EXPERT SYSTEMS PERILS AUD PROMISE

COhlPUTiNGPRACTICESEdgarH.SibleyPanelEditorBasedonareviewofsomeactualexpert-systemprojects,guidelinesareproposedforchoosingappropriateapplicationsandmanagingthedevelopmentprocess.EXPERTSYSTEMS:PERILSAUDPROMISEDANIELG.BOBROW,SANJAYMITTAL,andMARKJ.STEFIKByvirtueoftheirflamboyantquality,theexpres-sionsartificialintelligence(Al)andexpertsystemshavehelpedcontributetoanexpandingwaveofactivityandunrealisticexpectationsaboutthestateoftheart.Takingalong,hardlookatthoseexpectations,thisarticlecontraststhemwiththeresultsofsomeactualcasestudiesandproposesbothamorerealis-ticviewofthepracticeofbuildingexpertsystemsandsomeguidelinesforchoosingappropriateappli-cations.Thetermexpertsystemreferstocomputerpro-gramsthatapplysubstantialknowledgeofspecificareasofexpertisetotheproblem-solvingprocess.Thetermexpertisintendedtoimplybothnarrowspecializationandcompetence,wheresuccessofthesystemisoftenduetothisnarrowfocus.Justashumanexpertshavevaryinglevelsofexpertise,sotoodocomputersystems.Although,ingeneral,ex-pertsystemshavelessbreadthofscopeandflexibil-itythanhumanexperts,andforthisreasonaresometimescriticizedforcreatingunrealisticexpecta-tions,wefinditmoreproductivetoaskaboutthelevelandrangeofexpertiseofagivenprogram(i.e.,howwelldoesitdoonaspecificsetoftasks),ratherthanstrugglingwiththeimpreciseboundaryofwhatconstitutes“expert.”01966ACMOOOl-0782/86/0900-0880759Insomecircles,thetermsknowledge-basedsystemsorknowledgesystemsareusedinsteadofexpertsystem[lo]tofocusattentionontheknowledgethesystemscarry,ratherthanthequestionofwhetherornotsuchknowledgeconstitutesexpertise.Thesetermsalsoimplytheuseoftechnologyforexplicitrepre-sentationofknowledge.But,onceagain,thebound-arybetweenexplicitandimplicitrepresentationisimpreciseinthatknowledgecanberepresentedex-plicitlytodifferentdegreesandcantakedifferentforms.High-levelperformancecanbeachievedwithoutexplicitrepresentationofknowledgeasinanauto-pilot[%I;onemightevenaskwhetheraC-Compilerorapayrollprogramconstitutesanexpertsystem.Clearlytheybothembodyknowledge:intheonecase,ofalanguageandcomputerand,intheother,ofaccountingandtaxes.Whenconstructedwithconventionalprogrammingtechniques,bothwouldhaveaverylimitedrangeofcapabilities.However,itisalsopossibletobuildeitherasaknowledge-basedsystem.OnemightthenasktheAIpayrollsystemhypotheticalquestions,suchasthenetdifferenceintaxesifyouaddedtwomoredeductions.Intheirusualembodiments,systemslikethesearegenerallydesignedwithbuilt-incommitmentsastohowtheknowledgeembeddedinthemistobeused,thatis,tohave“compiledout”[builtintotheprograms)lim-000CommunicationsoftheACMSeptember1986Volume29Number9ComputingPracticesitedinput/outputbehavior.Tousethissameknowl-edgetogeneratetaxadvice,inthecaseofthepay-rollsystem,wouldrequirerecodingthesystemforthatpurpose.Forthisreason,thetermknowledge-busedisgenerallyreservedforsystemsthathaveex-plicitknowledgebasesandsomeflexibilityintheuseofthatknowledge.Inthisarticle,weexamineprimarilyknowledge-basedexpertsystems:systemsthatachieveexpert-levelperformanceusingexplicitrepresentationsofknowledge.Forpurposesofbrevity,wewillrefertothemsimplyasexpertsystems.APPROACHESTOBUILDINGEXPERTSYSTEMSDependingontheextentanddepthoftheexplicitrepresentationofknowledge,wecandelineatethreedifferentapproachestoexpert-systemdevelopment:thelowroad,themiddleroad,andthehighroad[6].Thelowroadinvolvesdirectsymbolicprogram-ming,usuallyintheLispprogramminglanguage.Ittakesadvantageofnewlow-costAImachineswithflexibleprogrammingenvironmentsenhancedbyuserinterfacesthatexploitwindowsystemsonlargedisplays.Theseenvironmentssupportastyleofpro-gramdevelopmentcalled“exploratoryprogram-ming”[28]inwhichthereisincremental,paralleldevelopmentofprogramspecificationandimple-mentation-anappropriatestyleforapplicationswheretheprimaryconcernisefficiencyandthere-quiredknowledgebaseissmallanddoesnotneedtobechangedveryfrequently.Thelowroadwasusedfortheearlyexpertsys-tems(e.g.,Dendral[19]),whichcombinedAItech-niquesforheuristicsearchwithLispcapabilitiesforsymbolicmanipulation.Dendralgeneratedandtestedhypothesesaboutchemicalstructuresandspectroscopicdata.Itneededtobeefficientbecausethesearchspaceofpossiblesolutionsisverylarge;moreover,programmingitdirectlyintoprocedureswaspracticalsincetheknowledgeusedforinter-pretingspectraldataisfairlystatic.Thehighroad,ontheotherhand,involvesbuild-ingasystemthatcontainsexplicitrepresentationoffairlycompleteknowledgeofsomesubjectmatter,andcanusethatknowledgeformorethanonepur-pose.Asystemiscalled“deep”whenitsknowledgerepresentstheprinciplesandtheoriesunderlyingthesubject;aconsequenceofthisdepthisthatsuchsystemsoftenrequirelongchainsofreasoningfromfirstprinciplestopracticalresults.Sophie[ll]isahigh-roadsystemthatperformsdiagnosticreasoningandqualitativesimulation,andcanreasonfromfirstprinciplesabouthowphysicaldeviceswork.Formanyclassesofdevices,Sophiecandeterminethebehavioralstatesthatthedeviceswilltraverseforagivensetofinputs.Whentheactualoutputofadevicedoesnotagreewiththepredictedoutput,theprogramusesthesamefunda-mentalknowledgetogeneratehypothesesabou