ContentslistsavailableatScienceDirect
Applied
Soft
Computing
jo u r n al hom e p a g e :w w w . e l s e v i e r . c o m / l o c a t e / a s o c
Parameter
selection
in
synchronous
and
asynchronous
deterministic
particle
swarm
optimization
for
ship
hydrodynamics
problems
Andrea
Serani
a,b,
Cecilia
Leotardi
a,
Umberto
Iemma
b,
Emilio
F.
Campana
a,
Giovanni
Fasano
c,
Matteo
Diez
a,∗aCNR-INSEAN,NationalResearchCouncil—MarineTechnologyResearchInstitute,Rome,Italy bDepartmentofMechanicalandIndustrialEngineering,RomaTreUniversity,Rome,Italy cDepartmentofManagement,Ca’FoscariUniversityofVenice,Venice,Italy
a
r
t
i
c
l
e
i
n
f
o
Articlehistory:
Received13November2015 Receivedinrevisedform15July2016 Accepted12August2016
Availableonline20August2016 Keywords:
Simulation-baseddesign Derivative-freeoptimization Globaloptimization Particleswarmoptimization Shiphydrodynamicsoptimization
a
b
s
t
r
a
c
t
Deterministicoptimizationalgorithmsareveryattractivewhentheobjectivefunctionis computation-allyexpensiveandthereforethestatisticalanalysisoftheoptimizationoutcomesbecomestooexpensive. Amongdeterministicmethods,deterministicparticleswarmoptimization(DPSO)hasseveralattractive characteristicssuchasthesimplicityoftheheuristics,theeaseofimplementation,anditsoftenfairly remarkableeffectiveness.TheperformancesofDPSOdependonfourmainsettingparameters:the num-berofswarmparticles,theirinitialization,thesetofcoefficientsdefiningtheswarmbehavior,and(for box-constrainedoptimization)themethodtohandletheboxconstraints.Here,aparametricstudyof DPSOispresented,withapplicationtosimulation-baseddesigninshiphydrodynamics.Theobjectiveis theidentificationofthemostpromisingsetupforbothsynchronousandasynchronousimplementations ofDPSO.Theanalysisisperformedundertheassumptionoflimitedcomputationalresourcesandlarge computationalburdenoftheobjectivefunctionevaluation.Theanalysisisconductedusing100 analyt-icaltestfunctions(withdimensionalityfromtwotofifty)andthreeperformancecriteria,varyingthe swarmsize,initialization,coefficients,andthemethodfortheboxconstraints,resultinginmorethan 40,000optimizations.Themostpromisingsetupisappliedtothehull-formoptimizationofahighspeed catamaran,forresistancereductionincalmwaterandatfixedspeed,usingapotential-flowsolver.
©2016ElsevierB.V.Allrightsreserved.
1. Introduction
Particleswarm optimization(PSO)wasoriginallyintroduced in [1], based on the social-behaviour metaphor of a flock of birds or a swarm of bees searching for food. PSO belongs to theclassofheuristicalgorithmsforsingle-objectiveevolutionary derivative-free global optimization. Derivative-free global opti-mizationapproachesareoftenpreferredtolocalapproacheswhen objectivesarenonconvexand/ornoisy,andwhenmultiplelocal optimacannotbeexcluded,asoftenencounteredin simulation-based design (SBD) optimization. Thecomputational burden of globaloptimizationtechniquesisusuallymuch largercompared tolocalmethods,sothattheaccuracyofthesolutionsoughtoften dependsontheavailablecomputationalresources.
Zhangetal.[2]presentsacomprehensivesurveyonthePSO variantsandtheirapplicationinseveralengineeringfields,suchas
∗ Correspondingauthor.
E-mailaddress:matteo.diez@cnr.it(M.Diez).
mechanicalorchemical. RecentapplicationsofPSO toshipSBD include medium-tohigh-fidelityhull-form andwaterjet design optimization of fast catamarans, by morphing techniques [3,4] andgeometrymodificationsbasedonKarhunen–Loèveexpansion (KLE)[5–7],andlow-tomedium-fidelityoptimizationof uncon-ventionalmulti-hullconfigurations[8].Whenglobaltechniquesare usedindesignoptimization,withCPU-timeexpensivesolvers,the optimizationprocessiscomputationallyexpensiveandits effec-tivenessandefficiencyremainanalgorithmicand technological challenge.Although complex SBD applicationsare oftensolved bymetamodels[9,10],theirdevelopmentandassessmentrequire benchmarksolutions,withsimulationsdirectlyconnectedtothe optimizationalgorithm.Thesesolutionsareachievedonlyif afford-ableandeffectiveoptimizationproceduresareavailable.
The original PSO makes use of random coefficients, aiming atsustaining thevarietyofthe swarmdynamics.Thisproperty impliesthatstatisticallysignificantresultscanbeobtainedonly throughextensivenumericalcampaigns.Suchanapproachcanbe tooexpensiveinSBDoptimizationforindustrialapplications,when CPU-time expensivecomputer simulations are used directlyas http://dx.doi.org/10.1016/j.asoc.2016.08.028
314 A.Seranietal./AppliedSoftComputing49(2016)313–334 analysistools.Furthermore,ifthedesignprobleminhandis
sched-uledwithinanaccurateprojectplanning,timeresourcesmightbe atightrequirementfortheoptimizationprocess.Forthese rea-sonsefficientdeterministicapproaches(suchasdeterministicPSO, DPSO)havebeendeveloped,andtheireffectivenessandefficiency inindustrialapplicationsinship hydrodynamicsproblemshave beenshown,includingcomparisonswithlocalmethods[11]and randomPSO[5].Moreover,theavailabilityofparallelarchitectures andhighperformancecomputing(HPC)systemshasofferedthe opportunitytoextendtheoriginalsynchronousimplementationof PSO(SPSO)toCPU-timeefficientasynchronousmethods(APSO), assessedontestfunctionsin[12],andappliedinseveral engineer-ingproblemssuchasmultidisciplinaryoptimizationofcommercial aircraft[13],biomechanics[14],andswarmrobotics[15].Using dis-tributedcomputing,synchronousimplementationsofPSOimply thatatiteration k+1thepositionandvelocityofanyparticleis updatedafterevaluatingthefunctionatalltheparticlespositions atiterationk.InanasynchronousimplementationofPSOthe posi-tionandvelocityofaparticleispossiblybasedonthefitnessvalue atasubsetofallparticlespositions.
The effectiveness and efficiency of PSO for box constrained optimization are significantly influenced by four main setting parameters:(a)thenumberofswarmparticlesinteractingduring theoptimization,(b)theinitializationoftheparticlesintermsof initiallocationandvelocity,(c)thesetofcoefficientsdefiningthe personalorsocialbehaviouroftheswarmdynamics,and(d)the methodtohandletheboxconstraints.Theseparametersandtheir effectsonPSOhavebeenstudiedbyanumberofauthors[16–18]. Morerecentlytheeffectsoftheparticleinitializationhavebeen studiedin[19,20],theeffectsofthecoefficientshavebeenshown in[21],whereasboundhandlingtechniqueshavebeenpresented in[22].AcomprehensivestudyonthePSOparameterselectionhas beenpresentedin[23]andapreliminaryassessmentofthe per-formancesofDPSO,varying(a),(b)and(c),ispresentedin[24]. Asurveyofapproachesforgeneralconstrainedoptimization prob-lemsinindustrialdesignandmultidisciplinarydesignoptimization maybefoundin[25],includingalsogeneralnonlinearconstraints. However,thediscussionontheapplicationofDPSOinSBD prob-lemsisstilllimited,lackingasystematicandcomparativeanalysis. Theobjectiveofthepresentworkistheidentificationofthemost effectiveandefficientparametersforbothsynchronousand asyn-chronousdeterministicparticleswarmoptimization(SDPSOand ADPSO),fortheiruseinSBDprocedures.Thefocusisonindustrial problems,directlyusingCPU-timeexpensiveanalyses.Thesemake thestatisticalanalysisoftheresultstooexpensiveandtherefore demandfordeterministicalgorithms.Due totheattractive fea-turesofDPSO(suchasthesimplicityoftheheuristics,theease ofimplementation, anditsoftenfairlyremarkableeffectiveness inindustrialproblems),thecurrentstudyislimitedtoDPSOand itsimplementations.AsystematiccomparisonofDPSOwithother deterministicandstochasticmethodsisbeyondthescopesofthe presentwork.
Theapproach includes a preliminaryparametric analysison 100analytical test functions[26–29]characterized bydifferent degreesofnon-linearitiesandnumberoflocalminima,with full-factorialcombinationof:(a)numberofparticles(usingapower oftwo times thenumber ofdesign variables); (b) initialization oftheparticlepositionandvelocity(usingHammersleysequence sampling(HSS)[30]);(c)setofcoefficients,chosenfromliterature [12,16,17,31,32];(d)inelasticandsemi-elasticwall-typeapproach forboxconstraints[22].Inorder tohandletheboxconstraints, wall-typeapproachesarepreferredinsteadofpenaltyapproaches orLagrangianfunctions,whichmightintroduceadditionalbiasin theanalysis.Thenumberofdesignvariablesrangesfromtwotofifty andthesimulationbudget(maximumnumberofobjectivefunction evaluations)isupto4096timesthenumberofdesignvariables.
Thepreliminaryparametricanalysisisconductedontwosubsets ofproblems,respectivelywithlessandmorethantendesign vari-ables.AIntelXeonE5-1620v23.70GHzisusedforthepreliminary tests.Threeabsolutemetricsaredefinedandappliedforthe evalua-tionofthealgorithmperformances,basedonthedistancebetween PSO-found solutions and analytical optima. Accordingwiththe numericaltests,themosteffectiveparameterchoiceamong(a), (b),(c),and(d)isidentified,basedontheassociatedrelative vari-abilityoftheresults.Then,themostpromisingsetupsforSDPSO andADPSOaredeterminedandappliedtoanindustrialproblem, namelyafastcatamaranhull-formoptimization,forcalmwaterand fixedspeed.TheobjectivefunctionistheratioRT/Wbetweenthe
totalresistance(RT)andtheshipweight(W).Thehull-form
modi-ficationisperformedusingaKLE-basedmorphingapproach[5–7], usingrespectivelyfour-andsix-dimensionaldesignspaces. Com-putersimulationsareperformedusingthepotentialflow(PF)solver CNR-INSEANWARP[33],onaclusterofIntelXeonE54622.80GHz. Eachfunctionevaluationtakesabout10minpernode. Addition-ally,theoptimizationresultsarecomparedwiththoseobtainedin earlierresearch,basedonahigh-fidelityURANSsolver[5].
2. PSOformulations
Considerthefollowingobjectivefunction:
f(x):RNdv−→R (1)
andtheglobaloptimizationproblem min
x∈Lf(x), L⊂R
Ndv (2)
whereLisaclosedandboundedsubsetofRNdvandNdvisthe
num-berofdesignvariables.Theglobalminimizationoftheobjective functionf(x)requirestofindavectora∈Lsuchthat:
∀
b∈L:f(a)≤f(b) (3)Then,ais aglobalminimum forthefunctionf(x)over L.Since thesolutionofEq.(2)isingeneralanNP-hardproblem,theexact identificationofaglobalminimummightbeverydifficult. There-fore,solutionswithsufficientgoodfitness,providedbyheuristic procedures,areoftenconsideredacceptableforseveralpractical purposes.Amongdifferentheuristicprocedures,PSOisoftenthe methodofchoiceforitscapabilitytooutreachasuitable approxi-matesolutionwithinafewiterations.InPSO,thepositionsofthe particlesrepresentthecandidate solutionsand willbedenoted byx∈L,withassociatedfitnessf(x).Moreover,inthispaperthe compactsetLrepresentstheboxconstraints.
2.1. Originalformulation
TheoriginalformulationofthePSOalgorithm,aspresentedin [16],reads
vk+1i =wvk i +c1r k 1,i(xi,pb−xki)+c2r k 2,i(xgb−xki) xk+1i =xk i +vk+1i (4)Theaboveequationsupdatevelocityandpositionoftheith parti-cleatthekthiteration,wherewistheinertiaweight;c1andc2are
respectivelythesocialandcognitivelearningrate;rk
1,iandr
k
2,iare
uniformlydistributedrandomnumbersin[0,1];xi,pbisthepersonal
bestpositioneverfoundbytheithparticleintheprevious itera-tionsandxgbistheglobalbestpositioneverfoundintheprevious
Fig.1.BlockdiagramsforparallelSDPSO(a)andADPSO(b).Thegreenboxesrepresentthefirstsetofparticlesevaluatedbythealgorithm[14].(Forinterpretationofthe referencestocolourinthisfigurelegend,thereaderisreferredtothewebversionofthearticle.)
Anoverallconstrictionfactorisusedin[34–38],inplaceofthe inertiaweightw.Accordingly,thesysteminEq.(4)isrecastinthe followingequivalentform
vki+1=vk i +c1r k 1,i(xi,pb−x k i)+c2r k 2,i(xgb−x k i) xki+1=xk i +v k+1 i (5)Inordertoprovidenecessary(butpossiblynotsufficient) condi-tionswhichavoiddivergenceofparticlestrajectories,thefollowing condition:
=
2 2−ϕ−ϕ2−4ϕ, whereϕ=c1+c2, ϕ>4 (6)
is indicated in [35], where setting the value of ϕ to 4.1, with =0.729,c1=c2=2.050issuggested[36].NotethatPSOschemes
includingboththeparameterswandhavebeenalsoproposed intheliterature.Furthermore,recentsurveysonPSOcoefficients, inthelightofpossibledivergenceofparticlestrajectories,canbe foundin[39,40].
2.2. Deterministicformulation
TwoessentialrequisitesinSBDproblemsaresurelyrepresented bytheefficiencyoftheoptimizationprocedureadopted,alongwith thepossibilitytoreplicatethesameprocedureondifferentsettings oftheproblem.In ordertocomplywithboththelatterfacts,a deterministicversion ofthePSO algorithm (namely DPSO)was formulatedin[11] bysettingrk
1,i=r
k
2,i=1inEq.(5).Ofcourse,
skippingrandomparametersdoesnotallowtoassessastatistical analysisoftheresultsonseveralpracticalproblems.Nevertheless, sinceSBDproblemsareusuallycomputationallyveryexpensive, andrepresentjustonestepwithinacomplexandformalproject plan,theuseofDPSOrevealedtobemandatorywithrespectto adoptingPSO.Ontheoverall,thefinalDPSOschemebecomes
vki+1=vk i +c1(xi,pb−xki)+c2(xgb−xki) xki+1=xk i +v k+1 i (7)InthecontextofSBDforshipdesignoptimization,asmentioned inSection1,theformulationofEq.(7)wascomparedtotheoriginal in[5].DPSOisthereforeusedforallthesubsequentanalyses.
Usingtheaboveformulation,itispossibletoprovethatthe nec-essary(butpossiblynotsufficient)conditionswhichensurethatthe trajectoryofeachparticledoesnotdiverge[41],are
0<<1 0<ω<2 (+1) (8) whereω=(c1+c2).Introducing ˇ= ω 2(+1) (9)andassuming>0asusuallyintheliterature,theconditionsofEq. (8)reduceto
0<<1
0<ˇ<1 (10)
2.3. Synchronousandasynchronousimplementations
AsreportedinSection1,twodifferentimplementationsofDPSO are comparatively applied in this paper. In particular, the syn-chronousimplementationofDPSO(SDPSO)updatesthepersonal bests{xi,pb}andtheglobalbestxgb,alongwithparticlesvelocity
andposition,attheendofeachiteration.SDPSOispresentedasa pseudo-codeinAlgorithm1,andasablockdiagraminFig.1a.The synchronousupdateensuresanyparticleofaperfectandcomplete informationofitsneighborhood.However,synchronousupdateis acostlychoice[12],asaparticleneedstowaitforthewholeswarm tobeupdated,beforeitcanmovetoanewpositionandcontinueits search.Hence,thefirstparticleevaluatedmightbeidleforalong time,waitingforthewholeswarmtobeupdated.
Algorithm1. SDPSOpseudo-code
1: InitializeaswarmofNpparticles
2: while(k<Maxnumberofiterations)do
3: fori=1,Npdo
4: Evaluatef(xk
i)
5: endfor
6: Update{xi,pb},xgb
7: Updateparticlepositionsandvelocities{xk+1 i },{vk+1i }
8: endwhile
9: Outputthebestsolutionfound
Furthermore,in parallelarchitectures,if theamountof time required to evaluatethe objective function at each iteration is notuniform(e.g.,duetoiterativeprocess/convergenceof analy-sistools),thewall-clocktimeandCPU-timereservationofSDPSO maysignificantlyincrease.IncontrasttoSDPSO,theasynchronous implementationofPSO(namelyADPSO[12])updatespersonaland globalbests,alongwithparticlesvelocityandposition,assoonas
316 A.Seranietal./AppliedSoftComputing49(2016)313–334
Fig.2. ExamplesofinitializationsinL=[−0.5,0.5]×[−0.5,0.5]with32particles.(Forinterpretationofthereferencestocolourintext,thereaderisreferredtotheweb versionofthearticle.)
theinformationrequiredfortheirupdateisavailable,anda parti-cleisreadyforanewanalysis.Thereforeinthisversionaparticle’s position/velocityisupdatedusingpartialor“imperfect” informa-tionofitsneighborhood.I.e.,someinformationcomefromparticles withposition/velocityupdatedinthecurrentiteration,while par-tialinformationisprovidedbyparticleswhoseposition/velocity wasupdatedinthepreviousiterations.Thislastaspectcontributes todiversityintheswarm,whichisadesirabletrait.ADPSOis pre-sentedasapseudo-codeinAlgorithm2,andasablockdiagramin Fig.1b.
ThedifferencesbetweenADPSOandSDPSOaresummarizedin Table1.
Algorithm2. ADPSOpseudo-code 1: InitializeaswarmofNpparticles
2: while(k<Maxnumberofiterations)do
3: fori=1,Npdo
4: evaluatef(xk
i)
5: Update{xi,pb},xgb
6: Updateparticlepositionsandvelocities{xk+1 i },{vk+1i }
7: endfor
8: endwhile
9: Outputthebestsolutionfound
3. PSOsettingsandevaluationmetrics
ThechoiceofPSOparametersusedinthecurrentanalysisis definedinthefollowingsection.Theirfull-factorialcombinationis considered,resultinginatotalof420PSOsetups.
3.1. Numberofparticles
Thenumberofparticlesused(Np)isdefinedas
Np=2mNdv, withm∈N[1,7] (11)
thereforerangingfromNp=2NdvtoNp=128Ndv.
3.2. Particlesinitialization
In order toavoid possible undesired bias due to thechoice of the swarm initialization, parameters from the literature are investigatedinthecurrentstudy.Inparticular,theinitializationof
Table1
SynchronousandasynchronousDPSOcharacteristics.
SDPSO ADPSO
Update Attheendofeachiteration Assoonastheobjective functionisevaluated Information Thewholeswarmuses
informationfromthesame iteration
Aparticlecoulduse informationfromboth currentandprevious iterations
particles’locationandvelocityfollowsadeterministicand homo-geneous distribution,according withtheHammersley sequence sampling(HSS)[30].Specifically,letq={q1,...,qNdv−1}beavector
ofprimenumberswithqi /= qj,
∀
i /= j.Anypositiveintegericanbeexpressedusingthesequence{qj}by
i=
r
k=0
akqkj (12)
whererisasuitableintegerandakisanintegerin[0,qj−1].Finally,
theithparticlelocationisdefinedas xi=
i Np ,q1(i),...,qNdv−1(i) fori=0,1,2,...,Np−1 (13) whereqj(i)= r k=0ak/qk+1j .Eq.(13)isappliedtothreedifferentsub-domains,definedas: 1.entiredomainL(reddotsinFig.2a)
2.domainbounds(blutrianglesinFig.2b)
3.domainandbounds(reddotsandbluetrianglesinFig.2c). Ontheotherhand,theinitialvelocityisdefinedbyeitherthe fol-lowing:
• nullvelocity:
vi=0,
∀
i∈[1,Np] (14)• non-nullvelocity,basedoninitialparticleposition: vi=
2 Ndv xi−l+u 2(15) wherelandurepresentthelowerandupperboundforx, respec-tively[5].
Combininginitialpositionand velocitysettingsresultsinsix differentinitializations,summarizedinTable2.
3.3. Coefficientset
WithasimilarapproachusedinSection3.2,inthispaperthe authorsarefirstconsideringacomparisonamongdifferentsetups
Table2
Swarminitialization.
HSS,sub-domain v=0 v /=0
Domain A.0 A.1
Bounds B.0 B.1
Table3
Coefficientset.
IDset Name c1 c2 ˇ
1 ShiandEberhart(1998) 0.729 2.050 2.050 0.864 2 CarlisleandDozier(2001) 0.729 2.300 1.800 0.864 3 Trelea(2003) 0.600 1.700 1.700 0.638 4 Clerc(2006) 0.721 1.655 1.655 0.693 5 PeriandTinti(2012) 0.754 2.837 1.597 0.953
ofDPSO,inordertofindapromisingconfiguration,tobeappliedfor thesolutionofanengineeringdesignproblem.Onthisguideline, ononehandinthispapertheauthorsarenotproposingtheirown setofpreferablePSOcoefficients.Ontheotherhand,afair com-parisonwithinapanelofdifferentDPSOcoefficientsetsrequiresto testoriginalsettingsfromtheliterature,withoutintroducingany hybridcombination/integrationamongthem.
Tothelatterpurpose,fivecoefficientsetsaretakenfrom liter-ature,asproposedbyseveralauthors.Thefirstsetistheoriginal byShiandEberhart[16],thesecondwassuggestedbyCarlisleand Dozier[12],thethirdwasproposedbyTrelea[17],thefourthisa furthersuggestionbyClerc[31],whereasthefifthwassuggestedby PeriandTinti[32].Theassociatednumericalvaluesareincludedin Table3,andtheyallsatisfyEq.(10).For possibleprosandcons ofeach coefficientsetchoice,thereadercanrefer totheabove references.
3.4. Boxconstraints
TheoriginalPSOprovidesanupdateofpositionandvelocityof particlesforunconstrainedproblems.Thus,incasepossible con-straintsarepresent,theoriginalformulationmightbeinadequate, evenincasesimpleboxconstraintsareconsidered.Thisimplies that,duringtheevolutionoftheswarm,simplyusingEq.(7)the particlescantraveloutsidethedomainbounds.Thiscanbeacritical issueinSBDproblems,whenthedomainboundscannotbeviolated duetophysical/geometrical/gridsconstraints.Abarrier(wall)ora penaltyapproachcanbeusedontheboundsoftheresearchspace, inordertoconfinetheparticles[25,42–44].Inparticular,theuse ofexactorsequentialpenaltyfunctionsmaybeinadequateinthe currentwork,foratleastacoupleofreasons:
• only box constraintsare used (i.e.simple) or inequality con-straints,whichhavebeenhandledintheliteratureofPSOand otherheuristicsinaspecificway(seeAlgorithms3and4inthis section);
• the use of general penalty approaches described in [25] is expectedtobeappealingwhenhighlynonlinearconstraintsare
involved.Moreover,acorrectimplementationofthelatter meth-odsdefinitelyrequiresafinetuningofthepenaltyparameters involved,includingtheiriterativeupdate.
Thus,accordingwiththeabovemotivations,andinordertoavoid thepossiblebiasintroducedbythepenaltyparameterupdate(see also [45] for details), herein theapproach presented in [22] is applied,sinceitisspecificallydesignedforthecurrentconstrained problems,intheframeworkofPSOliterature.
Specifically,theparticlesareconfinedwithinLusingan inelas-ticwall-typeapproach(IW).Ifaparticleisfoundtoviolateoneof theboundsinthetransitionfromkthto(k+1)thPSOiteration,itis placedontheboundsettingtozerotheassociatedvelocity compo-nent(seeFig.3a).Thisapproachhelpsthealgorithmtoexplorethe domainbounds.TheIWapproachisimplementedinAlgorithm3. Algorithm3. Inelasticwall-typeapproach(IW)
1: fori=1,numberofparticlesdo
2: forj=1,numberofvariablesdo
3: ifxk i,j>ujthen 4: xk i,j=uj,vki,j=0 5: else[ifxk i,j<lj] 6: xk i,j=lj,vki,j=0 7: endif 8: endfor 9: endfor
TheuseofIWhassomelimitations:intheunlikelyeventthatall theparticlestendtoleavethedomainfromthesamehyper-corner, theIWsetsallvelocitiestozeroandthePSOprogressmaystop pre-maturely.Forthisreason,asemi-elasticwall-typeapproach(SEW) isalsousedinthiswork.Specifically,incasetheparticleposition violatesaboundconstraint,thentheparticlepositionismodified inordertomakethatconstraintactive(i.e.theparticleismoved ontheboundaryofthatconstraint),whiletheassociated veloc-itycomponentisdefinedasinthefollowingAlgorithm4(seealso Fig.3b).
Algorithm4. Semi-elasticwall-typeapproach(SEW)
1: fori=1,numberofparticlesdo
2: forj=1,numberofvariablesdo
3: ifxk i,j>ujthen 4: vk i,j=−v k i,j/[(c1+c2)] 5: else[ifxk i,j<lj] 6: vk i,j=−vki,j/[(c1+c2)] 7: endif 8: endfor 9: endfor
Observethatthedampingfactor[(c1+c2)]−1inAlgorithm4is
usedtoconfinetheparticleinthefeasibledomain.
318 A.Seranietal./AppliedSoftComputing49(2016)313–334 3.5. NumberoffunctionevaluationsandPSOiterations
ThemaximumnumberoffunctionevaluationsNfeval(evaluation
budget)inthepresentworkisdefinedas
Nfeval=2nNdv, wheren∈N[7,12] (16)
andthereforerangesfrom128Ndvto4096Ndv.Asaconsequence,
accordingwiththesettinginEq.(11),thenumberofPSOiterations Niterperformedisgivenas
Niter= Nfeval Np = 2nN dv 2mN dv = 2n−m (17) 3.6. Evaluationmetrics
Threeabsoluteperformancecriteriaareusedtoassessthe algo-rithms,whicharedefinedasfollows[24]:
x=
1 Ndv Ndvj=1 x j,min−xj,min Rj 2 , f = fmin−fmin f max−fmin , t= 2 x+2f 2 (18)
wherexrepresentsanormalizedEuclideandistancebetweenthe
minimumpositionfoundbythealgorithm(xmin)andthe
analyt-icalminimumposition(x
min), andRj=|uj−lj|istherangeofthe
jthvariable.fistheassociatednormalizeddistanceintheimage
space,wherefministheminimumfoundbythealgorithm,f
minis
theanalyticalminimum,andf
maxistheanalyticalmaximumofthe
functionf(x)inthedomainL.tisacombinationofxandfand
isusedforanoverallassessment.
Additionally,therelative variability2
r,k [29]for eachmetric
x, f,t (Eq. (18)) is usedto assessthe impactof each
set-tingparameterskonthealgorithms’performance.LetPbeaset
of|P|problems,definingthealgorithm’ssettingparameter vec-torass=[s1,s2,...,sS]T ∈RS,therelativeperformancevariability
associatedtoitskthcomponentis 2 r,k= 2 k
|S| k=12k (19) where k2= 1 |˝|ω∈˝ [ ˆk(ω)] 2 − 1 |˝|
ω∈˝ ˆ k(ω) 2 (20)
with˝containingthepositionsωassumedbytheparametersk,
ˆ k(ω)= 1 |B|
s∈B ¯ (s), B={s:sk=ω} (21) and ¯ (s)=|P|1
p∈P [(s)]p (22) 4. Optimizationproblems 4.1. Analyticaltestfunctions
One hundred analytical test functions are used in the pre-liminary numerical experience, including a wide variety of problems,suchascontinuousanddiscontinuous,differentiableand
non-differentiable,separableandnon-separable,scalableand non-scalable,unimodalandmultimodal,withdimensionalityranging fromtwo tofifty [26–29,46].The numericalexperience is con-ducted setting apart problems withrespectively less and more than 10 design variables, as summarized in Appendix A (see TablesA.10andA.11).
4.2. Hull-formSBDoptimizationofahigh-speedcatamaran Thehigh-speedDelftcatamaran[3]isusedasSBDtest prob-lem.Thisisaninternationalbenchmarkgeometryusedforboth experimentalandnumericalstudies,includingcapabilitiesofCFD predictionsforcomplexflowswithassociatedvalidationandglobal shapeoptimizationforcalmwaterandwaveconditions.Forthe currentproblem,theobjectivefunctionf(x)isdefinedas
f(x)= RT
W (23)
whereRTis thetotal resistanceatFroudenumber(Fr)equalto
0.5incalmwater,andWistheweightforcemodulus.Geometry modificationshave tofit inabox, definedbymaximumoverall length,beam,and draught.Twofeasibledesign spacesare con-sidered.Thefirstincludesoveralldimensionbounds,whereasthe secondincludesoveralldimensionboundsand,inaddition, con-stantlengthbetweenperpendiculars(LBP)[5].Modificationsofthe parenthullareperformedusinghigh-dimensionalfree-form defor-mation(FFD)and95%-confidencedimensionalityreductionbased onKLEasshownin[6].Fourvariablesareusedforthefirstdesign spaceandsixforthesecond,referredtointhefollowingas4Dand 6Dspace,respectively.Newdesignsgareproducedas
g(x)=
⎛
⎝
1− Ndvj=1 xj
⎞
⎠
g0+ Ndvj=1 xjgj (24)
where−1≤xj≤1,
∀
j∈[1,Ndv]arethedesignvariables;Ndv=4forthefirstdesignspace,whereasNdv=6forthesecond;g0isthe
orig-inalgeometryandgjarethegeometriesassociatedtothedesign
spaceprincipal directions (oreigenmodes), asprovided by KLE fordimensionalityreduction.Fordetails,thereaderisreferredto [5,6].
SimulationsareconductedusingtheWARP(WAveResistance Program)code,developedatCNR-INSEAN.Waveresistance com-putationsarebasedonlinearpotentialflowtheoryanddetailsof equations,numericalimplementation,andvalidationofthe numer-icalsolveraregivenin[33].Thefrictionalresistanceisestimated usingaflat-plateapproximation,basedonlocalReynoldsnumber [47].Themodelisfreetosinkandtrim(2degreesoffreedom). Simulationsareperformedfortherightdemi-hull,sincethe prob-lemissymmetricalwithrespecttothexz-plane.Thefreesurface isdiscretizedasfollows:20×1panelsontheinner-upstream sub-domain,20×40ontheouter-upstream,20×1ontheinner-hull, 20×40ontheouter-hull,80×1ontheinner-downstream,80×2 onthetransom-downstream,80×40ontheouter-downstream; thebody is discretizedby125×50 panels (see Fig.4).Domain boundsaredefinedby1LBPupstream,4LBPdownstreamand2LBP sideways.
Sincetheanalyticalsolutionoftheoptimizationproblemisnot available,thealgorithmperformanceisevaluatedconsideringthe objectivereductionanditsconvergence.Thedesignsolutionsare finallycomparedwithearlieroptimizationresultsproducedusing ahigh-fidelityCFDsolver[5].
X Y Z hull body X Y Z outer hull outer upstream outer downstream inner downstream innerhull inner upstream
Fig.4.Panel-gridfortheCNR-INSEANWARPpotential-flowcode.
5. Numericalresults
5.1. TestfunctionsandDPSOparametersguidelineidentification Resultsonthe100analyticaltestfunctionsarepresentedinthe followingsectionsandusedtodefinetheguidelinesforSDPSOand ADPSO,adoptedlaterfortheSBDproblem.
5.1.1. SDPSO
Figs.5and7showtheperformancesofSDPSOversusthe bud-getoffunctionevaluations(Nfeval/Ndv),intermsofx,f,t,for
Ndv<10and≥10respectively.Averagevaluesarepresented,
condi-tionaltonumberofparticles,particleinitialization,coefficientset, andwall-typeapproach,respectively.Figs.6and8showtherelative variance2
rofx,f,tforNdv<10and≥10respectively,retained
byeachoftheaforementionedparameters.Theparticles initial-izationis foundthemostsignificantparametertoaffectSDPSO performance,especiallyforNdv≥10,whereasthecoefficientsset
(selectedherein)andthewall-typeapproachareshowntobethe leastrelevant.Tables4and5showthefivebestperformingsetups foreachx,f,andt,forNdv<10and≥10respectively,varying
thebudgetoffunctionevaluations.Averagevaluesandstandard deviationsforallSDPSOsetupsarealsoprovided.
5.1.2. ADPSO
Generally,ADPSOresultsarefoundsimilartoSDPSO. Specifi-cally,Figs.9and11showtheperformancesofADPSOversusthe budgetoffunctionevaluations,intermsofx,f,t,forNdv<10
and≥10respectively. Averagevaluesarepresented,conditional tonumberofparticles,particleinitialization,coefficientsset,and wall-typeapproachrespectively.Figures10and12showthe rela-tivevariance2
r ofx,f,t forNdv<10and≥10respectively,
retainedby each of the aforementionedparameters. The parti-cleinitializationisagainthemostsignificantparametertoaffect ADPSOperformance,especiallyforlowbudgetsandNdv≥10.The
coefficientsetand thewall-typeapproachareshown tohave a limited effect ontheperformance, compared toother parame-ters.Tables6and7summarizethefivebestperformingsetupsfor eachx,f,andt,for Ndv<10and ≥10,respectively,varying
thebudgetoffunctionevaluations.Overallaveragesandstandard deviationsforallADPSOsetupsarealsoincluded.
5.1.3. Suggestedguidelines
ThemostpromisingsetupsareselectedfromTables4–7,inorder todefineareasonableandrobustguidelinefortheuseofSDPSOand ADPSO.ThesearesummarizedinTable8.
Figs.13and14showtheperformanceofthesuggestedsetups, forSDPSOandADPSOandNdv<10and≥10,respectively.Average
performance, standard deviation, and best performing setup amongallcombinationsisalsoshownforeachbudget.The guide-linesetups(“Guide”)are foundalwaysvery closeorcoincident tothe“Best”.Inaddition,itmaybenotedhowADPSOisalways equivalentorslightlybetterthanSDPSO.
5.2. High-speedcatamaranSBDoptimization
A preliminarysensitivityanalysisfor each design variableis shown in Fig. 15 showing f(%) compared to the parent hull. Changesinfrevealareductionoftheobjectivefunctioncloseto 9%forthe4Ddesignspaceandcloseto10%forthe6D.Forthe cur-rentstudy,alldesignvariablesaredeemedrequiredforthehull formoptimization,eachoneprovidingasignificantvariationofthe designobjective.Accordingly,thesensitivityanalysisisnotapplied fordimensionalityreduction.
TheoptimizationisperformedwithbothSDPSOandADPSO,as pertheguidelinesuggestedinTable8.SDPSOandADPSOiterations areshowninFig.16,revealingaquitefastconvergence.Itmaybe notedthatthefinalconfigurationforthefirstdesignspace(4D)is fairlyclose(exceptforthesecondvariable)tothatobtainedusing metamodelsanda URANSsolverin[5],usedhereforreference (Fig.17a).Fortheseconddesignspace(6D)thedifferencesaremore significant(Fig.17b).Asageneralresult,onecanalsoobservebythe optimizationresults(seeTable9)thatSDPSOandADPSOwithboth IWandSEWleadtoareductionoftheobjectivefunctioncloseto 20%,forthe4Ddesignspace,andgreaterthan20%forthe6Ddesign space.Furthermore,theoptimumconfigurationleadstoa consid-erablereductionofwave’selevationcomparedtotheoriginalshape (Figs.18and19).Therearenotsignificantdifferencesbetweenthe resultsobtainedbySDPSOandADPSO,exceptforSDPSOwithIWfor the6Ddesignspace.InthiscasetheIWapproachinducesthe opti-mizationtostopafter6iterations.AsshowninFig.17,differences inoptimaldesignvariablesaremainlyduetoIWorSEW.
6. Discussion
TheanalysesofSDPSOandADPSOperformancesonthe ana-lyticaltestfunctionshaveshowntheimportanceoftheparticle initializationespeciallyforproblemswithahighnumberofdesign variables. In particular, the initialization on the only domain boundary is alwaystheworst and should beavoided, whereas the use of a non-null initial velocity is always advisable (see Figs.5b,7b,9band11b).
Thenumber ofparticles becomes moderatelyrelevantwhen thebudgetoffunctionevaluationsincreases,speciallyfor prob-lems with a smaller number of variables. More in detail,
320 A.Seranietal./AppliedSoftComputing49(2016)313–334
Fig.5.SDPSOaverageperformanceforNdv<10.
Fig.6. Relativevariance2
Fig.7. SDPSOaverageperformanceforNdv≥10.
Fig.8.Relativevariance2
322 A. Serani et al. / Applied Soft Computing 49 (2016) 313–334 Table4
BestperformingsetupsforSDPSO,Ndv<10.
Nfeval/Ndv Average(STD) BestSDPSO
x f t NNdvp Init Coef Wall x
Np
Ndv Init Coef Wall f
Np
Ndv Init Coef Wall t
128 0.193(0.078) 0.155(0.078) 0.205(0.081)
32 A.1 3 SEW 0.099 4 C.1 4 IW 0.033 4 C.1 4 SEW 0.086
4 C.1 4 SEW 0.102 4 C.0 4 SEW 0.034 4 C.1 4 IW 0.086
4 C.1 4 IW 0.106 2 A.1 4 SEW 0.035 2 A.1 4 SEW 0.091
4 C.1 3 IW 0.106 4 C.1 4 SEW 0.036 4 C.0 4 SEW 0.098
32 A.1 3 IW 0.107 2 A.1 4 IW 0.046 2 A.1 4 IW 0.100
256 0.178(0.076) 0.134(0.079) 0.185(0.081)
32 A.1 3 SEW 0.079 4 C.1 4 SEW 0.031 4 C.1 4 IW 0.082
64 A.1 3 SEW 0.082 4 C.1 4 IW 0.032 4 C.1 4 SEW 0.082
32 A.1 3 IW 0.092 4 C.0 4 SEW 0.032 16 A.1 3 SEW 0.086
64 A.1 3 IW 0.092 2 A.1 4 SEW 0.035 32 A.1 3 SEW 0.089
64 A.0 4 SEW 0.097 16 A.1 3 SEW 0.042 2 A.1 4 SEW 0.091
512 0.167 (0.077) 0.119 (0.080) 0.170 (0.082)
64 A.1 4 IW 0.070 4 C.1 4 SEW 0.031 32 A.1 3 SEW 0.068
32 A.1 3 SEW 0.072 4 C.1 4 IW 0.032 64 A.1 3 IW 0.071
64 A.1 2 IW 0.072 4 C.0 4 SEW 0.032 64 A.1 3 SEW 0.073
64 A.1 3 IW 0.074 2 A.1 4 SEW 0.035 64 A.1 4 IW 0.078
64 A.1 3 SEW 0.076 32 A.1 3 SEW 0.036 4 C.1 4 IW 0.081
1024 0.161 (0.078) 0.109 (0.083) 0.161 (0.085)
64 A.1 4 IW 0.063 64 A.1 3 SEW 0.021 64 A.1 3 SEW 0.052
64 A.1 3 SEW 0.063 32 A.1 3 SEW 0.021 64 A.1 4 IW 0.054
64 A.1 2 IW 0.065 64 A.0 4 IW 0.023 32 A.1 3 SEW 0.056
64 A.1 2 SEW 0.066 64 A.1 3 IW 0.023 64 A.1 3 IW 0.057
32 A.1 3 SEW 0.070 32 A.0 4 SEW 0.023 64 A.1 4 SEW 0.062
2048 0.157(0.080) 0.103(0.085) 0.154(0.088)
64 A.1 2 IW 0.053 64 A.1 3 SEW 0.020 64 A.1 2 IW 0.045
64 A.1 4 IW 0.059 64 A.1 4 SEW 0.020 64 A.1 4 IW 0.049
64 A.1 3 SEW 0.061 64 A.1 4 IW 0.020 64 A.1 3 SEW 0.050
64 A.1 2 SEW 0.063 64 A.0 4 SEW 0.020 64 A.1 2 SEW 0.052
64 A.1 3 IW 0.067 32 A.0 4 SEW 0.021 64 A.1 3 IW 0.055
4096 0.156(0.081) 0.100(0.086) 0.152(0.089)
64 A.1 2 IW 0.049 128 A.1 3 IW 0.020 64 A.1 2 IW 0.042
64 A.1 4 IW 0.058 128 C.1 3 SEW 0.020 64 A.1 4 IW 0.048
64 A.1 2 SEW 0.059 64 A.1 2 IW 0.020 64 A.1 2 SEW 0.049
64 A.1 3 SEW 0.060 128 C.1 4 SEW 0.020 64 A.1 3 SEW 0.050
128 C.1 4 IW 0.064 64 C.0 4 SEW 0.020 128 C.1 4 IW 0.053 Av. 0.169 (0.077) 0.120 (0.079) 0.171 (0.082)
32 A.1 3 SEW 0.076 4 C.1 4 SEW 0.032 4 C.1 4 SEW 0.081
64 A.1 3 SEW 0.078 4 C.1 4 IW 0.032 4 C.1 4 IW 0.081
64 A.1 2 IW 0.078 4 C.0 4 SEW 0.033 32 A.1 3 SEW 0.082
64 A.1 4 IW 0.079 2 A.1 4 SEW 0.035 64 A.1 3 SEW 0.082
A. Serani et al. / Applied Soft Computing 49 (2016) 313–334 323
BestperformingsetupsforSDPSO,Ndv≥10.
Nfeval/Ndv Average(STD) BestSDPSO
x f t NNdvp Init Coef Wall x
Np
Ndv Init Coef Wall f
Np
Ndv Init Coef Wall t
128 0.390 (0.101) 0.274 (0.150) 0.368 (0.128)
4 A.1 4 SEW 0.261 4 A.1 4 SEW 0.118 4 A.1 4 SEW 0.221
2 A.1 4 SEW 0.266 8 A.1 3 SEW 0.121 2 A.1 2 IW 0.224
8 A.1 3 SEW 0.267 4 A.1 3 SEW 0.122 2 A.1 4 SEW 0.225
4 A.1 3 SEW 0.267 2 A.1 4 SEW 0.122 8 A.1 3 SEW 0.227
8 C.1 3 SEW 0.268 4 C.1 4 SEW 0.122 4 A.1 3 SEW 0.227
256 0.379(0.102) 0.256(0.148) 0.354(0.128)
4 A.1 4 SEW 0.257 8 A.1 3 SEW 0.107 4 A.1 4 SEW 0.214
8 A.1 3 SEW 0.260 4 A.1 4 SEW 0.111 8 A.1 3 SEW 0.216
16 A.1 3 SEW 0.260 4 C.1 4 SEW 0.113 8 C.1 4 SEW 0.219
8 C.1 4 SEW 0.261 8 C.1 4 SEW 0.114 8 C.1 3 SEW 0.220
8 C.1 3 SEW 0.261 8 C.1 3 SEW 0.115 2 A.1 2 IW 0.220
512 0.367(0.103) 0.240(0.147) 0.340(0.128)
32 A.1 3 SEW 0.250 16 A.1 3 SEW 0.100 16 A.1 3 SEW 0.208
16 A.1 3 SEW 0.253 8 A.1 3 SEW 0.104 32 A.1 3 SEW 0.212
16 A.1 4 SEW 0.253 8 C.1 4 SEW 0.106 8 C.1 4 SEW 0.212
8 C.1 4 SEW 0.255 16 C.1 3 SEW 0.107 8 A.1 3 SEW 0.213
64 A.1 3 SEW 0.255 32 A.1 3 SEW 0.109 4 A.1 4 SEW 0.213
1024 0.361 (0.104) 0.230 (0.148) 0.332 (0.129)
64 A.1 3 SEW 0.242 32 A.1 3 SEW 0.095 32 A.1 3 SEW 0.200
32 A.1 3 SEW 0.244 16 A.1 3 SEW 0.097 64 A.1 3 SEW 0.204
16 A.1 4 SEW 0.246 32 C.1 3 SEW 0.099 16 A.1 4 SEW 0.205
32 A.1 4 SEW 0.247 16 C.1 4 SEW 0.101 16 A.1 3 SEW 0.205
16 A.1 3 SEW 0.250 64 A.1 3 SEW 0.103 32 A.1 4 SEW 0.207
2048 0.357 (0.105) 0.223 (0.149) 0.326 (0.130)
64 A.1 3 SEW 0.235 32 A.1 3 SEW 0.092 64 A.1 3 SEW 0.194
32 A.1 4 SEW 0.240 64 A.1 3 SEW 0.092 32 A.1 3 SEW 0.198
32 A.1 3 SEW 0.242 32 C.1 3 SEW 0.096 32 A.1 4 SEW 0.199
64 A.1 4 SEW 0.245 16 A.1 3 SEW 0.097 16 A.1 4 SEW 0.203
16 A.1 4 SEW 0.245 64 C.1 3 SEW 0.097 64 A.1 4 SEW 0.204
4096 0.354(0.105) 0.220(0.150) 0.323(0.131)
64 A.1 3 SEW 0.233 64 A.1 3 SEW 0.090 64 A.1 3 SEW 0.191
128 A.1 3 SEW 0.239 32 A.1 3 SEW 0.092 128 A.1 3 SEW 0.196
128 A.1 4 SEW 0.239 128 C.1 3 SEW 0.094 64 A.1 4 SEW 0.197
64 A.1 4 SEW 0.239 64 C.1 4 SEW 0.094 128 A.1 4 SEW 0.197
32 A.1 4 SEW 0.239 64 C.1 3 SEW 0.094 64 A.1 4 IW 0.197
Av. 0.368(0.102) 0.240(0.147) 0.340(0.128)
32 A.1 3 SEW 0.254 8 A.1 3 SEW 0.107 16 A.1 3 SEW 0.214
16 A.1 3 SEW 0.256 16 A.1 3 SEW 0.108 4 A.1 4 SEW 0.215
4 A.1 4 SEW 0.257 4 A.1 4 SEW 0.111 32 A.1 3 SEW 0.216
16 A.1 4 SEW 0.257 8 C.1 4 SEW 0.112 8 A.1 3 SEW 0.216
324 A.Seranietal./AppliedSoftComputing49(2016)313–334
Fig.9. ADPSOaverageperformanceforNdv<10.
Fig.10.Relativevariance2
Fig.11.ADPSOaverageperformanceforNdv≥10.
Fig.12.Relativevariance2
326 A. Serani et al. / Applied Soft Computing 49 (2016) 313–334 Table6
BestperformingsetupsforADPSO,Ndv<10.
Nfeval/Ndv Average(STD) BestADPSO
x f t NNdvp Init Coef Wall x
Np
Ndv Init Coef Wall f
Np
Ndv Init Coef Wall t
128 0.191(0.078) 0.155(0.076) 0.205(0.080)
4 C.1 4 IW 0.095 4 C.1 3 IW 0.023 4 C.1 4 IW 0.075
4 C.1 4 SEW 0.097 4 C.1 4 IW 0.024 4 C.1 4 SEW 0.084
8 A.1 3 IW 0.100 4 C.0 3 IW 0.034 4 C.1 3 IW 0.087
8 A.1 3 SEW 0.107 4 C.0 4 SEW 0.039 4 C.0 3 IW 0.094
2 A.1 4 SEW 0.110 4 C.0 4 IW 0.039 4 C.0 4 SEW 0.096
256 0.176(0.075) 0.134(0.076) 0.185(0.079)
4 C.1 4 IW 0.085 4 C.1 4 IW 0.021 4 C.1 4 IW 0.068
32 A.1 3 SEW 0.092 4 C.1 3 IW 0.023 4 C.1 4 SEW 0.080
4 C.1 4 SEW 0.093 4 C.0 4 IW 0.031 4 C.1 3 IW 0.085 8 A.1 3 IW 0.096 4 C.0 3 IW 0.033 4 C.0 4 SEW 0.087 32 A.1 4 IW 0.097 4 C.0 4 SEW 0.034 4 C.0 3 IW 0.090 512 0.163 (0.073) 0.118 (0.077) 0.168 (0.078) 32 A.1 4 IW 0.073 4 C.1 4 IW 0.020 4 C.1 4 IW 0.065
64 A.1 3 SEW 0.080 4 C.1 3 IW 0.023 16 A.0 4 IW 0.074
4 C.1 4 IW 0.082 16 A.0 4 IW 0.024 4 C.1 4 SEW 0.078
32 C.1 4 IW 0.084 4 C.0 4 IW 0.031 32 A.1 3 SEW 0.079
32 A.1 3 SEW 0.084 4 C.0 3 IW 0.033 64 A.1 3 SEW 0.082
1024 0.156 (0.075) 0.107 (0.079) 0.157 (0.081) 32 A.1 4 IW 0.064 4 C.1 4 IW 0.020 32 A.1 4 IW 0.057
64 A.1 3 SEW 0.071 16 A.0 4 IW 0.021 64 A.1 3 SEW 0.059
64 A.1 4 SEW 0.072 4 C.1 3 IW 0.023 4 C.1 4 IW 0.065
32 A.0 4 SEW 0.077 64 A.1 3 SEW 0.024 64 A.0 3 SEW 0.067
64 A.0 4 IW 0.077 64 A.0 3 SEW 0.024 64 A.0 4 SEW 0.067
2048 0.151(0.076) 0.100(0.082) 0.150(0.083)
64 A.1 4 SEW 0.064 4 C.1 4 IW 0.020 64 A.1 4 SEW 0.053
32 A.1 4 IW 0.064 32 C.1 4 SEW 0.021 32 A.1 4 IW 0.056
64 A.1 3 SEW 0.070 16 A.0 4 IW 0.021 64 A.1 3 SEW 0.057
64 A.1 1 IW 0.073 64 A.0 4 IW 0.022 64 A.0 4 IW 0.059
64 A.0 4 SEW 0.074 32 A.1 2 IW 0.022 64 A.0 4 SEW 0.060
4096 0.150(0.077) 0.097(0.083) 0.147(0.085)
64 A.1 4 SEW 0.063 128 A.1 4 SEW 0.020 64 A.1 4 SEW 0.052
32 A.1 4 IW 0.064 4 C.1 4 IW 0.020 32 A.1 4 IW 0.056
64 A.1 3 SEW 0.069 32 A.1 2 IW 0.021 64 A.1 3 SEW 0.056
64 A.1 1 IW 0.070 32 C.1 4 SEW 0.021 64 A.1 1 IW 0.057
128 A.0 2 SEW 0.072 16 A.0 4 IW 0.021 64 A.0 4 IW 0.059
Av. 0.164 (0.073) 0.119 (0.076) 0.169 (0.078) 32 A.1 4 IW 0.079 4 C.1 4 IW 0.021 4 C.1 4 IW 0.067 4 C.1 4 IW 0.084 4 C.1 3 IW 0.023 4 C.1 4 SEW 0.080
64 A.1 3 SEW 0.089 4 C.0 4 IW 0.033 32 A.1 4 IW 0.085
32 A.1 3 SEW 0.090 4 C.0 3 IW 0.033 4 C.1 3 IW 0.085
A. Serani et al. / Applied Soft Computing 49 (2016) 313–334 327
BestperformingsetupsforADPSO,Ndv≥10.
Nfeval/Ndv Average(STD) BestADPSO
x f t NNdvp Init Coef Wall x
Np
Ndv Init Coef Wall f
Np
Ndv Init Coef Wall t
128 0.386 (0.100) 0.268 (0.146) 0.363 (0.125)
4 A.1 4 SEW 0.253 4 A.1 3 SEW 0.112 4 A.1 4 SEW 0.213
2 A.1 4 SEW 0.255 2 C.1 4 SEW 0.113 2 A.1 4 SEW 0.216
4 C.1 4 SEW 0.256 4 A.1 4 SEW 0.114 4 C.1 4 SEW 0.217
4 C.1 2 SEW 0.257 2 A.1 1 IW 0.115 2 A.1 2 IW 0.218
2 A.1 2 SEW 0.260 4 C.1 4 SEW 0.117 4 A.1 3 SEW 0.220
256 0.374(0.101) 0.249(0.143) 0.349(0.125)
4 A.1 4 SEW 0.246 4 A.1 4 SEW 0.100 4 A.1 4 SEW 0.202
4 C.1 4 SEW 0.250 8 A.1 3 SEW 0.108 4 C.1 4 SEW 0.209
4 C.1 2 SEW 0.252 4 C.1 4 SEW 0.109 8 C.1 4 SEW 0.212
8 C.1 4 SEW 0.252 4 A.1 3 SEW 0.110 4 C.1 2 IW 0.213
2 A.1 4 SEW 0.254 8 C.1 3 SEW 0.110 2 A.1 4 SEW 0.215
512 0.361(0.101) 0.232(0.141) 0.333(0.124)
8 C.1 4 SEW 0.245 8 C.1 4 SEW 0.098 4 A.1 4 SEW 0.200
4 A.1 4 SEW 0.245 4 A.1 4 SEW 0.099 8 C.1 4 SEW 0.200
16 A.1 4 SEW 0.246 16 A.1 3 SEW 0.103 16 A.1 4 SEW 0.206
4 C.1 4 SEW 0.249 8 A.1 3 SEW 0.104 8 A.1 4 SEW 0.206
16 C.1 4 SEW 0.250 8 A.1 4 SEW 0.104 4 C.1 4 SEW 0.207
1024 0.354 (0.101) 0.221 (0.142) 0.324 (0.125)
32 A.1 4 SEW 0.237 8 C.1 4 SEW 0.095 16 A.1 4 SEW 0.197
16 A.1 4 SEW 0.239 16 A.1 4 SEW 0.097 8 C.1 4 SEW 0.198
16 C.1 4 SEW 0.242 16 C.1 4 SEW 0.097 16 C.1 4 SEW 0.199
32 A.1 4 IW 0.243 4 A.1 4 SEW 0.099 32 A.1 4 SEW 0.200
8 C.1 4 SEW 0.243 16 A.1 3 SEW 0.099 4 A.1 4 SEW 0.200
2048 0.349 (0.102) 0.214 (0.143) 0.317 (0.126)
32 A.1 4 SEW 0.228 32 A.1 4 SEW 0.091 32 A.1 4 SEW 0.187
64 A.1 3 SEW 0.233 32 C.1 4 SEW 0.093 64 A.1 3 SEW 0.194
16 A.1 4 SEW 0.238 64 A.1 3 SEW 0.095 16 A.1 4 SEW 0.195
32 A.1 4 IW 0.239 16 C.1 4 SEW 0.095 32 A.1 4 IW 0.196
16 C.1 4 SEW 0.239 16 A.1 4 SEW 0.095 16 C.1 4 SEW 0.196
4096 0.346(0.103) 0.210(0.144) 0.314(0.127)
32 A.1 4 SEW 0.227 64 C.1 4 SEW 0.088 32 A.1 4 SEW 0.186
64 A.1 3 SEW 0.231 32 A.1 4 SEW 0.089 64 A.1 3 SEW 0.190
128 A.1 3 SEW 0.236 64 A.1 3 SEW 0.092 64 A.1 4 IW 0.193
64 A.1 4 IW 0.236 32 C.1 4 SEW 0.092 128 A.1 3 SEW 0.193
64 A.1 4 SEW 0.237 64 A.1 4 SEW 0.092 64 A.1 4 SEW 0.193
Av. 0.362(0.100) 0.232(0.141) 0.333(0.124)
4 A.1 4 SEW 0.246 4 A.1 4 SEW 0.101 4 A.1 4 SEW 0.203
8 C.1 4 SEW 0.248 8 C.1 4 SEW 0.104 8 C.1 4 SEW 0.206
16 A.1 4 SEW 0.249 8 A.1 3 SEW 0.108 4 C.1 4 SEW 0.209
32 A.1 4 SEW 0.250 4 C.1 4 SEW 0.109 16 A.1 4 SEW 0.211
328 A.Seranietal./AppliedSoftComputing49(2016)313–334
Table8
SuggestedguidelineforSDPSOandADPSO.
Ndv Np/Ndv Initialization Coefficientsset Wall-type
SDPSO <10≥10 44 C.1A.1 44 IW/SEWSEW
ADPSO <10≥10 44 C.1A.1 44 IW/SEWSEW
Fig.13. PerformanceofsuggestedguidelinesusingSDPSO.
Figs. 5a, 7a,9a, and 11a showthat a higher number of parti-clesperformsbetterwithahighbudgetoffunctionevaluations. Thiscanbeexplainedrecallingtherelationbetweenthenumber ofparticles and thenumber ofiterations, fora fixedbudget of
functionevaluations.Ontheonehand,alownumberofparticles requires alarger numberof iterations,providing a fast conver-gence butpossibly alsoa faststagnation. On theotherhand, a highnumber ofparticlesrequiresa lowernumberofiterations,
Fig.15.Sensitivityanalysis.
Fig.16.ConvergenceofSDPSOandADPSO.
330 A.Seranietal./AppliedSoftComputing49(2016)313–334 Table9
SBDresults.
Wall-type RT[N] W[N] f[–] f[%]
Designspace Original 50.15 852.5 5.88e−2 –
4D
SDPSO IW 39.92 850.9 4.69e−2 −20.24 SEW 40.36 850.6 4.67e−2 −20.57 ADPSO IWSEW 39.8540.68 851.1851.7 4.68e−24.71e−2 −20.41−19.90
6D
SDPSO IWSEW 39.4134.29 849.1835.4 4.64e−24.10e−2 −21.10−30.27 ADPSO IWSEW 34.3134.08 835.5830.8 4.11e−24.09e−2 −30.10−30.30
Fig.19.6DSBDresults.
providingaslowerconvergenceandpossiblydelayingstagnation. Thisisthereasonwhythenumberofparticlesshouldbe deter-mined considering theavailable budget, that is usually low in theSBDcontext,especiallywhenCPU-timeexpensivesolversare used.
The coefficient set and the wall-type approach seem to be the less relevant parameters. Nevertheless, it can be said that Clerc’scoefficients(=0.721,c1=c2=1.655[31])andtheSEW-type approacharethemosteffectiveandefficientchoicesforbothSDPSO andADPSO(seeFigs.5c,7c,11c,5d,7d,9d,and11d).Moreover,the
useofSEWissuggestedinordertopreventanearlystopofthe swarmparticledynamics,asshownforthe6Ddesignspaceofthe navalengineeringproblem.
The asynchronous mechanism of ADPSO shows equivalent or slightly better performance (in terms of number of objec-tive function evaluations and objective reduction) compared to SDPSO, probably due to the diversity in the information update. This aspect of ADPSO provides a further interesting opportunityfortheexploitationofparallelarchitecturesin HPC systems.
332 A.Seranietal./AppliedSoftComputing49(2016)313–334 7. Conclusions
A guideline for an effective and efficient useof SDPSO and ADPSO,in a computationalframeworkcharacterizedby limited resources,has been suggested. A parametric analysis hasbeen performedvaryingthe numberof particles,theinitialization of theswarm, the setof coefficients, and the wall-type approach fortheboxconstraints.Theassessmentisbasedon100 analyt-ical test functions (with dimensionality from two to fifty) and three differentabsoluteperformance criteria. Allpossible com-binationsofDPSOparametersledto420optimizationsforeach function.ThemostpromisingDPSOsetupshavebeendiscussed, identified, and successfully applied to a ship SBD optimization problem, namely thehull-form optimization of the high-speed Delftcatamaran,advancingincalmwateratfixedspeed,usinga potential-flowsolver.Theoptimizationaimedatthereductionof theratiobetweenthetotalresistanceandtheshipweight,using four-andsix-dimensionaldesignspaces.
Theparticlesinitializationhasbeenfoundthemostsignificant parameterfortheDPSOperformance,especiallyforanumberof designvariablesNdv≥10andlowbudgetsoffunctionevaluations.
Conversely,thecoefficientssetandthewall-typeapproachhave beenfoundhavinga little influenceon theDPSO performance, comparedtotheotherparameters.
Forproblemswithlessthen10variables,thesuggestedSDPSO andADPSOsetupsfollow:(a)numberofparticlesNp equalto4
timesthenumberofdesignvariables;(b)particleinitializationwith HSSdistributionondomainandboundswithnon-nullvelocity;(c)
setof coefficientsproposedin[31],i.e.,=0.721,c1=c2=1.655;
(d)inelasticandsemi-elasticwall-typeapproach,forSDPSOand ADPSOrespectively.
Forproblemswithmorethen10designvariables,thesuggested setupsforSDPSOandADPSOfollow:(a)number ofparticlesNp
equalto4timesthenumberofdesignvariables;(b)particles ini-tialization includingHSS distribution on domainwith non-null velocity;(c)setofcoefficientsproposedin[31];(d)semi-elastic wall-typeapproach.
Theperformanceoftheguidelinesuggestedforthetest func-tions are very close or coincident to the best setup for each evaluation budget, among all 420configurations available (see “Guide”and“Best”inFigs.13and14).Finally,theguidelinesetups haveprovedtoperformwellonboththe4Dand6DshipSBD opti-mizationproblems.
Acknowledgments
Thepresent researchissupportedby theUS Officeof Naval Research,NICOPGrantN62909-15-1-2016,underthe administra-tionofDr.Woei-MinLinandDr.Ki-HanKim,andbytheItalian Flagship Project RITMARE, coordinated by the Italian National ResearchCouncilandfundedbytheItalianMinistryofEducation. AppendixA. Listoftestfunctions
TablesA.10andA.11summarizetheanalytical testfunctions withNdv<10andNdv≥10,respectively,usedinthecurrentwork.
TableA.10
AnalyticaltestfunctionswithNdv<10.
fp Name Dimension Bounds Optimum
Ndv [l,u]j,...,Ndv f
min
f1,2 5nloc.minima(Levy) 2,5 [−5,5]Ndv 0.000
f3,4 10nloc.minima(Levy) 2,5 [−5,5]Ndv 0.000
f5,6 15nloc.minima(Levy) 2,5 [−5,5]Ndv 0.000
f7 Ackley 2 [−5,4]Ndv 0.000 f8,9 Alpine 2,5 [−9,7]Ndv 0.000 f10 Beale 2 [−4.5,4.5]Ndv 0.000 f11 Booth 2 [−10,10]Ndv 0.000 f12 Bukinn.6 2 [−15,−5]×[−3,3] 0.000 f13 Colville 4 [−10,10]Ndv 0.000 f14,15 CosineMixture 2,4 [−1,0.5]Ndv −0.100·Ndv f16,17 Dixon-Price 2,5 [−1,1]Ndv 0.000 f18 Easom 2 [−100,100]Ndv −1.000 f19,20 Exponential 2,4 [−9,7]Ndv −1.000 f21 Freudenstein-Roth 2 [−5,5]Ndv 0.000 f22 Goldstein-Price 2 [−2,2]Ndv 3.000 f23,24 Griewank 2,5 [−9,7]Ndv 0.000 f25 Hartmann.3 3 [0,1]Ndv −3.860 f26 Hartmann.6 6 [0,1]Ndv −3.320 f27 Matyas 2 [−9,7]Ndv 0.000 f28,29 MultiModal 2,5 [−1,0.5]Ndv 0.000 f30 Powell 8 [−4,5]Ndv 0.000 f31 Quartic 2 [−10,10]Ndv −0.352 f32 Rosenbrock 2 [−30,30]Ndv 0.000 f33 Schaffern.2 2 [−100,90]Ndv 0.000 f34 Schaffern.6 2 [−100,90]Ndv 0.000 f35 Schubertpenalty1 2 [−10,10]Ndv −186.731 f36 Schubertpenalty2 2 [−10,10]Ndv −186.731 f37 Shekeln.5 4 [0,10]Ndv −10.153 f38 Shekeln.7 4 [0,10]Ndv −10.403 f39 Shekeln.10 4 [0,10]Ndv −10.536
f40 Six-HumpCamelBack 2 [−2.5,2.5]×[−1.5,1.5] −1.032
f41 Sphere 2 [−5,4]Ndv 0.000
f42,43 Styblinski-Tang 2,4 [−5,5]Ndv −39.166·N
dv
f44 TestTubeHolder 2 [−10,10]Ndv −10.872
f45 Three-HumpCamelBack 2 [−5,4]Ndv 0.000
f46 Treccani 2 [−5,4]Ndv 0.000
f47 Tripod 2 [−100,100]Ndv 0.000
f48 Vincent 5 [0.25,10]Ndv −Ndv
f49 Xin-SheYangn.2 5 [−2,]Ndv 0.000
TableA.11
AnalyticaltestfunctionswithNdv≥10.
fp Name Dimension Bounds Optimum
Ndv [l,u]j,...,Ndv fmin
f1,2 5nloc.minima(Levy) 10,20 [−5,5]Ndv 0.000
f3,4 10nloc.minima(Levy) 10,20 [−5,5]Ndv 0.000
f5,6 15nloc.minima(Levy) 10,20 [−5,5]Ndv 0.000
f7,8,9 Ackley 10,30,50 [−5,4]Ndv 0.000 f10,11 Alpine 10,20 [−9,7]Ndv 0.000 f12,13,14 Dixon-Price 10,25,50 [−1,1]Ndv 0.000 f15,16 Griewank 10,20 [−9,7]Ndv 0.000 f17,18,19 Mishran.11 10,25,50 [−10,9]Ndv 0.000 f20,21 MultiModal 10,20 [−1,0.5]Ndv 0.000 f22,23,24 Pathological 10,25,50 [−100,90]Ndv 0.000 f25 Paviani 10 [2.0001,9.9999]Ndv −45.778 f26,27 Powell 16,24 [−4,5]Ndv 0.000 f28,29,30 Rastrigin 10,30,50 [−5.12,5.12]Ndv 0.000 f31,32,33 Rosenbrock 10,25,50 [−30,30]Ndv 0.000 f34,35,36 Schwefel 10,25,50 [−500,500]Ndv −0.100·N dv f37,38,39 Trigonometricn.2 10,25,50 [−500,500]Ndv 1.000 f40,41 Vincent 15,30 [0.25,10]Ndv −Ndv f42,43,44 Xin-SheYangn.2 10,25,50 [−2,]Ndv 0.000 f45,46,47 Xin-SheYangn.4 10,25,50 [−10,10]Ndv −1.000 f48,49,50 Zacharov 10,25,50 [−5,10]Ndv 0.000 References
[1]J.Kennedy,R.Eberhart,Particleswarmoptimization,in:Proceedingsofthe FourthIEEEConferenceonNeuralNetworks,Piscataway,NJ,1995,pp. 1942–1948.
[2]Y.Zhang,S.Wang,G.Ji,Acomprehensivesurveyonparticleswarm optimizationalgorithmanditsapplications,Math.Probl.Eng.(2015).
[3]M.Kandasamy,D.Peri,Y.Tahara,W.Wilson,M.Miozzi,S.Georgiev,E. Milanov,E.F.Campana,F.Stern,Simulationbaseddesignoptimizationof waterjetpropelledDelftcatamaran,Int.Shipbuild.Prog.60(1)(2013) 277–308.
[4]X.Chen,M.Diez,M.Kandasamy,E.F.Campana,F.Stern,Designoptimization ofthewaterjet-propelledDelftcatamaranincalmwaterusingURANS,design ofexperiments,metamodelsandswarmintelligence,in:Proceedingsof12th InternationalConferenceonFastSeaTransportation(FAST2013),Amsterdam, TheNetherlands,2013.
[5]X.Chen,M.Diez,M.Kandasamy,Z.Zhang,E.F.Campana,F.Stern,High-fidelity globaloptimizationofshapedesignbydimensionalityreduction,metamodels anddeterministicparticleswarm,Eng.Optim.47(4)(2015)473–494.
[6]M.Diez,E.F.Campana,F.Stern,Design-spacedimensionalityreductionin shapeoptimizationbyKarhunen–Loèveexpansion,Comput.MethodsAppl. Mech.Eng.283(2015)1525–1544.
[7]M.Diez,A.Serani,E.F.Campana,S.Volpi,F.Stern,Designspace dimensionalityreductionforsingle-andmulti-disciplinaryshape optimization,in:AIAA/ISSMOMultidisciplinaryAnalysisandOptimization (MA&O),AVIATION2016,WashingtonDC,USA,June13–17,2016.
[8]B.Rosenthal,D.Kring,Optimaldesignofaliftingbodyswath,in:Proceedings of12thInternationalConferenceonFastSeaTransportation(FAST2013), Amsterdam,TheNetherlands,2013.
[9]S.Volpi,M.Diez,N.Gaul,H.Song,U.Iemma,K.K.Choi,E.F.Campana,F.Stern, Developmentandvalidationofadynamicmetamodelbasedonstochastic radialbasisfunctionsanduncertaintyquantification,Struct.Multidiscip. Optim.51(2)(2015)347–368.
[10]M.Diez,S.Volpi,A.Serani,F.Stern,E.F.Campana,Simulation-baseddesign optimizationbysequentialmulti-criterionadaptivesamplinganddynamic radialbasisfunctions,in:EUROGEN2015,InternationalConferenceon EvolutionaryandDeterministicMethodsforDesign,Optimizationand ControlwithApplicationstoIndustrialandSocietalProblems,Glasgow,UK, September,2015.
[11]E.F.Campana,G.Liuzzi,S.Lucidi,D.Peri,V.Piccialli,A.Pinto,Newglobal optimizationmethodsforshipdesignproblems,Optim.Eng.10(4)(2009) 533–555.
[12]A.Carlisle,G.Dozier,Anoff-the-shelfPSO,in:ProceedingoftheWorkshopon ParticleSwarmOptimization,2001,pp.1–6.
[13]G.Venter,J.Sobieszczanski-Sobieski,Aparallelparticleswarmoptimization algorithmacceleratedbyasynchronousevaluations,in:Proceedingsofthe 6thWorldCongressesofStructuralandMultidisciplinaryOptimization,Rio deJaneiro,Brazil,2005.
[14]B.Koh,A.George,R.Haftka,B.Fregly,Parallelasynchronousparticleswarm optimization,Int.J.Numer.MethodsEng.67(2006)578–595.
[15]N.A.A.Aziz,Z.Ibrahim,Asynchronousparticleswarmoptimizationforswarm robotics,Proc.Eng.41(2012)951–957.
[16]Y.Shi,R.Eberhart,Parameterselectioninparticleswarmoptimization,in: ProceedingsofEvolutionaryProgrammingVII(EP98),1998,pp.591–600.
[17]I.Trelea,Theparticleswarmoptimizationalgorithm:convergenceanalysis andparameterselection,Inf.Process.Lett.85(2003)317–325.
[18]R.Poli,J.Kennedy,T.Blackwell,Particleswarmoptimization,SwarmIntell.1 (1)(2007)33–57.
[19]A.Engelbrecht,Particleswarmoptimization:velocityinitialization,in:2012 IEEECongressonEvolutionaryComputation,2012,pp.1–8.
[20]P.Cazzaniga,M.S.Nobile,D.Besozzi,Theimpactofparticlesinitializationin PSO:parameterestimationasacaseinpoint,in:2015IEEEConferenceon ComputationalIntelligenceinBioinformaticsandComputationalBiology (CIBCB),2015,pp.1–8.
[21]J.C.Bansal,P.K.Singh,M.Saraswat,A.Verma,S.S.Jadon,A.Abraham,Inertia weightstrategiesinparticleswarmoptimization,in:2011ThirdWorld CongressonNatureandBiologicallyInspiredComputing(NaBIC),2011,pp. 633–640.
[22]S.Helwig,J.Branke,S.Mostaghim,Experimentalanalysisofboundhandling techniquesinparticleswarmoptimization,IEEETrans.Evol.Comput.17(2) (2013)9–24.
[23]A.RezaeeJordehi,J.Jasni,Parameterselectioninparticleswarmoptimisation: asurvey,J.Exp.Theor.Artif.Intell.25(4)(2013)527–542.
[24]A.Serani,M.Diez,C.Leotardi,D.Peri,G.Fasano,U.Iemma,E.F.Campana,On theuseofsynchronousandasynchronoussingle-objectivedeterministic particleswarmoptimizationinshipdesignproblems,in:Proceedingsofthe 1stInternationalConferenceinEngineeringandAppliedSciences Optimization,Kos,Greece,June4–6,2014.
[25]E.F.Campana,G.Fasano,D.Peri,PenaltyfunctionapproachesforShip MultidisciplinaryDesignOptimization,Eur.J.Ind.Eng.6(2012)765–784.
[26]S.Lucidi,M.Piccioni,Randomtunnelingbymeansofacceptance-rejection samplingforglobaloptimization,J.Optim.TheoryAppl.62(2)(1989) 255–277.
[27]E.F.Campana,G.Fasano,A.Pinto,Dynamicanalysisfortheselectionof parametersandinitialpopulation,inparticleswarmoptimization,J.Global Optim.48(3)(2010)347–397.
[28]M.Jamil,X.-S.Yang,Aliteraturesurveyofbenchmarkfunctionsforglobal optimisationproblems,Int.J.Math.Model.Numer.Optim.4(2013).
[29]E.F.Campana,M.Diez,U.Iemma,G.Liuzzi,S.Lucidi,F.Rinaldi,A.Serani, Derivative-freeglobalshipdesignoptimizationusingglobal/local hybridizationoftheDIRECTalgorithm,Optim.Eng.17(1)(2015) 127–156.
[30]T.T.Wong,W.S.Luk,P.A.Heng,SamplingwithHammersleyandHalton points,J.Graph.Tools2(2)(1997)9–24.
[31]M.Clerc,StagnationAnalysisinParticleSwarmOptimizationorWhat HappensWhenNothingHappens.Tech.rep.,2006 http://hal.archives-ouvertes.fr/hal-00122031.
[32]D.Peri,F.Tinti,Amultistartgradient-basedalgorithmwithsurrogatemodel forglobaloptimization,Commun.Appl.Ind.Math.3(1)(2012).
[33]P.Bassanini,U.Bulgarelli,E.F.Campana,F.Lalli,Thewaveresistanceproblem inaboundaryintegralformulation,Surv.Math.Ind.4(1994)151–194.
[34]Y.Shi,R.Eberhart,Amodifiedparticleswarmoptimizer,in:IEEEInternational ConferenceonEvolutionaryComputation,Anchorage,Alaska,1998.
[35]M.Clerc,Theswarmandthequeen:towardsadeterministicandadaptive particleswarmoptimization,in:Proceedingsofthe1999Congresson EvolutionaryComputation,1999,CEC99,vol.3,1999,pp.1951–1957.
[36]R.Eberhart,Y.Shi,Comparinginertiaweightsandconstrictionfactorin particleswarmoptimization,in:CongressonEvolutionaryComputing,vol.1, 2000,pp.84–88.
[37]R.Eberhart,Y.Shi,Particleswarmoptimization:developmentsapplications andresources,in:Proceedingsofthe2001CongressonEvolutionary Computation,2001,pp.81–86.
334 A.Seranietal./AppliedSoftComputing49(2016)313–334 [38]M.Clerc,J.Kennedy,Theparticleswarm:explosion,stabilityandconvergence
inamulti-dimensionalcomplexspace,IEEETrans.Evol.Comput.6(1)(2002) 58–73.
[39]C.W.Cleghorn,A.P.Engelbrecht,Ageneralizedtheoreticaldeterministic particleswarmmodel,SwarmIntell.8(1)(2014)35–59.
[40]C.W.Cleghorn,A.P.Engelbrecht,Particleswarmvariants:standardized convergenceanalysis,SwarmIntell.9(2)(2015)177–203.
[41]M.Diez,A.Serani,C.Leotardi,E.F.Campana,D.Peri,U.Iemma,G.Fasano,S. Giove,AProposalofPSOParticles’InitializationforCostlyUnconstrained OptimizationProblems:ORTHOinitLectureNotesinComputerScience,vol. 8794,SpringerInternationalPublishing,2014,pp.126–133.
[42]K.Deb,Anefficientconstrainthandlingmethodforgeneticalgorithms, Comput.MethodsAppl.Mech.Eng.186(2–4)(2000)311–338.
[43]M.Clerc,ConfinementsandBiasesinParticleSwarmOptimization.Tech.rep., 2006http://clerc.maurice.free.fr/pso.
[44]D.Bratton,J.Kennedy,Definingastandardforparticleswarmoptimization, in:Proc.IEEESwarmIntell.,2007,pp.120–127.
[45]M.Corazza,G.Fasano,R.Gusso,ParticleSwarmOptimizationwith non-smoothpenaltyreformulation,foracomplexportfolioselection problem,Appl.Math.Comput.224(2013)611–624.
[46]A.Levy,A.Montalvo,S.Gomez,A.Galderon,TopicsinGlobalOptimization LectureNotesinMathematics,vol.909,Springer-Verlag,1981.
[47]H.Schlichting,K.Gersten,Boundary-LayerTheory,Springer-Verlag,Berlin, 2000.