Académique Documents
Professionnel Documents
Culture Documents
Outline
Time,schedulesandresources
BLG521E AdvancedArtificialIntelligence
Lecture7:PlanningandActing intherealworld
Multiagentplanning
AdvancedArtificialIntelligence(BLG521E)@ITU ::ComputerEngineeringDepartment,Dr.SanemSarelTalay
Time,schedulesandresources
Untilknow:
whatactionstodo
Time,schedulesandresources
Jobshopscheduling:
jobsconsistsofasequenceofactions Actionduration required resources Determineaschedulethatminimizesthetotaltime requiredtocompletealljobs(respectingresource constraints).
IntheReal World:
actions occur at certain moments in time. actionsoccuratcertainmomentsintime. actionshaveabeginningandanend. actionstakeacertainamountoftime.
AdvancedArtificialIntelligence(BLG521E)@ITU ::ComputerEngineeringDepartment,Dr.SanemSarelTalay
AdvancedArtificialIntelligence(BLG521E)@ITU ::ComputerEngineeringDepartment,Dr.SanemSarelTalay
AutomobileAssemblyProblem
Init(Chassis(C1) Chassis(C2) Engine(E1,C1,30) Engine(E1,C2,60) Wheels(W1,C1,30) Wheels(W2,C2,15)) Goal(Done(C1) Done(C2)) Action(AddEngine(e,c,m) PRECOND:Engine(e,c,d) Chassis(c) EngineIn(c) g ( ) ( ) g ( ) EFFECT:EngineIn(c) Duration(d)) Action(AddWheels(w,c) PRECOND:Wheels(w,c,d) Chassis(c) EFFECT:WheelsOn(c) Duration(d)) Action(Inspect(c) PRECOND:EngineIn(c) WheelsOn(c) Chassis(c) EFFECT:Done(c) Duration(10))
AdvancedArtificialIntelligence(BLG521E)@ITU ::ComputerEngineeringDepartment,Dr.SanemSarelTalay
SolutionfoundbyPOP
Slack of 15
AdvancedArtificialIntelligence(BLG521E)@ITU ::ComputerEngineeringDepartment,Dr.SanemSarelTalay
BLG521E - Lecture 7
Planningvs.scheduling
Howdoestheproblemdifferfromastandardplanning problem? Whendoesanactionstartandwhendoesitend? Soplanningdurationisalsoconsidered
Duration(d)
ESandLS
Earliestpossible(ES)andlatestpossible(LS)start times. LSES=slackofanaction forallactionsdeterminestheschedulefortheentire p problem.
ES(Start)=0 ES(B)=maxA<B ES(A)+Duration(A) LS(Finish)=ES(Finish) LS(A)=minA<B LS(B)Duration(A)
AdvancedArtificialIntelligence(BLG521E)@ITU ::ComputerEngineeringDepartment,Dr.SanemSarelTalay
Schedulingwithresources
Resourceconstraints=requiredmaterialorobjects toperformtask
Reusableresources Aresourcethatisoccupiedduringanactionbutbecomesavailable whentheactionisfinished. Requireextensionofactionsyntax: Resource: R(k)
kunitsofresourcearerequiredbytheaction. Isaprerequisitebeforetheactioncanbeperformed. Resourcecannotbeusedforktimeunitsbyother.
Carexamplewithresources
Init(Chassis(C1) Chassis(C2) Engine(E1,C1,30) Engine(E1,C2,60) Wheels(W1,C1,30) Wheels(W2,C2,15) EngineHoists(1) WheelStations(1) Inspectors(2)) Goal(Done(C1) Done(C2)) Action(AddEngine(e,c,m) PRECOND:Engine(e,c,d) Chassis(c) EngineIn(c) EFFECT:EngineIn(c) Duration(d), RESOURCE:EngineHoists(1)) Action(AddWheels(w,c) Action(AddWheels(w c) PRECOND:Wheels(w,c,d) Chassis(c) EFFECT:WheelsOn(c) Duration(d) RESOURCE:WheelStations(1)) Action(Inspect(c) PRECOND:EngineIn(c) WheelsOn(c) Chassis(c) EFFECT:Done(c) Duration(10) RESOURCE:Inspectors(1)) aggregation
AdvancedArtificialIntelligence(BLG521E)@ITU ::ComputerEngineeringDepartment,Dr.SanemSarelTalay
AdvancedArtificialIntelligence(BLG521E)@ITU ::ComputerEngineeringDepartment,Dr.SanemSarelTalay
Carexamplewithresources
Schedulingwithresources
Aggregation=groupindividualobjectsinto quantitieswhentheobjectsareundistinguishable withrespecttotheirpurpose.
Reducescomplexity
Heuristic:minimumslackalgorithm
Selectanactionwithallpredecessors scheduledandwiththeleast slackfortheearliestpossiblestart.
AdvancedArtificialIntelligence(BLG521E)@ITU ::ComputerEngineeringDepartment,Dr.SanemSarelTalay
AdvancedArtificialIntelligence(BLG521E)@ITU ::ComputerEngineeringDepartment,Dr.SanemSarelTalay
BLG521E - Lecture 7
Hierarchicaltasknetworkplanning
Reducecomplexity hierarchicaldecomposition Ateachlevelofthehierarchyacomputationaltaskisreducedtoa small numberofactivitiesatthenextlowerlevel Thecomputationalcostofarrangingtheseactivitiesislow Hierarchicaltasknetwork(HTN)planningusesarefinement ofactions throughdecomposition through decomposition e.g.buildingahouse=gettingapermit+hiringacontractor+doing theconstruction+payingthecontractor Refineduntilonlyprimitiveactionsremain PureandhybridHTNplanning ActiondecompositionsareusedasplanrefinementsinPOP
Representationdecomposition
Generaldescriptionsarestoredin a planlibrary
Eachmethod=Decompose(a,d); a=actionandd=POplan.
Finishactionhasalleffectsofactionsnotnegated by otheractions
=externaleffects Primaryeffects(usedtoachievegoal)vs.secondaryeffects
AdvancedArtificialIntelligence(BLG521E)@ITU ::ComputerEngineeringDepartment,Dr.SanemSarelTalay
AdvancedArtificialIntelligence(BLG521E)@ITU ::ComputerEngineeringDepartment,Dr.SanemSarelTalay
Buildhouseexample
Buildhouseexample
Action(Buyland,PRECOND:Money,EFFECT:Land Money) Action(GetLoan,PRECOND:Goodcredit,EFFECT:Money Mortgage) Action(BuildHouse,PRECOND:Land,EFFECT:House) Action(GetPermit,PRECOND:LAnd,EFFECT:Permit) Action(HireBuilder,EFFECT:Contract) Action(Construction,PRECOND:Permit Contract,EFFECT:HouseBuilt Permit), Action(PayBuilder,PRECOND:Money HouseBuilt,EFFECT:Money House Contract), Decompose(BuildHouse, ecompose( uildHouse, Plan::STEPS{S1:GetPermit,S2:HireBuilder,S3:Construction,S4PayBuilder} ORDERINGS:{Start<S1<S3<S4<Finish,Start<S2<S3}, LINKS
External precond
External effects
Start Land S1,Start Money S4,S1 Permit S3,S2 Contract S3, HouseBuilt house Money S3 S4,S4 Finish,S4 Finish
AdvancedArtificialIntelligence(BLG521E)@ITU ::ComputerEngineeringDepartment,Dr.SanemSarelTalay
AdvancedArtificialIntelligence(BLG521E)@ITU ::ComputerEngineeringDepartment,Dr.SanemSarelTalay
Propertiesofdecomposition
Shouldbecorrectimplementationofactiona CorrectifplandiscompleteandconsistentPOplanforthe problemofachievingtheeffectsofa giventhe preconditionsofa Adecompositionisnotnecessarilyunique Performs information hiding: Performsinformationhiding: STRIPSactiondescriptionofhigherlevelactionhidessome preconditionsandeffects Ignoreallinternaleffectsofdecomposition Doesnotspecifytheintervalsinsidetheactivityduring whichpreconditionsandeffectsmusthold InformationhidingisessentialtoHTNplanning.
AdvancedArtificialIntelligence(BLG521E)@ITU ::ComputerEngineeringDepartment,Dr.SanemSarelTalay
Testgoal
AdvancedArtificialIntelligence(BLG521E)@ITU ::ComputerEngineeringDepartment,Dr.SanemSarelTalay
BLG521E - Lecture 7
AdaptingPOP toHTNplanning
RememberPOP?
Modifythesuccessorfunction:applydecompositionto currentplan
NEWSuccessorfunction:
SelectnonprimitiveactionainP Se ect o p t e act o a ForanyDecompose(a,d) methodinlibrarywherea anda unifywithsubstitution
Replacea withd=subst(,d)
AdvancedArtificialIntelligence(BLG521E)@ITU ::ComputerEngineeringDepartment,Dr.SanemSarelTalay
AdvancedArtificialIntelligence(BLG521E)@ITU ::ComputerEngineeringDepartment,Dr.SanemSarelTalay
POP+HTNexample
a
POP+HTNexample
a
AdvancedArtificialIntelligence(BLG521E)@ITU ::ComputerEngineeringDepartment,Dr.SanemSarelTalay
AdvancedArtificialIntelligence(BLG521E)@ITU ::ComputerEngineeringDepartment,Dr.SanemSarelTalay
Howtohookupd ina ?
Removeactiona fromPandreplacewith d
Foreachstepsind selectanactionthatwillplaytheroleofs (eithernews orexistings fromP) Possibilityofsubtasksharing
WhataboutHTN?
AdditionalmodificationtoPOP arenecessary BADnews:pureHTNplanningisundecidable dueto recursivedecompositionactions.
Walk=makeonestepandwalk
Connectorderingstepsfora tothestepsind
Putallconstraintssothatconstraintsoftheform B<a aremaintained. B < a are maintained Watchoutfortoostrictorderings!
Connectthecausallinks
IfBp>aisacausallinkinP,replaceitbyasetofcausallinksfromBtoallstepsind withpreconditionspthatweresuppliedbythestartstep Dothesameforap>C
Resolveproblemsby
Ruleoutrecursion Boundthelengthofrelevantsolutions HybridizeHTNwithPOP
YetHTNcanbeefficient
AdvancedArtificialIntelligence(BLG521E)@ITU ::ComputerEngineeringDepartment,Dr.SanemSarelTalay AdvancedArtificialIntelligence(BLG521E)@ITU ::ComputerEngineeringDepartment,Dr.SanemSarelTalay
BLG521E - Lecture 7
PlanningandActinginNondeterministicdomains
Sofar:fullyobservable,staticanddeterministicdomains.
Agentcanplanfirstandthenexecuteplanwitheyesclosed
Handlingindeterminacy
Sensorlessplanning(conformantplanning)
Findplanthatachievesgoalinallpossiblecircumstances(regardlessof initialstateandactioneffects).
Conditionalplanning(Contingencyplanning)
Constructconditionalplanwithdifferentbranchesforpossible contingencies.
E Executionmonitoringandreplanning i i i d l i
Whileconstructingplanjudgewhetherplanrequiresrevision.
Degreeofuncertaintydefinedbyindeterminacy
Bounded:actionscanhaveunpredictableeffects,yetcanbelistedinactiondescription axioms.
Whenflippingacoin,outcomeiseitherheadortail
Continuousplanning
Planningactiveforalifetime:adapttochangedcircumstancesand reformulategoalsifnecessary.
Unbounded:preconditionsandeffectsunknownortoo largetoenumerate.
Taxidrivingagent
AdvancedArtificialIntelligence(BLG521E)@ITU ::ComputerEngineeringDepartment,Dr.SanemSarelTalay
AdvancedArtificialIntelligence(BLG521E)@ITU ::ComputerEngineeringDepartment,Dr.SanemSarelTalay
Sensorlessplanning
Anabstractexample
Initialstate=< chair, table,cansofpaint,unknowncolors > goalstate = < color(table)=color(chair) > Classicalplanningagentcouldnothandlethisproblem Sensorless planning(Conformant planning)
Openanycanofpaintandapplyittobothchairandtable. O f i t d l it t b th h i d t bl
Conditionalplanning(Contingencyplanning)
Sensecoloroftableandchair,iftheyarethesamethenfinish; elsesenselabelspaintifcolor(label)= color(Furniture)thenapplycolortotheotherpiece; elseapplycolortoboth
Executionmonitoringandreplanning
Sameasconditionalandcanfixerrors(missedspots)
Continuousplanning
Canrevisegoalwhenwewanttofirsteatbeforepaintingthetableandthechair.
AdvancedArtificialIntelligence(BLG521E)@ITU ::ComputerEngineeringDepartment,Dr.SanemSarelTalay
AdvancedArtificialIntelligence(BLG521E)@ITU ::ComputerEngineeringDepartment,Dr.SanemSarelTalay
Conditionalplanning
Dealwithuncertaintybycheckingthe environmenttoseewhatisreallyhappening. Assume infullyobservableand nondeterministicenvironments: nondeterministic environments:
Theoutcomeofanactionisunknown. Conditionalstepswillcheckthestateoftheenvironment. Howtoconstructaconditionalplan?
Example,thevacuumworld
AdvancedArtificialIntelligence(BLG521E)@ITU ::ComputerEngineeringDepartment,Dr.SanemSarelTalay
AdvancedArtificialIntelligence(BLG521E)@ITU ::ComputerEngineeringDepartment,Dr.SanemSarelTalay
BLG521E - Lecture 7
Conditionalplanning
Actions:left,right,suck Propositionstodefinestates:AtL,AtR,CleanL, CleanR Howtoincludeindeterminism?
Actionscanhavemorethanoneeffect E.g.movingleftsometimesfails Action(Left,PRECOND:AtR,EFFECT:AtL) Becomes:Action(Left,PRECOND:AtR,EFFECT:AtLAtR) Actionscanhaveconditionaleffects Action(Left,PRECOND:AtR,EFFECT:AtL(AtLwhen cleanL: cleanL) Bothdisjunctiveandconditional
AdvancedArtificialIntelligence(BLG521E)@ITU ::ComputerEngineeringDepartment,Dr.SanemSarelTalay
Conditionalplanning
Conditionalplansrequireconditionalsteps:
If<test> thenplan_A elseplan_B if AtLCleanL then Right else Suck Plansbecometrees
G Gamesagainstnature: i
Findconditionalplansthatworkregardlessofwhichactionoutcomes actuallyoccur. Assumevacuumworld Initialstate=AtR CleanL CleanR DoubleMurphy:possibilityofdepositdirtwhenmovingtoother squareandpossibilityofdepositingdirtwhenactionisSuck.
AdvancedArtificialIntelligence(BLG521E)@ITU ::ComputerEngineeringDepartment,Dr.SanemSarelTalay
Gametree
State node
SolutionofGamesagainstNature
Solutionisasubtreethat Hasagoalnodeateveryleaf Specifiesoneactionateachofitsstatenodes Includeseveryoutcomebranchateachofthechance nodes. Inpreviousexample:
[Left,if AtL CleanL CleanR then [ ]else Suck]
chance node
AndOrSearchalgorithm
function AND-OR-GRAPH-SEARCH(problem) returns a conditional plan or failure return OR-SEARCH(INITIAL-STATE[problem], problem, []) function OR-SEARCH(state, problem, path) returns a conditional plan or failure if GOAL-TEST[problem](state) then return the empty plan if state is on path then return failure for action,state_set in SUCCESSORS[problem](state) do plan AND SEARCH( t t l AND-SEARCH(state_set, problem, [ t t | plan] ) t bl [state l ] if plan failure then return [action | plan] return failure function AND-SEARCH(state_set, problem, path) returns a conditional plan or failure for each si in state_set do plani OR-SEARCH(si, problem,path ) if plan = failure then return failure return [ if s1 then plan1 else if s2 then plan2 else if sn-1 then plann-1 else plann]
AndOrSearch algorithm
Howdoesitdealwithcycles? Whenastatethatalreadyisonthepathappears,return failure
Nononcyclicsolution
Ensuresalgorithmtermination
Thealgorithmdoesnotcheckwhethersomestateisalreadyonsomeotherpathfromtheroot.
AdvancedArtificialIntelligence(BLG521E)@ITU ::ComputerEngineeringDepartment,Dr.SanemSarelTalay
AdvancedArtificialIntelligence(BLG521E)@ITU ::ComputerEngineeringDepartment,Dr.SanemSarelTalay
BLG521E - Lecture 7
AndOrSearch algorithm
Sometimesonlyacyclicsolutionexists e.g.tripplemurphy:sometimesthemoveisnotperformed
[Left,if CleanL then []else Suck]isnotasolution
CPin PartiallyObservableEnvironments
Fullyobservable:conditionaltestscanaskany questionandgetananswer Partiallyobservable
Theagenthaslimitedinformationabouttheenvironment. Modeled by a state set = belief states Modeledbyastateset=beliefstates E.g.assumevacuumagentwhichcannotsensepresenceorabsence ofdirtinothersquaresthantheoneitison. alternativemurphy:dirtcanbeleftbehindwhenmovingtoother square. Solutioninfullyobservableworld:keepmovingleftandright, suckingdirtwheneveritappearsuntilbothsquaresarecleanand Iminsquareleft.
Uselabeltorepeatpartsofplan(butinfiniteloops)
[L1:Left,if AtR then L1 else if CleanL then []else Suck]
AdvancedArtificialIntelligence(BLG521E)@ITU ::ComputerEngineeringDepartment,Dr.SanemSarelTalay
AdvancedArtificialIntelligence(BLG521E)@ITU ::ComputerEngineeringDepartment,Dr.SanemSarelTalay
PO:alternatedoublemurphy
Beliefstates Representation
Setsoffullstatedescriptions
{(AtRCleanRCleanL) (AtRCleanRCleanL)} Exponential
Logicalsentencesthatcapturethesetofpossible worldsinthebeliefstate(OWA)
AtR CleanR
Knowledgepropositionsdescribingtheagents knowledge(CWA)
K(AtR) K(CleanR)
AdvancedArtificialIntelligence(BLG521E)@ITU ::ComputerEngineeringDepartment,Dr.SanemSarelTalay
AdvancedArtificialIntelligence(BLG521E)@ITU ::ComputerEngineeringDepartment,Dr.SanemSarelTalay
Beliefstates
Choice2and3areequivalent(letscontinuewith3) Symbolscanappearinthreeways:positive,negativeor unknown:3n possiblebeliefstatesforn propositionsymbols. YET,setofbeliefsetsisapowersetofthephysicalstates whichismuchlargerthan3n Hence3isrestrictedasrepresentation
AnyschemecapableofrepresentingeverypossiblebeliefstatewillrequireO(2n) bittorepresent eachoneintheworstcase. ThecurrentschemeonlyrequiresO(n)
SensinginConditional Planning
Howdoesitwork?
Automaticsensing
Ateverytimesteptheagentgetsallavailablepercepts
Activesensing
Perceptsareobtainedthroughtheexecutionofspecificsensoryactions. checkDirt andcheckLocation
AdvancedArtificialIntelligence(BLG521E)@ITU ::ComputerEngineeringDepartment,Dr.SanemSarelTalay
AdvancedArtificialIntelligence(BLG521E)@ITU ::ComputerEngineeringDepartment,Dr.SanemSarelTalay
BLG521E - Lecture 7
Monitoringandreplanning
Executionmonitoring:checkwhether everythingisgoingasplanned.
Unboundedindeterminancy:someunanticipated circumstanceswillarise. Anecessityinrealisticenvironments.
Monitoringandreplanning
Whensomethingunexpectedhappens:replan
Toavoidtoomuchtimeonplanningtrytorepairtheoldplan.
Types ofmonitoring:
Actionmonitoring:verifywhetherthenextactionwill work. Planmonitoring:verifytheentireremainingplan.
AdvancedArtificialIntelligence(BLG521E)@ITU ::ComputerEngineeringDepartment,Dr.SanemSarelTalay
AdvancedArtificialIntelligence(BLG521E)@ITU ::ComputerEngineeringDepartment,Dr.SanemSarelTalay
Replanningagent
functionREPLANNINGAGENT(percept) returnsanaction static: KB,aknowledgebase(+actiondescriptions) plan,aplaninitially[] whole_plan,aplaninitially[] goal,agoal TELL(KB,MAKEPERCEPTSENTENCE(percept,t)) current STATEDESCRIPTION(KB,t) ifplan =[] whole_plan plan PLANNER(current, goal,KB) ifPRECONDITIONS(FIRST(plan))notcurrentlytrueinKBthen candidates SORT(whole_plan,orderedbydistancetocurrent) findstate s incandidates suchthat failure repair PLANNER(current,s,KB) continuation thetailofwhole_plan startingat s whole_plan plan APPEND(repair,continuation) returnPOP(plan)
Repairexample
AdvancedArtificialIntelligence(BLG521E)@ITU ::ComputerEngineeringDepartment,Dr.SanemSarelTalay
AdvancedArtificialIntelligence(BLG521E)@ITU ::ComputerEngineeringDepartment,Dr.SanemSarelTalay
Repairexample:painting
Init(Color(Chair,Blue)Color(Table,Green) ContainsColor(BC,Blue) PaintCan(BC) ContainsColor(RC,Red) PaintCan(RC)) Goal(Color(Chair,x) Color(Table,x)) Action(Paint(object,color) PRECOND:HavePaint(color) EFFECT:Color(object,color)) Action(Open(can) PRECOND:PaintCan(can) ContainsColor(can,color) EFFECT:HavePaint(color)) [Start;Open(BC);Paint(Table,Blue),Finish]
AdvancedArtificialIntelligence(BLG521E)@ITU ::ComputerEngineeringDepartment,Dr.SanemSarelTalay
Repairexample:painting
Supposethattheagentnowperceivesthatthecolorsoftable andchairaredifferent Figureoutpointinwhole_plan toaimfor
CurrentstateisidenticalasthepreconditionbeforePaint
Repairactionsequencetogetthere.
Repair=[]andplan=[Paint,Finish]
AdvancedArtificialIntelligence(BLG521E)@ITU ::ComputerEngineeringDepartment,Dr.SanemSarelTalay
BLG521E - Lecture 7
Planmonitoring
Checkthepreconditionsforsuccessoftheentire plan. Exceptthosewhichareachievedbyanotherstepinthe plan. Executionofdoomedplaniscutofearlier. Tradeoff between detection of failure and expensive Tradeoffbetweendetectionoffailureandexpensive monitoring Limitationofreplanningagent: Itcannotformulatenewgoalsoracceptnewgoalsin additiontothecurrentone
Continuousplanning
Agentpersistsindefinitelyinanenvironment
Phasesofgoalformulation,planningandacting
AdvancedArtificialIntelligence(BLG521E)@ITU ::ComputerEngineeringDepartment,Dr.SanemSarelTalay
AdvancedArtificialIntelligence(BLG521E)@ITU ::ComputerEngineeringDepartment,Dr.SanemSarelTalay
Blocks worldexample
Initialstate: (a)
Action(Move(x,y), PRECOND:Clear(x) Clear(y) On(x,z) EFFECT:On(x,y) Clear(z) On(x,z) Clear(y)
Blocks worldexample
Assumethatperceptsdontchangeandthisplanisconstructed OrderingconstraintbetweenMove(D,B) andMove(C,D) Startislabelofcurrentstateduringplanning. Beforetheagentcanexecutetheplan,natureintervenes:
DismovedontoB
AdvancedArtificialIntelligence(BLG521E)@ITU ::ComputerEngineeringDepartment,Dr.SanemSarelTalay
AdvancedArtificialIntelligence(BLG521E)@ITU ::ComputerEngineeringDepartment,Dr.SanemSarelTalay
Blocks worldexample
StartcontainsnowOn(D,B) Agentperceives:Clear(B) andOn(D,G)arenolongertrue
Updatemodelofcurrentstate(start)
Blocks worldexample
Extending:wheneveracausallinkcanbesuppliedbyaprevious step Allredundantsteps(Move(D,B)andits causallinks)areremoved fromtheplan Executenewplan,performactionMove(C,D)
Thisremovesthestepfromtheplan
AdvancedArtificialIntelligence(BLG521E)@ITU ::ComputerEngineeringDepartment,Dr.SanemSarelTalay
AdvancedArtificialIntelligence(BLG521E)@ITU ::ComputerEngineeringDepartment,Dr.SanemSarelTalay
BLG521E - Lecture 7
Blocks worldexample
Executenewplan,performactionMove(C,D)
AssumeagentisclumsyanddropsConA
Blocks worldexample
SimilartoPOP Oneachiterationfindplanflawandfixit Possibleflaws:Missinggoal,Openprecondition,Causalconflict, Unsupportedlink,Redundantaction,Unexecutedaction, unnecessaryhistoricalgoal
AdvancedArtificialIntelligence(BLG521E)@ITU ::ComputerEngineeringDepartment,Dr.SanemSarelTalay
AdvancedArtificialIntelligence(BLG521E)@ITU ::ComputerEngineeringDepartment,Dr.SanemSarelTalay
Multiagent planning
Sofarweonlydiscussedsingleagentenvironments. Otheragentscansimplybeaddedtothemodelof theworld:
Poorperformancesinceagentsarenotindifferentof otheragents intentions
Cooperation:Jointgoalsandplans
Multiplanningproblem:assumedoubletennis examplewhereagentswanttoreturnball.
Agents(A,B) Init(At(A,[Left,Baseline]) At(B,[Right,Net]) Approaching(Ball,[Right, Baseline]) Partner(A,B) Partner(B,A)) Goal(Returned(Ball) At(agent,[x,Net])) Goal(Returned(Ball) At(agent [x Net])) Action(Hit(agent,Ball) PRECOND:Approaching(Ball,[x,y]) At(agent,[x,y]) Partner(agent, partner) At(partner,[x,y]) EFFECT:Returned(Ball)) Action(Go(agent,[x,y]) PRECOND:At(agent,[a,b]) EFFECT:At(agent,[x,y]) At(agent,[a,b]))
AdvancedArtificialIntelligence(BLG521E)@ITU ::ComputerEngineeringDepartment,Dr.SanemSarelTalay
Ingeneraltwotypesofmultiagentenvironments:
Cooperative Competitive
AdvancedArtificialIntelligence(BLG521E)@ITU ::ComputerEngineeringDepartment,Dr.SanemSarelTalay
Cooperation:Jointgoalsandplans
Jointgoal:winningthematch Asolutionisajointplan consistingofactionsforbothagents. Plan1:
A:[Go(A,[Right,Baseline]),Hit(A,Ball)] B:[NoOp(B),NoOp(B)]
Multibodyplanning
Planningproblemfacedbyasinglecentralizedagentthatcan dictateactions toeachofseveralphysicalentities.
Hencenottrulymultiagent
Plan2:
A:[Go(A,[Left,net),NoOp(A)] B:[Go(B,[Right,Baseline]),Hit(B,Ball)]
10
BLG521E - Lecture 7
Multibodyplanning
Alternativetosetofalljointactions:addextra concurrencylinestoactiondescription Concurrentaction
Action(Hit(A,Ball) CONCURRENT:Hit(B,Ball) PRECOND:Approaching(Ball,[x,y]) At(A,[x,y]) EFFECT:Returned(Ball)) EFFECT R t d(B ll))
Multibodyplanning
PlannersimilartoPOP withsomesmallchangesin possibleorderingrelations
Temporalorderingrelations
A<B(before) A=B(concurrent) A B ( before or concurrent) A B(beforeorconcurrent)
Requiredactions(carryingobjectbytwoagents)
Action(Carry(A,cooler,here,there) CONCURRENT:Carry(B,cooler,herethere) PRECOND:)
Concurrentactionsareinstantiated(neworexisting) Prohibitiedconcurrentactions:constraintstoberesolved
AdvancedArtificialIntelligence(BLG521E)@ITU ::ComputerEngineeringDepartment,Dr.SanemSarelTalay
AdvancedArtificialIntelligence(BLG521E)@ITU ::ComputerEngineeringDepartment,Dr.SanemSarelTalay
Coordinationmechanisms
Toensureagreementonjointplan:useconvention. Convention=aconstraintontheselectionofjointplans
sticktoyourcourtoroneplayerstaysatthenet oneplayeralwaysstaysatthenet
Flockingexample
Threerules:
Separation: Steerawayfromneighborswhenyougettooclose Cohesion Steertowards theaveragepositionofneighbors Alignment Steertowards theaverageorientationofneighbors Steer towards the average orientation of neighbors
Canbedomainspecificorindependent Couldarisethroughevolutionaryprocess(flockingbehavior).
AdvancedArtificialIntelligence(BLG521E)@ITU ::ComputerEngineeringDepartment,Dr.SanemSarelTalay
Coordinationmechanisms
Intheabsenceofconventions:
Communication
e.g. Shout Mine!OrYours!intennisexample
Competitiveenvironments
Agentscanhaveconflictingutilities
e.g.zerosumgameslikechess
Planrecognition
Theagentmust:
Recognizethatthereareotheragents Computesomeoftheotheragentsplans Computehowtheotheragentsinteractwithitsownplan Compute how the other agents interact with its own plan Decideonthebestactioninviewoftheseinteractions.
Th b d Theburdenofarrivingatasuccessfulljointplancan f i i t f ll j i t l beplacedon
Agentdesigner(agentsarereactive,noexplicitmodelsofother agents) Agent(agentsaredeliberative,modelofotheragentsrequired)
AdvancedArtificialIntelligence(BLG521E)@ITU ::ComputerEngineeringDepartment,Dr.SanemSarelTalay
AdvancedArtificialIntelligence(BLG521E)@ITU ::ComputerEngineeringDepartment,Dr.SanemSarelTalay
11