gents a intelligent 2: lectureparsons/courses/840-fall-2003/notes/le… · 1.1 reactivity if a prog...
TRANSCRIPT
LECTURE2:INTELLIGENTAGENTS
AnIntroductiontoMultiagentSystems
http://www.csc.liv.ac.uk/˜mjw/pubs/imas/
Lecture2AnIntroductiontoMultiagentSystems
1WhatisanAgent?
�Themainpointaboutagentsistheyareautonomous:capableofactingindependently,exhibitingcontrolovertheirinternalstate.
�Thus:anagentisacomputersystemcapableofautonomousactioninsomeenvironment.
ENVIRONMENT
SYSTEM
output input
http://www.csc.liv.ac.uk/˜mjw/pubs/imas/1
Lecture2AnIntroductiontoMultiagentSystems
�Trivial(non-interesting)agents:
–thermostat;
–UNIXdaemon(e.g.,biff).
�Anintelligentagentisacomputersystemcapableofflexibleautonomousactioninsomeenvironment.
Byflexible,wemean:
–reactive;
–pro-active;
–social.
http://www.csc.liv.ac.uk/˜mjw/pubs/imas/2
Lecture2AnIntroductiontoMultiagentSystems
1.1Reactivity
�Ifaprogram’senvironmentisguaranteedtobefixed,theprogramneedneverworryaboutitsownsuccessorfailure—programjustexecutesblindly.
Exampleoffixedenvironment:compiler.
�Therealworldisnotlikethat:thingschange,informationisincomplete.Many(most?)interestingenvironmentsaredynamic.
�Softwareishardtobuildfordynamicdomains:programmusttakeintoaccountpossibilityoffailure—askitselfwhetheritisworthexecuting!
�Areactivesystemisonethatmaintainsanongoinginteractionwithitsenvironment,andrespondstochangesthatoccurinit(intimefortheresponsetobeuseful).
http://www.csc.liv.ac.uk/˜mjw/pubs/imas/3
Lecture2AnIntroductiontoMultiagentSystems
1.2Proactiveness
�Reactingtoanenvironmentiseasy(e.g.,stimulus�responserules).
�Butwegenerallywantagentstodothingsforus.
�Hencegoaldirectedbehaviour.
�Pro-activeness=generatingandattemptingtoachievegoals;notdrivensolelybyevents;takingtheinitiative.
�Recognisingopportunities.
http://www.csc.liv.ac.uk/˜mjw/pubs/imas/4
Lecture2AnIntroductiontoMultiagentSystems
1.3SocialAbility
�Therealworldisamulti-agentenvironment:wecannotgoaroundattemptingtoachievegoalswithouttakingothersintoaccount.
�Somegoalscanonlybeachievedwiththecooperationofothers.
�Similarlyformanycomputerenvironments:witnesstheINTERNET.
�Socialabilityinagentsistheabilitytointeractwithotheragents(andpossiblyhumans)viasomekindofagent-communicationlanguage,andperhapscooperatewithothers.
http://www.csc.liv.ac.uk/˜mjw/pubs/imas/5
Lecture2AnIntroductiontoMultiagentSystems
2OtherProperties
�Otherproperties,sometimesdiscussedinthecontextofagency:
–mobility:theabilityofanagenttomovearoundanelectronicnetwork;
–veracity:anagentwillnotknowinglycommunicatefalseinformation;
–benevolence:agentsdonothaveconflictinggoals,andthateveryagentwillthereforealwaystrytodowhatisaskedofit;
–rationality:agentwillactinordertoachieveitsgoals,andwillnotactinsuchawayastopreventitsgoalsbeingachieved—atleastinsofarasitsbeliefspermit;
–learning/adaption:agentsimproveperformanceovertime.
http://www.csc.liv.ac.uk/˜mjw/pubs/imas/6
Lecture2AnIntroductiontoMultiagentSystems
2.1AgentsandObjects
�Areagentsjustobjectsbyanothername?
�Object:
–encapsulatessomestate;
–communicatesviamessagepassing;
–hasmethods,correspondingtooperationsthatmaybeperformedonthisstate.
http://www.csc.liv.ac.uk/˜mjw/pubs/imas/7
Lecture2AnIntroductiontoMultiagentSystems
�Maindifferences:
–agentsareautonomous:agentsembodystrongernotionofautonomythanobjects,andinparticular,theydecideforthemselveswhetherornottoperformanactiononrequestfromanotheragent;
–agentsaresmart:capableofflexible(reactive,pro-active,social)behavior,andthestandardobjectmodelhasnothingtosayaboutsuchtypesofbehavior;
–agentsareactive:amulti-agentsystemisinherentlymulti-threaded,inthateachagentisassumedtohaveatleastonethreadofactivecontrol.
http://www.csc.liv.ac.uk/˜mjw/pubs/imas/8
Lecture2AnIntroductiontoMultiagentSystems
Objectsdoitforfree...
�agentsdoitbecausetheywantto;
�agentsdoitformoney.
http://www.csc.liv.ac.uk/˜mjw/pubs/imas/9
Lecture2AnIntroductiontoMultiagentSystems
2.2AgentsandExpertSystems
�Aren’tagentsjustexpertsystemsbyanothername?
�Expertsystemstypicallydisembodied‘expertise’aboutsome(abstract)domainofdiscourse(e.g.,blooddiseases).
�Example:MYCINknowsaboutblooddiseasesinhumans.
Ithasawealthofknowledgeaboutblooddiseases,intheformofrules.
AdoctorcanobtainexpertadviceaboutblooddiseasesbygivingMYCINfacts,answeringquestions,andposingqueries.
http://www.csc.liv.ac.uk/˜mjw/pubs/imas/10
Lecture2AnIntroductiontoMultiagentSystems
�Maindifferences:
–agentssituatedinanenvironment:MYCINisnotawareoftheworld—onlyinformationobtainedisbyaskingtheuserquestions.
–agentsact:MYCINdoesnotoperateonpatients.
�Somereal-time(typicallyprocesscontrol)expertsystemsareagents.
http://www.csc.liv.ac.uk/˜mjw/pubs/imas/11
Lecture2AnIntroductiontoMultiagentSystems
2.3IntelligentAgentsandAI
�Aren’tagentsjusttheAIproject?
Isn’tbuildinganagentwhatAIisallabout?
�AIaimstobuildsystemsthatcan(ultimately)understandnaturallanguage,recogniseandunderstandscenes,usecommonsense,thinkcreatively,etc—allofwhichareveryhard.
�So,don’tweneedtosolveallofAItobuildanagent...?
http://www.csc.liv.ac.uk/˜mjw/pubs/imas/12
Lecture2AnIntroductiontoMultiagentSystems
�Whenbuildinganagent,wesimplywantasystemthatcanchoosetherightactiontoperform,typicallyinalimiteddomain.
�WedonothavetosolvealltheproblemsofAItobuildausefulagent:
alittleintelligencegoesalongway!
�OrenEtzioni,speakingaboutthecommercialexperienceofNETBOT,Inc:
Wemadeouragentsdumberanddumberanddumber...untilfinallytheymademoney.
http://www.csc.liv.ac.uk/˜mjw/pubs/imas/13
Lecture2AnIntroductiontoMultiagentSystems
3Environments
�Accessiblevsinaccessible.
Anaccessibleenvironmentisoneinwhichtheagentcanobtaincomplete,accurate,up-to-dateinformationabouttheenvironment’sstate.
Mostmoderatelycomplexenvironments(including,forexample,theeverydayphysicalworldandtheInternet)areinaccessible.
Themoreaccessibleanenvironmentis,thesimpleritistobuildagentstooperateinit.
http://www.csc.liv.ac.uk/˜mjw/pubs/imas/14
Lecture2AnIntroductiontoMultiagentSystems
�Deterministicvsnon-deterministic.
Aswehavealreadymentioned,adeterministicenvironmentisoneinwhichanyactionhasasingleguaranteedeffect—thereisnouncertaintyaboutthestatethatwillresultfromperforminganaction.
Thephysicalworldcantoallintentsandpurposesberegardedasnon-deterministic.
Non-deterministicenvironmentspresentgreaterproblemsfortheagentdesigner.
http://www.csc.liv.ac.uk/˜mjw/pubs/imas/15
Lecture2AnIntroductiontoMultiagentSystems
�Episodicvsnon-episodic.
Inanepisodicenvironment,theperformanceofanagentisdependentonanumberofdiscreteepisodes,withnolinkbetweentheperformanceofanagentindifferentscenarios.
Episodicenvironmentsaresimplerfromtheagentdeveloper’sperspectivebecausetheagentcandecidewhatactiontoperformbasedonlyonthecurrentepisode—itneednotreasonabouttheinteractionsbetweenthisandfutureepisodes.
http://www.csc.liv.ac.uk/˜mjw/pubs/imas/16
Lecture2AnIntroductiontoMultiagentSystems
�Staticvsdynamic.
Astaticenvironmentisonethatcanbeassumedtoremainunchangedexceptbytheperformanceofactionsbytheagent.
Adynamicenvironmentisonethathasotherprocessesoperatingonit,andwhichhencechangesinwaysbeyondtheagent’scontrol.
Thephysicalworldisahighlydynamicenvironment.
http://www.csc.liv.ac.uk/˜mjw/pubs/imas/17
Lecture2AnIntroductiontoMultiagentSystems
�Discretevscontinuous.
Anenvironmentisdiscreteifthereareafixed,finitenumberofactionsandperceptsinit.RussellandNorviggiveachessgameasanexampleofadiscreteenvironment,andtaxidrivingasanexampleofacontinuousone.
http://www.csc.liv.ac.uk/˜mjw/pubs/imas/18
Lecture2AnIntroductiontoMultiagentSystems
4AgentsasIntentionalSystems
�Whenexplaininghumanactivity,itisoftenusefultomakestatementssuchasthefollowing:
Janinetookherumbrellabecauseshebelieveditwasgoingtorain.
MichaelworkedhardbecausehewantedtopossessaPhD.
�Thesestatementsmakeuseofafolkpsychology,bywhichhumanbehaviourispredictedandexplainedthroughtheattributionofattitudes,suchasbelievingandwanting(asintheaboveexamples),hoping,fearing,andsoon.
�Theattitudesemployedinsuchfolkpsychologicaldescriptionsarecalledtheintentionalnotions.
http://www.csc.liv.ac.uk/˜mjw/pubs/imas/19
Lecture2AnIntroductiontoMultiagentSystems
�ThephilosopherDanielDennettcoinedthetermintentionalsystemtodescribeentities‘whosebehaviourcanbepredictedbythemethodofattributingbelief,desiresandrationalacumen’.
�Dennettidentifiesdifferent‘grades’ofintentionalsystem:
‘Afirst-orderintentionalsystemhasbeliefsanddesires(etc.)butnobeliefsanddesiresaboutbeliefsanddesires....Asecond-orderintentionalsystemismoresophisticated;ithasbeliefsanddesires(andnodoubtotherintentionalstates)aboutbeliefsanddesires(andotherintentionalstates)—boththoseofothersanditsown’.
�Isitlegitimateorusefultoattributebeliefs,desires,andsoon,tocomputersystems?
http://www.csc.liv.ac.uk/˜mjw/pubs/imas/20
Lecture2AnIntroductiontoMultiagentSystems
�McCarthyarguedthatthereareoccasionswhentheintentionalstanceisappropriate:
‘Toascribebeliefs,freewill,intentions,consciousness,abilities,orwantstoa
machineislegitimatewhensuchanascriptionexpressesthesameinformationaboutthemachinethatitexpressesaboutaperson.Itisusefulwhenthe
ascriptionhelpsusunderstandthestructureofthemachine,itspastorfuture
behaviour,orhowtorepairorimproveit.Itisperhapsneverlogicallyrequiredevenforhumans,butexpressingreasonablybrieflywhatisactuallyknownabout
thestateofthemachineinaparticularsituationmayrequirementalqualitiesorqualitiesisomorphictothem.Theoriesofbelief,knowledgeandwantingcanbe
constructedformachinesinasimplersettingthanforhumans,andlaterappliedtohumans.Ascriptionofmentalqualitiesismoststraightforwardformachinesof
knownstructuresuchasthermostatsandcomputeroperatingsystems,butismostusefulwhenappliedtoentitieswhosestructureisincompletelyknown’.
http://www.csc.liv.ac.uk/˜mjw/pubs/imas/21
Lecture2AnIntroductiontoMultiagentSystems
�Whatobjectscanbedescribedbytheintentionalstance?
�Asitturnsout,moreorlessanythingcan...consideralightswitch:
‘Itisperfectlycoherenttotreatalightswitchasa(verycooperative)agentwiththecapabilityoftransmittingcurrentatwill,whoinvariablytransmitscurrentwhenitbelievesthatwewantittransmittedandnototherwise;flickingtheswitchissimplyourwayofcommunicatingourdesires’.(YoavShoham)
�Butmostadultswouldfindsuchadescriptionabsurd!
Whyisthis?
http://www.csc.liv.ac.uk/˜mjw/pubs/imas/22
Lecture2AnIntroductiontoMultiagentSystems
�Theanswerseemstobethatwhiletheintentionalstancedescriptionisconsistent,
...itdoesnotbuyusanything,sinceweessentiallyunderstandthemechanismsufficientlytohaveasimpler,mechanisticdescriptionofitsbehaviour.(YoavShoham)
�Putcrudely,themoreweknowaboutasystem,thelessweneedtorelyonanimistic,intentionalexplanationsofitsbehaviour.
�Butwithverycomplexsystems,amechanistic,explanationofitsbehaviourmaynotbepracticable.
�Ascomputersystemsbecomeevermorecomplex,weneedmorepowerfulabstractionsandmetaphorstoexplaintheiroperation—lowlevelexplanationsbecomeimpractical.
Theintentionalstanceissuchanabstraction.
http://www.csc.liv.ac.uk/˜mjw/pubs/imas/23
Lecture2AnIntroductiontoMultiagentSystems
�Theintentionalnotionsarethusabstractiontools,whichprovideuswithaconvenientandfamiliarwayofdescribing,explaining,andpredictingthebehaviourofcomplexsystems.
�Remember:mostimportantdevelopmentsincomputingarebasedonnewabstractions:
–proceduralabstraction;
–abstractdatatypes;
–objects.
Agents,andagentsasintentionalsystems,representafurther,andincreasinglypowerfulabstraction.
�Soagenttheoristsstartfromthe(strong)viewofagentsasintentionalsystems:onewhosesimplestconsistentdescriptionrequirestheintentionalstance.
http://www.csc.liv.ac.uk/˜mjw/pubs/imas/24
Lecture2AnIntroductiontoMultiagentSystems
�Thisintentionalstanceisanabstractiontool—aconvenientwayoftalkingaboutcomplexsystems,whichallowsustopredictandexplaintheirbehaviourwithouthavingtounderstandhowthemechanismactuallyworks.
�Now,muchofcomputerscienceisconcernedwithlookingforabstractionmechanisms(witnessproceduralabstraction,ADTs,objects,...)
Sowhynotusetheintentionalstanceasanabstractiontoolincomputing—toexplain,understand,and,crucially,programcomputersystems?
�Thisisanimportantargumentinfavourofagents.
http://www.csc.liv.ac.uk/˜mjw/pubs/imas/25
Lecture2AnIntroductiontoMultiagentSystems
�Other3pointsinfavourofthisidea:
CharacterisingAgents
�Itprovidesuswithafamiliar,non-technicalwayofunderstanding&explaingagents.
NestedRepresentations
�Itgivesusthepotentialtospecifysystemsthatincluderepresentationsofothersystems.
Itiswidelyacceptedthatsuchnestedrepresentationsareessentialforagentsthatmustcooperatewithotheragents.
http://www.csc.liv.ac.uk/˜mjw/pubs/imas/26
Lecture2AnIntroductiontoMultiagentSystems
Post-DeclarativeSystems
�Thisviewofagentsleadstoakindofpost-declarativeprogramming:
–inproceduralprogramming,wesayexactlywhatasystemshoulddo;
–indeclarativeprogramming,westatesomethingthatwewanttoachieve,givethesystemgeneralinfoabouttherelationshipsbetweenobjects,andletabuilt-incontrolmechanism(e.g.,goal-directedtheoremproving)figureoutwhattodo;
–withagents,wegiveaveryabstractspecificationofthesystem,andletthecontrolmechanismfigureoutwhattodo,knowingthatitwillactinaccordancewithsomebuilt-intheoryofagency(e.g.,thewell-knownCohen-Levesquemodelofintention).
http://www.csc.liv.ac.uk/˜mjw/pubs/imas/27
Lecture2AnIntroductiontoMultiagentSystems
Anaside...
�Wefindthatresearchersfromamoremainstreamcomputingdisciplinehaveadoptedasimilarsetofideas...
�Indistributedsystemstheory,logicsofknowledgeareusedinthedevelopmentofknowledgebasedprotocols.
�Therationaleisthatwhenconstructingprotocols,oneoftenencountersreasoningsuchasthefollowing:
IFprocessiknowsprocessjhasreceivedmessagem�
THENprocessishouldsendprocessjthemessagem�.
�InDStheory,knowledgeisgrounded—givenapreciseinterpretationintermsofthestatesofaprocess;returntothislater...we’llexaminethispointindetaillater.
http://www.csc.liv.ac.uk/˜mjw/pubs/imas/28
Lecture2AnIntroductiontoMultiagentSystems
5AbstractArchitecturesforAgents
�AssumetheenvironmentmaybeinanyofafinitesetEofdiscrete,instantaneousstates:
E��e�e��������
�Agentsareassumedtohavearepertoireofpossibleactionsavailabletothem,whichtransformthestateoftheenvironment.
Ac����� �������
�Arun,r,ofanagentinanenvironmentisasequenceofinterleavedenvironmentstatesandactions:
re�� �e�
�� �e�
�� �e��� �����u�� �eu
http://www.csc.liv.ac.uk/˜mjw/pubs/imas/29
Lecture2AnIntroductiontoMultiagentSystems
�Let:
–�bethesetofallsuchpossiblefinitesequences(overEandAc);
–�
Acbethesubsetofthesethatendwithanaction;and
–�
Ebethesubsetofthesethatendwithanenvironmentstate.
http://www.csc.liv.ac.uk/˜mjw/pubs/imas/30
Lecture2AnIntroductiontoMultiagentSystems
StateTransformerFunctions
�Astatetransformerfunctionrepresentsbehaviouroftheenvironment:
��
Ac
���E�
�Notethatenvironmentsare...
–historydependent.
–non-deterministic.
�If��r���,thentherearenopossiblesuccessorstatestor.Inthiscase,wesaythatthesystemhasendeditsrun.
�Formally,wesayanenvironmentEnvisatripleEnv��E�e���
where:Eisasetofenvironmentstates,e�Eistheinitialstate;and�isastatetransformerfunction.
http://www.csc.liv.ac.uk/˜mjw/pubs/imas/31
Lecture2AnIntroductiontoMultiagentSystems
Agents
�Agentisafunctionwhichmapsrunstoactions:
Ag�
E�Ac
Anagentmakesadecisionaboutwhatactiontoperformbasedonthehistoryofthesystemthatithaswitnessedtodate.Let��
bethesetofallagents.
http://www.csc.liv.ac.uk/˜mjw/pubs/imas/32
Lecture2AnIntroductiontoMultiagentSystems
Systems
�Asystemisapaircontaininganagentandanenvironment.
�Anysystemwillhaveassociatedwithitasetofpossibleruns;wedenotethesetofrunsofagentAginenvironmentEnvby
��Ag�Env�.
�(Weassume��Ag�Env�containsonlyterminatedruns.)
http://www.csc.liv.ac.uk/˜mjw/pubs/imas/33
Lecture2AnIntroductiontoMultiagentSystems
�Formally,asequence
�e� ��e�� ���e�������
representsarunofanagentAginenvironmentEnv��E�e���if:
1.eistheinitialstateofEnv
2.��Ag�e�;and
3.foru��,
eu����e� ������ �u����where
�u�Ag��e� �������eu��
http://www.csc.liv.ac.uk/˜mjw/pubs/imas/34
Lecture2AnIntroductiontoMultiagentSystems
6PurelyReactiveAgents
�Someagentsdecidewhattodowithoutreferencetotheirhistory—theybasetheirdecisionmakingentirelyonthepresent,withnoreferenceatalltothepast.
�Wecallsuchagentspurelyreactive:
actionE�Ac
�Athermostatisapurelyreactiveagent.
action�e���
��
��
��
offife=temperatureOKonotherwise.
http://www.csc.liv.ac.uk/˜mjw/pubs/imas/35
Lecture2AnIntroductiontoMultiagentSystems
7Perception
�Nowintroduceperceptionsystem:
ENVIRONMENT
action
AGENT
see
http://www.csc.liv.ac.uk/˜mjw/pubs/imas/36
Lecture2AnIntroductiontoMultiagentSystems
�Theseefunctionistheagent’sabilitytoobserveitsenvironment,whereastheactionfunctionrepresentstheagent’sdecisionmakingprocess.
�Outputoftheseefunctionisapercept:
seeE�Per
whichmapsenvironmentstatestopercepts,andactionisnowafunction
actionPer
�
�A
whichmapssequencesofperceptstoactions.
http://www.csc.liv.ac.uk/˜mjw/pubs/imas/37
Lecture2AnIntroductiontoMultiagentSystems
8AgentswithState
�Wenowconsideragentsthatmaintainstate:
action see
nextstate
AGENT
ENVIRONMENT
http://www.csc.liv.ac.uk/˜mjw/pubs/imas/38
Lecture2AnIntroductiontoMultiagentSystems
�Theseagentshavesomeinternaldatastructure,whichistypicallyusedtorecordinformationabouttheenvironmentstateandhistory.LetIbethesetofallinternalstatesoftheagent.
�Theperceptionfunctionseeforastate-basedagentisunchanged:
seeE�Per
Theaction-selectionfunctionactionisnowdefinedasamapping
actionI�Ac
frominternalstatestoactions.Anadditionalfunctionnextisintroduced,whichmapsaninternalstateandpercepttoaninternalstate:
nextI�Per�I
http://www.csc.liv.ac.uk/˜mjw/pubs/imas/39
Lecture2AnIntroductiontoMultiagentSystems
8.1Agentcontrolloop
1.Agentstartsinsomeinitialinternalstatei.
2.Observesitsenvironmentstatee,andgeneratesaperceptsee�e�.
3.Internalstateoftheagentisthenupdatedvianextfunction,becomingnext�i�see�e��.
4.Theactionselectedbytheagentisaction�next�i�see�e���.
Thisactionisthenperformed.
5.Goto(2).
http://www.csc.liv.ac.uk/˜mjw/pubs/imas/40
Lecture2AnIntroductiontoMultiagentSystems
9TasksforAgents
�Webuildagentsinordertocarryouttasksforus.
�Thetaskmustbespecifiedbyus...
�Butwewanttotellagentswhattodowithouttellingthemhowtodoit.
http://www.csc.liv.ac.uk/˜mjw/pubs/imas/41
Lecture2AnIntroductiontoMultiagentSystems
9.1UtilitiesFunctionsoverStates
�Onepossibility:associateutilitieswithindividualstates—thetaskoftheagentisthentobringaboutstatesthatmaximiseutility.
�Ataskspecificationisafunction
uE�IR
whichassociatedarealnumberwitheveryenvironmentstate.
http://www.csc.liv.ac.uk/˜mjw/pubs/imas/42
Lecture2AnIntroductiontoMultiagentSystems
�Butwhatisthevalueofarun...
–minimumutilityofstateonrun?
–maximumutilityofstateonrun?
–sumofutilitiesofstatesonrun?
–average?
�Disadvantage:difficulttospecifyalongtermviewwhenassigningutilitiestoindividualstates.
(Onepossibility:adiscountforstateslateron.)
http://www.csc.liv.ac.uk/˜mjw/pubs/imas/43
Lecture2AnIntroductiontoMultiagentSystems
9.2UtilitiesoverRuns
�Anotherpossibility:assignsautilitynottoindividualstates,buttorunsthemselves:
u��IR
�Suchanapproachtakesaninherentlylongtermview.
�Othervariations:incorporateprobabilitiesofdifferentstatesemerging.
�Difficultieswithutility-basedapproaches:
–wheredothenumberscomefrom?
–wedon’tthinkintermsofutilities!
–hardtoformulatetasksintheseterms.
http://www.csc.liv.ac.uk/˜mjw/pubs/imas/44
Lecture2AnIntroductiontoMultiagentSystems
UtilityintheTileworld
�Simulatedtwodimensionalgridenvironmentonwhichthereareagents,tiles,obstacles,andholes.
�Anagentcanmoveinfourdirections,up,down,left,orright,andifitislocatednexttoatile,itcanpushit.
�Holeshavetobefilledupwithtilesbytheagent.Anagentscorespointsbyfillingholeswithtiles,withtheaimbeingtofillasmanyholesaspossible.
�TILEWORLDchangeswiththerandomappearanceanddisappearanceofholes.
�Utilityfunctiondefinedasfollows:
u�r�
��numberofholesfilledinr
numberofholesthatappearedinr
http://www.csc.liv.ac.uk/˜mjw/pubs/imas/45
Lecture2AnIntroductiontoMultiagentSystems
9.3ExpectedUtility&OptimalAgents
�WriteP�r�Ag�Env�todenoteprobabilitythatrunroccurswhenagentAgisplacedinenvironmentEnv.
Note:�
r���Ag�Env�P�r�Ag�Env����
�ThenoptimalagentAgoptinanenvironmentEnvistheonethatmaximizesexpectedutility:
Agopt����� Ag� ��
r���Ag�Env�
u�r�P�r�Ag�Env� �(1)
http://www.csc.liv.ac.uk/˜mjw/pubs/imas/46
Lecture2AnIntroductiontoMultiagentSystems
9.4BoundedOptimalAgents
�Someagentscannotbeimplementedonsomecomputers
(AfunctionAg�E�Acmayneedmorethanavailablememory
toimplement.)
�Write��mtodenotetheagentsthatcanbeimplementedonmachine(computer)m:
��m��Ag�Ag���andAgcanbeimplementedonm��
�Wecanreplaceequation(1)withthefollowing,whichdefinestheboundedoptimalagentAgopt:
Agopt����� Ag� �m�
r���Ag�Env�
u�r�P�r�Ag�Env� �(2)
http://www.csc.liv.ac.uk/˜mjw/pubs/imas/47
Lecture2AnIntroductiontoMultiagentSystems
9.5PredicateTaskSpecifications
�Aspecialcaseofassigningutilitiestohistoriesistoassign0(false)or1(true)toarun.
�Ifarunisassigned1,thentheagentsucceedsonthatrun,otherwiseitfails.
�Callthesepredicatetaskspecifications.
�Denotepredicatetaskspecificationby�.
Thus��������.
http://www.csc.liv.ac.uk/˜mjw/pubs/imas/48
Lecture2AnIntroductiontoMultiagentSystems
9.6TaskEnvironments
�Ataskenvironmentisapair�Env���,whereEnvisanenvironment,and
��������
isapredicateoverruns.
Let��bethesetofalltaskenvironments.
�Ataskenvironmentspecifies:
–thepropertiesofthesystemtheagentwillinhabit;
–thecriteriabywhichanagentwillbejudgedtohaveeitherfailedorsucceeded.
http://www.csc.liv.ac.uk/˜mjw/pubs/imas/49
Lecture2AnIntroductiontoMultiagentSystems
�Write���Ag�Env�todenotesetofallrunsoftheagentAginenvironmentEnvthatsatisfy�:
���Ag�Env���r�r���Ag�Env�and��r�����
�WethensaythatanagentAgsucceedsintaskenvironment
�Env���if
���Ag�Env����Ag�Env�
http://www.csc.liv.ac.uk/˜mjw/pubs/imas/50
Lecture2AnIntroductiontoMultiagentSystems
TheProbabilityofSuccess
�LetP�r�Ag�Env�denoteprobabilitythatrunroccursifagentAgisplacedinenvironmentEnv.
�ThentheprobabilityP���Ag�Env�that�issatisfiedbyAginEnvwouldthensimplybe:
P���Ag�Env���
r����Ag�Env�P�r�Ag�Env�
http://www.csc.liv.ac.uk/˜mjw/pubs/imas/51
Lecture2AnIntroductiontoMultiagentSystems
Achievement&MaintenanceTasks
�Twomostcommontypesoftasksareachievementtasksandmaintenancetasks:
1.AchievementtasksArethoseoftheform“achievestateofaffairs�”.
2.MaintenancetasksArethoseoftheform“maintainstateofaffairs�”.
http://www.csc.liv.ac.uk/˜mjw/pubs/imas/52
Lecture2AnIntroductiontoMultiagentSystems
�AnachievementtaskisspecifiedbyasetGof“good”or“goal”states:G�E.
Theagentsucceedsifitisguaranteedtobringaboutatleastoneofthesestates(wedonotcarewhichone—theyareallconsideredequallygood).
�AmaintenancegoalisspecifiedbyasetBof“bad”states:B�E.
TheagentsucceedsinaparticularenvironmentifitmanagestoavoidallstatesinB—ifitneverperformsactionswhichresultinanystateinBoccurring.
http://www.csc.liv.ac.uk/˜mjw/pubs/imas/53
Lecture2AnIntroductiontoMultiagentSystems
10AgentSynthesis
�Agentsynthesisisautomaticprogramming:goalistohaveaprogramthatwilltakeataskenvironment,andfromthistaskenvironmentautomaticallygenerateanagentthatsucceedsinthisenvironment:
syn� ����������� �
(Thinkof�asbeinglikenullinJAVA.
�Synthesisalgorithmis:
–soundif,wheneveritreturnsanagent,thenthisagentsucceedsinthetaskenvironmentthatispassedasinput;and
–completeifitisguaranteedtoreturnanagentwheneverthereexistsanagentthatwillsucceedinthetaskenvironmentgivenasinput.
http://www.csc.liv.ac.uk/˜mjw/pubs/imas/54
Lecture2AnIntroductiontoMultiagentSystems
�Synthesisalgorithmsynissoundifitsatisfiesthefollowingcondition:
syn��Env�����Agimplies��Ag�Env�����Ag�Env� �
andcompleteif:
�Ag���s.t.��Ag�Env�����Ag�Env�impliessyn��Env����� ���
http://www.csc.liv.ac.uk/˜mjw/pubs/imas/55