1-Case-Based-Reasoning-Foundational-Issues--Method

Publishedin:AICommunications,Vol.7Nr.1,March1994,pp39-591Case-BasedReasoning:FoundationalIssues,MethodologicalVariations,andSystemApproachesAgnarAamodtUniversityofTrondheim,CollegeofArtsandScience,DepartmentofInformatics,N-7055Dragvoll,Norway.Phone:[phone];fax:[phone];e-mail:[e-mail]’InvestigacióenIntel·ligènciaArtificial,CSIC,CamídeSantaBàrbara,17300Blanes,Catalonia,Spain.e-mail:[e-mail]-basedreasoningisarecentapproachtoproblemsolvingandlearningthathasgotalotofattentionoverthelastfewyears.OriginatingintheUS,thebasicideaandunderlyingtheorieshavespreadtoothercontinents,andwearenowwithinaperiodofhighlyactiveresearchincase-basedreasoninginEurope,aswell.Thispapergivesanoverviewofthefoundationalissuesrelatedtocase-basedreasoning,describessomeoftheleadingmethodo-logicalapproacheswithinthefield,andexemplifiesthecurrentstatethroughpointerstosomesystems.Initially,ageneralframeworkisdefined,towhichthesubsequentdescriptionsanddiscussionswillrefer.Theframeworkisinfluencedbyrecentmethodologiesforknowledgeleveldescriptionsofintelligentsystems.Themethodsforcaseretrieval,reuse,solutiontesting,andlearningaresumma-rized,andtheiractualrealizationisdiscussedinthelightofafewexamplesystemsthatrepresentdifferentCBRapproaches.Wealsodiscusstheroleofcase-basedmeth-odsasonetypeofreasoningandlearningmethodwithinanintegratedsystemarchitecture.1.IntroductionOverthelastfewyears,case-basedreasoning(CBR)hasgrownfromaratherspecificandiso-latedresearchareatoafieldofwidespreadinterest.Activitiesarerapidlygrowing-asseenbythein-creasedrateofresearchpapers,availabilityofcommercialproducts,andalsoreportsonapplica-tionsinregularuse.InEurope,researchersandap-plicationdevelopersrecentlymetattheFirstEuro-peanWorkshoponCase-basedreasoning,whichtookplaceinGermany,November1993.Itgath-eredaround120peopleandmorethan80papersonscien-tificandapplication-orientedresearchwerepresented.1.1.Backgroundandmotivation.Case-basedreasoningisaproblemsolvingpara-digmthatinmanyrespectsisfundamentallydiffer-entfromothermajorAIapproaches.Insteadofre-lyingsolelyongeneralknowledgeofaproblemdomain,ormakingassociationsalonggeneralizedrelationshipsbetweenproblemdescriptorsandconclusions,CBRisabletoutilizethespecificknowledgeofpreviouslyexperienced,concreteproblemsituations(cases).Anewproblemissolvedbyfindingasimilarpastcase,andreusingitinthenewproblemsituation.AsecondimportantdifferenceisthatCBRalsoisanapproachtoincre-mental,sustainedlearning,sinceanewexperienceisretainedeachtimeaproblemhasbeensolved,makingitimmediatelyavailableforfutureprob-lems.Thispaperpresentsanoverviewofthefield,intermsofitsunderlyingfoundation,itscurrentstate-of-the-art,andfuturetrends.ThedescriptionofCBRprinciples,methods,andsystemsismadewithinageneralanalyticscheme.Otherauthorshaverecentlygivenoverviewsofcase-basedrea-soning(Ch.1in[51],Introductorysectionof[18],[36,61]).Ouroverviewdiffersinfourmajorwaysfromtheseaccounts:First,weinitiallyspecifyageneraldescriptiveframeworktowhichthesubse-quentmethoddescriptionswillrefer.Second,weputastrongemphasisonthemethodologicalissuesofcase-basedreasoning,andlessonadiscussionofsuitableapplicationtypesandontheadvantagesofCBRoverrule-basedsystems.(Thishasbeentakenverywellcareofinthedocumentscitedabove).Third,westrivetomaintainaneutralviewofexistingCBRap-AamodtandPlaza:Case-BasedReasoningAICOMVol.7Nr.1March19942proaches,unbiasedbyaparticu-lar'school'1.Andfinally,weincluderesultsfromtheEuropeanCBRarena,whichunfortunatelyhavebeenmissinginAmericanCBRreports.Whatiscase-basedreasoning?Basically:Tosolveanewproblembyrememberingaprevioussimilarsituationandbyreusinginformationandknowledgeofthatsituation.Letusillustratethisbylookingatsometypicalproblemsolvingsituations:-Aphysician-afterhavingexaminedaparticu-larpatientinhisoffice-getsaremindingtoapatientthathetreatedtwoweeksago.Assum-ingthattheremindingwascausedbyasimilarityofimportantsymptoms(andnotthepatient'shair-color,say),thephysicianusesthediagnosisandtreatmentofthepreviouspatienttodeterminethediseaseandtreatmentforthepatientinfrontofhim.-Adrillingengineer,whohaveexperiencedtwodramaticblowoutsituations,isquicklyre-mindedofoneofthesesituations(orboth)whenthecombinationofcriticalmeasurementsmatchesthoseofablowoutcase.Inparticular,hemaygetaremindingtoamistakehemadeduringapreviousblow-out,andusethistoavoidrepeatingtheerroronceagain.-Afinancialconsultantworkingonadifficultcreditdecisiontask,usesaremindingtoapre-viouscase,whichinvolvedacompanyinsimi-lartroubleasthecurrentone,torecommendthattheloanapplicationshouldberefused.1.2.Case-basedproblemsolving.Astheaboveexamplesindicate,reasoningbyre-usingpastcasesisapowerfulandfrequentlyapp-liedwaytosolveproblemsforhumans.Thisclaimisalsosupportedbyresultsfromcognitivepsychologicalresearch.Partofthefoundationforthecase-basedapproach,isitspsychologicalplau-sibility.Severalstudieshavegivenempiricalevi-denceforthedominatingroleofspecific,pre-viouslyexperiencedsituations(whatwecallcases)inhumanproblemsolving(e.g.[53]).Schank[54]developedatheoryof1O