learning and inference in dynamic...
TRANSCRIPT
LearningandInferenceinDynamicEnvironments
KatsumiInoue�PrinciplesofInforma8csResearchDivisionNa8onalIns8tuteofInforma8cs,Japan
� InoueLabmembers:KatsumiInoue,NicolasSchwind,MorganMagnin*1,TendaOkimoto*2,TonyRibeiro*1,MaximeClement,Kotaro
Okazaki,TaisukeSato*3
(*1:ÉcoleCentraledeNantes,*2:KobeUniversity,*3:AICenter,AIST)
NIIIABMee(ng,October28,2015
SNS
Disaster
Security
Transportation
Unexpected Events
Economics
Big,Changing,Con8nuous,
Real-Time,andIncompleteData
DataMining(Sta8c)
Real-TimeProcessing
(NoInference)
Agent
DynamicEnvironments
History/Experience
𝑶
Goal𝑶
BackgroundKnowledge𝑩
Ac@on
Model𝑯
Time-SeriesData 𝑰
InternalState
𝑩∧𝑰∧𝑯⊨𝑶
?
Aninternalmodelisconstructedbylearningfromenvironmentandinterac8onwithothersystems.Themodelisusedforchoosingthenextac8on.
Theeffect/resultofanac8onaffectstheenvironmentandupdates8me-seriesdata,history,experienceandgoals.Thenthemodelisupdatedaccordingly.
Agent
DynamicEnvironments
History/Experience
𝑶
Goal𝑶
BackgroundKnowledge𝑩
Ac@on
Model𝑯
Time-SeriesData 𝑰
InternalState
𝑩∧𝑰∧𝑯⊨𝑶
?
OtherAgents𝑰Ac@on
Agent
DynamicEnvironments
History/Experience
𝑶
Goal𝑶
BackgroundKnowledge𝑩
Ac@on
Model𝑯
Time-SeriesData 𝑰
InternalState
𝑩∧𝑰∧𝑯⊨𝑶
?Theagentalsointeractswithotheragents,anditsinternalmodelisrefinedbysuchinterac8ons. Ac@on
Agent
DynamicEnvironments
History/Experience
𝑶
Goal𝑶
BackgroundKnowledge𝑩
Ac@on
Model𝑯
Time-SeriesData 𝑰
InternalState
𝑩∧𝑰∧𝑯⊨𝑶
?
OtherAgents𝑰Ac@on
Ac@onDynamicTransi8onGoal:DevelopAItechniquesthatcaniden8fythemodelofasystemindynamicenvironmentsandcanachievetasksevenwhenunknown
situa8onsareencountered.
Agent
DynamicEnvironments
History/Experience
𝑶
Goal𝑶
BackgroundKnowledge𝑩
Ac@on
Model𝑯
Time-SeriesData 𝑰
InternalState
𝑩∧𝑰∧𝑯⊨𝑶
?
OtherAgents𝑰Ac@on
Ac@onMachineLearning
DynamicTransi8on
TheMachineLearningPartconstructsthemodeloftheagentbyabduc8onandinduc8onfromthegoal,data,historyandbackgroundknowledge.
DecisionMaking
LearningFromInterpreta8onTransi8on(LFIT)
LearningDynamicalandComplexNetworks• Learningdynamicsystemsinvolvingposi8veandnega8vefeedbacks• LearningBooleannetworksfromstatetransi8ondiagrams• LearningCellularAutomatafromtracesofconfigura8onchange
Ø Inoue,K.,Ribeiro,T.,Sakama,C.:“LearningfromInterpreta8onTransi8on”,MachineLearning,94(1):51-79,2014.
Ø Völker,M.,Inoue,K.:“LogicProgrammingforCellularAutomata”,ICLP2015.
currentpa\ern 111 110 101 100 011 010 001 000newstateforcentercell 0 1 1 0 1 1 1 0
• c(x,t+1)←c(x-1,t)∧c(x,t)∧¬c(x+1,t).• c(x,t+1)←c(x-1,t)∧¬c(x,t)∧c(x+1,t).• c(x,t+1)←¬c(x-1,t)∧c(x,t)∧c(x+1,t).• c(x,t+1)←¬c(x-1,t)∧c(x,t)∧¬c(x+1,t).• c(x,t+1)←¬c(x-1,t)∧¬c(x,t)∧c(x+1,t).
t 0 1 2 3 4
0
1
2
3
4
5
6
7
8
9
Wolfram’sRule110(Turing-complete)
LearningRobustBooleanNetworks
Li,F.etal.:Theyeastcell-cyclenetworkisrobustlydesigned,PNAS,101(14),2004.
• Mosttransi8onsfrom212statesbelongtothesamebasinofa2racKon.
• Fromthisstatetransi8on,LFITlearned54statetransi8onrulesin0.8sec.
• AnimprovedlearningalgorithmusingBDDlearnedthesamerulesin0.18sec.
Ø Inoue,K.,RibeiroT.,Sakama,C.:“LearningfromInterpreta8onTransi8on”,MachineLearning,94(1):51-79,2014.
Ø Ribeiro,T.,Inoue,K.,Sakama,C.:“ABDD-BasedAlgorithmforLearningfromInterpreta8onTransi8on”,Post-Proc.ILP2013,LNAI,Vol.8812,pp.47-63,2014.
“CanMachinesLearnLogics?”
InputAgentA(human/computer)Output
S:formulasLogicalSystemLT(⊆Th(S))
MachineM(S,T)LearningSystem CK
• Giveninput(S,T),amachineMproducesanaxioma8csystemK.• LFITcanlearnmeta-levelone-stepdeduc8onrules,e.g.,MP.Thescenario
canbeappliedtolearningabduc8onandothernon-standardlogics.Ø Sakama,C.,Inoue,K.:“CanMachinesLearnLogics?”,AGI2015,LNAI,Vol.9205,pp.
341-351,2015.
1 2 6
3
4
pickup mov
e(4)
3
4
5 5a
5b 0.7
0.9
0.1
RevisingPlansinAdap8veSystems
EnvironmentSystem
Domainmodel
action
reaction
model revision
Behaviouralmodellearning/revisionthroughprobabilis@crulelearning
Ø Sykes,D.,Corapi,D.,Magee,J.,Kramer,J.,Russo,A.,Inoue,K.:“Learningrevisedmodelsforplanninginadap8vesystems”,ICSE2013:63-71.
Ø Marpnez,D.,Ribeiro,T.,Inoue,K.,Alenyà,G.,Torras,C.:“LearningProbabilis8cAc8onModelsfromInterpreta8onTransi8ons”,ICLP2015.
DNAsynthesis
p53
cyclinH
CAK
p21/WAF1
cdk2
cyclinE/cdk2
Rb
cdk7
DNAdamage
cyclinE
(1)(2)
(3)
(4)
(5)
(6)
promoted
promoted
suppressed
suppressed
(7)
suppressed
*
Cellcyclewithcyclin-dependentkinases(Schneideretal.,2002)
PathwayComple8onbyMeta-LevelAbduc8on
Ø Inoue,K.,Doncescu,A.,Nabeshima,H.:“Comple8ngcausalnetworksbymeta-levelabduc8on,MachineLearning,91(2):239-277,2013.
Comple8ngSBGNNetworkswithGeneExpressionData(collabora8onwithLRI/Paris-Sud&INRA-CNRS)
Biologicalphenomena Formalsystem
Geneexpressiondata SBGN-AFnetwork
(Discre8zed)observa8ons
Discre8za8on
backgroundtheory
Causalnetworktransla8on
HypothesisNewinterac8onsorpredicates
Meta-LevelAbduc8onbySOLAR
Biologicalanalysisusingdatabaseor
literature
Inferenceengine
Ø Yamamoto,Y.,Rougny,A.,Nabeshima,H.,Inoue,K.,Moriya,H.,Froidevaux,C.,Iwanuma,K.:“Comple8ngSBGN-AFNetworksbyLogic-BasedHypothesisFinding”,FMMB2014,LNBI,Vol.8738,pp.165-179,2014.
• SBGN-AF:SystemsBiologyGraphicalNota8onAc8vityFlow
• Applica8ons:FSHR-inducedsignalingnetworks
Agent
DynamicEnvironments
History/Experience
𝑶
Goal𝑶
BackgroundKnowledge𝑩
Ac@on
Model𝑯
Time-SeriesData 𝑰
InternalState
𝑩∧𝑰∧𝑯⊨𝑶
?
OtherAgents𝑰Ac@on
Ac@onMachineLearning
DecisionMaking
DynamicTransi8on
TheDecisionMakingPartchoosesthebestnextac8onbysolvingconstraintsa8sfac8onproblems(CSP)andconstraintop8miza8onproblems(COP).
DynamicConstraintOp8miza8onProblems
• Mostreallifeproblemsaredynamic,e.g.,transporta8on,teamforma8on,scheduling.
• Thosemodelscanberepresentedas(dynamic)(hard&sou)constraintnetworks.
• Goal:Minimizepenaltyandmaximizereward.• Requiresfastcomputa8onofnewsolu8ons,yet
somequalityguaranteesshouldbeprovided.
DynamicCOP:Applica8onsandApproaches
• TeamformaKon:Makingrobustteamsofagents.• Nursererostering:Whenanurseisabsent,buildanewschedulewith
minimalandfairchanges.• Timetabling:Reconstruct8metablesaccordingtositua8onchanges.
Ø Okimoto,T.,Schwind,N.,Clement,M.,Ribeiro,T.,Inoue,K.,Marquis,P.:“HowtoFormaTask-OrientedRobustTeam”,AAMAS2015:395-403.
Ø Clement,M.,Okimoto,T.,Schwind,N.,Inoue,K.:“FindingResilientSolu8onsforDynamicMul8-Objec8veConstraintOp8miza8onProblems”,ICAART2015:509-516.
Twodifferentapproaches:• ProacKve:weprepareandtakeac8onsbefore
changeshappen.• ReacKve:wedonotknowwhatwillhappen
andcanonlyreacttothechanges.
TaskOrientedRobustTeamForma8on• Goal:Makeateamtoachieveagivensetoftasks
suchthatitcanbeeffec8veevenifsomeagentsbreakdown.
• Ateamisk-robustifitcans8llperformalltasksauerlosinganykagents.
• Robustnessisaproac8vesolu8ontoprepareforunpredictablechanges(thelossofsomeagents).
• TORTFdoesnotprovemorecomputa8onallydemandingthanthetask-efficientteamforma8onproblem,i.e.,robustnessisforfree.
Ø Okimoto,T.,Schwind,N.,Clement,M.,Ribeiro,T.,Inoue,K.,Marquis,P.:“HowtoFormaTask-OrientedRobustTeam”,AAMAS2015:395-403.
CyberSecurityTrade-OffProblem• “Youcan'thave100percentsecurityandthenalsohave100percentprivacy
andzeroinconvenience.”(BarackObama,2013)• Intercep8onandcommunica8onsdatareten8onmeasures,evenifthe
purposeissocialsecurity,areunderthedifficulttrade-offbetweensecurity,privacyandcost.”
• Howtosolvethistrade-offandbuildthesocietalconsensus?• WeusedMulK-ObjecKveConstraintOpKmizaKon(MO-COP)techniques.
PRIVACY
SECURITY COST
Ø Okimoto,T.,Ikegai,N.,Ribeiro,T.,Inoue,K.,Okada,H.,Maruyama,H.:CyberSecurityProblemBasedonMul8-Objec8veDistributedConstraintOp8miza8onTechnique,WSR2013.
SelectedSolu8onsforMul8-Objec8veProblems• Mo8va8on:ProductConfigura8onExample
→Many(some8mes,thousandsof)choicesaregiventotheuser.→Eachalterna8veinvolvesmanycriteria(e.g.,foracar,price,life8me,safety,brandreputa8on).→Impossibleforuserstochoosetheirpreferredproduct!
• Characteriza8onofrepresentaKvesoluKons(Figureb)/Efficient
approxima8onprocedures(Figurec)• Interes8ngbenefitsforanitera8veuse,forlarge-scaleproblems.Ø Schwind,N.,Okimoto,T.,Konieczny,S.,M.,Inoue,K.:toappear,2016.
Agent
DynamicEnvironments
History/Experience
𝑶
Goal𝑶
BackgroundKnowledge𝑩
Ac@on
Model𝑯
Time-SeriesData 𝑰
InternalState
𝑩∧𝑰∧𝑯⊨𝑶
?
OtherAgents𝑰Ac@on
Ac@onMachineLearning
DecisionMaking
DynamicTransi8on
TheModelCheckingPartverifiesthattheagentworkscorrectlybypredic8ngfutureresultsindynamicenvironments.
ModelChecking
ResilienceofConstraint-BasedDynamicSystems
Ø Schwind,N.,Okimoto,T.,Inoue,K.,Chan,H.,Ribeiro,T.,Minami,K.,Maruyama,H.:“SystemsResilience:AChallengeProblemforDynamicConstraint-BasedAgentSystems”,AAMAS2013,pp.785-788.
Ø ReceivedThe3rdPrizeofBestChallengesandVisionsPapers.
• Formaliza8onofgenericsystemsthroughCOP variables:componentsofthesystem constraints:interac8onsbetweenthesecomponents opKmizaKonfuncKon:evaluatesthequalityofaconfigura8onofthesystem
• Resistance+Recoverability+Func8onality=Resilience
BeliefPropaga8oninMul8-AgentSystems
• Wewanttopredictthepropaga8onoffallaciousbeliefsinsocialnetworks(brandcrisismanagement,e.g.,Domino’sPizzacrisisinApril2009)
• Wewanttotrackthetruthwhenseveralagentsdescribethesamesitua8onbuthaveconflic8ngbeliefs
Ø Schwind,N.,Inoue,K.,Bourgne,G.,Konieczny,S.,Marquis,P.:“BeliefRevisionGames”,AAAI-15:1590-1596,2015.
• Formaliza8onoftheframework,named"BeliefRevisionGames"• Setofappealingproper8es:
– 3Preserva8onproper8es:Convergence,Responsiveness,Monotonicity
– 6agents’revisionpolicies(rangingfromcreduloustoskep8calones)
• Inves8ga8onoftheextenttowhichtheseproper8esaresa8sfiedbytherevisionpolicies
• Robustframework,consistentwithnaturalexpecta8ons:→credulousagentsaremoreresponsivethanskep8calones.→withskep8calagents,convergenceisguaranteed.
BeliefPropaga8oninMul8-AgentSystems
ResilienceProper8esinOscillatoryBiologicalSystems
Comet, etal.,“Simplified models for the mammalian circadian clock“,Evry Spring school on Modelling complex biological systems in the context of genomics, pp.85-106, 2012.
• UseparametricKmePetrinetsinordertoanalyzepreciselythedynamicbehaviorofbiologicaloscillatorysystems,e.g.,themammaliancircadianclock.
• Analyzeresilienceproper8esthatendorsemajorchangesintheirenvironment,e.g.jet-lags,day-nightalterna8ngwork-8me.
Ø Andreychenko,A.,Magnin,M.,Inoue,K.:“ModelingofResilienceProper8esinOscillatoryBiologicalSystemsUsingParametricTimePetriNets”,CMSB2015:239-250.
Agent
DynamicEnvironments
History/Experience
𝑶
Goal𝑶
BackgroundKnowledge𝑩
Ac@on
Model𝑯
Time-SeriesData 𝑰
InternalState
𝑩∧𝑰∧𝑯⊨𝑶
?
OtherAgents𝑰Ac@on
Ac@onMachineLearning
DecisionMaking
DynamicTransi8onModelChecking
Program/ContentAudiences
Viewer’sJourneyLog
𝑶
TargetGRP 𝑶
BackgroundKnowledgeDatabase 𝑩
ADDistribu@on
ADPlanningtoeachIP𝑯
Ra@ngDataExogenousData
𝑰
InternalState
𝑩∧𝑰∧𝑯⊨𝑶
?
OtherAgents𝑰AudienceReac@ons
ContentsProgramming
MachineLearning
DecisionMaking
DynamicTransi8onModelChecking
Adapta8ontoIPBroadcas8ngSystems
InferenceandLearninginDynamicEnvironments
AItechniquesthatcaniden8fythemodelofasystemindynamicenvironmentsandcanachievetasksevenwhen
unknownsitua8onsareencountered.
Modeling(MachineLearning)
DecisionMaking(ConstraintOp(miza(on)
Predic@on/Verifica@on(ModelChecking)
ResilientSystemsRobustness/Sustainability
DynamicSchedulingCyberSecurity
OpinionConstruc8onMul8-AgentLearning