as planning - cse.iitd.ernet.insak/courses/foav/planning-as-mc-slides.pdf · planning as model chec...
TRANSCRIPT
PlanningasModelChecking
MarcoPistore
DepartmentofInformaticsandTlc.UniversityofTrento-Italy
e-mail:[email protected]
1
The“Classical”PlanningProblem
•Domain=States(blockpositions)+Actions(moves)
•InitialState(“RedonTable”,“BlueonTable”,“GreenonBlue”)
•GoalState(“BlueonGreen”,“GreenonRed”,“RedonTable”)
•Plan(“MoveGreenonRed”,then“MoveBlueonGreen”)
PlanningProblem:Givenadomain(statesandactions),aninitialandgoalstate,theplanningproblemistheproblemtofindaplanofactionsthatleadsfromtheinitialstatetothegoal
2
The“Classical”PlanningProblem(cont.)
Abasicassumption...
3
The“Classical”PlanningProblem(cont.)
Abasicassumption...nouncertainty...
4
The“Classical”PlanningProblem(cont.)
Abasicassumption...nouncertainty...
5
PlanningunderUncertainty
Whathappensif...
6
PlanningunderUncertainty
Whathappensif...
The“Classical”PlanningAnswer:Planforthenominalcase!
7
PlanningunderUncertainty
Whathappensif...
The“Classical”PlanningAnswer:Planforthenominalcase!
BUT...
8
PlanningunderUncertainty
Whathappensif...
The“Classical”PlanningAnswer:Planforthenominalcase!
But:
•thissolutionisnotalwaysviable(planningunderuncertainty)
9
PlanningunderUncertainty
Whathappensif...
The“Classical”PlanningAnswer:Planforthenominalcase!
But:
•thissolutionisnotalwaysviable(planningunderuncertainty)
•thereisamuchbetterapproach(planningasmodelchecking)
10
Outline
•Whyplanningunderuncertainty?
•“Classical”planningasmodelchecking
•Planningwithtemporallyextendedgoals
•Conclusions
11
Outline
•Whyplanningunderuncertainty?
•“Classical”planningasmodelchecking
•Planningwithtemporallyextendedgoals
•Conclusions
12
PlanningunderUncertainty:Non-Determinism
Planforthenomicalcase?
Inmanydomains:
•actionshavenon-nominaloutcomesthatarehighlycritical.
13
PlanningunderUncertainty:Non-Determinism
Planforthenomicalcase?
Inmanydomains:
•actionshavenon-nominaloutcomesthatarehighlycritical.
•thereareactionswithnonominaloutcome.
14
PlanningunderUncertainty:Non-Determinism
Planforthenomicalcase?
Inmanydomains:
•actionshavenon-nominaloutcomesthatarehighlycritical.
•thereareactionswithnonominaloutcome.
Difficulties:
•aplanmayresultinmanydifferentexecutions.•theplannermustgenerateplansthathaveconditional
behaviours.•...
15
PlanningunderUncertainty:PartialObservability
Theassumptionofclassicalplanning:observationsnotneeded!
16
PlanningunderUncertainty:PartialObservability
Theassumptionofclassicalplanning:observationsnotneeded!
Butinseveralrealisticproblems,observationsareneeded.
17
PlanningunderUncertainty:PartialObservability
Theassumptionofclassicalplanning:observationsnotneeded!
Butinseveralrealisticproblems,observationsareneeded.
Difficulties:
•thestateofthesystemisonlypartiallyvisibleatrun-time.
•differentstatesareindistinguishableforthecontroller,namelyobservationsreturnsetsofstatesratherthansinglestates.
18
PlanningunderUncertainty:Extendedgoals
Theassumptionofclassicalplanning:goalsaresetsofstates!
19
PlanningunderUncertainty:Extendedgoals
Theassumptionofclassicalplanning:goalsaresetsofstates!
Butinseveralrealisticproblems,goalsareconditionsontheentireexecutionpathofaplan!
20
PlanningunderUncertainty:Extendedgoals
Theassumptionofclassicalplanning:goalsaresetsofstates!
Butinseveralrealisticproblems,goalsareconditionsontheentireexecutionpathofaplan!
•Goalsmayinvolvetemporalconditions(e.g.,airconditioner,safetyconditions)
•Goalsmayspecifyrequirementsofdifferentstrenghtthattakeintoaccountnondeterminismandpossiblefailures.
21
PlanningunderUncertainty:Extendedgoals
Theassumptionofclassicalplanning:goalsaresetsofstates!
Butinseveralrealisticproblems,goalsareconditionsontheentireexecutionpathofaplan!
•Goalsmayinvolvetemporalconditions(e.g.,airconditioner,safetyconditions)
•Goalsmayspecifyrequirementsofdifferentstrenghtthattakeintoaccountnondeterminismandpossiblefailures.
Difficulties:
•Extendedgoalsaddafurthercomplexitytothealreadycomplicatedproblem.
22
PlanningunderUncertainty:DifferentDimensions
probabilistic
non-determ.
deterministic full obs.
partial obs.
no obs.
reachability goals
extended goals
23
WhyPlanningunderUncertainty?
DomainPlanner
Plan
Controller
SystemActions
Observations
Goal
24
BecauseitisUseful!RealCase(I)
3456789
AAB
SDBSDBSDBSDBSDBSDBSDB
OPERATOR
. . . . . .
SCHEDULER
PROCESS i
PROCESS n
Safety Logic
PROCESS 1
COMMANDS
MANUAL
PERIPHERAL CONTROLS
PERIPHERAL STATUS
DEVICESPHERIPHERAL
Sourcesofuncertainty:
•operator,traindynamics,faults(actionswithuncertaineffects)•localsensors,neighborscontrollers(partialobservability)
25
BecauseitisUseful!ARealCase(II)
Dir. Gas
Interfaccia Utente
ON/OFF
ALARM/RESET
ALTA PRESSIONE
FN
ON/OFF/RESET
REQUEST
OPERATION
TERMICA
Evapor
Condensatore
Dir. Gas
Alta
pre
ssio
ne
Con
trol
lo
Con
trol
lo V
alvo
la
Ven
tole
Tem
pera
tura
Ven
tole
Bassa Pressione
Comandi
Compressore
Termica
Compressore
Controllo led
Pressione tasti
Tem
pera
tura
term
oreg
olaz
ione
Flussostato
Condensatore
Bassa Pressione
Flussostato
Controller
Sourcesofuncertainty:
•operator,temperature,faults(actionswithuncertaineffects)•unaccessiblevariables(partialobservability)
26
Outline
•Whyplanningunderuncertainty?
•“Classical”planningasmodelchecking
•Planningwithtemporallyextendedgoals
•Conclusions
27
ModelChecking
ModelChecking:atechniquetovalidateaformalmodelofasystemagainstalogicalspecification.
temporal formula
finite−state model
p
q
G(p −> Fq)
ModelChecker
p
yes!
no!
counterexample
q
28
PlanningbyModelChecking
PlanningbyModelChecking:atechniquetosynthesizeaplanfromaformalmodelofadomainandalogicalspecificationofagoal.
planning domain
goal
α
β
no plan!
yes!
Planner
reach r
p
q rα
βχ
plan
α
29
ModelCheckingandPlanning(informal)
•Themodelcheckingproblem:givenamodelMofasystemandapropertyϕ,where...
–ModelMisrepresentedasaFSM.
–Propertyϕisatemporallogicformula.
checkwhetherthepropertyissatisfiedinthemodel:M|=ϕ
•Theplanningproblem:givenamodelMofasystemandagoalϕ,findaplanπthatachievesthegoal:Mπ|=ϕ
•Theplanningvalidationproblemisamodelcheckingproblem
30
PlanningunderUncertaintybyModelChecking
Keyingredients:
2Planningdomainsasnon-deterministicstate-transitionsystems
2Goalsasformulasintemporallogic
2Plangenerationby(BDD-based,symbolic)modelcheckingtechniques
Results:
2Well-founded:formalframework,completeandcorrectalgorithms
2General:planninginnondeterministicdomains,underpartialobservability,andfor(temporally)extendedgoals,...
2Practical:implementationintheModelBasedPlanner(MBP)-automaticplanningforproblemsoflargesize
31
TheFirstResults:PlanningforReachabilityGoals
•Domains:nondeterministicautomata
•Goal:setofdesiredfinalstates
•Plans:memory-lesspoliciesthatmapstatestoactions
•Solutions:
1.Weaksolutions:“optimisticplans”[ECP97]
2.Strongsolutions:“safeplans”[AIPS98]
3.Strongcyclicsolutions:“iterativetrial-and-errorstrategies”[AAAI98]
A.Cimatti,M.Pistore,M.Roveri,P.Traverso.
Weak,Strong,andStrongCyclicPlanningviaSymbolicModelChecking.ArtificialIntelligence,147(1–2),2003
32
AnExample
DoorUncontrollable
Sensorsare not perfect
33
AnExample
34
Reachabilitygoals
Strongsolutions:Plansthatareguaranteedtoreachthegoal•allexecutiontracesreachthegoal
35
Reachabilitygoals
Weaksolutions:Plansthatmayachievethegoal•atleastoneexecutiontracereachesthegoal
36
Reachabilitygoals
StrongCyclicsolutions:trial-and-errorstrategies
•goalisreachablefromallthestatesofexecutiontraces•solutionsthatareguaranteedtoreachthegoalunderthe
fairnessassumptionof“noinfinitebadluck”
37
Implementation:TheMBPplanner
•MBP:AModelBasedPlanner(http://sra.itc.it/tools/mbp/)
•MBPbuiltontopofastate-of-the-artsymbolicBDD-basedmodelchecker,NUSMV(http://sra.itc.it/tools/nusmv/)
•MBPhastheplanningalgorithmsforweak,strong,andstrongcyclic(localandglobal)planning
•MBPhasalgorithmsforconformantplanning,forplanningunderpartialobservability,andforplanningfortemporallyextendedgoals
38
BDD-basedSymbolicModelChecking:Intuitions
SymbolicModelChecking[McMillan’93]basedonBDDs[Bryant’86]:
2Exploresetsofstatesrepresentedsymbolicallyasbooleanformulas
2BooleanformulasasOrderedBinaryDecisionDiagrams(BDDs)
2OBDDsrepresentstheassignmentssatisfying(andfalsifying)abooleanformula
2Operationsoversetsofstates(e.g.union,intersection)asbooleanoperations(e.g.conjunction,disjunction)implementedastransformationsoverBDDs
01
Loaded
0
Locked0
11
Locked1
10
0
Loaded
01
01
01
Loaded
0Locked
1
0 1
Loaded and not Locked Loaded or Locked Not Locked Loaded
39
PlannigasSymbolicModelChecking:Intuitions
Locked1
10
0
01
Loaded
0Locked
1
0 1
01
Loaded
0Locked
1
1 0
lockunload
unlocklock
unlock34 1
5
loadadjust
2
40
MoreResultsonPlanningforReachabilityGoals
•StrongSolutionswithPartialObservability:uncertaintyinobservations[Bertoli&Cimatti&Roveri&TraversoIJCAI01]
•PlanningforTemporallyExtendedGoals[Pistore&TraversoIJCAI01-AAAI02]
•Optimistic,Pessimistic&StrongCyclicplanninginUMOP[Jensen&VelosoJAIR00]
•Adversarialweak,strong,andstrongcyclicsolutions:environmentevents[Jensen,Veloso,BowlingECP01]
•SetBranchBDDbasedsearchenvironmentevents[Jensen&VelosoAIPS03]
41
Outline
•Whyplanningunderuncertainty?
•“Classical”planningasmodelchecking
•Planningwithtemporallyextendedgoals
•Conclusions
42
MotivationsforExtendedGoals
Themainmotivationsforintroducingextendedgoalsare:
•safeplanning:
–safetyconditions(“avoiddangerousstates”)complementthemaingoal.
•planningforreactivesystems:
–infiniteplanthatreactstoeventsintheenvironment(maildelivery,elevatorsystem,...).
•non-determinism:
–needtoexpress(reachability/maintainability)goalsofdifferentstrength(preferences).
43
Anexample
storelab
dep
Goal“reachdepandavoidlab”:
44
Anexample
storelab
dep
Goal“reachdepandavoidlab”:
•“Doreachdepanddoavoidlab”isunsatisfiable
45
Anexample
storelab
dep
Goal“reachdepandavoidlab”:
•“Doreachdepanddoavoidlab”isunsatisfiable•“Doreachdepandtrytoavoidlab”issatisfiablebyplan→
46
Anexample
storelab
dep
Goal“reachdepandavoidlab”:
•“Doreachdepanddoavoidlab”isunsatisfiable•“Doreachdepandtrytoavoidlab”issatisfiablebyplan→
•“Trytoreachdepanddoavoidlab”issatisfiablebyplan→
47
Planningforextendedgoals
Objectives:
•Planninginnon-deterministicdomainsforextendedgoals
•Dealinginpracticewithnon-determinismandcomplexgoalsindomainsoflargesize
Problems:
•Howcanweexpressextendedgoals?
•Whichkindofplansmustbegenerated?
•Planningalgorithm?
•Howcantheplanningalgorithmdealinpracticewithdomainsoflargesize?
48
The“PlanningbyModelChecking”approach
2ExtendedgoalsasformulasintheCTLtemporallogic:temporalconditionson“allpossiblestates”andon“somestates”resultingfromactionexecutions.
2Plansencodingconditional,iterative,andhistorydependentbehaviours,strictlymoreexpressivethanmemory-lesspolicies
2PlanningalgorithmsbasedonBDD-basedSymbolicModelCheckingtechniques,designedtodealwithlargestatespaces
2ImplementationintheModelBasedPlanner(MBP),aplannerbasedonthestate-of-theartsymbolicmodelcheckerNuSMV
2Experimentalresultsshowthattheplanningalgorithmworksinpractice
49
ExtendedGoalsareCTLformulas
CTL:g::=b|g∧g|g∨g|AFg|EFg|AGg|EGg|
A(gUg)|E(gUg)|A(gWg)|E(gWg)
Intuition:CTLcombines
•temporaloperators:F(eventually),G(always),U(until)...
FGU
•pathquantifiers:A(forallevolutions),E(forsomeevolution)
AE
50
ReachabilityGoalsinCTL
GivenaplaningdomainΣ(astatetransitionsystem),andagoalφ(aCTLformula),findaplanπsuchthatΣπ|=φ
1.WeakSolutions:φisEFp-plansthatmayreachthegoal
2.StrongSolutions:φisAFp-plansguaranteedtoreachthegoal
3.StrongCyclicsolutions:φisA(EFpWp)-iterativetrial-and-errorstrategieswhoseexecutionsalwayshaveapossibilityofterminatingand,whentheydo,theyareguaranteedtoreachthegoal.
51
MaintainabilityGoalsinCTL
GivenaplaningdomainΣ(astatetransitionsystem),andagoalφ(aCTLformula),findaplanπsuchthatΣπ|=φ
1.WeakMaintain:φisEGp-plansthatmaymaintainthegoal
2.StrongMaintain:φisAGp-plansguaranteedtomaintainthegoal
3.StrongCyclicMaintain:φisAGEFp-“maintainthepossibilityofreachingp”
52
ExamplesofCTLgoals
DoreachdepanddoavoidlabAFdep∧AG¬lab
DoreachdepandtrytoavoidlabAFdep∧EG¬lab
TrytoreachdepanddoavoidlabEFdep∧AG¬lab
53
Plansforextendedgoals:anexample
lab store
dep
•Thelabisadangerousroom—itmayharmtherobot
•Thegoalis“Continuously,trytoreachdepanddoreachstore”
•CTLgoal:AG(EFdep∧AFstore)
54
Plansforextendedgoals:anexample
store
dep
lab
Goal“Continuously,tryreachdepanddoreachstore”
•Satisfying“tryreachdep”(EFdep),...
55
Plansforextendedgoals:anexample
storelab
dep
Goal“Continuously,tryreachdepanddoreachstore”
•Satisfying“doreachstore”(AFstore),......
56
Plansforextendedgoals:anexample
store
dep
labstorelab
dep
?57
Plansforextendedgoals:anexample
lab
dep
store
Goal“Continuously,tryreachdepanddoreachstore”
•Satisfying“tryreachdep”...
•Satisfying“doreachstore”...
58
Plansforextendedgoals:anexample
Context 1Context 2
ExecutionContextsarenecessaryforthedifferentintentionsoftheexecutor:
•Context1:“tryreachdep”
•Context2:“doreachstore”
59
Plans
Aplanisdefinedintermsofanactionfunctionact:S×C⇀A,andacontextfunctionsense:S×C×S⇀C
statecontextactionnextstatenextcontext
swcontext1go-rightswcontext2
swcontext1go-rightdepcontext1/2
swcontext2go-upstorecontext1
...............
Context 1Context 2
store
storestore lablab
depdep sw
ne
sw
ne
60
Plans
APlanforadomainDisatuple〈C,c0,act,sense〉,where:
•Cisasetofexecutioncontexts,•c0istheinitialexecutioncontext,•act:S×C⇀Aistheactingfunction,•sense:S×C×S⇀Cisthesensingfunction.
Theidea:
•act(s,c)returnstheactiontobeexecutedbytheplan,•sense(s,c,s
′)associatestoeachreachedstates
′thenew
executioncontexts.
61
PlanExecution
•AtransitionofplanπinDisatuple(s,c)a
→(s′,c
′)suchthat:
–sa
→s′,
–a=act(s,c),and
–c′=sense(s,c,s
′).
•Arunofplanπfromstates0isaninfinitesequence(s0,c0)
a0
→(s1,c1)a1
→(s2,c2)a2
→(s3,c3)···
•TheexecutionstructureΣpofplanπhas:
–states(s,c)
–transitions(s,c)→(s′,c
′)
•Planthatsatisfiesagoal:Planπsatisfiesgoalgfrominitialstates0,writtenπ,s0|=g,if(s0,c0)|=ΣπgaccordingtothestandardsemanticsofCTL.
⇒planvalidationasmodelchecking:Σπ|=g
62
PlanningAlgorithm
functionsymbolic-plan(g):Plan
aut:=build-aut(g)
assoc:=build-assoc(aut)
plan:=extract-plan(aut,assoc)
returnplan
1.build-autconstructsanautomatonthatcontrolsthesymbolicsearch(statesarecontexts)
2.build-assocassociatesasetofstatesintheplanningdomaintoeachstateinthecontrolautomaton.
3.extract-planconstructsaplanfromthestatesassociatedtothecontexts.
63
Phase1:buildthecontrolautomaton
Whenwebuildthecontrolautomatonforthegivengoal:
•thecontrolstatesarethecontextsoftheplanthatisbeingbuilt
•thetransitionsrepresentthepossibleevolutionsofthecontextswhenactionsareexecuted.
64
Phase1:buildthecontrolautomaton
TwocontextsareneededforgoalAG(EFdep∧AFstore):
•onecorrespondingEFdep
•onecorrespondingAFstore
65
Phase1:buildthecontrolautomaton
store
store
InordertosatisfycontextAFstore,findanactionsuchthat:
•ifstoreholds,then:
–contextEFdepissatisfiableforALLtheoutcomes
•ifstoredoesnotholdthen:
–contextAFstoreissatisfiableforALLtheoutcomes
66
Phase1:buildthecontrolautomaton
dep
SOMEALL−OTHER
dep
InordertosatisfycontextEFdep,findanactionsuchthat:
•ifdepholds,then:
–contextAFstoreissatisfiableforALLtheoutcomes
•ifdepdoesnotholdthen:
–contextEFdepissatisfiableforSOMEoftheoutcomes–contextAFstoreissatisfiableforALLtheotheroutcomes
67
Phase1:buildthecontrolautomaton
dep
store
SOMEALL−OTHER
store dep
68
Phase2:search
Inthesearchphasethealgorithmassociatestoeachcontextthesetofstatesthatadmitaplanforthecontext.
•Initially,allthedomainstatesareassociatedtoeachcontext
•Theassociationisiterativelyrefined:
–Acontextischosen
–Thecorrespondingstatesarecomputed,basedonthecurrentassociation
–Theassociationforthecontextisupdated
•Thesearchterminateswhenafixpointisreached
69
Phase2:search
~dep
dep
~store
store
SOMEALL−OTHER
70
Phase2:search
~dep
dep
SOMEALL−OTHER
71
Phase2:search
~store
store
72
Phase2:search
~dep
dep
SOMEALL−OTHER
73
Step3:planextraction
•Findsuitableactionsforthestatesassociatedtothecontexts.
•Alltheinformationnecessarytoextracttheplanhasbeenalreadycomputedinthesearchphase.
•Reachabilityanalysisallowsforsimplerplans.
74
Step3:planextraction
~dep
dep
~store
store
SOMEALL−OTHER
go−south go−south
go−eastgo−northgo−north
go−west
75
Propertiesofthealgorithm
•Thealgorithmalwaysterminates.
•Thealgorithmiscorrectandcomplete:
–wheneverplansexist,thealgorithmfindsone;
–wheneverthereisnoplan,thealgorithmreturnsfail.
•Thecriticalstepforperformanceis“symbolicsearch”.
•ThealgorithmforextendedgoalsisimplementedinMBP
76
PlanningforCTLgoals:ExperimentalEvaluation
Aims:
•Testthescalabilityoftheapproach(domainsize,nondeterminism,goalcomplexity)
•ComparisonwithSimplan[Kabanzaet.al.](LTLgoals,explicitstate,handcodedstrategies)
•Comparisonwithspecialpurposestrong(cyclic)MBP
Results:
•Deterministiccase:SimPlanhandcodedstrategieswin•Nondeterministiccase:
–MBPperformancesdonotdegradewithnondeterminism–MBPoutperformsSimPlanevenwithhandcodedstrategies–Planningforextendedgoalscomparablewithspecialpurpose
strong(cyclic)algorithms
77
PlanningforCTLgoals:conclusions
•Goalswithtemporalconditionsonthewholeexecutionpath
•Goalsthattakeintoaccountnondeterminism(“forall”,“forsome”actionoutcomes)
•ImplementationintheMBPplanner(http://sra.itc.it/tools/mbp/)
•Experimentalevaluationshowsthattheapproachispractical
78
PlanningforCTLgoals:Conclusions
•Goalswithtemporalconditionsonthewholeexecutionpath
•Goalsthattakeintoaccountnondeterminism(“forall”,“forsome”actionoutcomes)
•ImplementationintheMBPplanner(http://sra.itc.it/tools/mbp/)
•Experimentalevaluationshowsthattheapproachispractical
...BUT...
79
LimitsofCTLgoals
RD
U A
S
Alarm!
Alarm!
SwitchToD
TrytoreachtheDirectposition,and...ifyouFail,DoreachtheAlarmposition
80
LimitsofCTLgoals
RD
U A
S
Alarm!
Alarm!
SwitchToD
TrytoreachtheDirectposition,and...ifyouFail,DoreachtheAlarmposition
Problem(1)...EFD–doesnotcaptureintentionality!!
81
LimitsofCTLgoals
RD
U A
S
Alarm!
Alarm!
SwitchToD
TrytoreachtheDirectposition,and...ifyouFail,DoreachtheAlarmposition
Problem(2)...CTLformulasdonotcapturepreferencesandfailurehandling!!
82
EaGLe:anExtendedGoalLanguage
EaGLeisanewExtendedGoalLanguagethat:
•canexpresstheintentionalaspectsnotcapturedinCTLandLTL(e.g.,“doeverythingpossibletoreachp”);
•candealwithfailureofgoalsandwithfailurerecovery(e.g.,“trytoachieveagoaland,ifyoufail,tryadifferentgoal”).
83
SyntaxofEaGLe
•reachability(basic)goals:DoReachp,TryReachp
•maintenance(basic)goals:DoMaintainp,TryMaintainp
•conjunction:gAndg′
•failure:gFailg′
•controloperators:gTheng′,Repeatg
84
DoReach
Goal“DoReachp”:
•requiresaplanthatguaranteestoreachpdespitenondeterminism
•failsifnosuchplanexists.
p
p
p
p
p
p
p
p
p
pp
success
failure
85
TryReach
Goal“TryReachp”:
•requiresaplanthatdoesitsbesttoreachp;
•failswhenthereisnopossibilitytoreachp.
p
pp
p
p
p
pp
p
p
p
86
Fail
Goal“g1Failg2”:
•dealswithfailure/recoveryandwithpreferencesamonggoals.
•Theplantriestosatisfygoalg1;wheneverafailureoccurs,goalg2isconsideredinstead.
•Example:DoReachpFailDoReachqvsTryReachpFailDoReachq
p
pq
q
q
q q qqq
p
p
87
TheRailwaysSwitchExample
RD
U A
S
Alarm!
Alarm!
SwitchToD
TryReachDFailDoReachA
R DUA
88
And/Then/Repeat
Goal“g1Andg2”:
•requirestosatisfyg1andg2inparallel.
Goal“g1Theng2”:
•requirestosatisfyg1andthentosatisfyg2.
Goal“Repeatg”:
•requirestosatisfyginacyclicway.
89
Asimpleexample
lab
store dep
office
Sensors arenot perfect
Uncontrollabledoor
Continuously,pickanobjectfromthestoreandtrytodeliverittotheoffice;ifyoufail,deliverittothedep.Donotenterthelab.
90
Apossibleplan
lab
store dep
office
Sensors arenot perfect
Uncontrollabledoor
init:/*inthestore*/
pickobject
/*trytoreachtheoffice*/
goeast;gosouth
if(room=office)then
dropobject
/*gobacktothestore*/
gowest;gonorth;gotoinit
else
/*reachthedep*/
gowest;gowest;dropobject
/*gobacktothestore*/
goeast;gotoinit
91
TheEaGleGoal
lab
store dep
office
Sensors arenot perfect
Uncontrollabledoor
Repeat(DoReach(store∧objpicked)
And(TryReach(office∧objdelivered)
FailDoReach(dep∧objdelivered)))
AndDoMaintain¬lab
92
TheEaGleControlAutomaton
dep
office
store lab
lab
lab
lab
dep
office
store
DoMaintainDoReach
TryReach
DoReach
93
PlanningwithEaGle
•Oncethecontrolautomatonhasbeenbuilt,thesamealgorithmusedforCTLgoalscanbeapplied.
•Thealgorithmisterminating,correct,andcomplete.
•ThealgorithmhasbeenimplementedinMBP.
•TheperformanceissimilartotheoneforCTLgoals...
•...butthequalityofthegeneratedplansismuchhigher.
94
Outline
•Whyplanningunderuncertainty?
•“Classical”planningasmodelchecking
•Planningwithtemporallyextendedgoals
•Conclusions
95
PlanningasModelChecking
goal
planning domain
q
r
α
βPMC
rαq
βα
χp
plan
96
Relatedwork
•TheMDP-planningapproach(e.g.,GPT[Bonet&Geffner])(includingMDP-planningbasedonADDs[Hoeyetal.])
•TheSAT-planningapproach([Rintanen],[Giunchiglia])
•Interleavingplanningandexecution([Koenig&Simmons])
•UMOP[Jensen&Veloso],basedonSymbolicModelChecking
•Plannersbasedonothermodelcheckingtechniques:
–Simplan[Kabanza]–ModelCheckingwithtimedautomata[Goldmanetal.]
•Automatatheoreticapproachtosynthesis
97
DirectionsforFutureResearch
null
reachability
GOALS
full partial
OBSERVABILITY
maintainability
.....................
MDP + MBP
TIME & RESOURCESEXECUTION
CONTROLLER SYNTHESIS
98
ThankstomanypeopleatTrento,butespeciallyto...
•PiergiorgioBertoli
•AlessandroCimatti
•MarcoRoveri
•PaoloTraverso
99