Goal-based Explanation Evaluation 1

Goal-basedExplanationEvaluation

DavidB.Leake

ComputerScienceDepartment

IndianaUniversity

101LindleyHallBloomington,IN47405leake@cs.indiana.edu

Iwouldliketothankmydissertationadvisor,RogerSchank,forhisveryvaluableguidanceonthisresearch,andtothanktheCognitiveSciencereviewersfortheirhelpfulcommentsonadraftofthispaper.TheresearchdescribedherewasconductedprimarilyatYaleUniversity,supportedinpartbytheDefenseAdvancedResearchProjectsAgency,monitoredbytheOfficeofNavalResearchundercontractN0014-85-K-0108andbytheAirForceOfficeofScientificResearchundercontractF49620-88-C-0058.

CorrespondanceandrequestsforreprintsshouldbesenttoDavidB.Leake,ComputerScienceDepartment,IndianaUniversity,Bloomington,IN47405-4101.

Abstract

Manytheoriesofexplanationevaluationarebasedoncontext-independentcriteria.Suchtheo-rieseitherrestricttheirconsiderationtoexplanationtowardsafixedgoal,orassumethatallvalidexplanationsareequivalent,sothatevaluationcriteriacanbeneutraltothegoalsunderlyingtheattempttoexplain.However,explanationcanservearangeofpurposesthatplacewidelydivergentrequirementsontheinformationanexplanationmustprovide.Wearguethatunderstandingwhatdeterminesexplanations’goodnessrequiresadynamictheoryofevaluation,basedonanalysisoftheinformationneededtosatisfythemanygoalsthatcanpromptexplanation;thisviewconformstothecommon-senseideathatpeopleacceptandapplyexplanationspreciselyifthoseexplanationsgivetheinformationtheyneed.Weexaminearangeofgoalsthatcanunderlyexplanation,andpresentatheoryforevaluatingwhetheranexplanationprovidestheinformationanexplainerneedsforthesegoals.WeillustrateourtheorybysketchingitsimplementationinthecomputerprogramACCEPTER,whichdoesgoal-basedevaluationofthegoodnessofexplanationsforsurprisingeventsinnewsstories.

i

1Introduction

Theuseofexplanationiscentraltotheoriesinmanyareasofartificialintelligence,suchastextunderstanding(e.g.(Granger,1980;Schank,1982;Wilensky,1983;Schank,1986;Hobbsetal.,1990)),planrepairandindexing(e.g.,(Hammond,1989)),andguidinggeneralization(e.g.,(Mitchelletal.,1986;DeJong&Mooney,1986)).Inaddition,recentexperimentshavesupportedsomeoftheseexplanation-basedprocessesaspsychologicalmodels(Ahnetal.,1987).However,thebenefitsofexplanation-basedprocessingdependontheexplanationstakenasstartingpoint.Inreal-worldsituations,manycandidateexplanationsmaybeavailableforasingleevent,promptingthequestionofhowtoselecttheexplanationonwhichtobasefurtherprocessing.

Sinceexplanationscanbeusedinmanyways,itseemsreasonablethatratherthanseekingauniversal“best”explanation,anexplainershouldtailorexplanationtowardsservingthegoalforwhichitisintended.Nevertheless,theinfluenceofcontextonexplanationhasreceivedlittlestudyinpsychologyandartificialintelligence.Inpsychology,thecentralcurrentforresearchonpeople’schoiceofexplanationsisattributiontheory(Heider,1958),whichgenerallyaccountsforchoiceofexplanationswithoutreferencetowhatmotivatedtheexplanationeffort.Likewise,inartificialintelligence,criteriaforexplanations’goodnesstendtotakeacontext-independentview,basingtheirjudgementsoncriteriathatmakenoreferencetothegoalsoverarchingtheexplanationeffort(e.g.,(Granger,1980;Wilensky,1983;Pazzani,1988;Rajamoney&DeJong,1988;Thagard,1989;Ng&Mooney,1990;Hobbsetal.,1990)).Thesetheoriesfocusonchoosingtheexplanationsmostlikelytobevalid.However,eveniftheysucceedinidentifyingexplanationsthatarevalid,theymaynotgivesufficientinformationtoselecttheexplanationthatbestsatisfiestheexplainer’sgoals.Forexample,supposethattheeventtoaccountforisacompanybankruptcy.Supposefurtherthatthecompanywasbankruptastheresultoftwofactors,eitherofwhichwouldhavebeensufficienttocausethebankruptcy:thecompany’sbadmanagement,andexcessivelocaltaxes.Consequently,forthisexamplethereareatleasttwovalidexplanations:“thebankruptcywascausedbybadmanagement”and“thebankruptcywascausedbyhightaxes.”

Bankruptcieshaveadverseeffectsonthecommunitiesinwhichtheyoccur,soalocalpoliticianmighttrytoexplainthebankruptcy,withtheunderlyinggoalofpreventingitsrecurrence.Forthisgoal,therelativevalueoftheexplanationswouldbequitedifferent.“Thebankruptcywascausedbybadmanagement”isunhelpfultothepolitician,sinceitsuggestsnocourseofaction—

1

governmenthasnocontrolovermanagementinlocalindustries.However,“thebankruptcywascausedbyhightaxes”couldbethesourceofthegeneralizationthatextremetaxratesarelikelytoforcecompaniesintobankruptcy,whichcouldbeusefulasmotivationforthepoliticiantolowertaxes.Forashareholderinacompanyinalow-taxstate,theusefulnessofthetwoexplanationsisreversed:theshareholderbenefitsmorefromconsideringthedangersofbadmanagers,sincebadmanagementmightcausecompanystocktolosevalue.

Wemightthinkthateitherexplainercouldsimplyrequirethatanexplanationidentifyallcausallyrelevantfactors,inthatwayassuringthatallimportantconditionsaretakenintoaccount.However,real-worldeventsaresimplytoocomplicatedforcompleteexplanation.Forexample,anycompanybankruptcyactuallydependsoncountlessfactors,suchascompetingproducts,con-sumerdemand,thecompany’slaborsituation,itspasthistory(affectingitscashreserves),theowners’managementstyle,andthecurrentbankruptcylaws.Inourview,goodexplanationsmusthighlightafewimportantfactorsfromthemanythatarecausallyinvolved:thosethatgivetheexplainertheinformationitneeds.Sincenoexplanationcanbeexhaustive,goal-basedexplana-tionevaluationisneededforexplainerstorecognizewhethercandidateexplanationsaddresstheappropriatefactors.Insufficientexplanationscanberejected,orelaborateduntiltheyprovidetheneededinformation.

Inwhatfollows,wedevelopatheoryofgoal-basedexplanationevaluation.Weelaborateontheroleofoverachinggoalsinexplanation,demonstratinghowdifferentgoalsleadtodifferentrequirementsforexplanations’goodness,andhowthesecanbecharacterizedalongasmallsetofdimensions.Wethendescribetheimplementationofgoal-basedevaluationinacomputermodel,ACCEPTER,whichillustratesourtheory(Leake,1988a,1988b,1989b,inpress).Wearguethatgoal-basedevaluationimprovesanexplanation-basedsystem’scapabilitytofocusexplanation,toguideexplanationconstruction,andtofunctioneffectivelydespitesomeimperfectionsinitsdomaintheory.

2Overview

Webeginbysamplingpriorexplanationevaluationapproachesfrompsychology,philosophy,andartificialintelligence.Inbothpsychologyandinartificialintelligencethesetheoriestendeither

2

tolookattheevaluationprocessascontext-independent,ortoexamineitwithinasinglefixedcontext.Somephilosophicalaccountshavetakenaviewclosetoours,arguingthatexplanationdependsoncontext,andwegiveasamplingofsomeofthoseperspectives.

Wetheninvestigatehowaspecificaspectofcontext,theoverarchinggoaloftheexplainertouseanexplanation,determinestheinformationthattheexplainerrequireswhenconfrontedwithananomaloussituation.Weshowthatgoal-basedevaluationpermitsamoreperspicaciouschoicebetweencandidateexplanationsthanpriorapproaches,andalsoextendstherangeofexplanationsanexplainercanuse,byenablingprincipleduseofpartialexplanations.

Weconcentrateonevaluationofexplanationsofanomalousevents;thoseeventsoftenpromptnewgoalsbecausetheyrevealunexpectedaspectsofasituation.Thenewgoalspromptplans,whosesuccessdependsonappropriateinformation,inturnpromptingexplanationpurposestoextractparticularinformationfromanexplanation.Inordertocharacterizetheinformationthatexplanationpurposesrequire,wedevelopasmallsetofevaluationdimensionsfortestingtheuse-fulnessofhypothesizedcausesofanevent,andthatareusedtobuildchecksfortheexplanationpurposes.Thusourmodelaccountsforexplanationevaluationasfollows:

Anomaly

Goals

Plans

Explanationpurposes

Dimensionchecks

Evaluationdecision

Inthispaper,weconcentrateourdiscussiononexplanationpurposesandevaluationdimensions.

Weshowthatchecksalongdifferentcombinationsofthoseevaluationdimensionsaresufficienttoevaluateexplanationsforawiderangeofexplanationpurposes.WesketchACCEPTER’spre-liminaryimplementationofchecksforcertainevaluationdimensions,andshowhowitsevaluationdynamicallyreflectschangesincontextandinoverarchinggoals.Finallywearguethattheef-fectivenessofanyexplanation-basedprocessingdependsontheabilitytoassurethatexplanationssatisfytheneedsforinformationthatarisefromsystemgoals.

3

3Previousperspectives

3.1Psychologicalapproaches

MuchpsychologicalresearchonchoiceofexplanationsoriginatedinHeider’sseminalworkonattributiontheory(Heider,1958).Attributiontheoryinvestigateshowpeopledecidewhethertoexplainanactionintermsoffeaturesofitsactor,orfeaturesoftheenvironment.(Mostworkinattributiontheoryassumesthateitherpersonalorsituationalfactorswillapply,butnotboth.)OneimportantresultisKelley’scovariationprinciple,whichgivesahypothesisforhowpeoplemakethedecisionbetweenattributinganoutcometopersonalorsituationalfactors(Kelley,1967).Thecovariationprinciplesuggeststhatpeoplelookatcovariationacrossdifferenttime,people,andotherentitiesinordertodecidewhichtypeoffactorapplies.Forexample,ifJohndislikesamovie,butmostotherviewersareenthusiastic,Kelley’scovariationprinciplesuggeststhatJohn’sdislikeshouldbeexplainedbyaspectsofJohn,ratherthanaspectsofthemovie.

Althoughattributiontheorygivescriteriafordecidingwhichclassoffactorstoimplicate,itdoesnotsuggesthowtodecidewhichparticularpersonalorenvironmentalfactorsareimportant.Intheexampleofthemovie,itsaysthatagoodexplanationmustinvolvesomeaspectofJohn,butdecidingwhichisbeyonditsscope.ThisproblemwaspointedoutbyLalljeeandAbelson,whoobservethatpeoplewouldusuallytrytofindamorespecificreasonforthedislike(Lalljee&Abelson,1983).Forexample,someonewhohadinvitedJohntothemovie,expectinghimtolikeit,wouldprobablytrytodetermineJohn’sparticularobjections.Themorespecificinformationismoreuseful:itallowspredictingJohn’sreactionnexttime,toavoidinvitinghimtoanotherinappropriatemovie.

Attributiontheoryalsofailstoconsidertheeffectofcontextonpreferencesforexplanations.Thecovariationprincipledoesnottakeanexplainer’spriorexpectationsintoaccount,orthereasonforexplaining.However,(Lalljeeetal.,1982)showsthattheexplanationspeopleseek,ratherthanbeingdeterminedbyabstractcriteria,varywithcircumstances:unexpectedbehaviorrequiresmorecomplexexplanationsthanexpectedbehavior,andislikelytorequiremoreofbothsituationalandpersonalelements.

4

Aknowledgestructureapproach:LalljeeandAbelsonquestionattributiontheory’sbasicap-proach,andsuggestthatitsproblemscanbeovercomebyadoptingaknowledgestructureapproachtoattribution.Theydistinguishtwotypesofexplanation:constructiveandcontrastive.Inconstruc-tiveexplanation,peopleexplaineventsintermsofknowledgestructuressuchasscriptsandplans(Schank&Abelson,1977).Incontrastiveexplanation,theyexplainthembyshowingwhyeventsdeviatedfromexpectationsprovidedbyknowledgestructures.Forexample,“Johnlefthisbicycleunlocked”mightbeexplainedintermsofgoalreversal:perhapsratherthanhavingthenormalgoalofwantingtoprotectit,heactuallywantedtogetridofit.

Ourtheoryfollowsthebasiclinesofthisapproach,accountingforanomaliesthroughbothconstructiveexplanation(tounderstandthesurprisinginformation)andcontrastive(toaccountforthefailureinpreviousexpectationsorbeliefs).Inaddition,welookathowpreferenceforexplana-tionsisaffectedbygoalsbeyondthegeneraldesiretofillknowledgegaps,tofurtherspecificaimsoverarchingtheunderstandingprocess.

Excusetheory:Researchonattributionhasexaminedtheinfluenceofoneclassofoverarchinggoalontheattributionprocess:thegoaltoabsolvetheexplainerfromblame.Excusetheorystudieshowthedesiretoformexcusesmakespeoplemanipulatethetypesoffactorstouseinattribution,toblameexternalinfluencesfortheirownbadperformance(forexample,(Mehlman&Snyder,1983)or(Snyderetal.,1983)).Excusetheory’sdemonstrationthatgoalsinfluenceexplanationisconsistentwithourview.However,weconsiderevaluationwithinaknowledgestructureframework,ratherthanattributiontheory,inordertoaddressthespecificknowledgerequiredfromanexplanation,andconsidertheinfluenceofawiderrangeoverarchinggoals.

3.2Philosophicalviews

Muchofthephilosophicalinvestigationofexplanationhasattemptedtoestablishformalrequire-mentsforexplanations,independentlyofcontext,butsomeapproacheshavetakenaviewmuchclosertoourgoal-basedperspective.Forexample,Hansonpointsoutthatdifferentexplainerswouldfocusondifferentaspectsofafatalvehicleaccident:

Thereareasmanycausesofxasthereareexplanationsofx.Considerhowthecause

5

ofdeathmighthavebeensetoutbyaphysicianas‘multiplehaemorrhage’,bythebarristeras‘negligenceonthepartofthedriver’,byacarriage-builderas‘adefectinthebrakeblockconstruction’,byacivicplanneras‘thepresenceoftallshrubberyatthatturning’.(Hanson,1961,page54)

Likewise,Mackie(1965)discussestheneedofexplanationstoaidinmakingthedistinctionsimportanttoaparticularexplainer.Inthisview,explanationisconductedagainstthecausalfieldofsituationstobedistinguishedbytheexplanation,andvarieswiththatfield.Thismakesrequire-mentsforexplanationstronglycontext-dependent:anexplanationof“whydidthecarcrashgoingaroundtheturn?”wouldbedifferentifthecausalfieldwereotherturns(inwhichcasetheexpla-nationmightbetheturn’ssharpness),orifitwereotherinstancesofthecargoingaroundthesameturn(inwhichcasetheexplanationmightbethatthistime,thedriverwassleepy).VanFraassen(1980)arguesinfavorofHanson’sview,andaccountsforthechoiceofcauseswithanideasimilartothecausalfield.However,ourapproachdiffersinexaminingtheneedsthatgeneratetheback-groundforexplanation,andshowinghowparticulargoalsdetermineconcreterequirementsthatanexplanationmustsatisfy.

3.3AIapproaches

Inartificialintelligence,thequestionofexplanations’goodnesshasbeeninvestigatedinthreemainareas.Researchinexpertsystemsexplanationhasconcentratedonthequestionofexplanations’goodnessforexplainingsystembehavior,forthebenefitofthesystemuser;researchinexpla-nationfortextunderstandinghasconcentratedonhowtoselectvalidexplanationsfromarangeofhypotheses;andexplanation-basedlearningresearchhasprimarilyconsideredtheproblemofdeterminingexplanations’goodnessforlearningtoimproveperformanceonconceptrecognitionandsearch.Wediscussbelowtheapproachesofeacharea.3.3.1Expertsystemsexplanation

Researchonexpertsystemsexplanationhasinvestigatedtheproblemofgeneratingexplanationsofexpertsystembehaviorthataresufficienteithertoeducatetheuseraboutthetaskdomain,orto

6

justifysystemdecision-making(Shortliffe,1976).Thedesiretoprovidetheneededinformationhasprompteddevelopmentofmanyexplanationsystems,includingsystemsthatcanshownotjusttheirdecisionpaths,butthereasoningunderlyingthosepaths(Swartout,1983);systemstodeviseexplanationsthatareappropriatetotheuser’spriorknowledgelevel(Paris,1987);andsystemsthattreattheexplanationprocessasacontinuingdialogue,allowingclarificationsandelaborationstobeofferedinresponsetofollow-upquestions(Moore&Swartout,1989).Ourinvestigationtakesacomplementaryview,somewhatclosertothatoftheuserofsuchsystemsthanthesystemsthemselves:weexaminehowanunderstanderwithparticularknowledgeandgoalscandecidewhetheragivenexplanationissufficient,orwhethertoaskadditionalquestions.3.3.2Selectingexplanationsforunderstanding

Asdiscussedintheintroduction,muchAIresearchonexplanationevaluationhasbeenaddressedtowardsdecidingexplanations’likelyvalidity.Forexample,researchonexplanationinstoryun-derstandinghasreliedprimarilyonfixedstructuralcriteriaforchoosingbetweencompetingex-planations,suchasfavoringexplanationsinvolvingshortexplanatorychains,orfavoringexplana-tionswiththemoststructuralcoherence(e.g.,(Granger,1980),(Wilensky,1983),(Ng&Mooney,1990)).

However,theexamplesinourintroductionshowthatvalidityaloneisnotenoughtoassureanexplanation’sgoodness:anexplanationmaybevalidwithoutbeinguseful.Infact,validitymayactuallybeundesirableinthecontextofcertainexplainergoals,asisshownbythefollowingscenefromthemovieBreakingAway:

Aused-carsalesmanistakingaprospectivebuyeroutonatestdrive.Hestopssud-denlytoavoidabicyclist,andthecardies.Hetriesfranticallytostartit,buthecan’t.Heexplainstothebuyer:“Imusthaveputexpensivegasinitbymistake.Thisbabyjusthatesexpensivegas.”

Althoughthesalesmanknowshisexplanationisinvalid,heselectsitbecauseitserveshispurposebetterthanthetrueexplanationofthecar’sbadcondition.Notonlydoesitdivertblamefromthe

In(Leake,inpress),weargueforevaluatingexplanations’validityintermsofcontent-basedcriteriaratherthansolelystructuralones,andpresentacontent-basedapproachtodecidingvalidity.

7

car’sconditiontothesalesman’smistake,butitalsointroducesanewfactortoswaytheprospectivebuyertowardsbuyingit:thatthecarisinexpensivetooperate.3.3.3Operationalityinexplanation-basedlearning

Explanation-basedlearning,inwhichanexplanationisusedtoguidegeneralization,isapowerfulmeansofcategoryformation,oftenallowinglearningfromsingleexamples(Mitchelletal.,1986;DeJong&Mooney,1986).Thedesiretoassurethatusefulconceptsareformedhaspromptedstudyofwhatdeterminesanexplanation’soperationality—itssuitabilityforusebythesystemperfor-manceelement.(Thenotionofoperationalityoriginatedwithasomewhatdifferentformulationin(Mostow,1983).)

InmostEBLresearch,operationalitycriteriaaimedtowardsefficientrecognitionofconceptinstances,orcontrolofsearchinproblemsolving.Criteriausedforjudgingtheoperationalityofaconceptformulationrangefromstaticannotationofpredicateswiththeireaseofevaluation(Mitchelletal.,1986),todynamictechniquesbasedonsystemknowledge(DeJong&Mooney,1986),totechniquesdirectlyreflectingtheutilityofanexplanation,usingestimatesandactualmeasurementsofrecognitioncostversusbenefittothesystem(Minton,1988).

Researchonoperationalityforconceptrecognitionhasledtotheidentificationofsignificantcharacteristicsofoperationalityforthattask,suchastheoperationality/generalitytradeoff(e.g.,(Segre,1987)).Becauseoperationalityhasbeenstudiedforsofewtasks,therehasbeensometendencytoassumethatsuchpropertiesapplytooperationalityjudgementsforanytask.However,(Keller,1988)pointsoutthatexistenceofanoperationality/generalitytradeoffdependsonspecificassumptionsunderlyingtherecognitiontask,andthatthetradeoffdoesnotexistforeveryuseofexplanations—amoregeneralexplanationmayactuallybemoreoperationalaswell.Forexam-ple,ateachergeneratinganexplanationforastudentmayfindamoregeneralexplanationmoreoperational,forthetaskofextendingtherangeofassignmentsastudentcansolve.Thusinordertounderstandoperationality,weneedtoconsiderthewidevarietyoftasksforwhichexplanationsareused.

Keller’sprogramMetaLEX(Keller,1987)reflectsthedynamicnatureofoperationalitybynotrelyingonfixedoperationalitycriteria.Instead,itsoperationalitycriterionisaninputtothesystem,andcanreflectcurrentgoalsandgoalpriorities.Consequently,thespiritofthisworkis

8

quitesimilartoours,thoughitdoesnotaddressthequestionofidentifyingtherangeofpossiblepurposesforexplanation.(Southeretal.,1989)alsopresentsanargumentcloseinspirittoours—thatitisessentialtobeabletogenerateexplanationsfromagivenviewpoint—andidentifiesclassesofexplanationsthatstudentsmightseekwhenstudyingcollege-levelbotany.However,sincethatworkconcentratesontutoringapplications,itdoesnotneedtoconnecttheclassestoover-archinggoalsthatmakethemimportant,beyondsimplydoingwellinacourse.

Theaboveapproachesconcernthelevelofelaborationtousewhenexplainingagivenconcept.Althoughmanyexplanation-basedsystemstakeasinputtheconcepttoexplain,approacheshavealsobeenadvancedtoselectusefulconceptstoacquire.Theyarebeyondthescopeofthispaper;foradiscussionofsomeofthoseissues,see(Kedar-Cabelli,1987),whichexamineshowthepurposeoffindingobjectstouseinplansguidesselectionofconceptstoexplain,or(Schank,1982;Riesbeck,1981;Hammond,1989;Leake,1991),whichaddresshowtheneedtorepairanexpectationfailurepromptstheexplanationeffort.

4Atheoryofgoal-basedevaluation

Ourworkconcentratesonevaluatingexplanationsgeneratedforanomalousstatesoreventsthatanunderstandernoticesineverydayunderstanding—conflictswithitsbeliefsandexpectations.Theseanomaliesshowthattheunderstander’sworldmodelisflawedinsomeway,sincetheworlddiffersfromitsexpectationsorbeliefs.Inordertorespondappropriatelytoprotectitsgoals,andtotakeadvantageofchangesintheworld,theunderstanderneedsadditionalinformation.Forexample,someonefiredfromajobmayhavemanyreasonstotrytodiscoverwhythefiringoccurred,suchasavoidingfuturefirings,protestingthecurrentone,orpreparingtocounterbadreferenceswhenapplyingfornewemployment.Toillustratetherangeofpurposesthatmayarise,andtheireffectonexplanation,weconsiderthefollowingexample:

CompanyXwasbeleagueredbyhightaxes,foreigncompetition,andout-datedequip-ment,despitelowlaborcostsduetobeingnon-union.Rumorsofproblemsspread,andthecompany’sstockplummeted,butitsmanagersannouncedtheirdecisionnottohavelayoffs.Thenextweek,itwasrumoredthattheywouldlayoff20percentoftheirworkforce.

9

Becausethelayoffsviolatethemangers’pledge,theyareanomalous,andmightpromptexplana-tion.Actorsinvolvedintheincidentwouldbelikelytohavegoalsaffectedbyit,andachievementofthosegoalswouldoftenrequireparticularinformation.Theseinformationrequirementsdeter-minewhichexplanationsaresufficient:

1.Someonewhoknewaboutthemanagers’pledge,andconsequentlydidnotbelievethattherewouldbelayoffs,wouldwanttoverifythatthelayoffswouldreallyoccur.Thegoalofmaintainingaccuratebeliefspromptsaplantosubstantiatethelayoffs.Thispromptstheexplanationpurposeofconvincinglyconnectingthelayoffstotrustedinformation.Forthispurpose,asatisfactoryexplanationmightbeanewspaperfoundasecretcompanymemofromthecompanypresident,describingthetimetableforthelayoffs.

2.Thesamepersonmighthavethegoalofavoidingfutureincorrectpredictionsinsimilarsituations.Oneplantoachievethatgoalistoexplainwhythecurrentbeliefwentwrong,andrepairthesourceoftheproblem.Herethepurposeofexplanationwouldbeaccountingforthebadpredictionintermsoffalsepriorbeliefs.Iftheproblemwasthatthemanagerhadbeenbelievedtrustworthy,butwasactuallydishonest,thatexplanationwouldaccountfortheexpectationfailure,andpreventtheexplainerfrombeingmisleadbythemanageragain.3.Aworkerwhowassurprisedtobelayedoffmightwanttoavoidbeingunemployed,byfindinganewjobbeforebeinglayedoffnexttime.Forthispurpose,asuitableexplanationmightbethelayoffswereinevitablebecauseofpressuretoreducecoststoshoreupthecompany’sfallingstock:giventhatthisisavalidexplanation,thedangeroffuturelayoffscouldbepredictedbywatchingthestockprice.

4.Inalocalpolitician,thelayoffsmightpromptthegoalofimprovingthearea’seconomichealth,byendingthelayoffs.Thisgivesrisetotheexplanationpurposeoffindingfeasiblerepairpointsforthecurrentsituation,byexplaininghowthelayoffsarecausedorenabledbyfactorsthatareundergovernmentcontrol,andwhoseremovalwillrestorethedesiredsituation.Forexample,iflayoffsresultedfromhightaxes,apossibilitywouldbetolowertaxesinordertomakethefactoryagainprofitable,andhavetheworkerscalledback.5.Aworkerwhowonderedwhethertolookforanewjob,orsimplywaittobecalledback,wouldwanttoknowhowlongthelayoffswerelikelytolast.Thiswouldprompttheexpla-10

nationpurposeofclarifyingthesituationtohelpformpredictionsofthelayoffs’duration,whichmightinvolvefindingifthelayoffsresultedprimarilyfromlong-termfactors,suchasforeigncompetition,versusshort-termones,suchastemporaryoverstocks.

6.Aworkerstillemployed,whowantedhisjobtobemoresecure,mightwanttopreventfuturelayoffs.Thiswouldprompttheexplanationpurposeoffindingpotentialcausesofthelayoffsthattheworkercanaffect.Ifthefactory’spreviouslackofunionizationresultedinacontractthatgaveemployeesnosecurity,enablingthelayoffs,theworkercouldrespondbystartingaunion.(Therequirementsforpreventinganeventarenotnecessarilythesameasforitsrepair:unionizationislikelytobeaneffectivepreventativeforlayoffs,butafterworkershavealreadybeenlayedoff,anewunionisunlikelytomakethecompanyrescindthelayoffs.)7.Themanagerwhoorderedthelayoffsmightwanttoavoidnegativepublicity,anddecidetodosobydeflectingblame.Oneexplanationpurposeforthisgoalwouldbeidentifyingotheractors’contributionstotheoutcome,asbyexplainingtherolethatanoutsideexperthadinshapingthedecision.

8.Theownerofanotherfactorymightwanttoimproveitsprofitability.Thisgoalmighttriggertheplanofanalyzingothermanagers’decisions,tolearnbettermanagementstrategies.Thiscouldprompttheexplanationpurposeofdecidingwhytheothermanagerschosetohavelayoffs,asopposedtoalternativeresponsestothecompany’sproblems,suchasincreasingadvertisingtoincreasedemandforproducts.Accountingfortheirsurprisingchoicemightshowimportantfactorsthattheownerwouldneedtoconsiderinsimilarfuturedecisions.Todeterminethosefactors,theownerwouldhavetheexplanationpurposeoffindingparticulartypesofcauses:thegoalsandgoalprioritiesthatenteredintothedecisionoflayoffsversusadvertising.

9.Abusinessconsultantmightwanttodevelopatheoryofhowdemographictrendsforcelayoffsindifferentindustries.Aplanforthatgoalistoshowhowthetheoryaccountsforparticularepisodesoflayoffs,whichwouldprompttheexplanationpurposeoffindingcausesofthecurrentlayoffsthatarerelevanttothetheory.

Eventhoughalltheaboveactorsareexplainingthesameevent,eachhasdifferentneedsforinformation,promptingadifferentexplanationpurpose.Althoughwecanimaginesingleexplana-11

tionsthatwouldbeusableformultiplepurposes,anexplanation’susefulnessformultiplepurposesisnotassured:noneofthesampleexplanationsaresufficientforanypurposesbeyondtheoneeachillustrates.

Ourexamplesidentifyninemajorexplanationpurposes.Althoughwedonotclaimthattheyformanexhaustivelist,thesepurposesaresufficienttodemonstratethatthereisawiderangeofgoal-basedpurposesforexplanation,including:1.Connecteventtoexpected/believedconditions.2.Connecteventtopreviouslyunexpectedconditions.3.Findpredictorsforanomaloussituation.

4.Findrepairpointsforcausesofanundesirablestate.5.Clarifycurrentsituationtopredicteffectsorchooseresponse.6.Findcontrollable(blockableorachievable)causes.7.Findactors’contributionstooutcome.

8.Findmotivationsforanomalousactionsordecisions.9.Findawithin-theoryderivation.

Ourbasicmodelofgoal-drivenexplanationisthatanomaliesshowthattheworldisdifferentfromexpected,sothatgoalsandplansmayneedtobere-considered.Thedesiretoprofitfromthesituation,ortoavoidbadeffectsthatmightresult,triggersselectionofnewgoals,whichtriggerplanstoachievethem.Theseplansgenerateneedsforinformation.Basedonthoseneeds,anexplainerhasanexplanationpurposetoconstructanexplanationreflectingcertainaspectsofthesituation,andevaluationmustconfirmthatthoseaspectsareincludedintheexplanation.Table1sketchesthisprocess,showinghoweachofthepurposesabovecanarisefromageneralgoal,andastrategytoachievethegoal.Eachexplanationpurposedeterminesspecifictypesofinformationthatanexplanationmustprovide.

12

Sometimesananomalywillsuggestmorethanonegoal,causingmultipleexplanationpur-posestobeactivesimultaneously.Forexample,ifsomethinghappensthatisbothsurprisinganddesirable,theexplainermightwanttofindhowtorecognizewhenitislikelytooccuragain,toprofitfromit,andtoknowwhomtocreditwithbringingitabout,toencouragethatpersontobringitaboutagain.Ourtheorydealsonlywithevaluationofexplanationsoncethepurposehasbeenselected;thesequenceofstepsfromanomalydetectiontonewgoals,andplansgivingrisetorequirementsforinformationtoachievethem,isatopicforfutureresearch.

Thefollowingsectionsdiscusstheimplementationofevaluationcriteriainacomputersystemthatjudgesexplanationsforarangeofexplanationpurposes,acceptingthemiftheyincludetheinformationneededforthosepurposes.

5Acomputermodel

ACCEPTERisastoryunderstandingprogramthatrequestsexplanationswhenitdetectsanomalies—conflictsbetweennewinformationanditspriorbeliefsorexpectations.Sinceananomalouseventindicatesthatthesystem’sunderstandingisincorrect,itsignalsthattheremaybeunexpectedrisksoropportunities,andpromptsnewgoalsforanunderstander.GoalstoguideevaluationaregiventoACCEPTERasinput,anditevaluatesexplanations’appropriatenessforthosegoals.Thesys-temprocessesabout20simplestoriesofanomalousevents,includingstoriesofthespaceshuttleChallenger’sdisaster,thewarshipVincennes’accidentalshootdownofanIranianairliner,celebritydeaths,andautomobiledefects.Fortheseanomalies,itevaluatesatotalofabout30explanations.Theappropriatenessofeachexplanationcanbeevaluatedforpredictingsimilarevents,preventingtheirrecurrence,repairofthecurrentsituation,andassignmentofblameorresponsibility.Theprogramhasbeenusedbothasastand-alonesystem,andastheexplanationevaluationcomponentinSWALE,astoryunderstandingsystemthatusesinformationfromACCEPTERtoguideexplanationofnoveleventsinthestoriesitprocesses(Schank&Leake,1989;Leake&Owens,1986;Kass,1986;Kass&Leake,1988).SWALEusestheresultantexplanationstoac-countfornovelevents,toformnewexplanatoryschemas,andtoguideindexingofthoseschemasinmemory.

AfterdiscussingACCEPTER’sevaluationcriteria,wereturntoACCEPTER’sroleinSWALE,

13

Table1

Asketchofhowthenineexplanationpurposesarise.Goalstodealwithasurprisingsituationtriggerplansforwhichinformationisneeded,promptinganexplanationpurpose.

Goal

Plan

ExplanationpurposeConnecteventtoexpected/-believedconditions.Connecteventtopreviously

unexpectedconditions.Findpredictorsforanoma-loussituation.Findrepairpoints.

Clarifysituationtopredictef-fects.

Findachievablecauses.Findactors’contributionstooutcome.

Findmotivationsforadver-sary’sunexpectedactions.Buildwithin-theoryderiva-tion.

Preventbadeffectsofactingonfalseinformation.

1.Confirmreasonablenessofnewinformation.

2.Findsourceofflawinpre-viousinformation.

Predictsimilareventsintimetoprepare.Executerepair.

Dealwithimportantramifica-tionsofanomaly.Re-causeevent.

Punishcurrentactorstodeterfutureperpetrators.

Predictandrespondtohisac-tions.

Usetheorytoaccountforun-expecteddata.

Minimizebadeffects/maxi-mizegoodeffectsinsimilarfuturesituations.

Usemalfunctioningdevice.Protectcurrentplans.Re-achievethegoodeffectscausedbyanomalousevent.Preventrecurrenceofsurpris-ingbadstate.Counteradversary.Refine/demonstrateatheory.

14

toexaminetheadvantagesofgoal-basedevaluationtoanexplanation-basedunderstandingsystem.

5.1ACCEPTER’sbasicalgorithm

ACCEPTERtakesasinputastoryrepresentedintermsofConceptualDependencytheoryprim-itives(Schank,1972),orintermsofschemaspackagingsequencesofthoseprimitivestorep-resentstereotypedsequencesofactions.Forexample,aschemamightrepresenttheeventsin-volvedineatingatarestaurant,suchasentering,beingseated,beingbroughtmenus,ordering,etc.TheseschemasarerepresentedinACCEPTER’smemoryasmemoryorganizationpackets(MOPs)(Schank,1982).

ACCEPTERprocessesstoriesonefactatatime,updatingitsbeliefsandgeneratingexpecta-tionsforlaterinputsfromthestory.Forthisroutineunderstanding,ACCEPTERusesschema-basedunderstandingprocessmodelledon(Cullingford,1978),integratingnewinformationintoadynamicmemory(Schank,1982;Lebowitz,1980;Kolodner,1984).Asitintegratesinputfactsintomemory,itchecksforanomalies—conflictsbetweentheinputsanditsbeliefsorexpectations.ACCEPTERimplementsatheoryofpattern-basedanomalydetectionthatallowsproblemstobedetectedwithcontrolledinference,butthatprocessisbeyondthescopeofthispaper.See(Leake,1989a)foranoverview,or(Leake,inpress)foramorecompleteaccount.

Whenananomalyisfound,ACCEPTERpresentstheanomalytotheuser,alongwithpossibleexplanations.Theseexplanationsareformedbyretrievingandinstantiatingexplanationpatterns(XPs)(Schank,1986)fromanXPlibraryinACCEPTER’smemory.Explanationpatternsrepresentbothspecificexplanationsofpriorepisodes,andmoregeneralstereotypedcommonsenseexplana-tions,suchasIfacarwon’tstart,itmayhaveadeadbattery,orIfastudentfailsatest,itmaybebecauseofnotstudyingenough,orIfacarhasadefect,itmaybebecauseofmanufacturer’sbadqualitycontrol.(Wedescribethestructureofexplanationpatternsbelow.)

FromthelistpresentedbyACCEPTER,theuserselectsacandidateexplanationpatterntoin-stantiate,andinputsanexplanationpurposetobetakenintoaccountwhenjudgingtheinformationthattheexplanationprovides.ACCEPTERthenperformstheevaluation,signallinganyproblemsitencounters.IntheSWALEsystem,ACCEPTER’sproblemcharacterizationsareusedtoindexintoalibraryofexplanationmodificationstrategies,toselectastrategyappropriatetotheproblem

15

(Kass,1986).Thestrategyisthenapplied,anditsresultre-evaluatedtoidentifyproblemsthatremain,ornewproblemsintroducedbytherepair.Inthestand-alonesystem,theusercanrespondtoACCEPTER’sproblemreportbychoosinganalternativeexplanationthatavoidstheproblem,orselectingdifferentrole-fillersforanyuser-instantiatedrolesintheexplanation.Forexample,oneexplanationACCEPTERevaluatesfortheChallengerexplosionisthatRussiamighthavesab-otagedthelaunch.ACCEPTERsignalsaplausibilityproblem,sinceRussiawouldnotrisksuchadangerousconfrontationwiththeUnitedStates.AfterACCEPTERpointsoutthatproblem,theusercanre-instantiatetheexplanationwithanothersaboteur.IftheuserselectsLibya,ACCEPTERconsidersthenewexplanationmorereasonablebecauseofLibya’swillingnesstotakerisks.Figure1summarizesACCEPTER’sbasicunderstandingprocess.

5.2Representationofexplanations

Explanationpatternstracethereasoningneededtoaccountforanevent.Theyrepresentexplana-tions’structureinbelief-supportchains,whicharebeliefdependencynetworksthatincludefourcomponents:initialstatesorevents,hypothesizedasleadingtoanevent;internalbeliefs,inferredfromtheinitialhypothesesonthewaytotheconclusionstobederived;theconclusionsthemselves;andplausibleinferencelinkstracingtheinferenceprocessfrominitialbeliefstoconclusions.Forexample,theXPearlydeathfromlifeinthefastlanetracesthecircumstancesleadingtothedeathofstarssuchasJanisJoplinandJohnBelushi.Itsbelief-supportchainstartswithtwoba-sichypotheses:thatthedeceasedwasaperformer,andwasverysuccessful.OneofACCEPTER’sinferencerulesencodesthetendencyforstarperformerstobeunderconsiderablestress.Start-ingwiththeinferredbeliefthattheperformerisunderhighstress,thedesiretoreducestresscanbeinferred,andfromthisdesirethetendencytotakerecreationaldrugsasanescape.Takingrecreationaldrugssometimesleadstoanaccidentaldrugoverdose,fromwhichdeathisalikelyoutcome.Figure2isaschematicdiagramofthebelief-supportnetworkassociatedwiththisrea-soningchain.

Wecanimaginesituationsinwhichanylinkintheabovereasoningwouldfailtohold.Forexample,stressdoesnotnecessarilyleadtotakingdrugs,notdoesrecreationaldrugusenecessarilyleadtoanoverdose.Eachofthelinksissimplyaplausibleinferencelink,showinghowthe

16

Figure1:ACCEPTER’sbasicunderstandingprocess.

17

Figure2:ACCEPTER’sbelief-supportchainfortheXPearlydeathfromlifeinthefastlane.

18

hypothesesmaketheconclusionmorelikely.Foradetaileddescriptionofthestructureofbelief-supportchains,andthetypesoflinksused,see(Schank&Leake,1989).

6Explanationevaluationforroutineunderstanding

ACCEPTER’sexplanationevaluationisguidedbytwotypesofgoals.First,itsevaluationservesthebasicgoalofanunderstandertoaccountforeventsinthestoriesitprocesses,andtomaintainaccuratepredictions.Second,itservesdynamicallychanginggoalsbeyondroutineunderstanding,byevaluatingwhetherexplanationsprovidetheinformationnecessarytoachievethem.Inthissec-tion,wediscusshowACCEPTER’sevaluationcriteriajudgewhetheranexplanationprovidestheinformationneededtosatisfythegoalsofroutineunderstanding.Thefollowingsectiondiscussesevaluationcriteriaforexplanationsservingothergoals.

Whenanunderstanderencountersanomalies,theypromptknowledgegoals(Ram,1989)toreconciletheanomalieswithotherknowledge,bothtocorrecttheunderstander’spictureofthecurrentsituation,andtoavoidformingfaultyexpectationsinthefuture.Correctingitspictureofthecurrentsituationrequiresdeterminingwhythesurprisingevent(orstate)wasreasonable,inordertointegratetheanomalouseventintopreviousbeliefs;avoidingformingfuturebadexpectationsrequiresidentifyingtheflawinpriorreasoning(Leake,1988b;Collins&Birnbaum,1988).Evaluationforshowinganevent’sreasonableness:

Findingwhythesurprisingevent(orstate)

wasreasonablerequiresshowingwhytheeventshouldhavebeenexpected,giventhepriorsitu-ation.Toverifythatanexplanationprovidestheneededsubstantiation,ACCEPTERchecksthattheexplanation’santecedentsareconsistentwithpreviousbeliefs,expectations,orknownpatterns(Leake,1989a),andthatthebelief-supportlinksderivingtheeventarealsoconsistentwithitspre-dictiveknowledge.Forexample,ifwearesurprisedbyacar’shighprice,asalesmanmightshowthatthehighpriceiscausedbyqualitymanufacturingandunusualamenities.

Evaluationforidentifyingflawsinpriorreasoning:Avoidingformingfuturebadexpectationsrequiresexplainingtheflawinpriorreasoning.Iftheexpectationwasgeneratedbyapplyinga

19

standardknowledgestructure,theexplanationmustshowwhythatknowledgestructurewasinap-plicable,bypointingtoaspectsofthesituationthatweredistinctivefromthenorm,butpreviouslyoverlooked.(Inreal-worldsituations,itissimplyimpossibletoconsiderallpotentiallyrelevantfactorsinadvance.)Iftheexpectationarosefromreasoningaboutacausalchainorphysicalpro-cess,theexplanationcanidentifyblockagesinthechain,orunusualconditionsinterferingwiththeprocess.Forexample,ifweexpectbreadtorise,wemightexplainitsfailuretodosobytheyeastbeingputintohottapwater,whichwasmuchhotterthanwehadbelieved,andbeingkilledbythehightemperature.Inthiscase,thecookprobablythoughtthetemperatureofhottapwaterwaslower,andtheexplanationshowsthatthisbeliefmustberevised.Wecallthefactorsthatwerepreviouslyunexpected,orevendisbelieved,distinctivecausesoftheevent—theydistinguishtherealsituationandwhatwaspreviouslyconsidered.

Inordertoverifywhetheranexplanationconnectsaneventtodistinctivecauses,asystemneedstobeabletodeterminewhetherafactisdistinctiveinthecurrentcontext.Forexample,manycausesleadtotheChallengerexplosion,suchasdesignofthesolidrocketboosters(whichreliedonO-ringstoholdaseal),launchapprovalprocedures(whichallowedtheengineers’warningstobeoverridden),thehightemperatureoftheboosters’flames(thathelpedthemburnthroughtheseals),andthecoldnessofthelaunchday(whichmadethesealsbrittle,interferingwiththeseals’positioning,andmakingiteasierforflamestopenetrate).However,mostofthesefactorswereroutine:noproblemshadarisenfromthedesignandlaunchprocedures,ortheboosterflametemperatures,inpastlaunches.Consequently,anexplanationbasedonthemwouldnotsaywhytheshuttleexplodedonthatparticularlaunchbutnotonpreviouslaunches.

However,thecoldweatheronthelaunchdaydiddifferfromcircumstancesofpreviouslaunches.Sincethecolddaywasdistinctive,itwouldbereasonabletofearanotherexplosiononcolddays,andtoexpectsuccessfullaunchesonwarmerdays.Infact,whentheSpaceShuttlesweregroundedaftertheChallengerexplosion,manypeoplearguedthatthelaunchesshouldbeallowedtocontinueonanydaysthatwereaboveacertaintemperature.

ACCEPTER’sevaluationcriteriaareexpressedintermsofasetofrequirementsthatthean-tecedentsofanexplanationmustsatisfyinorderfortheexplanationtoprovidetheneededinforma-tion.Wecalltheclassesoffactorsevaluationdimensions;thedistinctivenessofcausesfrompriorexpectationsisoneofthesedimensions.Wediscusstheevaluationdimensionsindetailbelow.

20

7Beyondroutineunderstanding:evaluationformorespecificgoals

Theprevioussectionsketchedevaluationfortwoexplanationpurposes:connectinganeventtoexpected/believedconditions,inordertoconfirmitsreasonableness,andconnectinganeventtounexpectedordisbelievedconditions,inordertorepairfaultyknowledge.BoththesepurposesarisefromACCEPTER’sbasicunderstandinggoalofmaintaininganaccuratemodelofthestoriesitreads.

ACCEPTER’ssecondsetofexplanationpurposesisprovidedexternally,andreflectsothergoals.InaplanningsystemthatusedACCEPTERtomaintainitsworldmodel,thepurposeswouldbetriggeredbythegoalsandplansoftheoverarchingplanner.Inthecurrentstand-aloneversionofACCEPTER,thepurposesareselectedbyahumanuser.

ACCEPTERevaluateswhetherexplanationsincludetheinformationneededforfindingpre-dictors,findingrepairpoints,findingcontrollablecauses,andfindingactors’contributionstoanoutcome.Thesystemalsocombinesitsevaluationforotherpurposestojudgeexplanationsforahigher-levelgoal:blockingfutureoccurrenceofanundesirableoutcome.

ACCEPTER’sevaluationprocessistotracethroughthebelief-supportnetworkexplaininganevent,examiningthecausesinthenetworktoseeifsomesubsetofthosecausesprovidestheneededinformation.Ifsomesubsetofthosecausesincludesallneededtypesoffactors(e.g.,ifACCEPTER’spurposeistoconnectaneventtounexpectedconditions,anditfindsadistinctivecause),theexplanationisaccepted.Allrelevantcausesthesystemfinds,andadescriptionoftheteststheysatisfy,areoutputbythesystem,tofacilitateapplyingtheexplanation’sinformationinplansforoverarchinggoals.

Whileitwouldbepossibletosimplydeviseindependentevaluationproceduresforeachpur-pose,parsimonysuggestsanalyzingthepurposestofindsharedcomponentsofthetestsfordif-ferentexplanationpurposes,andusingthosecomponentsasbuildingblocksforevaluationpro-cedures.ACCEPTERbuildsitspurpose-specificevaluationproceduresfromteststhatevalu-atecausesalongnineevaluationdimensions:timeliness,knowability,distinctiveness,predictive

Althoughwesimplifythediscussionbypresentingthevaluesonthesedimensionsasyes/nodecisions,allactuallyfallalongacontinuum.

21

power,causalforce,independence,repairability,blockability,anddesirability.Inwhatfollows,wewilldemonstratehowexplanationpurposesrelatetoparticularinformationneeds,andhowtheneededinformationischaracterizedintermsofcombinationsoftheevaluationdimensions.Forexample,wewillshowthatinordertorepairanundesirabledevicestate,weneedtofindanycausesoftheproblemthatarestillineffect(timeliness),thatwecanrepair(repairability),andthatareunusualcomparedtothenormaldevicestate(distinctiveness).

Afteraninitialsetofdimensionswasdefinedforafewpurposes,wefoundthatneedsformanyofourotherninepurposescouldbedescribedbysimplyusingdifferentcombinationsofdimen-sionsfromtheinitialset,andwebelievethatonlyasmallsetofadditionaldimensionswouldbeneededforaddingotherpurposestothesystem.Thesectionsbelowdescribesomeofthedimen-sionsimplementedforACCEPTER’sexplanationpurposes,andsketchtheirimplementation.

7.1Evaluationdimensionsforprediction

Whenaneventsurprisesus,andwewanttoanticipateititinthefuture,weneedtoexplainwhatcauseditsoccurrence,inpredictitinsimilarfuturecircumstances.Inorderforagroupofcausestobeusefulforprediction,itmustholdthat(1)occurrenceofthecausesmakestheeventlikely,(2)thecausehappenslongenoughinadvanceoftheeventforthepredictionittriggerstobeuseful,(3)oneofthecausesisunusualcomparedtotheexpectedsituation,sothatitgivesevidenceforthesurprisingeventasopposedtothepreviouslyexpectedone,and(4)thecausesarefactorsthatwearelikelytobeawareofinthefuture.Theserequirementscorrespondtofourevaluationdimensions:(1)predictivepower,(2)timeliness,(3)distinctiveness,and(4)knowability.Predictiveness:Evenifsomethingcausedanevent,itmaynotbepredictiveofthatevent.Forexample,eveniftheblowoutofatirewascausedbydrivingathighspeed,wewouldnotexpectablowoutthenexttimeweseeaspeedingdriver.Todeterminewhethercausesarepredictiveofanoutcome,ACCEPTERreliesonannotationsofthecausalrulesusedtoconnectthecausestotheoutcome:ifalltherulesconnectingthecausestotheoutcomearepredictive,thecausesareconsideredpredictive.

22

Timeliness:Apredictionisonlyusefulifitgivesenoughwarningtoletusdealbetterwiththepredictedevent.Forexample,supposeaNASAengineerexplainstheexplosionoftheSpaceShuttleChallengerby“theboostersburnedthrough,allowingflamestoreachthemainfueltank,causingexplosion.”Thatexplanationletsusexpectanexplosionthenexttimetheboostersburnthrough,butbythenitwillbetoolatetoabortthelaunch.However,anengineerwhorealizesthattheburnthroughwascausedbytheboostersealsbeingbrittle,duetocoldweather,couldpredictdifficultiesincoldweather,andavoidfutureexplosionsbyrefusingtoapprovelaunchesinweatherbelowacertaintemperature.

Forsomegoals,timelinessrequirementsdependoninformationwithintheexplanation.Forexample,ifthegoalispreventionofabadoutcome,theexplainerneedstopredicttheoutcomewhilethereisstilltimetoblockoneofthecauses.Inthiscase,ACCEPTERdeterminesfromtheexplanationtheamountofwarningneeded,byexaminingtheexplanation’sbelief-supportchaintofindtheearliestcausestheexplainercanprevent.Forothergoals,factorsexternaltotheexplanationdeterminetheneededamountofwarning.Forexample,someonehopingtoprofitfrompredictingstockpriceswillneedtopredicttheirfluctuationsintimetobuyorsellbeforethechangetakesplace.Thewarningneededisunrelatedtowhythestockpricechanged;itonlydependsonhowfastordersareprocessed.AlthoughACCEPTERhasnomechanismfordecidingneededtimelinessforthesegoals,theusercanmakeACCEPTER’sjudgementreflecttheirneedsbysimplyinputingthedesiredamountofwarning.

Tocalculatehowmuchwarningoneofthecausesgives,ACCEPTERaddsuptemporalsepa-rationsalongthelinksoftheexplanation’sbelief-supportchain,tofindouthowfarinadvanceoftheoutcomethecauseprobablyoccurred.

Distinctiveness:Asdiscussedabove,distinctivenessjudgeswhetherasurprisingeventwasitselfcausedbysomethingsurprising.Ifso,thatsurprisingcausemaybeusefulasapredictivefeatureoftheoutcome.

ACCEPTERjudgesacause’sdistinctivenessbycheckingwhetheritdeviatesfromstereotype-basedexpectationsthatwereineffectbeforetheanomaly.Theseexpectationsincludeexpectationstriggeredbyapplicationofschemasforstandardeventsinagivencontext,suchastheMOPforstandardeventsinarestaurantmeal,andstereotypesfortheactivitiesofbroadclassesofactors

23

(e.g.,thatathletesintrainingshouldavoidwildparties).(Foradescriptionofthesestereotypesandhowtheyareapplied,see(Leake,1989a)or(Leake,inpress).)Todeterminewhetherafactisdistinctiveinacontext,theprogramusesitsroutineunderstandingprocesstogeneratetheexpec-tationsforthatcontext,andcomparesthefacttothoseexpectations.

ThefollowingexampleshowsoutputfromACCEPTER’sformationofdistinctivenessjudgements.TheexplanationbeingconsideredisthatChallenger’sexplosionwascausedbythecombinationofthelaunch(whichenabledtheboosterflames)andcoldweather(whichcausedtheseals’brit-tleness).Thelaunchisroutine,soitisnotausefulpredictivefeature,eventhoughitiscausallyrelevant:

CheckingwhetherCHALLENGER’SSPACE-LAUNCHsatisfiestest(s)for\"DISTINCTIVENESS\".

ApplyingtestforDISTINCTIVENESStoCHALLENGER’SSPACE-LAUNCH.

UsingroutineunderstandingtocheckwhetherCHALLENGER’SSPACE-LAUNCHisstandardincontextofCHALLENGER’SROCKET-STRUCTURE.

BuildingupnewmemorycontextwithexpectationsfromCHALLENGER’SROCKET-STRUCTURE.

IntegratingCHALLENGER’SSPACE-LAUNCHintothatcontext.

CHALLENGER’SSPACE-LAUNCHsatisfiestherole-fillingpattern\"ROCKETsroutinelyfillroleSPACECRAFTinSPACE-LAUNCH\soit’sroutine....testfailed.

However,thecoldweatherisunexpected,soitisjudgeddistinctive,allowingproblemstobepredictedonfuturecolddays.

Knowability:Nomatterhowearlyapredictiveeventmaybe,itwillnothelpustopredictunlesswecanfindoutthatitoccurred.Forexample,weknowthattakeoverannouncementscausestock

Minoreditingoftheoutputhasbeendoneforreadability.

24

priceincreases,soitwouldbehelpfultopredictthemlongenoughinadvancetobuystock.Thedecisiontostartatakeoverhappensatleastafewhoursbeforetheannouncement,soitistimely,anditisalsopredictiveoftheannouncement.Unfortunately,itisusuallyuselessforprediction:thetakeoverdecisioniskeptsecret,sotheinvestorwillnotbeabletofindoutaboutituntiltheannouncementtakesplace.Knowabilitymeasureshoweasyitistoknowwhenaneventoccurs.Predictiveknowledgeneedstobeindexedunderfeaturesthatthesystemislikelytoknowabout,orthatitcancheckforroutinenesswithreasonablecost.

ACCEPTERdistinguishesbetweenthreelevelsofknowability:observablecauses,whicharelikelytobenoticedinroutineunderstanding,testablecauses,thatcouldbeknownbycarryingouttestingplansinthesystem’smemory,andundetectablecauses,whichcannotbedetectedbyknowntests,eveniftheoutcomeisimportantenoughtomaketestingworthwhile.(Ofcourse,differenttestscanrequireverydifferentlevelsofeffort,andthiscouldbeaddedtothetests’representa-tion,toallowACCEPTERtomakefiner-graineddistinctions.)Dependingontheimportanceofpredictingtheoutcome,theuserselectswhichlevelofknowabilitytorequire.ACCEPTERusestheheuristicsbelowtodecidewhethercausesareobservableortestable.Causesthatareneitherobservablenortestableareassumedtobeundetectable.

Judgingobservability:NodesinACCEPTER’smemorynetaremarkedwithinformationontheirusualobservabilitytosomeonenearby.Forexample,mostactionsareobservabletopeoplewhoarepresent,butthoughtsarenot.Observabilityinformationaboutspecificobjectandeventfeaturesisalsostored,indexedundertheobjectsandevents.Forexample,aperson’shaircolorisusuallyobservable,buthisbloodpressureisnot.Whennoobservabilityinformationisindexedundertheobjectoreventtype,observabilityinformationisinheritedfromhigher-levelabstractionsinACCEPTER’smemory.

Oneoftheprogram’sexamplesconcernsthedeathoftheracehorseSwale,whodiedunexpect-edlyatthepeakofhiscareer.WhenACCEPTERevaluatesusefulnessofdeathfromhorserace+heartdefectforpredictingfuturedeaths,itacceptsthehorseraceasobservable,because,initsmemorynet,horseracingisaspecificationofpublicperformance,andmostpublicperformancescanbeobserved.However,heartsarespecificationsofinternalorgans,andthephysicalstateofinternalorgansisnotusuallyobservable,sonotallthecausesareobservable—theexplanationisnotsufficienttoallowprediction.(Wenotethatthiscriterionisequivalenttopastapproachesfor

25

staticoperationalitycriteria(Mitchelletal.,1986),andsuffersfromthesameproblems;wepointtowardsanalternativedirectioninthesummarysectionbelow.)

Judgingtestability:Sometimespredictinganoutcomeisworththeeffortofperformingtests.Ifweknowalow-costplanforcheckingapredictivefeature,wecanperformthatplanperiodicallytogivewarningofproblemsahead.Forexample,anemployeeatYalechangedherbehaviorafterexplainingthecauseofenginedamagetohercar:

Xburnedoutanenginebydrivingwhenhercarwaslowonoil.Afterthatincident,shestartedcheckingtheoillevelwhenevershegotgas,tocorrectlowoilbeforedamaginganotherengine.

Peoplehavestandardtestsforfeaturesofsituationsthathaveimportanteffects,butarenotdirectlyobservable.Ifwewanttoknowthetemperatureoutside,wemightgotoawindowthathasathermometermountedonitsledge.Ifwethinkthatthebatteryofacarisdead,wemighttryturningontheradioorthelights.Ifwewanttoknowwhetherasteakissufficientlydone,wecanpressit,andseehowitspringsback.ACCEPTERjudgestestabilityofanexplanation’scausesbysearchingitsmemoryfortestingplansthatcandeterminewhetherthecauseholds.(Thetestsinitsmemoryaresimplyplaceholders;howtocarryoutthetestsisnotrepresentedinthesystem).IfACCEPTERfindsaplan,itjudgesthecausetestable.

TheoutputbelowshowsACCEPTERjudgingtheexplanationdeathfromhorserace+heartdefectforSwale’sdeath,toseeifitcouldbeusedbyanownertopredictandavoiddeathsinotherhorses.Theheartdefectisnotobservable,butistestablebydoinganelectrocardiogram.Consequently,theexplanationshowsthatanownercouldpredictproblems:byhavingavetdoanEKGonthehorsestheownerbuys,tofindanyheartdefects.

ApplyingtestforKNOWABILITYtoSWALE’SHEART’SHEREDITARY-DEFECTIVEORGANIC-STATE.

Searchingupabstractionnetforobservabilityinformation.SWALE’SHEART’SHEREDITARY-DEFECTIVEORGANIC-STATE

isprobablynotobservable,sinceitisa(n)ORGANIC-STATEofa(n)INTERNAL-ORGAN.

26

Searchingupabstractionnetforpointerstostandardtests.SWALE’SHEART’SHEREDITARY-DEFECTIVEORGANIC-STATEistestable,sinceORGANIC-STATEsofHEARTscan

bedetectedbythestandardtestELECTROCARDIOGRAM....testpassed.

7.2Evaluationdimensionsforrepair

Iftheanomalythatpromptsexplanationisadevicefailure,wemaywishtorepairthedeviceaspectsthatcausedthefailure.Thispromptstheexplanationpurposeoffindingrepairpoints.Fourevaluationdimensionsareimportantwhencheckingifanantecedentisagoodrepairpoint.Themostobviousrequirementsarethattheantecedenthavecausalforce(itmusthavecausedthebaddevicebehavior),andthatitberepairablebytheexplainer.However,notallcausesareworthrepairing.Ifsomeonewascarryingatelevision,andtrippedonsomeunevensteps,causinghimtodropanddamageit,theconditionofthestepswouldbeacauseofthedamage,andmightberepairable.However,repairingthestepswouldnotfixthetelevision.Thecausetorepairmustbeonewhosepresenceispredictiveofthedevicefailure’srecurrence,accordingtothepredictivenesscriteriadescribedabove.

Finally,evenifthereisacontinuingproblemthatwecanrepair,itwillnothelpifsomethingelsewillcausetheproblemagainassoonaswefixit.Forexample,droppingthetelevisionmighthavecausedapowersupplydefect,burningoutafuse.Fixingthefusebyitselfispointless:thebadpowersupplywillsimplyburnoutthereplacement.Thusweneedtofixthepowersupplyaswell.Thisexampleshowsthattheexplanationneedstotracebacktoacausewithindependencefrompriorcauses.Tojudgeindependence,ACCEPTERsimplyassumesthatacauseisindepen-dentfrompriorcausesunlessitiscausedbythecurrentstateofsomeobject.Causalforceandrepairabilityarediscussedinmoredetailbelow.

Causalforce:ACCEPTER’sinferencerulesdistinguishtwoclassesofconnectionsbetweenevents(orstates).Oneeventcausesasecondifitactuallybringsaboutthesecondevent,aswhen

27

aheartattackcausesdeath.Aneventispredictiveofanother,withoutnecessarilybeingacause,ifthefirstpromptsanexpectationforthesecond.Forexample,ifweareshockedbyanawfulmealinarestaurant,someonemightexplainitbysaying“allNewHavenrestaurantsarebad.”Theexplanationmayletuspredictfuturebadmeals,butitdoesn’tsaywhatcausesthelowquality.ACCEPTER’sinferencerulesareannotatedwithwhethertheydescribeacausalconnection,orapredictiveone(causalconnectionsthatholdinagivensituationmaybenon-predictive,iftheytraceanunlikelyoutcome:drinkingmilkcausesanallergicreactioninsomepeople,butwedonotnormallypredictthatreaction).

Repairability:Findingthecauseofabadstateonlyhelpsrepairifwealsoknowhowtofixit.Forexample,tracinganenginefailurebacktosomeobscurecomponentofthetransmissionwillbelittleusetomostpeoplewhoaretryingtodoaroad-sideemergencyrepair.However,iftheyfindthatawirehasshakenloose,orahosehascomeunattached,theycanmaketherepair.Tojudgerepairabilityofadevice,ACCEPTERsearchesmemoryforarepairplanindexedunderthedevice,andthestatebeingconsideredasapotentialrepairpoint.Ifatelevisionhadananomalouslybadpicture,ACCEPTERwouldlookatthecausesofthebadpicturegivenbytheexplanation,andseeifitcouldretrieveaplantofixanyofthem.Forexample,iftheexplanationattributedtheproblemtobadatmosphericconditions,itwouldbeunlikelytofindarepairplan.However,iftheproblemwerecausedbytheantennapointingthewrongway,aplanforfixingorientationsofsmallobjects—grabbingthemandmovingthem—couldberetrievedandapplied.TheoutputbelowgivesasampleofACCEPTER’srepairabilityjudgements.WhenitevaluatestheexplanationthatJohn’s“brandA”carwasdefectivebecauseofabadtransmission,itfirstcheckstoseewhetheritcanretrieveaplanforrepairingthetransmission.Itfindsnostoredrepairplansfortransmissions,orforanyoftheabstractionsoftransmissionsinitsmemory:

ApplyingtestforREPAIRABILITYtoTRANSMISSION-743’SLOWMECHANICAL-CONDITION.

Searchingupabstractionnetforpointerstostandardrepairplans....testfailed.

28

TRANSMISSION-743’SLOWMECHANICAL-CONDITIONisprobablynotrepairable,sincenostandardrepairplansarestoredunderanyofitsabstractions.

SinceACCEPTERfindsnostandardrepairplans,itcheckswhetherlaterstepsinthebelief-supportchainarerepairable.TheXP’sbelief-supportchainshowsthattheenginedefectiscausedbyboththefactthatthetransmissionisdefective,asabove,andthatthetransmissionisacom-ponentoftheengine.AlthoughACCEPTERfindsnorepairstrategyspecificallydirectedtowardsfixingthefactthatatransmissionisanenginecomponent,itdoesfindastrategyforrepairinganycomponentrelationshipthatcausesproblems:replacethecomponent.Thisstrategysuggeststhatreplacingthetransmissionisapossiblerepairplan:

ApplyingtestforREPAIRABILITYtoTRANSMISSION-743’SPART-OF-RELATIONSHIPtoBRAND-A’SENGINE.

CheckingrepairabilityoffeaturesofTRANSMISSION-743’SPART-OF-RELATIONSHIPtoBRAND-A’SENGINE.

Searchingupabstractionnetforpointerstostandardrepairplans.

BRAND-A’SENGINEASCONTAINEROFTRANSMISSION-743’S

PART-OF-RELATIONSHIPtoBRAND-A’SENGINEisrepairable,sinceCONTAINERsofPART-OF-RELATIONSHIPscanusuallyberepairedbythestandardplanREPLACE-COMPONENT....testpassed....Detailisacceptable.

7.3Evaluationdimensionsforcontrol

Whenananomalousstateoreventisundesirable,anunderstandermaywanttopreventitinthefu-ture.Oneplanforthisgoalistoblocksomeoftheevent’scausesdirectly;anotheristodiscourage

29

anyactorsresponsibleforthesituationfromcontributingtoitsrecurrence.(Likewise,ifanout-comeisdesirable,anexplainermaywishtofindhowtoachieveit,eitherdirectlyorbyinfluencingothers,butwewillnotconsiderthatpurposehere.)

Directlypreventingfutureoccurrenceofaneventinvolvesfindingpremiseswithcausalforce—thatcausetheoutcome—andthattheexplainercanblock(blockability).Anexplanationthatshowsblockablecausescanbeusedintwoways:first,theexplainermaybeabletopermanentlydisableoneofthecauses,sothatanyrepetitionofthesamecausalchainisimpossible.(Forexample,ifahousewasburglarizedbecauseathiefcouldenterthroughanopenwindow,thevictimcanblockrecurrencebykeepingthewindowsshut.)Second,ifitisimpracticaltopermanentlyblockanyofanevent’scauses(itmaybetoounpleasanttokeepwindowsclosedinthesummer),explanationneedstoshownotonlyhowtheoutcomecouldbeblocked,butalsohowtopredictaspecificinstanceoftheoutcomelongenoughinadvancetoblocktheprobleminthecurrentinstance.ThisprocedureisbasicallytheanticipateandavoidstrategysuggestedbyHammondforavoidingfailuresincase-basedplanning(Hammond,1989).Wediscussbelowsomeheuristicsforjudgingblockability,andhowtheyareappliedtojudgeexplanationsforanticipatingandavoidinganundesirableoutcome.

Blockability:Decidingwhatanexplainercanpreventisdifficult;thingsthatseemuncontrollableatfirstglancemayactuallybeeasytoinfluence.Forexample,ifapicnicisrainedout,itisnaturaltoaccepttherainasbeyondourcontrol,butitmightmightbeavoidablebyschedulingthepicnicduringthedryseason.Nevertheless,averysimpleheuristicissufficientforjudginganactor’sabilitytocontrolmanyevents:wecanassumethatactorswhovoluntarilyfillactorrolesinevents,orprovideotherrole-fillersforthem,canprobablyblockthoseevents.

Areal-worldexampleoftheimportanceofactorrolesinblockinganoutcomecomesfromtheChallengerexplosion.EachareaofNASAattemptedtofindwaystopreventsimilarsitua-tionsarising,anddifferentdivisionsfocusedondifferentcausesoftheproblem,dependingontheexplainers’influence.Aftertheexplosion,accordingtotheastronautJohnCreighton:

[Everyonehadadifferentidea]ofwhatwedidn’tthinkworked.Ifyou’reanengineman,youwanttheenginefixed;ifyou’reinchargeofsomethingelse,youwantthat

30

fixed.

ACCEPTER’sblockabilitychecksconsiderthreewaysapersoncanbeinvolvedinanevent:asanactorwhoisimmediatelyinvolved;asadirectoroftheaction,whoisnotimmediatelyinvolved,butwhohasauthorityoveritsprogression;andasasupplieroftheobjectsoractorsthatthedirectorselects.

Actorsinaneventmaybeabletoblockitbyrefusingtoparticipate.Directorsmayblockitbyorderingactorsundertheircontrolnottoparticipate,bycontrollingthesettingforevents(locale,timeofoccurrence,orfeaturesoftheenvironment),bychangingwaysofselectingtheobjectsused,toavoidusingobjectsthatareparticularlylikelytocontributetoabadoutcome,orbychangingobjectsupplierstoavoidsuchobjects.Supplierscanalsoblockoutcomesbycontrollingtheobjectstheymakeavailable.Table2sketcheshoweachofthesemeansofexplainercontrolenteredintostrategiesforpreventinganotherspaceshuttledisaster.

ACCEPTER’sbasicprocedurefordecidingblockabilitybyagivenpersonistousetheexpla-nationtofindaperson’sinvolvementintheoutcome,andthenseeifthatpersonisinvolvedasanactor,director,orsupplierofanobject(whichACCEPTERchecksbyseeingifthatpersonisitsowner).IfsoACCEPTERassumesthatthepersoncouldblocktheevent.Itthencheckswhethertimelypredictionispossible,toallowthepersontoanticipatetheoutcomeandexertcontroltoblockitsoccurrence.

Forexample,theoutputbelowshowshowACCEPTERdecidesthatthespaceshuttle’sastro-nautscouldpreventfutureexplosions.TheantecedentbeingconsideredisChallenger’slaunch,andACCEPTERfirstverifiesthatthelaunchisacauseoftheexplosion,toseeifblockingthelaunchcouldbeaneffectivewaytopreventtheoutcome:

Checkingwhethersomeantecedentsatisfiesthefollowingtests:CAUSALFORCETEST(doesfactcauseconsequent?),and

BLOCKABILITY+TIMELINESS(canCHALLENGER’SSPACE-LAUNCH’sASTRONAUTblockafterpredictingoutcome?).

ApplyingtestforCAUSALFORCE(doesfactcauseconsequent?)toCHALLENGER’SSPACE-LAUNCH.

Newsweek,October10,1988.

EachcausallinkisrepresentedasanXPinACCEPTER’smemory;theoutputonlyshowsthenamesofthelinks.

31

Table2

StrategiesforblockingrecurrenceoftheChallengerdisaster

Real-lifeapplicationofthepreventionstrategiesforcontrolbyactors,directorsandsuppliers.

Controlbyactors

Refusetoparticipate

Aftertheexplosion,astronautssaidthattheywouldn’tflyuntiltheshuttleswererepaired.

Controlbydirectors

Changesettingfortheevent

–Changelocalefortheevent

SomepeoplesuggestedchangingthelaunchsitetoHawaii,wheretheweatheriswarmer.–Changethetimeoftheevent

Engineersadviseddelayinglaunchingwhentheweatherwasbelow53degrees.–Changefeaturesoftheenvironment

NASAinstalledheatersonthelaunchpadtowarmboostersbeforelaunch.

Changerole-fillerchoice

–Applyteststoruleoutbadobjects

SomepeoplesuggestedNASAshouldinspecttheboostersealsafterdeliveryoftheboosters.NASArejectedthisbecausetestswouldrequiredisassemblythatmightintroducenewdefects.–Changesupplierofobject

SomeadvocatedstoppingusingboostersmadebyMortonThiokol.

Controlbysuppliers

Blockaccesstoaclassofrole-filler

NewshuttlemanufacturingwasfrozenbycongresswhiletheChallengerinvestigationwenton,inordertoavoidrepeatsoftheexplosion.

Changedesign

MortonThiokolredesignedtheboosters,withbetterseals.

32

CheckingtheconnectionbetweenCHALLENGER’SSPACE-LAUNCHandCHALLENGER’SEXPLOSION.

TestingifCAUSAL-MOP-SCENE:LAUNCH->IGNITIONsatisfiestestforLINKCAUSATION....testpassed.

TestingifCAUSAL-MOP-SCENE:IGNITION->HIGH-PRESSUREsatisfiestestforLINKCAUSATION....testpassed.

Testingif

BRITTLE-SEAL+CONTAINER-SEAL+CONTENTS-PRESSURE=>CONTAINER-EXPLOSIONsatisfiestestforLINKCAUSATION....testpassed.

TestingifEXPLOSION-IN-COMPONENT=>EXPLOSION-IN-WHOLEsatisfiestestforLINKCAUSATION....testpassed.

...Alllinksareacceptable....testpassed.

Sincetheexplanationshowsthatthelaunchcausesarocketbooster’sexplosion,whichinturncausestheshuttleasawholetoexplode,blockingthelaunchwouldblocktheexplosion.Thisidentifiesitasacausethatwouldbeworthwhiletoblock,soACCEPTERcheckswhetheranastronautcouldblockit.Itfirstcheckswhethertheastronautcontrolstheavailabilityofobjectsrequiredtofillrolesinthelaunch(whichtheastronautdoesnot),orwhethertheastronauthascontroloverwhethertoparticipateinthelaunch:

CheckingifCHALLENGER’SSPACE-LAUNCHisblockable.

CheckingifCHALLENGER’SSPACE-LAUNCH’sASTRONAUTcontrolledoutcomebycontrollinganeededobject....Noactor-controlledobjectsfound.

CheckingifCHALLENGER’SSPACE-LAUNCH’sASTRONAUTcontrolledoutcomethrougharolehefilled.

33

CHALLENGER’SSPACE-LAUNCH’sASTRONAUTmight

havebeenabletopreventCHALLENGER’SSPACE-LAUNCH,byrefusingtobeitsASTRONAUT,sincethatisavoluntaryactorrole.

Fromthisinformation,ACCEPTERconcludesthatthelaunchisablockablecause.Thenextquestioniswhetheranexplosioncanbepredictedbeforethelaunch,sotheastronautcanstopthelaunchintime.Becausethelastcauseoftheexplosionthattheastronautcontrolsisthelaunch,preventingfutureexplosionsisonlypossibleiftheexplanationshowshowtopredictexplosionbeforethelaunchoccurs.Toknowhowmuchwarningwouldbeneeded,ACCEPTERfirstusestheexplanationtodeterminehowfarinadvanceoftheexplosionthelaunchoccurs,andthencheckswhethertheexplanationallowspredictionwiththatamountofwarning.(ACCEPTERusesacoarse-grainedtemporalrepresentation,withtemporalseparationsrepresentedasNONE,MINUTES,HOURS,DAYS,WEEKSandYEARS.)

CheckingifXPallowspredictionofoutcomebeforeCHALLENGER’SSPACE-LAUNCHoccurs.

CalculatingamountofwarningneededtopredictbeforeCHALLENGER’SSPACE-LAUNCHoccurs.

CHALLENGER’SSPACE-LAUNCHleadstoROCKET-IGNITION-41immediately.

ROCKET-IGNITION-41leadstoGAS-42’SHIGHPRESSUREMINUTESafterwards.

GAS-42’SHIGHPRESSUREleadstoSOLID-ROCKET-43’SEXPLOSIONMINUTESafterwards.

SOLID-ROCKET-43’SEXPLOSIONleadstoCHALLENGER’SEXPLOSIONimmediately.

PredictingbeforeCHALLENGER’SSPACE-LAUNCHrequiresfindingapredictivefeatureatleastMINUTESbeforeCHALLENGER’SEXPLOSION.

34

CheckingdetailforpredictingCHALLENGER’SEXPLOSIONMINUTESbeforeithappens.

Sincethelaunchisminutesbeforetheexplosion,anexplanationforpreventionmustallowpredic-tiontobedoneatleastminutesbeforetheexplosionaswell.Todecidewhethertheexplanationdoesso,ACCEPTERappliesthetestsforpredictivenessdescribedintheprevioussection.Thecoldnessofthesealoccursearlyenoughtoprovidewarning,andisdistinctive,predictive,andknowable.Consequently,ACCEPTERacceptstheexplanationasbeingusefulforanastronaut’sprediction,allowinghimtomonitorandintervene:oncolddays,theastronautcanrefusetofly.

7.4Evaluationdimensionsforactors’contributionstoanoutcome

Tomaintaineffectiveperformance,anyactorintheworldneedswaysofjudgingtheappropriate-nessofitsactions.Whenasurprisinglygoodorbadoutcomeoccurs,asystemmaybenefitbyexplainingwhocontributedtotheevent.Byidentifyingtheactorsinvolved,andascribingpraiseorblamefortheirroles,itmaybepossibletoinfluencetheirfuturebehavior.Also,bylearningaboutthe(possiblyunexpected)goodandbadramificationsofothers’behavior,thesystemcanlearnaboutpossibleproblemsandopportunitiestoconsiderwhenplanningitsownfutureactions,formingnewstrategiesforguidingitsownbehavior.

Thesegoalsprompttheexplanationpurposeofidentifyingactors’contributionstoanaction,andascribingpraiseorblame.Thiscanbedoneaccordingtothedesirabilityoftheoutcome,andofactionsleadingtoit.Inaddition,anactorcanalsobeblamed,evenifitwasimpossibletopredictorcontrolanoutcome,ifthatactorcontributedtotheoutcomethroughanundesirableact.Thebadresultstrengthenstheact’soriginalproscription:forexample,wemightblameadrugdealerforanaddict’sdeathbyoverdose,evenifdeathsfromoverdosearerelativelyunlikely.

TheoutputbelowshowsACCEPTER’sevaluationofblamefortheexplanationthattherace-horseSwalediedbecauseatraineraccidentallygavehimanoverdoseofperformance-enhancingdrugs.Itfirstcheckswhethertheactorshouldhaveanticipatedtheproblem,andcouldhavepre-ventedit:

35

CheckingPERFORMANCE-DRUG-OVERDOSE-BY-TRAINERforASSIGNING-BLAME.

ApplyingtestforBLOCKABILITY+TIMELINESS(canATHLETIC-TRAINER-1497blockafterpredictingoutcome?)toATHLETIC-TRAINER-1497’SM-PERFORMANCE-ENHANCEMENT-DRUGS.

ATHLETIC-TRAINER-1497couldhavepreventedinitiation

ofATHLETIC-TRAINER-1497’SM-PERFORMANCE-ENHANCEMENT-DRUGSsinceitsDRUGGERcontrolsinitiation.

However,ACCEPTERdeterminesthatperformanceenhancingdrugsarenotpredictiveoffataloverdose,sothetrainerwouldnothaveexpectedtheoutcome.Consequently,thetrainercannotbeblamedwithintentionallycausingSwale’sdeath.

ACCEPTERthencheckswhethergivingperformance-enhancingdrugsisinitselfundesirable.InACCEPTER’smemorynet,administeringperformanceenhancingdrugsisaspecificationofthecategoryforillegalactions,soitjudgesthedruggingasundesirable.Givingperformanceenhancingdrugsisannotatedasusuallybeingavoluntaryaction,soACCEPTERdecidesthattheexplanationgivescauseforblame.

7.5Summaryofevaluationdimensions

Table3summarizesourevaluationdimensions,andthesimpleheuristicsACCEPTERusestotestalongthem.Foradiscussionofthestrategiesnotdiscussedhere,andhowtheycombinetosatisfytheinformationrequirementsforotherpurposes,see(Leake,inpress).

Itshouldbenotedthatwhilethesystemusesdynamiccriteriatoevaluateforsomedimensions(e.g.,timeliness,distinctiveness,andrepairability),itscriteriaforotherdimensionsarestatic(e.g.,theobservabilitycomponentofknowability).Wedidnotinvestigatedynamiccharacterizationsofalldimensions,simplybecauseourmaineffortwasdevotedtoinvestigatingtherelationshipbetweenevaluationgoalsandtheneededdimensions.However,westronglyagreewiththear-gumentsin(DeJong&Mooney,1986)thatsuchcriteriamustbeabletodynamicallytakeintoaccountcurrentsystemknowledge.RicherandmoredynamiccharacterizationsofACCEPTER’sdimensionsinvolvemanyissuesforfutureresearch.Forexample,arichercharacterizationofknowabilitycriteriawouldhavetoinvolvereasoningaboutcompetingwaystogatherinformation,

36

Table3.

ACCEPTER’sheuristicsfortestinganeventorstateA,statedasacauseinanexplanation,alongeachevaluationdimension.

TracealongtheXP’sderivationofoutcomefromA,summingthe

Timeliness

standarddelaysfromantecedenttoconsequentofeachinferencerule.

Useroutineunderstandingmechanismtobuildupstandardexpec-Distinctiveness

tationsfromthebackgroundsituation,andintegrateAintothatcontext.Aisdistinctiveifitisunexpectedoranomalousinthatcontext.

CheckwhetherA,oritsabstractionsinmemory,isannotatedas

Knowability

usuallyobservable,orsearchforaplaninmemorythatcanbeappliedtodeterminewhetherAhasoccurred.

Predictivepower/causalforce

CheckwhethertheXPderivestheoutcomefromAbyasequenceofpredictivelinks(respectively,causallinks).

CheckifAiscausedbyastatestillineffect.Ifnot,assumeAisindependentofpriorcauses.

SearchmemoryforastandardrepairplanforanyabstractionofA.Lookfordirectinvolvementoftheactorasanactor,director,or(forblockabilityonly)supplierofA.

CheckifactiondescribedbyAisaspecificationofillegal-activity.

IndependenceRepairabilityControllabilityDesirability

37

theirlikelycosts,andtheirchanceofsuccess;(Hunter,1990)discussessomeofthesedirectionsinthecontextofknowledgeplanning.

8Thevalueofgoal-basedevaluation

Theprecedingsectionssketchedourapproachtogoal-basedexplanationevaluation.Inthissection,weputthatapproachinperspective,discussinghowagoal-basedevaluationmodulecancontributetoanoverarchingexplanation-basedsystem.Weadvancethreemainpoints:First,thatgoal-basedevaluationallowslearningtobefocusedmoreeffectivelythaninpreviousapproaches.Second,thatitfacilitatesconstructionofusefulexplanationsinacase-basedexplanation,and,moregenerally,makesitpossibletoreliablybuildexplanationsforsystemswithmultipletypesofgoals.Third,thatitoffersavaluablenewwaytodealwiththeimperfecttheoryprobleminexplanation-basedlearning,byallowinganexplanation-basedsystemtolearneffectivelydespitecertainimperfectionsinitsdomaintheory.

Focusingexplanationtowardsknowledgegaps:Ourexplanationprocesscentersaroundex-plaininganomalies.Inourdiscussionofevaluationforroutineunderstanding,wedescribedAC-CEPTER’srequirementthatexplanationsaccountnotonlyfortheevent,butforwhyexpectationsfailed.Thisdiffersfromapproachessuchas(Mooney,1990),whichconcentrateonaccountingforwhytheeventoccurred.Thedifferenceisimportanttodecidingwhichbeliefstorepair,andwhattolearnfromanewsituation.IfwewanttoexplainwhybasketballteamAdefeatedteamBinaclosegame,theexplanationswewillseekwillbequitedifferentifweexpectedBtowin(inwhichcasewemightrefertoinjuriesofB’splayers),orifweexpectedAtowinbyalargemargin(inwhichcasewemightrefertoteamAbeingover-confident).Ingeneral,goal-basedevaluationallowslearningtofocusontheaspectsofasituationthatareimportant:thoserelevanttosystemgoals.ACCEPTERcanjudgeasingleexplanationacceptableforonepurpose,butnotforanother.forexample,oneofACCEPTER’sstoriesinvolvestherecallofadefectivecar.Whenthesystem’spurposeistofindwaysofpredictingdefects,itacceptstheexplanationthatthemanu-facturerhasbadqualitycontrol;whenitspurposeistofindrepairpoints,itrejectsthatexplanationasinsufficient,andacceptsanexplanationthatpointstothespecificpartthatfailed.

38

Guidingexplanationconstructionincase-basedexplanation:Howtocontrolexplanationconstructionisadifficultissuethathasreceivedsurprisinglylittleattention.Manyexplanation-basedsystemsconstructexplanationsbyundirectedchaining,whichrisksoverwhelminginferencecostforanybutthemostsimpleexplanations(Rieger,1975).TheSWALEsystemaddressesthisproblembyusingcase-basedreasoningtobuildnewexplanations.Ratherthanexplainingfromscratch,itappliespriorexperiencebyretrievingandadaptingexplanationsfromsimilarpastepisodes.Thecase-basedapproachfacilitatesconstructionofcomplicatedexplanations,byre-usingpriorreasoningwheneverpossible.Inaddition,whennewandoldsituationsarequitesimi-lar,thecase-basedapproachcangenerateexplanationsthataremorelikelytobevalid:ratherthanbeingarbitraryrulecombinations,thehypothesesitbuildsaresupportedbypriorexperience.Unliketraditionalexplanation-basedsystems,whichcanrelyonallcandidateexplanationsbeingintheproperformfortheirsingletask(e.g.,showingsufficientconditionsforconceptmem-bershipfortheconceptrecognitiontaskin(Mitchelletal.,1986)),acase-basedexplainerreusesexplanationsthatmayhavebeenconstructedinverydifferentcontexts,andforverydifferentgoals.Consequently,unlikesingle-purposesystemsthatcantailorexplanationconstructiontotheirgoals,andrelyoninitialexplanationconstructiontoprovideanappropriatetypeofexplanation,acase-basedexplainercannotbeassuredofstartingwithanexplanationthatincludesappropri-ateinformation(Leake,1989b).Consequently,itmusthaveameansfordeterminingwhetheranexplanationcontainsthespecificinformationitneeds,andtoidentifygapstofillthroughexplana-tionadaptation.Ourevaluationprocessprovidesthatguidance,enablingacase-basedexplainertoreliablyusecasesfromamulti-purposecaselibrary.Thisguidancecouldalsobeappliedtoanyexplanationconstructionsystemthatmustbuildexplanationsformultiplepurposes.

Dealingwiththeimperfecttheoryproblem:Explanation-basedlearningresearchtraditionallyconsidersexplanationstobedeductiveproofsofconceptmembership(Mitchelletal.,1986).Inthisframework,thestructureoftheproofassuresthattheexplanationpointstoasufficientsetoffactorsforconceptmembership.However,ifweseektoexplainreal-worldevents,noexplanationcanincludeallthecausally-relevantfactors.AsMitchelletal.observe,real-worlddomaintheoriesareoftenbothincompleteandintractable.

Responsestotheimperfecttheoryproblemproposemethodsforrepairingthetheory’sdefects.Forexample,(Dietterich&Flann,1988)suggeststhatwhenadomaintheoryallowsmultiple

39

incompatibleexplanations,inductionoverexplanationsforasetoftrainingexamplescanbeusedtofindaspecializeddomaintheorythatexplainsthepositiveexamples,butnoneofthenegativeones.(Rajamoney,1988)advocatesexperimentationtodeterminehowtoextendorrepairadomaintheoryinresponsetoproblemssuchasmultipleincompatibleexplanations,ortheinabilitytoconstructanycompleteexplanation.Unfortunately,theseapproachesarenotalwayspractical:itmaybenecessarytoactwithouthavingobservedenoughexamplestorepairadomaintheoryusinginductionoverexplanations,andmaybeinfeasibletoperformtheexperimentsneededtodoexperimentation-basedtheoryrevision.

However,theinabilitytoconstructacompleteexplanation,ortoruleoutincompatiblealter-nativeexplanations,doesnotnecessarilyinterferewithhumanuseofexplanation.Peopleoftenaccomplishtheirgoalsusingfragmentaryexplanations.Someonewithlittleautomotiveknowl-edgemaybuyausedcar,andnoticethatitoftenfailstostartoncolddays.Althoughthebuyerwouldprobablybeunabletogenerateacompleteexplanation,itwouldstillbepossibletoformthepartialexplanationthatcoldisoneofthecausestheproblem.Thishypothesisdoesnotgivesufficientconditionsforthefailure,sincethecarsometimesstartsdespitethecold,anditwouldprobablynotbefeasibleforanovicetoidentifytheotherfactorsthatareinvolved.Nevertheless,thepartialexplanationisstilluseful:itcanbeusedtodecidetokeepthecarinthegarageoncoldnights.

Likewise,choosingbetweencompetingincompatibleexplanationsmaynotbepossible,ormaynotbeworthwhileevenifitispossible:apolicemanarrestingamurdererdoesnotneedtochoosebetweentheincompatibleexplanationsthatthemurdererwasmotivatedbyjealousy,orbyfinancialgain,aslongasthemurdererproposedbyallcompetingexplanationsisthesameperson.Thusimperfectionsintheorydonothavetointerferewiththeuseofanexplanation.Bygivingaprincipledaccountofwhichaspectsofexplanationsareimportant,andwhicharenot,ourgoal-basedevaluationcriteriadeterminewhenanimperfectexplanationcansimplybeapplied,withouthavingtoextendthetheorytoresolvedefectsorchoosebetweencompetingalternatives.Allowingexplanation-basedprocessingofpartialandimperfectexplanationssignificantlyextendsthecircumstancesinwhichasystemcanapplyexplanation-basedtechniques.

40

9Conclusion

Evaluationofexplanationsisadynamicprocess:anexplanation’sgoodnessdependsonwhetheritsatisfiestheexplainer’scurrentneedsforinformation.Anexplanationthatisgoodforonepurposemaybeirrelevanttoanother,ormaygiveinadequateinformationforit.Forsomepurposes,averyvagueorincompleteexplanationmaybesufficient;forothers,certainaspectswillneedtobedescribedingreatdetail.

Inordertojudgethegoodnessofanexplanation,weneedtoknowhowitwillbeused,andwhatinformationthatuserequires.Thispapersketcheshowgoalspromptplansthatinturntriggerbasicexplanationpurposes,andtracestheirinformationrequirements.Tojudgewhetherexplanationssatisfytheserequirements,ACCEPTERusesheuristicsforjudgingcausesalongcombinationsofsimpleevaluationdimensions.

Theinformationrequiredfortheexplanationpurposeswehavedescribedcanbecharacterizedintermsofnineevaluationdimensions(timeliness,knowability,distinctiveness,predictivepower,causalforce,independence,repairability,controllability,anddesirability),givingacompactwayofdescribingneedsforinformation,andsuggestingclassesofevaluationheuristicstorefineinfutureresearch.

Byprovidingagoal-sensitivewaytojudgetheinformationcontainedinaparticularexplana-tion,ACCEPTER’sapproachextendsconsiderablytheapplicabilityofexplanation-basedprocess-ing,givingameansfordecidingwhethertoacceptandlearnfromanimperfectexplanation.Itisafirststeptowardsamodelofexplanationasadynamic,goal-drivenprocess,integratedfullywiththesystemtasksitserves.

41

References

Ahn,W.,Mooney,R.,Brewer,W.,&DeJong,G.(1987).Schemaacquisitionfromoneexample:

Psychologicalevidenceforexplanation-basedlearning.ProceedingsoftheNinthAnnual

ConferenceoftheCognitiveScienceSociety(pp.50–57).Seattle:CognitiveScienceSociety.Collins,G.,&Birnbaum,L.(1988).Anexplanation-basedapproachtothetransferofplan-ningknowledgeacrossdomains.Proceedingsofthe1988AAAISpringSymposiumonExplanation-basedLearning(pp.107–111).Stanford:AAAI.

Cullingford,R.(1978).ScriptApplication:ComputerUnderstandingofNewspaperStories(Tech-nicalReport116).NewHaven:YaleUniversityComputerScienceDepartment.DeJong,G.,&Mooney,R.(1986).Explanation-basedlearning:Analternativeview.Machine

Learning,1(1),145–176.Boston:Kluwer.Dietterich,T.,&Flann,N.(1988).Aninductiveapproachtosolvingtheimperfecttheoryproblem.

Proceedingsofthe1988AAAISpringSymposiumonExplanation-basedLearning(pp.42–

46).Stanford:AAAI.

Granger,R.(1980).AdaptiveUnderstanding:CorrectingErroneousInferences(TechnicalReport

171).NewHaven:YaleUniversityComputerScienceDepartment.

Hammond,K.(1989).Case-BasedPlanning:ViewingPlanningasaMemoryTask.SanDiego:

AcademicPress.

Hanson,N.(1961).PatternsofDiscovery.Cambridge:CambridgeUniversityPress.

Heider,F.(1958).ThePsychologyofInterpersonalRelations,volumeXVofCurrentTheoryand

ResearchinMotivation.NewYork:Wiley.Hobbs,J.,Stickel,M.,Appelt,D.,&Martin,P.(1990).Interpretationasabduction(Technical

Report499).MenloPark:SRIInternational.Hunter,L.(1990).Planningtolearn.ProceedingsoftheTwelfthAnnualConferenceoftheCogni-tiveScienceSociety(pp.261–268).Cambridge:CognitiveScienceSociety.

42

Kass,A.(1986).Modifyingexplanationstounderstandstories.ProceedingsoftheEighthAn-nualConferenceoftheCognitiveScienceSociety(pp.691–696).Amherst:CognitiveScience

Society.

Kass,A.,&Leake,D.(1988).Case-basedreasoningappliedtoconstructingexplanations.Kolod-ner,J.(Ed.),ProceedingsoftheCase-BasedReasoningWorkshop(pp.190–208).ClearwaterBeach:DARPA.

Kedar-Cabelli,S.(1987).Formulatingconceptsaccordingtopurpose.ProceedingsoftheSixth

AnnualNationalConferenceonArtificialIntelligence(pp.477–481).Seattle:AAAI.Keller,R.(1987).TheRoleofExplicitContextualKnowledgeinLearningConceptstoImprove

Performance.PhDthesis,RutgersUniversity.Keller,R.(1988).Operationalityandgeneralityinexplanation-basedlearning:Separatedi-mensionsoroppositeendpoints?Proceedingsofthe1988AAAISpringSymposiumon

Explanation-basedLearning(pp.153–157).Stanford:AAAI.

Kelley,H.H.(1967).Attributiontheoryinsocialpsychology.Levine,D.(Ed.),NebraskaSympo-siumonMotivation.Lincoln,NE:UniversityofNebraskaPress.

Kolodner,J.(1984).RetrievalandOrganizationalStrategiesinConceptualMemory.Hillsdale,

NJ:LawrenceErlbaumAssociates.

Lalljee,M.,&Abelson,R.(1983).Theorganizationofexplanations.Hewstone,M.(Ed.),Attri-butionTheory:SocialandFunctionalExtensions.Oxford:Blackwell.

Lalljee,M.,Watson,M.,&White,P.(1982).Explanations,attributions,andthesocialcontextof

unexpectedbehavior.EuropeanJournalofSocialPsychology,12,17–29.

Leake,D.(1988a).Evaluatingexplanations.ProceedingsoftheSeventhNationalConferenceon

ArtificialIntelligence(pp.251–255).Minneapolis:AAAI.

Leake,D.(1988b).Usingexplainerneedstojudgeoperationality.Proceedingsofthe1988AAAI

SpringSymposiumonExplanation-basedLearning(pp.148–152).Stanford:AAAI.

43

Leake,D.(1989a).Anomalydetectionstrategiesforschema-basedstoryunderstanding.Proceed-ingsoftheEleventhAnnualConferenceoftheCognitiveScienceSociety(pp.490–497).Ann

Arbor:CognitiveScienceSociety.

Leake,D.(1989b).Theeffectofexplainergoalsoncase-basedexplanation.Hammond,K.

(Ed.),ProceedingsoftheCase-BasedReasoningWorkshop(pp.290–294).PensacolaBeach:DARPA.

Leake,D.(inpress).EvaluatingExplanations.Hillsdale,NJ:LawrenceErlbaumAssociates.Leake,D.,&Owens,C.(1986).Organizingmemoryforexplanation.ProceedingsoftheEighth

AnnualConferenceoftheCognitiveScienceSociety(pp.710–715).Amherst:CognitiveSci-enceSociety.

Lebowitz,M.(1980).GeneralizationandMemoryinanIntegratedUnderstandingSystem(Tech-nicalReport186).NewHaven:YaleUniversityComputerScienceDepartment.Mackie,J.(1965).Causesandconditions.AmericanPhilosophicalQuarterly,(4).

Mehlman,R.,&Snyder,C.(1983).Excusetheory:Atestoftheself-protectiveroleofattributions.

JournalofPersonalityandSocialPsychology,49(4),994–1001.

Minton,S.(1988).Quantitativeresultsconcerningtheutilityofexplanation-basedlearning.Pro-ceedingsoftheseventhnationalconferenceonartificialintelligence(pp.564–569).SaintPaul:AAAI.

Mitchell,T.,Keller,R.,&Kedar-Cabelli,S.(1986).Explanation-basedgeneralization:Aunifying

view.MachineLearning,1(1),47–80.Mooney,R.(1990).AGeneralExplanation-basedLearningMechanismanditsApplicationto

NarrativeUnderstanding.SanMateo,CA:MorganKaufmann.Moore,J.,&Swartout,W.(1989).Areactiveapproachtoexplanation.ProceedingsoftheEleventh

InternationalJointConferenceonArtificialIntelligence(pp.1504–1510).Detroit:IJCAI.

44

Mostow,D.J.(1983).Machinetransformationofadviceintoaheuristicsearchprocedure.Michal-ski,R.,Carbonell,J.,&Mitchell,T.(Eds.),MachineLearning:AnArtificialIntelligence

Approach.Cambridge:TiogaPublishingCompany.

Ng,H.,&Mooney,R.(1990).Ontheroleofcoherenceinabductiveexplanation.Proceedingsof

theEighthNationalConferenceonArtificialIntelligence(pp.337–342).Boston:AAAI.Paris,C.(1987).

Combiningdiscoursestrategiestogeneratedescriptionstousersalonga

naive/expertspectrum.ProceedingsoftheTenthInternationalJointConferenceonArtificialIntelligence(pp.626–632).Milan,Italy:IJCAI.

Pazzani,M.J.(1988).Selectingthebestexplanationforexplanation-basedlearning.1988Spring

SymposiumSeries:Explanation-BasedLearning(pages165–169).Stanford:AAAI.Rajamoney,S.(1988).Experimentation-basedtheoryrevision.Proceedingsofthe1988AAAI

SpringSymposiumonExplanation-basedLearning(pp.7–11).Stanford:AAAI.

Rajamoney,S.,&DeJong,G.(1988).Activeexplanationreduction:Anapproachtothemul-tipleexplanationsproblem.ProceedingsoftheFifthInternationalConferenceonMachineLearning(pp.242–255).AnnArbor:MachineLearning.

Ram,A.(1989).Question-drivenunderstanding:Anintegratedtheoryofstoryunderstanding,

memoryandlearning(TechnicalReport710).NewHaven:YaleUniversityComputerScience

Department.

Rieger,C.(1975).Conceptualmemoryandinference.ConceptualInformationProcessing.Ams-terdam:North-Holland.

Riesbeck,C.(1981).Failure-drivenremindingforincrementallearning.ProceedingsoftheSev-enthInternationalJointConferenceonArtificialIntelligence(pp.115–120).Vancouver,B.C.:IJCAI.

Schank,R.(1972).Conceptualdependency:Atheoryofnaturallanguageunderstanding.Cogni-tivePsychology,3(4),552–631.Schank,R.(1982).DynamicMemory:ATheoryofLearninginComputersandPeople.Cam-bridge:CambridgeUniversityPress.

45

Schank,R.(1986).ExplanationPatterns:UnderstandingMechanicallyandCreatively.Hillsdale,

NJ:LawrenceErlbaumAssociates.Schank,R.,&Abelson,R.(1977).Scripts,Plans,GoalsandUnderstanding.Hillsdale,NJ:

LawrenceErlbaumAssociates.Schank,R.,&Leake,D.(1989).Creativityandlearninginacase-basedexplainer.Artificial

Intelligence,(40).AlsoinCarbonell,J.(Ed.),1990,MachineLearning:Paradigmsand

Methods,Cambridge:MITPress.

Segre,A.M.(1987).Ontheoperationality/generalitytradeoffinexplanation-basedlearning.Pro-ceedingsoftheTenthInternationalJointConferenceonArtificialIntelligence(pp.242–248).Milan,Italy:IJCAI.

Shortliffe,E.(1976).Computer-basedmedicalconsultations:MYCIN.NewYork:American

Elsevier.Snyder,C.,Higgens,R.,&Stucky,R.(1983).Excuses:MasqueradesinSearchofGrace.New

York:Wiley.Souther,A.,Acker,L.,Lester,J.,&Porter,B.(1989).Usingviewtypestogenerateexplanationsin

intelligenttutoringsystems.ProceedingsoftheEleventhAnnualConferenceoftheCognitive

ScienceSociety(pp.123–130).AnnArbor:CognitiveScienceSociety.

Swartout,W.(1983).XPLAIN:asystemforcreatingandexplainingexpertconsultingprograms.

ArtificialIntelligence,(21),285–325.

Thagard,P.(1989).Explanatorycoherence.TheBehavioralandBrainSciences,12(3),435–502.VanFraassen,B.(1980).TheScientificImage,chapter5.Oxford:ClarendonPress.Wilensky,R.(1983).PlanningandUnderstanding.Reading,MA:Addison-Wesley.

46