Particle Swarm Opti mi zati on Method in Mul ti obj ecti ve

Probl ems
K. E . Parsopoul os
Department of Mathematics,
University of Patras Artificial
Intelligence Research Center (UPAIRC),
University of Patras,
GR- 26110 Patras, Greece
k ostasp @ mat h. upat ras. g r
M. N . V rahat i s
Department of Mathematics,
University of Patras Artificial
Intelligence Research Center (UPAIRC),
University of Patras,
GR- 26110 Patras, Greece
vrahat i s @ mat h. upat r as. g r
ABS TRACT
This paper constitutes a first study of the Particle Swarm
Optimization (PSO) met hod in Mu|tiobjective Optimi~.a-
tion (MO) problems. The ability of PSO to detect Pareto
Optimal points and capture the shape of the Pareto Front
is studied through experiments on well-known non-trivial
test functions. The Weighted Aggregation technique with
fixed or adaptive weights is considered. Furthermore, criti-
cal aspects of the VEGA approach for Multiobjective Opti-
mization using Genetic Algorithms are adapted to the PSO
framework in order to develop a multi--swarm PSO t hat can
cope effectively with MO problems. Conclusions are derived
and ideas for further research are proposed.
Keywords
Particle Swazm Optimization, Multiobjective Optimization
1. I NTRODUCTI ON
Multiobjective Optimization (MO) problems are very com-
mon, especially in engineering applications, due to the mul-
ticriteria nature of most real-world problems. Design of
complex hardware/software systems [20], atomic structure
determination of proteins [2], potential function parameter
optimization [18], x-ray diffraction pattern recognition [14],
curve fitting [1] and production scheduling [19] are such ap-
plications, where two or more, sometimes competing and/ or
incommensurable objective functions have to be minimized
simultaneously. In contrast to the single-objective optimiza-
tion case, where the optimal solution is clearly defined, in
MO problems there is a whole set of trade-offs giving rise to
numerous Pareto Optimal solutions. These points are opti-
mal solutions for the MO problem when all objectives are
simultaneously considered.
Although t.he traditional Gradient-based optimization te-
Permission to malce digital or hard copies of all or part of this work for
personal or classroom use is granted without fee provided that copies are
not made or dislributed for profit or commercial advantage and that copies
bear this notice and the full citation on the first page. To copy otherwise, to
republish, to post on servers or to redistribute to lists, requires prior specific
permission and/or a fee.
4C 2002, Madrid, Spain
(~) 2002 ACM 1-58113-445-2/02/03 _|5.00.
chniquee can be used to detect Pareto Optimal solutions,
this approach suffers from two critical drawbacks; (I) the ob-
jectives have to be aggregated in one single objective func-
tion and (II) only one solution can be detected per opti-
mization run. The inherent difficulty to foreknow which
aggregation of the objectives is appropriate in addition to
the heavy computational cost of Gradient-based techniques,
necessitates the development of more efficient and rigorous
methods. Evolutionary Algorithms (EA) seem to be eepe-
dally suited to MO problems due to their abilities, to search
simultaneously for multiple Pareto Optimal solutions and,
perform better global search of the search space [20].
The Particle Swarm Optimization (PSO) is a Swarm Intel-
ligence method that models social behavior to guide swarms
of particles towards the most promising regions of the search
space [3]. PSO has proved to be efficient at solving Uncon-
strained Global Optimization and engineering problems [4,
10, 11, 12, 13, 17]. It is easily implemented, using either
binary or floating point encoding, and it usually results in
faster convergence rates than the Genetic Algorithms [7].
Although PSO' s performance, in single-objective optimiza-
tion tasks, has been extensively studied, there are insulB-
cient results for MO problems thus far. In this paper a first
study of the PSO' s performance in MO problems is pre-
sented through experiments on well-known test functions.
In the next section the basic concepts and terminology of
MO are briefly presented. In Section 3 the PSO method is
described and briefly analyzed. In Section 4 the performance
of PSO in terms of finding the Pareto Front in Weighted
Aggregation cases is exhibited, while in Section 5 a modified
PSO is used to perform MO similar to the VEGA system. In
the last section conclusions are derived and farther research
directions are proposed.
2. BASI C CONCEPTS
Let X be an n-dlmensional search space and ]~(z), i =
1, . . . , k, be k objective functions defined over X. Assuming,
g~Cz) ~< 0, i---- 1, . . . , m,
be m inequality co~t rai nt s, the MO problem can be stated
as finding a vector z* = (z~, z~, . . . , z*) E X t hat saris-
flee the constraints and optimizes (without loss of generality
we consider only the minimization case) the vector func-
tion f(z) ~- i l l (z), f 3( z) , . . . , ]k(z)]. The objective func-
tions may be in conflict, thus, in most cases it is impossible
603
t o obt ai n for all obj ect i ves t he gl obal mi ni mum at t he s ame
poi nt . The goal of MO i s t o pr ovi de a set of Pa r e t o Opt i ma l
sol ut i ons t o t he af or ement i oned pr obl em.
Let u = ( ul , . . - , uk) a nd v -- ( vl , . . . , vk) be two vect or s.
Then, ~t dominateJ v if and onl y if ul ~< vl, i = 1, - - . , k, and
< vi for at l east one component . Thi s pr ope r t y i s known
as Pare~o Domi nance and i t is used t o def i ne t he Pa r e t o
Opt i ma l poi nt s. Thus, a s ol ut i on z of t he MO pr obl em
is sai d t o be Pareto Optimal if and onl y i f the~re does not
exi st anot her s ol ut i on y such t ha t f ( y) domi nat es f ( z) . The
set of all Pa r e t o Opt i ma l sol ut i ons of an MO pr obl e m is
cal l ed Pareto Optimal Set a nd we denot e i t wi t h 7 ~*. The
s e t ~ * = { ( / ~ ( - . ) , . . . , I k ( z ) ) l z e r * } is c ~ e d P a r ~ t o
]W'ont. A P a r e t o F r o n t 7 3 =* is c a l l e d c o r t v e z if a n d o n l y if
V u, v E 7 3 ~ r * , V A ~ (0, 1), B u7 e 7~ =" : AI[u][+(1--A)IIvl[ >t
I1~11- Respect i vel y, i t is cal l ed concaue i f and onl y i f V u, v
73. ~ * , V A E ( 0 , 1 ) , ~ w E 7~3 c~ : ~ l l u l l + (1 - A ) l l u l l ~< I1~11.
A P a r e t o F r o n t c a n b e c o n v e x , c o n c a v e o r p a r t i a l l y c o n v e x
a nd/ or concave a nd/ or di scont i nuous. The l ast t hr ee cases
pr es ent t he gr eat es t di ~c ul t y for mos t MO t echni ques.
I n t he next sect i on, t he PSO t echni que is br i ef l y pr esent ed.
3. PARTI CLE S WARM OPTI MI ZATI ON
The PSO is a Swa r m I nt el l i gence me t hod t ha t differs f r om
wel l - known Evol ut i ona r y Comput a t i on al gor i t hms , such as
t he Genet i c Al gor i t hms [5, 6, 7], i n t ha t t he popul at i on is
not ma ni pul a t e d t hr ough oper at or s i ns pi r ed by t he huma n
DNA pr ocedt t r es. I ns t ead, i n PSO, t he popul a t i on dynami cs
s i mul at e t he behavi or of a Kbixds' fl ock" , wher e soci al shar -
i ng of i nf or mat i on t akes pl ace a nd i ndi vi dual s pr of i t f r om t he
di scoveri es a nd pr evi ous exper i ence of al l ot her compani ons
dur i ng t he sear ch for f ood. Thus , each compani on, cal l ed
particle, in t he popul at i on, whi ch is cal l ed stuarm, is as s umed
t o ~fly" over t he sear ch s pace l ooki ng for pr omi s i ng r egi ons
on t he l ands cape. For exampl e, i n t he s i ngl e- obj ect i ve mi n-
i mi zat i on case, such r egi ons possess l ower f unct i on val ues
t he e ot her s pr evi ous l y vi si t ed. I n t hi s cont ext , each par t i cl e
is t r e a t e d as a poi nt i nt o t he sear ch space, whi ch adj us t s i t s
own "flying" accor di ng t o i t s f l yi ng exper i ence as wel l as t he
f l yi ng exper i ence of ot her par t i cl es.
Fi r s t , l et us defi ne t he not a t i on a dopt e d i n t hi s paper :
as s umi ng t ha t t he sear ch space is D- di me ns i ona l , t he i - t h
par t i cl e of t he s war m is r epr es ent ed by t he D- di me ns i ona l
vect or Xi ---- ( z l x , z ~2, . . . ,zCD) a nd t he bes t par t i cl e i n t he
s war m, i.e. t he par t i cl e wi t h t he s mal l es t f unct i on val ue, is
de not e d by t he i ndex g. The bes t pr evi ous pos i t i on (i.e. t he
pos i t i on gi vi ng t he bes t f unct i on val ue) of t he / - t h par t i -
cle is r ecor ded and r epr es ent ed as P~ ---- ( p~t, p~2, . . . ,p~D),
whi l e t he pos i t i on change ( vel oci t y) of t he i - t h par t i cl e is
r epr es ent ed as Vi ~ (v~x, v~2, . . . , V~D). Fol l owi ng t hi s not a-
t i on, t he par t i cl es ar e ma ni pul a t e d accor di ng t o t he follow-
i ng equat i ons
U i d = W J I J i d "4- C X r l ~ i d -- ~ i d ) -4- c 2 r 2 ~ s d -- ~ i d ) , ( 1 )
z ~ = z , d + X v ~ , ( 2 )
wher e d = 1 , 2 , . . . , D; N is t he size of t he popul at i on;
i ---- 1, 2, . . . , N; X is a cons t r i ct i on f act or whi ch cont r ol s a nd
cons t r i ct s t he vel oci t y' s magni t ude; ~ is t he i ner t i a wei ght ;
cl a nd c2 ar e t wo pos i t i ve cons t ant s ; rx a nd r2 ar e t wo r an-
dom number s wi t hi n t he r ange [0, 1].
Equa t i on (1) det er mi nes t he i - t h par t i cl e' s new vel oci t y
as a f unct i on of t hr ee t er ms : t he par t i cl e' s pr evi ous vel oc-
i t y; t he di s t ance bet ween t he bes t pr evi ous pos i t i on of t he
par t i cl e and i t s cur r ent posi t i on, and fi nal l y; t he di s t ance
bet ween t he s wa r m' s bes t exper i ence ( t he pos i t i on of t he
be s t par t i cl e in t he s war m) and t he 4-t h par t i cl e' s cur r ent
posi t i on. Then, accor di ng t o Equa t i on (2), t he i - t h par t i cl e
' qiies " t owar ds a new posi t i on. I n gener al , t he per f or mance
of each par t i cl e is me a s ur e d accor di ng to a f i t ness f unct i on,
whi ch is pr obl e m- de pe nde nt . I n opt i mi zat i on pr obl ems , t he
fi t ness f unct i on is us ual l y t he obj ect i ve f unct i on under con-
si der at i on.
The rol e of t he i ner t i a wei ght w is cons i der ed t o be cr uci al
for t he PSO' s conver gence. The i ner t i a wei ght is empl oyed
t o cont r ol t he i mpa c t of t he pr evi ous hi s t or y of vel oci t i es on
t he cur r ent vel oci t y of each par t i cl e. Thus , t he pa r a me t e r
w r egul at es t he t r a de - of f bet ween gl obal a nd l ocal expl o-
r at i on abi l i t y of t he swar m. A l ar ge i ner t i a wei ght f aci l i t at es
gl obal expl or at i on ( sear chi ng new ar eas) , whi l e a s mal l one
t ends t o f aci l i t at e l ocal expl or at i on, i.e. f i ne - t uni ng t he cur -
r ent sear ch area. A s ui t abl e val ue for t he i ner t i a wei ght w
bal ances t he gl obal mad l ocal expl or at i on abi l i t y and, conse-
quent l y, r educes t he numbe r of i t er at i ons r equi r ed t o l ocat e
t he opt i mum sol ut i on. A gener al r ul e of t humb suggest s t ha t
i t is be t t e r t o i ni t i al l y set t he i ner t i a t o a l ar ge val ue, i n or der
t o make be t t e r gl obal expl or at i on of t he sear ch space, a nd
gr adual l y decr ease i t t o get mor e r ef i ned sol ut i ons. Thus ,
a t i me- decr eas i ng i ner t i a wei ght val ue is used. The i ni t i al
s wa r m can be gener at ed ei t her r a ndoml y or us i ng a Sobol
sequence gener at or [15], whi ch ensur es t ha t t he par t i cl es will
be uni f or ml y di s t r i but e d wi t hi n t he sear ch space.
Fr om t he above di scussi on, i t is obvi ous t ha t PSO r esem-
bl es, t o s ome ext ent , t he " mut at i on" ope r a t or of Genet i c
Al gor i t hm~ t hr ough t he pos i t i on upda t e Equa t i ons (1) mad
(2). Not e, however , t ha t i n PSO t he ~mut at i on" ope r a t or is
gui ded by t he par t i cl e' s own ~flying" exper i ence a nd benef i t s
f r om t he s war m' s "flying" exper i ence. I n ot her wor ds, PSO
is cons i der ed as per f or mi ng mut a t i on wi t h a "conscience",
as poi nt e d out by Ebe r ha r t a nd Shi [4].
I n t he next sect i on, s ome wel l - known be nc hma r k MO
pr obl ems ar e des cr i bed and r esul t s f r om t he appl i cat i on of
PSO usi ng We i ght e d Aggr egat i on appr oaches ar e exhi bi t ed.
4. T HE WE I GHT E D AGGRE GAT I ON AP-
P ROACH
The Weighted Aggregation is t he mos t c ommon a ppr oa c h
for copi ng wi t h MO pr obl ems . Accor di ng t o t hi s appr oach,
al l t he obj ect i ves ar e s umme d t o a wei ght ed c ombi na t i on
F = ~-]~=1 wJ i ( z ) , wher e w~, i = 1, . . . , k, ar e non- ne ga t i ve
wei ght s. I t is us ual l y a s s ume d t ha t ~- - x wi = 1. Thes e
wei ght s can be ei t her f i xed or dyna mi c a l l y a da pt e d dur i ng
t he opt i mi zat i on.
I f t he wei ght s ar e f i xed t he n we ar e i n t he case of t he
Conuentional Weighted Aggregation ( CWA) . Us i ng t hi s ap-
pr oach onl y a si ngl e Par et o Opt i ma l poi nt can be obt a i ne d
per opt i mi zat i on r un a nd a pr i or i knowl edge of t he sear ch
space is r equi r ed i n or der t o choose t he a ppr opr i a t e wei ght s
[9]. Thus , t he sear ch has t o be r e pe a t e d sever al t i mes t o ob-
t ai n a des i r ed numbe r of Paxet o Opt i ma l poi nt s. Yet , t hi s is
not s ui t abl e i n mos t pr obl ems due t o t i me l i mi t at i ons a nd
heavy c omput a t i ona l cost s. Mor eover , CWA is unabl e t o
det ect sol ut i ons i n concave r egi ons of t he Pa r e t o Fr ont [9].
Ot her Wei ght ed Aggr egat i on approar2aes have been pr o-
pos ed t o al l evi at e t he l i mi t at i ons of t he CWA. For a t wo-
604
obj ect i ve MO probl em, t he weights can be modi fi ed duri ng
t he opt i mi zat i on, accordi ng t o t he following equat i ons,
~/)l(t) = sign ( si n( 2~r t / F) ) , Io2 (t) = 1 -- wl (t),
where t is t he i t er at i on' s i ndex and F is t he wei ght s' change
frequency. Thi s is t he wel l - known Bang- Bang Weighted Ap-
grepation ( BWA) approach, accordi ng t o which, t he weights
are changed abr upt l y due t o t he usage of t he sign(-) func-
tion. Al t er nat i vel y, t he weights can be gr adual l y modi fi ed
accordi ng t o t he equat i ons,
w ~ ( t ) -- I s i n ( ~ t / F ) l , ~ 2 ( t ) -- ] - w x ( t ) .
Thi s is cal l ed Dynamic Weighted Aggregation (DWA). The
slow change of t he weights forces t he opt i mi zer t o keep mov-
i ng on t he Par et o Fr ont , i f it is convex, per f or mi ng bet t er
t han in t he BWA case. I f t he Par et o Fr ont is concave t hen
t he per f or mance usi ng DWA and BWA is al most i dent i cal
when Genet i c Al gor i t hms are used [9].
The t hr ee different approaches t hat are ment i oned above
have been appl i ed in exper i ment s using t he PSO t echni que,
wi t h F ---- 100 for t he BWA and F ---- 200 for t he DWA case
respect i vel y. The benchmar k probl ems t hat were used are
d ~ m e d i n [ 8 , 21]:
* Funct i on F1 (convex, uni form Par et o Front ):
Y~ = 1-. E~ = ~ ~-L f ~ = ~ E ~ _ - I ( ~ , - 2) 2.
* Funct i on F2 (convex, non- uni f or m Par et o Front ):
I,
. Funct i on Fs (concave Par et o Front ):
f l = z l , 9 ---- 1 - ] - ~ E ~ ' - - 2 z , , f 2 ----- 9 ( 1 - ( ] 1 / 9 ) 2 ) .
F u n c t i o n F4 ( n e i t h e r p u r e l y c o n v e x n o r p u r e l y c o n c a v e
Par et o F r ont ) : Yl = "-1, g = 1 + ~ E ~ = 2 =' ,
Funct i on Fs ( Par et o Front t hat consists of separ at ed
convex par t s) : f l = z l , g = I + ~- I ~-~-~=2 z, ,
= -
Al t hough si mpl e and consi st i ng of onl y two obj ect i ves,
t hese pr obl ems are consi dered difficult (especi al l y Fs, F4 and
F~) due t o t he shape of t he Par et o Front (purel y or par t i al l y
concave, di scont i nuous et c. ) . In order to have compar abl e
resul t s in fi ndi ng t he Par et o Fr ont of t he benchmar k prob-
lems wi t h t he resul t s pr ovi ded in [9] for t he Evol ut i onar y
Al gor i t hms case, we used t he pseudocode pr ovi ded in [9]
t o bui l d and mai nt ai n t he archi ve of Par et o sol ut i ons and
we per f or med all si mul at i ons usi ng t he same par amet er val-
ues. Thus, we per f or med 150 i t er at i ons of PSO for each
pr obl em, wi t h z G [0, 1] 2. The PSO par amet er s were fixed
c~ ---- c2 = 0.5 and t he i ner t i a w was gr adual l y decreased
from 1 t owar ds 0.4. The size of t he swarm depended on t he
pr obl em but never exceeded 40.
The first exper i ment s were done usi ng t he CWA appr oach
wi t h a smal l swarm of 10 part i cl es. The desi red number of
Par et o Opt i mal poi nt s was 20 for t he funct i ons F1, F2, F.~
and 40 for F4. Thus we had t o r un t he al gor i t hm 20 t i mes
for t he first t hr ee funct i ons and 40 for t he fourt h. The ob-
t ai ned Ps r et o Front s are exhi bi t ed in Fi gures 1 and 2 and
t hey are si mi l ar to t hose obt ai ned from t he Evol ut i onar y
Al gor i t hm in [9] but wi t h very low comput at i onal cost and
fast convergence r at e (less t han 2 mi nut es were needed for
each funct i on).
~ m A m q m ~
m
g
g
O q b P ~ O O O D ~
' Fi gur e 1: CWA a ppr oa c h r e s ul t s f or FI a nd F2.
|
t . . . . . . . . . .
Fi gur e 2: CWA a ppr oa c h r e s ul t s f or F3 a nd F4.
As it was expect ed, t he CWA appr oach was abl e to det ect
t he Par et o Fr ont in F1, where it is convex and uniform,
in F2, where it is convex and non- uni f or m, and in F4 i t
det ect ed onl y t he convex par t s. In Fs (concave case) it was
unabl e t o det ect Par et o Opt i mal poi nt s ot her t han t he two
ends of t he Par et o Front . The obt ai ned Par et o Front s for
t he exper i ment s usi ng t he BWA s a d DWA approaches are
exhi bi t ed in Fi gures 3--6.
~r ~l mm~ Ps~o~
l l . l g , m
l U l ~ I
i .
u i IJ | i i ( l u u
Fi gur e 3: BWA ( l ef t ) a nd DWA ( r i ght ) a ppr oa c he s '
r e s ul t s f or t he f unc t i on FI .
It is clear t hat PSO succeed in capt ur i ng t he shape of
Par et o Fr ont in each case. The resul t s are bet t er usi ng t he
DWA approach, wi t h t he except i on of t he case of t he con-
cave Par et o Fr ont of t he funct i on F3, at which t he BWA
appr oach per f or med bet~.er. Swar m size was equal t o 20 for
all si mul at i ons, except for t he funct i on F~, for which i t was
set to 40.
The MO pr obl ems can be al t er nat i vel y sol ved usi ng popu
l at i on- based non- Par et o appr oaches i nst ead of Wei ght s Ag-
6 0 5
=~
%
Fi gur e 4: BWA (l eft ) a nd DWA ( r i ght ) a ppr oa c he s '
r es ul t s f or t he f unc t i on Fs.
i
Fi gur e 6: BWA (l eft ) a nd DWA, ( r i ght ) a ppr oa c he s '
r es ul t s f or t he f unc t i on Fs.
Fi gur e 5: BWA (l eft ) a nd DWA ( r i ght ) a ppr oa c he s '
r es ul t s f or t he f unc t i on F4.
gregating approaches. One such approach is VEGA (Vector
Eval uat ed Genetic Al gori t hm), developed by Scha~er [16].
In t he next section, a modification of t he PSO al gori t hm
borrowing ideas from VEGA is used in MO problems.
5. A P O P U LA T I O N - B A S E D N O N - P A R E T O
A P P R O A C H
According to t he VEGA approach, fractions of t he next
generation, or subpopul at i ons, are selected from t he old gen-
eration according t o each of t he objectives, separately. Aft er
shuffling all these sub-popul at i ons together, crossover and
mut at i on are applied t o generate t he new popul at i on [16].
The mai n ideas of VEGA were adopt ed and modified t o
fit t he PSO framework, developing t he VEPSO algorithm.
We used two swarms t o solve t he five benchmar k problems
F1-F5. Each swarm was eval uat ed according t o one of t he
objectives but , i nformat i on coming from t he ot her swarm
was used t o det ermi ne the change of t he velocities. Specif-
ically, t he best particle of t he second swarm was used for
t he det ermi nat i on of t he new velocities of t he first swarm' s
particles, using Equat i on (1), and vice versa. Alternatively,
t he best positions of the second swarm can be used, in con-
j unct i on wi t h t he best particle of t he second swarm, for t he
evaluation of t he velocities of t he first swarm, and vice versa.
The obt ai ned Paret o Fronts for all benchmark problems are
exhibited in Figures 7- 11. The left part of each figure is
t he Paret o Front obt ai ned using only t he best particle of t he
ot her swarm, while t he right part is obt ai ned using bot h t he
best particle and t he best previous positions of t he other
swarm. Wi t h t he exception of Funct i on Fa, no significant
difference between t he two approaches was observed. For
each experiment, two swarms of size 20 were used and t he
al gori t hm ran for 200 iterations.
%
t . ,
%l
i
Fi gur e 7: VEPSO r es ul t s f or t he f unc t i on Ft .
6. C O N C LU S I O N S
A first st udy of t he performance of t he PSO technique in
MO problems has been presented. The PSO met hod solved
efficiently well known test problems, including difficult cases
for MO techniques, such as concavi t y and/ or discontinuity
of t he Paret o Front. We used low dimensional objectives
in order t o investigate t he simplest cases first. Besides, it
is a general feeling t hat two objectives are sufficient t o re-
flect essential aspects of MO [21]. Promi si ng results were
obt ai ned even when t he size of t he swarm was very small.
In addition t o t he Wei ght ed Aggregat i on cases, a modified
version of PSG (VEPSO) t hat resembles t he VEGA ideas
was also developed and applied on t he same probl ems, wi t h
promising results.
Furt her research will include investigation of t he perfor-
mance of PSO in hi gher-di mensi onal problems wi t h more
t han two objectives. Especially in t he case of t he Popul at i on-
Based Non- Par et o Approach, a r andom selection of t he ob-
jective, in probl ems wi t h more t han two objectives, seems
very interesting. Theoret i cal work is also requi red t o fully
under st and t he swarm' s dynami cs and behavior duri ng t he
optimization.
7. R EFER ENCES
[1] H. Ahonen, P.A. Desouza, and V.K. Garg. A Genetic
Al gori t hm for Fi t t i ng Lorent zi an Li ne-Shapes in
Mossbauer-Spect ra. Nuclear Instruments ~ Methods
~n Physics Research, 124(4): 633-638, 1997.
[2] T.S. Bush, C.B..A. Catlow, and P.D. Battle.
Evol ut i onary Pr ogr ammi ng Techniques for Predi ct i ng
Inorgani c Crystal-Structuxes. J. Materials Chemistry,
5 ( 8 ) : 1 2 6 9 - 1 2 7 2 , 19 9 5.
[3] 11.C. Eber har t and J. Kennedy. A New Opt i mi zer
Using Particle Swarm Theory. In Proc. 6th Int. Syrup.
6 0 6
"%
*%.
" 4h
1 u u ~ ~ ~ u u u 1
|
'%
%
\
%,
Fi gur e 8: VEPSO r esul t s for t he f unct i on F2.
"1
% %
%
%
1 1 i s u u u u u a l u i
n l q
Fi gur e 10~ VEPSO r esul t s f or t he f unct i on F4.
I I
1111
I I
qi
%.
m , i
I
i
Fi gur e 9: VEPSO r esul t s for t he f unct i on Fa.
.
%
%
i L
F i g u r e 11: VE P S O r e s ul t s f or t h e f u n c t i o n F e .
on Micro Mach. ~ Hum. 8ci., pages 39-43, 1995.
[4] It.C. Eberhart and Y.H. Shi. Evolving Artificial
Neural Networks. In Proc. Int. Conf. on Neural
Networks and Brain, 1998.
[5] lt.C. Ebexhart, P.K. Simpson, and R.W- Dobbins.
Computational Intelligence PC Tools. Academic Press
Professional, Boston, 1996.
[6] J. Kennedy and R.C. Eberhart. Particle Swarm
Optimization. In Proc. of the I EEE Int. Conf. Neural
Networks, pages 1942-1948, 1995.
[7] J. Kennedy and It.C. Eberhart. Stoarm Intelligence.
Morgan Kaufmann Publishers, 2001.
[8] J.D. Knowles and D.W. Corne. Approximating the
Nondominated Front Using the Pareto Archived
Evolution Strategies. Euolutionary Computation,
8(2):149-172, 2000.
[9] Y. Jin, M. Olhofer, and B. SendhoE Dynamic
Weighted Aggregation for Evolutionary
Multi-Objective Optimization: Why Does It Work
and How?. In Prac. GECCO ~001 Con]., 2001.
[10] K.E. Parsopoulos, V.P. Plagiaaakos, G.D. Magoulas,
and M.N. Vra~atis. Objective Function "Stretching"
to Alleviate Convergence to Local Minima. Nonlinear
Analysis, TMA, 47(5):3419-3424, 2000.
[11] K.E. Parsopoulos, V.P. Plagianakos, G.D. Magoulas,
and M.N. Vrallatis. Stretching Technique for
Obtaining Global Minimizers Through Particle Swarm
Optimization. In Proc. Particle Swarm Optimization
Workahop, pages 22-29, 2001.
[12] K. E Parsopoulos and M.N. Vrahatis. Modification of
the Particle Swarm Optimizer for Locating All the
Global Minima. V. Kurkova, N. Steele, It. Neruda,
M. Kaxny (Eds.), Artificial Neural Networks and
Genetic Algorithms, pages 324-327, Springer, 2001.
[13] K.E. Parsopoulos and M.N. Vrahatis. Particle Swarm
Optimizer in Noisy sad Continuously Changing
Environments. M.H. Hamza (Ed.), Artificial
Intelligence and Soft Computing, pages 289-294,
IASTED/ ACTA Press, 2001.
[14] W. Paszkowicz. Application of the Smooth Genetic
Algorithm for Indexing Powder Patterns: Test for the
Orthorhombic System. Materials Science Forum,
228(1&2):19-24, 1996.
[15] W.H. Press, W.T. Vetterling, S.A. Teukolsky, and
B.P. Fl~--ery. Numerical Recipes in Fortran 77,
Cambridge University Press, Cambridge, 1992.
[16] J.D. Schaffer. Multiple Objective Optimization with
Vector Evaluated Genetic Algorithms. In Genetic
Algorithma and their Applicationa: Pron. f irst Int.
Con]. on Genetic Algorithma, pages 93--100, 1985.
[17] Y. Shi, R.C. Eberhart, and Y. Chen. Implementation
of Evolutionary Fuzzy Systems. I EEE T)'ans. Fuzzy
Systems, 7(2):109-119, 1999.
[18] A.J. Skinner and J.Q. Broughton. Neural Networks in
Computational Materials Science: Training
Algorithms. Modelling and Simulation in Material
Science and Engineering, 3(3):371-390, 1995.
[19] K. Swinehart, M. Yasin, and E. Guimaraes. Applying
an Analytical Approach to Shop-Floor Scheduling: A
Case-Study. Int. J. Materiab gJ Product Technology,
11(1-2):98-107, 1996.
[20] E. Zitzler. Evolutionary Algorithms for Multiobjective
Optimization: Methods and Applications. Ph.D.
thesis, Swiss Fed. Inst. Techn. Zurich, 1999.
[21] E. Zitzler, K. Deb, and L. Thiele. Comparison of
Multiobjective Evolution Algorithms: Empirical
Results. Evolutionar~ Computation, 8(2):173-195,
2000.
607