IntegrativeNegotiationInComplexOrganizationalAgentSystems
TrackingNumber:163
ComputerScienceDepartment
UniversityofMassachusettsat
Amherst
XiaoqinZhang
ComputerScienceDepartment
UniversityofMassachusettsat
Amherst
VictorLesser
AutomatedReasoningGroupHoneywellLaboratories
TomWagner
wagner
shelley@cs.umass.edulesser@cs.umass.edu
negotiatetomaximizetheirownlocalutility;incooperativenegoti-ation,agentsworktofindasolutionthatincreasestheirjointutility–thesumoftheutilitiesofallinvolvedagents.Inthecompeti-tivenegotiationclass,significantwork[8,9]hasbeendoneintheareaofboundedrationalself-interestedagents(BRSI).Saidagentsareself-interestedandsocialwelfareisnotaconcern–eachagentworkstomaximizeitsownutilitythoughcontracting,biddinganddecommiting.Inthecooperativenegotiationclass,significantworkhasbeendoneintheareaofconflictresolutionthroughnegotiation[2,5,11].Inthiswork,thereisnonotionofindividualagentutility–agentsare“completely-cooperative”witheachotherandcooper-atetosolveproblemstogether.
Wefeelthatasthesophisticationofmulti-agentsystemsincreases,MASwillbeneithersimplemarketsystemswhereeachagentispurelyself-interested,seekingmaximizeitslocalutility,nordis-tributedproblemsolvingsystemswhereallagentsarecompletely-cooperativeworkingtomaximizetheachievementofasetofglobalgoals.Thiswilloccurfortworeasons.Onereasonisthatagentsfromdifferentandseparateorganizationalentitieswillcometo-gethertodynamicallyformvirtualorganization/teamforsolvingspecificproblemsthatarerelevanttoeachoftheirorganizationalentities[7].Howtheseagentsworkintheirteamwilloftendepen-dentontheexistenceofbothlongtermandshort-termrelationshipsandontheconfrontationalattitudeoftheirunderlyingorganiza-tionalentities.Wealsofeelthatevenforagentsfromself-interestedorganizations,itmightbebeneficialforthemtobepartiallyco-operativewhentheyareinthesituationswheretheywillhavere-peatedtransactionswithotheragentfromotherorganizationalenti-ties.Additionally,agentsmaybeinvolvedconcurrentlywithmorethanonevirtualorganizationswhiledoingtasksfortheirownorga-nizationalentity.Secondly,wefeelthatevenagentsworkingsolelywithagentsoftheirownorganizationalentity,itstillmaybead-vantageousforthemtotakevaryingattitudesinthespectrumoffullycooperativetototallyself-interestedinorderfortheorganiza-tiontobestachieveitsoverallgoals.Thisperspectiveisbasedonabounded-rationalargument:itisnotpossiblefromacomputationalnorcommunicationperspectiveforanagenttobefullycooperative,sinceagentsneedtotakeintoaccounttheutilitiesofallagentsintheorganizationandthestateofachievementofallorganizationalgoalstobefullycooperative.Thus,itisourfeelingthatitmaybebestfortheorganizationtohaveagentsbeingpartiallycooperativeinitslocalnegotiationwithotheragentsratherthanbeingfullyco-operativeinordertomoreeffectivelydealwithuncertaintyofnothavingacompletelyinformedandup-to-dateviewofthestateof
ThismaterialisbaseduponworksupportedbytheNationalSci-enceFoundationunderGrantNo.IIS-9812755andtheAirForceResearchLaboratory/IFTDandtheDefenseAdvancedResearchProjectsAgencyunderContractF30602-99-2-0525.TheU.S.GovernmentisauthorizedtoreproduceanddistributereprintsforGovernmentalpurposesnotwithstandinganycopyrightannotationthereon.Disclaimer:Theviewsandconclusionscontainedhereinarethoseoftheauthorsandshouldnotbeinterpretedasnecessarilyrepresentingtheofficialpoliciesorendorsements,eitherexpressedorimplied,oftheDefenseAdvancedResearchProjectsAgency,AirForceResearchLaboratory/IFTD,NationalScienceFounda-tion,ortheU.S.Government.
Permissiontomakedigitalorhardcopiesofallorpartofthisworkforpersonalorclassroomuseisgrantedwithoutfeeprovidedthatcopiesarenotmadeordistributedforprofitorcommercialadvantageandthatcopiesbearthisnoticeandthefullcitationonthefirstpage.Tocopyotherwise,torepublish,topostonserversortoredistributetolists,requirespriorspecificpermissionand/orafee.
Copyright2001ACMX-XXXXX-XX-X/XX/XX...$5.00.
Degree of Concernfor Own Outcomes
CompetiveCooperative
Sharing
(Compromising)
Avoidant
Accommodative
Degree of Concern for Other’s Outcomes
Figure1:Thedualconcernmodel
theentireagentorganization.
Multi-agentsystemwillthusconsistoflargegroupsoflooselycoupledagentsthatworktogetherontasks.Therelationshipsamongagentswilldependontheirorganizationalrolesandmaybeofanytypefrompurelyself-interestedtototallycooperative.Thisisthecomplexorganizationalproblemspacethe[12,13]frameworkisdesignedtorepresent.Notethatthisworkpertainstocomplexagents.WeassumethatAgentsareautonomous,heterogeneous,persistent,computingentitiesthathavetheabilitytochoosewhichtaskstoperformandwhentoperformthem.Agentsarealsora-tionallybounded,resourcebounded,andhavelimitedknowledgeofotheragents.1Agentscanperformtaskslocallyiftheyhavesufficientresourcesandtheymayinteractwithotheragents.Theagentswillhavechoicesaboutwithwhomtocollaborate,howtonegotiate,whattochargeforservices,etc.Further,thenegotiationstrategywillbedependentontherelationshipsamongthenegotiat-ingpartiesandtheparticularnegotiationissue.
Wethereforefeelthatinacomplexagentsociety,anagentwillneedtoworkwithotheragentsfromavarietyofdifferentorgani-zationalpositions.Forexample,anagentfromitsowngroup,anagentwhohasahigherpositionandthusmoreauthority,anagentfromacooperativecompany,oranagentfromacompetingcom-panyandsoforth.Theagent’sattitudetowardnegotiationisnotjustsimplyeithercompetingorcooperative,theagentneedstoqualita-tivelyreasonabouteachnegotiationsession,e.g.,howimportantitsownoutcomeiscomparedtotheotheragents’outcomes,soitcanchooseanappropriatenegotiationstrategy.
Figure1describesthisdualconcernmodel[6].Whentheagentonlyattachesimportancetoitsownoutcome,itsattitudetowardnegotiationiscompetitive(self-interested);whenanagentattachesthesamedegreeofimportancetoitsownoutcomeasitdoestotheoutcomesoftheotheragent,itsattitudeiscooperative;whentheagentattachesmoreimportancetotheoutcomesofotheragentsandnoimportancetoitsownoutcome,itsattitudeisaccommodative;iftheagentattachesnoimportancetoanyoutcomes,itsattitudeisavoidant(thenegotiationisnotworthitstimeandeffort).Fromthismodel,wefindthattherearepotentiallymanyoptionsbetweenthetwoextremesofself-interestedandcooperative.Theseotheroptionsdependontheimportancetheagentattachestotheincreaseofitsownutilityrelativetotheimportanceitattachestotheotheragents’utilityincreases.
Inthispaper,wepresentanintegrativemechanismthatenablesanagenttoqualitativelymanageitsattitudetowardseachnegotia-tionsession.Thismechanismisnotpurelyself-interestedorpurelycooperative,butsupportsrangesofthesebehaviorssothattheagent
Foreachbelongingtoanagent,ithasapreferencefunctionorutilitycurve,aparticularquantityofthe
,that,describesi.e.,itspreferencefor
thatwhereistheutilityassociated
suchwith
andisnotdirectlyinterchangeablewithunless.Differentagentsmayhavedifferentpreferencesfor
thesame
.Preferencesintheframeworkaredefinedbytherelationbetweentaskperformanceandorganizationalgoalsordirectives.
Anagent’soverallutilityatanygivenmomentintimeisa
functionofitsdifferentutilities:
.Wemakenoassumptionsaboutthepropertiesofthatitenablesagentstodeterminepreferenceordominance,onlybetweentwodifferentagentstateswithrespecttos.
MQTasksareabstractionsoftheprimitiveactionsthattheagentmaycarryout.tasks:
Mayhavedeadlines,,fortaskperformancebe-yondwhichperformanceofsaidtaskyieldsnousefulresults.
Mayhaveearlieststarttimes,,fortaskperformancebeforewhichperformanceofsaidtaskyieldsnousefulre-sults.Eachtaskconsistsofoneormorealternatives,whereonealternativecorrespondstoadifferentperformanceprofileofthetask.Inmanyways,thisextensionsimplifiesreasoningwiththepreliminarymodelpresentedin[12]whileatthesametimeincreasingtherepresentationalpoweroftheframeworkbycouplingdifferentdurationswiththeotherper-formancecharacteristics.Eachalternative:
–Requiressometimeordurationtoexecute,denoted.–Producessomequantityofoneormoreproductionset(s,calledanMQ,where),whichisdenoted.by:
Thesequantities
arepositiveandreflectthebenefitderivedfromperformingthetask,e.g.,progresstowardagoalortheproductionofanartifactthatcanbeexchangedwithotheragents.Inthismodel,thetwoareequivalent.–Akintothe,tasksmaycalleds.anTheMQspecificationconsumptionofsetthealsoanddenoted
sconsumeconsumedquantitiesbyataskof
is,where.Consumptionsetsmodel
tasksconsumingresources,orbeingdetrimentaltoanorga-nizationalobjective,oragentscontractingworkouttootheragents,e.g.,payinganotheragenttoproducesomedesiredre-sultoranotheragentaccumulatingfavorsorgoodwillastheresultoftaskperformance.Consumptionsetsarethenegativesideoftaskperformance.
–Allquantities,e.g.,,
fromanexpectedvaluestandpoint.
,,arecurrentlyvieweddefinesquantitiesthatarerequiredfortaskperfor-mance.Ifatasklackssufficientsforexecutionitisdeemedun-executableandwillnotbeperformedinanyfash-ion.Thismeansitwillhavezeroduration,consumezero
s,andwillproducezeros.
Spacelimitationsprecludeafullpresentationofthemodel,butitissufficientforunderstandinghowourintegrativenegotiationframeworkisbuiltupontheMQframework2.
related
MQandrelationalMQ.Theseclassesareconceptualandusedtoclearlydifferentiatemotivationsfortaskperformancefromatti-tudestowardnegotiationissues–inreality,theyarebothsimplyMQs.Goal
relatedMQs,theagentcollectsMQsforitsown
utilityincrease.Inthissense,agentB’sperformanceoftasktismo-tivatedby“self-interested”reasonsifpaymentisviaagoal
xtransferredwithit,and
foragentB,theutilitycurveofMQ
MQ
relatedMQs,itis“self-interested”fortheagent’sonly
concernsisitsownutilityincrease.
Consideramodifiedcase.Supposethatbyhavingtasktaccom-plishedagentA’sownutilityincreasesby20units.IfagentBtakesthisfactintoconsiderationwhenitmakesitsdecisionabouttaskt,agentBiscooperativewithagentAbecauseagentBisalsocon-cernedaboutagentA’soutcome(inadditiontoitsown).IfwewantagentBtoconsiderA’sutility,weneedtointroduceanotherMQdesignedtomodelB’s(revised)preferenceforAtohaveautil-ityincreasealso.ToreflecttheB’sattitudetowardA’soutcome,weintroducearelationalMQ,thepreferenceforwhichrepresentshowcooperativeagentBiswithagentAconcerningtaskt.Let
whenagentbeBtheperformsrelationaltaskMQtfortransferredagentA.fromSinceagentAtoisagentarela-B
tionalMQ,itsonlypurposeistomeasuretherelationshipbetweenagentsAandB.WhileagentBmayactuallyhaveanorganizationalgoaltoaccumulatesofthistype3,inthispaper,forsimplic-ityofpresentation,wewillassumethatagentBdoesnothaveanorganizationallevelgoaltocooperatewithagentA.Accordingly,whenmeasuringtheutilityofagentBtowardproblemsolving,wewillnotconsidertheutilityproducedbyanyrelationalMQssuchas
toagent.B,LikewisewewillwithnotagenttabulateA.WhenthenegativeagentAchangetransfers
inutilityofagentAbecausethechangeinutilityisnotrelatedtoproblemsolv-ingprogressbutisinsteadrelatedtothetransferofarelationalMQ.Thereasonforthisapproachisthatinthispaperourperformancemetricissocialwelfareasitisconventionallyused,whichisintermsofprogresstowardjointgoals.Fromthisview,theutilityproducedbyarelationalMQcanbeseenasvirtualutility.Though
producesvirtualutility,isimportantbecauseitcarriesthe
informationofhowimportanttasktisforagentA4andmakesitpossibleforagentBtoconsideragentA’soutcomewhenitmakesitsowndecisions.Actually,howB’s(virtual)utility,meaningutilitythatisnotismappedincludedintointheagentso-cialwelfarecomputation5dependsonhowcooperativeagentBiswithagentA.Supposethat20unitstaskt,representingtheutilityagentAgainedarebytransferredhavingagentwithBperformtaskt,transferredtoagentB,Figure2showsfourdifferentfunctionsformappingFunctiona,bandcarelinertofunctions:
agentB’sutility.If.(a),theutilityagentAgainedbytransferring6t),thenagent(Bdenotesiscom-pletelycooperativetoagentA;If(b),,thenagentBis
related
MQ)toagentB,agentBandagentAcannegotiateaboutwhattypeofgoal
related
MQsarefixedandagentsdonotnegotiateaboutthem,sowecandemonstratehowtherelationalMQworks.
Themappingfunctioncouldalsobeanonlinearfunction(d)thatdescribesamorecomplicatedattitudeofagentBtoagentA,i.e.,agentBbeingfullycooperativewithagentAforsomeperiodandthenbecomingself-interested.Anagentcanadjusttheutilitymap-pingfunctiontoreflectitsrelationshipwithanotheragent,whichcouldbeit’sadministrator,colleague,friend,clientorcompetitor.Byadjustingsomeparametersinthemappingfunction,moresub-tlerelationshipscouldbemanaged.Theagentcoulddifferentiateafriendlycolleaguefromanunfriendlycolleague,alsoitcoulddrawdistinctionsbetweenabestfriendandanordinaryfriend.Differentfromthegoal
Outside AgentOutside AgentOutside AgentShipping_ComputerShipping_ProductWholeSale_ComputerminTransport AgentProduce_ComputerminenablesShipping_Computertime: 6Purchase_PartsGet_HardwareGet_SoftwareenablesInstall_Softwaretime: 10Hardware Producer AgentGet_Hardwaretime: 10enablestime: 10Computer Producer AgentFigure3:AgentSociety
canbelearnedfromexperience[10].
IntheMQframework,theMQschedulerenablestheagenttoop-timizeitsscheduleandmaximizeitlocalutility.Whiletheframe-workdirectlysupportstheconceptofrelationalsandbeingmotivatedtocooperateonthatbasis,theuseoftransferenceinthispaperextendstheMQframeworktointerconnectthelocalschedulingproblemsoftwoormoreagentsinadynamicfashion(basedonthecurrentcontext).Priortothiswork,nomeaningful
transferenceortheimplicationsofit.workhadbeendonein
Inthissection,weintroduceanexampleofathree-agentsocietyandshowhowtheintegrativenegotiationmechanismworksusingtheMQframework.
Parts”taskfrom
anoutsideagentwithxunitsMQ
Hardware”taskalwaysgoestoHardwareProducerAgent
Product”withafixedrewardof3unitsMQ
taskalwaysgoestoTransportAgentwithafixedrewardof3unitsMQComputer”taskcomestoComputerPro-ducerAgentwitharewardof20unitsMQ
Hardware”and“Ship-ping
Hardware”and7units
ferredwithtask“Shipping
trans-
4.THESCENARIO
TherearethreeagentsinthissocietyasshowninFigure3:1.ComputerProducerAgent(c):receives“Produce
Computer”
task,ComputerProducerAgentneedstogenerateanexternalrequestforhardware(“Get
Computer”)throughatrans-portagent.
2.HardwareProducerAgent(h):receivestask“Get
Parts”
taskfromanoutsideagent.
3.TransportAgent(t):receivestask“Shipping
Product”
taskfromanoutsideagents.Inthisexample,everyagentcollectsthesametypeofgoal
$”.Theutilitycurvefor“MQ
hc/t”withthetaskrepresentstheutilityincreaseofCom-puterProducerAgentbyhavingthistaskaccomplished.HowitismappedintoHardwareProducerAgent’svirtualutilitydependsonHardwareProducerAgent’sattitudetowardstheutilityincreaseofComputerProducerAgentregardingtask“Get
Computer”taskcouldbefinishedearlierthanitsdead-line,ComputerProducerAgentcouldgetmorethan20unitsre-ward.Theextrautilityincreasecouldbeestimatedandreflectedby
tc/t”fortheothermorethan7unitstransferred”MQ
twoagents.SupposethefollowingtaskisreceivedbyComputerProducerAgent:
taskname:PurchaseAearlieststarttime:10deadline:70
reward:20unitsMQ
$”).
earlyfinishrewardrate(e):Iftheagentcanfinishthetaskbythetime(ft)asitpromisedinthecontract,itwillgettheextraearlyfinishreward:max(e*r*(dl-ft),r)7inadditionto
$.Itslocalutilityincreasesby20unitsaftertheaccomplish-
mentofthistask.HencethefollowtwotaskrequestsaresenttoHardwareProducerAgentandTransportAgentrespectively:
taskname:Get
Aearlieststarttime:10deadline:20
reward:3unitsMQhc/t”
earlyfinishrewardrate:e=0.01
9
taskname:Shipping
Aearlieststarttime:30deadline:40
reward:3unitsMQ
tc/t”earlyfinishrewardrate:e=0.01
Inthisexample,welookatthreedifferentattitudeswithalinerfunction:.
1.k=1,HardwareProducerAgentiscompletely-cooperativetoComputerProducerAgentregardingtask“Get
Hardware”.
3.k=0,HardwareProducerAgentisself-interestedtoCom-puterProducerAgentregardingtask“Get
Parts
PartsHardware
Hardware
$,3],[MQ
Parts
$,4]
taskname:PurchaseB
earlieststarttime:10deadline:20processtime:10MQPS:[MQ
Parts$insteadof9units,thebestMQscheduleproducedisasfollowing:
HardwareProducerAgentwillhave7unitsutilityincreaseaftertheaccomplishmentofthisschedule.
Computer”itreceives.
AsimilarreasoningprocessalsoappliestotheTransportAgent.
Theaboveexampleshowshowanagentreactsinanegotiationpro-cessdependsonitsattitudetowardstheotheragentregardingthisissue,andalsoisaffectedbytheothertasksonitagenda.Themorecooperativeanagentis,themoreitwillsacrificeitsownutilityfortheotheragent’sutilityincrease.Thisintegrativenegotiationmechanismenablestheagenttomanageandreasonaboutdifferentcooperativeattitudesitcouldhavewithanotheragentregardingacertainissue.
5.EXPERIMENT
TheexampleinSection4showsthatanagentneedstosacrificesomeofitsownutilitytobecooperativewithanotheragent.Thequestionis:Couldcooperativeagentsmakethesocialwelfare10better?Isitalwaystruethatacooperativeagentcouldimprovethesocialwelfare?Whenshouldanagentbecooperativeandhowcooperativeitshouldbe?
Toexplorethesequestions,thefollowingexperimental11workwasdonebasedonthescenariodescribedinSection4.HardwarePro-ducerAgenthasachoiceofthreedifferentattitudestowardCom-puterProducerAgent:completely-cooperative(C),half-cooperative(H),andself-interested(S),TransportAgenthasthesamethreechoices,sothereare9combinations:SS(bothagentsareself-interested),SC(HardwareProducerAgentisself-interestedwhileTransportAgentiscompletely-cooperative),SH(HardwarePro-ducerAgentisself-interestedwhileTransportAgentishalf-cooperative),HS,HC,HH,CS,CH,CC.Thedataiscollectedover48groupsofexperiments;ineachgroupofexperiments,theagentsworkonthesameincomingtasksetundertheninedifferentsituations.Thetasksineachsetforeachgroupexperimentarerandomlygener-atedwithdifferentrewards,deadlinesandearlyrewardrateswithincertainranges.
Table1showsthecomparisonofeachagent’sutilityandtheso-cialwelfareunderthesedifferentsituations.Thepercentagenum-bersarethenormalizedutilitynumbersbasedontheutilitygainedwhenagentisself-interested.Table1showsthatwhenbothHard-wareProducerAgentandTransportAgentarecompletely-cooperativetoComputerProducerAgent(CC),thesocietygainsthemostso-cialwelfare.Evenwhenbothagentareonlyhalf-cooperative(HH),thesocialwelfareisstillverygood.However,whenoneagentiscompletely-cooperative,theotheragentisself-interested(CS,SC),thesocialwelfaredoesnotimprovemuchcomparedtothecom-pletelyself-interested(SS)case.Thereasonforthelackofsig-nificantimprovementisthat,inthisexample,toaccomplishtask
“Produce
Hardware”andtask“Shipping
Computer”maystillfailbecausetheotheragentisnotcoop-erative,theutilityofComputerProducerAgentdoesnotincreaseasexpected,andtheglobalutilitydoesnotimprove.Thishap-penswhenthecompletionofataskisspreadovermorethantwoagents,theinformationfromComputerProducerAgentaboutitsutilityincreaseisonlyanestimation,itdependsnotonlyontask“Get
Computer”forTransportAgent.Inthissituation,
Producer
Percentage
SS
842
HH
301
CS
390
SH
632
CH
3.681.36
500
2.24
467
2.84
587
1.000
415
Percentage1.000
766
0.86
798
0.63
845
1.02
772
0.70
Table1:comparisonofperformance
Percentage1.000
2022
0.94
1686
0.98
1702
0.95
1905
0.94
Percentage1.0001.141.011.031.19
Object
330
HH-SS
0
CS-SS
Ho
Result
330
RejectHo
p
0.01
0.0008
0.01
0.0965
=180
0
=0FailtorejectHo
Table2:resultsfromstatisticaltests
ifHardwareProducerAgenthasnoknowledgeabouttheattitudeofTransportAgent,itmaynotbeagoodideatobecompletely-cooperativetowardsComputerProducerAgent.TheabovedataalsoshowsthattheutilityofTransportAgentdoesnotdecreasesasmuchasHardwareProducerAgentwhenitbecomescoopera-tiveorhalf-cooperative,thereasonisthefollowing.Intheexper-imentalsetup,task“Shipping
Hardware”,soitispossibleforTransportAgenttoac-ceptmoretaskswithoutlosingtoomanyhighrewardtasksfromtheoutside.
Table2showssomestatisticalresultsaboutthedifferencebe-tweenthesocialwelfareunderdifferentcooperativesituationsus-ingt-test.Forexample,thefirstlineinTable2showsthatwiththe0.01Alpha-level,wecanacceptthestatementthatthesocialwel-fareofthesystemwhenbothagentsarecooperativeisatleast20%betterthanwhenbothagentsareself-interested12.
Table3showstheexpectedutilitiesofHardwareProducerAgentandtheexpectedsocialwelfareunderthethreepossiblesituations:whenHardwareProducerAgentisself-interested,completely-co-operativeandhalf-cooperative.WhenHardwareProducerAgentchoosesoneattitude,TransportAgentmayadoptoneofthethreedifferentattitudes.Forexample,whenHardwareProducerAgentchoosestobeself-interested,theglobalsituationcouldbeSS,SC,orSH.Theutilitynumberinthetableintheexpectedvalueoftheutilitiesunderthesethreedifferentsituations.Table4showssim-ilarinformationforTransportAgent.Table3tellsusthatwhenacooperativeoperationinvolvesmorethantwoagentsandwhentheotheragents’attitudesareunknown,beingcompletely-cooperativemeanssacrificingitsownutilitysignificantlyandthusisnotagoodidea.However,itisagoodchoiceforanagenttobehalf-cooperative,sacrificinglessofitsownutilityformoreglobalutilityincrease.Thisisanexamplewherethelackofacompleteglobalviewcanbepartiallycompensatedforbyhavinganagentactinginapartiallycooperativeattituderatherthanbeingfullycooperative.FortheTransportAgentwhichdoesnotneedtosacrificetoomuchtobecompletely-cooperative,itshouldalwayschoosetobecompletely-cooperative.
Percentage
583
Completely-Cooperative
487
0.68
18311679
Percentage
1.13
UtilityofTransport
Agent
Self-Interested
803
Half-Cooperative
0.971.0
SocialWelfare
1.0
1846
1.05
Table4:theutilityofTransportAgentandthesocialwelfare
realisticdomainswheretaskscarryreal-timeconstraintsandtherearepotentiallycomplexinterrelationshipamongtasksdistributedoverdifferentagents.Otherrelatedworkincludesthecooperativenegotiationworkontaskallocation[15],wheretheagentsusethemarginalutilitygainandmarginalutilitycosttoevaluateifitworthtoacceptataskcontractinordertoincreasetheglobalutility.How-everinthiswork,theagentactsasina“completely-cooperative”modedescribedinthispaperandthereisnochoiceonhowcoop-erativeitwanttobe.
[6]RoyJ.LewickiandJosephA.LittererNegotiation1985,
RichardD.Irwin,Inc.Homewood,Illinois.
[7]EugenioOliveira,AnaPaulaRocha.Agentsadvanced
featuresfornegotiationinElectronicCommerceandVirtualOrganisationsformationprocessInC.SierraandF.Dignum,editors,BookonEuropeanperspectivesonAMEC.Springer-Verlag,June2000.
[8]TuomasSandholmandVictorLesserIssuesinAutomated
NegotiationandElectronicCommerce:ExtendingtheContractNetFramework.ProceedingsoftheFirstInternationalConferenceonMulti–AgentSystems,pp.328-335,MITPress,1995.
[9]Sandholm,T.andLesser,V.1996.AdvantagesofaLeveled
CommitmentContractingProtocol.ThirteenthNationalConferenceonArtificialIntelligence(AAAI-96),pp.126-133,Portland,OR,.
[10]SandipSenReciprocity:afoundationalprinciplefor
promotingcooperativebehavioramongself-interestedagentsinProc.oftheSecondInternationalConferenceon
MultiagentSystems,pages322–329,AAAIPress,MenloPark,CA,1996.
[11]SandipSenandEdmundH.DurfeeAFormalStudyof
DistributedMeetingSchedulingGroupDecisionandNegotiation,volume7,pages265-289,1998.
[12]Wagner,ThomasandLesser,Victor.RelatingQuantified
MotivationsforOrganizationallySituatedAgents.InIntelligentAgentsVI:AgentTheories,Architectures,andLanguages,Springer
[13]ThomasWagnerandVictorLesserEvolvingReal-Time
LocalAgentControlforLarge-ScaleMASIntelligentAgentsVIII(ProceedingsofATAL-01)editor,”J.J.MeyerandM.TambeSpringer-Verlag,Berlin,LectureNotesinArtificialIntelligence,2002
[14]Vincent,R.;Horling,B.;Lesser,V.AnAgentInfrastructure
toBuildandEvaluateMulti-AgentSystems:TheJavaAgentFrameworkandMulti-AgentSystemSimulator.InLectureNotesinArtificialIntelligence:InfrastructureforAgents,Multi-AgentSystems,andScalableMulti-AgentSystems.Volume1887,Wagner&Rana(eds.),Springer,pp.102127,2000
[15]Zhang,XiaoQin;Podorozhny,Rodion;Lesser,Victor.
Cooperative,MultiStepNegotiationOvera
Multi-DimensionalUtilityFunction.InProceedingsoftheIASTEDInternationalConference,ArtificialIntelligenceandSoftComputing(ASC2000),136-142,Banff,Canada,July,2000,IASTED/ACTAPress.
7.CONCLUSIONANDFUTUREWORK
Weintroduceanintegrativenegotiationmechanismwhichen-ablesagentsinteractoveraspectrumtomanageofnegotiationat-titudesfromself-interestedtocompletely-cooperativeinauniformreasoningframework,namelytheMQframework.Theagentnotonlycanalsochoosetobeself-interestedorcooperative,butcouldchoosehowcooperativeitwantstobe.Thisprovidestheagentacapabilitytodynamicallyadjustitsnegotiationattitudeinacom-plexagentsociety.Experimentalworkshowsitmaynotbeagoodideatoalwaysbecompletely-cooperativeinasituationinvolvinganunknownagent’sassistance;inthatcase,choosingtobehalf-cooperativemaybegoodforboththeindividualagentandalsoforthesociety.Inthefutureweplantoexploreadditionalquestionsusingthisframework,suchas:howshouldanagentchooseitne-gotiationattitudebasedonitslearningfrompastexperience?Howdoesdifferentattitudesaffectstheagent’sperformanceandtheso-cialwelfareindifferentorganizationalcontexts?andsoforth.
8.REFERENCES
[1]Axelrod,R.TheEvolutionofCooperation.BasicBooks.[2]Conry,S.E.,Kuwabara,K.,Lesser,V.R.andMeyer,R.A.
MultistageNegotiationinDistributedConstraintSatisfactionInIEEETransactionsonSystems,ManandCybernetics,Volume21,Number6,pp.1462-1477,November,1992.[3]AlyssaGlassandBarbaraGrosz2000.SociallyConscious
Decision-Making.IntheProceedingsofAgents2000
Conference,Barcelona,Spain,June3-7,2000.pp.217-224.[4]Horling,Bryan,Lesser,Victor,Vincent,Regis.Multi-Agent
SystemSimulationFramework.In16thIMACSWorldCongress2000onScientificComputation,Applied
MathematicsandSimulation,EPFL,Lausanne,Switzerland,August2000.
[5]Lander,S.andLesser,V.UnderstandingtheRoleof
NegotiationinDistributedSearchAmongHeterogeneousAgents.InProceedingsoftheInternationalJointConferenceonArtificialIntelligence,1993.
因篇幅问题不能全部显示，请点此查看更多更全内容