HAL Id: hal-00394945
https://hal.archives-ouvertes.fr/hal-00394945
Preprint submitted on 16 Jun 2009
HAL is a multi-disciplinary open access archive for the deposit and dissemination of sci- entific research documents, whether they are pub- lished or not. The documents may come from teaching and research institutions in France or abroad, or from public or private research centers.
L’archive ouverte pluridisciplinaire HAL, est destinée au dépôt et à la diffusion de documents scientifiques de niveau recherche, publiés ou non, émanant des établissements d’enseignement et de recherche français ou étrangers, des laboratoires publics ou privés.
A method to improve Standard PSO
Maurice Clerc
To cite this version:
Maurice Clerc. A method to improve Standard PSO. 2009. �hal-00394945�
Tehnialreport. DRAFTMC2009-03-13
MaurieCler
Abstrat
Inthisreport,IpresentmypersonalmethodtodesignamoreaurateversionofPSO,assumingweknowwhatkindof
problemswewillhavetosolve. Toillustratethemethod,Ionsideralassofproblemsthataremoderatelymultimodal
andnotofveryhighdimensionality(typiallysmallerthan30). ThestartingversionisStandardPSO2007(SPSO07).
Iuseamodiedveloityupdateequationisused,speitotheonsideredlassofproblems.Then,inordertoimprove
therobustnessofthealgorithm,moregeneralmodiationsare added(inpartiular,betterinitialisations,andthe use
of two kindof partiles). Onthe whole, the resulting algorithm is indeed animprovement of SPSO07, though the
improvementisnotalwaysveryimpressive. Inpassing,Igiveapreisedenitionofbetterthan,andexplainwhythe
lassialmeanbestperformaneriterionmayeasily beompletelymeaningless.
1 Why this paper?
AquestionIoftenfaeisthis: YouarealwaysreferringtoStandardPSO2007(SPSO07,[1℄). Isn'titpossible
to designa better version?. The answeris ofourse yes. Itis even written in thesoure ode available on
line:
ThisPSOversiondoesnotintendtobethebest oneonthemarket
Ontheonehand,itisindeednotverydiulttowriteabetterPSO.Ontheotherhand,asareviewerIreada
lotofpapersinwhihtheauthorslaimtheyhavesueededindoingthat,thoughitisnotreallytrue.,mainly
fortworeasons:
•
theystartfromabadversionofPSO(usuallyaglobalbest one). Althoughtheymayreallyimproveit,theresultingalgorithmisstillnotasgoodasSPSO07.
•
orthey start from areasonablygood version(or even from SPSO 07 itself), but they update it havingapparentlynolearideaofwhatthemodiationisfor. Atypialexampleistheusuallaimwewantto
improvetheexploitation/explorationbalane withoutgivinganyrigorousdenitionofwhatthisbalane
is. Asaresult,themodiedalgorithmmayormaynotbebetter,moreorlessin arandomway,andwe
donoeventknowwhy.
Also,ofourse,thereisthewellknownproblemofomparisonoftwoalgorithms. Often,wendsomethinglike
ournewalgorithmBisthereforebetterthanalgorithmA ... withoutanyleardenitionofbetterthan. Let
usrstsayafewwordsaboutthispoint.
2 My algorithm is better than yours
Whatdoessuhalaimmean? First,suhalaimisvalidonlyifweagreeonaommonbenhmark. Letussay,
forexample,weonsiderasubsetoftheCEC2005benhmark[2℄. Then,wehavetoagreeonagivenriterion.
For examplefor eah funtion, the riterionmay be
Cr
= suess rateoverN
runs,for agiven aurayε
,after at most
F
tnessevaluationsforeahrun, or,in aseofatile, meanofthe best results. Notethat the riterionmaybeaprobabilistione(forexampleCr ′
=theprobabilitythatA
isbetterthanB
onthisproblemaordingto theriterion
Cr
isgreater than0.5). What isimportant is tolearly dene themeaning ofA
is betterthan
B
on one problem. A lot of researhersindeed use suh an approah, and then perform niestatistialanalyses(nullhypothesis,p-test,andsoon),inordertodeideinprobability whetheranalgorithm
AisbetterthananalgorithmB,onthewholebenhmark.
However, in the proess, they miss a simple and frustrating fat: there is no omplete order in
R D
, forall
D > 1
. Why is that important? It maybe usefulto express it moreformally. Letus say thebenhmarkontains
P
problems. We build twoomparison vetors. FirstC A,B = (c A,B,1 , . . . , c A,B,P )
withc A,B,i = 1
ifA
is better thanB
on problemi
(aording to the unique riterion dened),c A,B,i = 0
otherwise. SeondC B,A = (c B,A,1 , . . . , c B,A,P )
withc B,A,i = 1
ifB
is betterthanA
ontheproblemi
,andc B,A,i = 0
otherwise.Wehave toompare thetwonumerialvetors
C A,B
andC B,A
. Now, preisely beause there isno ompleteorderin
R D ,
weansaythatA
isbetterthanB
ifandonlyifforanyi
wehavec A,B,i ≥ c B,A,i
foralli
,andifthereexists
j
sothatc A,B,j > c B,A,j
.This is similar to the denition of the lassialPareto dominane. As we have
P
valuesof oneriterion,theproessofomparing
A
andB
anbeseenasamultiriterion(or multiobjetive)problem. Itimpliesthat most of the time no omparison is possible, exept by using an aggregation method. For example, here, weouldountthenumberof
1
s in eah vetor,andsay that theone withthe largersumwins. But thepointisthat anyaggregationmethod isarbitrary,i.e. foreah methodthere isanotheronethatleadstoadierent
onlusion 1
.
Letusonsider anexample:
•
thebenhmarkontains5unimodalfuntionsf 1
tof 5
,and5multimodalonesf 6
tof 10
•
thealgorithmA
is extremelygoodonunimodalfuntions(veryeasy,sayforagradientmethod)•
thealgorithmB
isquitegood formultimodalfuntions,butnotforunimodalones.You nd
c A,B,i = 1
fori = 1, 2, 3, 4, 5
, and also for6
(just beause, for example,the attration basin of theglobaloptimumisverylarge,omparedtotheonesoftheloal optima),and
c B,A,i = 1
fori = 7, 8, 9, 10
. Yousaythen"
A
isbetterthanB
". Anusertrustsyou,andhoosesAforhisproblems. Andasmostofinteresting realproblemsaremultimodal,hewillbeverydisappointed.So,wehavetobebothmoremodestandmorerigorous. Thatiswhytherststepinourmethodofdesigning
an improvedPSO is thehoie ofasmall benhmark. Butwewill saythat
A
is betterthanB
only ifit istrueforalltheproblemsofthisbenhmark.
3 Step 1: a small representative benhmark
Thisisthemostspeipartofthemethod,foritisdependsonthekindorproblemswedowanttosolvelater
withourimprovedPSO.Letusonsiderthefollowinglassofproblems:
•
moderatelymultimodal,orevenunimodal(butofoursewearenotsupposed toknowitinadvane)•
notoftoohighdimensionality(saynomorethan30)Forthis lass, to whih alot of realproblems belong, I havefoundthat agood small benhmarkmaybethe
followingone(seethetable 1):
•
CEC2005Sphere(unimodal)•
CEC2005Rosenbrok(oneglobaloptimum,atleastoneloaloptimumassoonasthedimensionisgreaterthan3)
•
Tripod(twoloaloptima,oneglobaloptimum,verydeeptive[3℄)Thesethreefuntionsaresupposedtoberepresentativeofourlassofproblem. Ifwehaveanalgorithmthat
is good onthem, then itisveryprobablyalso good foralot ofother problems ofthesamelass. Ouraim is
thentodesignaPSOvariantthatisbetterthanSPSO07forthesethreefuntions. OurhopeisthatthisPSO
variantwillindeed bealsobetterthanSPSO 07onmoreproblemsof thesamekind. Andifitistrueevenfor
somehighlymultimodalproblems,and/orforhigherdimensionality,well,weanonsiderthatasaniebonus!
1
Forexample,itispossibletoassigna"weight"toeahproblem(whihrepresentshow"important"isthiskindofproblemfor
theuser) andtolinearly ombinethe
c A,B,i
andc B,A,i
. Butif,forasetof(nonidential)weights,A
isbetterthanB
,thenitalwaysexistsanotheroneforwhih
B
isbetterthanA
.Tab.1: Thebenhmark. Moredetailsaregivenin 9.1
Searh Required Maximumnumber
spae auray oftnessevaluations
CEC2005Sphere
[ − 100, 100] 30
0.000001 15000CEC2005Rosenbrok
[ − 100, 100] 10
0.01 50000Tripod
[ − 100, 100] 2
0.0001 100004 Step 2: a highly exible PSO
MymaintoolisaPSOversion(Code),whihisbasedonSPSO07. HoweverIhaveaddedalotofoptions,in
ordertohaveaveryexibleresearhalgorithm. Atually,Ioftenmodifyit, butyoualwaysanndthelatest
version(namedBalanedPSO)onmytehnial site[4℄. WhenIused itforthis paper,themain optionswere:
•
two kind of randomness (KISS [5℄, and the standard randomnessprovided in LINUX C ompiler). Inwhat follows,IalwaysuseKISS,sothattheresultsanbemorereproduible
•
seveninitialisationmethodsforthepositions(in partiularavariantoftheHammersley'sone[6℄)•
sixinitialisationmethodsfortheveloities(zero,ompletelyrandom,randomaround aposition,et.)•
twolampingoptionsfortheposition(atually, justlampinglikeinSPSO 07,ornolampingandnoevaluation)
•
possibilitytodeneasearhspaegreaterthanthefeasiblespae. Ofourse,ifapartileiesoutsidethe feasiblespae,itstnessisnotevaluated•
sixloalsearhoptions(noloalsearhasinSPSO07,uniformin thebestloalarea,et.). Notethatitimpliesarigorousdenitionofwhat aloal areais. See[7℄
•
twooptionsfortheloopoverpartiles(sequentialorat random)•
sixstrategiesThestrategiesarerelatedtothelassialveloityupdateformula
v (t + 1) = wv (t) + R (c 1 ) (p (t) − x (t)) + R (c 2 ) (g (t) − x (t))
(1)One anuse dierentoeients
w
, and dierent random distributionsR
. Themost interesting point is thatdierentpartilesmayhavedierentstrategies.In theCsoureode, eah optionhasanidentier toeasily desribethe optionsused. Forexample, PSO
P1V2means: SPSO07, inwhihtheinitialisationofthepositionsisdonebymethod1,andtheinitialisation
oftheveloitiesbythemethod2. Pleaserefertotheonlineodeformoredetails. Inourase,wewillnowsee
nowhowaninterestingPSOvariantan bedesignedbyusing justthreeoptions.
5 Step 3: seleting the right options
Firstofall,wesimulateSPSO07,bysettingtheparametersandoptionstotheorrespondingones. Theresults
over500runs aregivenin thetable 2. Inpassing,itisworthnotingthat theusualpratieoflaunhingonly
25oreven100runs isnotenough,forreally badrunsmayourquiterarely. ThisisobviousforRosenbrok,
asweansee from the table 3. Any onlusionthat is drawn after just 100runs is risky, partiularly if you
onsiderthemeanbest value. Thesuessrateismorestable. Moredetails aboutthispartiularfuntion are
givenin9.5.
Tab.2: StandardPSO2007. Resultsover500runs
Suessrate Meanbest
CEC2005Sphere 84.8%
10 − 6
CEC2005Rosenbrok 15% 12.36
Tripod 46% 0.65
Tab.3:ForRosenbrok,themeanbestvalueishighlydependingonthenumberofruns(50000tnessevaluations
foreahrun). Thesuessrateismorestable
Runs Suess rate Meanbestvalue
100 16% 10.12
500 15% 12.36
1000 14,7% 15579.3
2000 14% 50885.18
5.1 Applying a spei improvement method
Whenweonsider thesurfaes oftheattrationbasins,theresultforTripodisnotsatisfying(thesuessrate
should begreater then50%). Whatoptions/parametersould we modify in order to improvethe algorithm?
Letusallthethree attrationbasinsas
B 1
,B 2
, andB 3
. Theproblemisdeeptivebeausetwoofthem,sayB 2
andB 3
, leadto only loal optima. If, for apositionx
inB 1
(i.e. in the basinof theglobal optimum) theneighbourhood best
g
is eitherinB 2
or inB 3
, then, aordingto theequation 1,evenifthe distanebetweenx
andg
is high, the positionx
may be easily modiedsuh that it isnot inB 1
anymore. This isbeauseinSPSO07theterm
R (c 2 ) (g (t) − x (t))
issimplyU (0, c 2 ) (g (t) − x (t))
,whereU
istheuniformdistribution.However,weareinterestedonfuntionswithasmallnumberofloaloptima,andthereforewemaysuppose
that thedistane betweentwooptimais usuallynotverysmall. So,in ordertoavoidtheabovebehaviour,we
usetheideaisthatthefurtheraninformeris,thesmallerisitsinuene(thisanbeseenasakindofnihing).
Wemaythentrya
R (c 2 )
that is infat aR (c 2, | g − x | )
,and dereasingwith| g − x |
. TheoptionalformulaIusetodothatin myexible PSOis
R (c 2, | g − x | ) = U (0, c 2 )
1 − | g − x | x max − x min
λ
(2)
Experimentssuggestthat
λ
should notbetoohigh, beauseinthat ase,although thealgorithmbeomesalmost perfet for Tripod, the result for Sphere beomes quite bad. In pratie,
λ = 2
seemsto bea goodompromise. With thisvaluetheresultfor Sphereisalsoimprovedasweanseefrom thetable4. Aording
to ournomenlature, this PSO is alled PSO R2. Theresultfor Rosenbrokmaybenowslightlyworse, but
we haveseenthat wedo notneed toworry toomuh aboutthe meanbest,ifthe suessrate seemsorret.
Anyway,wemaynowalsoapplysomegeneral improvementoptions.
Tab.4: ResultswithPSO R2(distanedereasing distribution,aordingto theequation2
Suess rate Meanbest
CEC2005Sphere 98.6%
0.14 × 10 − 6
CEC2005Rosenbrok 13.4% 10.48
Tripod 47.6% 0.225
5.2 Applying some general improvement options (initialisations)
Theaboveoptionwasspeiallyhoseninordertoimprovewhatseemedtobetheworstresult,i.e. theonefor
theTripod funtion. Now,wean triggersomeotheroptionsthat areoftenbeneial,at leastformoderately
multimodalproblems:
•
modiedHammersleymethodfortheinitialisationofthepositionsx
•
One-rand method for the initialisation of the veloity of the partile whose initial position isx
, i.e.v = U (x min , x max ) − x
. Notethatin SPSO07,themethodistheHalf-di one,i.e.v = 0.5 (U (x min , x max ) − U (x min , x max ))
ThismodiedalgorithmisPSOR2P2V1. Theresultsaregiveninthetable5,andarelearlybetterthanthe
onesofSPSO07. Theyarestillnotompletelysatisfying(f. Rosenbrok),though. So,weantryyetanother
option,whihanbealled bi-strategy.
Tab.5: Resultswhenapplyingalsodierentinitialisations,forpositionsandveloities(PSOR2P2V1)
Suess rate Meanbest
CEC2005Sphere 98.2%
0.15 × 10 − 6
CEC2005Rosenbrok 18.6% 31132.29
Tripod 63.8% 0.259
5.3 Bi-strategy
The basiidea is verysimple: we usetwo kindsof partiles. In pratie, during the initialisation phase, we
assignoneoftwopossiblebehaviours,with aprobabilityequalto0.5. Thesetwobehavioursaresimply:
•
theoneofSPSO07. Inpartiular,R (c 2 ) = U (0, c 2 )
•
ortheoneofPSOR2(i.e. byusingequation2)TheresultingalgorithmisPSO R3P2V1. As weanseefrom thetable6,forallthethree funtionsnowwe
obtainresultsthatarealsolearlybetterthantheonesofSPSO07. Suess ratesareslightlyworseforSphere
andRosenbrok,slightlybetterforTripod,sonolearomparisonispossible. Howevermoretests(notdetailed
here) showthat this variant is morerobust, aswe anguess by looking at themean best values, sowe keep
it. Twoquestions,though. Isitstillvalidfordierentmaximumnumberoftnessevaluations(searheort).
Andisittrueformoreproblems,eveniftheyarenotreallyin thesamelass,in partiulariftheyarehighly
multimodal? Bothanswersarearmative,astakledinnextsetions.
Tab.6: Resultsbyaddingthebi-strategyoption(PSOR3P2V1)
Suessrate Meanbest
CEC2005Sphere 96.6%
< 10 − 10
CEC2005Rosenbrok 18.2% 6.08
Tripod 65.4% 0.286
6 Now, let's try
6.1 Suess rate vs Searh eort
Here, on the same three problems, we simply onsider dierent maximum numbers of tness evaluations
(
F E max
),andweevaluatethesuessrateover500runs. Asweanseefrom thegure1,foranyF E max
thesuessrate ofour variantis greaterthanthe oneofSPSO 07. SO,weansafelysaythat it is reallybetter,
at least on this small benhmark. Of ourse, it is not alwaysso obvious. Giving along list of resultsis out
of thesopeof this paper, whih isjust aboutadesign method, but weanneverthelesshavean ideaof the
performaneonafewmoreproblems.
6.2 Moderately multimodal problems
Table7andgure2areaboutmoderatelymultimodalproblems. Thisisasmallseletion,toillustratedierent
ases:
•
learimprovement,i.e. nomatterwhatthenumberoftnessevaluationsis,buttheimprovementissmall (Shwefel,PressureVessel). AtuallySPSO07isalreadyprettygoodontheseproblems(forexample,forPressureVessel,SOMAneedsmorethan50000tnessevaluationstosolveit[8℄),sooursmallmodiations
annotimproveitalot.
•
questionableimprovement,i.e. dependingonthenumberoftnessevaluations(Compressionspring)•
lear bigimprovement(Geartrain). Forthisproblem, andafter 20000tness evaluations, not onlythe suessrateofPSOR3P2V1is92.6%,butitndstheverygoodsolution(19, 16, 43, 49)
(oranequivalentpermutation),85timesover500runs. Thetnessofthissolutionis
2.7 × 10 − 12
(SOMAneedsabout200,000evaluationstondit).
Even when the improvement is not very important, the robustness is inreased. For example, for Pressure
Vessel, with 11000tness evaluations, themean best is 28.23 (standard dev. 133.35)with SPSO 07, asit is
18.78(standarddev. 56.97)withPSOR3P2V1.
Tab. 7: Moremoderatelymultimodalproblems. See9.2fordetails
Searh Required
spae auray
CEC2005Shwefel
[ − 100, 100] 10
0.00001Pressurevessel 4variables 0.00001
(disreteform) objetive7197.72893
Compressionspring 3variables 0.000001
objetive2.625421
(granularity0.001for
x 3
)Geartrain 4variables
10 − 9
6.3 Highly multimodal problems
Table8andgure3areforhighlymultimodalproblems. ThegoodnewsisthatourmodiedPSOisalsobetter
evenforsomehighly multimodalproblems. It isnottrueallthetime(seeGriewankorCellularphone),butit
wasnotitsaim,anyway.
7 Claims and suspiion
We have seen that it is possible to improve Standard PSO 2007 by modifying the veloity update equation
and the initialisationshemes. However,this improvementis not valid arossall kindsof problems, and not
valid aross all riterions (in partiular, it may be depending on the number of tness evaluations). Also,
the improvement is not always very impressive. Thus, this study inites us to be suspiious when reading
an assertion like My PSO variant is far better than Standard PSO. Suh a laim has to be veryarefully
supported,byarigorousdenitionofwhatbettermeans,andbysigniantresultsonagoodrepresentative
benhmark,onalargerangeofmaximumnumberoftnessevaluations. Also,wehavetobeveryarefulwhen
using the meanbest riterionfor omparison, for it maybe meaningless. And, of ourse, theproposed PSO
variantshouldbeompared totheurrent StandardPSO,andnotto anoldbad version.
(a)Sphere
(b)Rosenbrok
()Tripod
Fig.1: SuessprobabilityvsSearheort. Forany
F E max
thevariantisbetter(a)Shwefel (b)Pressurevessel
()Compressionspring (d)Geartrain
Fig.2: OntheShwefeland Pressurevessel problemsPSOR3P2V1isslightlybetterthanSPSO 07for any
numberoftnessevaluations. OntheCompressionspringproblem, itistrueonlywhenthenumberoftness
evaluationsisgreaterthanagivenvalue(about19000). So,onthisproblem, eitherlaimSPSO 07isbetter
orPSOR3P2V1isbetter iswrong
(a)Rastrigin
(b)Griewank
()Akley (d)Cellularphone
Fig.3: Suessprobabilityforsomehighlymultimodalproblems. Althoughdesignedformoderatelymultimodal
problems,PSOR3P2V1isevensometimesgoodfortheseproblems. Butnotalways
Tab.8: Highlymultimodalproblems. See9.3fordetails
Searh Required
spae auray
CEC2005Rastrigin
[ − 5, 5] 10
0.01CEC2005Griewank
[ − 600, 600] 10
0.01(notrotated)
CEC2005Akley
[ − 32, 32] 10
0.0001(notrotated)
Cellularphone
[0, 100] 20 10 − 8
8 Home work
Thespei improvementmodiationof SPSO07usedherewasformoderatelymultimodalproblems,in low
dimension. LetusallthemM-problems. Now,whatouldbeaneetivespeimodiationforanotherlass
of problems? Take, for examplethe lassof highly multimodal problems, but still in low dimension(smaller
than30). Letusall themH-problems.
First,wehavetodeneasmallrepresentativebenhmark. Hint: inludeGriewank10D,fromtheCEC2005
benhmark(no needto usethe rotatedfuntion). Seond, wehaveto understand in whihwaythe diulty
ofanH-problemisdierentfromthatofanM-problem. Hint: onanH-problem,SPSO07isusuallylesseasily
trapped into a loal minimum, just beause theattration basins are small. Onthe ontrary, if apartile is
insidethegood attrationbasin(theoneoftheglobaloptimum),itmayevenleaveitprematurely. Andthird,
we haveto nd what optionsare neededto opewith the foundspei diulty(ies). Hint: just makesure
the urrent attration basin is well exploited, aquik loal searh may be useful. A simple way is to dene
aloal areaaround thebest known position, andto sample itsmiddle (PSOL4) 2
. With just this option,an
improvementseemspossible,asweanseefrom gure4fortheGriewankfuntion. However,itdoesnotwork
verywellforRastrigin.
Allthiswillprobablybethetopiofafuturepaper,butforthemoment,youanthinkatityourself. Good
luk!
2
Let
g = (g 1, g 2, . . . , g D )
be the best known position. On eah dimensioni
, letp i
andp ′ i
are the nearest oordinates ofknownpoints,"ontheleft",and"onthe right"of
g i
. Theloal areaH
istheD
-retangle(hyperparallelepid)artesianprodut⊗ i
g i − α (g i − p i ) , g i + α p ′ i − g i
with,inpratie,
α = 1/3
.Thenitsenterissampled. Usually,itisnotg
.Fig.4: Griewank,omparisonbetweenSPSO07andPSOL4. Forahighlymultimodalproblem,averysimple
loalsearhmayimprovetheperformane.
9 Appendix
9.1 Formulae for the benhmark
Tab.9: Benhmarkdetails
Formula
Sphere
− 450 +
30
X
d=1
(x d − o d ) 2
TherandomosetvetorO = (o 1 , · · · , o 30 )
isdenedbyitsCode.
Thisisthesolutionpoint.
Rosenbrok
390 +
10
X
d=2
100 z d− 2 1 − z d 2
+ (z d − 1 − 1) 2
Therandomosetvetor
O = (o 1 , · · · , o 10 )
with
z d = x d − o d + 1
isdenedbyitsCode.Thisisthesolutionpoint
Thereis alsoaloalminimumon
(o 1 − 2, · · · , o 30 )
. Thetnessvalueisthen394.
Tripod
1 − sign(x 2 )
2 ( | x 1 | + | x 2 + 50 | )
+ 1+sign(x 2 2 ) 1 −sign(x 2 1 ) (1 + | x 1 + 50 | + | x 2 − 50 | ) + 1+sign(x 2 1 ) (2 + | x 1 − 50 | + | x 2 − 50 | )
sign (x) = − 1 x ≤ 0
Thesolutionpointis(0, − 50)
Oset forSphere/Parabola(Csoureode)
statidoubleoset_0[30℄={-3.9311900e+001,5.8899900e+001,-4.6322400e+001,-7.4651500e+001,-1.6799700e+001,
-8.0544100e+001, -1.0593500e+001, 2.4969400e+001, 8.9838400e+001, 9.1119000e+000, -1.0744300e+001, -
2.7855800e+001,-1.2580600e+001,7.5930000e+000,7.4812700e+001,6.8495900e+001,-5.3429300e+001,7.8854400e+001,
-6.8595700e+001, 6.3743200e+001, 3.1347000e+001, -3.7501600e+001, 3.3892900e+001, -8.8804500e+001, -
7.8771900e+001,-6.6494400e+001,4.4197200e+001,1.8383600e+001,2.6521200e+001,8.4472300e+001};
Oset forRosenbrok (Csoureode)
statidoubleoset_2[10℄={8.1023200e+001,-4.8395000e+001,1.9231600e+001,-2.5231000e+000,7.0433800e+001,
4.7177400e+001,-7.8358000e+000,-8.6669300e+001,5.7853200e+001};
9.2 Formulae for the other moderately multimodal problems
9.2.1 Shwefel
Thefuntionto minimiseis
f (x) = − 450 +
10
X
d=1 d
X
k=1
x k − o k
! 2
Thesearhspaeis
[ − 100, 100] 10
. ThesolutionpointistheosetO = (o 1 , . . . , o 10 )
,wheref = − 450
.Oset (Csoureode)
statidoubleoset_4[30℄=
{3.5626700e+001,-8.2912300e+001,-1.0642300e+001,-8.3581500e+001,8.3155200e+001,4.7048000e+001,
-8.9435900e+001,-2.7421900e+001,7.6144800e+001,-3.9059500e+001};
9.2.2 Pressure vessel
Justinshort. Formoredetails,see[9,10,11℄. Therearefourvariables
x 1 ∈ [1.125, 12.5]
granularity0.0625 x 2 ∈ [0.625, 12.5]
granularity0.0625 x 3 ∈ ]0, 240]
x 4 ∈ ]0, 240]
andthree onstraints
g 1 := 0.0193x 3 − x 1 ≤ 0 g 2 := 0; 00954x 3 − x 2 ≤ 0
g 3 := 750 × 1728 − πx 3 2 x 4 + 4 3 x 3
≤ 0
Thefuntion tominimiseis
f = 0.06224x 1 x 3 x 4 + 1.7781x 2 x 2 3 + x 2 1 (3.1611x + 19.84x 3 )
Theanalytialsolutionis
(1.125, 0.625, 58.2901554, 43.6926562)
whihgivesthetnessvalue7,197.72893. To taketheonstraintsinto aount,apenaltymethodisused.9.2.3 Compression spring
Formoredetails,see[9,10,11℄. Therearethreevariables
x 1 ∈ { 1, . . . , 70 }
granularity1 x 2 ∈ [0.6, 3]
x 3 ∈ [0.207, 0.5]
granularity0.001
andveonstraints
g 1 := 8C f F πx max 3 x 2
3 − S ≤ 0
g 2 := l f − l max ≤ 0 g 3 := σ p − σ pm ≤ 0 g 4 := σ p − F K p ≤ 0 g 5 := σ w − F max K −F p ≤ 0
with
C f = 1 + 0.75 x x 3
2 − x 3 + 0.615 x x 3
2
F max = 1000 S = 189000
l f = F max K + 1.05 (x 1 + 2) x 3
l max = 14 σ p = F K p σ pm = 6
F p = 300
K = 11.5 × 10 6 8x x 4 3
1 x 3 2
σ w = 1.25
andthefuntion tominimiseis
f = π 2 x 2 x 2 3 (x 1 + 1) 4
The best known solution is
(7, 1.386599591, 0.292)
whih gives the tness value 2.6254214578. To take the onstraintsinto aount,apenaltymethodisused.9.2.4 Geartrain
Formoredetails,see[9,11℄. Thefuntion tominimiseis
f (x) = 1
6.931 − x 1 x 2
x 3 x 4
2
Thesearhspaeis
{ 12, 13, . . . , 60 } 4
. Thereareseveralsolutions,dependingontherequiredpreision. Forexample
f (19, 16, 43, 49) = 2.7 × 10 − 12
9.3 Formulae for the highly multimodal problems
9.3.1 Rastrigin
Thefuntionto minimiseis
f = − 230 +
10
X
d=1
(x d − o d ) 2 − 10 cos (2π (x d − o d ))
Thesearhspaeis
[ − 5, 5] 10
. ThesolutionpointistheosetO = (o 1 , . . . , o 10 )
,wheref = − 330
.Oset (Csoureode)
statidoubleoset_3[30℄=
{1.9005000e+000,-1.5644000e+000,-9.7880000e-001,-2.2536000e+000,2.4990000e+000,-3.2853000e+000,
9.7590000e-001,-3.6661000e+000,9.8500000e-002,-3.2465000e+000};
9.3.2 Griewank
Thefuntionto minimiseis
f = − 179 + P 10
d=1 (x d − o d ) 2
4000 −
10
Y
d=1
cos
x d − o d
√ d
Thesearhspaeis
[ − 600, 600] 10 .
Thesolutionpointis theosetO = (o 1 , . . . , o 10 )
,wheref = − 180
.Oset (Csoureode)
statidoubleoset_5[30℄=
{-2.7626840e+002,-1.1911000e+001,-5.7878840e+002,-2.8764860e+002,-8.4385800e+001,-2.2867530e+002,
-4.5815160e+002,-2.0221450e+002,-1.0586420e+002,-9.6489800e+001};
9.3.3 Akley
Thefuntionto minimiseis
f = − 120 + e + 20e − 0.2 q
1 D
P 10
d=1 (x d − o d ) 2
− e D 1 P 10
d=1 cos(2π(x d −o d ))
Thesearhspaeis
[ − 32, 32] 10 .
Thesolutionpointis theosetO = (o 1 , . . . , o 10 )
,wheref = − 140
.Oset (Csoureode)
statidoubleoset_6[30℄=
{-1.6823000e+001,1.4976900e+001,6.1690000e+000,9.5566000e+000,1.9541700e+001,-1.7190000e+001,
-1.8824800e+001,8.5110000e-001,-1.5116200e+001,1.0793400e+001};
9.3.4 Cellular phone
Thisproblemarisesinarealappliation,onwhihIhaveworkedintheteleommuniationsdomain. However,
here, all onstraints has been removed, exept of ourse the ones given by the searh spae itself. We have
a square at domain
[0, 100] 2
, in whih we want to putM
stations. Eah stationm k
has two oordinates(m k,1 , m k,2 )
. These arethe2M
variablesoftheproblem. Weonsider eah integer pointofthedomain, i.e.(i, j) , i ∈ { 0, 1, . . . , 100 } , j ∈ { 0, 1, . . . , 100 }
. Oneahintegerpoint,theeldinduedbythestationm k
isgivenby
f i,j,m k )=
1
(i − m k,1 ) 2 + (j − m k,2 ) 2 + 1
andwewanttohaveatleastoneeld thatisnottooweak. Finally,thefuntion tominimiseis
f = 1
P 100 i=1
P 100
j=1 max k (f i,j,m k )
In this paper, we set
M = 10
. Therefore the dimension of the problem is 20. The objetive value is0.005530517. This is not the true minimum, but enough from an engineering point of view. Of ourse, in
reality we do not know the objetivevalue. We just run the algorithm several times for a given number of
tnessevaluations,andkeepthebestsolution. Fromthegure5weanseeasolutionfoundbySPSO07after
20000 tness evaluations. Atually, for this simplied problem, more eient methods do exist (Delaunay's
tessellation, for example), but those an not be used as soon as we introdue a third dimension and more
onstraints,sothattheeldisnotspherialanymore.
Fig. 5: Cellularphone problem. A possible (approximate) solution for 10 stations, found by SPSO 07 after
20000tnessevaluations
9.4 A possible simpliation
Wemaywonderwhetherthetwoinitialisationmethodsused in5.2are reallyusefulornot. Letus tryjust the
bi-strategyoption,by keepingthe initialisations ofSPSO 07. Resultsare in thetable 10. Whenweompare
theresultswiththosegiveninthetable6,weanseethat forthethree funtions,theresultsarenotasgood.
However,theyarenotbadatall. So,forsimpliity,itmaybeperfetly aeptabletousejustPSO R3.
Tab.10: Resultswithjustthebi-strategyoption(PSOR3)
Suessrate Meanbest
CEC2005Sphere %
CEC2005Rosenbrok %
Tripod 60.6% 0.3556
9.5 When the mean best may be meaningless
OntheRosenbrokfuntion,wehavequiklyseenthatthemeanbest dependsheavilyonthenumberof runs
(see table 3), and therefore is not an aeptable performane riterion. Here is a moredetailed explanation
of this phenomenon. First we show experimentally that the distribution of the errors for this funtion is
not Gaussian, and, more preisely, that the probability of a verybad run (i.e. a veryhigh tness value) is
not negligible. Then, and more generally, assuming that for agivenproblem this property is true, asimple
probabilistianalysisexplainswhythesuessrateisamorereliableriterion.
9.5.1 Distributionof the errorsforRosenbrok
We runthe algorithm 5000 times, with 5000 tness evaluations for eah run,i.e. just enough to havea non
zerosuessrate. Eahtime,wesavethebest valuefound. Weanthenestimatetheshapeofthedistribution
of these 5000 values, seen as ourrenes of a random variable. Contrary to what is sometimes said, this
distributionisfar fromnormal(Gaussian)one. Indeed,themain peak isveryaute, andthere aresomevery
highvalues. Evenifthesearerare,itimpliesthatthemeanvalueisnotreallyrepresentativeoftheperformane
of the algorithm. It would be better to onsider the value on whih the highest peak (the mode) lies. For
SPSO 07,itis about7(the rightvalueis0), and themean is25101.4(there areafewverybad runs). As we
ansee from gure6,wehaveaquite niemodel byusing theunion ofapowerlaw(onthe left ofthemain
peak),andaCauhylaw(ontheright).
f requency = α class m k k+1
ifclass ≤ m
= 1 π (class γ
− m) 2 +γ 2
elsewith
γ = 1.294
,m = 7
,andk = 6.5
. Note thataseondpowerlawfortherightpartofthe urve(insteadof theCauhyone) would notbe suitable: although itould bebetterfor lassvaluessmaller thansay15, it
wouldforgettheimportantfatthattheprobabilityofhighvaluesisfarfromzero. Atually,eventheCauhy
modelisoverlyoptimisti,aswean seefromthemagniedversion(lasses40-70)ofthegure6,but atleast
theprobabilityisnotvirtuallyequaltozero,aswiththepowerlawmodel.
ForPSO R3P2V1,the mode isabout6,i. e. just slightlybetter. However, themean isonly3962.1. It
showsthat thisversionis abit morerobust(verybad runsdonotexist). Forbothalgorithms,thesmallpeak
(around10,astherightvalueis4)orrespondsto aloal optimum. Thesmallvalley (around3)is alsodue
totheloal optimum: sometimes(but veryrarely)theswarmisquiklytrappedintoit. It showsthatassoon
as there are loal optima the distribution has neessarily somepeaks, at least for asmall number of tness
evaluations.
9.5.2 Meanbestvs suessrateas riterion
Arun issaidtobesuessful ifthenalvalueissmallerthanasmall
ε
,andbad ifthenalvalueisgreaterthanabig
M
. Foronerun,letp M
betheprobabilityof thatrun beingbad. Then, theprobability, overN
runs,thatatleastoneof theruns isbadis
p M,N = 1 − (1 − p M ) N
Thisprobabilityinreasesquiklywiththenumberofruns. Now,let
f i
bethenalvalueoftheruni
. Theestimateofthemeanbest valueisusuallygivenby
µ N = P N
i=1 f i
N
Let us saythe suessrate is
ς
. It meanswehaveςN
suessful runs. Letus onsider another sequeneof
N
runs,exatly thesame, exept thatk
runs are replaed by bad ones. Letm
be themaximumof theorresponding
f i
in therstsequeneofN
runs. Theprobabilityofthiseventisp M,N,1 = p k M (1 − p M ) N − k
Forthenewsuessrate
ς ′
,wehaveς ≥ ς ′ ≥ ς − k N
Forthenewestimate
µ ′ N
ofthemeanbest, wehaveµ ′ N > µ N + k M − m N
Weimmediatelyseethatthereisaproblemwhenabigvalue
M
ispossiblewithanonnegligibleprobability:when thenumberofruns
N
inreasesthesuess ratemay slightlyderease,but then themeandramatially inreases. Letussupposethat, foragivenproblemandagivenalgorithm,thedistributionoftheerrorsfollowsaCauhylaw. Thenwehave
(a)Globalshape
(b)Zoom"onlasses40to70
Fig.6: Rosenbrok. Distributionofthebest valueover5000runs. Onthezoom,weanseethat theCauhy
model, although optimisti, givesa better idea of thedistribution than the powerlawmodel forlass values
greaterthan40
p M = 0.5 − 1 π arctan
M γ
With theparametersof themodel of thegure6, wehavefor example
p 5000 = 8.3 × 10 − 5
. OverN = 30
runs,theprobabilitytohaveatleastonebadrun(tnessvaluegreaterthan
M = 5000
)islow,just2.5 × 10 − 3
.Letussaywendanestimateofthemeantobe
m.
OverN = 1000
runs,theprobabilityis0.08
,whihisquitehigh. It may easily happen. Insuh aase,eveniffor allthe otherruns the best valueisabout
m
, thenewestimateisabout
(4999m + 5000) /1000
,whihmaybeverydierentfromm
. Inpassing,andifwelookatthetable3,thissimpliedexplanationshowsthat forRosenbrokaCauhylawbasedmodel isindeed optimisti.
In otherwords,ifthe numberof runs istoosmall, you may neverhaveabad one,andtherefore, wrongly
estimate the mean best, even when it exists. Note that in ertain asesthe mean maynot even exist at all
(for example, in ase of a Cauhy law), and therefore any estimate of a mean best is wrong. That is why
it is important to estimate the mean for dierent
N
values(but of ourse with the same number of tnessevaluations). Ifitseemsnotstable, forgetthis riterion,and justonsider thesuessrate, or,asseenabove,
themode. Astherearealotof papersinwhihtheprobableexisteneof themeanisnotheked,itisworth
insistingonit: if thereisnomean,givinganestimate ofit isnottehnially orret. Worse,omparing two
algorithmsbasedonsuhanestimate issimplywrong.
Referenes
[1℄ PSC,PartileSwarmCentral,http://www.partileswarm.info.
[2℄ CEC, Congress on Evolutionary Computation Benhmarks,http://www3.ntu.edu.sg/home/epnsugan/,
2005.
[3℄ L.Gagne,Steadystateevolutionaryalgorithmwithanoperatorfamily, in EISCI,(Kosie,Slovaquie),
pp.373379,2002.
[4℄ M.Cler,MathStuaboutPSO,http://ler.maurie.free.fr/pso/.
[5℄ G.MarsagliaandA. Zaman,Thekissgenerator, teh.rep.,Dept.ofStatistis,U. ofFlorida,1993.
[6℄ T.-T. Wong, W.-S. Luk, and P.-A. Heng, Sampling with Hammersley and Halton points, Journal of
Graphis Tools,vol.2(2), pp.924,1997.
[7℄ M.Cler,Themythialbalane,orwhenPSOdoenotexploit,Teh.Rep.MC2008-10-31,2008.
[8℄ I.Zelinka,SOMA-Self-OrganizingMigratingAlgorithm,inNewOptimizationTehniquesinEngineering,
pp.168217,Heidelberg,Germany: Springer, 2004.
[9℄ E.Sandgren,Nonlinearintegeranddisreteprogramminginmehanialdesignoptimization, 1990.ISSN
0305-2154.
[10℄ M.Cler,PartileSwarmOptimization. ISTE(InternationalSientiandTehnialEnylopedia),2006.
[11℄ G.C.OnwuboluandB.V.Babu,NewOptimizationTehniquesinEngineering.Berlin,Germany: Springer,
2004.