On the Role of Humans in Enterprise Control Systems: the Experience of INSPECT AndreValente,Jim Blythe,YolandaGilandWilliamSwartout InformationSciences Institute UniversityofSouthernCalifornia Marinadel Rey,CA90292 valente,blythe,gil,swartout @isi.edu f g Abstract to new situations, user preferences, and institutionalprac- tices. These characteristics indicatethatenterprisecontrol Inthispaper,weusetheexampleofasuccessfulmixed- tools need to be adaptable and responsive to changes in initiativeplan evaluation tool for the domain of air cam- context. paign planning to argue that the human-in-the-loopis an Basedonthisinsightwedrewfromourexperienceswith importantfeature of enterprise control systems. Our tool, INSPECT andothersystems todesignareusablesystem for called INSPECT, evaluates air campaign plans and alerts critiquingand evaluating plans, which can be instantiated the user about inconsistencies and potentialproblems. A andappliedtoany planningdomain, suchas aircampaign generalizationof INSPECT called PSMTool is also capable planningorarmycourseofactioncritiquing. Weusedthis of limited interactionwitha subject matter expert to cap- asaplatformtodevelopaknowledgeacquisitiontoolcalled turenewcritiquesofplans. Thepaperdescribes ourwork PSMToolthatallowsdomainexpertstoenternewevaluation onINSPECTandPSMTool,analyzesthekeycontributionsof criteriaforplans. PSMToolmakesuseofgeneralontologies thesetools, and draws someconclusions abouttherole of thatwedevelopedforplanningandplancritiquingtohelp mixed-initiativetoolsinenterprisecontrolsystems. usersdefineandorganizethecriteriathroughasimpledia- log. InpreliminaryexperimentsatFortLeavenworthBCBL conductedthroughDARPA’sHPKBproject,wefoundthat enduserswereabletousePSMToolwithverylittletraining 1 Introduction toaddnewcritiquesthatwithoutthetoolcouldonlybeadded byaknowledgeengineer. Thissystemextendstheapproach WorkingwithAirForce experts from the“Checkmate” ofkeepinghumanusers intheloopand buildingtoolsthat GroupatthePentagon(HQ USAF XOOC),we developed caneasilybeadaptedinresponsetonewconditions. INSPECT, atoolforevaluatingandcritiquingtheair-related Inthispaper, wefirst describetheINSPECT system: the portion of a military campaign plan. INSPECT integrates problem it solves, its design, architecture and implemen- several AI technologies. It was built using the EXPECT tation. We emphasize some of the lessons learned in this frameworkforknowledge-basedsystemsdevelopment,that respect. ThenwedescribePSMToolandthefurtherlessons incorporates knowledge acquisition techniques, a descrip- from building such a highly adaptable tool. Finally we tion logic-based knowledge representation system, and a presentourlessonslearnedforthecontructionofenterprise sophisticatedproblem-solvinglanguageandreasoner. controlsystems. OurexperienceindevelopingINSPECT andseveralother systems led us to argue formixed-initiativesystems as the 2 The Air Campaign Planning Domain and rightapproachforbuildingenterprisecontrolsystems. The theDesign ofINSPECT aircampaignplanningdomainexpertsweworkedwithfelt thatitwasessentialthatanysystemshouldkeephumanusers “in theloop.” For example, they argued that tools for air INSPECT wasdevelopedasatooltosupporttheaircam- campaignplanevaluationshouldallowuserstounderstand paignplanningprocess. Thetoolhadtoprovidesupportto their plans, identify critical factors, and analyze tradeoffs the adoption and use of a methodology for military plan- amongoptions. Moreimportantly,thesedomainexpertsar- ning called strategies-to-tasks [12, 11]. In this approach, guedthatevaluationcriterianeedtobeadaptedinresponse high-level national objectives are refined into lower level 1 Report Documentation Page Form Approved OMB No. 0704-0188 Public reporting burden for the collection of information is estimated to average 1 hour per response, including the time for reviewing instructions, searching existing data sources, gathering and maintaining the data needed, and completing and reviewing the collection of information. Send comments regarding this burden estimate or any other aspect of this collection of information, including suggestions for reducing this burden, to Washington Headquarters Services, Directorate for Information Operations and Reports, 1215 Jefferson Davis Highway, Suite 1204, Arlington VA 22202-4302. Respondents should be aware that notwithstanding any other provision of law, no person shall be subject to a penalty for failing to comply with a collection of information if it does not display a currently valid OMB control number. 1. REPORT DATE 3. DATES COVERED 1999 2. REPORT TYPE 00-00-1999 to 00-00-1999 4. TITLE AND SUBTITLE 5a. CONTRACT NUMBER On the Role of Humans in Enterprise Control Systems: the Experience of 5b. GRANT NUMBER INSPECT 5c. PROGRAM ELEMENT NUMBER 6. AUTHOR(S) 5d. PROJECT NUMBER 5e. TASK NUMBER 5f. WORK UNIT NUMBER 7. PERFORMING ORGANIZATION NAME(S) AND ADDRESS(ES) 8. PERFORMING ORGANIZATION University of California,Information Sciences Institute ,4676 Admiralty REPORT NUMBER Way,Marina del Rey,CA,90292 9. SPONSORING/MONITORING AGENCY NAME(S) AND ADDRESS(ES) 10. SPONSOR/MONITOR’S ACRONYM(S) 11. SPONSOR/MONITOR’S REPORT NUMBER(S) 12. DISTRIBUTION/AVAILABILITY STATEMENT Approved for public release; distribution unlimited 13. SUPPLEMENTARY NOTES The original document contains color images. 14. ABSTRACT 15. SUBJECT TERMS 16. SECURITY CLASSIFICATION OF: 17. LIMITATION OF 18. NUMBER 19a. NAME OF ABSTRACT OF PAGES RESPONSIBLE PERSON a. REPORT b. ABSTRACT c. THIS PAGE 8 unclassified unclassified unclassified Standard Form 298 (Rev. 8-98) Prescribed by ANSI Std Z39-18 National Goal Maintain survival askedthattheproblemsshouldbepresentedasconstructive criticism,includingsuggestionsofpossiblefixestotheprob- lem. Notsurprisingly,followingthesesuggestionsallowed ustomakeourtoolmoreattractivetousers. National Security Deter/defeat aggression Objectives 3 INSPECT: AnAir CampaignPlanEvalua- tionTool National Military Deter/defeat large-scale, Based on thedesigndecisions described above, we de- Objectives regional, military velopedINSPECT(INtelligentSystemforaircampaignPlans aggression EvaluationbasedonexpeCT). ThearchitectureofINSPECT isshowninFigure2. Campaign Objective Attain air superiority EXPERTS knowledge about: Operational Objective Suppress enemy •air campaign plans air-sortie generation •plan analysis/evaluation Destroy key hardened EXPECT airbase-support facilities Operational Task OPERATION CONCEPT: F15E squadrons with LGBS, supported by EC-13, KC-135 and E-3A INSPECT Figure 1. Strategies-to-Tasks Example (from [Todd 94]) Model of Air Campaign Plans Library of Critiques for Air Campaign Plans militaryobjectivesandthenintooperationaltasks. Figure1 illustratesthehierarchyofobjectives andhow itlinkslow air campaign evaluation leveloperationalactivitiestohigherlevelobjectives. Byen- plans (agenda) couragingplannerstothinkabouthowlow-leveldecisions Plan USERS relatetohigherlevelobjectives,thisapproachpromotesair Editor (air campaign campaignplanningthatismorerationalandhelpsavoidthe planners) sortsofmistakesthatcanarisefromthinkingabouttarget- ingtoolocally. Theresultshouldbeaircampaignplansthat achieve the desired results while reducing risks and min- imizing unnecessary damage. These plans are also more Figure 2. Architecture of INSPECT. structured than traditional plans and capture more of the Afterenteringaircampaignobjectiveswithaplaneditor, rationalebehindtheplan. auserinvokesINSPECTtoevaluatetheplan. INSPECTlooks Our goal was to design an application that integrated for several types of problemsin theplan. For example, it wellwiththismethodologyandtheaircampaign planning checks the hierarchical structure, identifies incomplete or process. Ourdomainexpertshadanumberofusefulsugges- incoherent objectives, and performs rough feasibilityesti- tionsaboutthedesignofthetool,mostnotablyaboutissues mates based on the resources available for the campaign. regardingitsinteractionwiththeuser. First,theyrequested INSPECT shows the user an agenda with all the problems that theuser should notbe locked in by thetool: the tool foundwiththeplan. Theagendaitemsaremarkedaccord- shouldpresentsuggestions,notdefinitiveanswers,andusers ing to their seriousness using a convention very familiar shouldbeencouragedtothinkofalternatives. Second,they to Air Force pilots: WARNING, CAUTION, and NOTE requestedthatallproblemsidentifiedintheplanshouldbe (warningsrequiringimmediate attentions,and notes being accompaniedbyanexplanationandjustification,sothatthe non-critical). In addition to pointing out these problems, user could understand thereasoning behind thetool’srec- INSPECTcanprovideadetailedexplanationofeachproblem ommendationandmakeaninformeddecision. Third,they andalsosuggestwaystofixit. Figure3showsasnapshot 2 oftheagenda,includinganexplanationandsuggestedfixes lemoccurswhenaparentobjectiveissubsumedbyone foroneoftheagendaitems. ofitschildobjectives. Thereareanumberofbenefitsforaircampaignplanning thatstemfromusingatoollikeINSPECT: Catch errors introduced during manual plan de- (cid:15) velopment: Air campaign plans are very large and complex, andneed tobechangedovertime. Because theplancreationprocess ismanual, itisconduciveto introducingerrorsand inconsistenciesin theplan. In fact, INSPECT has found unintentionalerrors in every plan generated by our domain experts which would otherwisehavegoneunnoticed. Raise the floor on plan quality: An evaluation tool (cid:15) likeINSPECThelpsusersavoidcreatinginconsistentor low-qualityplans. Because it works with the higher levels of the plan, it can detect problems that could percolatedowntothelowerlevelsandbeaccentuated astheplandetailsareworkedout. Enforce “good” practices in plan construction: (cid:15) INSPECT’s evaluations enforce the strategy-to-tasks methodology, and a number of style guidelines that experiencedaircampaignplannersdevelopedusingit. Figure 3. INSPECT’s Evaluation of an Air Each of our domain experts liked to hear about the CampaignPlan. The explanation(lowerframe) evaluationcriteriasuggestedbyotherexperts,asanew refers to the selected item on the agenda (up- insight on how they could improve their own work per frame). onplanningaircampaigns. OurexpertslikedthatIN- SPECT wouldpointout thatthey were followingtheir ThetypesofproblemsdetectedbyINSPECTinclude: ownstandardsasitevaluatedtheirplans. Objective with no child/parent. According to the (cid:15) Training new planners: INSPECT’s knowledge base (cid:15) captures the knowledgeof experienced planners, and strategy-to-taskapproachtoaircampaignplanning,all objectives must be subordinated to higher objectives noviceuserscanlearnas theyusethetool. INSPECT’s evaluationspointoutwhatexperiencedplannerswould andbedecomposedfurther(untilapre-specified“leaf” see as serious flaws in theplan, and theexplanations level). ofeach agendaitemaredesignedtobackuptheeval- No objective fulfilling one of the basic tenets of air (cid:15) uations with sources of information in the air cam- power. AirForcedoctrinesuggeststhatanaircampaign paign planning domain (doctrine, typical capabilities plan must contain objectives for all the tenets of air power,suchasforcedeploymentandforceprotection. of weapon systems, etc.). INSPECT’s suggested fixes show what an experienced planner would do differ- Objectivewithtoomanyparents. Anobjectivewithtoo (cid:15) ently. manyparentsisanindicationthattheparentobjective iseithertoogeneral(andshouldthusbedivided)orthat 4 BuildingINSPECT withEXPECT theconnectionsaremeanttoemphasizepriorityrather thanreflectatruedecomposition. Incompatiblesequencerestrictions. Thisproblemoc- INSPECT integratesseveralAItechnologies. Itwasbuilt (cid:15) curswhentemporalconstraintsoftwoormoreobjec- usingtheEXPECTframeworkforknowledge-basedsystems tivesarecontradictory. development,thatincorporatesknowledgeacquisitiontech- Noadequateaircraftcurrentlyavailableforanobjec- niques,adescriptionlogic-basedknowledgerepresentation (cid:15) tive. Thisproblemoccurswhenanobjectiverequiresa system, and a sophisticatedproblem-solvinglanguageand typeofmissionforwhichnoneoftheaircraftavailable reasoner. EXPECT also has a language to express problem isconsideredadequate. solvinggoals that is based on case grammars. Below, we Incoherent decomposition. In principle, an objective brieflydescribethesetechnologies,andhowtheywereused. (cid:15) must be decomposed into objectives which are more Theknowledgeacquisitionbottleneckisfrequentlycited specificormoredetailedthantheirparent. Thisprob- as amajor impedimentto broaddisseminationofAI tech- 3 nology. TheEXPECTproject[6,7]isaddressingthisproblem methods. The EXPECT system then puttogether these two bydevelopingaknowledgeacquisitionframeworkthatem- typesofknowledge,indicatingwhethertherewereanygaps powers people to augment, modify and adapt knowledge orproblems. TheresultofthisprocessisanEXPECT model basedsystemswithoutneedingtounderstandthedetailsof that records all the dependencies between procedural and thesystem’simplementation. ThekeytoEXPECT’sapproach domain knowledge. This model was then passed through isthatitcapturesthedesignrationaleforknowledgebased theEXPECT compiler,thattransformeditintoefficientLisp systems, and uses that design knowledge to guide a user codethatisabletosolvethespecifiedproblem. inaugmentingthesystem. InadditiontoINSPECT, EXPECT has been used to build several knowledge based systems in domains such as transportationplanningand battlefield 4.1 Representing Air Campaign Plans assessment. and Objectives Most knowledge acquisition tools have a fixed set of Averyprominentcontributionofourworkresultedfrom guidelinesorexpectationsabouthowknowledgeshouldbe integratingINSPECTwiththeplaneditortool. Wedesigned addedtoasystem. Theproblemwiththisapproachisthat arepresentationforaircampaign plansandobjectivesthat itis inflexible,and limitstherangeof systems thatcan be wouldallowbothusers and toolsto exchangeinformation supported. EXPECTtakesamoreflexibleapproach: itauto- about the plan. This representation has been adopted by maticallyderives aknowledge-basedsystem fromabstract other planningtools in the air campaign planning domain domainfactsandproblem-solvingmethods. Thederivation throughouttheARPIandJFACCprograms,andisnowseen process is recorded so that EXPECT captures the normally as an important input to an ongoing effort in the US Air implicitdependenciesinaKBS,suchaswhatfactualknowl- Forcetocreatea commonrepresentationofobjectivesand edgeisneededtosupportproblemsolving,andhowfactual tasksforairoperationsplanning. knowledge is used in problem solving. EXPECT provides In integrating INSPECT with the plan editing tool tools that use this informationto guide the user in adding (ACPT), we found a representation gap. Objectives in knowledgeand tools(such as a natural languageexplana- ACPTwererepresentedwithasentencelike“Gainairsupe- tionfacility)thathelpmakeEXPECT’srepresentationsmore riorityinthewesternregion”,or“Destroypetroleumdistri- understandabletonon-computerexperts. Forexample, the butionfacilitiesbeforethe15thdayofthecampaign”. This systemunderstandshowvarioustypesofinstancesareused was an unconstrained string, and the planner could write in problem solving, so when a new instance is added the whatevercametohis/hermind. acquisitiontoolscan make surethatenoughinformationis Inordertobeabletoautomateanyinterestingevaluation specified abouttheinstance so that itcan be used. In this oftheplan,weneededtocapturetheobjectivestatementin way,EXPECTallowsausertoaddknowledgetoaknowledge- aformalrepresentationlanguage. Parsingand interpreting based system without requiring him to understand all the thenaturallanguagesentencewastoocomplex(andanew detailsofhowtheknowledgeinteracts. problem by itself). At the same time, the users were not The EXPECT system is fullyintegrated withthe LOOM willingtowritetheirobjectivesinaformsubstantiallydif- knowledge representation system [9]. LOOM is an im- ferent from theonethey already used. The representation plementation of description logics, which emphasizes ef- weproposedwasthereforeamiddle-ground:weusedacase ficiency and expressiveness instead of completeness. In grammar.1 Thebasicideaofcasegrammarsisthatthereis EXPECT, LOOMisusedtorepresentthefactualanddefini- normallyalimitednumberofroles(calledthematicorcase tionalknowledgeaboutadomain. Forexample,inINSPECT roles)thatanargumentofaverbcanplaywithrelationtothe thereareLOOMdefinitionsaboutwhataretheelementsof verb. Thatwasdefinitelytruefortheobjectivestatements, air campaign plans, what are objectives, what are known forseveralreasons. First,wefoundthattheobjectivestate- typesofaircraft,whatkindsofmissionstheyfly,etc. This mentsfollowedaveryregulargrammaroftheform knowledgehasprovedtobeanimportantbyproductofthe . Second, we found out that there were<sveevrerba>l INSPECT development. It has been used as a basis for the r<ergoullaersit>iesontheuseofthisstructure. Forexample, only developmentofabroadontologyofaircampaignplanning, a handfulof verbs (less than 30) are used. Third, each of whichisbeingusedandfurtherdevelopedundertheJFACC theseverbsintroduceslimitationswithrespect tothetypes DARPAProgram. ofrolesthatcanbeused. Forinstance,mostoccurrencesof INSPECT was built using EXPECT as follows. General theactiontype refertoa(physical)objecttypelike knowledge about air campaign plans, their structure and “missilelaunchDseitsetsr”oyor“militaryheadquarters”. Wewere contents, as well as general domain knowledge about air fightwascodedintoaLOOMknowledgebase. Procedural 1Whilecasegrammarshavebeendismissedasageneralsolutionfor naturallanguageinterpretation, theycanbeaninterestingandpowerful knowledge on how to evaluate the plan according to the deviceinrestrictedsettingssuchastheonewehaveintheaircampaign critiquesspecifiedwasacquiredandrepresentedasEXPECT planningdomain. 4 abletoestablishreasonablyexhaustivelistsofterminalsfor theAirForceonspecifyingvalidtypesoftasksandobjec- eachofthemaintypesspecifiedforrolefillers. Fourth,we tives. Indeed,thesuccesswiththisrepresentationhasledus foundthatcertainroleswereactuallymodifiersthatareused toparticipateinthedevelopmentofspecializedrepresenta- tospecifyrestrictionsorconstraintsontheobjective. There tionsforotherelementsofaircampaignplans,aswellasin arethreetypesofrestrictions,fortime(e.g., ), extendingtheexistingrepresentationofobjectives. space/area( )andresourcweisth(in21days inW)e.sAterdniaRgergamionshowing the struucstuinrgeoBf-5th2es 5 AddingnewcritiqueswithPSMTool pfrroompobseadsegrXamYZmarisshowninFigure4. Action-type (verb) Action/ INSPECT supports the user in simple modification and Role name activity maintenancetasksbyvirtueoftheunderlyingEXPECT sys- tem. However we wanted to provide support for adding Role-specification* newcritiquestoacritiquingtoolsuchasINSPECT. Fromour Object Role object experiences withINSPECT as wellas critiquersforlogistics [Area restriction] planning[8]andarmycoursesofaction[2]weobservedthat Aspect/ plancritiquesoftenfollowoneofa set of genericpatterns state whichcouldbecapturedusinganontologyofplanningand [Time restriction] critiquing. This ontologycan beused by a computer pro- Action gramtoprovideguidanceforaddingnewcritiquesthrough [Sequence restriction] capability dialogue with users. It can also help to organize the cri- tiquesand givea baseline estimate ofthe completeness of thecritiquingsystem,allofwhichhelpstoincreaseuserac- Figure4.Overallstructure ofthecasegrammar ceptanceoftheplansthataregenerated. Moredetailsabout to represent air campaign objectives. the ontologies and their coverage of real-world critiquing taskscanbefoundin[2]. There were several benefits to this representation. On We implemented a knowledge acquisition tool called thesystemic side, it allowed theintegrationof ACPT and PSMToolthatusestheseprinciplestoacquirenewcritiques INSPECT.Asyntax-orientededitorwasbuiltthathelpsthe from users. The tool is domain-independent but requires usersentervalidsentencesbyofferinglistsofvalidcomple- that the domain has been aligned with the generic ontol- tions(accordingtothegrammar)forthetextbeingentered. ogyofplanningandcritiquing,whichitusestoexpressthe Thisprovidesaproactivesupportforusingforthegrammar new critiques. Even with this tool, specifying a new cri- intheeditionofobjectives,withoutunnecessarilyconstrain- tique involves specifying problem-solving knowledge, so ingtheplanner,whostillhasthelibertytowritefreetextif PSMTool also makes use of an editor developed to enter he/shedeemsnecessary. Thecasegrammarbecameashared problem-solving knowledge for the Expect system using representationthat allowedother applicationsto make use English-likesyntax[4]. WetestedPSMToolat FortLeav- ofthemoresemanticrepresentation. enworthwiththehelpofthe BattleCommand BattleLabs Somewhatunforeseenwereanumberofmethodological andfoundthatArmyofficers withverylittletrainingwere benefits,i.e.,thebenefitsofusingagrammartotheplanning abletoaddsignificantnewcritiquestoanArmybattlecourse processitself,independentlyofanytoolsused. Thesebene- ofactioncritiquer. Inthissectionwedescribethedesignand fitsweresoimportantthatthestructuredrepresentationtook implementationofPSMToolandbrieflydescribetheresults alifeofitsownandisoftenseenasakeycontributionofour oftheusertests. workonINSPECT.First,theknowledgeacquisitionprocess PSMTooluses twoontologiestorepresent general plan involvedinbuildingtherepresentationforcedtheexpertsto critiquing strategies: an ontology of plans and an ontol- explain and reflect over the way they write air campaign ogy of critiques. Generic problem-solving knowledge is objectives. For instance, theycame to theconclusionthat attachedtothecritiquesaswenowdescribe. Theplanning frequently occurring objectives like “Conduct operations” ontology,calledPlanet[1],wasdevelopedundertheDarpa- shouldnotbeallowedbecausetheyinfactdonotmeanany- ranHPKBprojectandisageneralontologythatallowsboth thing—inaaircampaignplan,basicallyeverythingcanbe machine-generatedandhuman-generatedplanstoberepre- seenasconductingoperations. Second,theresultinggram- sented and also explicitlyrepresents differentassumptions mar embeds the notion of “reasonable” objectives, which made by planners. It has also been used as the basis of had never been made explicit before then. Third, the ad- atranslationserviceforsoftwareagentscollaboratingona ditionalstructureprovidedbythegrammarwasconsidered planningtask[3]. particularlyuseful for training. Fourth, the case grammar Thecritiquingontologyrepresents domain-independent became an inputto an ongoingstandardization process in critiquesof two types: thosebased on thestructureof the 5 plan and those based on its use of resources. Consider, alternativesthatallowuserstocreateandmodifyprocedures forexample,theproblemsdetectedbyINSPECT thatarede- throughnavigationcanbefoundin[4]. scribedintheprevioussection. 6 ExperienceswithPSMTool Thecritiquesobjectivewithnochild,objectivewithtoo (cid:15) manyparentsandobjectivewithincompatiblesequence restrictionsarebasedonthestructureoftheplan. They PreliminaryexperimentswererunwithPSMToolatFort areindependentoftheaircampaigndomainandthefirst Leavenworth. Four subjects who had been given oneday twoareincludedinourontologyofcritiquesalongwith oftrainingwithExpectandtherepresentationofthecourse theproblem-solvingknowledgerequiredtoimplement ofaction(COA)domainwerepresentedwithfournewcri- them. tiquestoaddtothesystem. Thesubjectswerearmyofficers Theobjectiveincoherent decompositionis also based whowerereasonablyfamiliarwiththeCOAgenerationand (cid:15) on plan structure but requires domain knowledge to whoseuseofcomputersrangedfromreadingemailtohaving know if a parent objectiveis subsumed by one of its beenexposedtojava. Twoofthecritiqueswereaddedusing parents. In this case the critique ontology provides PSMTool and two withoutthe tool. In order to minimize supportforevaluatingaset ofobjectiveswithrespect learningeffects onourresults, twoofthesubjects worked totheirparents, andthedetailsarefleshedoutas part withoutPSMToolfortwocritiquesandthenworkedwiththe ofthedomaindefinition. tool,andtwoofthesubjectsworkedtheotherwayaround. TheobjectiveNoadequateaircraftcurrentlyavailable Our results, although preliminary, show that not only (cid:15) for an objective is a resource-based critique that re- can users add more critiques using the tool than without, quires domain knowledge to implement. Again, the theyarealsofasterandlesserror-pronewhenusingthetool. critiquingontologyprovidespartialsupport. Moreover,thesubjectsreportedthecritiquesthattheyadded usingthetoolasbeingsimpler,eventhoughtheywereinfact PSMToolusestheseontologiestoclassifyanewcritique isomorphic. This providessome supportfor theidea that, added by the user in a question-answeringprocess. Once byclassifyingthenewcritiqueswiththeontologiesasthey the critique is classified in the ontologies, the appropriate areadded,PSMToolhelpsuserstoorganizethecritiquesin genericproblem-solvingknowledgecanbeappliedandthe theirownminds. domain-dependentknowledgethatisrequiredcan beiden- In a second experiment we tested whether PSMTool tified. In many cases, this process also breaks down the couldbeused withlittleornotraining. A fifthsubject, an knowledgethatneedstobeacquiredintosmallchunksthat armyofficerwithnoprogrammingexperience,wasaskedto areeasierforadomainexperttoexpress. addtwoofthesamecritiquesasinthepreviousexperiment, Figure5showsthePSMToolinterfacewhileanew cri- but withoutthe day of trainingin Expect and thedomain, tiqueisbeingadded. Thisisacritiquefromthearmycourse instead the tool was demonstrated and explained for ap- of action domain that checks that thefriendly forces have proximately45minutes. Thesubjectsuccessfullyaddedthe enough force ratio for each of the tasks in the battle, ac- critiques,intimecomparablewiththatoftheearliersubjects cordingtostandardpractice. Thewindowontheleftshows whohadreceivedtraining. how a three-part script is being followed to add the cri- We are currently designing more thorough user exper- tique. In the first part, four questions were answered that iments to test these highly encouraging results. However allowed PSMTool to classify the critique, showing that it theyseemtoindicatethatthegoalofallowingsubjectmat- is a quantity-based check made on each task in the plan. terexpertstoaddnewevaluationcriteriatoaplancritiquer Inthesecondpart,PSMToolexplainshowthecritiquewill bydirectlyinteractingwithan automated system isattain- be implemented based on thisclassification, and identifies able,eveniftheexpertshavelittletrainingwiththesystem. pieces of problem-solvingknowledge to acquire from the Ifthisisthecase,wecanhopetogreatlyincreasetherange user tocomplete thecritique. In the thirdpart, which has of scenarios in which an enterprise control system can be notyet been reached, PSMToolwillrunthecritiqueonan usefullyapplied,byallowingittobemodifiablebyitsusers. existingcourseofactionsotheusercanchecktheresults. Inthecaseoftheforceratiocritique,PSMToolasksfor 7 Lessons Learned: Mixed-Initiative Tools twopiecesofproblem-solvingknowledge: howtoestimate forEnterpriseControlSystems the amount of force ratio required for a task and how to estimatetheamountavailableforatask. Intheright-hand windowinFigure5,theEnglish-basededitorisbeingusedto Inordertoobtaincost savingsandscalability, theideal specifyhowtoestimatetheamountofforceratioavailable scenarioforenterprisecontrolsystemsseemstoarguefora to a task. More details about the editor, how it produces completelyautomatedcontrolsystem. However,ourexperi- an English paraphrase of the procedure body and a set of encewithINSPECThasshownthatinlarge-scale,real-world 6 Figure 5. Adding a new critique with PSMTool. The window on the left shows the questions that were asked to de(cid:12)ne the critique and the new knowledge that is required. In the window on the right, some of that knowledge is entered using the English-based editor. domains, completely automated control is often impracti- campaignplanning,wherelivesareatstake. cableduetothescopeandbreadthofknowledgerequired. As a result of this view on enterprise control systems, Wearenotalonewiththisview. Forexample,[5,10]argue wearguethatthesesystemsshouldadoptamixed-initiative that for planning — a key task in enterprise control — a approachwithtwomaincharacteristics. First,theymustbe mixed-initiativeapproachwheremachinesandpeoplework designedtoworkwithpeople,ratherthancompletelytaking togetherisoftenmoredesirable. over processing. This in turn means that the products of The key insight is that people can understand the en- ourtoolsmustbeunderstandablebypeopleanditmustbe terpriseproblemsand theircontext morebroadlythan ma- possibleforpeopletoeasilyinputinformationanddecisions chines,andthuswillbeabletomakebetterjudgmentsabout intothetools. Second,becauseitisimpossibletoanticipate certain decisions. Machines, on the other hand, are able inadvancealltheknowledgeasystemmightneedinabroad to carry out tasks with a well-defined context much more domain, and because knowledge frequently changes, our effectively. Anotherissueisadaptability: humansareable goal is to provide knowledge acquisition tools that allow tounderstandthat theirknowledgeabouta certain kindof for users to augment and adapt a system’s knowledge in control process is no longer valid and seek to revise this responsetonewsituationsandnewneeds. knowledgeat thelightof the new information. Machines simplydonothavethatcapabilityatthemoment,andthus Acknowledgments behavewithbrittleness. Morover,evenifwedodeveloptechniquesthatcanover- cometheselimitations,itisanopenquestionwhethergiving Theworkreportedhererelatestoresearchsponsoredby allthepowertothemachineisadesirablechoice. Humans theDefenseAdvancedResearchProjectsAgency(DARPA) tend to be afraid of giving up control on certain key de- underFt. HuachucaContractDABT63-95-C-0059andAir cisions, and particularly so when they do not have access ForceResearchLaboratoryAgreementsF30602-97-C-0118 to an explanation as to why the decision was made, what and F30602-97-C-0068. Many thanks to all Checkmate otherdecisions were possible, and what are otherchoices. memberswhohelpedusintheprocess;mostColPlebanek, Black-boxesaresimplynotacceptableformanykeycontrol who allowed us to have this interaction, LtCol Cardenas, decisions, particularly in applications such as military air who coordinated the knowledge acquisition sessions and 7 wasacentralexpert,andtheexpertsweworkedwithmore closely: Maj Allison, LtCol Alred, LtCol Cardenas, Maj Cunico,andMajJackson. Weare alsogratefultothestaff of the Battle Command Battle Labs at Fort Leavenworth fortheirassistanceinrunninguser experiments,especially CaptRasch andColDuquette. Theviewsandconclusions containedinthisarticlearethoseoftheauthorsandshould notbeinterpretedasrepresentingtheofficialpolicies,either expressed or implied, of the Defense Advanced Research ProjectsAgencyortheU.S.Government. References [1] J. Blythe and Y. Gil. Planet: A shareable and reusable ontology for representing plans. Technical report, Expect internalreport,1999. [2] J. Blythe andY. Gil. A problem-solving method for plan evaluationandcritiquing. InProc.TwelfthKnowledgeAc- quisition for Knowledge-BasedSystems Workshop, Banff, Alberta,1999. [3] J. Blythe, Y. Gil, H. Chalupsky,andR. MacGregor. Sup- portingtranslationamongplanningagents. InSubmittedto theFifthInternationalConferenceonArtificialIntelligence PlanningSystems,2000. [4] J.BlytheandS.Ramachandran.Knowledgeacquisitionus- ingandenglish-basedmethodeditor.InProc.TwelfthKnowl- edgeAcquisitionfor Knowledge-BasedSystemsWorkshop, Banff,Alberta,1999. [5] G. Ferguson,J. Allen, andB. Miller. Trains-95: Towards a mixed-initiative planning assistant. In B. Drabble, edi- tor, Proc.Third International Conferenceon Artificial In- telligencePlanningSystems,UniversityofEdinburgh,May 1996.AAAIPress. [6] Y. Gil and E. Melz. Explicit representationsof problem- solvingstrategiestosupportknowledgeacquisition.InProc. Thirteenth National Conference on Artificial Intelligence. AAAIPress,1996. [7] Y. Gil and B. Swartout. Expect: Explicit representations for flexible acquisition. In Proc.Ninth KnowledgeAcqui- sitionforKnowledge-BasedSystemsWorkshop,Banff, Al- berta,1995. [8] Y. Gil andW. R. Swartout. Expect: A reflectivearchitec- tureforknowledgeacquisition. InProceedingsofthe1994 WorkshopoftheARPA-RomeLaboratoryKnowledge-Based PlanningandSchedulingInitiative, Tucson,AZ, February 1994. [9] R.MacGregorandR.Bates. InsidetheLOOMdescription classifier.SIGARTBulletin,2(3):88–92,June1991. [10] K. Myers. Strategic advice for hierarchical planners. In ProceedingsoftheInternationalConferenceonKnowledge Representation,1996. [11] D.Thaler.Strategiestotasks,aframeworkforlinkingmeans andends.technicalreport,RANDCorporation,1993. [12] D.Todd. Strategies-to-tasksbaselineforusafplanning. In- ternal document, Strategic Planning Division, HQ United StatesAirForce,1994. 8