consensusasshownbyExample1belowExercise31inp517of3InparticularconvergencetoconsensusfailseveninthespecialcaseoftheequalneighbormodelThemainideaisthattheagreementalgorithmcancloselyemulatean ID: 89933
Download Pdf The PPT/PDF document "Convergence in multiagent coordination c..." 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.
Equalneighbormodel:Here,aij(t)=1=ni(t);ifj2Ni(t),0;ifj=2Ni(t),whereNi(t)=fjj(j;i)2E(t)gisthesetofagentsjwhosevalueistakenintoaccountbyiattimet,andni(t)isitscardinality.ThismodelisalinearversionofamodelconsideredbyVicseketal.[15].NotethatheretheconstantofAssumption1isequalto1=n.Pairwiseaveragingmodel([5]):Thisisthespecialcaseofboththesymmetricmodelandoftheequalneighbormodelinwhich,ateachtime,thereisasetofdisjointpairsofagentswhocommunicate(bidirectionally)witheachother.Ificommunicateswithj,thenxi(t+1)=xj(t+1)=(xi(t)+xj(t))=2.Notethatthesumx1(t)++xn(t)isconserved;therefore,ifconsensusisreached,ithastobeontheaverageoftheinitialvaluesofthenodes.Theassumptionbelowisreferredtoaspartialasynchro-nismin[3].Wewillseethatitissometimesnecessaryforconvergence.Assumption3.(Boundedintercommunicationintervals)Ificommunicatestojaninnitenumberoftimes[thatis,if(i;j)2E(t)innitelyoften],thenthereissomeBsuchthat,forallt,(i;j)2E(t)[E(t+1)[[E(t+B1).III.CONVERGENCERESULTSINTHEABSENCEOFDELAYS.Wesaythattheagreementalgorithmguaranteesasymp-toticconsensusifthefollowingholds:foreveryx(0),andforeverysequencefA(t)gallowedbywhateveras-sumptionshavebeenplaced,thereexistssomecsuchthatlimt!1xi(t)=c,foralli.Theorem1.UnderAssumptions1,2(connectivity),and3(boundedintercommunicationintervals),theagreementalgorithmguaranteesasymptoticconsensus.Theorem1subsumesthespecialcasesofsymmetryoroftheequalneighbormodel,andthereforesubsequentconver-genceresultsandproofsforthosecases.Theorem1ispresentedin[14]andisprovedin[13];asimpliedproof,forthespecialcaseofxedcoefcientscanbefoundin[3].Themainidea,whichappliestomostresultsofthistype,isasfollows.Letm(t)=minixi(t)andM(t)=maxixi(t).SinceeachA(t)isstochastic,itisstraightforwardtoverifythatm(t)andM(t)arenondecreas-ingandnonincreasing,respectively.ItthensufcestoverifythatthedifferenceM(t)m(t)isreducedbyaconstantfactoroverasufcientlylargetimeinterval;theintervalischosensothateveryagentgetstoinuence(indirectly)everyotheragent;bytracingthechainofsuchinuences,andusingtheassumptionthateachinuencehasanontrivialstrength(ourassumptionthatwheneveraij(t)isnonzero,itisboundedbelowby0),theresultfollows.Intheabsenceoftheboundedintercommunicationintervalassumption,thealgorithmdoesnotguaranteeasymptotic consensus,asshownbyExample1below(Exercise3.1,inp.517of[3]).Inparticular,convergencetoconsensusfailseveninthespecialcaseoftheequalneighbormodel.Themainideaisthattheagreementalgorithmcancloselyemulateanonconvergentalgorithmthatkeepsexecutingthethreeinstructionsx1:=x3,x3:=x2,x2:=x1,oneaftertheother.Example1.Letn=3,andsupposethatx(0)=(0;0;1).Let1beasmallpositiveconstant.Considerthefollowingsequenceofevents.Agent3communicatestoagent1;agent1formstheaverageofitsownvalueandthereceivedvalue.Thisisrepeatedt1times,wheret1islargeenoughsothatx1(t1)11.Thus,x(t1)(1;0;1).Wenowletagent2communicatestoagent3,t2times,wheret2islargeenoughsothatx3(t1+t2)1.Inparticular,x(t1+t2)(1;0;0).Wenowrepeattheabovetwoprocesses,innitelymanytimes.Duringthekthrepetition,1isreplacedbyk(andt1;t2getadjustedaccordingly).Furthermore,bypermutingtheagentsateachrepetition,wecanensurethatAssumption2issatised.Afterkrepetitions,itcanbecheckedthatx(t)willbewithin11kofaunitvector.thus,ifwechoosetheksothatP1k=1k1=2,asymptoticconsensuswillnotbeobtained.Ontheotherhand,inthepresenceofsymmetry,theboundedintercommunicationintervalassumptionisunnec-essary.Thisresultisprovedin[9]and[4]forthespecialcaseofthesymmetricequalneighbormodelandin[11],[7],forthemoregeneralsymmetricmodel.AmoregeneralresultwillbeestablishedinTheorem4below.Theorem2.UnderAssumptions1and2,andforthesym-metricmodel,theagreementalgorithmguaranteesasymp-toticconsensus.IV.PRODUCTSOFSTOCHASTICMATRICESANDCONVERGENCERATETheorem1and2canbereformulatedasresultsontheconvergenceofproductsofstochasticmatrices.Corollary1.ConsideraninnitesequenceofstochasticmatricesA(0);A(1);A(2);:::,thatsatisesAssumptions1and2.IfeitherAssumption3(boundedintercommunicationintervals)issatised,orifwehaveasymmetricmodel,thenthereexistsanonnegativevectordsuchthatlimt!1A(t)A(t1)A(1)A(0)=1dT:(Here,1isacolumnvectorwhoseelementsareallequaltoone.)AccordingtoWolfowitz'sTheorem([16])convergenceoccurswheneverthematricesarealltakenfromanitesetofergodicmatrices,andthenitesetissuchthatanyniteproductofmatricesinthatsetisagainergodic.Corollary1extendsWolfowitz'theorembynotrequiringthematricesA(t)tobeergodic,thoughitislimitedtomatriceswithpositivediagonalentries. increasingsequenceoftimes,witht0=0andtk+1tk!1.Iftkttk+1,theagentsupdateaccordingtox1(t+1)=(x1(t)+x2(tk))=2;x2(t+1)=(x1(tk)+x2(t))=2:Wewillthenhavex1(t1)=11andx2(t1)=1,where1-277;0canbemadearbitrarilysmall,bychoosingt1largeenough.Moregenerally,betweentimetkandtk+1,theabsolutedifferencejx1(t)x2(t)jcontractsbyafactorof12k,wherethecorrespondingcontractionfactors12kapproach1.IfthekarechosensothatPkk1,thenQ1k=1(12k)]TJ/;༔ ; .96; Tf; 11.;ɸ ; Td;[000;0,andthedisagreementjx1(t)x2(t)jdoesnotconvergetozero.Accordingtotheprecedingexample,theassumptionofboundeddelayscannotberelaxed.Ontheotherhand,theassumptionofboundedintercommunicationintervalscanberelaxed,inthepresenceofsymmetry,leadingtothefollowinggeneralizationofTheorem2,whichisanewresult.Theorem4.UnderAssumptions1,2(connectivity),and4(boundeddelays),andforthesymmetricmodel,theagreementalgorithmwithdelays[cf.Eq.(1)]guaranteesasymptoticconsensus.Proof.LetMi(t)=maxfxi(t);xi(t1);:::;xi(tB+1)g;M(t)=maxiMi(t);mi(t)=minfxi(t);xi(t1);:::;xi(tB+1)g;m(t)=minimi(t):Aneasyinductiveargument,asinp.512of[3],showsthatthesequencesm(t)andM(t)arenondecreasingandnonincreasing,respectively.Theconvergenceproofrestsonthefollowinglemma.Lemma1:Ifm(t)=0andM(t)=1,thenthereexistsatimetsuchthatM()m()1nB.GivenLemma1,theconvergenceproofiscompletedasfollows.Usingthelinearityofthealgorithm,thereexistsatime1suchthatM(1)m(1)(1nB)(M(0)m(0)).ByapplyingLemma1,withtreplacedbyk1,andusinginduction,weseethatforeverykthereexistsatimeksuchthatM(k)m(k)(1nB)k(M(0)m(0)),whichconvergestozero.This,togetherwiththemonotonic-itypropertiesofm(t)andM(t),impliesthatm(t)andM(t)convergetoacommonlimit,whichisequivalenttoasymptoticconsensus.q.e.d.ProofofLemma1:Fork=1;:::;n,wesaythatPropertyPkholdsattimetifthereexistatleastkindicesiforwhichmi(t)kB.Weassume,withoutlossofgenerality,thatm(0)=0andM(0)=1.Then,m(t)0forallt,becauseofthemonotonicityofm(t).Furthermore,thereexistssomeiandsome2fB+1;B+2;:::;0gsuchthatxi()=1.Usingtheinequalityxi(t+1)xi(t),weobtainmi(+ B)B.ThisshowsthatthereexistsatimeatwhichpropertyP1holds.Wecontinueinductively.SupposethatknandthatPropertyPkholdsatsometimet.LetSbeasetofcardinalitykcontainingindicesiforwhichmi(t)kB,andletScbethecomplementofS.Letbethersttime,greaterthanorequaltot,atwhichaij()6=0,forsomej2Sandi2Sc(i.e.,anagentjinSgetstoinuencethevalueofanagentiinSc).Suchatimeexistsbytheconnectivityassumption(Assumption2).Notethatbetweentimestand,theagents`inthesetSonlyformconvexcombinationsbetweenthevaluesoftheagentsinthesetS(thisisaconsequenceofthesymmetryassumption).SinceallofthesevaluesareboundedbelowbykB,itfollowsthatthislowerboundremainsineffect,andthatm`()kB,forall`2S.Fortimess,andforevery`2S,wehavex`(s+1)x`(s),whichimpliesthatx`(s)kBB,fors2f+1;:::;+Bg.Therefore,m`(+B)(k+1)B,forall`2S.Considernowanagenti2Scforwhichaij()6=0.Wehavexi(+1)aij()xj(ij())mi()kB+1:Usingalsothefactxi(s+1)xi(s),weobtainthatmi(+B)(k+1)B.Therefore,attime+B,wehavek+1agentswithm`(+B)(k+1)B(namely,theagentsinS,togetherwithagenti).ItfollowsthatPropertyPk+1issatisedattime+B.ThisinductiveargumentshowsthatthereisatimeatwhichPropertyPnissatised.Atthattimemi()nBforalli,whichimpliesthatm()nB.Ontheotherhand,M()M(0)=1,whichprovesthatM()m()1nB.q.e.d.Thesymmetrycondition[(i;j)2E(t)iff(j;i)2E(t)]usedinTheorem4issomewhatunnaturalinthepresenceofcommunicationdelays,asitrequiresperfectsynchronizationoftheupdatetimes.Alooserandmorenaturalassumptionisthefollowing.Assumption5.ThereexistssomeB-521;0suchthatwhenever(i;j)2E(t),thenthereexistssomethatsatisesjtjBand(j;i)2E().Assumption5allowsforprotocolssuchasthefollowing.Agentisendsitsvaluetoagentj.Agentjrespondsbysendingitsownvaluetoagenti.Bothagentsupdatetheirvalues(takingintoaccountthereceivedmessages),withinaboundedtimefromreceivingotheragent'svalue.Inarealisticsetting,withunreliablecommunications,eventhisloosesymmetryconditionmaybeimpossibletoenforcewithabsolutecertainty.Onecanimaginemorecomplicatedprotocolsbasedonanexchangeofacknowledgments,butfundamentalobstaclesremain(seethediscussionofthetwo-armyprobleminpp.32-34of[2]).Amorerealisticmodelwouldintroduceapositiveprobabilitythatsomeoftheupdatesarenevercarriedout.(Asimplepossibilityisto