您好,欢迎访问三七文档
Penalty/BarrierMultiplierMethodsforConvexProgrammingProblemsAharonBen-Tal1andMichaelZibulevsky2FacultyofIndustrialEngineeringandManagementTechnion|IsraelInstituteofTechnologyHaifa32000,IsraelRevisedversion|March1995AbstractWestudyaclassofmethodsforsolvingconvexprograms,whicharebasedonnonquadraticAugmentedLagrangiansforwhichthepenaltyparametersarefunc-tionsofthemultipliers.Thisgivesrisetolagrangianswhicharenonlinearinthemultipliers.Eachaugmentedlagrangianisspeci edbyachoiceofapenaltyfunction’andapenalty-updatingfunction .Therequirementson’aremild,andallowfortheinclusionofmostofthepreviouslysuggestedaugmentedlagrangians.Moreim-portantly,anewtypeofpenalty/barrierfunction(havingalogarithmicbranchgluedtoaquadraticbranch)isintroducedandusedtoconstructane cientalgorithm.Convergenceofthealgorithmsisprovedforthecaseof beingasublinearfunc-tionofthedualmultipliers.Thealgorithmsaretestedonlarge-scalequadraticallyconstrainedproblemsarisinginstructuraloptimization.1ResearchsupportedbyagrantfromtheIsrael-U.S.A.BinationalScienceFoundation(BSF).2ResearchsupportedbyTheCenterforAbsorptioninScience,MinistryofImmigrantAbsorp-tion,StateofIsrael1IntroductionMethodsofmultipliersforconstrainedconvexprograms,involvingnonquadraticaugmentedLagrangiansaregettingrenewedattention.Newtheoreticalresultsaregivenforashiftedlogarithmicmultipliermethodin[12]and[13].Amoregeneralschemeisstudiedin[9].Implementationsreportinggoodnumericalresultsaregivenin[3]and[7].Fortheexponentialmethodofmultipliersnewconvergenceresults,includingrateofconvergence(forthelinearprogrammingcase)areobtainedin[18].HereweintroduceaclassofmethodscalledPenaltyBarrierMultiplier(PBM)methods,whicharebasedonnonquadraticaugmentedlagrangians.AmemberinthePBMclassisspeci edbyapenalty/barrierfunction’andapenaltyupdatingfunction ,responsibleforupdatingthepenaltyparametersineachiteration.Therequirementson’arerathermild,andallowfortheinclusionasspecialcases,theexponentialandtheshiftedlogarithmicfunctions.Moreimportantly,inx4wesug-gestanewtypeofpenalty/barrierfunctionmadeofalogarithmicbranchgluedsmoothlytoaquadraticbranch.APBMalgorithmbasedonthislog-quadraticaugmentedlagrangianprovedtobeverye cientandcapableofsolvinglarge-scaleproblemstoahighdegreeofaccuracy(seex7,x8).Therequirementonthepenaltyupdatingfunction isthatitisasublinearfunctionofthemultipliers.Thisre-quirementwasinspiredbyasuggestioninthepaperbyTsengandBertsekas[18].Wepointoutthatanaugmentedlagrangianresultingfromsuchachoiceof isanonlinearfunctionofthemultipliers.Inx4weshowthatthePBMmethodisassociatedwitha\proximalpointalgorithm,whichsimultaneouslysolvesthedualconvexprogrammingproblem.Thedistance-likefunctionappearingintheproxi-maltermisrelatedtotheconjugatefunctionof’.Convergencepropertiesoftheproximalpointalgorithmisstudiedinx5,andtheresultsobtainedthereareinstru-mentalinprovingthatthePBMalgorithmproducesasequenceofpointswhichareasymptoticallyprimalfeasible.FullconvergenceanalysisofthePBMmethodiscon-tainedinx6.Itisshown(seeTheorem1)thattheprimal-dualsequencesgeneratedbythealgorithmarebounded,andeachoftheirlimitpointsisapairofsolutionsfortheprimalanddualproblems.Inx7wediscussimplementationissuesofthePBMmethodandinx8weapplythealgorithmtolarge-scalequadraticallycon-strainedconvexproblemswhichariseintwotypesofstructuraloptimization|(1)TrussTopologyDesignand(2)ShapeDesign.Forthe rstapplication,thelargestproblemsolvedhas462variablesand16290quadraticconstraints.Forthesecondapplication,thelargestproblemsolvedhas6498variablesand3136quadraticcon-straints.Thecomputationalresults3demonstratethatthePBMmethodsolvessuchproblemsinalmosta xednumberofNewtonsteps(typically30)independentlyoftheproblemsdimensions.3AdditionalcomputationalresultsfortheTrussTopologyDesignproblemwerereportedinanearlierversionofthispaper,see[3].12ProblemFormulationandAssumptionsWestudytheordinaryconvexprogrammingproblem(inthesenseofRockafellar[14,p.277])(P)f =infff(x):gi(x) 0;i=1;:::;mgwherethefunctionsf,gi:::;gm:IRn!Rareclosedproperconvexfunctions.WeletSdenotethefeasiblesetofPandS thesetofoptimalsolutions.Itisassumedthroughoutthepaperthat(A1)S isnonemptyandcompact.Associatedwithproblem(P)isitsLagrangianL(x;u):=f(x)+mXi=1uigi(x);itsdualfunctionG(u):=inffL(x;u):x2IRngandthedualconcaveproblem(D)G =supfG(u):u 0g.OursecondworkingassumptionisSlater’scondition:(A2)exists^x2domfsuchthatgi(^x)0;i=1;:::;m:Followingthisassumption,itiswellknownthattheoptimalsolutionsetofthedualproblem(D)isnonemptyandcompact,andthatf =G .Moreover,foreach 0G ,thelevelsetfu2IRm:u 0;G(u) 0giscompact.Notethatwedonotassumedi erentiabilityofthefunctionsf;g1;:::;gm.3ThePenalty/BarrierMultiplier(PBM)MethodWetransformtheconstraintsofproblem(P)usingareal-valuedfunction’havingthefollowingproperties(’0)’isastrictlyincreasingtwicedi erentiablestrictlyconvexfunctionwithdom’=( 1;b);0b 12(’1)’(0)=0(’2)’0(0)=1(’3)limt!b’0(t)=1(’4)limt! 1’0(t)=0(’5)’00(t) 1Mforallt2[0;b],forsomeM0.Wenextshowthatproperties(’1){(’4)implyanimportantpropertyoftherecessionfunction’1of’.Recall[14,x8]that’1(s):=lim !1’(t+ s) ’(t
本文标题:Penaltybarrier multiplier methods for convex progr
链接地址:https://www.777doc.com/doc-3311719 .html