一种冲突导向的基于模型的自治方法_第1页
一种冲突导向的基于模型的自治方法_第2页
一种冲突导向的基于模型的自治方法_第3页
一种冲突导向的基于模型的自治方法_第4页
一种冲突导向的基于模型的自治方法_第5页
已阅读5页,还剩62页未读 继续免费阅读

下载本文档

版权说明:本文档由用户提供并上传,收益归属内容提供方,若内容存在侵权,请进行举报或认领

文档简介

Graph-basedPlanningBrianC.Williams

Sept.25th&30th,200216.412J/6.834JOutlineIntroductionTheGraphPlanPlanningProblemGraphConstructionSolutionExtractionPropertiesTerminationwithFailureGraphPlanGraphplanwasdevelopedin1995byAvrimBlumandMerrickFurst,atCMU.RecentimplementationsbyotherresearchershaveextendedthecapabilityofGraphplantoreasonwithtemporallyextendedactions,metricsandnon-atomicpreconditionsandeffects.Approach:GraphPlanConstructscompactencodingofstatespacefromoperatorsandinitialstate,whichprunesmanyinvalidplans–PlanGraph.Generatesplanbysearchingforaconsistentsubgraphthatachievesthegoals.PropositionInitStateActionTime1PropositionTime1ActionTime2PlangraphsfocustowardsvalidplansPlangraphsexcludemanyplansthatareinfeasible.Plansthatdonotsatisfytheinitialorgoalstate.Planswithoperatorsthatthreateneachother.Plangraphsareconstructedinpolynomialtimeandareofpolynomialinsize.Theplangraphdoesnoteliminateallinfeasibleplans.Plangenerationstillrequiresfocusedsearch.Example:GraphandSolutionnoGarbcleanHquietdinnerpresentcarrydollycookwrapcarrydollycookwrap

cleanHquiet

noGarbcleanHquietdinnerpresent1Prop1Action2Prop2Action3PropOutlineIntroductionTheGraphPlanPlanningProblemGraphConstructionSolutionExtractionPropertiesTerminationwithfailure8GraphPlanPlanningProblemStateAconsistentconjunctionofpropositions(positiveliterals)E.g.,(and(cleanhands)(quiet)(dinner)(present)(noGarbage))Doesn’tcommittothetruthofallpropositionsInitialStateProblemstateattimei=0E.g.,(and(cleanHands)(quiet))GoalStateApartialstate,representedbyaconjunctionofliterals.E.g.,(and(noGarbage)(dinner)(present))Theplannermustputthesysteminafinalstatethatsatisfiestheconjunction.9GraphPlanPlanningProblemActionsE.g.,(:operatorcarry

:precondition

:effect(:and(noGarbage)(not(cleanHands)))Preconditions:propositionsthatmustbetruetoapplyoperator.Aconjunctionofpropositions(nonegatedpropositions).Effects:Propositionsthatoperatorchanges,givenpreconditions.Aconjunctionofpropositions(adds)andtheirnegation(deletes).carrynoGarb

cleanHParameterizedschemasandobjectsareusedtocompactlyencodeactions.AddEdgeDeleteEdge10GraphPlanPlanningProblem(:operatorcook :precondition(cleanHands)

:effect(dinner))(:operatorcarry:precondition

:effect(:and(noGarbage)(not(cleanHands)))ActionExecutionattimei:Ifallpropositionsof:preconditionappearinthestateati,Thenstateati+1iscreatedfromstateati,byaddingtoi,all“add”propositionsin:effects,removingfromi,all“delete”propositionsin:effects.carrynoGarb

cleanHcookdinnercleanHands11GraphPlanPlanningProblemNoopsEverypropositionhasano-opaction,

whichmaintainsitfromtimeitoi+1.E.g.,(:operatornoop-P:precondition(P):effect(P))Noop-PPPExample:DinnerDateProblemInitialConditions:(and(cleanHands)(quiet))Goal: (and(noGarbage)(dinner)(present))Actions: (:operatorcarry:precondition

:effect(and(noGarbage)(not(cleanHands))) (:operatordolly:precondition

:effect(and(noGarbage)(not(quiet))) (:operatorcook:precondition(cleanHands)

:effect(dinner)) (:operatorwrap:precondition(quiet)

:effect(present)) +noops

dinnerpresentcookwrapcarry

cleanHquietnoGarbcleanH

dinnerpresentPropat1Actionat1Propat2Actionat2Propat2noop-dinnernoop-presentSetsofconcurrentactionsperformedateachtime[i]Usesmodelofconcurrencyasinterleaving.Ifactionsaandboccurattimei,thenitmustbevalidto

performeitherafollowedbyb,ORbfollowedbya.APlaninGraphPlan<Actions[i]>ACompleteConsistentPlanGiventhattheinitialstateholdsattime0,aplanisasolutioniff:Complete:Thepreconditionsofeveryoperatorattimei,

issatisfiedbyapropositionattimei.Thegoalpropositionsallholdinthefinalstate.Consistent:Theoperatorsatanytimeicanbeexecutedinanyorder,

withoutoneoftheseoperatorsundoingthe:preconditionsofanotheroperatorattimei.effectsofanotheroperatorattimei.

dinnerpresentcookwrapcarry

cleanHquiet

noGarbcleanH

dinnerpresentPropat1Actionat1Propat2Actionat2Propat3(noopdinner)(nooppresent)Exampleofa

CompleteConsistentPlanInitialConditions:(and(cleanHands)(quiet))Goal: (and(noGarbage)(dinner)(present))GraphPlanAlgorithmPhase1–PlanGraphExpansionCreatesgraphencodingpairwiseconsistencyandreachabilityofactionsandpropositionsfrominitialstate.Graphincludes,asasubset,allplansthatarecompleteandconsistent.Phase2-SolutionExtractionGraphtreatedasakindofconstraintsatisfactionproblem(CSP).Selectswhetherornottoperformeachactionateachtimepoint,byassigningCSPvariablesandtestingconsistency.OutlineIntroductionThePlanningProblemGraphConstructionSolutionExtractionPropertiesTerminationwithfailureGraphPropertiesAPlangraphcompactlyencodesthespaceofconsistentplans,whilepruning...partialstatesandactionsateachtimei

thatarenotreachablefromtheinitialstate.pairsofactionsandpropositions

thataremutuallyinconsistentattimei.plansthatcannotreachthegoals.Constructingtheplanninggraph…(Reachability)InitialpropositionlayerContainspropositionsininitialstate.Example:InitialState,Layer0

cleanHquiet

1Prop1Action2Prop2Action3PropConstructingtheplanninggraph…(Reachability)InitialpropositionlayerContainspropositionsininitialstateActionlayeriIfallaction’spreconditionsareconsistentini-1ThenaddactiontolayeriPropositionlayeri+1ForeachactionatlayeriAddallitseffectsatlayeri+1Example:AddActionsandEffectsnoGarbcleanHquietdinnerpresentcarrydollycookwrap

cleanHquiet

1Prop1Action2Prop2Action3PropConstructingtheplanninggraph…(Reachability)InitialpropositionlayerWritedownjusttheinitialconditionsActionlayeriIfallaction’spreconditionsappearconsistentini-1ThenaddactiontolayeriPropositionlayeri+1ForeachactionatlayeriAddallitseffectsatlayeri+1RepeatuntilallgoalpropositionsappearCanasolutionexist?noGarbcleanHquietdinnerpresentcarrydollycookwrap

cleanHquiet

1Prop1Action2Prop2Action3PropDoallgoalpropositionsappear?Constructingtheplanninggraph…(Consistency)InitialpropositionlayerWritedownjusttheinitialconditionsActionlayeriIfaction’spreconditionsappearconsistentini-1(non-mutex)ThenaddactiontolayeriPropositionlayeri+1ForeachactionatlayeriAddallitseffectsatlayeri+1IdentifymutualexclusionsActionsinlayeriPropositionsinlayeri+1Repeatuntilallgoalpropositionsappearnon-mutex26MutualExclusion:ActionsActionsA,Baremutually

exclusiveatleveli

ifnovalidplancouldpossiblycontainbothati:TheyInterfereAdeletesB’spreconditions,orAdeletesB’seffects,orViceversaorOR....Whatcausestheexclusion?noGarbcleanHquietdinnerpresentcarrydollycookwrap

cleanHquiet

1Prop1Action2Prop2Action3PropnoGarbcleanHquietdinnerpresentcarrydollycookwrap

cleanHquiet

Whatcausestheexclusion?1Prop1Action2Prop2Action3PropnoGarbcleanHquietdinnerpresentcarrydollycookwrap

cleanHquiet

Whatcausestheexclusion?1Prop1Action2Prop2Action3PropnoGarbcleanHquietdinnerpresentcarrydollycookwrap

cleanHquiet

Whatcausestheexclusion?1Prop1Action2Prop2Action3Prop31MutualExclusion:ActionsActionsA,Baremutually

exclusiveatleveli

ifnovalidplancouldpossiblycontainbothati:TheyInterfereAdeletesB’spreconditions,orAdeletesB’seffects,orViceversaorTheyhavecompetingneeds:A&BhaveinconsistentpreconditionsLayer0:completeactionmutexsnoGarbcleanHquietdinnerpresentcarrydollycookwrap

cleanHquiet

1Prop1Action2Prop2Action3Prop33MutualExclusion:PropositionLayerPropositionsP,QareinconsistentatiifnovalidplancouldpossiblycontainbothatiIfati,allwaystoachievePexcludeallwaystoachieveQPQA1A2MNLayer1:addpropositionmutexsnoGarbcleanHquietdinnerpresentcarrydollycookwrap

cleanHquiet

1Prop1Action2Prop2Action3PropNone!Canasolutionexist?noGarbcleanHquietdinnerpresentcarrydollycookwrap

cleanHquiet

1Prop1Action2Prop2Action3PropDoallgoalpropositionsappearnon-mutex?OutlineIntroductionThePlanningProblemGraphConstructionSolutionExtractionPropertiesTerminationwithfailure37GraphplanCreateplangraphlevel1frominitialstateLoopIfgoal

propositionsofthehighestlevel(nonmutex)ThensearchgraphforsolutionIfsolutionfound,thenreturnandterminateElseextendgraphonemorelevelAkindofdoublesearch:forwarddirectionchecksnecessary(butinsufficient)conditionsforasolution,...Backwardsearchverifies...38SearchingforaSolutionRecursivelyfindconsistentactionsachievingallgoalsatt,t-1,...:ForeachgoalpropositionGattimetForeachactionAmakingGtrueattIfAisn’tmutexwithpreviouslychosenactionatt,ThenselectitFinally,IfnoactionofGworks,ThenbacktrackonpreviousG.FinallyIfactionfoundforeachgoalattimet,Thenrecurseonpreconditionsofactionsselected,t-1,Elsebacktrack.Searchingforasolution

noGarb

cleanHquietdinnerpresentcarrydollycookwrap

cleanHquiet

1Prop1Action2Prop2Action3PropSelectactionachievingGoalnoGarbSearchingforasolution

noGarb

cleanHquiet

dinnerpresentcarrydollycookwrap

cleanHquiet

1Prop1Action2Prop2Action3PropSelectactionachievingGoaldinner,ConsistentwithcarryBacktrack

onnoGarbSearchingforasolution

noGarbcleanH

quietdinnerpresentcarrydollycookwrap

cleanHquiet

1Prop1Action2Prop2Action3PropSelectactiondollyBacktrack

onnoGarbSearchingforasolution

noGarbcleanH

quiet

dinnerpresentcarrydollycookwrap

cleanHquiet

1Prop1Action2Prop2Action3PropSelectactionachievingGoaldinner,ConsistentwithdollySearchingforasolution

noGarbcleanH

quiet

dinner

presentcarrydollycookwrap

cleanH

quiet

1Prop1Action2Prop2Action3PropSelectactionachievingGoalpresent,Consistentwdolly,cookSearchingforasolution

noGarbcleanH

quiet

dinner

presentcarrydollycookwrap

cleanH

quiet

1Prop1Action2Prop2Action3PropRecurseonpreconditionsofdolly,cook&wrapAllsatisfiedbyinitialstateAvoidingredundancyNo-opsarealwaysfavoured.guaranteesthattheplanwillnotcontainredundantactions.SupposethePlanGraphwasExtendednoGarbcleanHquietdinnerpresentcarrydollycookwrapcarrydollycookwrap

cleanHquiet

noGarbcleanHquietdinnerpresent1Prop1Action2Prop2Action3PropCanavalidplanusecarry?noGarbcleanHquietdinnerpresentcarrydollycookwrapcarrydollycookwrap

cleanHquiet

noGarbcleanHquietdinnerpresent1Prop1Action2Prop2Action3PropUsingcarryatlevel2isvalidnoGarbcleanHquiet

dinner

presentcarrydollycookwrapcarrydollycookwrap

cleanHquiet

noGarbcleanHquiet

dinner

present1Prop1Action2Prop2Action3PropUsingcarryatlevel2isvalidnoGarbcleanHquiet

dinner

presentcarrydollycookwrapcarrydollycookwrap

cleanH

quiet

noGarbcleanHquiet

dinner

present1Prop1Action2Prop2Action3PropOutlineIntroductionThePlanningProblemGraphConstructionSolutionExtractionPropertiesTerminationwithfailurePlanGraphPropertiesPropositionsmonotonicallyincreaseoncetheyareaddedtoalayertheyareneverremovedinsuccessivelayers;Mutexesmonotonicallydecreaseonceamutexhasdecayeditcanneverreappear;Memoizedsets(tobediscussed)monotonicallydecreaseifagoalsetisachievableatalayer

Thenitwillbeachievableatallsuccessivelayers.Thegraphwilleventuallyreachafixpoint,thatis,

alevelwherefactsandmutexesnolongerchange.FixpointExample:

DoorDomainABFixpointExample:

DoorDomainMovefromroomAtoroomBpre:robotisinAandthedoorisopenadd:robotisinBdel:robotisinAOpendoorpre:doorisclosedadd:doorisopendel:doorisclosedClosedoorpre:doorisopenadd:dooriscloseddel:doorisopennoopnoopMoveMoveOpennoopClosenoopIn(B)In(A)ClosedOpenedLayer4NMoveMoveOpenIn(A)ClosedLayer1OpennoopnoopIn(A)ClosedOpenedLayer2In(B)noopnoopMoveOpennoopCloseIn(A)ClosedOpenedLayer3Layer4isthefixedpoint(calledlevelout)ofthegraphGraphSearchPropertiesGraphplanmayneedtoexpandwellbeyondthefixpointtofindasolution.Why?GripperExampleMovefromoneroomtoanotherpre:robotinthefirstroomadd:robotinthesecondroomdel:robotinthefirstroomPickaballupinagripperpre:robothasafreegripperandisinsameroomasballadd:robotholdingtheballdel:gripperfree,ballinroom.Dropaballinaroompre:robotholdingtheball,robotindestinationroomadd:ballindestinationroom,gripperfree.del:ballbeingheld.GripperExampleInGripper,thefixpointoccursatlayer4,allfactsconcerningthelocationsoftheballsandtherobotarepairwisenon-mutexafter4steps.Thedistancetothesolutionlayerdependsonthenumberofballstobemoved.ForlargenumbersofballsGraphplansearchescopiesofthesamelayersrepeatedly.Forexample,for30ballsthesolutionlayerisatlayer59,54layerscontainidenticalfacts,actionsandmutexes.SearchProperties(cont)Plansdonotcontainredundantsteps.BypreferringNo-opsGraphplanguaranteesparalleloptimality.Theplanwilltakeasshortatimeaspossible.Graphplandoesn’tguaranteesequentialoptimalityMightbepossibletoachieveallgoalsatsomelayerwithfeweractionsatthatlayer.Graphplanreturnsfailureifandonlyifnoplanexists.OutlineIntroductionThePlanningProblemGraphConstructionSolutionExtractionPropertiesTerminationwithfailureSimpleTerminationIfthefixpointisreachedandeither:OneofthegoalsisnotassertedorTwoofthegoalsaremutexThenGraphplancanreturn"Nosolution"withoutanysearchatall.ElsetheremaybehigherorderexclusionswhichGraphplancannotdetectRequiresmoresophisticatedadditionaltestfortermination.WhyContinueAfterthe

FixPoint?facts,actionsandmutexesnolonger

changeafterthefixpoint,N-aryexclusionsDOchange.Newlayersaddtimetothegraph.TimeallowsactionstobespacedsothatN-arymutexeshavetimetodecay.Whilenewthingscanhappenbetweentwosuccessivelayersprogressisbeingmade.Trackn-arymutexes.Terminateontheirfixpoint.MemoizationandTerminationAgoalsetatalayermaybeunsolvableExample,inGripper.Ifagoalsetatlayerkcannotbeachieved,

Thenthatsetismemoizedatlayerk.TopreventwastedsearcheffortCheckseachnewgoalsetatkagainstmemosofkforinfeasibility.Ifanadditionallayerisbuiltandsearchingitcreatesnonewmemos,thentheproblemisunsolvable.Recap:GraphPlanGraphplanwasdevelopedin1995byAvrimBlumandMerrickFurst,atCMU.Graphplansearchesacompactencodingofthestatespacefromtheoperatorsandinitialstate,whichprunesmanyinvalidplans,viola

温馨提示

  • 1. 本站所有资源如无特殊说明,都需要本地电脑安装OFFICE2007和PDF阅读器。图纸软件为CAD,CAXA,PROE,UG,SolidWorks等.压缩文件请下载最新的WinRAR软件解压。
  • 2. 本站的文档不包含任何第三方提供的附件图纸等,如果需要附件,请联系上传者。文件的所有权益归上传用户所有。
  • 3. 本站RAR压缩包中若带图纸,网页内容里面会有图纸预览,若没有图纸预览就没有图纸。
  • 4. 未经权益所有人同意不得将文件中的内容挪作商业或盈利用途。
  • 5. 人人文库网仅提供信息存储空间,仅对用户上传内容的表现方式做保护处理,对用户上传分享的文档内容本身不做任何修改或编辑,并不能对任何下载内容负责。
  • 6. 下载文件中如有侵权或不适当内容,请与我们联系,我们立即纠正。
  • 7. 本站不保证下载资源的准确性、安全性和完整性, 同时也不承担用户因使用这些下载资源对自己和他人造成任何形式的伤害或损失。

评论

0/150

提交评论