 1 ).
O n e a c h tim e s te p th e a n im a t re c e iv e s a s e n s o ry m e s s a g e w h ic h d e s c rib e s th e e ig h t
s u rro u n d in g c e lls . T h e m e s s a g e is e n c o d e d a s a 1 6 -b it b in a ry s trin g w ith tw o b its
re p re s e n tin g e a c h c a rd in a l d ire c tio n . A b la n k c e ll is re p re s e n te d b y 0 0 , fo o d (F ) b y 1 1
F ig .
e n v ir
C S c
a d ja c
*
O
O
O
O
T h e W
m e n t s
tro lle d
t to th
o o d s 1
h o w in g th e
a n im a t *
e fo o d g o a l.
F
O
O
1:
o n
o n
e n
O
O
a n d ro c k s (O ) b y 1 0 (0 1 h a s n o m e a n in g ). T h e m e s s a g e is o rd e re d w ith th e c e ll d ire c tly
a b o v e th e a n im a t re p re s e n te d b y th e firs t b it-p a ir, a n d th e n p ro c e e d in g c lo c k w is e
a ro u n d th e a n im a t.
T h e tria l
p re v io u s
o n e a c h
w h ils t th
W o o d s 7
F ifty -s e v e
ro c k s p o s
is b la n k .
o p tim u m
Z C S h a s
c a n n o t b e
h o w w e ll
W o o d s 7
is re p e a te
5 0 tria ls )
tria l. If it
e o p tim u m
n
1 0 ,0 0 0 tim e
o f h o w m a n y
m o v e d ra n d o m
is s a id to b e
s a n
s te p
ly W
1 .7
d
a
s it
ils o
s te p
re c o rd is k e p t o f a m o v in g a v e ra g e (o v e r th e
ta k e s fo r th e a n im a t to m o v e in to a fo o d c e ll
n c a lc u la te s p e rfo rm a n c e a t 2 7 s te p s p e r tria l,
s.
is a m o re c o m p le x a n d n o n -M a rk o v v e rs io n o f
c e lls e v e n ly s c a tte re d a ro u n d th e m a p c o n ta in
itio n e d ra n d o m ly in tw o o f th e e ig h t s u rro u n d in g
W ils o n s ta te s th a t ra n d o m s e a rc h w ill ta k e 4 1
is 2 .2 s te p s p e r tr ia l ( n o t s h o w n - s e e [ W ils o n 1
n o te m p o ra ry m e m o ry (s e e [C liff & R o s s 1 9 9 5 ][T
e x p e c te d to s o lv e W o o d s 7 o p tim a lly , h o w e v e r
[it] c a n d o [W ils o n 1 9 9 4 ]. A ll in p u ts a n d o th e r
a s W o o d s 1 .
A ll re s u lts in
4 .2
d
th is p a p e r a re th e a v e ra g e o f te n
W o o d s 1 o n a 5 8 x 1 8 g rid .
fo o d . E a c h o f th e s e h a s
c e lls . T h e re s t o f th e m a p
s te p s to fo o d , w h ils t th e
9 9 4 ]). It is n o te d th a t, s in c e
o m lin s o n & B u ll 1 9 9 8 ]), it
it is s till o f in te re s t to s e e
ta s k d e ta ils a re th e s a m e in
ru n s.
R e s u lts
F ig u re 2 s h o w s th e p e rfo rm a n c e o f th e s e lf-a d a p tiv e m u ta tio n o p e ra to r w ith in Z C S o n
W o o d s 1 . H e re th e in itia l p o s s ib le ra n g e o f m u ta tio n ra te s w a s c e n tre d a ro u n d th e
f i x e d r a t e o f 0 .0 0 2 u s e d b y W i l s o n , i .eµ. < 00 <.0 0 4 . I t c a n b e s e e n t h a t t h e u s e o f s e l f -
S e lf-A d a p tiv e M u ta tio n in Z C S C o n tro lle rs
3 4 3
a d a p ta tio n h a s h a d n o re a l b e n e fic ia l/d e trim e n ta l e ffe c ts o n p e rfo rm a n c e , if a n y th in g
le a rn in g is a little q u ic k e r (F ig . 2 a ). E x a m in a tio n o f th e a v e ra g e m u ta tio n ra te in th e
F i g . 2: S h o w i n g
ru le -b
0 .0 0 2
th e re
b y W
(a )
th e b e h a v io u r o f s e lf-a d a p tiv e m u ta tio n
in
W o o d s 1 .
a s e ( F ig . 2 b ) s h o w s th a t it r is e s s lig h tly f r o m th e m e a n o f 0 .0 0 2 0 u p to a r o u n d
5 . T h a t is , a lth o u g h n o s ig n ific a n t im p ro v e m e n ts in p e rfo rm a n c e a re s e e n h e re ,
is o b v io u s ly a s lig h t s e le c tiv e p re s s u re fo r a h ig h e r m u ta tio n ra te th a n th a t u s e d
ils o n .
T h e r e s u ltin g r u le - b a s e s a t th e e n d o f th e 1 0 ,0 0 0 tr ia ls h a v e b e e n e x a m in e d . A lth o u g h
th e a v e ra g e m u ta tio n ra te a p p e a rs ro u g h ly e q u a l to th e fix e d ra te u s e d b y W ils o n ,
a n a ly s is s h o w s a w id e ra n g e o f in d iv id u a l ra te s . T y p ic a lly , a la rg e p ro p o rtio n o f th e
F i g . 3: S h o w i n g
(a )
th e e ffe c ts o f a h ig h e r in itia l m u ta tio n
r u l e s h a v e n o m u t a t i o n , i . µe .i = 0 , w h i l s t o t h e r s h
T h e fo rm e r o f th e s e is a s s o c ia te d w ith ru le s fo
e n v iro n m e n t, w h ils t th e la tte r a re a s s o c ia te d w ith
H e n c e it a p p e a r s th a t th e d e g r e e o f e v o lu tio n a r
s itu a tio n /c e ll, d u r in g th e le a r n in g p r o c e s s , is d ir
t h e r u l e s i n t h e i n d u c t i v e / r e i n f o r c e m e n t c h a i nR . u n
tria ls s h o w th e a v e ra g e m u ta tio n ra te e v e n tu a lly
ra te s e e d
in
W o o d s 1 .
a v e m u ta tio n ra te s u p to a n d o v e r 1 .
r c e lls c lo s e s t to th e fo o d g o a l in th e
ru le s fo r c e lls fu rth e s t fro m th e g o a l.
y s e a r c h o n th e r u le s o f a p a r tic u la r
e c tly c o r r e la te d w ith th e p o s itio n o f
s o v e r a m u c h la rg e r n u m b e r o f
g o in g to z e ro (n o t s h o w n ).
3 4 4
L . B u ll a n d J . H u rs t
W e h a v e a ls o e x a m in e d th e ro b u s tn e s s o f th e s e lf-a d a p tiv e a p p ro a c h to s ta rtin g th e
s y s te m w ith a n in a p p ro p ria te m e a n m u ta tio n ra te . T h a t is , w e w e re in te re s te d in
w h e th e r s e lf-a d a p ta tio n c a n b e u s e d to re m o v e th e m u ta tio n p a ra m e te r fro m th e
(a )
F i g . 4: S h o w i n g
(b
th e e ffe c ts o f s e lf-a d a p tiv e m u ta tio n
d e s ig n e r’s c o n tro l s o m
p a rtic u la rly e v o lu tio n a
0 < µ < 0 .5 . I t c a n a g a in
fix e d m u ta tio n ra te (F
to w a r d s 0 .0 0 2 a f te r a
e w h a t a
ry ro b o
b e se e n
ig . 3 a ),
s lig h t d
n d h e n
tic s . F
th a t n
b u t th a
e la y (F
c e e a s e th
ig u re 3 s
o re a l b e n
t n o w th e
ig . 3 b ).
e u se
h o w s
e fit/d e
a v e ra
in
o f
re s
trim
g e
W o o d s 7 .
C la s
u lts
e n t
m u ta
s ifie rs
fro m
is fo u
tio n r
in c o m p le x ta s
a n in itia l ra n g
n d o v e r th e u s e
a te c o n tin u a lly f
k s,
e
o f a
a lls
R e s u lts fro m th e m o re c o m p le x a n d n o n -M a rk o v W o o d s 7 ta s k w e re v e ry s im ila r to
th o s e a b o v e . F ig u re 4 s h o w s th a t th e re is n o s ig n ific a n t c h a n g e in p e rfo rm a n c e w h e n
F i g . 5: S h o w i n g
(a )
th e e ffe c ts o f a h ig h e r in itia l m u ta tio n
th e in itia l m u ta tio n ra te s
T h e re is a ra p id in c re a s e
w h ic h th e n d e v ia te s a ro u n
in a d iffe re n t e n v iro n m e n
m u ta tio n ra te . T e s ts fo r
0 < µ < 0 .5 ) a ls o s h o w e d n o
a re se e d e d
in th e a v e
d 0 .0 0 3 , a
t w e se e a
ro b u s tn e s s
c h a n g e in
a ro u n
ra g e m
s lig h tly
d iffe re
w ith
p e rfo r
d
W ils o
u ta tio n
h ig h e r
n t fo rm
m u c h h
m a n c e a
n ’s
ra te
ra te
o f
ig h e
n d
ra te s e e d
in
W o o d s 7 .
f ix e d r a te o f 0 .0 0 2 ( F ig . 4 a ) .
, u p to a r o u n d 0 .0 0 4 ( F ig . 4 b ) ,
th a n s e e n in W o o d s 1 . T h a t is ,
s e lf-a d a p ta tio n o c c u rrin g in th e
r in itia l m u ta tio n r a te s ( e .g .
th a t th e a v e ra g e m u ta tio n ra te
S e lf-A d a p tiv e M u ta tio n in Z C S C o n tro lle rs
fa lls to w a rd
0 .0 0 2 , a lth o u g h
m o re q u ic k ly
th a n
in
W o o d s 1
3 4 5
(F ig u re 5 ).
D u e to th e c o m p le x ity o f th e e n v iro n m e n t a n a ly s is o f th e re s u ltin g ru le -b a s e s is m o re
d iffic u lt h e re , b u t th e s a m e g e n e ra l c o rre la te d e ffe c t in te rm s o f m u ta tio n /s tre n g th
c o n v e rg e n c e a p p e a rs to o c c u r a s d e s c rib e d a b o v e in W o o d s 1 .
5
C o n c lu s io n s
In th is p a p e r it h a s b e e n s h o w n th a t it is p
o p e ra to r w ith in M ic h ig a n -s ty le C la s s ifie r S y s te m
th e a im o f im p ro v in g th e ir p e rfo rm a n c e a s c o n
W e a re n o w m o v in g th e s e e x p e rim e n ts o n to a
th e In te llig e n t A u to n o m o u s S y s te m s L a b o ra to r
F u rth e r e n h a n c e m e n ts to th e s e lf-a d a p tiv e m
in v e s tig a te d , a s w e ll a s im p le m e n tin g th e m in
[W ils o n 1 9 9 5 ].
o s s ib le to u s e a s e lf-a d a p tiv e
s - s p e c ific a lly W ils o n ’s Z C S
tro lle rs fo r a u to n o m o u s m o b ile
re a l ro b o t p la tfo rm in c o n ju n c
y , F a c u lty o f E n g in e e rin g a t
e c h a n is m
a re a ls o c u rre n tly
th e m o re s o p h is tic a te d X C S
m u ta tio n
- w ith
ro b o ts .
tio n w ith
U W E .
b e in g
s y s te m
A c k n o w le d g e m e n ts
T h a n k s to
A n d y
T o m lin s o n
fo r a n u m b e r o f u s e fu l d is c u s s io n s d u rin g
th is w o rk .
R e fe r e n c e s
A n g e lin e , P .J ., F o g e l, D .B ., F o g e l, L .J . ( 1 9 9 6 ) A C o m p a r is o n o f S e lf - A d a p ta tio n
M e th o d s f o r F in ite S ta te M a c h in e s in a D y n a m ic E n v ir o n m e n t. I n L .J . F o g e l, P .J .
A n g e l i n e , & T . B c k ( e d s E . )v o l u t i o n a r y P r o g r a m m i n g V , M I T P r e s s , p p . 4 4 1 - 4 4 9 .
B
c k , T . ( 1 9 9 2 ) S e lf - A d a p ta tio n in G e n e tic A lg o r ith m s . I n F .J . V a r e la & P . B o u r g in e
( e d s .) T o w a r d a P r a c tic e o f A u to n o m o u s S y s te m s : P r o c e e d in g s o f th e F ir s t E u r o p e a n
C o n f e r e n c e o n A r t i f i c i a l L i ,f e M I T P r e s s , p p 2 6 3 - 2 7 1 .
C liff, D . & B u llo c k , S . (1 9 9 3 ) A d d in g
B e h a v io r 2 (1 ):4 7 -7 0 .
C liff, D . & R o s s , S . (1 9 9 5 ) A d d in g
3 (2 ): 1 0 1 -1 5 0 .
D o n n a rt, J -Y . & M e y e r,
P o s itio n in g w ith M o n a L y s
W ils o n ( e d s .) F r o m A n im a
C o n fe r e n c e o n S im u la tio n
J-A
a .
ls
o f
. (1 9 9 4
In P . M
to A n im
A d a p tiv
`F o v e a l V is io n ' to
T e m p o ra ry
) S
a e s
a ts
e B
M e m o ry
p a tia l E x p lo ra
, M . M a ta ric ,
4 : P r o c e e d in g
e h a v i o uM r ,I T P r
tio n
J-A
s o
e ss,
D o r i g o , M . & C o l o m b e t t i , M . ( 1 9 9 R9 ) o b o t S h a p i n g : A n
E n g in e e r in g . M IT P re s s .
W i l s o n ' s A nA i d m a a p t t . i v e
t o AZ dC a S p .t i v e B e h a v i o r
, M
. M
f th
p p
a p L e
e y e r,
e F o u
2 0 4 -2
a rn in g , a n d S e lfJ . P o lla c k & S .W .
r th In te r n a tio n a l
1 3 .
E x p e r im e n t in
B e h a v io r
3 4 6
L . B u ll a n d J . H u rs t
F o g e l , D . B . ( 1 9 9 2 )E v o l v i n g
C a lifo rn ia .
A r t i f i c i a l I n t e l l i g e n c .e P h D
H o lla n d , J .H . ( 1 9 7 5 )A d a p ta tio n
M ic h ig a n P re s s .
in
N a tu r a l a n d
d is s e rta tio n , U n iv e rs ity
A r t i f i c i a l S y s t e . m sU n i v e r s i t y
H o lla n d , J .H ., H o ly o a k , K .J ., N is b e tt, R .E . &
T h a g a rd ,
P r o c e s s e s o f I n f e r e n c e , L e a r n i n g a n d D i s c o v e. r M y I T P r e s s .
K o z a , J . R . ( 1 9 9 1 )G e n e t i c P r o g r a m m i n g . M I T
P .R .
o f
o f
( 1 9 I n8 d6 u) c t i o n :
P re ss.
L a n z i , P - L . , S t o l z m a n n , W . & W i l s o n , S . W . ( e d s . ) ( 2 0 P 0 r0 o ) c e e d i n g s o f t h e S e c o n d
I n t e r n a t i o n a l W o r k s h o p o n L e a r n i n g C l a s s i f i e r S y s t e ,m Ss p r i n g e r - V e r l a g .
R e c h e n b e r g , I . ( 1 9 7 3 )E v o lu tio n s s tr a te g ie ; O p tim ie r u n g te c h n is c h e r
P r in z ip e n d e r b io lo g is c h e n E v o lu tio.n F r o m m a n n - H o lz b o o g V e r la g .
R io
J-A
In te
3 2 6
S y s te m e
n a c h
lo , R . (1 9 9 1 ) L o o k a h e a d P la n n in g a n d L a te n t L e a rn in g in a C la s s ifie r S y s te m . In
. M e y e r & S . W . W i l s o n ( e d Fs . r ) o m A n i m a l s t o A n i m a t s : P r o c e e d i n g s o f t h e F i r s t
r n a t i o n a l C o n f e r e n c e o n S i m u l a t i o n o f A d a p t i v e B e h a v i ,o u M r I T P r e s s , p p 3 1 6 .
S to lz m a n n , W . (1 9 9 9 ) L a te n t L e a rn in g in K h e p ra R o b o ts w ith A n tic ip a to ry C la s s ifie r
S y s t e m s . I n A . S . W u ( e d . )P r o c e e d i n g s o f t h e 1 9 9 9 G e n e t i c a n d E v o l u t i o n a r y
C o m p u t a t i o n C o n f e r e n c e W o r k s h o p P r o g r a ,m M o r g a n K a u f f m a n , p p 2 9 0 - 2 9 7 .
T o m lin s o n , A . & B u ll, L . ( 1 9 9 8 ) A C o r p o r a te C la s s if ie r S y s te m . I n A .E . E ib e n , T .
c k , M . S c h o e n a u e r & H - P . S c h w e f e l ( e dP s a . )r a l l e l P r o b l e m S o l v i n g f r o m N a t u r e
- P P S N V, S p r i n g e r , p p . 5 5 0 - 5 5 9 .
B
W a t k i n s , C . ( 1 9 8 9 )L e a r n i n g
C a m b rid g e .
fr o m
D e la y e d
R e w a r d. s P h D
d is s e rta tio n , U n iv e rs ity
o f
W ils o n , S .W . ( 1 9 8 5 ) K n o w le d g e G r o w th in a n A r tif ic ia l A n im a l. I n J .J . G r e f e n s te tte
( e d .) P r o c e e d in g s o f th e F ir s t I n te r n a tio n a l C o n fe r e n c e o n G e n e tic A lg o r ith m s a n d
t h e i r A p p l i c a t i o n s, L a w r e n c e E r l b a u m A s s o c i a t e s , p p 1 6 - 2 3 .
W ils o n , S .W . ( 1 9 8 7 ) C la s s if ie r S y s te m s a n d
2 :1 9 9 -2 2 8 .
W ils o n , S .W . ( 1 9 9 4 )
C o m p u ta tio n 2 (1 ):1 -1 8 .
Z C S :
A
W ils o n , S .W . ( 1 9 9 5 ) C la s s if ie r
C o m p u ta tio n 3 (2 ):1 4 9 -1 7 7 .
t h e A n i m a t P r o b lM e m a c . h i n e L e a r n i n g
Z e ro th -le v e l
F itn e s s
C la s s ifie r
B a se d
o n
S y s t e mE v . o l u t i o n a r y
A c c u r a c E y v. o l u t i o n a r y
Using a Hybrid Evolutionary-A* Approach for
Learning Reactive Behaviours
Carlos Cotta and Jose M. Troya
Dept. of Lenguajes y CC.CC., University of Malaga,
Complejo Tecnologico 3.2.49 , Campus de Teatinos,
E-29071, Malaga, Spain
fccottap, troyag@lcc.uma.es
Abstract. A hybrid approach for learning reactive behaviours is presented in this work. This approach is based on combining evolutionary algorithms EAs with the A* algorithm. Such combination is done within
the framework of Dynastically Optimal Forma Recombination, and tries
to exploit the positive features of EAs and A* e.g., implicit parallelism,
accuracy and use of domain knowledge while avoiding their potential
drawbacks e.g., premature convergence and combinatorial explosion .
The resulting hybrid algorithm is shown to provide better results, both
in terms of quality and in terms of generalisation.
1
Introduction
The control of autonomous mobile agents is a complex task to which great eorts
are devoted due to its practical applications. In general, such control is achieved
by means of both planning and reactive components 13 . Each of these components has it own particularities, and can be examined in combination e.g., 1,
2, 7 or in isolation e.g., 11, 12 . In line with the latter, this work focuses on
the acquisition of reactive behaviours in mobile agents.
Reactive behaviours are driven by a stimulus-to-response mapping, i.e., the
agent receives some information about its local environment and decides the
action s to carry on exclusively on the basis of such information. This kind of
behaviour has usually the advantage of not requiring any underlying global model
of the world in which the agent is located. The obvious drawback of reactive
systems is the fact that they can get stuck into dead-ends, situations in which
the correct action does not only depend on the locally available information but
also the structure of the world at a higher-level hence the necessity of longterm planning capabilities. Nevertheless, reactive systems have been shown to
provide a very good performance in a wide variety of scenarios and remain a
very suitable option when response-time is critical.
There exist several techniques for designing reactive systems. These can be
typically classied into reinforcement learning and optimisation techniques. Algorithms such as Holland's bucket brigade, Sutton's temporal dierence learning
or Watkins's Q-learning lie in the rst class. Within the second class, evolutionary algorithms deserve special attention because of their power and exibility.
S . C a g n o n i e t a l. ( E d s .) : E v o W o r k s h o p s 2 0 0 0 , L N C S 1 8 0 3 , p p . 3 4 7 − 3 5 6 , 2 0 0 0 .
S p rin g e r-V e rla g B e rlin H e id e lb e rg 2 0 0 0
3 4 8
C . C o tta a n d J .M . T r o y a
Regarding the use of these techniques for this purpose e.g., 8, 9, 11 , a critical point is the use of as much domain knowledge as possible. Otherwise, the
user would be relying on a fortuitous matching between her algorithm and the
problem under consideration 14 . Such specialised algorithms are usually termed
hybrid evolutionary algorithms 6 .
This work presents a hybrid evolutionary algorithm for acquiring reactive
behaviours. In the proposed algorithm, domain knowledge is included by using a
specialised technique the A* algorithm as an internal operator. The remainder
of the article is organised as follows. First, the agent and the worlds used in the
experiments are described Sect. 2 . Next, the classical A* approach for solving
the posed problem is shown Sect. 3 . Then, the hybrid algorithm is introduced
Sect. 4 . Subsequently, experimental results are presented Sect. 5 . Finally,
some conclusions are extracted and future work is outlined Sect. 6 .
2 The Agent and its World
The agent used in this work is located in a two-dimensional toroidal grid-world
in which several obstacles are distributed. The purpose of the agent is to reach
a certain target point from its initial location within an allowed time. To do so,
the agent is capable of making some elementary actions such as moving straight
ahead a single grid square, turning 90o to its left, or turning 90o to its right.
Obviously, the agent must avoid obstacles while navigating through its world.
For this purpose, it is equipped with proximate sensors that can inform of the
presence or absence of obstacles in front of the agent, 90o to its left, or 90o to
its right see Fig. 1, left . In addition, these sensors can also detect whether the
target point is in any of these three locations or not.
Fig. 1. Left Structure of the agent used in experiments. Right Example world and
regions into which it is divided according to the location of the target point.
U s i n g a H y b r i d E v o l u t i o n a r -y A* Approach for Learning Reactive Behaviours
3 4 9
The agent is equipped with a direction sensor as well. This sensor allows
determining in which of four imaginary regions of the world the target point is
located. These regions are illustrated in Fig. 1 right. It must be noted that
these regions are not absolute but relative to the agent's actual orientation. For
example, the agent is facing North in Fig. 1 and hence the target point is in zone
1. Now, if the agent turned 90o to its right, the target would be in zone 0. Notice
also that these regions are determined taking into account the toroidal shape of
the world. Thus, if the agent were a few positions South from the location shown
in the previous example, the target point might happen to be in zone 2.
According to this description, the goal is to design a reactive behaviour allowing the agent reaching its target in as many situations as possible. Such reactive
behaviour can be de ned in a variety of ways, e.g., using a neural network 15,
a fuzzy rule-base 8, a cellular automata 3, etc. This work is in line with the
latter approach. To be precise, a lookup-table is sought relating every possible
sensorial input with a primitive action. At each time-step, the agent must look
up the action that corresponds to the current inputs and carry it out. Since each
proximate sensor can provide three dierent inputs OBSTACLE, NO-OBSTACLE,
TARGET, and the direction sensor can return four values, the resulting table has
33 4 = 108 entries. Since three primitive actions MOVE-AHEAD, TURN-LEFT,
TURN-RIGHT are available, this implies a search space of 3108 3 1051 tables.
3 A Classical Approach: A*
A classical approach for nding the lookup-table mentioned above is the utilisation of the A* algorithm. Based on incrementally constructing solutions in an
intelligent fashion, this technique constitutes a powerful tool for solving search
problems to optimality. Before getting into the application of this technique to
the design of reactive behaviours, some notation details must be given.
Let be the current world, and let be the con guration of the agent
position and orientation. Now, let be the sensorial input of the agent
when con gured according to . Let M^ be a possibly underspeci ed function
relating sensorial inputs with actions, and let M^ 1 M^ 2 whenever M^ 2 provides
the same outputs that M^ 1 does and M^ 2 is de ned in at least one case in which
M^ 1 is not. Finally, let be the maximum allowed time for reaching the target
and let be a function such that M^ = 1 k . This function
provides a trace of the agent trajectory across con guration space when behaving
according to M^ . The value is an indication of the nal status of the agent:
AT-TARGET, COLLISION, TIMED-OUT or UNKNOWN. This latter value is returned
whenever no action is speci ed in M^ for the current input.
Now, the application of the A* algorithm requires the availability of an optimistic evaluation function such that M^ provides a lower bound on
the number of steps necessary for reaching the target when the agent is congured as and behaves according to any M^ M^ M^ . It is easy to see that
making return the Manhattan distance from the agent's current location to
W
I W
W
h
i
W
0
0
3 5 0
C . C o tta a n d J .M . T r o y a
the target point fulls this requirement. Having dened this function, the whole
process is as follows:
1. Let P00 = 0 M^ 0 0 t0 , where 0 is the initial conguration of the agent, M^ 0
is a fully underspecied function, and t0 = W 0 M^ 0 . Let P = M^ 0 1
be the current best solution. Insert P00 in the node queue.
^ h t be the
2. If the node queue is empty, go to 3. Otherwise let P = M
rst element in the queue.
a Let W M^ = h1 i .
b if 6= UNKNOWN then
^ t0 + , where
i. if = COLLISION or = TIMED-OUT then P = M
0
^
t = h + k + W M .
^ h+k .
ii. if = AT-TARGET then P = M
iii. If P is better than P , update the latter and purge nodes in the
queue.
iv. Go to 2.
c Create three nodes P 3+1+1 , P 3+1+2 , and P 3+1+3 from P . Each node is
P 3+1+ = M^ h0 t , where M^ is obtained by extending M^ to return the rth possible action when the input is I W , h0 = h + k, and
t = W M^ . Insert these nodes in the queue keeping it ordered
according to the sum of the last two components of each node.
d Go to 2.
3. Return P
i
j
k
k
j
j
i
r
k
r
r
i
j
i
j
i
i
j
r
k
r
k
r
This algorithm will thus return the lookup-table allowing the agent reach the
target in minimal time from the given starting point. Since the problem has been
posed with the goal of obtaining a generalisable reactive behaviour, the process
must be slightly modied. To be precise, a training set is selected and the A*
algorithm tries to nd the table that minimises the sum of the times required
to reach the target in each training case or, if such a solution is not possible,
a table that rstly maximises the number of training cases solved and secondly
minimises the total time. Notice that no global model of the world i.e., highlevel knowledge about the distribution of obstacles is required. All information
used for nding the optimal solution is locally obtained through simulation.
This algorithm has been evaluated on a set of nine dierent worlds. These
worlds are named as Wxy, where x 2 f10 25 50g indicates the dimension of
the world each world is a x x grid , and y 2 fa b cg indicates the density of
obstacles 5, 10 and 20 respectively . For each world, a training set of ve
cases has been selected. Subsequently, the best solution found has been tested
for generality on a test set whose size depends on the dimension of the world
50, 400 and 2000 cases respectively . The results are shown in Table 1.
These results are very indicative of the two main drawbacks of the A* algorithm. On the one hand, it is very sensitive to the size of the task to be solved.
As it can be seen, the algorithm expended a high computational eort for solving
W25b and ran out of memory in three cases W25c, W50b, and W50c . Moreover, it did not nd any fully satisfactory solution for all training cases in W50b
Using a Hybrid Evolutionary- A* Approach for Learning Reactive Behaviours
3 5 1
Results of the A* algorithm on nine dierent worlds. The cost values are
measured as the number of single simulation steps carried out.
Table 1.
World Timeout Optimal solution Iterations
W10a
13.20
11615
W10b 25
13.80
19664
W10c
14.00
26429
W25a
32.40
75818
W25b 150
31.60
222788
W25c
26.40, 51.00
300000
W50a
54.60
119864
W50b 400 45.60, 479.40 150000
W50c
34.80, 1629.80 110000
Performance
Cost
on test set
289857
74
300391
88
281091
58
10352732
63
29083614
48
25000000
36
36861614
66
37000000
14
16000000
2
and W50c 1 and 4 training cases were left unsolved1. On the other hand, the
solutions found are not very generalisable. This is a direct consequence of the
internal functioning of the A* algorithm. Assume that the nal solution is found
when evaluating node . This node was obtained as successive extensions of
n3 n9
0
0 . Hence, it contains information regarding the best decisions
,1 ,2
to be taken only in the situations found during this optimal path, i.e., the path
from the root node of the implicitly de ned search tree to the optimal leaf node.
All that may have been learnt in solving other situations is discarded since these
situations do not take place in this optimal path.
i
Pj
i
P
j
i
P
j
P
Growth of the computational cost of the A* algorithm when the number of
training cases is increased.
Fig. 2.
1
These results were not bad a priori since there might exist no better solution. However, further experimentation with the hybrid EA showed that this was not the case.
3 5 2
C . C o tta a n d J .M . T r o y a
This generalisation problem could be solved by considering a larger training
set whose optimal solution covered all possible situations. However, the subsequent combinatorial explosion makes this approach unrealistic. This is illustrated
in Fig. 2. As it can be seen, the computational cost of the algorithm grows very
fast when the size of the training set is increased. For this reason, it is clear that
alternative approaches must be found. These will be discussed in next section.
4 The Hybrid EA-A* Approach
Evolutionary algorithms constitute a very suitable alternative to A* for nding
the lookup-table. A nave approach for applying EAs to this problem would
rstly consist of dening an encoding function for storing the lookup-table into
an individual, e.g., a linear chromosome in which the rows of the table are
consecutively arranged. Since this is an orthogonal representation 10 i.e., all
combinations of genes are feasible, the next step would simply involve selecting
any of the standard genetic operators that can be found in the literature e.g.,
single-point crossover SPX , uniform crossover UX , etc..
However, such a simple approach is likely to provide very poor results. Recall
that this is highly epistatic problem in which the value of each gene i.e., a specic action to be carried out when a certain sensorial input is received does not
contribute with a xed amount to the tness of an individual. On the contrary,
the goodness of the reactive behaviour dened is determined by the interplay
between all genes. For this reason, a blind recombination operator that randomly
shues the genetic material of recombined solutions will provably produce solutions with a phenotype reactive behaviour completely unrelated to the parents,
even when the latter are genotypically similar. In an extreme situation, it may
even reduce to macromutation.
The algorithm would be largely more eective if it were able to extract positive behavioural patterns from existing solutions and transmit them to the ospring. This can be achieved within the framework of Dynastically Optimal
Forma Recombination 4 DOR. This framework comprises a family of recombination operators of the form
DOR : S S S ! 0 1
1
where S is the search space and DOR
is the
P probability of generating
= 1, the
when recombining and . Besides the obvious z2S DOR
probability distribution induced by these operators verify that
x y z
z
x
x y z
y
DOR
0 f 2 f g ^ 8 2 f g : g 2
where is the tness function to be minimised without loss of generality and
f g is the dynastic potential 10 of and , i.e., the set of solutions that
can be built using nothing but the information contained in and .
Thus, the solutions created by DOR are the best that can be constructed
using the genetic material of the parents. On the one hand, this implies that
x y z
z
,
x y
w
,
x y
w
z
,
x y
x
y
x
y
Using a Hybrid Evolutionary- A* Approach for Learning Reactive Behaviours
3 5 3
DOR is a fully transmitting operator, i.e., no implicit mutation genetic information not present in any of the parents is introduced in the ospring. On the
other hand, the tness-oriented functioning of DOR makes valuable portions of
solutions be transmitted to ospring only if they contribute to a good resulting
behaviour. In other words, DOR is capable of identifying valuable high-order
formae, preventing their disruption. This intelligent combination of information
has provided very good results on epistatic problems 5 .
In order to implement DOR, it is required to use an embedded A*-like mechanism so as to nd the best solution in the dynastic potential of the parents. In
this case, the algorithm described in Sect. 3 can be used. It is only necessary to
modify step 2c by considering that the possible actions to be taken in a given
situation are just those present in any of the parents for . Notice that the
search carried out by this subordinate A* algorithm is thus restricted to small
portions of the search space and hence its computational cost is largely reduced
with respect to the original unrestricted version. Moreover, individuals in the
population tend to be more similar as the EA converges and, subsequently, the
dynastic potential of selected solutions tends to be smaller and DOR is less
computationally expensive.
This combination of EAs and A* has an additional advantage. Each individual carries an information that reects its past evolution in fact, the evolution
of its ancestors. This way, things that were learnt in the past are retained as
long as they do not negatively aect the present behaviour. This accumulated
history" eect is also present in a simple EA, but the learning capabilities of the
hybrid algorithm are larger. For this reason, solutions obtained with the hybrid
EA are expected to be more general than either the EA or the A* algorithm by
themselves. This will be studied in next section.
I
I
5 Experimental Results
Experiments have been done with a steady-state EA popsize = 100, c = 9,
m =1chromosomeLength using ranking selection + = 2 0 , = 0 0. This
algorithm has been run 40 times for each operator and test world. In order to
make a fair comparison between DOR and the other simpler operators, each run
is terminated when a xed number of simulation steps 105 in these experiments, where is the timeout value is reached. Thus, the internal calculations
performed by DOR are eectively accounted. As in Sect. 3, a training set of ve
cases is used in the tness function.
First of all, Fig. 3 shows how the hybrid EA is much more successful in solving
the training cases. As it can be seen, while standard operators only provide an
acceptable performance on the smallest instances and with the lowest obstacle
density, DOR consistently yields satisfactory results: above a 70 of the runs
provide a fully successful solution for the training set the percentage is 100
for 5 out of 9 test worlds. The exception is world W50c for which none of the
operators could nd a full solution it must be noted that such a solution may
p
p
:
:
:
3 5 4
C . C o tta a n d J .M . T r o y a
Fig. 3. Number of runs in which each operator provided a fully satisfactory solution
for the training set.
not exist. Nevertheless, DOR was capable of solving 3 out of the 5 training
cases while SPX could only solve one and UX could not solve any of them.
Table 2 shows a more detailed summary of the results. Notice that DOR
is not only more eective in nding satisfactory solutions, but also provides
higher-quality results. By comparing the median values2 provided by DOR with
the optimal best-known solutions see Table 1, it can be seen that DOR yields
near-optimal solutions. Moreover, the lower variance of DOR results with respect
to SPX and UX indicates a more stable algorithm.
Table 2. Comparison of dierent genetic operators on nine dierent environments. All
results correspond to series of forty runs.
SPX
UX
DOR
World Timeout mean
median mean
median mean
median
W10a
17.36 12.06 16.00 15.88 7.47 15.20 13.85 0.87 13.80
W10b 25
22.15 25.96 16.20 18.10 9.26 16.20 14.35 0.74 14.20
W10c
51.49 38.39 59.60 53.85 38.81 59.60 18.75 10.75 14.00
W25a
176.70 210.00 45.60 121.63 193.32 40.00 33.17 2.14 32.40
W25b 150 376.08 285.43 399.30 380.96 272.08 468.40 36.65 6.67 34.00
W25c
520.32 179.07 616.80 502.23 218.40 616.40 99.92 107.70 38.80
W50a
442.39 348.27 438.00 423.87 335.45 438.00 61.17 11.72 58.80
W50b 400 1532.52 554.56 1636.60 457.65 457.65 1636.60 337.32 506.03 103.8
W50c
2021.45 60.77 2031.20 2031.20 0.00 2031.20 1638.78 538.75 1650.60
2
The median value seems to be a more representative measure of the quality of the
results than the mean value since the former is much less sensitive to outliers. Furthermore, it provides an reasonable alternative to averaging the tness of solutions
that solve the whole training set with solutions that do not solve any training case.
Using a Hybrid Evolutionary- A* Approach for Learning Reactive Behaviours
3 5 5
Fig. 4. Percentage of the test set solved for each of the techniques considered.
Finally, the results obtained with the EA are tested for generality. Fig. 4
shows the results. Firstly, notice the poor results of standard EAs. The solutions
provided by UX and SPX do not reach 50 success in 6 out of 9 worlds. The
A* algorithm performs better than standard EAs, but its performance quickly
drops when the density of obstacles is increased. The hybrid EA provide the
overall best results, outperforming both A* and standard EAs on all worlds.
Moreover, this improvement is larger on instances with higher obstacle densities.
It must be noted that the results on W50c are not satisfactory for any algorithm
although the hybrid algorithm remains the best . This is a really hard instance
as mentioned before, and may require longer evolution times and or a larger
training set to cope with such a tough environment.
6 Conclusions
This work has presented a hybrid approach for learning reactive rule-bases. By
combining EAs with the A* algorithm, a synergetic system has been achieved.
This hybrid algorithm has been shown to provide higher-quality results than
standard EAs. These results are also better than those of the A* algorithm in
terms of their generalisation to previously unseen test cases. Furthermore, the
hybrid EA is capable of tackling instances in which the A* algorithm would
suer the eects of the combinatorial explosion.
Future work will try to extend these results to more sophisticated agents.
In this sense, notice that most details of the agent are encapsulated within the
simulation function and hence they do not aect the presented algorithm
qualitatively. Nevertheless, it is clear that issues regarding simulations of higher
computational cost are worth studying. Work is in progress in this area. Additionally, new environments and tasks to be solved will be tackled as well.
3 5 6
C . C o tta a n d J .M . T r o y a
Acknowledgement
This work is supported by the Spanish
Tecnologa
Comision Interministerial de Ciencia y
CICYT under grant TIC99-0754-C03-03.
References
1. K. Ali and A. Goel. Combining navigational planning and reactive control. In
Theories of Action, Planning, and Robot Control. Bridging the Gap: Proceedings
of the 1996 AAAI Workshop, pages 19, Menlo Park, CA, 1996. AAAI Press.
2. C.T.C. Arsene and A.M.S. Zalzala. Control of autonomous robots using fuzzy logic
controllers tuned by genetic algorithms. In Proceedings of the 1999 Congress on
Evolutionary Computation, pages 428435. IEEE NNC - EP Society - IEE, 1999.
3. T.D. Barfoot and D'Eleuterio G.M.T. An evolutionary approach to multiagent
heap formation. In Proceedings of the 1999 Congress on Evolutionary Computation,
pages 420427. IEEE NNC - EP Society - IEE, 1999.
4. C. Cotta, E. Alba, and J.M. Troya. Utilising dynastically optimal forma recombination in hybrid genetic algorithms. In A.E. Eiben, Th. Back, M. Schoenauer, and
H.-P. Schwefel, editors, Parallel Problem Solving From Nature V, volume 1498 of
Lecture Notes in Computer Science, pages 305314. Springer-Verlag, Berlin, 1998.
5. C. Cotta and J.M. Troya. Tackling epistatic problems using dynastically optimal recombination. In B. Reusch, editor, Computational Intelligence. Theory and
Applications, volume 1625 of Lecture Notes in Computer Science, pages 197205.
Springer-Verlag, Berlin Heidelberg, 1999.
6. L. Davis. Handbook of Genetic Algorithms. Van Nostrand Reinhold Computer
Library, New York, 1991.
7. J.-Y. Donnart and J.-A. Meyer. Learning reactive and planning rules in a motivationally autonomous animat. IEEE Transactions on Systems, Man, and Cybernetics, 263:381195, 1996.
8. F. Homann and G. Pster. Learning of a fuzzy control rule base using messy
genetic algorithms. In F. Herrera and J.L. Verdegay, editors, Genetic Algorithms
and Soft Computing, pages 279305. Physica-Verlag, Heidelberg, 1996.
9. J.R. Koza. Genetic Programming. MIT Press, Cambridge MA, 1992.
10. N.J. Radclie. The algebra of genetic algorithms. Annals of Mathematics and
Arti cial Intelligence, 10:339384, 1994.
11. A.C. Schultz and J.J. Grefenstette. Using a genetic algorithm to learn behaviours
for autonomous vehicles. In Proceedings of the AIAA Guidance, Navigation and
Control Conference, pages 739749, Hilton Head SC, 1992.
12. C. Thornton. Learning where to go without knowing where that is: the acquisition
of a non-reactive mobot behaviour by explicitation. Technical Report CSRP-361,
School of Cognitive and Computing Sciences, University of Sussex, 1994.
13. G. Weiss. Multiagents Systems: a Modern Approach to Distributed Arti cial Intelligence. The MIT Press, Cambridge MA, 1999.
14. D.H. Wolpert and W.G. Macready. No free lunch theorems for search. Technical
Report SFI-TR-95-02-010, Santa Fe Institute, 1995.
15. B. Yamauchi and R. Beer. Integrating reactive, sequential and learning behaviour
using dynamical neural networks. In D. Cli, P. Husbands, J.-A. Meyer, and S. Wilson, editors, From Animals to Animats 3: Proceedings of the Third International
Conference on Simulation of Adaptive Behaviour, pages 382391, Cambridge MA,
1994. MIT PressBradford Books.
Supervised Evolutionary Methods in
Aerodynamic Design Optimisation
D.J. Doorly, S Spooner and J. Peiro
Aeronautics Department, Imperial College, London SW7 2BY, UK
Abstract. This paper outlines the application of evolutionary search
methods to problems in aeronautical design optimisation. The procedures described are based on the genetic algorithm GA and may be
applied to other areas. Although easy to implement, a simple genetic algorithm is often found in applications to be of low e ciency and to su er
from premature convergence. To improve performance, two alternative
strategies are investigated. In the rst, a learning classi er scheme is
used to tune the GA for a particular class of problems. The second strategy uses a parallel distributed genetic algorithm supervised by single or
competing agents. The implementation of each procedure, and results for
typical design problems are outlined. The agent supervised distributed
genetic algorithm is found to provide a model with a very high degree of
adaptibility, and to lead to considerably improved e ciency.
1
Introduction
The principles of evolutionary computation are well established and are described
in texts such as 1, 3, 2, 4 . These methods continue to grow in diversity, and
are becoming more commonly used for many problems in engineering design. In
aerospace vehicle design, the genetic algorithm GA has been applied to a range
of problems, as described for example in 5, 6, 7, 8, 9 . The apparent robustness
of the GA, the ease with which it can be applied, and its ability to handle
discontinuous or even discrete data make it attractive as a search procedure.
There are other search procedures however including various gradient search
methods which are more ecient than the GA, albeit that they may only
work for a more restricted class of problems. Furthermore, although evolutionary
methods have been found to be quite adept at locating global optima in highly
multimodal problems, there is usually no guarantee that they will, and their
rate of convergence to the optimum solution can be very slow. Ideally one would
like an optimisation procedure to work well across a broad range of problems.
Bearing in mind the no free lunch `rule' however, some degree of matching of
technique to problem seems inevitable.
Three simple ways of improving an evolutionary search procedure for a particular type of application are:
1. tuning the parameters to suit the class of problem,
2. improving its ability to adapt to the problem,
3. hybridisation with other techniques appropriate to the class of problem.
S . C a g n o n i e t a l. ( E d s .) : E v o W o r k s h o p s 2 0 0 0 , L N C S 1 8 0 3 , p p . 3 5 7 − 3 6 6 , 2 0 0 0 .
S p rin g e r-V e rla g B e rlin H e id e lb e rg 2 0 0 0
3 5 8
D .J . D o o r ly , S . S p o o n e r , a n d J . P e ir ó
We discuss only the rst two of these here, and will use the GA as our basic
evolutionary procedure. The construction of a suitable hybrid is very problem
dependent, and we consider that an improved GA may either outperform a
hybrid routine or may at worst lead to more eective hybrids. Hybrid routines
will in any case be discussed elsewhere. The rest of the paper is organised as
follows. The common design problem of shape or form optimisation is introduced
in the context of aerodynamic or aeroelastic design. The application of a real
encoded GA to a typical problem is also briey outlined. After this, we examine
the use of a learning classier system to tune the GA parameters, for a particular
class of such design problems. The distributed genetic algorithm or DGA is then
introduced, and nally the use of agent and multi-agent supervision to improve
the adaptibility of the DGA is described.
2 Shape optimisation & outline of basic GA
A frequent task in aeronautical design is to nd the `best' aerodynamic shape
of an airfoil in 2D or a wing in 3D, subject to certain constraints. The basic
GA we use for this task is constructed as follows.
The encoding species the shape of a trial solution, though it may also specify structural parameters such as type of material, weight, rigidity etc. For
simplicity, let us consider the case of a 2D wing section or airfoil. Then an
array cj may be used to specify the ordinates of a B-spline control polygon,
which in turn describes an airfoil shape, as shown in g. 1.
We use a real array of 20 control point ordinates to encode the shape. For
...
...
C h ro m o so m e
P o p u la tio n
k
...
...
C h r o m o s o m e a r r a y : C k = ( Y 1 , ... ,Y n )
j
i
0.06
0.04
Y
y/c
0.02
i
0.0
Y
-0.02
j
-0.04
-0.06
0.0
Fig. 1.
0.1
0.2
0.3
0.4
0.5
x/c
0.6
0.7
0.8
0.9
1.0
Representation of airfoil geometry.
aeroelastic optimisation of wings instead 13, we encode the design as a sur-
S u p e rv is e d E v o lu tio n a ry M e th o d s in A e ro d y n a m ic D e s ig n O p tim is a tio n
3 5 9
face interpolated between a series of spanwise sections, with the encoding
specifying a standard basic prole, together with the chord, thickness, twist
and structural parameters for each section.
An initial population of random airfoils is generated, each initial shape is
ensured to be valid, i.e. encodings which produce surface crossings or violate
thickness constraints are discarded.
The tness is evaluated using a CFD ow solver or a CFD + structural
solver. Either an unstructured mesh based solution of the Euler equations,
or a viscous inviscid panel ow solver is used, 11 . In a direct optimisation,
the search is for the shape which best meets the requirements e.g. high
lift drag L D, whereas in an inverse optimisation, the shape which best
matches a given pressure coe cient distribution p is sought.
Fitness values are range-scaled and remapped to the interval 0 1 , with
individuals below the mean tness assigned a base value typically 0 4,
and those above the mean assigned a value scaled quadratically up to 1.
Roulette wheel is used for selection in the basic GA and DGA operations,
though the learning GA allows other methods binary tournament etc.
Two point crossover is again the basic crossover type, with one point, uniform etc. allowed in the learning GA. Mutation and crossover are applied
as separate operators. The probability of mutation of a given gene is low,
typically 0.005, but is commonly increased at later stages.
Elitist population replacement is applied, where the best of each generation is automatically carried through, here together with a slightly mutated
copy of the best. The population replacement routine ensures that no overreplication i.e. excess identical or almost identical duplication occurs.
C
O
:
3 Classier Learning Directed GA
The classier learning system is described in standard texts on machine intelligence much of the research in this area also follows from ideas put forward by
Holland 14 . The classier learning system adds a layer on top of the GA. Rules
to control operators mutation crossover, type of crossover, selection scheme
etc., and parameter values e.g. mutation rate are prescribed, and their effectiveness when implemented in the GA are assessed. In a static mode, the
entire set of rules can undergo genetic operations to evolve better rules and rule
combinations. In a dynamic mode, a system of reward paybacks can be used to
determine rule selection.
The objective of the procedure here was to train the GA in a static mode
for a particular problem type. The e ciency of the trained and basic GAs were
then compared rstly when given problems of a very similar nature to those
used in training, and secondly for slightly dierent problems. In the learning
GA outlined in g. 2, an initial population of 30 rules was used, which were
randomly initialised. The rules were encoded as 15 genes in an IF-THEN-AND
conguration, and were designed to work in groups, with the nal 2 genes being
3 6 0
D .J . D o o r ly , S . S p o o n e r , a n d J . P e ir ó
L e a r n in g G A
G A r u le c o n tr o l
S T A R T
G A
T e s t n e w r u le s e t
&
F lo w S o lv e r
M a tu r e r u le s
E n d :
D e v e lo p e d s o lu tio n
Fig. 2.
Classier Learning SystemGA
index references to other rules to be used in combination with the current one.
The rule gene string comprised:
two IF operators, with parameter values determining the generation number,
and average gene diversity respectively which triggered operation of the rule,
the following 11 gene values type THEN set operator type and parameters,
the nal two type AND determined the combination of rules to be used.
The procedure was repetitively applied to the inverse design of respectively
symmetric and non-symmetric airfoils. To recall, in inverse design, the problem
is to nd the shape which matches a specied pressure coe cient Cp distribution it is often required in real applications, and for comparative tests of search
methods purposes it is preferrable to direct design, as the target is given. Each
rule was tested twice, and the top 10 peforming rules were then isolated and
repetitively tested over 15 separate runs.
The GA was trained for inverse optimisation of symmetric and non-symmetric
airfoil sections, respectively using NACA airfoil types -0012,-0022,-0024 for symmetric training, and -23015, -4421 for non-symmetric training. Figure 3 shows
results for the inverse design of a non-symmetric airfoil 23015, comparing the
performance of the GA using: rules developed for this type of airfoil, rules developed for a dierent symmetric type, and the basic untrained GA. It can be
seen that the trained GA outperforms the baseline GA for inverse optimisation
whether it is trained on a class of similar airfoils, or a dissimilar class. However, the dierence in performance resulting from the class of problem used for
training shows how very specic training may be needed to obtain the highest
gains.
At present, the cost of CFD evaluations is generally so high, that the benets of training do not appear worthwhile. However other possibilities, such as
using a simpler approximate evaluation method for training only, or reducing
the complexity of the scheme, and hence degree of training, may yet render
it more practical in this area. Also, the classier system may still prove useful
within the context of an agent supervised DGA.
A v e ra g e b e s t fitn e s s (1 5 ru n s )
S u p e rv is e d E v o lu tio n a ry M e th o d s in A e ro d y n a m ic D e s ig n O p tim is a tio n
3 6 1
R u le s fro m n o n s y m m e tric a irfo il
tra in in g
R u le s fro m
s y m m e tric a irfo il
tra in in g
}
0
2 0 0
B a s e lin e
G e n e ra tio n n o .
Application of learning GA to inverse design of NACA-23015 airfoil. Comparison of performance with rule sets developed for similar airfoils, for di erent airfoils,
and basic GA performance.
Fig. 3.
4 Distributed Genetic Algorithm DGA
Previous work has already shown that the DGA outperforms the GA on many
test problems and in design optimisation, 10, 16. Applications of the DGA
to aeronautical design problems are also described in 11, 13. Brie y the DGA
di ers from the standard GA in that the population of trial solutions is split into
semi-isolated subpopulations or `demes'. The demes are considered analogous to
island populations, where geography acts as a barrier to exchange. Restrictions
on the recombination and genetic exchange between subpopulations are imposed
the exchange is limited to the migration of a few individuals often only the best
one or two from one neighbouring deme to another every m generations, with
typically m = 5 or m = 10 in our implementations.
Thus the parameter set for the basic GA operator probabilities, selection
mode, etc. is enlarged to include the number of individuals migrating, barriers
to acceptance of immigrants, exchange frequency, geographical exchange radius
and topology of the demes. With limited exchanges between demes, the DGA is
then ideally suited for coarse grain parallelisation g.4 whether on a parallel
supercomputer or network of workstations. Nang 12 surveys the parallel GA,
of which the `stepping stone' connected DGA shown in g. 4 is one type.
For a workstation network 11, provided the ow solution can be run on a
single workstation, the only communication required between processors involves
the exchange of a limited number of chromosomes, at intervals of several generations. The communication requirements are extremeley low, given that by far
the bulk of the computational e ort is devoted to the ow solution which performs the evaluation. Each processor may be responsible for a number of demes.
For a heterogeneous network of processors of di erent speeds or loadings, load
balancing can be achieved by varying the mapping of demes to processors, or by
altering the number of individuals treated by a given processor.
3 6 2
D .J . D o o r ly , S . S p o o n e r , a n d J . P e ir ó
C F D
Fig. 4.
G A
C F D
G A
( I ,J )
D E M E
( I + 1 ,J )
D E M E
C F D
G A
( I ,J - 1 )
D E M E
Distributed GA DGA mapped to processor array.
4.1 Application to aeronautical design
Inverse design Comparison of the convergence behaviour obtained using a distributed 13 and a conventional GA, with the same total population and number
of evaluations, for the viscous inverse design of a NLF1 -0115 airfoil is shown in
g. 5. For the DGA, the population of 180 was distributed onto 9 subpopulations
of 20, each residing on a dierent processor, as in g.4 above. The migration between the islands occurred in a stepping stone fashion, ie. migration occurred
only between immediate neighbours every ve generations. As the results indicate, the DGA greatly outperforms the conventional GA algorithmically. This
gain is then further multiplied almost ninefold with the distributed processor
implementation. The implementation was done using the MPI standard on a
network of workstations it has also been implemented on a multiprocessor machine .
Comparison of convergence rate of parallel DGA and single population version.
If computing speed up rate nearly 9 were applied, parallel DGA performance gain
would appear even more dramatic
Fig. 5.
S u p e rv is e d E v o lu tio n a ry M e th o d s in A e ro d y n a m ic D e s ig n O p tim is a tio n
3 6 3
Direct Airfoil Optimisation Results Application of the DGA to problems
of direct airfoil optimisation for low speed and transonic cases are described in
6, 11. For both inverse and direct optimisation problems however, the use of
an agent to supervise the operation of the DGA has been found to improve the
performance, as described next.
5 Agent Supervision of DGA
The better ability of the DGA at maintaining population diversity appears to
account for its notable gain in performance over the single population GA. Eventually however, the population on each island converges. Adding an agent to supervise the operation of the parallel GA provides a capability whereby the DGA
can adapt more generally than is possible with a sequential GA. Although adaptation can be built into a GA e.g. in the adaptive operator tness of Davis 4,
the use of agents provides a more general framework by decoupling the tasks of
higher level supervision from the lower level optimisation. The agent supervised
paradigm is very well suited to a distributed computing environment, where
agents can direct the operation of the GA on local or global populations, and
can additionally direct processing resources. At one extreme, the agent layer may
be combined with the DGA software to execute as a single albeit distributed
entity, or at another, it may run as an entirely separate distributed program,
communicating with the DGA by reading external output messages and writing
to action inputs.
A simple agent supervision of the DGA is as follows. The agent receives
status messages from the DGA, i.e. generation number, measures of population convergence, tness changes etc. , and parameter settings local mutation
and crossover rates types etc. . The agent layer then instructs the DGA to take
either global or local action. Examples of such actions could be the introduction of a mechanism to improve diversity between island populations, actions to
favour speci c local niches within islands, or actions to improve the parallel load
balancing by adjusting the deme placing or sizing.
5.1 Infection Agent
A simple implementation of agent supervision of the DGA which was used for
direct airfoil design employs a single agent supervisor 15 to act as a vector
for infections, with low population diversity encouraging epidemics. On infected
islands, individuals close to the global best have greatly reduced tness, and
undergo increased mutation results 15 shows the pattern of infection changes
dynamically. When implemented for the problem of inverse design optimisation, there was a gain in late solution convergence, beginning just beyond the
point when the population on all the islands initially converges towards a global
`champion'.
Applying the procedure to direct design viscous L D optimisation of a low
speed airfoil section, at operating points of 3 and 8 degrees incidence, and a
3 6 4
D .J . D o o r ly , S . S p o o n e r , a n d J . P e ir ó
Reynolds number of 4 million, with a moment constraint Cm 0:97 produced the shape and Cp distributions shown in g. 6, though this is not yet
fully a converged solution.
j
j
Fig. 6. LD optimization: Distribution of Cp on the surface of the `best' airfoil. This
computation was performed using XFOIL.
The addition of the infection in these problems was found to show a clear
improvement over the solution obtained in 11 for corresponding eort the
improvement in the tness after island convergence is shown in g. 8a further
below.
5.2 Evolutionary Agent Supervision
In the learning classi er method presented earlier, adaptation occurs through
repetitively solving a problem or class of problems. The agent supervision can
adapt dynamically however, especially if a large number of generations are to be
evaluated.
Agent or
Multi-Agent
A
COMPETE
A
A
Actions
(Local or Global)
DGA
Messages
Fig. 7.
Messages
Competing agent supervised DGA
For example, one may place several islands in a group under the control of one
agent, and other groups of islands under the control of other agents, with each
agent adopting dierent strategies. If migration between groups is eliminated, the
S u p e rv is e d E v o lu tio n a ry M e th o d s in A e ro d y n a m ic D e s ig n O p tim is a tio n
3 6 5
relative improvement over a number of generations may be compared. The worse
performing agents then modify either the parameter settings or the rules which
they apply to the DGA under their control, and may also replace their population
partly or fully with that of the best group. An example of the eectiveness of the
approach is shown in g.8b, where the simple agent infection approach described
previously is supervised by a pair of competing agents. A population of 200 10
individuals per deme demes connected in a 4 x 5 array was used in this case
for the inverse design of a NACA 0012 airfoil. The agents apply dierent genetic
operators to their respective groups of subpopulations, and their performance
as managers is compared after a certain number of generations. In the example,
only the mutation rate was altered by the supervising agents at a lower level,
the infection strategy was still implemented. As can be seen from g. 8b, these
early results are encouraging further work will consider the eects of controlling
dierent parameters, and dierent competition mechanisms.
18
30
16
D G A + A g e n t
C o m p e tin g A g e n ts
25
14
20
In fe c tio n
10
F itn e s s
F itn e s s
12
8
6
15
10
D G A
4
S in g le A g e n t
5
2
0
0
0
50
100
150
200
250
300
350
400
0
50
100
150
200
G e n e ra tio n
G e n e ra tio n
a
b
250
300
350
400
Agent supervision: a Comparison of DGA and simple agent supervised DGA
bComparison of simple agent and competing agent supervised DGA. Note dierence
in tness scales results in b are also averaged over more trials
Fig. 8.
6
Conclusion
The distributed genetic algorithm DGA has been applied to a number of aeronautical design optimisation problems. Earlier results indicated that the method
has better convergence behaviour than the single population GA the present
work outlines agent supervision strategies to improve the DGA further. A learning classier scheme applied to the single population GA showed some improvement, but appears very costly at present. In contrast, the use of competing
agents which evolve appears quite promising for further investigation. Applica-
3 6 6
D .J . D o o r ly , S . S p o o n e r , a n d J . P e ir ó
tions of the DGA to low speed airfoil optimisation, demonstrate that the method
is straightforward to implement, and can be easily applied to dierent problems.
References
1. Goldberg D E, Genetic Algorithms in Search, Optimisation and Machine Learning,
Addison-Wesley, 1988.
2. Back T, Evolutionary Algorithms in Theory and practice, Oxford, 1996.
3. Schwefel H P, Evolution and Optimum Seeking, Wilrey New York, 1995.
4. Davis L, Handbook of Genetic Algorithms, Van Nostrand Reinhold, New York
1991.
5. Obayashi, S., Yamaguchi Y., and Nakamura, T. Multiobjective genetic algorithm
for multidisciplinary design of transonic wing planfoem, J. Aircraft 34, 5, pp 690
693, 1997
6. Doorly D J, Ch. 13 of Genetic Algorithms in Engineering and Computer Science,
ed. G. Winter et al., Wiley, 1995.
7. Quagliarella D and DellaCioppa A, Genetic Algorithms Applied to the Aerodynamic Design of Transonic Airfoils, J. Aircraft 32, 889891, 1995.
8. Poloni C, Ch. 20 of Genetic Algorithms in Eng. and Comp. Sci., ed. G. Winter et
al., Wiley, 1995.
9. Yamamoto K, and Inoue O, Applications of Genetic Algorithms to Aerodynamic
Shape Optimisation, AIAA-95-1650-CP, 1995
10. Tanese R, Distributed Genetic Algorithms, PhD thesis, U. Michigan, 1989.
11. Doorly D J, Peiro J, Kuan T, and Oesterle J-P, Optimisation of Airfoils Using
Parallel Genetic Algorithms, in Proc. 15th Int. Conf. Num. Meth. Fluid Dyn.,
Monterey, 1996.
12. Nang J and Matsuo K, A Survey of Parallel Genetic Algorithms, J. SICE 33, 6,
500509, 1994.
13. Doorly D J, Peiro J, and Oesterle J-P, Optimisation of Aerodynamic and Coupled Aerodynamic-Structural Design using Parallel Genetic Algorithms, in Proc.
Sixth AIAANASAISSMO Symposium on Multidisciplinary Analysis and Optimization, 401409, 1996.
14. Holland J H, Adaptation in Natural and Articial Systems, MIT Press, 1992.
15. Doorly D J and Peiro , Supervised parallel genetic algorithms in Aerodynamic
Optimisation, AIAA paper 97-1852, 1997.
16. Oesterle J-P, Aeronautical optimisation using parallel genetic algorithms, MSc thesis, Aeronautics Dept.,Imperial College London, 1996.
An Evolutionary Algorithm for Large Scale Set
Covering Problems with Application to Airline
Crew Scheduling
Elena Marchiori1 and Adri Steenbeek2
Free University
Faculty of Sciences, Department of Mathematics and Computer Science
De Boelelaan 1081a, 1081 HV Amsterdam, The Netherlands
1
elena@cs.vu.nl
CWI,
P.O. Box 94079, 1090 GB Amsterdam, The Netherlands
2
adri@cwi.nl
Abstract. The set covering problem is a paradigmatic NP-hard com-
binatorial optimization problem which is used as model in relevant applications, in particular crew scheduling in airline and mass-transit companies. This paper is concerned with the approximated solution of large
scale set covering problems arising from crew scheduling in airline companies. We propose an adaptive heuristic-based evolutionary algorithm
whose main ingredient is a mechanism for selecting a small core subproblem which is dynamically updated during the execution. This mechanism
allows the algorithm to nd covers of good quality in rather short time.
Experiments conducted on real-world benchmark instances from crew
scheduling in airline companies yield results which are competitive with
those obtained by other commercial academic systems, indicating the
e ectiveness of our approach for dealing with large scale set covering
problems.
1 Introduction
The set covering problem SCP is one of the oldest and most studied NP-hard
problems cf. 14 .
Given a m-row, n-column, zero-one matrix a , and an n-dimensional integer vector w , the problem consists of nding a subset of columns covering all
the rows and having minimum total weight. A row i is covered by a column j
if the entry a is equal to 1. This problem can be formulated as a constrained
optimization problem as follows:
ij
j
ij
minimize
P
n
j =1
wx
j
j
subject to the constraints
8 x 2 f0 1g
: P =1 a x
j
n
j
ij
j
j = 1 : : : n
1 i = 1 : : : m:
S . C a g n o n i e t a l. ( E d s .) : E v o W o r k s h o p s 2 0 0 0 , L N C S 1 8 0 3 , p p . 3 6 7 − 3 8 1 , 2 0 0 0 .
S p rin g e r-V e rla g B e rlin H e id e lb e rg 2 0 0 0
3 6 8
E . M a rc h io ri a n d A . S te e n b e e k
The variable xj indicates whether column j belongs to the solution xj = 1
or not xj = 0. The m constraint inequalities are used to express the requirement that each row be covered by at least one column. The weight wj is a
positive integer that speci es the cost of column j . When all wj 's are equal to
1, then the SCP is called unicost SCP.
Relevant practical applications of the SCP include crew scheduling 1, 2, 12,
15: nd a set of pairings having minimum-cost which covers a given set of trips,
where a pairing is a sequence of trips that can be performed by a single crew.
A widely used approach to crew scheduling works as follows. First, a very large
number of pairings is generated. Next, a SCP is solved, having as rows the trips
to be covered, and as columns the pairings generated. When this approach is used
in mass-transit applications, very large scale SCP instances may arise, involving
thousands of rows and millions of columns.
The most successful heuristic algorithms for large scale SCP's are based on
Lagrangian relaxation 13. Lagrangian relaxation is used to compute the score of
a column according to its likelihood to be selected in an optimal solution. These
scores are employed in simple greedy heuristics for computing a solution. A very
eective heuristic algorithm for large scale SCPs based on this approach is 7.
We refer the reader to 8 for a recent survey on exact and heuristic algorithms
for SCP. All eective heuristics for large scale SCP's act on a subset of the
columns, called core, which is selected before the execution of the algorithm. In
the static approach the core remains the same during the execution cf. 6, 11,
while in the dynamic approach it is updated using an adaptive mechanism e.g.
7, 10, 9.
In this paper we propose a novel heuristic algorithm for large scale SCPs
arising from crew sheduling problems in airline companies. At each iteration a
near optimal cover is constructed using the information provided by the previous
iterations to guide the search. The nal solution is the best cover found in all
the iterations.
Given a problem instance, the algorithm extracts an initial core from the
set of columns given in the input. Then the algorithm consits of the iterated
application of the following three steps: 1 First, an approximated solution to
the actual SCP core is constructed by means of a novel greedy heuristic. 2
Next, a local search optimization algorithm is applied to the resulting solution.
3 Finally, some columns that occur in the best solution found in all iterations up
to now are selected for forming the initial partial solution for the next iteration.
The size of the core is determined by an adaptive size parameter, while
the selection of a column is speci ed by a suitable merit criterion. During the
execution, the score of the columns is modi ed as well as the size parameter,
and the core is dynamically updated.
This algorithm can be viewed as a hybrid 1 + 1 steady-state evolutionary
algorithm, where at each iteration a child is generated from the parent using
the above described heuristic, and the best between the parent and the child
survives.
A n E v o lu tio n a ry A lg o rith m
fo r L a rg e S c a le S e t C o v e rin g P ro b le m s
3 6 9
In order to assess the performance of the algorithm, we conduct extensive
experiments on real-world problem instances arising from crew scheduling in airlines, as well as on other benchmark instances from the literature. The results
of the experiments are rather satisfactory: our algorithm is able to nd covers
of very good quality in a short amount of time, yielding results which are competitive with those reported by the best industrial as well as academic methods
for solving large set covering problems.
The rest of the paper is organized as follows. In the next subsections we
briey discuss some related work, and set up the notation and terminology used
throughout the paper. In Section 2 we introduce the overall method and present
in detail the four main modules of the algorithm. In Section 3 we report the
results of extensive computational experiments. We conclude with some nal
remarks on the present investigation and on future work.
1.1 Related Work
An experimental comparison of the most e ective exact and heuristic algorithms
for the weighted SCP is given in a recent paper by Caprara et al 8 .
A rather e ective heuristic algorithm based on Lagrangian relaxation is the
CFT algorithm 7 by Caprara et al. This algorithm has been tested also on
large scale problem instances arising from crew scheduling in railway, yielding
rather satisfactory results. In 15 a approximation algorithm for solving large
0-1 integer programming problems is proposed. This algorithm is used in the
CARMEN system for airline crew scheduling, a industrial system used by several
major airlines.
Research based on evolutionary computation includes the following two papers.
Beasley and Chu in 6 introduce a genetic algorithm for the SCP. The authors
employ a representation where a chromosome is a bit string of lenght equal to
the number of columns, one bit for each column, representing the set of columns
whose bit in the string are equal to 1. The algorithm employs a heuristic repair
mechanism for transforming infeasible chromosomes into solutions. Moreover, a
core is used for constructing the chromosomes of the initial population.
A genetic algorithm based on a non-binary representation has been proposed
by Eremeev in 11 . Here a chromosome is a string of lenght equal to the number
of rows, where the i-th entry contains the index of a column covering the ith row. As a consequence, all chromosomes are feasible solutions, thus they do
not need to be repaired as in 6 . Moreover, heuristics are used for eliminating
redundant columns as well as for dening the crossover operator.
In Section 3 we will compare experimentally the above mentioned algorithms
with the algorithm introduced in this paper.
1.2 Notation and Terminology
In order to describe our method, we use the following terminology and notation.
3 7 0
E . M a rc h io ri a n d A . S te e n b e e k
In the sequel, the indexes i j denote a generic row and column, respectively. A
column will also be denoted by c, and a row by r, possibly subscripted. Moreover,
S denotes a set of columns.
Let cov S be the set of rows that are covered by the columns in S :
cov S = fi j a = 1 for some j 2 S g:
ij
For simplicity, we write cov j instead of cov fj g. We say that a column j
is redundant with respect to S if cov S n fj g = cov S .
A partial cover also called partial solution is a set of columns containing
no redundant column.
Let cov j S be the set of rows which are covered by column j , but are not
covered by any column in S n fj g:
cov j S = fi j a = 1 and a = 0 for all j 2 S n fj g g:
Moreover, let min weighti be the minimum weight of the columns that
cover i:
min weighti = minimum fw j i 2 cov j g:
We can now de ne the function cov val , called cover value, which is used to
evaluate a column j with respect to a partial cover S in order to select a column
to be added resp. removed to resp. from S :
0
ij 0
ij
j
cov val j S =
X
i2cov jS
min weighti:
A convenient property of cov val is that cov val j S = cov val j S n fj g.
This allows one to compute the cover value of a column without taking into account whether it belongs to the partial solution S or not. Moreover, we can characterize the redundancy of a column by means of the condition cov val j S = 0.
The cover value is used to de ne the selection value sel val j S of a column
j with respect to the partial cover S :
if j redundant wrt S ,
sel val j S = Lim
w =cov val j S otherwise.
The selection value of redundant columns is set to a very big constant Lim.
In this way, redundant columns do not have any chance of being selected.
j
2 The Overall Method
The algorithm we propose consists of an iterated procedure, where each iteration
generates an approximated solution using only columns from the actual core.
Roughly, at each iteration a greedy heuristic is used to construct incrementally a cover starting from a partial cover: in the rst iteration the partial cover
is empty, while in the following iterations the partial cover is a proper subset
of the best cover found in all iterations up to now. The cover found after the
A n E v o lu tio n a ry A lg o rith m
fo r L a rg e S c a le S e t C o v e rin g P ro b le m s
3 7 1
application of the greedy heuristic is given as input to an optimization procedure
which tries to improve the partial solution. The core is updated from time to time
during the execution. The nal result is the best cover found in all iterations.
The corresponding algorithm WSCP Weighted Set Covering Problem is
illustrated below in pseudo-code, where Sbest represents the best cover found
so far, S denotes the actual partial
P cover, and valueS is the sum of the weights
of the columns of S, that is, j2S wj . Therefore the optimal cover is the cover
S having minimum valueS.
FUNCTION WSCP
BEGIN
RECOMPUTE_CORE
Sbest - 1..ncol
S -
FOR 1 .. param.number_of_iterations DO
IF core_selection RECOMPUTE_CORE ENDIF
S - GREEDYS
S - OPTIMIZES
IF valueS = valueSbest THEN Sbest - S ENDIF
S - SELECT_PARTIAL_COVERSbest
ENDFOR
RETURN Sbest
END
2.1 Greedy Heuristic
Our greedy heuristic GREEDY is described in pseudo-code below. Lines starting
with "==" are comments. The algorithm constructs a solution a cover, starting
from a possibly empty partial cover S . Columns are added resp. removed to
resp. from S until S covers all the rows.
extend S until it is a cover:
FUNCTION GREEDY var S
BEGIN
WHILE S is not a cover DO
select and add one column to S
S - S + select_add
remove 0 or more columns from S
WHILE remove_is_okay DO
S - S - select_rmv
ENDWHILE
ENDWHILE
S is a cover, without redundant columns
return S
END
3 7 2
E . M a rc h io ri a n d A . S te e n b e e k
The function select add selects a column j not in S having minimum selection
value sel val j S .
The test remove is okay determines whether columns should be removed
from S . If S is empty it returns false if S contains at least one redundant
column then it returns true otherwise, with probability param:p rmv typical
value 0:3 it returns true, otherwise false.
Finally, the function select rmv selects a column in S having maximum selection value.
2.2 Local Optimization
The local optimization procedure OPTIMIZE is based on the following idea.
Given a cover S , suppose there is a column j 62 S such that S fj g contains
at least two columns other than j , say j1 : : : j , with l 2 that are redundant,
and
of their weights is greater than the weight of j , that is,
P such wthatk thew .sum
Then
S n fj1 : : : j g fj g is a better cover than S . In this
=1
case we call j a superior column. The gain of j is de ned by
l
k
:::l
j
j
l
X
gain j =
k
=1
w k ,w :
j
j
:::l
So a best superior column is the one having highest gain. Note that the optimization procedure operates on a cover containing no redundant columns. The
optimization algorithm OPTIMIZE in pseudo-code is given below.
S is a cover, without redundant columns
FUNCTION OPTIMIZE var S
BEGIN
Sup - select_superior
WHILE Sup not empty DO
select best column from Sup
best - select_best
Sup - Sup - best
add superior and remove redundant columns from S
IF best superior
S - S + best
S - S - select_redundant
ENDIF
ENDWHILE
S is a cover, without redundant columns
return S
END
First, the function select superior is used, which generates the list Sup
consisting of all the superior columns ordered in decreasing order according to
their gain. Next, the list Sup is scanned in the WHILE loop. At each iteration, the head of Sup is removed and memorized in the variable best using
A n E v o lu tio n a ry A lg o rith m
fo r L a rg e S c a le S e t C o v e rin g P ro b le m s
3 7 3
the function select best. If the selected column is still superior that is if the
test best superior is satised then it is added to S , and the set of redundant columns are removed from the resulting partial cover S using the function
select redundant.
2.3 Restoring Part of the Actual Best Solution
In the rst iteration of WSCP the heuristic GREEDY constructs a cover starting from the empty set in the following iterations, GREEDY builds a cover
starting from a subset of the best cover found so far. For a column j , we keep
track of the number chosen j of times that j has been part of a best solution. The function SELECT PARTIAL COVER considers the set E of so-called
elite columns, consisting of those columns j of the best solution Sbest such that
cov val j Sbest wj . Then SELECT PARTIAL COVER selects from E the
set of columnsPhaving low chosen j in our implementation chosen j has to be
smaller than j chosen j =neli 10, where neli is the number of elements of
E while the remaining columns of E are selected with a probability that is set
to a random value between 0:1 and 0:9.
2.4 Selecting the SCP Core
This is a fundamental step in the design of an algorithm for dealing with large
SCP instances. We introduce the following method for constructing an SCP core,
which has been implemented in the function RECOMPUTE CORE.
The SCP cover is constructed from the empty set by incrementally adding
columns according to the following criterion. Columns are selected in increasing
order according to their selection value. Suppose column j has been selected:
1. if j is an elite column then with probability close to 1 it is added to the
actual SCP core
2. otherwise, j is added if there exists a row i such that j covers i and wj
min weighti K0 , with K0 a given constant real value greater or equal
than 1
3. otherwise, j is added if there exists a row i such that j covers i and i is
covered by less than K1 columns of the actual SCP core, with K1 a given
constant integer value greater or equal than 1.
Note that K0 K1 are parameters which are chosen depending on the class
of problems one considers. Condition 3 implies that the SCP core contains for
each row, at least the rst K1 best columns according to the ordering induced
by the selection value function that cover that row.
The function core selection determines when the actual SCP core has to
be recomputed. In our implementation, we recompute the SCP core every 100
iterations of WSCP.
During the execution of GREEDY, when ninety per cent of a cover has been
constructed, the min weight of those rows that are not yet covered is increased
3 7 4
E . M a rc h io ri a n d A . S te e n b e e k
by a small quantity in our implementation min weighti is multiplied by 1:1.
This aects the selection value of the columns, hence their order of selection in
the construction of the SCP core changes during the execution of the overall
algorithm WSCP.
3 Experimental Evaluation
The algorithm WSCP has been tested on large set covering problems arising
from crew scheduling applications in various airline companies. Moreover, we
have considered the weighted SCP instances from the OR library maintained by
J.E. Beasley 1 . These instances are considered standard benchmarks for testing
the eectiveness of exact and heuristic algorithms for the SCP. In particular, they
have been used in 8 for comparing experimentally various exact and heuristic
algorithms for SCP.
WSCP has been implemented in C++. The algorithm was run on a Sun
Ultra 10 UltraSPARC-IIi 300MHz.
The results of the experiments are based on 10 runs on each problem instance
of the OR Library, and on 5 runs on the other instances. In each table, the
entry labeled Id contains the name of the problem instance. The label BK
denotes the best known solution for that instance Bst denotes the best result
found by the algorithm Fbst indicates the frequency of obtaining the best
solution
runs Apd denotes the average percentage deviation
Pk=110 zink ,thez performed
=10 z 100, where zk is the solution found in the k-th run,
and z is the optimal or best known solution. Tbst denotes the average cpu
time for obtaining the best solution Bst, while Tsol denotes the average cpu
time for nding a solution. Finally, Ibst and Isol denote the average number of
iterations of obtaining the best solution Bst, and a solution, respectively.
3.1 Experiments on Airline Crew Scheduling Problem Instances
We consider three sets of benchmark instances from real-world airline crew
scheduling problems. A set of instances from a major airline company, here
called AIR instances, the airline scheduling instances from Wedelin 15 , and the
instances from Balas and Carrera 3 . The characteristics of these problems are
reported in Tables 2, 3, and 7, respectively. Observe that in many instances, like,
e.g., the Wedelin instances, the weights of the columns are very large numbers,
because the weight represents the cost of a pairing and takes into account several
factors.
We compare experimentally WSCP with the industrial system used by an
airline company on the AIR instances, with the CFT algorithm by Caprara et al
7 , and with the Wedelin algorithm 15 . The results of the experiments are given
in Tables 2, 4, 5, and 7. Note that the results for the Wedelin and CFT algorithms
are taken from the paper 7 , where the cpu time is estimated in DECstation
1
see http:mscmga.ms.ic.ac.ukjeborlibscpinfo.html
A n E v o lu tio n a ry A lg o rith m
fo r L a rg e S c a le S e t C o v e rin g P ro b le m s
3 7 5
5000240 CPU seconds. Only the value of the best solution is reported. For the
CFT algorithm, the time for nding the best solution is given, while for the
Wedelin algorithm, only the overall execution time Texe of the algorithm is
reported. The authors do not specify the setting of the various parameters in
their algorithms, and the total number of trials performed.
Id Rows Columns Density Weight Range
A01
A02
A03
A04
A05
A06
A07
A08
A09
A10
A11
A12
258303
0.167
1319-35302
19441
0.135
1437-37206
40580
0.092
1337-37148
79481
0.123
1460-37142
72377
0.126
1411-37251
23741
0.135
1437-37037
32363
0.15
1319-36370
45286
0.18
1345-36370
50047
0.19
1361-36370
49525
0.18
1344-36370
389388
5.55
1800-18768
642613
1.45
1630-19000
Table 1. Characteristics of AIR instances
Id Industry
Bst
A01 16351667
A02 12879297
A03 15663720
A04 16110608
A05 16315241
A06 13162511
A07 13301520
A08 13510606
A09 13489489
A10 13571530
A11 247775
A12 732587
5265
3878
4965
4916
4656
1971
4203
4320
4287
4369
150
682
WSCP
Bst Fbst Apd Tbst Tsol Ibst Isol
16351667 1.0 0.0 550.9 550.9 919.8 919.8
12879297 1.0 0.0 131.0 131.0 822.4 822.4
15663688 1.0 0.0 254.2 254.2 1004.4 1004.4
16110608 1.0 0.0 363.1 363.1 1085.0 1085.0
16315070 0.3 0.0001 923.5 848.1 3501.6 3203.2
13162511 1.0 0.0 156.9 156.9 907.6 907.6
13301520 1.0 0.0 200.9 200.9 945.6 945.6
13510584 1.0 0.0 254.4 254.4 946.2 946.2
13489489 1.0 0.0 235.4 235.4 944.2 944.2
13571530 1.0 0.0 237.3 237.3 933.8 933.8
247775 1.0 0.0 224.2 224.2 1087.8 1087.8
732587 0.3 0.11 1064.9 896.8 1460.6 1167.4
Table 2. Results for AIR instances
On three AIR instances WSCP found a solution which is better than the
best solution found by the industrial system, while on the other instances WSCP
found solutions of equal value as those found by the industrial system.
3 7 6
E . M a rc h io ri a n d A . S te e n b e e k
Id
B727scratch
ALITALIA
A320
A320coc
SASjump
SASD9imp2
Rows Columns Density Weight Range
29
157
8.2
1600-11850
118 1165
3.1
2200-2110900
199 6931
2.3
1600-2111450
235 18753
1.9
1900-1812000
742 10.370
0.6
4720-55849
1366 25032
0.3
3860-35200
Table 3. Characteristics of Wedelin instances
Id
CFT
Wedelin
Bst Tbst
Bst Texe
B727scratch 94400 94.400 0.3 94400 4.7
ALITALIA 27258300 27258300 6.2 27258300 37.2
A320
1262100 1262100 79.5 1262100 216.9
A320coc 14495500 14495600 577.8 14495500 1023.7
SASjump 7338844 7339537 396.3 7340777 806.8
SASD9imp2 5262190 5263640 2082.1 5262190 1579.7
Table 4. Results of CFT and Wedelin on Wedelin instances
Id
BK
BK
WSCP
Bst Fbst Apd Tbst Tsol Ibst Isol
B727scratch 94400 94400 1.0 0.0 0.018 0.018 38.4 38.4
ALITALIA 27258300 27258300 1.0 0.0 0.63 0.63 106.8 106.8
A320
1262100 1262100 1.0 0.0 17.34 17.34 326.2 326.2
A320coc 14495500 14495500 0.2 0.0006 651.08 446.20 3494.5 2402.0
SASjump 7338844 7339541 0.1 0.02 269.3 200.98 4635.0 3454.6
SASD9imp2 5262190 5263590 0.1 0.04 741.9 608.452 4603.0 3671.4
Table 5. Results of WSCP on Wedelin instances
A n E v o lu tio n a ry A lg o rith m
fo r L a rg e S c a le S e t C o v e rin g P ro b le m s
Id
AA03
AA04
AA05
AA06
AA11
AA12
AA13
AA14
AA15
AA16
AA17
AA18
AA19
AA20
BUS1
BUS2
Rows Columns Density Weight Range
106 8661
4.05
91-3619
106 8002
4.05
91-3619
105 7435
4.05
91-3619
105 6951
4.11
91-3619
271 4413
2.53
35-2966
272 4208
2.52
35-2966
265 4025
2.60
35-2966
266 3868
2.50
35-2966
267 3701
2.58
35-2966
265 3558
2.63
35-2966
264 3425
2.61
35-2966
271 3314
2.55
35-2966
263 3202
2.63
35-2966
269 3095
2.58
35-2966
454 2241
1.89
120-877
681 9524
0.51
120-576
Table 6. Characteristics of Balas and Carrera instances
CFT
WSCP
Id Bst Tbst Bst Fbst Apd Tbst Tsol Ibst Isol
AA03 33155 61.0 33155 1.0 0.0 1.26 1.266 40.0 40.0
AA04 34573 3.6 34573 1.0 0.0 1.73 1.73 74.6 74.6
AA05 31623 3.1 31623 1.0 0.0 0.48 0.48 9.6 9.6
AA06 37464 5.2 37464 1.0 0.0 2.67 2.67 128.2 128.2
AA11 35384 193.7 35384 1.0 0.0 19.11 19.11 755.4 755.4
AA12 30809 53.8 30809 1.0 0.0 7.88 7.88 350.8 350.8
AA13 33211 8.3 33211 1.0 0.0 2.32 2.32 103.8 103.8
AA14 33219 30.3 33219 1.0 0.0 11.74 11.74 557.8 557.8
AA15 34409 18.8 34409 1.0 0.0 8.92 8.92 485.6 485.6
AA16 32752 33.6 32752 1.0 0.0 4.63 4.63 257.4 257.4
AA17 31612 10.9 31612 1.0 0.0 4.69 4.69 262.2 262.2
AA18 36782 13.5 36782 0.1 0.01 17.1 6.94 1108.0 433.0
AA19 32317 5.9 32317 1.0 0.0 2.73 2.73 175.4 175.4
AA20 34912 13.6 34912 1.0 0.0 4.76 4.76 318.4 318.4
BUS1 27947 5.0 27947 1.0 0.0 8.19 8.19 382.6 382.6
BUS2 67760 19.2 67760 1.0 0.0 37.24 37.24 616.2 616.2
Table 7. Results of CFT and WSCP on Balas and Carrera instances
3 7 7
3 7 8
E . M a rc h io ri a n d A . S te e n b e e k
On the instances from Wedelin the performance of WSCP is comparable to
the one of the CFT and Wedelin algorithms.
Finally, on the instances from Balas and Carrera, both WSCP and CFT are
always able to nd the optimal solution. In the AA instances WSCP is faster
that CFT, while in the BUS instances CFT nds the optimum in a shorter time.
The results of the experiments indicate that WSCP is a rather powerful tool
for solving large real-life airline crew scheduling problems.
3.2 Experiments on the OR Library SCP Instances
We consider the families A-D from 4, and the NRE-NRH from 5, consisting of
randomly generated SCP instances. Each class contains 5 instances. The values
of the characteristic parameters of these problem classes, like number of rows
and columns, are given in Table 8.
We compare experimentally WSCP with the genetic algorithms by Beasley
and Chu 6, and by Eremeev 11, and with the CFT algorithm by Caprara et
al 7. The results of the experiments are summarized in Tables 9, 10, and 11.
The results for the CFT, Beasley Chu, and Eremeev algorithms are from 11.
In particular, the cpu time is estimated in 100MHz Pentium CPU seconds.
All the algorithms are able to solve the instances of the classes A-D. On these
instances, WSCP seems to have a more robust behaviour that the two genetic
algorithms, nding the optimum in each of the 10 trials. The performance of
WSCP on the other problem instances of classes E-H is rather satisfactory, both
in terms of quality of the solutions as well as running time. On each instance,
WSCP is able to nd the optimum or best known solution, while the two genetic
algorithms BC and Er do not nd the optimum value on instances H1 and H2.
Moreover, WSCP nds the solutions for instances in the harder classes G and H
in a much shorter time than all the other algorithms.
Id
Rows
Columns
Density
Weight Range
A B C D E F G H
300 300 400 400 500 500 1000 1000
3000 3000 4000 4000 5000 5000 10000 10000
2
5
2
5 10 20 2
5
1-100 1-100 1-100 1-100 1-100 1-100 1-100 1-100
Table 8. Characteristics of Classes A, B, C, D
4 Conclusion
In this paper we have introduced a novel heuristic method for solving large
weighted set covering problems. The results of the experiments indicate that
WSCP is able to nd covers of satisfactory quality in short running time.
A n E v o lu tio n a ry A lg o rith m
Id CFT
Tbst
A 47.15
B 3.34
C 29.23
D 7.64
fo r L a rg e S c a le S e t C o v e rin g P ro b le m s
3 7 9
Beasley Chu
Fbst Apd Tsol
0.86 0.20 65.98
1.00 0.00 68.63
0.68 0.41 87.93
0.96 0.06 101.70
Eremeev
WSCP
Fbst Apd Tbst Tsol Fbst Apd Tbst Tsol Ibst Isol
0.44 0.35 82.00 71.8 1.00 0.00 0.98 0.98 108.0 108.0
1.00 0.00 20.80 20.80 1.00 0.00 0.30 0.30 7.8 7.8
0.74 0.26 53.50 52.40 1.00 0.00 0.76 0.76 72.6 72.6
0.94 0.08 26.62 23.33 1.00 0.00 0.40 0.40 26.0 26.0
Table 9. Results for Classes A, B, C, D
Id BK
CFT
Beasley Chu
Eremeev
Bst Tbst Bst Fbst Apd Tsol Bst Fbst Apd Tbst Tsol
E1 29 29 11.5 29 1.0 0.0 16.9 29 1.0 0.0 1.0 1.0
E2 30 30 180.5 30 0.4 2.0 266.9 30 1.0 0.0 94.8 94.8
E3 27 27 41.7 27 0.3 2.6 85.1 27 1.0 0.0 23.1 23.1
E4 28 28 11.6 28 1.0 0.0 238.5 28 1.0 0.0 11.0 11.0
E5 28 28 16.2 28 1.0 0.0 15.5 28 1.0 0.0 2.1 2.1
F1 14 14 14.7 14 1.0 0.0 33.8 14 1.0 0.0 7.9 7.9
F2 15 15 13.8 15 1.0 0.0 34.5 15 1.0 0.0 1.3 1.3
F3 14 14 110.0 14 1.0 0.0 117.9 14 1.0 0.0 55.4 55.4
F4 14 14 13.7 14 1.0 0.0 92.6 14 1.0 0.0 20.4 20.4
F5 13 13 89.0 13 0.3 5.4 67.1 13 0.3 5.4 497.4 151.2
G1 176 176 65.0 176 0.2 1.0 451.3 176 0.7 0.3 115.0 96.0
G2 154 154 346.6 155 0.5 1.5 159.3 154 0.5 0.65 318.3 226.6
G3 166 166 432.7 166 0.1 1.1 312.1 166 0.1 0.8 627.6 319.1
G4 168 168 105.0 168 0.2 1.4 665.4 168 0.4 0.7 160.0 172.5
G5 168 168 105.0 168 0.2 0.8 242.6 168 0.7 0.05 161.2 170.4
H1 63 63 642.1 64 1.0 1.6 743.0 64 1.0 1.6 90.5 90.5
H2 63 63 392.5 64 1.0 1.6 234.3 64 1.0 1.6 34.7 34.7
H3 59 59 690.4 59 0.9 0.2 796.6 59 1.0 0.0 493.2 493.2
H4 58 58 105.1 58 0.4 91.6 62.9 58 1.0 0.0 218.2 218.2
H5 55 55 68.8 55 0.9 0.2 198.6 55 1.0 0.0 25.2 25.2
Table 10. Results of CFT, Beasley and Chu, and Eremeev on Classes E, F, G, H
3 8 0
E . M a rc h io ri a n d A . S te e n b e e k
Id BK
E1 29
E2 30
E3 27
E4 28
E5 28
F1 14
F2 15
F3 14
F4 14
F5 13
G1 176
G2 154
G3 166
G4 168
G5 168
H1 63
H2 63
H3 59
H4 58
H5 55
WSCP
Bst Fbst Apd Tbst Tsol
29 1.0
30 1.0
27 1.0
28 1.0
28 1.0
14 1.0
15 1.0
14 1.0
14 1.0
13 1.0
176 1.0
154 0.5
166 0.2
168 0.4
168 0.9
63 0.2
63 1.0
59 0.3
58 0.8
55 1.0
Table 11. Results
Ibst Isol
0.0 1.8 1.8
2
2
0.0 2.7 2.7 62.9 62.9
0.0 2.3 2.3 48.3 48.3
0.0 2.1 2.1 31.1 31.1
0.0 1.8 1.8 5.0
5.0
0.0 3.6 3.6 19.6 19.6
0.0 3.6 3.6 9.0
9.0
0.0 6.2 6.2 153.2 153.2
0.0 3.6 3.6 13.1 13.1
0.0 34.1 34.1 2061.1 2061.5
0.0 2.2 2.2 29.7 29.7
1.1 8.9 4.0 315 107.9
0.6 30.2 14.9 1433.5 640.3
0.8 18.4 25.0 812.3 1114.7
0.6 5.9 5.7 207.7 197.4
1.1 9.5 11.7 161.5 269.2
0.0 50.4 50.4 1872 1872
1.1 25.2 21.6 778.3 678.9
0.3 28.1 23.9 1016.3 834.1
0.0 5.5 5.5 64.1 64.1
of WSCP on Classes E, F, G, H
In all the experiments we have worked with a core which is a proper subset
of the set of all columns. The size of the core depends on the problem instance.
However, in general a small fraction which varies from 10 per cent to 50 per
cent of the set of columns is used as core. Using small covers helps the e ciency
of the algorithm. Moreover, extensive experiments with di erent core sizes have
revealed a somehow counter intuitive phenomenon: in many instances, the quality of the results become worse by using a larger core, even if the same number
of iterations is used. This seems to indicate that the merit criterion used in
WSCP is not the best possible, because it can make the wrong decision when
all the columns are present in the core. We are actually investigating the use of
alternative merit criteria and their relationship with the selection of the core.
Future work concerns the investigation of how to tune automatically the parameters 0 1 for determining the core problem, and how the value of thesecan
be adaptively change during the execution.
K K
Acknowledgements
We would like to thank Thomas Baeck and Martin Schuetz for interesting discussions on the subject of this paper.
A n E v o lu tio n a ry A lg o rith m
fo r L a rg e S c a le S e t C o v e rin g P ro b le m s
3 8 1
References
1. E. Andersson, E. Housos, Kohl, and D. Wedelin. Crew pairing optimization. In
Operation Research in the Airline Industry. Kluwer Scientic Publishers, 1997.
2. J.P. Arabeyre, J. Fearnley, F.C. Steiger, and W. Teather. The airline crew scheduling problem: A survey. Transportation Science, 3 :140163, 1969.
3. E. Balas and M.C. Carrera. A dynamic subgradient-based branch-and-bound procedure for set covering problem. Operations Research, 44:875890, 1996.
4. J.E. Beasley. An algorithm for set covering problem. European Journal of Operational Research, 31:8593, 1987.
5. J.E. Beasley. A lagrangian heuristic for set covering problems. Naval Research
Logistics, 37:151164, 1990.
6. J.E. Beasley and P.C. Chu. A genetic algorithm for the set covering problem.
European Journal of Operational Research, 94:392404, 1996.
7. A. Caprara, M. Fischetti, and P. Toth. A heuristic method for the set covering
problem. In W.H. Cunningham, T.S. McCormick, and M. Queyranne, editors,
Proc. of the Fifth IPCO Integer Programming and Combinatorial Optimization
Conference. Springer-Verlag, 1996.
8. A. Caprara, M. Fischetti, and P. Toth. Algorithms for the set covering problem.
Technical report, DEIS Operation Research Technical Report, Italy, 03 1998.
9. S. Ceria, P. Nobili, and A. Sassano. A Lagrangian-based heuristic for large-scale
set covering problems. Mathematical Programming, 1995. to appear.
10. H.D. Chu, E. Gelman, and E.L. Johson. Solving large scale crew scheduling problems. European Journal of Operational Research, 97:260268, 1997.
11. A.V. Eremeev. A genetic algorithm with a non-binary represenation for the set
covering problem. In Proc. of OR'98, pages 175181. Springer-Verlag, 1998.
12. M.M. Etschmaier and D.F. Mathaisel. Airline scheduling: An overview. Transportation Science, 19 :127138, 1985.
13. M.L. Fisher. An application oriented guide to Lagrangian relaxation. Interfaces,
15 2 :1021, 1985.
14. M.R. Garey and D.S. Johnson. Computers and Intractability: A Guide to the
Theory of NP-completeness. Freeman, San Francisco, 1979.
15. D. Wedelin. An algorithm for large scale 0-1 integer programming with application
to airline crew scheduling. Annals of Operational Research, 57:283301, 1995.
Design, Implementation, and Application of a
Tool for Optimal Aircraft Positioning
J. Pfalzgraf 1 , K. Frank 1 , J. Weichenberger 1 , S. Stolzenberg 2
1
Department of Computer Science,University of Salzburg jpfalz@cosy.sbg.ac.at,
2
Deutsche Lufthansa AG, FrankfurtMain Siegfried.Stolzenberg@dlh.de
Abstract. Optimal positioning of aircraft at a specic airport is a very
dicult problem involving the modeling of many constraints. Lufthansa
AG formulated this problem eld for the airport FrankfurtMain. In this
contribution we describe the development of a tool for nding solutions to
positioning problems automatically. Our approach consists of two parts.
A generic airport model is developed where the notion of logical berings plays a basic role. The optimization task is treated by application
of modied and extended genetic algorithms. A system has been implemented which is capable of computing concrete positioning plans that
can be used by a human operator for further processing. This leads to a
considerable speed up in the generation of positioning plans for aircraft
in comparison with the former method. The application of the aircraft
positioning tool to a real world scenario airport Frankfurt is brie y
presented.
Keywords: optimal aircraft positioning, logical berings, genetic algorithms,
hybrid problem solving
1
Introduction
This contribution deals with the general problem of optimal positioning of aircraft at an airport. The problem formulation has been provided by Lufthansa
AG for the concrete case of airport FrankfurtMain. The main task in the eld
of aircraft positioning is to nd an optimal schedule for all incoming and outgoing aircraft with respect to their position at corresponding gates of an airport.
One has to take into account many constraints and requests, such as neighborhood relationships, runway crossings, critical passenger connections, aircraft
types, airline requests, special gates, security constraints, and others, depending
on particular situations. The basic task is to fulll all these constraints and requests in an optimal way so that the yield converges to a predened maximum.
Our work consists of two main parts. The development of a generic airport
model" not depending on the choice of a specic airport and thus reusable
and the application of extended and modied genetic algorithms to work on
the optimization problem. A rather complicated tness function cost function
S . C a g n o n i e t a l . ( E d s . ) : E v o W o r k s h o p s 2 0 0 0 , L N C S 1 8 0 3 , p p . 3 8 3 2 , − 2 30 90 0 .
S p rin g e r-V e rla g B e rlin H e id e lb e rg 2 0 0 0
D e s i g n , I m p l e m e n t a t i o n , a n d A p p l i c a t i o nof a Tool for Optimal Aircraft Positioning
3 8 3
has been developed in cooperation with Lufthansa. In establishing the airport
model, the concept of logical berings plays a basic role. This is a logical modeling approach which allows to describe what we call the logical state space
of a virtual airport distribution of logics. Thus, for example, a gate and a
position has its own local" logic, respectively, and they can communicate via
non-classical logical operations called transjunctions. In principle, it would be
possible to apply methods from multi-agent systems MAS. In our case a gate or
an airplane position could be interpreted as an agent, respectively. Application of
MAS techniques is intended future work. In terms of logical berings an agent"
corresponds to a local ber". For the treatment of the optimization problem we
develop and apply modied and extended genetic algorithms. A suitable tness
function has been devised which encodes the basic information and constraints
underlying the positioning task. A prototype of an aircraft positioning tool has
been implemented. A detailed description of the generic airport model, the evolutionary optimization approach, the system design and real world applications
can be found in FW99 .
As far as we know no such approach to treat the problem eld as described
above has been tried before. This has been conrmed by somebody from the
EvoNet project. And no such implemented system existed before, as Lufthansa
states. There is a considerable increase of performance by our system in comparison with the procedure in use so far. Our PC-based system needs about one
hour to calculate a positioning plan which can then be used by a human operator
to process it further. In contrast to that, with the former method, the human
planner needed about one week to establish a similar plan by hand" supported
by a graphical computing system, interactively.
2 On logical berings
The purpose of this section is twofold: to give a very brief introduction to the
elementary notions of logical berings for later use and to bring them to the
attention of the reader. The concept of logical berings originates in an industrial project a case study on so-called polycontextural logics and their possible
applications to complex communication and information systems. Subsequently
we present only the elementary notion of a logical bering and point to practical applications in the area of cooperating agents in particular cooperating
robots scenarios. A detailed introduction to logical berings with background
information and motivating comments can be found in Pfa91 .
The notion of a logical bering is inspired by the mathematical modeling
language of ber bundles. This very expressive and powerful notion integrates
dierent structures, namely geometric, topological, and algebraic structures. A
ber bundle consists of a whole bunch of bers which form a so-called total space
ditributed over a base manifold. Each ber is mapped via a projection map onto
its base point in the base space. Thus, for example, in a vector bundle the typical
ber is a vector space of a given dimension. So we can say that in a ber bundle
vertically one does algebra and horizontally geometry, topology.
3 8 4
J . P fa lz g ra f e t a l.
Concerning a logical bering, now the idea is to use a ber bundle and take
logics as bers, thus, vertically one does logic. In Pfa91 we take as typical ber
a classical two valued logic. Thus a logical bering is an abstract ber space
or bundle with typical ber F = L, a classical rst order logical space or a
sub- space generated by a set of formulas. The base space B of the bering will
often be denoted by B = I , the indexing set, the total space is denoted by E ,
and : E ,! I is the corresponding projection map. For i 2 I we have the ber
,1 i over i, namely ,1 i = fx 2 E j x = ig. It has the structure of a logic
as mentioned above. We note that E is decomposed into the bers ,1 i for all
i 2 I.
The simplest form of a bering is the "trivial bering" having total space
E = I L, base space I and i l = i. Therefore the ber over i 2 I is
L := ,1 i = fig L. Such a trivial bering is a parallel system" of logics L
over the index set I as base space. We can think of reasoning processes running
in parallel within each ber L = ,1 i . A ber L is interpreted as a local
logical system a subsystem of the whole bering . Transition communication
betweeen bers is described with the help of suitable maps cf. Pfa91 . Such a
trivial bering as previously considered will also be called "free parallel system"
with total space the disjoint union of the bers L . Each subsystem L has local
classical truth values = fT F g. The global set of truth values is denoted by
I . In a free parallel system I is just the disjoint union of the i for i 2 I .
Logical connectives can be introduced by taking "berwise" logical operations. For example, one can form logical expressions like the following for a
system with three bers using vector notation : x1 ^ y1 x2 y2 x3 _ :y3 ,
etc.. For more details we refer to Pfa91.
A special nonclassical bivariate operation arises naturally: a local pair xi yi
in Li Li , i 2 I , can be mapped into dierent subsystems L L :::: Taking
truth values as input we can observe that for the four possible input pairs in
i i for a locally dened bivariate operation there can be maximally four
dierent subsystems where that function can be evaluated. We can say that
the values will be distributed over the subsystems. Such an operation is called
transjunction. Below we give an example of a conjunctional" transjunction
we just display the truth table
i
i
i
i
i
i
i
i
i
T0 F0
T0 T F
F0 F F
Every transjunction can be described by such a table or T-F-pattern" together with the indices f g. If f g 6= fig then we obtain a transjunction in the previous example we have the type of a conjunction as can
be seen by omitting the indices. If f g = fig then we have a classical
conjunction remaining in subsystem Li . In work cited below, we introduced generalized transjunctions having more than 4 input pairs. The example which we
present in section 3 deals with such a generalized trunsjunction.
Design, Implementation, and Application of a Tool for Optimal Aircraft Positioning
3 8 5
We applied transjunctions for the logical control of cooperating robots scenarios. Cf.Pfa97 for a brief discussion of such an application. For further information on the subject we refer to PSS96a, PSS96b. In the framework of this
contribution here we apply methods from logical berings to support logically
the modeling of the aircraft positioning system as described subseqently.
Generally spoken, we consider the concept of logical berings as a natural
logical modeling approach for multi-agent systems MAS . This has been discussed in Mei99. Future work is planned, especially with respect to an extended
generic airport model using MAS techniques.
3 Development of an aircraft positioning system
3.1 Design concept of the positioning tool
First of all, we want to give a short problem description and explain what we
mean by aircraft positioning. The main task of aircraft positioning is to nd
schedules for all incoming and outgoing aircraft at an airport. Di cult constraints have to be considered, like runway crossings, optimal passenger connections, aircraft types, special ights, particular gates, airline requests, security
problems, and others. The main problem in the eld of aircraft positioning is
to nd an optimal schedule for all incoming and outgoing aircraft with respect
to their positions and corresponding gates. Three subtasks have to be distinguished: long-time, short-time and the actual day scheduling. To each subtask
corresponds an individual knowledge about the aircraft which inuences the positioning. The knowledge changes rapidly during scheduling. The external state
of the airport can change rapidly too, caused by construction work, for example. Thus an important design objective is to build a tool which is able to react
to rapidly changing situations. As previously mentioned, we decided to choose
logical berings as a logical modeling approach. This decision was naturally motivated by our problem analysis. Figure 1 shows the global system design.
Our system includes various agencies and data areas, namely the kernel
agency, the airport agency and the external data areas, like airport database,
temporary ightplan. Another part of our system is the output unit and the output communication unit, which will be used to visualize an airport utilization.
The communication between the various parts of our model will be handled
by a negotiation protocol. This allows us to handle the basic communication in
our system for the short-time planning task. An augmented nal version of the
tool will be able to treat the two remaining cases too, namely long time planning
and the actual day.
3.2 External data areas
The external data areas are specialized data storages for, e.g. airport description,
ight characteristics, airport characteristics , airline characteristcs. We use an
airport database, which includes basic airport information like the number of
3 8 6
J . P fa lz g ra f e t a l.
Fig. 1.
Global system design
Design, Implementation, and Application of a Tool for Optimal Aircraft Positioning
3 8 7
positions and gates. Besides that there are the infoserver, which is our main data
source for ight information, the airline preferences, the ight specialities and
the gate and position specialities. All these areas contain positioning relevant
data which are necessary to calculate an optimal solution for the positioning
problem.
3.3 Airport Agency
The airport agency is used to model a virtual airport. The virtual airport model
can be applied to an existing airport, in our case airport FrankfurtMain. The
design of the virtual airport is based on the concept of logical berings. The airport agency, respectively the virtual airport, consists of serveral units. Two units
are interfaces for communication with the enviroment and all others are internal
units. First, there is the precalculating ltering hierarchical list of positions and
gates p.f.H.L.P.G and second the output communication unit O.C.U which
will not be discussed in this contribution. These two units manage the data
transfer with the enviroment and prepare input data for the internal use in the
airport agency and the kernel agency which will be decribed later in this contribution. Figure 2 gives an overview of the design of the airport agency with their
units and the internal communication paths.
Fig. 2.
Design of the Airport Agency
The airport agency, shown in Figure 2, includes a number of virtual clusters,
e.g. "AWEST", "AOST" or "BWEST". These clusters include the bers of our
3 8 8
J . P fa lz g ra f e t a l.
airport. We use two dierent types of bers one for positions and the other one
for gates because they have a lot of dierent characteristics. Some typical characteristics of a position ber are the maximum valid wing code or the buertime,
on the other side some typical characteristics of a gate ber are the maximum
number of allowed passengers in the gate area or the time to bring the passengers
to the aircraft.
In terms of logical berings to each agent" corresponds a local" ber. These
bers are connected via communication paths with the p.f.H.L.P.G. and they can
have further connections with their neighbors. The connections communication
with neighbors are modeled by transjunctions. A transjunction can be used to
control the state spaces of neighbored connected bers. A special eect in an
application of a transjunction can be described as follows: if a local ber A
corresponds to, for example, an aircraft or group of passengers GoPax , and
there is a transjunction from A to a ber B, then the state space of ber B will
be downgraded by the transjunction. This downgrade of the state space of ber
B will be cancelled again as soon as ber A is no longer used by attached to"
an aircraft or GoPax.
Example:Transjunction rule 'A10' to 'A12'
if 11 then 5
The eect of this transjunction is, that the state space of 'A12' will be downgraded to SWC 5, if the current SWC of 'A10' is 11.
Here SWC is the short notation for wing code". One has to take into account
that dierent wing codes have to be distinguished there exist priorities which
must be taken into consideration.
The complete transjunction corresponding to this example has 3 truth values
in the local ber A and 12 values in local ber B and therefore represents a
generalized transjunction in the sense of section 2. The complete truth table of
the transjunction is displayed in FW99.
Figure 3 shows a special cluster with its communication paths and neighborhood relationships. Moreover, the complexity of the state space of an agent
depends on the number of allowed aircraft types for that agent.
3.4 Kernel Agency
The kernel agency includes currently 4 dierent algorithms, the adaptive longtime scheduling algorithm, the genetic short-time scheduling algorithm, the random scheduling algorithm, the conict solving algorithm. Another part of the
kernel is the kernel communication unit. In this contribution we discuss the
genetic short-time scheduling algorithm. Treatment of the other scheduling algorithms is planned as future work.
Genetic short-time scheduling algorithm This genetic algorithm Mic96,
Hof96 is designed to solve the problems occurring in short time scheduling. It
uses an already pre-optimized season plan generated by a human operator or by
Design, Implementation, and Application of a Tool for Optimal Aircraft Positioning
Fig. 3.
3 8 9
Internal structure of cluster 'AWest'
Fig. 4.
First attempt
a long time scheduler together with actual changes and additional information
including, among others, actually ying passengers, planes available, passengers
that need to get a connection. We decided to use evolutionary computing to
solve this problem.
The original idea was to treat the positioning problem with a usual genetic
algorithm, encoding the positions of a given aircraft in a standard way gure
4.
Although this approach works it causes a crucial problem: almost no valid
valid for a nal solution individuals are created this way since most solutions
will position one of the planes either on an impossible position plane is too
large, for example or will put planes at the same time on the same position.
Of course, such solutions would get a very low tness, so they are not likely to
produce o spring. Still we get the problem that the resulting algorithm spends
most of the time searching for a solution where the planes actually t and not
for an optimal positioning of the planes thus wasting valuable time.
3 9 0
J . P fa lz g ra f e t a l.
Fig. 5.
Second attempt
Therefore, another approach was chosen. First we changed the coding such
that the aircraft are mapped onto positions gure 5.
This enables us to use the complete rule set which is provided by the airport
agency. So we are now able to allow only such genetic operators which produce
a valid individuum. For example, if the aircraft number one would be a B747
and the only positions which were allowed for such a plane would be position
one, two, and three, then a mutation on plane number one could only mutate it
towards position one, two, or three.
This second model works very well with the mutation operator, but new
problems arise when using a crossover operator. The loss of data or the duplication of data can happen. These problems are well known since they also occur
when trying to nd solutions to a traveling salesman problem with the help of
GAs.
When doing a crossover not only the parts that are actually selected are
exchanged, but also other parts which are necessary to maintain consistency. To
this end we have to check for each exchanged plane whether its counterpart"
is also moved. If not, we have to set this aircraft also on the exchange list. On
rst view this seems to be a very useful trick to solve the problem, but it also
produces a problematic side e ect. Every time we wish to do a crossover at a
certain point we also exchange planes on quite randomly chosen other positions.
This can destroy building blocks, especially in a problem as big as the positioning
task. For this reason we decided not to implement a crossover operator in the
classical way.
The Condense Operator Mutation 1 This is the simplest mutation operator
which processes all planes that are not yet at valid positions. It aims at positioning such planes correctly.
The Replace Operator Mutation 2 This second mutation operator is almost
as simple as the rst one, but works in the opposite way. It is mainly used to
maintain the diversity in a population. To do this it moves a plane from one
position to another one. Technically this is also done in three steps. First one
aircraft is randomly selected. This aircraft is moved to the Temp place. Temp
denotes the set of planes not yet positioned. Then the algorithm selects randomly
Design, Implementation, and Application of a Tool for Optimal Aircraft Positioning
3 9 1
a position where the plane ts. Then the selected plane is force-positioned on
this selected position. It will be positioned there in any case removing disturbing
planes if necessary. Later the removed planes will be reordered using the random
scheduler.
Chromosome Repositioning Mutation 3 This is the most 'advanced' mutation
operator that will be used in our implementation of the GA. It does not operate
on single planes like the previous ones, but on complete positions.
Guided Replace Crossover 1 As already mentioned, many diculties can arise
when using crossover operators in a given problem. Therefore we decided not to
use them in a classical way, but in form of a new class of genetic operators. This
class works like a mixture of crossover and mutation. It behaves like a normal
crossover functionally, but it has a probability of occurrence like a mutation.
So this kind of crossover operators can be considered as guided mutations. The
Guided Replace operator is the simplest case of an operator of this type.
Guided Chromosome Repositioning Crossover 2 This crossover-like operator
has similarity to the operator Mutation 3". The idea is to make a transfer of a
perfect distribution of planes corresponding to a certain position in an individual
to another individual. The exact algorithm becomes a rather complicated ruleset.
A detailed description can be found in FW99 .
Cluster Crossover Crossover 3 This is the most advanced crossover operator.
Since the airport contains groups clusters of positions it is plausible to use
these in the optimization process. The internal algorithm of the operator has
close links with the second crossover operator. In fact, we can use this operator
for the cluster crossover if we select more than one position all positions in the
given cluster.
Fitness Function To nd an appropriate tness function was one of the main
problems of the complete optimization process. We designed the function in close
cooperation with the experts from Lufthansa. The optimization process is inuenced by many factors which have to be taken into account. Furthermore, the
approach should be exible and it should be possible to cope with an optimization task which depends on selected factors only. The complete tness function
consists of several constituents. is based on four dierent main optimization
criteria with respect to an individual :
F
F i
i
= + + +
F i
P i
C i
S i
Q i :
denotes the part that tries to optimize the number of passengers. is
responsible for reducing the connecting time. describes the service aspects
and the quality of the solution.
The Greek letters in the subsequent formulas denote parameters which can
be tuned problem dependent.
P i
C i
S i
Q i
3 9 2
J . P fa lz g ra f e t a l.
P i = IP + OP + IB + OB + " IY + OY
The passenger part takes care that a maximum amount of passengers can
leave or enter a plane directly from the gate IP and OP . Additionally, also
the yield of passengers can be taken into account IY and OY . P i models the
request that a maximum amount of aircraft are well positioned right in front of
the building IB and OB .
C i = CT +
GD:
The connexe connecting passengers part optimizes the time needed to get
from one plane to the other. To this end two options are possible. Either the
buer time maximal possible time to reach the plane - walking time is maximized CT or special gates are used for planes with many connecting passengers
GD.
S i =
BN +
BY +
P D:
S i represents the service part of the tness function. Three factors are
summarized here. BN and BY reduce the amount of bus transfers or the amount
of bus transfers for valuable passengers, respectively.
The third factor, P D, makes the time between boarding and take o as small
as possible.
Q i = P C + GC + CC .
Finally, Q i is responsible for the fact that the solution is admissible at all.
P C and GC model the constraints that all positions and gates are valid. CC is
responsible for the request that all connecting passengers are able to catch their
planes.
4 Test of the tool in a real world scenario
In FW99 a real world example is presented which corresponds to a typical scenario to be handled every day by Lufthansa and FAG the Frankfurt airport
operating center. The initial input to our system is a concrete schedule elaborated by the long-term scheduling team of FAG. The syntax of the input data
is of the following type example we display three lines only.
inbound LH 00201 sta 07:45 in type I in ight type P inpaxcode 3
outbound LH 03720 std 08:40 out type S outpaxcode 3 out ight type P
dest VIE air type A321 swc 11 pos V123 out gate B13
The complete input comprises about 1200 such 3-line units. Evaluating the
performance of the tool in its application to a real problem situation it turned
out that the system found better solutions than the human expert. In order
to produce an optimized solution to the position scheduling task the human
Design, Implementation, and Application of a Tool for Optimal Aircraft Positioning
3 9 3
operator needs about one week, whereas, working on the same task, our system
needs about one hour on a standard modern PC and it produces even better
positioning plans. More details and an example of an airport resource utilization
plan produced by the tool can be found in FW99.
5 Conclusions
In the previous sections we presented work on the hard problem of optimal
positioning of aircraft at an airport. Many real world constraints have to be considered. The original problem has been described by Lufthansa AG focusing on
airport Frankfurt Main. We developed a general generic airport model using,
among others, the concept of logical berings. The optimization problem was
treated by modi ed and extended genetic algortihms. On the basis of these approaches an aircraft positioning tool was developed and implemented, especially
tailored for computing positioning con gurations of the airport Frankfurt Main.
A prototypic rst version of the system is currently being tested with Lufthansa
at Frankfurt airport. Our new system achieves much better performance than
the methods applied before.
In addition to the previously described methods, in future work we intend
to use also methods from arti cial neural networks for modeling position constraints and optimization, make systematic applications of multi-agent systems
techniques and rule based systems.
References
FW99 K. Frank and J. Weichenberger. Design and implementation of an aircraft positioning tool using hybrid problem-solving methods. Master's thesis, Institut
fur Computerwissenschaften, Universitat Salzburg, Austria, 1999.
Hof96 Frank Ho mann. Automatischer Entwurf von Fuzzy-Reglern mit genetischen
Algorithmen. PhD thesis, Mathematisch-Naturwissenschaftliche Fakultat,
Christian-Albrechts Universitat zu Kiel, Germany, 1996.
Mei99 W. Meixl. Logical berings. a general decomposition method for many-valued
logics and a modeling approach for multi-agent systems. Master's thesis,
Institut fur Computerwissenschaften, Universitat Salzburg, Austria, 1999.
Mic96 Zbigniew Michalewicz. Genetic Algorithms + Data Structures = Evolution
Programs. Springer Verlag Berlin Heidelberg, 1996.
Pfa91 J. Pfalzgraf. Logical berings and polycontextural systems. In Fundamentals of Articial Intelligence Research, Ph.Jorrand, J.Kelemen eds. . Lecture
Notes in Computer Science 535, Subseries in AI, Springer Verlag, 1991.
Pfa97 J. Pfalzgraf. On geometric and topological reasoning in robotics. Annals of
Mathematics and Articial Intelligence, 19:279 318, 1997.
PSS96a J. Pfalzgraf, U. Sigmund, and K. Stokkermans. Towards a general approach
for modeling actions and change in cooperating agents scenarios. special issue
of IGPL Journal of the Interest Group in Pure and Applied Logics , IGPL
4 3 445-472, 1996.
r
nd rss n, .
nn tt , . . 6
lch, . . 54
l , . 27
rd
r, . . . 26 7
r
, . . 54
ckls, . . 4
l
l
, .
7 , 255,
i, . .
7
, ..
7 , 255
h n, . 27
hish l, .
7
h ng, .
rn , . 224
rn , . 2 5, 247
tt , . 47
rl, . . 57
l
in, .
rr ir d il , . . 44
indl , . 26 7
g rt, . . 26 7
st r, . .
r nk, .
2
rn l
, . .
r
r¨
, . . .
l
, . 47
5
, . . 7
i, . 27 ,
, 5
d r, . 224
rchiri, . 6 7
c r , .
ng, . 22
rkl, . 2 7
idd nd rf, . 2 7
ils, .
7 , 255
i
shi, . 2 7
s r, . 7 7
rt, . . 7 7
dl c, . 6
ri, . 2 4
rd hl
, . 2, 6 5,
rdin, . 2, 6 5,
l
, ..
5
tt l
ng s , .O . 7
O t s, . 224
Ok
r, . 2 7
2
rt, . 27 7
r , . . 54
iji, . 2 7
rst, .
r s n, ,
h r, . . 7
lss , . 47
s n, . 7
k lkis, .
, . . 6
cht r, .
7
ir , . 57
rkins, . 54
tr , . . 4
f lgr f, .
2
rt r, . . 54
r h, . 4
7 , 255
rl
ss n, . 6 5
n, . . 6
d ng , . 247
init , . 7
ich rds n, .
5
3
r
d
sin, . . 7
ss, . 27 7
h rpls, . 2 5
i ns n, .
5
n
rd , . 2 5, 247
pir , . . 7
p n r, . 57
q il
lr , . 2 5
t n k, . 6 7
t l n rg, .
2
nss n, .
nski, . . 54
n, . . 7
ng, . .
5
k, . 2 7
h ilr, . 54
i kin, . . 7
r , . . 47
rq h rt, .
7
rgh s , . . 54
ilnt , . 247
k
n, . 2 5
ich n rg r, .
2
, . . 22, 7 , 5
ng, . . 22, 7
ng, . . 54
, . 6
ng, .
5