Druzdzel and Roger R Flynn Decision Systems Lab oratory Sc ho ol of Information Sciences and In telligen Systems Program Univ ersit of Pittsburgh Pittsburgh 15260 marekflynn sis p itt e du httpwwwsispi tt ed u ds app ear in Encyclop di ID: 22525
Download Pdf The PPT/PDF document "Decision Supp ort Systems Marek J" is the property of its rightful owner. Permission is granted to download and print the materials on this web site for personal, non-commercial use only, and to display it on your personal computer provided you do not modify the materials and that you retain all copyright notices contained in the materials. By downloading content from our website, you accept the terms of this agreement.
DecisionSupportSystemsMarekJ.DruzdzelandRogerR.FlynnDecisionSystemsLaboratorySchoolofInformationSciencesandIntelligentSystemsProgramUniversityofPittsburghPittsburgh,PA15260fmarek,flynng@sis.pitt.eduhttp://www.sis.pitt.edu/dslToappearinEncyclopediaofLibraryandInformationScience,SecondEdition,AllenKent(ed.),NewYork:MarcelDekker,Inc.,20021 ContentsIntroduction3DecisionsandDecisionModeling4TypesofDecisions........................................4HumanJudgmentandDecisionMaking............................4ModelingDecisions........................................5ComponentsofDecisionModels.................................5DecisionSupportSystems6NormativeSystems7NormativeandDescriptiveApproaches.............................7Decision-AnalyticDecisionSupportSystems..........................8Equation-BasedandMixedSystems..............................10UserInterfacestoDecisionSupportSystems11SupportforModelConstructionandModelAnalysis.....................11SupportforReasoningabouttheProblemStructureinAdditiontoNumericalCalculations11SupportforBothChoiceandOptimizationofDecisionVariables..............12GraphicalInterface........................................12Summary122 IntroductionMakingdecisionsconcerningcomplexsystems(e.g.,themanagementoforganizationaloperations,industrialprocesses,orinvestmentportfolios;thecommandandcontrolofmilitaryunits;orthecontrolofnuclearpowerplants)oftenstrainsourcognitivecapabilities.Eventhoughindividualinteractionsamongasystem'svariablesmaybewellunderstood,predictinghowthesystemwillreacttoanexternalmanipulationsuchasapolicydecisionisoftendicult.Whatwillbe,forexample,theeectofintroducingthethirdshiftonafactory\roor?Onemightexpectthatthiswillincreasetheplant'soutputbyroughly50percent.Factorssuchasadditionalwages,machineweardown,maintenancebreaks,rawmaterialusage,supplylogistics,andfuturedemandneedalsobeconsidered,however,astheyallwillimpactthetotalnancialoutcomeofthisdecision.Manyvariablesareinvolvedincomplexandoftensubtleinterdependenciesandpredictingthetotaloutcomemaybedaunting.Thereisasubstantialamountofempiricalevidencethathumanintuitivejudgmentanddeci-sionmakingcanbefarfromoptimal,anditdeterioratesevenfurtherwithcomplexityandstress.Becauseinmanysituationsthequalityofdecisionsisimportant,aidingthedecienciesofhumanjudgmentanddecisionmakinghasbeenamajorfocusofsciencethroughouthistory.Disciplinessuchasstatistics,economics,andoperationsresearchdevelopedvariousmethodsformakingrationalchoices.Morerecently,thesemethods,oftenenhancedbyavarietyoftechniquesoriginatingfrominformationscience,cognitivepsychology,andarticialintelligence,havebeenimplementedintheformofcomputerprograms,eitherasstand-alonetoolsorasintegratedcomputingenvironmentsforcomplexdecisionmaking.Suchenvironmentsareoftengiventhecommonnameofdecisionsupportsystems(DSSs).TheconceptofDSSisextremelybroad,anditsdenitionsvary,dependingontheauthor'spointofview.ToavoidexclusionofanyoftheexistingtypesofDSSs,wewilldenethemroughlyasinteractivecomputer-basedsystemsthataidusersinjudgmentandchoiceactivities.An-othernamesometimesusedasasynonymforDSSisknowledge-basedsystems,whichreferstotheirattempttoformalizedomainknowledgesothatitisamenabletomechanizedreasoning.Decisionsupportsystemsaregaininganincreasedpopularityinvariousdomains,includingbusi-ness,engineering,themilitary,andmedicine.Theyareespeciallyvaluableinsituationsinwhichtheamountofavailableinformationisprohibitivefortheintuitionofanunaidedhumandecisionmakerandinwhichprecisionandoptimalityareofimportance.Decisionsupportsystemscanaidhumancognitivedecienciesbyintegratingvarioussourcesofinformation,providingintelligentaccesstorelevantknowledge,andaidingtheprocessofstructuringdecisions.Theycanalsosupportchoiceamongwell-denedalternativesandbuildonformalapproaches,suchasthemethodsofengineeringeconomics,operationsresearch,statistics,anddecisiontheory.Theycanalsoemployarticialintel-ligencemethodstoaddressheuristicallyproblemsthatareintractablebyformaltechniques.Properapplicationofdecision-makingtoolsincreasesproductivity,eciency,andeectivenessandgivesmanybusinessesacomparativeadvantageovertheircompetitors,allowingthemtomakeoptimalchoicesfortechnologicalprocessesandtheirparameters,planningbusinessoperations,logistics,orinvestments.Whileitisdiculttooverestimatetheimportanceofvariouscomputer-basedtoolsthatarerelevanttodecisionmaking(e.g.,databases,planningsoftware,andspreadsheets),thisarticlefocusesprimarilyonthecoreofaDSS,thepartthatdirectlysupportsmodelingdecisionproblemsandidentiesbestalternatives.Wewillbrie\rydiscussthecharacteristicsofdecisionproblemsandhowdecisionmakingcanbesupportedbycomputerprograms.WethencovervariouscomponentsofDSSsandtherolethattheyplayindecisionsupport.Wewillalsointroduceanemergentclassofnormativesystems(i.e.,DSSsbasedonsoundtheoreticalprinciples),andinparticular,decision-analyticDSSs.Finally,wewillreviewissuesrelatedtouserinterfacestoDSSsandstresstheimportanceofuserinterfacestotheultimatequalityofdecisionsaidedbycomputerprograms.3 DecisionsandDecisionModelingTypesofDecisionsAsimpleviewofdecisionmakingisthatitisaproblemofchoiceamongseveralalternatives.Asomewhatmoresophisticatedviewincludestheprocessofconstructingthealternatives(i.e.,givenaproblemstatement,developingalistofchoiceoptions).Acompletepictureincludesasearchforopportunitiesfordecisions(i.e.,discoveringthatthereisadecisiontobemade).Amanagerofacompanymayfaceachoiceinwhichtheoptionsareclear(e.g.,thechoiceofasupplierfromamongallexistingsuppliers).Shemayalsofaceawell-denedproblemforwhichshedesignscreativedecisionoptions(e.g.,howtomarketanewproductsothattheprotsaremaximized).Finally,shemayworkinalessreactivefashionandviewdecisionproblemsasopportunitiesthathavetobediscoveredbystudyingtheoperationsofhercompanyanditssurroundingenvironment(e.g.,howcanshemaketheproductionprocessmoreecient).Thereismuchanecdotalandsomeempiricalevidencethatstructuringdecisionproblemsandidentifyingcreativedecisionalternativesdeterminetheultimatequalityofdecisions.Decisionsupportsystemsaimmainlyatthisbroadesttypeofdecisionmaking,andinadditiontosupportingchoice,theyaidinmodelingandanalyzingsystems(suchascomplexorganizations),identifyingdecisionopportunities,andstructuringdecisionproblems.HumanJudgmentandDecisionMakingTheoreticalstudiesonrationaldecisionmaking,notablythatinthecontextofprobabilitytheoryanddecisiontheory,havebeenaccompaniedbyempiricalresearchonwhetherhumanbehaviorcomplieswiththetheory.Ithasbeenratherconvincinglydemonstratedinnumerousempiricalstudiesthathumanjudgmentanddecisionmakingisbasedonintuitivestrategiesasopposedtotheoreticallysoundreasoningrules.Theseintuitivestrategies,referredtoasjudgmentalheuristicsinthecontextofdecisionmaking,helpusinreducingthecognitiveload,butalasattheexpenseofoptimaldecisionmaking.Eectively,ourunaidedjudgmentandchoiceexhibitsystematicviolationsofprobabilityaxioms(referredtoasbiases).FormaldiscussionofthemostimportantresearchresultsalongwithexperimentaldatacanbefoundinananthologyeditedbyKahneman,Slovic,andTversky[16].Dawes[2]providesanaccessibleintroductiontowhatisknownaboutpeople'sdecision-makingperformance.Onemighthopethatpeoplewhohaveachievedexpertiseinadomainwillnotbesubjecttojudgmentalbiasesandwillapproachoptimalityindecisionmaking.Whileempiricalevidenceshowsthatexpertsindeedaremoreaccuratethannoviceswithintheirareaofexpertise,italsoshowsthattheyalsoareliabletothesamejudgmentalbiasesasnovicesanddemonstrateapparenterrorsandinconsistenciesintheirjudgment.Professionalssuchaspracticingphysiciansuseessentiallythesamejudgmentalheuristicsandarepronetothesamebiases,althoughthedegreeofdeparturefromthenormativelyprescribedjudgmentseemstodecreasewithexperience.Inadditiontolaboratoryevidence,thereareseveralstudiesofexpertperformanceinrealisticsettings,showingthatitisinferioreventosimplelinearmodels(aninformalreviewoftheavailableevidenceandpointerstoliteraturecanbefoundinthebookbyDawes[2]).Forexample,predictionsoffutureviolentbehaviorofpsychiatricpatientsmadebyapanelofpsychiatristswhohadaccesstopatientrecordsandinterviewedthepatientswerefoundtobeinferiortoasimplemodelthatincludedonlythepastincidenceofviolentbehavior.Predictionsofmarriagecounselorsconcerningmaritalhappinesswereshowntobeinferiortoasimplemodelthatjustsubtractedtherateofghtingfromtherateofsexualintercourse(again,themarriagecounselorshadaccesstoalldata,includinginterviewswiththecouples).Studiesyieldingsimilarresultshavebeenconductedwithbankloanocers,physicians,universityadmissioncommittees,andsoon.4 ModelingDecisionsThesuperiorityofevensimplelinearmodelsoverhumanintuitivejudgmentsuggeststhatonewaytoimprovethequalityofdecisionsistodecomposeadecisionproblemintosimplercomponentsthatarewelldenedandwellunderstood.Studyingacomplexsystembuiltoutofsuchcomponentscanbesubsequentlyaidedbyaformal,theoreticallysoundtechnique.Theprocessofdecomposingandformalizingaproblemisoftencalledmodeling.Modelingamountstondinganabstractrepresentationofareal-worldsystemthatsimpliesandassumesasmuchaspossibleaboutthesystem,andwhileretainingthesystem'sessentialrelationships,omitsunnecessarydetail.Buildingamodelofadecisionproblem,asopposedtoreasoningaboutaprobleminaholisticway,allowsforapplyingscienticknowledgethatcanbetransferredacrossproblemsandoftenacrossdomains.Itallowsforanalyzing,explaining,andarguingaboutadecisionproblem.Thedesiretoimprovehumandecisionmakingprovidedmotivationforthedevelopmentofavarietyofmodelingtoolsindisciplinesofeconomics,operationsresearch,decisiontheory,decisionanalysis,andstatistics.Ineachofthesemodelingtools,knowledgeaboutasystemisrepresentedbymeansofalgebraic,logical,orstatisticalvariables.Interactionsamongthesevariablesareexpressedbyequationsorlogicalrules,possiblyenhancedwithanexplicitrepresentationofuncertainty.Whenthefunctionalformofaninteractionisunknown,itissometimesdescribedinpurelyprobabilisticterms;forexample,byaconditionalprobabilitydistribution.Onceamodelhasbeenformulated,avarietyofmathematicalmethodscanbeusedtoanalyzeit.Decisionmakingundercertaintyhasbeenaddressedbyeconomicandoperationsresearchmethods,suchascash\rowanalysis,break-evenanalysis,scenarioanalysis,mathematicalprogramming,inventorytechniques,andavarietyofoptimizationalgorithmsforschedulingandlogistics.Decisionmakingunderuncertaintyenhancestheabovemethodswithstatisticalapproaches,suchasreliabilityanalysis,simulation,andstatisticaldecisionmaking.Mostofthesemethodshavemadeitintocollegecurriculaandcanbefoundinmanagementtextbooks.Duetospaceconstraints,wewillnotdiscusstheirdetailsfurther.ComponentsofDecisionModelsWhilemathematicallyamodelconsistsofvariablesandaspecicationofinteractionsamongthem,fromthepointofviewofdecisionmakingamodelanditsvariablesrepresentthefollowingthreecomponents:ameasureofpreferencesoverdecisionobjectives,availabledecisionoptions,andameasureofuncertaintyovervariablesin\ruencingthedecisionandtheoutcomes.Preferenceiswidelyviewedasthemostimportantconceptindecisionmaking.Outcomesofadecisionprocessarenotallequallyattractiveanditiscrucialforadecisionmakertoexaminetheseoutcomesintermsoftheirdesirability.Preferencescanbeordinal(e.g.,moreincomeispreferredtolessincome),butitisconvenientandoftennecessarytorepresentthemasnumericalquantities,especiallyiftheoutcomeofthedecisionprocessconsistsofmultipleattributesthatneedtobecomparedonacommonscale.Evenwhentheyconsistofjustasingleattributebutthechoiceismadeunderuncertainty,expressingpreferencesnumericallyallowsfortrade-osbetweendesirabilityandrisk.Thesecondcomponentofdecisionproblemsisavailabledecisionoptions.Oftentheseoptionscanbeenumerated(e.g.,alistofpossiblesuppliers),butsometimestheyarecontinuousvaluesofspeciedpolicyvariables(e.g.,theamountofrawmaterialtobekeptinstock).Listingtheavailabledecisionoptionsisanimportantelementofmodelstructuring.Thethirdelementofdecisionmodelsisuncertainty.Uncertaintyisoneofthemostinherentandmostprevalentpropertiesofknowledge,originatingfromincompletenessofinformation,imprecision,5 andmodelapproximationsmadeforthesakeofsimplicity.Itwouldnotbeanexaggerationtostatethatreal-worlddecisionsnotinvolvinguncertaintyeitherdonotexistorbelongtoatrulylimitedclass.1Decisionmakingunderuncertaintycanbeviewedasadeliberation:determiningwhatactionshouldbetakenthatwillmaximizetheexpectedgain.Duetouncertaintythereisnoguaranteethattheresultoftheactionwillbetheoneintended,andthebestonecanhopeforistomaximizethechanceofadesirableoutcome.Theprocessrestsontheassumptionthatagooddecisionisonethatresultsfromagooddecision-makingprocessthatconsidersallimportantfactorsandisexplicitaboutdecisionalternatives,preferences,anduncertainty.Itisimportanttodistinguishbetweengooddecisionsandgoodoutcomes.Byastrokeofgoodluckapoordecisioncanleadtoaverygoodoutcome.Similarly,averygooddecisioncanbefollowedbyabadoutcome.Supportingdecisionsmeanssupportingthedecision-makingprocesssothatbetterdecisionsaremade.Betterdecisionscanbeexpectedtoleadtobetteroutcomes.DecisionSupportSystemsDecisionsupportsystemsareinteractive,computer-basedsystemsthataidusersinjudgmentandchoiceactivities.Theyprovidedatastorageandretrievalbutenhancethetraditionalinformationaccessandretrievalfunctionswithsupportformodelbuildingandmodel-basedreasoning.Theysupportframing,modeling,andproblemsolving.TypicalapplicationareasofDSSsaremanagementandplanninginbusiness,healthcare,themilitary,andanyareainwhichmanagementwillencountercomplexdecisionsituations.Deci-sionsupportsystemsaretypicallyusedforstrategicandtacticaldecisionsfacedbyupper-levelmanagement|decisionswithareasonablylowfrequencyandhighpotentialconsequences|inwhichthetimetakenforthinkingthroughandmodelingtheproblempaysogenerouslyinthelongrun.TherearethreefundamentalcomponentsofDSSs[22].Databasemanagementsystem(DBMS).ADBMSservesasadatabankfortheDSS.ItstoreslargequantitiesofdatathatarerelevanttotheclassofproblemsforwhichtheDSShasbeendesignedandprovideslogicaldatastructures(asopposedtothephysicaldatastructures)withwhichtheusersinteract.ADBMSseparatestheusersfromthephysicalaspectsofthedatabasestructureandprocessing.Itshouldalsobecapableofinformingtheuserofthetypesofdatathatareavailableandhowtogainaccesstothem.Model-basemanagementsystem(MBMS).TheroleofMBMSisanalogoustothatofaDBMS.ItsprimaryfunctionisprovidingindependencebetweenspecicmodelsthatareusedinaDSSfromtheapplicationsthatusethem.ThepurposeofanMBMSistotransformdatafromtheDBMSintoinformationthatisusefulindecisionmaking.SincemanyproblemsthattheuserofaDSSwillcopewithmaybeunstructured,theMBMSshouldalsobecapableofassistingtheuserinmodelbuilding.Dialoggenerationandmanagementsystem(DGMS).ThemainproductofaninteractionwithaDSSisinsight.Astheirusersareoftenmanagerswhoarenotcomputer-trained,DSSsneedtobeequippedwithintuitiveandeasy-to-useinterfaces.Theseinterfacesaidinmodel1AsBenjaminFranklinexpresseditin1789inalettertohisfriendM.LeRoy,\inthisworldnothingcansaidtobecertain,exceptdeathandtaxes"(TheCompleteWorksofBenjaminFranklin,JohnBigelow(ed),NewYorkandLondon:G.P.Putnam'sSons,1887,Vol.10,page170).6 building,butalsoininteractionwiththemodel,suchasgaininginsightandrecommendationsfromit.TheprimaryresponsibilityofaDGMSistoenhancetheabilityofthesystemusertoutilizeandbenetfromtheDSS.Intheremainderofthisarticle,wewillusethebroadertermuserinterfaceratherthanDGMS.WhileavarietyofDSSsexists,theabovethreecomponentscanbefoundinmanyDSSarchitecturesandplayaprominentroleintheirstructure.InteractionamongthemisillustratedinFig.1.Essentially,theuserinteractswiththeDSSthroughtheDGMS.ThiscommunicateswiththeDBMSModelBaseDatabaseMBMSDBMSDGMSDSSUserFigure1:ThearchitectureofaDSSs(afterSage,Ref.[22]).andMBMS,whichscreentheuserandtheuserinterfacefromthephysicaldetailsofthemodelbaseanddatabaseimplementation.NormativeSystemsNormativeandDescriptiveApproachesWhetherornotonetruststhequalityofhumanintuitivereasoningstrategieshasaprofoundim-pactonone'sviewofthephilosophicalandtechnicalfoundationsofDSSs.Therearetwodistinctapproachestosupportingdecisionmaking.Therstaimsatbuildingsupportproceduresorsystemsthatimitatehumanexperts.ThemostprominentmemberofthisclassofDSSsareexpertsystems,computerprogramsbasedonruleselicitedfromhumandomainexpertsthatimitatereasoningofahumanexpertinagivendomain.Expertsystemsareoftencapableofsupportingdecisionmakinginthatdomainatalevelcomparabletohumanexperts.Whiletheyare\rexibleandoftenabletoaddresscomplexdecisionproblems,theyarebasedonintuitivehumanreasoningandlacksoundnessandformalguaranteeswithrespecttothetheoreticalreliabilityoftheirresults.Thedangeroftheexpertsystemapproach,increasinglyappreciatedbyDSSbuilders,isthatalongwithimitatinghumanthinkinganditsecientheuristicprinciples,wemayalsoimitateitsundesirable\raws[13].Thesecondapproachisbasedontheassumptionthatthemostreliablemethodofdealingwithcomplexdecisionsisthroughasmallsetofnormativelysoundprinciplesofhowdecisionsshouldbemade.Whileheuristicmethodsandadhocreasoningschemesthatimitatehumancognitionmayinmanydomainsperformwell,mostdecisionmakerswillbereluctanttorelyonthemwheneverthecostofmakinganerrorishigh.Togiveanextremeexample,fewpeoplewouldchooseto\ryairplanesbuiltusingheuristicprinciplesoverairplanesbuiltusingthelawsofaerodynamicsenhancedwithprobabilisticreliabilityanalysis.ApplicationofformalmethodsinDSSsmakesthesesystems7 philosophicallydistinctfromthosebasedonadhocheuristicarticialintelligencemethods,suchasrule-basedsystems.ThegoalofaDSS,accordingtothisview,istosupportunaidedhumanintuition,justasthegoalofusingacalculatoristoaidhuman'slimitedcapacityformentalarithmetic.Decision-AnalyticDecisionSupportSystemsAnemergentclassofDSSsknownasdecision-analyticDSSsappliestheprinciplesofdecisiontheory,probabilitytheory,anddecisionanalysistotheirdecisionmodels.Decisiontheoryisanaxiomatictheoryofdecisionmakingthatisbuiltonasmallsetofaxiomsofrationaldecisionmaking.Itexpressesuncertaintyintermsofprobabilitiesandpreferencesintermsofutilities.Thesearecom-binedusingtheoperationofmathematicalexpectation.Theattractivenessofprobabilitytheory,asaformalismforhandlinguncertaintyinDSSs,liesinitssoundnessanditsguaranteesconcerninglong-termperformance.Probabilitytheoryisoftenviewedasthegoldstandardforrationalityinreasoningunderuncertainty.Followingitsaxiomsoersprotectionfromsomeelementaryinconsis-tencies.Theirviolation,ontheotherhand,canbedemonstratedtoleadtosurelosses[23].Decisionanalysisistheartandscienceofapplyingdecisiontheorytoreal-worldproblems.Itincludesawealthoftechniquesformodelconstruction,suchasmethodsforelicitationofmodelstructureandprobabilitydistributionsthatallowminimizationofhumanbias,methodsforcheckingthesensitivityofamodeltoimprecisioninthedata,computingthevalueofobtainingadditionalinformation,andpresentationofresults.(See,forexample,Ref.[27]forabasicreviewoftheavailabletechniques.)Thesemethodshavebeenundercontinuousscrutinybypsychologistsworkinginthedomainofbe-havioraldecisiontheoryandhaveproventocopereasonablywellwiththedangersrelatedtohumanjudgmentalbiases.Normativesystemsareusuallybasedongraphicalprobabilisticmodels,whicharerepresentationsofthejointprobabilitydistributionoveramodel'svariablesintermsofdirectedgraphs.Directedgraphs,suchastheoneinFig.2,areknownasBayesiannetworks(BNs)orcausalnetworks[19].Bayesiannetworksoeracompactrepresentationofjointprobabilitydistributionsandarecapableofpracticalrepresentationoflargemodels,consistingoftensorhundredsofvariables.Bayesiannetworkscanbeeasilyextendedwithdecisionandvaluevariablesformodelingdecisionproblems.Theformerdenotevariablesthatareunderthedecisionmaker'scontrolandcanbedirectlyma-nipulated,andthelatterencodeusers'preferencesovervariousoutcomesofthedecisionprocess.Suchamendedgraphsareknownasin\ruencediagrams[15].BoththestructureandthenumericalprobabilitydistributionsinaBNcanbeelicitedfromahumanexpertandareare\rectionoftheexpert'ssubjectiveviewofareal-worldsystem.Ifavailable,scienticknowledgeaboutthesystem,bothintermsofthestructureandfrequencydata,canbeeasilyincorporatedinthemodel.Onceamodelhasbeencreated,itisoptimizedusingformaldecision-theoreticalgorithms.Decisionanal-ysisisbasedontheempiricallytestedparadigmthatpeopleareabletoreliablystoreandretrievetheirpersonalbeliefsaboutuncertaintyandpreferencesfordierentoutcomes,butaremuchlessreliableinaggregatingthesefragmentsintoaglobalinference.Whilehumanexpertsareexcellentinstructuringaproblem,determiningthecomponentsthatarerelevanttoitandprovidinglocalestimatesofprobabilitiesandpreferences,theyarenotreliableincombiningmanysimplefactorsintoanoptimaldecision.Theroleofadecision-analyticDSSistosupportthemintheirweaknessesusingtheformalandtheoreticallysoundprinciplesofstatistics.TheapproachtakenbydecisionanalysisiscompatiblewiththatofDSSs.Thegoalofdecisionanalysisistoprovideinsightintoadecision.Thisinsight,consistingoftheanalysisofallrelevantfactors,theiruncertainty,andthecriticalnatureofsomeassumptions,isevenmoreimportantthantheactualrecommendation.Decision-analyticDSSshavebeensuccessfullyappliedtopracticalsystemsinmedicine,business,8 Figure2:ExampleofaBayesiannetworkmodelingteachingexpendituresinuniversityoperations.andengineering.2Asthesesystemstendtonaturallyevolveintothreenotnecessarilydistinctclasses,itmaybeinterestingtocomparetheirstructureandarchitecturalorganization.Systemswithstaticdomainmodels.Inthisclassofsystems,aprobabilisticdomainisrep-resentedbyalargenetworkencodingthedomain'sstructureanditsnumericalparameters.Thenetworkcomprisingthedomainmodelisnormallybuiltbydecisionanalystsanddomainexperts.Anexamplemightbeamedicaldiagnosticsystemcoveringacertainclassofdisor-ders.Queriesinsuchasystemareansweredbyassigningvaluestothosenodesofthenetworkthatconstitutetheobservationsforaparticularcaseandpropagatingtheimpactoftheob-servationthroughthenetworkinordertondtheprobabilitydistributionofsomeselectednodesofinterest(e.g.,nodesthatrepresentdiseases).Suchanetworkcan,onacase-by-casebasis,beextendedwithdecisionnodesandvaluenodestosupportdecisions.Systemswithstaticdomainmodelsareconceptuallysimilartorule-basedexpertsystemscoveringanareaofexpertise.Systemswithcustomizeddecisionmodels.Themainideabehindthisapproachisautomaticgenerationofagraphicaldecisionmodelonaper-casebasisinaninteractiveeortbetweentheDSSandthedecisionmaker.TheDSShasdomainexpertiseinacertainareaandplaystheroleofadecisionanalyst.Duringthisinteraction,theprogramcreatesacustomizedin\ruencediagram,whichislaterusedforgeneratingadvice.Themainmotivationforthisapproachisthepremisethateverydecisionisuniqueandneedstobelookedatindividually;anin\ruencediagramneedstobetailoredtoindividualneeds[14].2SomeexamplesofapplicationsaredescribedinaspecialissueofCommunicationsoftheACMonpracticalapplicationsofdecision-theoreticmethods(vol.38,no.3,March1995).ThereaderscanexperimentwithGeNIe[7],adevelopmentsystemfordecision-analyticDSSsdevelopedattheDecisionSystemsLaboratory,UniversityofPittsburgh,availableathttp://www2.sis.pitt.edu/genie.9 Systemscapableoflearningamodelfromdata.Thethirdclassofsystemsemployscomputer-intensivestatisticalmethodsforlearningmodelsfromdata[1,11,12,21,26].Whenevertherearesucientdataavailable,thesystemscanliterallylearnagraphicalmodelfromthesedata.Thismodelcanbesubsequentlyusedtosupportdecisionswithinthesamedomain.Thersttwoapproachesaresuitedforslightlydierentapplications.Thecustomizedmodelgener-ationapproachisanattempttoautomatethemostlaboriouspartofdecisionmaking,structuringaproblem,sofardonewithsignicantassistancefromtraineddecisionanalysts.Asessionwiththeprogramthatassiststhedecisionmakerinbuildinganin\ruencediagramislaborious.Thismakesthecustomizedmodelgenerationapproachparticularlysuitablefordecisionproblemsthatareinfre-quentandseriousenoughtobetreatedindividually.Becauseinthestaticdomainmodelapproachanexistingdomainmodelneedstobecustomizedbythecasedataonly,thedecision-makingcycleisrathershort.Thismakesitparticularlysuitableforthosedecisionsthatarehighlyrepetitiveandneedtobemadeundertimeconstraints.Apracticalsystemcancombinethethreeapproaches.Astaticdomainmodelcanbeslightlycustomizedforacasethatneedsindividualtreatment.Oncecompleted,acustomizedmodelcanbeblendedintothelargestaticmodel.Learningsystemscansupportboththestaticandthecustomizedmodelapproach.Ontheotherhand,thelearningprocesscanbegreatlyenhancedbypriorknowledgefromdomainexpertsorbyapriormodel.Equation-BasedandMixedSystemsInmanybusinessandengineeringproblems,interactionsamongmodelvariablescanbedescribedbyequationswhich,whensolvedsimultaneously,canbeusedtopredicttheeectofdecisionsonthesystem,andhencesupportdecisionmaking.Onespecialtypeofsimultaneousequationmodelisknownasthestructuralequationmodel(SEM),whichhasbeenapopularmethodofrepresentingsystemsineconometrics.Anequationisstructuralifitdescribesaunique,independentcausalmechanismactinginthesystem.Structuralequationsarebasedonexpertknowledgeofthesystemcombinedwiththeoreticalconsiderations.Structuralequationsallowforanatural,modulardescriptionofasystem|eachequationrepresentsitsindividualcomponent,aseparableandindependentmechanismactinginthesystem|yet,themainadvantageofhavingastructuralmodelis,asexplicatedbySimon[24],thatitincludescausalinformationandaidspredictionsoftheeectsofexternalinterventions.Inaddition,thecausalstructureofastructuralequationmodelcanberepresentedgraphically[24],whichallowsforcombiningthemwithdecision-analyticgraphicalmodelsinpracticalsystems[9,20].Structuralequationmodelsoersignicantadvantagesforpolicymaking.Oftenadecisionmakerconfrontedwithacomplexsystemneedstodecidenotonlythevaluesofpolicyvariablesbutalsowhichvariablesshouldbemanipulated.Achangeinthesetofpolicyvariableshasaprofoundimpactonthestructureoftheproblemandonhowtheirvalueswillpropagatethroughthesystem.Theuserdetermineswhichvariablesarepolicyvariablesandwhicharedeterminedwithinthemodel.AchangeintheSEMsorthesetofpolicyvariablescanbere\rectedbyarapidrestructuringofthemodelandpredictionsinvolvingthisnewstructure[25].Ourlong-termproject,theEnvironmentforStrategicPlanning(ESP)[6],isbasedonahybridgraphicalmodelingtoolthatcombinesSEMswithdecision-analyticprinciples.ESPiscapableofrepresentingbothdiscreteandcontinuousvariablesinvolvedindeterministicandprobabilisticrelationships.ThepowerfulfeaturesofSEMsallowESPtoactasagraphicalspreadsheetintegratingnumericalandsymbolicmethodsandallowingtheindependentvariablestobeselectedatwillwithouthavingtoreformulatethemodeleachtime.Thisprovidesanimmense\rexibilitythatisnotaorded10 byordinaryspreadsheetsinevaluatingalternatepolicyoptions.UserInterfacestoDecisionSupportSystemsWhilethequalityandreliabilityofmodelingtoolsandtheinternalarchitecturesofDSSsareimpor-tant,themostcrucialaspectofDSSsis,byfar,theiruserinterface.Systemswithuserinterfacesthatarecumbersomeorunclearorthatrequireunusualskillsarerarelyusefulandacceptedinpractice.ThemostimportantresultofasessionwithaDSSisinsightintothedecisionproblem.Inaddition,whenthesystemisbasedonnormativeprinciples,itcanplayatutoringrole;onemighthopethatuserswilllearnthedomainmodelandhowtoreasonwithitovertime,andimprovetheirownthinking.AgooduserinterfacetoDSSsshouldsupportmodelconstructionandmodelanalysis,reasoningabouttheproblemstructureinadditiontonumericalcalculationsandbothchoiceandoptimizationofdecisionvariables.Wewilldiscusstheseinthefollowingsections.SupportforModelConstructionandModelAnalysisUserinterfaceisthevehicleforbothmodelconstruction(ormodelchoice)andforinvestigatingtheresults.Evenifasystemisbasedonatheoreticallysoundreasoningscheme,itsrecommendationswillbeasgoodasthemodeltheyarebasedon.Furthermore,evenifthemodelisaverygoodapproximationofrealityanditsrecommendationsarecorrect,theywillnotbefollowediftheyarenotunderstood.Withoutunderstanding,theusersmayacceptorrejectasystem'sadviceforthewrongreasonsandthecombineddecision-makingperformancemaydeteriorateevenbelowunaidedperformance[17].Agooduserinterfaceshouldmakethemodelonwhichthesystem'sreasoningisbasedtransparenttotheuser.Modelingisrarelyaone-shotprocess,andgoodmodelsareusuallyrenedandenhancedastheirusersgatherpracticalexperienceswiththesystemrecommendations.Itisimportanttostrikeacarefulbalancebetweenprecisionandmodelingeorts;somepartsofamodelneedtobeveryprecisewhileothersdonot.Agooduserinterfaceshouldincludetoolsforexaminingthemodelandidentifyingitsmostsensitiveparts,whichcanbesubsequentlyelaboratedon.Systemsemployedinpracticewillneedtheirmodelsrened,andagooduserinterfaceshouldmakeiteasytoaccess,examine,andreneitsmodels.Somepointerstoworkonsupportforbuildingdecision-analyticsystemscanbefoundin[8,10,18,28].SupportforReasoningabouttheProblemStructureinAdditiontoNu-mericalCalculationsWhilenumericalcalculationsareimportantindecisionsupport,reasoningabouttheproblemstruc-tureisevenmoreimportant.Oftenwhenthesystemanditsmodelarecomplexitisinsightfulforthedecisionmakertorealizehowthesystemvariablesareinterrelated.Thisishelpfulindesigningcreativedecisionoptionsbutalsoinunderstandinghowapolicydecisionwillimpacttheobjective.Graphicalmodels,suchasthoseusedindecisionanalysisorinequation-basedandhybridsys-tems,areparticularlysuitableforreasoningaboutstructure.Undercertainassumptions,adirectedgraphicalmodelcanbegivenacausalinterpretation.ThisisespeciallyconvenientinsituationswheretheDSSautonomicallysuggestsdecisionoptions;givenacausalinterpretationofitsmodel,11 itiscapableofpredictingeectsofinterventions.Acausalgraphfacilitatesbuildinganeectiveuserinterface.Thesystemcanrefertocausalinteractionsduringitsdialoguewiththeuser,whichisknowntoenhanceuserinsight[3].SupportforBothChoiceandOptimizationofDecisionVariablesManyDSSshaveanin\rexiblestructureinthesensethatthevariablesthatwillbemanipulatedaredeterminedatthemodel-buildingstage.Thisisnotverysuitableforplanningofthestrategictypewhentheobjectofthedecision-makingprocessisidentifyingboththeobjectivesandthemethodsofachievingthem.Forexample,changingpolicyvariablesinaspreadsheet-basedmodeloftenrequiresthattheentirespreadsheetberebuilt.Ifthereisnosupportforthat,fewuserswillconsideritasanoption.Thisclosestheworldofpossibilitiesfor\rexiblereframingofadecisionproblemintheexploratoryprocessofsearchingforopportunities.SupportforbothchoiceandoptimizationofdecisionvariablesshouldbeaninherentpartofDSSs.GraphicalInterfaceInsightintoamodelcanbeincreasedgreatlyattheuserinterfacelevelbyadiagramrepresentingtheinteractionsamongitscomponents;forexample,adrawingofagraphonwhichamodelisbased,suchasinFig.2.Thisgraphisaqualitative,structuralexplanationofhowinformation\rowsfromtheindependentvariablestothedependentvariablesofinterest.Asmodelsmaybecomeverylarge,itisconvenienttostructurethemintosubmodels,groupsofvariablesthatformasubsystemofthemodeledsystem.Suchsubmodelscanbeagainshowngraphicallywithinteractionsamongthem,increasingsimplicityandclarityoftheinterface.Fig.3showsasubmodel-levelviewofamodeldevelopedinourESPproject.NotethatthegraphinFig.2isanexpandedversionoftheTeachingExpendituressubmodelinFig.3.Theusercannavigatethroughthehierarchyoftheentiremodelinherquestforinsight,openingandclosingsubmodelsondemand.Somepointerstoworkonuserinterfacesofdecision-analyticsystemscanbefoundin[4,5,28].Summarysupportsystemsarepowerfultoolsintegratingscienticmethodsforsupportingcomplexdecisionswithtechniquesdevelopedininformationscience,andaregaininganincreasedpopularityinmanydomains.Theyareespeciallyvaluableinsituationsinwhichtheamountofavailableinfor-mationisprohibitivefortheintuitionofanunaidedhumandecisionmakerandinwhichprecisionandoptimalityareofimportance.Decisionsupportsystemsaidhumancognitivedecienciesbyintegratingvarioussourcesofinformation,providingintelligentaccesstorelevantknowledge,aidingtheprocessofstructuring,andoptimizingdecisions.NormativeDSSsoeratheoreticallycorrectandappealingwayofhandlinguncertaintyandpreferencesindecisionproblems.Theyarebasedoncarefullystudiedempiricalprinciplesunderlyingthedisciplineofdecisionanalysisandtheyhavebeensuccessfullyappliedinmanypracticalsystems.Webelievethattheyoerseveralattractivefeaturesthatarelikelytoprevailinthelongrunasfarasthetechnicaldevelopmentsareconcerned.BecauseDSSsdonotreplacehumansbutratheraugmenttheirlimitedcapacitytodealwithcomplexproblems,theiruserinterfacesarecritical.TheuserinterfacedetermineswhetheraDSS12 Figure3:Asubmodel-levelviewofadecisionmodel.willbeusedatallandifso,whethertheultimatequalityofdecisionswillbehigherthanthatofanunaideddecisionmaker.AcknowledgmentsWorkonthisarticlewassupportedbytheNationalScienceFoundationunderFacultyEarlyCareerDevelopment(CAREER)Program,grantIRI{9624629,bytheAirForceOceofScienticResearchundergrantsF49620{97{1{0225andF49620{00{1{0112,andbytheUniversityofPittsburghCentralResearchDevelopmentFund.Figures2and3aresnapshotsofGeNIe,ageneralpurposedevelopmentenvironmentforgraphicaldecisionsupportsystemsdevelopedbytheDecisionSystemsLaboratory,UniversityofPittsburghandavailableathttp://www.sis.pitt.edu/genie.WewouldliketothankMs.NanetteYurcikforherassistancewithtechnicalediting.References[1]GregoryF.CooperandEdwardHerskovits.ABayesianmethodfortheinductionofprobabilisticnetworksfromdata.MachineLearning,9(4):309{347,1992.[2]RobynM.Dawes.RationalChoiceinanUncertainWorld.HartcourtBraceJovanovich,Pub-lishers,1988.[3]MarekJ.Druzdzel.ProbabilisticReasoninginDecisionSupportSystems:FromComputationtoCommonSense.PhDthesis,DepartmentofEngineeringandPublicPolicy,CarnegieMellonUniversity,Pittsburgh,PA,December1992.13 [4]MarekJ.Druzdzel.Explanationinprobabilisticsystems:Isitfeasible?willitwork?InProceedingsoftheFifthInternationalWorkshoponIntelligentInformationSystems(WIS{96),pages12{24,Deblin,Poland,2{5JUne1996.[5]MarekJ.Druzdzel.Fiveusefulpropertiesofprobabilisticknowledgerepresentationsfromthepointofviewofintelligentsystems.FundamentaInformatic,SpecialissueonKnowledgeRepresentationandMachineLearning,30(3/4):241{254,June1997.[6]MarekJ.Druzdzel.ESP:Amixedinitiativedecision-theoreticdecisionmodelingsystem.InWorkingNotesoftheAAAI{99WorkshoponMixed-initiativeIntelligence,pages99{106,Or-lando,FL,18July1999.[7]MarekJ.Druzdzel.SMILE:StructuralModeling,Inference,andLearningEngineandGeNIe:Adevelopmentenvironmentforgraphicaldecision-theoreticmodels.InProceedingsoftheSix-teenthNationalConferenceonArticialIntelligence(AAAI{99),pages902{903,Orlando,FL,July18{221999.[8]MarekJ.DruzdzelandF.JavierDez.Criteriaforcombiningknowledgefromdierentsourcesinprobabilisticmodels.InWorkingNotesoftheworkshoponFusionofDomainKnowledgewithDataforDecisionSupport,SixteenthAnnualConferenceonUncertaintyinArticialIn-telligence(UAI{2000),pages23{29,Stanford,CA,30June2000.[9]MarekJ.DruzdzelandHerbertA.Simon.CausalityinBayesianbeliefnetworks.InProceedingsoftheNinthAnnualConferenceonUncertaintyinArticialIntelligence(UAI{93),pages3{11,SanFrancisco,CA,1993.MorganKaufmannPublishers,Inc.[10]MarekJ.DruzdzelandLindaC.vanderGaag.Buildingprobabilisticnetworks:\Wheredothenumberscomefrom?"guesteditors'introduction.IEEETransactionsonKnowledgeandDataEngineering,12(4):481{486,July{August2000.[11]ClarkGlymourandGregoryF.Cooper,editors.Computation,Causation,andDiscovery.AAAIPress,MenloPark,CA,1999.[12]DavidE.Heckerman,DanGeiger,andDavidM.Chickering.LearningBayesiannetworks:Thecombinationofknowledgeandstatisticaldata.MachineLearning,20(3):197{243,1995.[13]MaxHenrion,JohnS.Breese,andEricJ.Horvitz.DecisionAnalysisandExpertSystems.AIMagazine,12(4):64{91,Winter1991.[14]SamuelHoltzman.IntelligentDecisionSystems.Addison-Wesley,Reading,MA,1989.[15]RonaldA.HowardandJamesE.Matheson.In\ruencediagrams.InRonaldA.HowardandJamesE.Matheson,editors,ThePrinciplesandApplicationsofDecisionAnalysis,pages719{762.StrategicDecisionsGroup,MenloPark,CA,1984.[16]DanielKahneman,PaulSlovic,andAmosTversky,editors.JudgmentUnderUncertainty:HeuristicsandBiases.CambridgeUniversityPress,Cambridge,1982.[17]PaulE.Lehner,TheresaM.Mullin,andMarvinS.Cohen.Aprobabilityanalysisoftheusefulnessofdecisionaids.InM.Henrion,R.D.Shachter,L.N.Kanal,andJ.F.Lemmer,editors,UncertaintyinArticialIntelligence5,pages427{436.ElsevierSciencePublishersB.V.(NorthHolland),1990.[18]Tsai-ChingLu,MarekJ.Druzdzel,andTze-YunLeong.Causalmechanism-basedmodelcon-struction.InProceedingsoftheSixteenthAnnualConferenceonUncertaintyinArticialIntel-ligence(UAI{2000),pages353{362,SanFrancisco,CA,2000.MorganKaufmannPublishers,Inc.14 [19]JudeaPearl.ProbabilisticReasoninginIntelligentSystems:NetworksofPlausibleInference.MorganKaufmannPublishers,Inc.,SanMateo,CA,1988.[20]JudeaPearl.Causality:Models,Reasoning,andInference.CambridgeUniversityPress,Cam-bridge,UK,2000.[21]JudeaPearlandThomasS.Verma.Atheoryofinferredcausation.InJ.A.Allen,R.Fikes,andE.Sandewall,editors,KR{91,PrinciplesofKnowledgeRepresentationandReasoning:ProceedingsoftheSecondInternationalConference,pages441{452,Cambridge,MA,1991.MorganKaufmannPublishers,Inc.,SanMateo,CA.[22]AndrewP.Sage.DecisionSupportSystemsEngineering.JohnWiley&Sons,Inc.,NewYork,1991.[23]LeonardJ.Savage.TheFoundationsofStatistics(SecondRevisedEdition).DoverPublications,NewYork,NY,1972.[24]HerbertA.Simon.Causalorderingandidentiability.InWilliamC.HoodandTjallingC.Koopmans,editors,StudiesinEconometricMethod.CowlesCommissionforResearchinEco-nomics.MonographNo.14,chapterIII,pages49{74.JohnWiley&Sons,Inc.,NewYork,NY,1953.[25]HerbertA.Simon,JayantR.Kalagnanam,andMarekJ.Druzdzel.Performancebudgetplan-ning:Thecaseofaresearchuniversity.Inpreparation,2000.[26]PeterSpirtes,ClarkGlymour,andRichardScheines.Causation,Prediction,andSearch.SpringerVerlag,NewYork,1993.[27]DetlofvonWinterfeldtandWardEdwards.DecisionAnalysisandBehavioralResearch.Cam-bridgeUniversityPress,Cambridge,1988.[28]HaiqinWangandMarekJ.Druzdzel.Userinterfacetoolsfornavigationinconditionalproba-bilitytablesandelicitationofprobabilitiesinBayesiannetworks.InProceedingsoftheSixteenthAnnualConferenceonUncertaintyinArticialIntelligence(UAI{2000),pages617{625,SanFrancisco,CA,2000.MorganKaufmannPublishers,Inc.15