Lectures AIProgramming:Introduction IntroductiontoRoboRescue 3 AgentsandAgentsArchitecture 4 Communication 5 MultiagentDecisionMaking 6 CooperationAndCoordination1 7 CooperationAndCoordination2 8 MachineLearning 9 KnowledgeRepresentation 10 PuttingItAllTogether 1 2 TDDD10AIProgramming AgentsandAgentArchitectures CyrilleBerger 2/83 Lecturegoals Lecturecontent Acquireknowledgeonwhatisanagent. Acquireknowledgeonwhatarethe differentagentarchitectureandhow theytakedecision Agents AnOverviewofDecisionMaking AgentArchitectures DeliberativeArchitecture ReactiveArchitecture StateMachines HybridArchitecture Summary 3/83 4/83 Whatisanagent? Agentsareautonomous:capableofacting independentlyandexhibitingcontrolover theirinternalstate Anagentisa(computer)systemthatis situatedinsomeenvironmentandthatis capableofautonomousactioninthis environmentinordertomeetitsdelegated objectives. Agents 6 Whatisanagent? Whatisanagent? Shouldanagentbeabletolearn? Shouldanagentbeintelligent? 7 8 IntelligentAgentProperties SocialAbility Reactivity Cooperationisworkingtogetherasateamtoachievea sharedgoal. Intelligentagentsareabletoperceivetheirenvironment,and respondinatimelyfashiontochangesthatoccurinitinorderto meetitsdelegatedobjectives. Oftenpromptedeitherbythefactthatnooneagentcanachievethegoal alone,orthatcooperationwillobtainabetterresult(e.g.,getresult faster). Proactivity Coordinationismanagingtheinterdependencies betweenactivities. Negotiationistheabilitytoreachagreementson mattersofcommoninterest. Intelligentagentsareabletoexhibitgoal-directedbehavior bytakingtheinitiativeinordertomeetitsdelegated SocialAbility Intelligentagentsarecapableofinteracting(cooperating, coordinatingandnegotiating)withotheragents(andpossible humans)inordertomeetitsdelegatedobjectives. Typicallyinvolvesofferandcounter-offer,withcompromisesmadeby participants. 9 AgentsasIntentionalSystems ThephilosopherDanielDennettcoinedthetermintentionalsystemto describeentities“whosebehaviourcanbepredictedbythemethodof attributingbelief,desiresandrationalacumen”. Isitlegitimateorusefultoattributebeliefs,desires,andsoon,tocomputer systems? Withverycomplexsystemsamechanisticexplanationofitsbehaviourmay notbepracticaloravailable.But,themoreweknowaboutasystem,theless weneedtorelyonanimistic,intentionalexplanationsofitsbehaviour. Ascomputersystemsbecomeevermorecomplex,weneedmorepowerful abstractionsandmetaphorstoexplaintheiroperation—lowlevel explanationsbecomeimpractical.Theintentionalstanceissuchan abstraction,whichprovideuswithaconvenientandfamiliarwayof describing,explaining,andpredictingthebehaviourofcomplexsystems. 10 ObjectOrientedProgrammingvsMulti-AgentSystems Object-OrientedProgramming Objectsarepassive,i.e. anobjecthasnocontrol overmethodinvocation Objectsaredesignedfor acommongoal Typicallyintegratedinto asinglethread 11 Multi-AgentSystems Agentsareautonomous, i.e.pro-active Agentscanhave diverginggoals,e.g. comingfromdifferent organizations Agentshaveownthread ofcontrol 12 Agent-OrientedProgramming Structuralunit Relationto previouslevel Machine Language Structured Programming ObjectOriented Programming Agent-Oriented Programming Program Subroutine Object Agent Boundunitof program Object+ independent Subroutine+ persistentlocal threadof state control+ initiative AgentOrientedProgramming (YoavShoham) Basedontheagentdefinition:“Anagentisanentity whosestateisviewedasconsistingofmental componentssuchasbeliefs,capabilities,choices,and commitments.” Thementalconstructswillappearintheprogramming languageitself. Thesemanticswillberelatedtothesemanticsofthe mentalconstructs. Acomputationwillconsistofagentsperforming speech-actsoneachother. 13 AGENT0(1/2) 14 AGENT0(2/2) Eachcommitmentrulecontains AGENT0isimplementedinLISP EachagentinAGENT0has4components: amessagecondition amentalcondition anaction asetofcapabilities(thingstheagentcando) asetofinitialbeliefs asetofinitialcommitments(thingstheagentwilldo) asetofcommitmentrules Oneach‘agentcycle’... Themessageconditionismatchedagainstthemessagesthe agenthasreceived Thementalconditionismatchedagainstthebeliefsof theagent Iftherulefires,thentheagentbecomescommittedtothe action(theactiongetsaddedtotheagent’scommitmentset) Thekeycomponent,whichdetermineshow theagentacts,isthecommitmentruleset 15 16 ExampleAGENT0Rule ExampleAGENT0Rule COMMIT( (agent,REQUEST,DO(time,action) ),;;;msgcondition (B,[now,Friendagent]AND CAN(self,action)AND NOT[time,CMT(self,anyaction)] ),;;;mentalcondition self, DO(time,action) ) Onerulecouldbe: ifIreceiveamessagefromagentwhich requestsmetodoactionattime,andI believethat: agentiscurrentlyafriend Icandotheaction Attime,Iamnotcommittedtodoinganyother action actionattime thencommittodoing 17 18 Individualdecisionmaking Explicitdecisionmaking AnOverviewofDecisionMaking Decisiontrees Rules Automata Singleagenttaskspecificationlanguages Decisiontheoreticdecisionmaking MarkovDecisionProcesses(MDP) PartiallyObservableMarkovDecisionProcesses(POMDP) Declarative(logic-based)decisionmaking TheoremProving Planning Constraintsatisfaction 20 Multiagentdecisionmaking Explicit Mutualmodeling Norms OrganizationsandRoles Multiagenttaskspecificationlanguages AgentArchitectures Decisiontheoretic DecentralizedPOMDPs(Dec-POMDP) Gametheoretic Auctions Declarative Multiagentplanning Distributedconstraintsatisfaction 21 AgentArchitectures AgentArchitectures “[A]particularmethodologyforbuilding [agents].Itspecifieshow…theagentcan bedecomposedintotheconstructionofa setofcomponentmodulesandhowthese modulesshouldbemadetointeract.”(P. Maes1991) Threetypes: deliberative(symbolic/logical) reactive hybrid. 23 24 DeliberativeArchitecture Wedefineadeliberative(orreasoning)agentor agentarchitecturetobeonethat: containsanexplicitlyrepresented,symbolicmodeloftheworldand makesdecisionsviasymbolicreasoning. DeliberativeArchitecture Viewsagentsasknowledge-basedsystems. Wecansaythatadeliberativeagentmakesan actioninthreesteps: Sense Plan Act 26 PracticalReasoning PracticalReasoning “Practicalreasoningisamatterof weighingconflictingconsiderationsfor andagainstcompetingoptions,wherethe relevantconsiderationsareprovidedby whattheagentdesires/values/cares aboutandwhattheagentbelieves.” Bratman Humanpracticalreasoningconsistsof twoactivities: deliberation-decidingwhatstateofaffairswewant toachieve; means-endsreasoning-decidinghowtoachieve thesestatesofaffairs. Theoutputofdeliberationisintentions. 27 28 Intentions(1/4) Intentions(2/4) Agentsneedtodeterminewaysof achievingintentions. Agentsbelievetheirintentionsare possible. IfIhaveanintentiontoφyouwouldexpectmeto devoteresourcestodecidinghowtobringabout Anagentbelievesthereisatleastsomewaythatthe intentionscouldbebroughtabout. Intentionsprovideafilterforadopting otherintentions,whichmustnotconflict. Agentsdonotbelievetheywillnot bringabouttheirintentions. IfIhaveanintentiontoφ,youwouldnotexpect metoadoptanintentionψsuchφandψare mutuallyexclusive. Itwouldnotberationalofmetoadoptan intentiontoφifIbelievedφwasnotpossible. 29 Intentions(3/4) 30 Intentions(4/4) Undercertaincircumstances,agentsbelieve theywillbringabouttheirintentions. ItwouldnotnormallyberationalofmetobelievethatI wouldbringmyintentionsabout;intentionscanfail.Moreover, itdoesnotmakesensethatifIbelieveφisinevitablethatI wouldadoptitasanintention. Agentstrackthesuccessoftheirintentions,and areinclinedtotryagainiftheirattemptsfail. Ifanagent'sfirstattempttoachieveφfails,thenallother thingsbeingequal,itwilltryanalternativeplantoachieveφ. 31 Agentsneednotintendalltheexpected sideeffectsoftheirintentions. IfIbelieveφ⇒ψandIintendthatφ,Idonot necessarilyintendψalso.(Intentionsarenotclosed underimplication.) Thislastproblemisknownasthesideeffectorpackage dealproblem.Imaybelievethatgoingtothedentist involvespain,andImayalsointendtogotothedentist -butthisdoesnotimplythatIintendtosufferpain! 32 Intentions-Summary Means-EndsReasoning Intentionsdrivemeans-endsreasoning Intentionspersist Intentionsconstrainfuturedeliberation Intentionsinfluencebeliefsuponwhich futurepracticalreasoningisbased Given: arepresentationofgoal/intentiontoachieve arepresentationofactionsitcanperform arepresentationoftheenvironment generateaplantoachievethegoal. 33 AgentControlLoopVersion1 whiletrue observethe 3 update internalworld 4 deliberateabout whatintentions toachievenext 5 usemeans-ends reasoningtoget aplanforthe intention 6 executethe 7 endwhile 1 2 34 Deliberation(1/2) whiletruedo getnext perceptp; 3B:=brf(B,p); 4 I:= deliberate(B); 5 P:= plan(B,I); 6execute(P); 7 endwhile Howdoesanagentdeliberate? 1 2 beginbytryingtounderstandwhattheoptions availabletoyouare choosebetweenthem,andcommittosome. Chosenoptionsarethenintentions. 35 36 AgentControlLoopVersion2 Deliberation(2/2) whiletruedo 2 getnext perceptp; 3B:=brf(B,p); 4I:= deliberate(B); 5P:= plan(B,I); 6execute(P); 7endwhile Thedeliberatefunctioncanbedecomposedinto twodistinctfunctionalcomponents: optiongeneration 1 inwhichtheagentgeneratesasetofpossiblealternatives representoptiongenerationviaafunction,options,whichtakesthe agent’scurrentbeliefsandcurrentintentions,andfromthemdetermines asetofoptions(=desires). filtering inwhichtheagentchoosesbetweencompetingalternatives,and commitstoachievingthem. Inordertoselectbetweencompetingoptions,anagentusesafilter function. whiletruedo getnextpercept p; 3 B:=brf(B,p); 4 D:=options(B,I); 5 I:= filter(B,D,I); 6 P:=plan(B,I); 7 execute(P); 8 endwhile 1 2 37 ExampleofLogicBasedAgent(1/3) 38 ExampleofLogicBasedAgent(2/3) Cleaningrobotwith: BeliefsBare: Perceptsp={dirt,X,Y} ActionsA={turnRight, forward,suck...} {dirt,0,2}{dirt,1,2} {pos,0,0,East} OptionsD: Start:(0,0,North) Goal:searchingand cleaningdirt {clean,0,2} {clean,1,2} 39 40 ExampleofLogicBasedAgent(3/3) CommitmentStrategies Afterfilteringthe intentionis: Blindcommitmentablindlycommittedagentwillcontinue tomaintainanintentionuntilitbelievestheintentionhas actuallybeenachieved.Blindcommitmentisalsosometimes referredtoasfanaticalcommitment. Single-mindedcommitmentasingle-mindedagentwill continuetomaintainanintentionuntilitbelievesthateither theintentionhasbeenachieved,orelsethatitisnolonger possibletoachievetheintention. Open-mindedcommitmentanopen-mindedagentwill maintainanintentionaslongasitisstillbelievedpossible. {clean,0,2} PlanP: {turnRight,forward,forward, suck} 41 42 AgentControlLoopVersion3 Commitment whiletruedo 2 getnextpercept p; 3 B:=brf(B,p); 4 D:= options(B,I); 5 I:= filter(B,D,I); 6 P:=plan(B,I); 7 execute(P); 8 endwhile Anagenthascommitmentbothtoends(i.e. ofwishestobringabout),andmeans(i.e., themechanismviawhichtheagentwishes toachievethestateofaffairs). Currently,ouragentcontrolloopis overcommitted,bothtomeansandends. Modification:replanifeveraplangoes wrong. 1 43 whiletruedo getnextpercept 3 B:= 4 D:= 5 I:= 6 whilenotempty(P)do 7 a:=first(P); 8 execute(a); P:=rest(P); 9 getnextperceptp; 10 B:=brf(B,p); 11 ifnotsound(P,B,I)then 12 P:=plan(B,I); 13 endwhile 14 endwhile 1 2 44 AgentControlLoopVersion4 Commitment whiletruedo getnextperceptp; 3 B:=brf(B,p); 4 D:=options(B,I); 5 I:=filter(B,D,I); 6 whilenotempty(P)do 7 a:=first(P); execute(a); 8 P:=rest(P); 9 getnextperceptp; 10 B:=brf(B,p); 11 ifnotsound(P,B,I) then 12 P:=plan(B,I); 13 endwhile 14 endwhile 1 Stillovercommittedtointentions:never stopstoconsiderwhetherornotits intentionsareappropriate. Modification:stoptodetermine whetherintentionshavesucceededor whethertheyareimpossible(singlemindedcommitment). 2 whiletruedo getnextperceptp; 3 B:=brf(B,p); 4D:=options(B,I); 5 I:=filter(B,D,I); 6 whilenotempty(P)or succeeded(B,I)or impossible(B,I)do 7 a:=first(P);execute(a); 8 P:=rest(P); 9 getnextperceptp; 10B:=brf(B,p); 11 ifnotsound(P,B,I)then 12 P:=plan(B,I); 13 endwhile; 14 endwhile 1 2 45 46 AgentControlLoopVersion5 IntentionReconsideration whiletruedo getnextperceptp; 3 B:=brf(B,p); 4D:=options(B,I); 5 I:=filter(B,D,I); 6 whilenotempty(P)or succeeded(B,I)or impossible(B,I)do 7 a:=first(P); execute(a); 8 P:=rest(P); 9 getnextperceptp; 10B:=brf(B,p); 11ifnotsound(P,B,I)then 12 P:=plan(B,I); 13 endwhile; 14 endwhile 1 Ouragentgetstoreconsideritsintentionsonce everytimearoundtheoutercontrolloop,i.e., when: 2 ithascompletelyexecutedaplantoachieveitscurrentintentions;or itbelievesithasachieveditscurrentintentions;or itbelievesitscurrentintentionsarenolongerpossible. Thisislimitedinthewaythatitpermitsanagent toreconsideritsintentions. Modification:Reconsiderintentionsafter executingeveryaction. 47 whiletruedo getnextperceptp; 3 B:=brf(B,p); 4 D:=options(B,I); 5 I:=filter(B,D,I); 6 whilenotempty(P)or succeeded(B,I)or impossible(B,I)</font>do 7 a:=first(P);execute(a); 8 P:=rest(P); 9 getnextperceptp; 10 B:=brf(B,p); 11 D:=options(B,I); 12 I:=filter(B,D,I); 13 ifnotsound(P,B,I)thenP :=plan(B,I); 14 endwhile; 15 endwhile 1 2 48 AgentControlLoopVersion6 IntentionReconsideration Butintentionreconsiderationiscostly!Adilemma: anagentthatdoesnotstoptoreconsideritsintentionssufficiently oftenwillcontinueattemptingtoachieveitsintentionsevenafterit isclearthattheycannotbeachieved,orthatthereisnolongerany reasonforachievingthem; anagentthatconstantlyreconsidersitsintentionsmayspend insufficienttimeactuallyworkingtoachievethem,andhence runstheriskofneveractuallyachievingthem. Solution:incorporateanexplicitmeta-level controlcomponent,thatdecideswhetherornotto reconsider. whiletruedo getnextperceptp; 3 B:=brf(B,p); 4 D:=options(B,I); 5 I:=filter(B,D,I); 6 whilenotempty(P)or succeeded(B,I)or impossible(B,I)do 7 a:=first(P);execute(a); 8 P:=rest(P); 9 getnextperceptp; 10 B:=brf(B,p); 11 D:=options(B,I); 12 I:=filter(B,D,I); 13 ifnotsound(P,B,I)then 14 P:=plan(B,I); 15 endwhile; 16 endwhile 1 2 whiletruedo getnextperceptp; 3B:=brf(B,p); 4D:=options(B,I); 5I:=filter(B,D,I); 6whilenotempty(P)or succeeded(B,I)or impossible(B,I)</font>do 7a:=first(P);execute(a); 8P:=rest(P); 9getnextperceptp; 10 B:=brf(B,p); 11 ifreconsider(B,I)then 12 D:=options(B,I);I:= filter(B,D,I); 13 ifnotsound(P,B,I)then 14 P:=plan(B,I); 15 endwhile 16 endwhile 1 2 49 OptimalIntentionReconsideration 50 KinnyandGeorgeff'sResults Ifɣislow(i.e.,theenvironmentdoesnotchangequickly), thenboldagentsdowellcomparedtocautiousones.This isbecausecautiousoneswastetimereconsideringtheir commitmentswhileboldagentsarebusyworkingtowards -andachieving-theirintentions. Ifɣishigh(i.e.,theenvironmentchangesfrequently),then cautiousagentstendtooutperformboldagents.Thisis becausetheyareabletorecognizewhenintentionsare doomed,andalsototakeadvantageofserendipitous situationsandnewopportunitieswhentheyarise. KinnyandGeorgeff'sexperimentally investigatedeffectivenessofintention reconsiderationstrategies. Twodifferenttypesofreconsideration strategywereused: boldagentsneverpausetoreconsiderintentions,and cautiousagentsstoptoreconsideraftereveryaction. Dynamismintheenvironmentisrepresented bytherateofworldchange,ɣ. 51 52 Therepresentation/reasoningproblem CritizismofSymbolicAI Howtosymbolicallyrepresentinformationaboutcomplex real-worldentitiesandprocesses. Howtotranslatetheperceivedworldintoanaccurate, adequatesymbolicdescription,intimeforthatdescription tobeuseful …vision,speechrecognition,learning. Howtogetagentstoreasonwiththisinformationintime fortheresultstobeuseful …knowledgerepresentation,automatedreasoning,planning. Duringcomputation,thedynamicworldmightchangeandthusthesolutionnot validanymore! Howtorepresenttemporalinformation,e.g.,howasituationchangesovertime? Therearemanyunsolvedproblemsassociatedwith symbolicAI. “Mostofwhatpeopledointheirdaytodaylivesisnotproblem-solving orplanning,butratheritisroutineactivityinarelativelybenign,but certainlydynamic,world.”(Brooks,1991) Theseproblemshaveledsomeresearcherstoquestion theviabilityofthewholeparadigm,andtothe developmentofreactivearchitectures. Althoughunitedbyabeliefthattheassumptions underpinningmainstreamAIareinsomesensewrong, reactiveagentresearchersusemanydifferenttechniques. 53 54 Brooks'DesignCriterias ReactiveArchitecture Anagentmustcopeappropriatelyandina timelyfashionwithchangesinitsenvironment. Anagentshouldberobustwithrespecttoits environment. Anagentshouldbeabletomaintainmultiple goalsandswitchbetweenthem. Anagentshoulddosomething,itshouldhave somepurposeinbeing. 56 Brooks-BehaviourLanguages Situatednessandembodiment:Theworldis itsownbestmodelanditgivestheagenta firmgroundforitsreasoning. Intelligenceandemergence:“Intelligent” behaviourarisesasaresultofanagent's interactionwithitsenvironment.Also, intelligenceis“intheeyeofthebeholder”; itisnotaninnate,isolatedproperty. Brookshasputforwardthreetheses: 1 2 3 Brooks-KeyIdeas Intelligentbehaviourcanbegeneratedwithout explicitrepresentationsofthekindthatsymbolicAI proposes. Intelligentbehaviourcanbegeneratedwithout explicitabstractreasoningofthekindthatsymbolic AIproposes. Intelligenceisanemergentpropertyofcertain complexsystems. 57 TheSubsumptionArchitecture(1/2) 58 TheSubsumptionArchitecture Traditionaldecompositionintofunctionalmodules: Toillustratehisideas,Brooksbuiltsome robotsbasedonhissubsumptionarchitecture. Asubsumptionarchitectureisahierarchyof task-accomplishingbehaviours. Eachbehaviourisarathersimplerule-like structure. Eachbehaviour“competes”withothersto exercisecontrolovertheagent. Decompositionbasedontaskachievingbehaviors: 59 60 TheSubsumptionArchitecture TheSubsumptionArchitecture Lowerlayersrepresentmoreprimitivekindsof behaviour,(suchasavoidingobstacles),andhave precedenceoverlayersfurtherupthehierarchy. Theresultingsystemsare,intermsoftheamount ofcomputationtheydo,extremelysimple. Someoftherobotsdotasksthatwouldbe impressiveiftheywereaccomplishedbysymbolic AIsystems. 61 62 ExampleofReactiveArchitecture StateMachines 63 FiniteStateMachines Example:lightswitch Itisamachinethatisinonestate amongafinitenumberofstates Afinitestatemachineisdefinedby AfinitenumberofstatesS AfinitenumberoftransitionsTbetweenstates Aninitialstates₀∊S Thecurrentstates∊S 65 Example:ambulance 66 Benefitsanddrawbacks Benefits: Simple Predictable Flexible Fast Verifiable/Provable Drawbacks: Complexityincreasefasterthanthenumberofstates andtransitions 67 68 Statecharts FSMCalculator ExtendFiniteState-Machines: Hierarchicalstates ConcurentStates Dataflow 69 70 AdvantagesOfReactiveSystems StateChartCalculator Simplicity,i.e.moduleshavehigh expressiveness Computationaltractability Robustnessagainstfailure,i.e. possibilityofmodelingredundancies Overallbehavioremergesfrom interactions 71 72 ProblemsWithReactiveSystems Thelocalenvironmentmustcontainenough informationtomakeadecision. Hardtotakenon-localinformationintoaccount. Behavioremergesfrominteractions⇒Howto engineerthesysteminthegeneralcase? Howtomodellong-termdecisions? Howtoimplementedvaryinggoals? Hardtoengineer,especiallylargesystemswith manylayersthatinteracts. HybridArchitecture 73 HybridArchitecture HybridArchitecture Ahybridsystemisneitheracompletelydeliberativenor completelyreactiveapproach. Anobviousapproachistobuildanagentoutoftwo(or more)subsystems: Inalayeredarchitecture,anagent'scontrolsubsystems arearrangedintoahierarchy,withhigherlayersdealing withinformationatincreasinglevelsofabstraction. Akeyprobleminlayeredarchitecturesiswhatkindof controlframeworktoembedtheagent'ssubsystemsin, tomanagetheinteractionsbetweenthevariouslayers. adeliberativeone,containingasymbolicworldmodel,whichdevelopsplans andmakesdecisionsinthewayproposedbysymbolicAI;and areactiveone,whichiscapableofreactingtoeventswithoutcomplex reasoning. Horizontallayering-Layersareeachdirectlyconnectedtothe sensoryinputandactionoutput.Ineffect,eachlayeritselfactslikean agent,producingsuggestionsastowhatactiontoperform. Verticallayering-Sensoryinputandactionoutputareeachdealtwithby atmostonelayereach. Often,thereactivecomponentisgivensomekindof precedenceoverthedeliberativeone.Thiskindof structuringleadsnaturallytotheideaofalayered architecture. 75 76 HybridArchitecture HybridArchitecture 77 78 AgentArchitecturesSummary Summary Originally(1956-1985),prettymuchallagents designedwithinAIweresymbolicreasoningagents Itspurestexpressionproposesthatagentsuseexplicit logicalreasoninginordertodecidewhattodo Problemswithsymbolicreasoningledtoareaction againstthis—theso-calledreactiveagents movement,1985–present From1990-present,anumberofalternativesproposed: hybridarchitectures,whichattempttocombinethebest ofreasoningandreactivearchitectures 80 ReactiveAgentArchitectures DeliberativeArchitectures Properties Properties Internalstate(usingsymbolicrepresentation) Search-baseddecisionmaking Goaldirected Noexplicitworldmodel Rule-baseddecisionmaking Benefits Benefits Efficient Robust Niceandclear(logics)semantics Easytoanalyzebyprovingproperties Problems Problems Thelocalenvironmentmustcontainenoughinformationto makeadecision. Easytobuildsmallagents,hardtobuildagentswith manybehaviorsorrules.Emergentbehavior. Can’treactinatimelymannertoeventsthatrequiresimmediateactions. Intractablealgorithms. Hardtocreateasymbolicrepresentationfromcontinuoussensordata. Theanchoringproblem. 81 HybridAgentArchitectures Properties Triestocombinethegoodpartsofbothreactive anddeliberativearchitectures. Usuallylayeredarchitectures. Benefits Attackstheproblemondifferentabstractionlevels. Hasthebenefitsofbotharchitecturetypes. Problems Harddocombinethedifferentparts. 83 82