AlsoFinlandDistinguishedProfessorHelsinkiSchoolofEconomicsPOBox1210FIN00101HelsinkiFinlandKalyanmoyDebhse 2simplymaximizingMobjectivefunctionsoneatatimeThisisbecausethesearchoftheworst ID: 383311
Download Pdf The PPT/PDF document "AHybridIntegratedMulti-ObjectiveOptimiza..." is the property of its rightful owner. Permission is granted to download and print the materials on this web site for personal, non-commercial use only, and to display it on your personal computer provided you do not modify the materials and that you retain all copyright notices contained in the materials. By downloading content from our website, you accept the terms of this agreement.
AHybridIntegratedMulti-ObjectiveOptimizationProcedureforEstimatingNadirPointKalyanmoyDeb1??,KaisaMiettinen2,andDeepakSharma11DepartmentofMechanicalEngineeringIndianInstituteofTechnologyKanpur,PIN208016,Indiafdeb,dsharmag@iitk.ac.in2DepartmentofMathematicalInformationTechnologyP.O.Box35(Agora),FI-40014UniversityofJyvaskyla,Finlandkaisa.miettinen@jyu.fiAbstract.AnadirpointisconstructedbytheworstobjectivevaluesofthesolutionsoftheentirePareto-optimalset.Alongwiththeidealpoint,thenadirpointprovidestherangeofobjectivevalueswithinwhichallPareto-optimalsolutionsmustlie.Thus,anadirpointisanimportantpointtoresearchersandpractitionersinterestedinmulti-objectiveopti-mization.Besides,ifthenadirpointcanbecomputedrelativelyquickly,itcanbeusedtonormalizeobjectivesinmanymulti-criteriondecisionmakingtasks.Importantly,estimatingthenadirpointisachallengingandunsolvedcomputingproblemincaseofmorethantwoobjectives.Inthispaper,wereviseapreviouslyproposedserialapplicationofanEMOandalocalsearchmethodandsuggestanintegratedapproachforndingthenadirpoint.Alocalsearchprocedurebasedonthesolutionofabi-levelachievementscalarizingfunctionisemployedtoextremesolu-tionsinstabilizedpopulationsinanEMOprocedure.Simulationresultsonanumberofproblemsdemonstratetheviabilityandworkingoftheproposedprocedure.1IntroductionAnadirpointsignies,inprinciple,oppositetothatmeantbyanidealpoint,inthecontextofmulti-objectiveoptimization.AnidealpointisanM-dimensionalobjectivevector(whereMisthenumberofobjectives)constructedwithbestfeasibleobjectivevaluesandisacomparativelyeasytocompute.Forminimiza-tionproblems,inprinciple,thiscallsforsolvingMsingle-objectiveminimizationproblemsandcollectingeachoptimalobjectivevaluestoformtheidealpoint.Ontheotherhand,anadirpointisconstructedwiththeworstobjectivevaluesofPareto-optimalsolutions.Inminimizationproblems,thistaskisdierentfrom ??AlsoFinlandDistinguishedProfessor,HelsinkiSchoolofEconomics,POBox1210,FIN-00101,Helsinki,Finland,(Kalyanmoy.Deb@hse.) 2simplymaximizingMobjectivefunctionsoneatatime.ThisisbecausethesearchoftheworstvalueofanobjectivemustberestrictedwithinthePareto-optimalsolutions.Thisisthereasonwhytheestimationofnadirpointhasbeenfoundtobeacomplextask[13,11]andtheredoesnotexistanyprovablealgo-rithmforthetask,evenforlinearmulti-objectiveoptimizationproblemshavingthreeormoreobjectives.Withtheadventofecientevolutionaryoptimizationproceduresformulti-objectiveoptimization,someattentionhasbeenmadeintherecentpastindevel-opingproceduresforestimatingthenadirpoint.Simplisticideas,suchasndingasetofPareto-optimalsolutionsbyanEMOprocedureandthenchoosingtheextremesolutionsforestimatingthenadirpoint,tomoresophisticatedideas,suchasreplacingthefocusofEMOtondawide-spreadedsetofsolutionsontheentirePareto-optimalfronttondonlythecriticalextremePareto-optimalpoints[4,16],aresuggested.MostoftheseEMOmethodologieshaveshowntondanapproximationofthenadirpoint,ratherthantoestimatetheexactnadirpoint.Recentstudies[6,5]suggestedatwo-stepserialprocedureofemployingamodiedNSGA-IIproceduretoidentifyextremenearPareto-optimalsolutionsandthenalocalsearchproceduretoconvergetothetrueextremePareto-optimalpoints.Inthisstudy,wesuggestandsimulateahybridintegratedapproachinwhichalocalsearchprocedureisusedwithinthemodiedNSGA-IIalgorithmspar-inglytoachievethenadirpointestimationtask.Thesuggestedlocalsearchprocedureisbasedonutilizingareferencepointbasedapproach,aso-calledachievementscalarizingfunction[17]whichiswidelyusedintheMCDMeld.Usingthisscalarizedfunction,anypointintheobjectivespacecanbeprojectedontheParetooptimalfrontandthescalarizingfunctiondoesnotneedanyarti-cialinformationlikeweights[14].Intheprocedureproposed,theachievementscalarizingfunctionisusedinabi-levelmannertoguaranteegettingreliableenoughinformationaboutextremevaluesintheParetooptimalfrontforesti-matingthenadirpoint.BasedonastatisticalanalysisoftheperformanceoftheNSGA-IIprocedure,theexecutionofthelocalsearcheventisdecideddy-namicallyateverygeneration.BothNSGA-IIandlocalsearchproceduresareterminatedusingstatisticalperformancecriteria.Simulationresultsonanumberoftestproblemsandthreeengineeringproblemsarepresentedtodemonstratetheecacyoftheproposedprocedure.2NadirObjectiveVectorWeconsidermulti-objectiveoptimizationproblemsinvolvingMcon\rictingob-jectives(fi:S!R)asfunctionsofdecisionvariablesx:minimizeff1(x);f2(x);:::;fM(x)g;subjecttox2S;(1)whereSRndenotesthesetoffeasiblesolutions.Problem(1)givesrisetoasetofPareto-optimalsolutionsoraPareto-optimalfront(P),providingatrade-o 3amongtheobjectives.Inthesenseofminimizationofobjectives,Pareto-optimalsolutionscanbedenedasfollows[14]:Denition1Adecisionvectorx2Sandthecorrespondingobjectivevectorf(x)arePareto-optimaliftheredoesnotexistanotherdecisionvectorx2Ssuchthatfi(x)fi(x)foralli=1;2;:::;Mandfj(x)fj(x)foratleastoneindexj.Inwhatfollows,weassumethatthePareto-optimalfrontisbounded.Wenowdeneanadirobjectivevector,thatis,anadirpoint,asfollows.Denition2Anobjectivevectorznad=(znad1;:::;znadM)TconstructedusingtheworstvaluesofobjectivefunctionsinthecompletePareto-optimalfrontPiscalledanadirobjectivevector.Hence,forminimizationproblemswehaveznadj=maxx2Pfj(x).Estimationofthenadirobjectivevectoris,ingeneral,adiculttask.Unliketheidealobjectivevectorz=(z1;:::;zM)T,whichcanbefoundbyminimizingeachobjectiveindividuallyoverthefeasiblesetS(or,zj=minx2Sfj(x)),thenadirpointcannotbeformedbymaximizingobjectivesindividuallyoverS.Tondthenadirpoint,Pareto-optimalityofsolutionsusedforconstructingthenadirpointmustberstestablished.Thismakesthetaskofndingthenadirpointadicultone.Toillustratethisaspect,letusconsiderabi-objectiveminimizationproblemshowninFigure1.Ifwemaximizef1andf2individually,weobtainpointsAandB,respectively.Thesetwopointscanbeusedtoconstructtheso-calledworstobjectivevector,zw.Inmanyproblems(eveninbi-objectiveoptimizationproblems),thenadirobjectivevectorandtheworstobjectivevectorarenotthesamepoint,whichcanalsobeseeninFigure1. Pareto- optimal frontABWorst objective vectorIdealpointobjective spaceFeasiblef1f2 Nadir objective vector Fig.1.Thenadirandworstobjectivevectors. ZnadZ'BIdealpointZ*AB'CA'C'G'F'E'D' 1 0.8 0.6 0.4 0.2 0.4 0.6 0.8 1 0 0.2 1.2 1.4 1.6 0.2 0.4 0.6 0.8 1 1.2 1.4 0f1f2 1.2 1.4f3 Fig.2.Payotablemaynotproducethetruenadirpoint. 43ExistingMethods3.1PayoTableMethodBenayounetal.[1]introducedtherstinteractivemulti-objectiveoptimizationmethodforestimatingthenadirpointbyusingapayotable.Tobemorespe-cic,eachobjectivefunctionisrstminimizedindividuallyandthenatableisconstructedwherethei-throwofthetablerepresentsvaluesofallobjectivefunctionscalculatedatthepointwherethei-thobjectiveobtaineditsminimumvalue.Thereafter,themaximumvalueofthej-thcolumncanbeconsideredasanestimateoftheupperboundofthej-thobjectiveinthePareto-optimalfrontandthesemaximumvaluestogethermaybeusedtoconstructanapproximationofthenadirobjectivevector.Themaindicultyofsuchanapproachisthatsolutionsarenotnecessarilyuniqueandthuscorrespondingtotheminimumsolutionofanobjectivetheremayexistmorethanonesolutionshavingdier-entvaluesofotherobjectives,inproblemshavingmorethantwoobjectives.Intheseproblems,thepayotablemethodmaynotresultinanaccurateestimationofthenadirobjectivevector.Toillustrate,considerathree-objectiveproblemshowninFigure2.MinimizationoftherstobjectivewillresultinanysolutiononthetrapeziumCBB0F0C0C.IfthepointmarkedinasmallcircleonlineCBisobtainedbyanoptimizationalgorithmandsimilarlyothertwocirclesonlinesCAandABareobtainedforminimizationsoff2andf3,respectively,awrongestimate(z0)ofthenadirpoint(znad)willbemade.3.2EvolutionaryApproachesThenadirpointisassociatedwithPareto-optimalsolutionsand,thus,deter-miningasetofPareto-optimalsolutionswillfacilitatetheestimationofthenadirpoint.SinceanEMOalgorithmisaimedatndingasetofPareto-optimalsolutions,itmaybeanidealwaytondthenadirobjectivevector.Severalapproachesareproposedrecently.Inthenaiveapproach,rstawell-distributedsetofPareto-optimalsolutionscanbeattemptedtondbyanEMO[4].Thereafter,anestimateofthenadirobjectivevectorcanbemadebypickingtheworstvaluesofeachobjective[16].InthecontextoftheproblemdepictedinFigure2,thismeansrstndingawell-representedsetofsolutionsontheplaneABCandthenestimatingthenadirpointfromthem.SinceEMOalgorithmsarenotfoundtoconvergewellandmaintainawell-diversesetofsolutionsformorethanthreeobjectives[7],theaccuracyoftheestimatednadirpointusingthenaiveapproachisquestionable.SzczepanskiandWierzbicki[16]havesimulatedtheideaofsolvingmultiplebi-objectiveoptimizationproblemssuggestedin[8]usinganEMOapproachandconstructthenadirpointbyaccumulatingallbi-objectivePareto-optimalfrontstogether.Asdiscussedinourearlierstudy[5],suchatechniqueisnotgenericandrequiresadditionalobjectiveandvariable-spacenichingtechniquestocorrectlyestimatethenadirpoint.Moreover,theprocedurerequires M2 5bi-objectiveoptimizations,makingitadauntingtaskparticularlyforproblemshavingmorethanthreeobjectives.However,theideaofconcentratingonapreferredregiononthePareto-optimalfront,insteadofndingtheentirePareto-optimalfront,canbepushedfurther.AnemphasiscanbeplacedinanEMOapproachtondonlythecriticalextremepointsofthePareto-optimalfront.Ourearlierstudy[4]suggestedtwoapproachesinthecrowdingdistanceoperatoroftheNSGA-IIprocedureandcon-cludedinfavoroftheextremizedcrowdingdistanceapproach.Intheextremized-crowdedNSGA-IIapproach[4],weemphasizedinconcentratingonthebestandworstsolutionsofeachobjective.Inthisapproach,solutionsonaparticularnondominatedfrontarerstsortedfromminimum(withrankR(m)i=1)tomaximum(withrank=Nf)basedoneachobjective.Therankofsolutioniforthem-thobjectiveR(m)iisassignedasmaxfR(m)i;Nf R(m)i+1g.TwoextremesolutionsforeveryobjectivegetarankequaltoNf(numberofsolutionsinthenondominatedfront),thesolutionsnexttotheseextremesolutionsgetarank(Nf 1),andsoon.Afterarankisassignedtoasolutionbyeachobjective,themaximumvalueoftheassignedranksisdeclaredasthecrowdingdistance.Likeotherevolutionaryoptimizationstudies,theproposedextremizedcrowdedNSGA-IIapproachdidnotensureconvergingtothetrueextremesolutionsex-actly,asevolutionaryalgorithmsareexpectedtondanear-optimalsolution,ratherthanatrueoptimalsolutioninanitenumberofsolutionevaluations.However,inthepursuitofestimatingthenadirpointforthepurposeofnormal-izingobjectivesforexecutingdierentmulti-objectiveoptimizationalgorithmsorforknowingthetruerangeofPareto-optimalsolutionsfordecision-making,itisimportanttondthetrueextremePareto-optimalpoints,sothatthenadirpointcanbeestimatedaccurately.Inarecentstudy[6],theextremizedcrowdedNSGA-IIapproachisendedwithabi-levellocalsearchoperationonallextremesolutionstotakethemarbitraryclosertothetrueextremesolutions,sothatthenadirpointcanbeestimatedmoreaccurately.Inthispaper,were-addresstheissueoftheserialapplicationofNSGA-IIandthelocalsearchprocedureandsuggestahybridintegratedapproachforanaccurateestimationofthenadirpoint.4ProposedIntegratedApproachInsteadofapplyingthelocalsearchontheextremesolutionsobtainedbytheex-tremizedcrowdedhybridNSGA-IIprocedure,weproposeanintegratedNSGA-IIapproachinwhichateverygenerationtheextremesolutionsofthebestnon-dominatedfrontaremodiedbythelocalsearchproceduretopushthemtowardstheirtruevalues.Althoughthetaskincreasesthenumberofsolutionevaluationsofeachgeneration,webelievethattheattainedaccuracyoftheintegratedproce-dureisbetterandhasasmallerchanceofgettingstucktointermediatesolutions,whichmaynotleadtoanaccurateestimationofthenadirpoint.Inthefollowing,weoutlineaniterationoftheproposedintegratedNSGA-IIprocedureinwhichthepopulationPtisthecurrentparentpopulationofsizeN.Everymember(i) 6ofPtisalreadyrankedbasedonitsnon-dominationlevel(NDi)anditscrowd-ingdistance(CDi)withinthepopulationmembersofitsownnon-dominationlevel.Step1:PopulationPtisusedtocreateanospringpopulationQtbyusingbinarytournamentselection,recombinationandmutationoperators.TwosolutionsarechosenatrandomfromPtandahierarchicalselectionbasedonNDfollowedbyCDisusedtocompletethetournamentselectionoperation.Thereafter,twosuchselectedsolutionsarerecombinedusingthesimulatedbinarycrossoveroperator[3,2]tocreatetwoospringsolutions,eachofwhichisthenmutatedbyusingthepolynomialmutationoperator[2].Theseoperatorsinvolvethefollowingparameters:recombinationprobabilitypc,SBXindexc,mutationprobabilitypm,andmutationindexm.Step2:PopulationsPtandQtarecombinedtogetherandrankedintodierentlevelsofnon-domination:Pt[Qt=fF1;F2;:::g.ThesetF1containsnon-dominatedsolutionsoflevelone,andsoon.Step3:Dependingonacheckonwhethertoperformthelocalsearchornot(whichwedescribealittlelater),inthesetF1,weidentifytheworstsolution(x(j))withrespecttoeachobjectivej,andmodifyitbyusingalocalsearchprocedure.Themodiedsolution(y(j))replacestheworstpopulationmem-ber.ForMobjectives,thereareMsuchlocalsearchoperationsperformedineachiterationoftheproposedprocedure.Theestimatednadirpoint(zest)atgenerationtisthenformedfromtheextremesolutionsobtainedbythelocalsearches.AllmembersofthesetF1arethenassignedanon-dominationrank(ND)valueequaltoone(beingtherst-rankedsolutions)andacrowdingdistance(CD)valuebasedontheextremizedcrowdedrankingproceduredescribedearlier[6].Formembersofothernon-dominationlevels(l2),wedonotperformthelocalsearch,butassignanon-dominationrank(ND)equaltolandcrowdingdistance(CD)valuecomputedasabove.Step4:AnewpopulationPt+1isthencreatedbycopyingsolutionsfromthebestnon-dominationlevelF1onwardsoneatatimetillwehaveNpop-ulationmembers.Whenwereachanon-dominationlevelwhichcannotbeentirelyaccepted(tonotincreasethesizeofPt+1overN),weusethecrowd-ingdistance(CD)valuesofthesettodeterminewhichsolutionsshouldbeaccepted.WesimplysortthemembersofthesetaccordingtotheirCDvaluefromhighesttolowestandchooseasmanyweneedtollupthepopulationfromthetopofthesortedlist.ThisprocedureissimilartotheoriginalNSGA-IIprocedure,exceptthatthecrowdingdistancecomputationisdierentsuitingtheneedforemphasizingex-tremesolutionsforthetaskofestimatingthenadirpointandthatalocalsearchprocedureisusedtoupdatetheextremeobjective-wisesolutionstomakesurethatthenadirpointcanbeestimatedwithadesiredaccuracy.Wenowdescribethelocalsearchprocedurehere.Thebest(fminj)andworst(fmaxj)valuesofeachobjectivejofthesetF1arerstnoted.Weapplyabi-levellocalsearchprocedurefromeachworstsolution(solutionx(j)forwhich 7thej-thobjectivehastheworstvalueinF1)tondthecorrespondingoptimalsolutiony(j)usingthefollowingbi-leveloptimizationprocedure.Theupper-leveloptimization(describedin(2))usesanobjectivevector(z,referredhereasareferencepoint)asavariablevectorandmaximizesthej-thobjectivevalueoftheoptimalsolutionobtainedbysolvingthecorrespondingaugmentedachievementscalarizingproblem[14](werefertothistaskasthelower-leveloptimizationtask,describedin(3)):maximize(z)fj(z);subjecttozif(j)iEA 0:5(fmaxi fmini);i=1;2;:::;M;zif(j)iEA+1:5(fmaxi fmini);i=1;2;:::;M:(2)Thetermfj(z)istheoptimalvalueofthej-thobjectivefunctionoftheoptimalsolutiontothefollowinglower-leveloptimizationproblemforwhichziskeptxed[17]:minimize(y)maxMi=1fi(y) zi fmaxi fmini+PMk=1fk(y) zk fmaxk fmink;subjecttoy2S;(3)Figure3illustratesthislocalsearchprocedure. f*2(A')for B2(C) for ASearch space for reference pointSearch space for reference pointf*f*1(B') QBADC A' Pf2 f1 B' Fig.3.Eacharrowcorrespondstoalower-levelsearchforaspeciedreferencepoint(C,A'orB').Theupper-levelsearchndsareferencepointhavingoptimalworstobjective(suchasA'orB').Inthelower-leveloptimiza-tionproblem,thesearchisperformedontheoriginalde-cisionvariablespace.Theso-lutiony(j)(z)tothislower-leveloptimizationproblemdeterminestheoptimalobjec-tivevectorffromwhichweextractthej-thcomponentanduseitintheupper-leveloptimizationproblem.Thus,foreveryreferencepointz(asolutionfortheupper-levelproblem),thecorrespondingoptimalaugmentedachieve-mentscalarizingfunctionisfoundinthelower-levelloop.Theupper-leveloptimizationisinitializedwiththeNSGA-IIsolutionz(0)=f(x(j))andthelower-leveloptimizationisinitializedwiththeNSGA-IIsolutiony(0)=x(j).Wenowdiscussthetermi-nationcriterionofeachopti-mizationprocedure.ForterminatingtheoverallNSGA-IIprocedure,wecompute 8anormalizeddistance(ND)metricasfollows:D=vuut 1 MMXi=1zesti zi zwi zi2:(4)Here,thevectorszandzwaretheidealandworstobjectivevectorsoftheop-timizationproblem,respectively.ThesequantitiescanbecomputedoncebeforetheNSGA-IIprocedurebysolving2Mdierentsingle-objectiveoptimizationsofminimizingandmaximizingeachobjectiveatatime.SincetheexactnalvalueoftheDmetricisnotknownapriorionanar-bitraryproblem,werecordthechangeinDforthepast(=50usedhere)generations.Say,Dmax,Dmin,andDavg,arethemaximum,minimum,andav-erageDvaluesforthepastconsecutivegenerations.IfthechangeD=(Dmax Dmin)=Davgissmallerthanathreshold(=1(10 4)isusedhere),theNSGA-IIprocedureisterminated.WeusethesamenormalizeddistancemetrictodecidewhetherthelocalsearchneedstobeperformedinaparticulargenerationofNSGA-II.Atagen-eration,thechangelDinnormalizeddistanceoverthepastl(=20usedhere)generationsisrecorded.IflD(=0:005usedhere),thelocalsearchisperformed.Thisreducesthenumberoflocalsearchesperformedfromnotsogoodsolutions.Whenthebestnon-dominatedfronthasstabilizedsomewhat,theextremesolutionsofthesetaremodiedusingthelocalsearchprocedure.Bothupperandlower-leveloptimizationtasksinthelocalsearchoperationusesapoint-by-pointsearchapproachwhichisterminatedbasedonthechosenoptimizationalgorithmandcodeusedforthepurpose.Inalloursimulations,wehaveusedKNITROforthelower-leveloptimizationtaskinwhichwehavesetaterminationconditionontheKKTerrorvalue(10 6)oramaximumof100iterationswhicheverhappensrst.Fortheupper-leveloptimizationtask,wehaveusedKNITRO'sSQPsolver.Theupper-leveltaskisterminatedifthenormoftheNewton'sdirectionislessthanofequalto0.001oramaximumiterationof100iselapsed.AftertheNSGA-IIrunisterminated,weconstructthenadirpointfromtheworstobjectivevaluesofthenalnon-dominatedsetF1.5SimulationResultsInthissection,wepresentsimulationresultsoneightproblemshavingthreeormoreobjectives.Inmostoftheseproblems,thenadirpointwasdiculttoobtainusingthepay-otable.Inallproblems,weuseapopulationofsizemax(60;20n)(nisthenumberofvariables),crossoverandmutationprobabilitiesof0.9and1=n,crossoverandmutationindicesof10and50,respectively,and=10 4.Ineachcase,wemake10dierentrunsfromdierentinitialpopulations,buteverytimetheprocedureisfoundtoconvergenearaparticularsetofextremepoints,therebyleadingtondingasimilarnadirpointeverytime. 95.1ProblemKMTherstproblemKM,adaptedfrom[12],isthefollowing:minimize8: x1 x2+51 5(x21 10x1+x22 4x2+11)(5 x1)(x2 11)9=;;subjectto3x1+x2 120;2x1+x2 90;x1+2x2 120;0x14;0x26:(5)Thetruenadirpointofthisproblemisreportedtobeznad=(5;4:6; 14:25)T[9].Table1showsthethreeextremesolutions(x)foundbyourproposedap-proach.Itisclearthatwhentheworstobjectivevaluesarecollectedtogether,weobtainanidenticalpoint(uptotwodecimalpoints)asthatinthetruenadirpoint.Figure4showsthatthenormalizeddistancevaluegetsstabilizedatTable1.Extremepointsfoundbythepro-posedapproachonproblemKM. x Estimatedznad 0.000 0.000 5.000 2.200 -55.000 0.000 6.000 -1.000 4.600 -25.001 3.500 1.501 0.000 -3.100 -14.251 Terminated at gen. 87D stabilized for 50 gen. 0 10 20 30 40 50 60 70 80 90 Fig.4.VariationofDwithgenerationonKM.around40generationandsinceD=50isused,ittookanother50generationstoterminatethehybridprocedure.Interestingly,theDvaluereachesthenalstabilizedvalueveryquickly,therebyindicatingtheeciencyoftheproposedprocedure.5.2ProblemSW1ThesecondproblemSW1isasfollows[16]:minimize8:f1(x)= (100 7x1 20x2 9x3)f2(x)= (4x1+5x2+3x3)f3(x)= x39=;;subjectto11 2x1+x2+13 5x39;x1+2x2+x310;xi0;i=1;2;3:(6) 10Thepreviousstudy[16]reportedthetruenadirpointtobeznad=( 3:6364;0;0)T.Table2showstwoextremesolutions(x)(hence,thetruenadirpoint)foundbyourproposedapproach.Figure5showstheprogressoftheproposedapproach.Table2.ExtremepointsfoundbytheproposedapproachonproblemSW1. x Estimatedznad 0.0000 3.1818 3.6364 -3.6364 -26.8182 -3.6364 0.0000 0.0000 0.0000 -100.0000 0.0000 0.0000 Stabilized for 50 gen.Terminated at gen. 87 0 10 20 30 40 50 60 70 80 90 Fig.5.VariationofDwithgener-ationonSW1.5.3ProblemSW2ThethirdproblemSW2originatesfrom[16]:minimize8-2.4;䌡-2.4;䌡:9x1+19:5x2+7:5x37x1+20x2+9x3 (4x1+5x2+3x3) (x3)9-2.4;䌡-2.4;䌡=-2.4;䌡-2.4;䌡;;subjectto1:5x1 x2+1:6x39;x1+2x2+x310;xi0;i=1;2;3:(7)Thetruenadirpointforthisproblemisreportedtobeznad=(94:5;96:3636;0;0)T[16].Theoriginalstudy[16]foundaclosepoint(94:4998;95:8747;0;0)Tusingmultiple,bi-objectiveoptimizationsimulationusinganEMOprocedure.Theoutcomeisnotidenticaltothetruenadirpoint.Table3showsthethreeex-tremesolutionsfoundbyourproposedapproach.Weobtainthetruenadirpoint.Table3.ExtremepointsfoundbytheproposedapproachonproblemSW2. x Estimatedznad 4.0000 3.0000 0.0000 94.5000 88.0000 -31.0000 0.0000 0.0000 3.1818 3.6363 89.3182 96.3636 -26.8182 -3.6363 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 11DuetoanidenticalbehaviorofDvariationwithgenerationnumberonthisandsubsequentproblems,wedonotshowthegureshere.5.4ProblemKSS1ThelinearKSS1problem[13]wasfoundtobedicultforestimatingthenadirpoint:maximize8:11x2+11x3+12x4+9x5+9x6 9x711x1+11x3+9x4+12x5+9x6 9x711x1+11x2+9x4+9x5+12x6+12x79=;;subjecttoP7i=1xi=1;xi0;i=1;2;:::;7:(8)Thetruenadirpointisreportedtobeznadir=(0;0;0)T[13].Table4showsthethreeextremesolutionsfoundbyourproposedapproach.OurapproachndsaTable4.ExtremepointsfoundbytheproposedapproachonproblemKSS1. x Estimatedznad 1.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 11.000 11.000 0.000 0.994 0.000 0.000 0.000 0.001 0.004 10.910 -0.026 11.006 0.000 0.000 1.000 0.000 0.000 0.000 0.000 11.000 11.000 0.000 nearnadirpointwithaslighterrorinthesecondobjectivevalue(asshowninFigure6theerrorisnotvisuallydetectable).Thisproblemisadicultonetosolveforestimatingtheexactnadirpoint,becauseoftheslowslopeleadingtoeachofthethreeextremepoints,asshownbyasetofrepresentativesolutionsobtainedthroughaclusteredNSGA-II,inwhichNSGA-II'scrowdingdistancemethodisreplacedbythek-meanclusteringmethod[2].Inthisproblem,itiseasytogetstucktoanon-dominatedpointclosetooneormoreextremepoints.Ourapproachseemstohavefoundtheexactextremevaluesforrstandthirdobjectivesandmanagedtogettoanear-bypointaroundtheextremeofthesecondobjective.5.5ProblemKSS2Next,weconsideranotherlinearproblemKSS2[13]:maximize(x1;x2;x3);subjecttox1+2x2+2x38;2x1+2x2+x38;3x1 2x2+4x312;xi0;i=1;2;3:(9) 12 Nadir point Extreme points (3)by proposed approachClusteredpointsNSGA-II 0 0 2 4 6 8 10 12 0 2 4 6 8 10 12 10f3f1f2 8 6 4 2 12 Fig.6.Pareto-optimalfrontshowslongnarrowregionsnearextremepointsinproblemKSS1. Nadir point frontPareto-opt 15 20 25 30 35 40 5000 10000 15000 20000 25000 30000 10Normal StressDeflection 0.012 0.008 0.004CostProposed approachNSGA-II 5 0.0158 0Normal Stress Fig.7.Pareto-optimalfrontandtwoob-tainedpointsforproblemWELD.Table5.ExtremepointsfoundbytheproposedapproachonproblemKSS2. x Estimatedznad 0.000 3.818 0.166 0.000 3.818 0.166 3.344 0.000 0.432 3.344 0.000 0.433 3.253 0.628 0.000 3.253 0.628 0.000 Thenadirpointisreportedtobeznad=(0;0;0)T.Table5presentstheex-tremesolutionsobtainedbyourapproach.Thetruenadirpointisfoundbyourapproachinthisproblem.Nowweconsiderthreemoreproblems,borrowedfromengineeringelds.Oneachoftheseproblems,theexactnadirpointisnotknown,butwhereverpossibleweexplaintheaccuracyofthenadirpointobtainedbyourapproach.5.6ProblemWELDTheWELDproblemhasfourvariablesandthreeobjectives,andisformulatedin[6].Ourpreviousstudy[6]introducedtheWELDproblemwhichhasfourvariablesandthreeobjectives.Thenadirpointwasestimatedtobeznad=(36:4209;0:0158;30000)T.Table6presentstwoextremepointsfoundbyourpro-posedapproachofthispaper.Theextremepointsforthesecondandthirdobjec-tivesarefoundtobeidenticalinthisproblem,indicatingthatalthoughtheprob-lemhasthreeobjectivefunctions,thePareto-optimalfrontistwo-dimensional,asisalsoconrmedbytheoriginalNSGA-IIpointsinFigure7.Thenadirpointestimatedbyourapproachis(36:4221;0:0158;30000:1284)T,whichisclosetothatobtainedbytheearlierstudy[6]. 13Table6.ExtremepointsfoundbytheproposedapproachonproblemWELD. x Estimatedznad 1.7356 0.4788 10.0000 5.0000 36.4221 0.000439 1008.0000 0.2444 6.2175 8.2915 0.2444 2.3810 0.015759 30000.1284 5.7ProblemCARTheseven-variable,three-objectiveCARproblemisdescribedin[10]. pointsNSGA-IIapproach (2 pts.)Proposed pointNadir 44 3.7 3.8 3.9 4 10.8 11.2 11.6 12 12.4f1f2f3 24 28 32 36 40 3.6 Fig.8.ExtremeobjectivevectorscoverstheentirePareto-optimalfrontforprob-lemCAR.Nopreviousstudyexistsonthisprob-lemforndingthenadirpoint.InTable7,wepresenttwoextremepointsobtainedbyourprocedure.Thus,thenadirpointestimatedbyourapproachforthisproblemisznad=(42:767;4:000;12:521)T.Fig-ure8showsthecompletePareto-optimalfrontwithasetofrepresen-tativeclusteredNSGA-IIsolutions.ItisclearfromtheplotthattheabovetwoextremepointsareadequatetocovertheextremeobjectivevaluesofthePareto-optimalfrontandisabletolocatethenadirpointoftheprob-lem.Table7.ExtremepointsfoundbytheproposedapproachonproblemCAR. x Estimatedznad 1.500 1.350 1.500 1.500 2.625 1.200 1.200 42.767 3.585 10.611 0.500 1.226 0.500 1.208 0.875 0.884 0.400 23.589 4.000 12.521 5.8ProblemWATERFinally,weconsidertheWATERproblem[15],whichisalsodescribedin[2].Forthisproblem,theexactnadirpointisnotknown.However,sincetherearethreevariablesandveobjectives,someredundancyintheobjectivesisexpectedforthePareto-optimalsolutions.AnapplicationofNSGA-IItothisproblem[2](page388)wasfoundtoindicatesomecorrelationsamongtheobtainedrepresen-tativesolutions.Table8presentstheextremepointsobtainedforthisproblem 14Table8.ExtremepointsfoundbytheproposedapproachonproblemWATER. x Estimatedznad 0.010 0.100 0.100 1.038 0.020 0.949 0.075 5.649 0.450 0.098 0.010 0.916 0.900 0.936 0.033 0.002 0.114 0.100 0.010 0.918 0.228 0.951 0.031 0.285 0.098 0.010 0.100 0.918 0.197 0.095 2.671 5.713 byourapproach.Weobservethattheextremepointsforobjectivesf4andf5comefromanidenticalsolution.Theestimatednadirpointusingourprocedureisznad=(1:038;0:900;0:951;2:671;5:713)T.6ConclusionsInthispaper,wehaveextendedourpreviousstudyonaserialimplementationofanEMOprocedurefollowedbyanMCDMbasedlocalsearchapproachtondextremepointsaccuratelyforestimatingthenadirpointofamulti-objectiveoptimizationproblem.Thenadirpointinmulti-objectiveoptimizationisusedinnormalizingobjectiveswhichisnecessaryfordierentmulti-criterionopti-mizationalgorithms.Besides,thetaskofestimatingthenadirpointforthreeormoreobjectivesisaopenresearchtaskinmulti-criterionoptimizationliterature.Nadirpointscanonlybeestimatedaccuratelyif(i)objective-wiseextremesand(ii)Pareto-optimalsolutionsarefound.Duetothistwo-prongedrequirements,wehavesuggestedabi-levellocalsearchtask.Thelocalsearchisemployedwithextremenon-dominatedsolutionsonlywhenthebestnon-dominatedfronthasstabilizedsomewhat,therebymakingtheoverallmethodcomputationallytractable.Onasetofvetestproblemsandthreeengineeringdesignproblems,theproposedintegratedprocedurehasabletondtheexactnadirpointquiteaccurately.Thisworkisalsoimportantfromanotherpointofview.Thisworkdemon-strateshowalocalsearchapproachcanbeintegratedwithanevolutionarypopulation-basedapproachandusedsparinglyforacomplexoptimizationtoensureaccurateconvergence.7AcknowledgmentsAuthorsacknowledgethesupportfromtheAcademyofFinland,FoundationofHelsinkiSchoolofEconomics,andtheJennyandAnttiWihuriFoundationduringthecourseofthisstudy. 15References1.R.Benayoun,J.deMontgoler,J.Tergny,andP.Laritchev.Linearprogrammingwithmultipleobjectivefunctions:Stepmethod(STEM).MathematicalProgram-ming,1(3):366{375,1971.2.K.Deb.Multi-objectiveoptimizationusingevolutionaryalgorithms.Chichester,UK:Wiley,2001.3.K.DebandR.B.Agrawal.Simulatedbinarycrossoverforcontinuoussearchspace.ComplexSystems,9(2):115{148,1995.4.K.Deb,S.Chaudhuri,andK.Miettinen.Towardsestimatingnadirobjectivevectorusingevolutionaryapproaches.InProceedingsoftheGeneticandEvolu-tionaryComputationConference(GECCO-2006),pages643{650.NewYork:TheAssociationofComputingMachinery(ACM),2006.5.K.DebandK.Miettinen.Areviewofnadirpointestimationproceduresusingevolutionaryapproaches:Ataleofdimensionalityreduction.TechnicalReportKanGALReportNumber2008004,KanpurGeneticAlgorithmsLaboratory(Kan-GAL),IndianInstituteofTechnologyKanpur,India,2008.6.K.Deb,K.Miettinen,andS.Chaudhuri.Estimatingnadirobjectivevector:Hybridofevolutionaryandlocalsearch.TechnicalReportElectronicWorkingPaperW-440,HelsinkiSchoolofEconomics,Finland,2008.7.K.DebandD.Saxena.SearchingforPareto-optimalsolutionsthroughdimen-sionalityreductionforcertainlarge-dimensionalmulti-objectiveoptimizationprob-lems.InProceedingsoftheWorldCongressonComputationalIntelligence(WCCI-2006),pages3352{3360,2006.8.M.EhrgottandD.Tenfelde-Podehl.ComputationofidealandnadirvaluesandimplicationsfortheiruseinMCDMmethods.EuropeanJournalofOperationalResearch,151:119{139,2003.9.P.Eskelinen,K.Miettinen,K.Klamroth,andJ.Hakanen.Paretonavigatorforinteractivenonlinearmultiobjectiveoptimization.ORSpectrum,inpress.10.L.Gu,R.J.Yang,C.H.Tho,L.Makowski,O.Faruque,andY.Li.Optimizationandrobustnessforcrashworthinessofsideimpact.InternationalJournalofVehicleDesign,26(4),2001.11.H.IsermanandR.E.Steuer.Computationalexperienceconcerningpayotablesandminimumcriterionvaluesovertheecientset.EuropeanJournalofOpera-tionalResearch,33(1):91{97,1988.12.K.KlamrothandK.Miettinen.Integratingapproximationandinteractibedecisionmakinginmulticrietriaoptimization.OperationsResearch,56:222{234,2008.13.P.Korhonen,S.Salo,andR.Steuer.Aheuristicforestimatingnadircriterionvaluesinmultipleobjectivelinearprogramming.OperationsResearch,45(5):751{757,1997.14.K.Miettinen.NonlinearMultiobjectiveOptimization.Kluwer,Boston,1999.15.T.Ray,K.Tai,andK.C.Seow.Anevolutionaryalgorithmformultiobjectiveoptimization.EngineeringOptimization,33(3):399{424,2001.16.M.SzczepanskiandA.P.Wierzbicki.Applicationofmultiplecrieterionevolu-tionaryalgorithmtovectoroptimization,decisionsupportandreferencepointapproaches.JournalofTelecommunicationsandInformationTechnology,3:16{33,2003.17.A.P.Wierzbicki.Theuseofreferenceobjectivesinmultiobjectiveoptimization.InG.FandelandT.Gal,editors,MultipleCriteriaDecisionMakingTheoryandApplications,pages468{486.Berlin:Springer-Verlag,1980.