论文.docx

论文.docx

《论文.docx》由会员分享，可在线阅读，更多相关《论文.docx（45页珍藏版）》请在冰豆网上搜索。

论文

TheKnowledgeEngineeringReview,Vol.29:

1,51–77.&CambridgeUniversityPress,2013doi:

10.1017/S0*******13000076Firstpublishedonline21March2013

知识工程回顾，卷29：

1，51-77,剑桥大学出版社，2013DOI：

10.1017/S0*******13000076最先于2013年3月在网上公布。

Areviewofcase-basedreasoningincognition–actioncontinuum:

asteptowardbridgingsymbolicandnon-symbolicartificialintelligence

回顾认知连续行动基于案例加强对桥的象征和非符号人工智能的推理。

PINARO¨ZTU¨RKandAXELTIDEMANN

PINAR0ZTUR·K和阿克塞尔TIDEMANN

DepartmentofComputerandInformationScience,NorwegianUniversityofScienceandTechnology（NTNU）,SemSaelandsvei7-9,NO-7491Trondheim,Norway;e-mail:

pinar@idi.ntnu.no,axel.tidemann@

计算机与信息科学系，挪威科技大学（NTNU），SEMSaelandsvei7-9，NO-7491挪威特隆赫姆;电子邮件：

pinar@idi.ntnu.no，axel.tidemann@。

Abstract

Intheoriesandmodelsofcomputationalintelligence,cognitionandactionhavehistoricallybeeninvestigatedonseparategrounds.Weconjecturethatthemainmechanismofcase-basedreasoning（CBR）appliestocognitivetasksatvariouslevelsandofvariousgranularity,andhencecanrepresentabridge—oracontinuum—betweenthehigherandlowerlevelsofcognition.CBRisanartificialintelligence（AI）methodthatdrawsupontheideaofsolvinganewproblemreusingsimilarpastexperiences.Inthispaper,were-formulatethenotionofCBRtohighlightthecommonalitiesbetweenhigher-levelcognitivetaskssuchasdiagnosis,andlower-levelcontrolsuchasvoluntarymovementsofanarm.Inthisview,CBRisenvisagedasagenericprocessindependentfromthecontentandthedetailedformatofcases.Diagnosticcasesandinternalrepresentationsunderlyingmotorcontrolconstitutetwoinstantiationsofthecaserepresentation.Inordertoclaimsuchagenericmechanism,theaccountofCBRneedstoberevisedsothatitspositioninnon-symbolicAIbecomesclearer.ThepaperreviewstheCBRliteraturethattargetslowerlevelsofcognitiontoshowhowCBRmaybeconsideredasasteptowardbridgingthegapbetweensymbolicandnon-symbolicAI.

1Introduction

Intheoriesandmodelsofcomputationalintelligence,cognitionandactionhavehistoricallybeeninvestigatedonseparategrounds.Withinthemethodareaofartificialintelligence（AI）,symbol-processingapproacheshaveemphasizedhigher-levelcognition,whilesituatedandembodiedAIhasmorefocusonthebodilyactivitiesofanagent.TheissueofrepresentationhasbecomeaneffectivedeparturepointcontributingtotheescalationoftensionbetweenthesetwolinesofAI.Inrecentyears,however,atrendhasemergedthatislookingforsimilaritiesandcomplementaritiesbetweenunder-lyingmechanismsofhigher-levelcognitionandlower-levelmotorbehavior.Ongoingresearchincognitivescienceaimsatprovidingcluesofaseamlesslinkbetweencognitionandaction,resultingfromtheclosecooperationbetweenresearchersacrossdisciplinessuchascognitivepsychology,linguistics,neuroscienceandAI,whopreviouslyinvestigatedthesameproblemsinisolationfromeachother.Animportanthypothesisthathasariseninthiscontextistheequivalenceoftheneuralmechanismsunderlyingovertmotorbehaviorandmotorimagery（i.e.thinkingofamotoractionwithoutexecutingit）,whichimplicatestheexistenceofunderlyingrepresentations.Bothcognitivepsychology（Robinetal.,2007）,philosophy（Grush,2004）andneuroscience（Filimonetal.,2007;Zimmermann-Schlatteretal.,2008）findingsprovideconvincingevidencesupportingthishypothesis.Knowledgemodelsofhigher-levelreasoningtasksincludegeneralizedmodelssuchasrule-based,frame-based,ornetwork-basedmodels,butalsoinstance-basedmodelssuchascollectionsofsituation-specificexperiences.Thelatteraretypicallycapturedbycaserepresentationsandcase-basedreasoning（CBR）methods（Kolodner,1993;deMa´ntarasetal.,2005）.CBRdrawsfromresearchincognitivescienceonhumanmemoryandwasoriginallytargetedatexplainingexperience-basedbehaviorathigherlevelsofcognition（Schank,1982）.Itrestsontheassumptionthatsimilarproblemsarelikelytohavesimilarsolutions,whichcanbeusedtosolvenewsimilarproblems.TwokeyindicatorsofCBRarecontext-sensitiverepresentationscouplingaconcreteproblemsituationwithacorrespondingsolution,andsimilarity-basedsearchforreusablecases.Besidesaplethoraofapplicationsinhigher-levelcognitionsuchasmedicaldiagnosisandtherapy,design,jobschedulingandplanning,CBRhasalso,morerecently,beenappliedinhighlydynamicdomainsthatcallfornon-symbolicrepresentationssuchasnavigation,real-timestrategygames,robotsoccerandactuatormovement.AlthoughCBRhasprovidedoutstandingresultsinhigher-levelcognitivetasks,ithassofarscarcelybeenemployedforlower-levelcontrolandevenlesssoincooperationwithneuroscientistsstudyingmovementplanning,whohaveprovidedinterestingfindingsaboutinternalrepresentationpertinenttoactuator-levelplanning.ThispaperhasaspecialfocusonhowCBRisappliedatthelowerlevelsofcognition–actioncontinuum,andthroughareviewofthecurrentstateoftheart,arguesthatCBRisagenericmechanismthatcanbeusedintheentirecognition–actioncontinuum.ThenotionofrepresentationhastriggeredamajordebateinAI.Brooks（1991）,forexample,bannedrepresentationsinrobotbehaviorandmaintainedthat‘theworldisitsownbestrepresenta-tion’and‘explicitrepresentationsandmodelsoftheworldsimplygetintheway.Itturnsouttobebettertousetheworldasitsownmodel’.Weconjecturethatunificationofsymbolicandnon-symbolicAIisconditionedonareconciliationofthedifferentviewsofrepresentation,andwhere/whentheyareusedintheentirecognition–actioncontinuum.Ourstandpointisthatrepresentationsinthebrainexistformotor-levelcontrolaswellandareusedinaCBRprocess.Planninginbothhigherandlowerlevelsofcognition–actioncontinuumrelyonrepresentationspertinenttotheirowntargetdomains,althoughthetwotypesofAIhave,atthefirstsight,irreconcilableperspectivesontherepresentationsunderlyingthesemodels.FollowingMarkmanandDietrich（2000）,thispaperstressesthatthenon-symbolic‘models’shouldbeincludedinthedefinitionofrepresentationbecausetheyarestoredforlateruseandcanberecalleduponencounteringsimilarsituations.OurhypothesisisthatthecomputationalframeworkforCBRmayencompasslower-levelcognitivetaskssuchasmotorcontrol,aswellashigher-levelreasoningtasks.Thereasoningisthatactionsmaybeatvariousabstractionlevels（O¨ztu¨rk,2009）whiletheunderlyingactionselectionmechanismmaybesimilarinvitalways.Proposingmarriage,orderingabloodtest,goingtothecinemaandliftingateacupareallactions,albeitatratherdifferentlevelsofgranularity.Theideaweadvanceisthat‘actionselection’innon-deterministicdomainsincludingpurpo-sefulactuatormovementiscaseselection.ButwhyisapplyingCBRtosuchlow-levelmechanismsasmovementlearningandcontrolimportant?

Fromthebiologicalplausibilityperspective,theaforementionedresearchfindingsinneuroscienceindicatethatthemotorsystemdealsnot（only）withpurelyreactivestimulus-responsetypeofprocessesbutoftenwithstimulus-response-consequencetypeofprocesseswheretheconsequencesarecomputedthroughanticipation,animportantpartofthewholeenterprise（Herwig&Waszak,2009;Shinetal.,2010）.Thenotionofanticipationinitselfisavitalissuewithrespecttothediscussionwhethertouserepresentationsornot,providingsupportingevidencetotheubiquityofrepresentationsintheentirecognition–actioncontinuum.Furthermore,studiesinneuroscienceonimitation-basedmotorcontrolalsosupporttheexistenceofmultiplecontext-sensitiverepresentationsunderlyingactuatormovement.Theoutlineofthepaperisasfollows.Followingtheintroduction,maincharacteristicsofaCBRsystemandthetasksofthesystemdesignerareintroducedinSection2.Section3conveystherationalebehindourhypothesisthatcognitionandactionconstituteacontinuuminvolvingactionselectionatvariouslevels,whyCBRunderliestheentirecognition–actioncontinuumandwhyitprovidesabridgetounifysymbolicandsub-symbolicAI.Section4reviewsCBRsystemsthatdealwithreal-timedynamicdomainssuchasnavigation,robotsoccerandreal-timestrategicalgames.Section5discussestheroleofimitationinvariousdisciplines,withaparticularfocuson

52P.O¨ZTU¨RKANDA.TIDEMANN

neuroscience,followedbyadiscussionoftheroleofrepresentationsinhumanmotorcontrolandplanninginSection6.Section7presentsanarchitectureaboutimitation-basedmotorcontrolandlearning,suggestedbyneuroscientists.ThisisapreparationfortheotherpartofthereviewofCBRsystems,whichfocusesonthelowerlevelsofcognition–actiontasks,inSections8,9and10.Finally,Section11discusseswhythereviewedCBRsystemsstrengthenthepossibilitythatCBRplaysanimportantroleinconnectingsymbolicandsub-symbolicAI.

2MaincharacteristicsanddesigntasksofCBRsystems

ThemainargumentofusingCBRingeneralisthatitdoesnotneedanextensiveanddeepdomainmodelandreliesinsteadonexperience-based,compiledknowledge,whichhumansareknowntogatherduringandafterproblemsolving.ThedistinctiveroleofCBRinsymbolicAIisitsabilityoffastreasoningcomparedwithreasoningwithdeepdomainknowledge.Anexperienceofsolvingaproblemenrichestheexperientialmemoryofhumans（andanimals,ingeneral）and,ifproperlystored,experiencescanbeusefulinsolvingnew,butsimilarproblems.ACBRprocesshasanumberofcharacteristicsubprocesses,asshowninFigure1.