/
1IntroductionANashequilibriumofabimatrixgameisapairofstrategiesinwhich 1IntroductionANashequilibriumofabimatrixgameisapairofstrategiesinwhich

1IntroductionANashequilibriumofabimatrixgameisapairofstrategiesinwhich - PDF document

tatyana-admore
tatyana-admore . @tatyana-admore
Follow
356 views
Uploaded On 2016-08-24

1IntroductionANashequilibriumofabimatrixgameisapairofstrategiesinwhich - PPT Presentation

4NashequilibriumForWSNEthecurrentbestresultwasgivenbyFearnleyetal10and ndsa2 3WSNEwhere000473ItbuildsonanapproachofKontogiannisandSpirakis13which ndsa2 3WSNEinpolynomialtimeusing ID: 455466

4-Nashequilibrium.For-WSNE thecurrentbestresultwasgivenbyFearnleyetal.[10]and ndsa(2 3)-WSNE where=0:00473.ItbuildsonanapproachofKontogiannisandSpirakis[13] which ndsa2 3-WSNEinpolynomialtimeusing

Share:

Link:

Embed:

Download Presentation from below link

Download Pdf The PPT/PDF document "1IntroductionANashequilibriumofabimatrix..." 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.


Presentation Transcript

1IntroductionANashequilibriumofabimatrixgameisapairofstrategiesinwhichthesupportsofbothplayersconsistonlyofbestresponses.TheapparenthardnessofcomputinganexactNashequilibrium[6,5]eveninabimatrixgamehasledtoworkoncomputingapproximateNashequilibria,andtwonotionsofapproximateNashequilibriahavebeendeveloped.The rstandmorewidelystudiednotionisofan-approximateNashequilibrium(-Nash).Here,norestrictionisplaceduponthesupports;anystrategycanbeinthesupportsprovidedeachplayerachievesanexpectedpayo thatiswithinofabestresponse.Therefore,-Nashequilibriahaveapracticaldrawback:aplayermightplaceprobabilityonastrategythatisarbitrarilyfarfrombeingabestresponse.Thesecondnotion,de nedtorectifythisproblem,iscalledan-wellsupportedapproximateNashequilibrium(-WSNE).Here,thecontentofthesupportsarerestricted,butlessstringentlythaninanexactNashequilibrium.Speci cally,bothplayerscanonlyplacepositiveprobabilityonstrategiesthathavepayo withinofapurebestresponse.Observethatthelatternotionisastrongerequilibriumconcept:every-WSNEisan-Nash,buttheconverseisnottrue.Approximatewell-supportedequilibriarecentlyplayedanimportantroleinunderstandingthehardnessofcomputingNashequilibria.Theyaremoreusefulinthesecontextsthan-Nashequilibriabecausetheirde nitionismorecombina-torialandmorecloselyresemblesthebestresponseconditionthatcharacterizesexactNashequilibria.Indeed,approximatewell-supportedequilibriawereintro-ducedin[12,6]inthecontextofPPADreductionsthatshowthehardnessofcomputing(approximate)Nashequilibria.TheyweresubsequentlyusedasthenotionofapproximateequilibriumbyChenetal.[5]thatshowedthePPAD-hardnessofcomputinganexactNashequilibriumevenforbimatrixgames.Anotheractiveareaofresearchistoinvestigatethebest(smallest)thatcanbeguaranteedinpolynomialtime.For-Nash,thecurrentbestalgorithm,duetoTsaknakisandSpirakis[18],achievesa0:3393-Nashequilibrium;see[8,7,3]forotheralgorithms.Fortheimportantclassofwin-losegames{gameswithpayo sinf0;1g{[18]givesa1 4-Nashequilibrium.For-WSNE,thecurrentbestresultwasgivenbyFearnleyetal.[10]and ndsa(2 3�)-WSNE,where=0:00473.ItbuildsonanapproachofKontogiannisandSpirakis[13],which ndsa2 3-WSNEinpolynomialtimeusinglinearprogramming.ThealgorithmofKontogiannisandSpirakisproducesa1 2-WSNEofwin-losegamesinpolynomialtime,whichisbest-known(themodi cationsofFearnleyetal.donotleadtoanimprovedapproximationguaranteeforwin-losegames).2 wecanhopefor{aprobabilisticargument[13]showsthatthereexist-WSNEwithsupportsofcardinalityO(1 2logn),forany�0.72ALowerBoundontheSupportSizeofWellSupportedNashEquilibriaWebeginbyformallyde ningbimatrixwin-losegamesandwell-supportedNashequilibria.Abimatrixgameisa2-playergamewithmnpayo matricesAandB;wemayassumethatmn.Thegameiscalledwin-loseifeachmatrixentryisinf0;1g.Apairofmixedstrategiesfp;qgisaNashequilibriumifeverypurerowstrategyinthesupportofpisabestresponsetoqandeverypurecolumnstrategyinthesupportofqisabestresponsetop.Arelaxationofthisconceptisthefollowing.Apairofmixedstrategiesfp;qgisan-wellsupportedNashequilibriumifeverypurestrategyinthesupportofp(resp.q)isan-approximatebestresponsetoq(resp.p).Thatis,foranyrowriinthesupportofpwehaveeiTAqmax`e`TAq�and,foranycolumncjinthesupportofqwehavepTBejmax`pTBe`�:Inthissectionweproveourmainresult.Theorem1.Forany2 3,thereexistwin-losegamesforwhichevery-well-supportedNashequilibriumhassupportsofcardinality (3p logn).Toprovethisresult,we rstformulateourwin-losegamesgraphically.Thiscanbedoneinastraight-forwardmanner.Simplyobservethatwemayrepresenta2-playerwin-losegamebyadirectedbipartitegraphG=(R[C;E).Thereisavertexforeachrowandavertexforeachcolumn.Thereisanarc(ri;cj)2Eifandonlyif(B)ij=bij=1;similarlythereisanarc(cj;ri)2Eifandonlyif(A)ij=aij=1.Consequently,wearesearchingforagraphwhosecorrespondinggamehasnohighqualitywell-supportedNashequilibriumwithsmallsupports.Weshow 7ForNashequilibria,Althofer[1]andLiptonandYoung[15]independentlyprovedsimilarresultsforzero-sumgames,andLiptonatal.[16]laterprovedsimilarresultsforgeneral-sumbimatrixandmulti-playergames.4 theexistenceofsuchagraphprobabilistically.TheConstruction.LetT=(V;E)bearandomtournamentonNnodes.NowcreatefromTanauxiliarybipartitegraphG(T)=(R[C;A)correspondingtoa2-playerwin-losegameasfollows.Theauxiliarygraphhasavertex-bipartitionR[CwherethereisavertexofRforeachnodeofTandthereisavertexofCforeachsetofkdistinctnodesofT.(Observethat,forclaritywewillrefertonodesinthetournamentTandverticesinthebipartitegraphG.)TherearetwotypesofarcinG(T):thoseorientedfromRtoCandthoseorientedfromCtoR.Forarcsoftheformertype,eachvertexX2Cwillhavein-degreeexactlyk.Speci cally,letXcorrespondtothek-tuplefv1;:::;vkgwherevi2V(T),forall1ik.Thentherearearcs(vi;X)inGforall1ik.NextconsiderthelattertypeofarcinG.Foreachnodeu2Rthereisanarc(X;u)inGifandonlyifudominatesX=fv1;:::;vkginthetournamentT,thatisif(u;vi)arearcsinTforall1ik.Thiscompletestheconstructionoftheauxiliarygraph(game)G.WesaythatasetofverticesW=fw1;:::;wtgiscoveredifthereexistsavertexysuchthat(wj;y)2A,forall1jt.Furthermore,abipartitegraphisk-coveredifeverycollectionofkverticesthatlieonthesamesideofthebipartitioniscovered.NowwithpositiveprobabilitytheauxiliarygraphG(T)isk-covered.Lemma1.Forallsucientlylargenandk3p logn,thereexistsatourna-mentTwhoseauxiliarybipartitegraphG(T)isk-covered.Proof.Observethatthepayo matricesthatcorrespondtoG(T)havem=Nrowsandn=�Nkcolumns.Furthermore,byconstruction,anysetofkverticesinRiscovered.Thus, rstwemustverifythatanysetofkverticesinCisalsocovered.SoconsideracollectionX=fX1;:::;XkgofkverticesinC.SinceeachXi2Ccorrespondstoak-tupleofnodesofT,weseethatXcorrespondstoacollectionofatmostk2nodesinT.Thus,foranynodeu=2[iXi,wehavethatuhasanarcinTtoeverynodein[iXiwithprobabilityatleast2�k2.Thuswithprobabilityatmost(1�1 2k2)N�k2thesubsetXofCnotcoveredinG(T).Applyingtheunionboundwehavethatthereexiststhedesiredtournamentifnk1�1 2k2N�k21(1)5 Nowsetk=log1 3n.Thereforelogn1 k=log2 3n=k2.Inaddition,becausen=�Nk,wehavethatNk en1 k.Hence,N�k2�n1 k.(Notethat,sinceNkthisimpliesthatG(T)isde ned.)Consequently,nk1�1 2k2N�k2nk1�1 2k2n1 knke�1 2k2n1 knke�1 ek2log2n1 kThus,takinglogarithms,weseethatInequality(1)holdsifek2log2klognn1 k(2)Butn1 k=ek2,soInequality(2)clearlyholdsforlargen.Theresultfollows.ApropertyoftheauxiliarygraphG(T)thatwillbeveryusefultousisthatitcontainsnocycleswithlessthansixvertices.Lemma2.TheauxiliarygraphG(T)containsnodigonsandno4-cycles.Proof.SupposeG(T)containsadigonfw;Xg.Thearc(w;X)impliesthatX=fx1;:::;xk�1;wg.Ontheother-hand,thearc(X;w)impliesthatwdominatesXinTand,thus,w=2X.SupposeG(T)containsa4-cyclefw;X;z;YgwherewandzareinRandwhereX=fx1;:::;xk�1;wgandY=fy1;:::;yk�1;zgareinC.ThenzmustdominateXinTandwmustdominateYinT.ButthenwehaveadigoninTas(w;z)and(z;w)mustbearcsinT.ThiscontradictsthefactthatTisatournament.Lemmas1and2arealreadysucienttoproveamajordistinctionbetweenapproximate-Nashequilibriaandwell-supportedNashequilibria.Recallthattherealwaysexist1 2-Nashequilibriawithsupportsofcardinalityatmosttwo[8].Insharpcontrast,forsupportsofcardinalityatmosttwo,noconstantap-proximationguaranteecanbeobtainedfor-well-supportedNashequilibria.Theorem2.Forany�0,thereexistwin-losegamesforwhichnopairofstrategyvectorswithsupportsizesatmosttwoisa(1�)-well-supportedNashequilibrium.Proof.Taketheauxiliarywin-losegameG(T)fromLemma1forthecasek=2.Nowconsideranypairofstrategyvectorsp1andp2withsupportsofcardinality6 Supposep1andp2forma-well-supportedequilibriumforsome2 3.Thenitmustbethecasethateachri2S1hasexpectedpayo atleast1��1 3againstp2.Similarly,eachcj2S2hasexpectedpayo atleast1��1 3againstp1.Butthiscannothappen.ConsiderthesubgraphofG(T)inducedbyS1[S2whereeachri2S1hasweightwi=p1(ri)andeachcj2S2hasweightwj=p2(cj).WeconvertthisintoanunweightedgraphHbymakingLwvcopiesofeachvertexv,forsomelargeintegerL.NowHisanLLbipartitegraphwithminimumin-degree(1�)L�1 3L.Thus,byTheorem3,Hcontainsa4-cycle.Thisisacontradiction,byLemma2.utWeremarkthatthe2 3inTheorem1cannotbeimprovedusingthisprooftechnique.Speci callytheminimumin-degreerequirementof1 3kinTheorem3istight.Toseethis,takeadirected6-cycleCandreplaceeachvertexinCby1 3kcopies.ThuseacharcinCnowcorrespondstoacompletek 3k 3bipartitegraphwithallarcorientationsinthesamedirection.ThegraphHcreatedinthisfashionisbipartitewithallin-degrees(andallout-degrees)equalto1 3k.ClearlytheminimumlengthofadirectedcycleinHissix.3ConclusionAnoutstandingopenproblemiswhetheranyconstantapproximationguaranteebetterthan1isachievablewithconstantcardinalitysupports.Wehaveshownthatsupportsofcardinalitytwocannotachievethis;apositiveresolutionofConjecture1wouldsucetoshowthatsupportsofcardinalitythreecan.How-ever,Conjecture1seemsahardgraphproblemanditiscertainlyconceivablethatitisfalse.8Ifso,thatwouldleadtotheintriguingpossibilityofaverymajorstructuraldi erencebetween-Nashand-WSNE;namely,thatforany�0,thereexistwin-losegamesforwhichnopairofstrategieswithconstantcardinalitysupportsisa(1�)-well-supportedNashequilibrium.Theexistenceofsmallsupport-WSNEclearlyimpliestheexistenceofofpolynomialtimeapproximationalgorithmsto ndsuchequilibria.Obtainingbetterapproximationguaranteesusingmorecomplexalgorithmsisalsoanin-terestingquestion.Asdiscussed,thebestknownpolynomial-timeapproxima-tionalgorithmforwell-supportedequilibriainwin-losegames ndsa1 2-wellsupportedequilibrium[13]bysolvingalinearprogram(LP).Forgameswith 8Forexample,theconjectureresemblesaquestionabouttheexistenceofk-existentiallycompletetriangle-freegraphsfork�3referredtoin[9],whichtheauthorsconsidertobewideopen.9

Related Contents


Next Show more