From: AAAI Technical Report SS-92-01. Compilation copyright © 1992, AAAI (www.aaai.org). All rights reserved. Completable Scheduling: An Integrated Approach to Planning and Scheduling MelindaT. Gervasioand GeraldF. DeJong Beckman Institute for Advanced Science and Technology University of Illinois at Urbana-Champaign 405 N. Mathews Ave., Urbana, IL 61801 gervasio@cs.uiuc.edu dejong@cs.uiuc.edu Abstract Theplannlng problem has traditionallybeentreated separately fromthe schedulingproblem.However, as morerealistic domainsare tackled, it becomes evidentthat the problem of decidingonanordered set of tasksto achievea set of goalscannot be treated independently of the problem of actually allocating resources to the tasks. Doingso wouldresult in losingthe robustnessand flexibility neededto deal withimperfectlymodeled domains.Completable schedulingis an approachwhich integrates the twoproblemsby allowingan a priori planning moduleto defer particularplanningdecisions, and consequentlythe associatedschedulingdecisions, until execution time. This allows a completableschedulingsystemtomaximizeplan flexibility by allowingrantimeinformationto be takeninto considerationwhenmakingplanningand scheduling decisions.Furthermore, throughthe criterion of achievability placedon deferreddecisions,a completablescheduling systemis able to retain muchof the goal--directednessand guaranteesof achievement affordedbya priori planning.The eompletableschedulingapproachis further enhancedby the use of contingentexplanation--based learning, whichenables a completableschedulingsystemto learn generalcompletable plans fromexampleand improveits performancethroughexperience.Initial experimental results showthat completable schedulingoutperformsclassical schedulingas well as pure reactive schedulingin a simpleschedulingdomain. Introduction Theplanning problemhas traditionally beentreated separately from the scheduling problem. Planning deals with the determinationof an ordered set of actions for achievinga set of goals. In the context of scheduling domains,planning deals with determiningan ordered set of tasks for a set of jobs. In contrast, schedulingdeals with the actual assignmentof tasks to machinesand is generally concernedmorewith finding the best of several alternative task-machine assignments than with finding aparticular task-machine assignment. As more realistic scheduling domainshave been addressed, however, it has becomeapparent that planning and scheduling cannot be treated independently. The complexity of real-world domainsmakesperfect characterizations difficult to construct and often unwieldy.To this end, researchers in both planning and scheduling have investigated reactive approacheswhich allow for decision-making during execution [Agre87, Firby87, Kaelbling88, Muscettola90, Ow88,Presser89]. However,the classical approachof first doingplanningand then scheduling still remains a problem. Consider giving a classical systemin a process planningdomainthe job of man- 117 ufacturing a particular part. Its planner mustdecide a priori on an orderedset of actions or operations whichwill result in the conversionof the raw material into the desired product. Its scheduleris then given the responsibility of actually allocating resources and carrying out the operations on the machines. However,becausethe planner commitsthe system to a particular set of operations, the scheduler maynot execute the best plan. For example, the planner maynot be able to guaranteethat the efficient newmilling machinewill be available and so choosethe older, slower one. However,during execution, the moreefficient machinemayturn out to be available, but the scheduler does not havethe option to alter the plan. Furthermore, an over-constrained classical plan may prevent quick fixes to unpredictable runtime situations, For example,a chosendrill bit mayturn out to be unavailable,thus rendering invalid those subsequentactions involving a correspondingly sized bolt and wrench. A simple fix wouldbe to use a different drill bit, and switchto the appropriatelysized bolt and wrench, but a scheduler with a completely--determinedplan does not have this capability. A purely reactive approach,with no a priori planning, has its ownproblems. Most manufacturing domains are fairly well-behaved; there is muchinformation available a priori and fairly accurate predictions can be madeabout the behavior of the world under particular circumstances. A purely reactive approach which performs no projection cannot take advantageof this informationto constrain its actions and prevent thrashing. Withthe planning problemand the scheduling problem combined, the runtime decision-making problem also becomesa larger and more complexone. This research beganas an attempt to address the problems with classical, a priori planningand purereactivity. In particular, completable planning was developed as an approach which combinedthe goal--directedness and provably correct plans of classical planningwith the flexibility and ability to utilize runtime information afforded by reactive planning. This enabled a completableplanner to moreefficiently deal with the problemsarising from imperfect a priori information whilestill retaining the benefits of planningbeforehandin relatively well-behaved environments. Morerecently, we have been investigating scheduling, and we have found that many of the techniquesoriginally developedas part of the completable approachto planningare also useful for solving someof the problems which arise from scheduling in realistic domainswhere perfect a priori information about the environmentis unavailable. For example,in the scheduling scenario above, a completablescheduling systemcould defer the deci- sion ofwhich milling machine touseaswell asthechoice of execution.In attachingtwopartsusinga bolt, all that a system bit, bolt, andwrench sizes. During execution, itcanthen use mayknow is that it will be givena bolt of somesize, butit may additional information regarding resource availability toadnot knowprecisely whatsize. However, providedit has acdress thedeferred planning decisions andmake theassociated cessto differentiy-sized bits andwrenches, it canplana driUscheduling decisions aswell. ing operationfollowedby a boltingoperationwithoutspeciInthis paper, wepresent anintegrated approach toplanning fying the precise bit and wrenchto use. Duringexecution, andscheduling called completable scheduling. Wewill first when it is giventhe bolt, it canthendetermine the appropriate give anoverview ofthemain ideas behind completable plan- valuesfor bit size andwrenchsize. Completable scheduling ning, andthen discuss theextension toscheduling. Wewill allowsthe use of conjecturedvariablesto act as placeholders thendiscuss howcomplctable schedules arelearned through for unspecifiedparametersettings providedachievability anexplanation-based learning strategy called contingent proofscan be constructedfor their eventualinstantiation. By EBL. Finally, wewill briefly discuss theimplementation, inallowingdeferred parametersettings, completableschedulcluding some preliminary results andongoing experiments.ing enablesa systemto both plan aheadandyet remainflexible enoughto deal with someuncertainty. Second,a schedulemaybe incompletedueto an undeterCOMPLETABLE SCHEDULING minablenumberof iterations. Animperfectworldmodelmay includeincompletecharacterizationsof operationsandtheir Overview of Completable Planning effects. Consequently, for an actionthat requiresrepetitionto In completableplanning[Gervasio90a,Gervasio90b,Gervaachievesomegoal, the precise numberof iterations needed sio91], a classical planneris augmented with a reactivecommaynot be determinableprior to execution.For example,the ponentwhichprovidesit withthe ability to defer planningdedepthto whicha millingoperationcuts througha part is decisions until executiontime. As an augmentedclassical pendenton the face cutter used.Prior to execution,a system planningapproach,completableplanningretains the advanmaynot knowwhichface cutter will be set up on the machine. tages of classical planningwhilebuyinginto the advantages However, it knows that regardlessof whiehface cutter is used, providedby reactivity. Fromclassical planning,completable the desiredcut can be achievedby simplyrepeatingthe millplanningborrowsthe ability to constructprovably--correct ing operationas manytimesas necessary.Bynot tying itself plans for providinggoal--directedbehavior. Fromreactive to a particularface cutter andconsequently a particular numplanning,it borrows planningflexibility andthe ability to utiber of iterations, duringexecutionthe systemcan chooseto lize runtimeinformationin makingplanningdecisions. Comuse the currentset-up andsavethe cost of changingset-ups, ptetableplanningachievesthe integrationthroughthe achievor it canelect to changeto a moreefficient set-up. Completability criterion, whichrequires everydeferred goal to be able schedulingpermitsthe defermentof iteration decisions provenachievable.Provingachievability requires proving providedincrementalprogress whichconvergesto the goal that there exists a plan whichwill achievethe goal. Ourrecan be proven.Through this deferment,a completable sy stem searchhas shownthat provingthe existenceof a plandoesnot can use imperfectoperationdescriptionsas well as makeoptinecessarilyentail determining the planitself, andthe intuition mizationsto a schedulebasedon runtimeinformation. is that provingachievabilityrequires muchsimplerandmore Third, a schedulemaybe incompletedueto an unidentifireadily available a priori knowledge than full-blownplanable operationchoiceor task-machine assignment.This case ning. Acompletableplanneris thus able to construct proarises whenthere are multiple waysof achievingthe same vably--correctplans in spite of incomplete a priori informagoal fromdifferent states andthe systemlacks the necessary tion, andin doingso providegoal-directednes s to its reactive a priori information for identifyingwhichparticularstate will component whileallowingitself to defer decisionsandutilize bereached.This case also arises whenthere are multipleways runtimeinformationin addressingdeferred goals. of achievinga goal, withdifferentsituationsresultingin different preferencesamongthe various alternatives, andthe Deferring Decisions systemdoes notknow aprioriwhichsituation will be reached. Thedefermentof schedulingdecisionsis a powerfultool in For example,in planningto shapean object, a systemmight dealing with imperfecta priori information.Thecomplexity use someor all of variouscutting operations,suchas milling, of real-worlddomainsmakesit difficult to constructperfect planing,sawing,or grinding.Whether there are severalpossimodels.Evenwhenperfect modelsexist, their use often exble states requiringdifferent operationsormultipleapplicable ceedsreasonablecomputational bounds.Arealistic scheduloperationswith unknown preferences,a systemcan use addiing systemis thus left to contendwith imperfectknowledge. tional runtimeinformationto makea more-informed operaThereare four types of incompleteness whichcan result from tion choice. Completable schedulingallowsa systemto defer usingclassical planningtechniqueson imperfectinformation. operationchoiceprovidedit canprovethat there exists a way First, a schedulemaybe incomplete dueto an unspecifiable to reachthe next state regardless of whichof the possible parameter setting. Withthe lack of a fine-grainedandtractastates is reached.This defermentis useful for tworeasons. ble worldmodel,a schedulingsystemmaynot be able to deFirst, it enablesa systemto use the samescheduleto achieve termineprecise parametersettings a priori. For example,the a goal fromanyof severaldifferent states. Second,it allows parametersof an operationmaybe dependenton the propera systemto applypreferencesto a set of possibleoperations ties of a particular object, whichmaynot be knownprior to using morecompleteand accurate runtimeinformation. 118 Fourth, a schedulemaybe incompletedue to an unorderable set of operations.Imperfecta priori information mayresult in insufficientconstraintsfor completely orderinga set of operations.For example,in the constructionof twopans,the only precedenceconstraints maybe betweenthe milling, drilling, andtappingoperationsfor eachpanmi.e,the operations for the different panscan be orderedin any way.Dependingupona priori knownfactors such as the pans involvedandthe difficulty of changingset-ups as well as a priori unknown factors suchas the initial set-up andmachine availability, particularorderingswill be moredesirablethan others. Bydeferring the decision until all the factors are known,a systemcan utilize runtimeinformationto makedecisions for moreoptimalorderings. Completable scheduling permitsthe defermentof orderingdecisionsprovidedthe different orderingsare all capableof achieving the goal. In doing so, a completable plannercan utilize runtimeinformationin makingmore-informedordering decisions for an unconstrainedset of actions. Proving Achievability Whileimperfecta priori information is the primaryreasonfor deferringdecisions,achievabilityis the primarycriterion for deferment in completable scheduling.Byrequiringthat a deferred goal be provenachievable,completableschedulingenables the construction of incompleteyet provably-correct plans. Previous workon achievability involved finding proofsfor the existenceof plans to achievedeferredgoals. Achievabllityproofs for deferred parametersettings and numberof iterations are discussedin [Gervasio90a,Gcrvasio90b], andfor deferredoperatorchoicein [ Gervasi091 ]. In [Gervasio91],completableplanning wasalso extendedto probabllisticdomains by relaxingthe originalcriterion of absolute achievabilityto probableachicvabllity. Scheduling domainsgive rise to further newissues in achievability.In planning,the mainfocusis findinga plan,or sequenceof actions,whichachievesthe goalfroma giveninitial state. In scheduling,the existenceof severalpossibleschedules is takenas a given, andthe focusis choosingonefrom amongthemusing someset of preferencecriteria, maximizing particular performance measures.Examplesof performmacegoals are meetingdeadlinesand minimizing idle time. Thus,simplydefininga goal to be achievableif there exists a planfor it is insufficientfor scheduling.Achievability must also be relatedto the ideaof optimization andrelative preferencesbetweenpossiblecoursesof action. For example,proving the achievabilityof the goalassociatedwithan unordered set of actions is implicit in the constructionof a nonlinear plan--i.e, actionsare left unordercd if thereare noconstraints requiring precedencerelations betweenthem.Thusthere exists a planfor achievingthe goal.However, thereis the interesting issue of decidingona completeorderingduringexecution. This involvesseekingout additional informationfor evaluatingthe differentoptionsas wellas carryingout the operations themselves.In tying the conceptof achievabilityto optimization,wecanalso better investigatea primarymotivation for combining classical andreactivetechniques:the ability to utilize mntime informationin planning.Goal-directed, 119 robust behaviorin the face of uncertaintyis onereasonfor augmenting a classical plannerwith reactive abilities. However, anotherreasonto integratethe twoapproaches, is to take advantageof the wealthof informationwhichbecomesavailable at runtime.This additionalinformationfacilitates planning byhelpingto focusthe searchfor an appropriateaction. LEARNING COMPLETABLE SCHEDULES Explanation-basedlearning [DeJong86,Mitchell86] has beendemonstrated to be useful in improving the performance of various planningsystems[Bennett90,Chien89,Fikes72, Hammond86, Minton85],and in [Gervasio90a,Gervasio91] wepresent an explanation-basedlearning strategy called contingent EBLfor learning completableplans. Learning completable schedulesbasically involveslearningto distinguish betweena priori planning decisions and decisions whichhaveto be madeor are better madeduringexecution. Learningwhento defer decisionsinvolvesfirst identifying the deferred decision, then constructingan achievability prooffor the associateddeferredgoal. Thena completorfor makingthe deferreddecisionduringexecutionmustbe incorporatedinto the learnedgeneralplan. Identifying Deferred Decisions Amaindifference betweenclassical plans andcompletable plans is the existenceof deferred decisionsin completable plans. In constructingan explanationfor howa giventraining exampleachievesa target goal, an EBLsystemmustexplain howeach action is chosenfor execution.In planning, this usuallymeansverifyingthat previousactionsachievethe preconditionsnecessaryfor the executionof an action. However, withthe additionof reactiveabilities andthe optionto utilize runtimeinformation,a systemneedsto distinguish between a priori satisfied preconditionsandruntime-verifiedpreconditions. Oursolutionis to allowthe systemto distinguishbetweena priori informationandruntime-gathered information andto prefera classicalproofof correctnessto an explanation of achievability.Thus,in explaininghowan action is chosen for execution,a systemfirst attemptsto explainits preconditions witha priori availableinformation.If this is unsuccessful, thenthe actionbeingexplainedis taggedas a potential deferred decision, andthe systemattemptsto construct an achievabilityexplanationfor the precondition.Onlyif it is successfulis the learningprocessallowedto continue.Thefinal explanation will thus containthe identifieddeferreddecisions as well as their supportingachievabilityexplanations. Tyingthe conceptof achievabilityto optimization addsfurther concerns.Anexplanationof executabilityis no longer enough.Explanationsfor preferences mayalso need to be constructed,andas with other deferreddecisions,the associated runtimeverified conditionsneedto be distinguished froma priori satisfied conditions.Aswithproofsof correctnessandexplanations of achievability,explanations of preferences mayalso be constructedin standardEBLfashion. Constructing Achievability Proofs Toconstruct provably-correctplans, a completableplanner mustconstructachievabilityproofsfor the deferredgoalsof its incompleteplans. Whilethe mechanicsof constructing proofsof correctnessvs. proofsof achievabilityare essentially the same--b(~huse standardEBLon a given domaintheory-there are somerequirementsneededfor a domaintheory to be usedin provingachievability. Therearefourtypesof deferreddecisionsandeachrequires particular kindsof informationfor provingachievability. First, deferred parametersettings mustbe represented,and this is doneusingconjecturedvariables. Thesevariablesmay onlybe introducedin the contextof the rules usedto construct their corresponding achievabilityexplanations,thus guaranteeing that everyconjecturedvariable in an explanationhas a supponiugaehievability proof. Second,a systemmustbe able to reasonaboutthe incrementalprogressachievedby a repeatedaction. Thisrequiresactioncharacterizationsto inelude statementsregardingthe changesmadewith respect to somemeasurablequantity. This can then be used to reason about progress towardsthe goal. Third, the incompletely known situation requiringa deferredoperator choicemustbe representedin sucha waythat the systemcan reasonaboutthe spaceof possibilities. Achievabilitycan then be measured in termsof the eoverageprovided bythe alternative actionsover this space.Finally, provingachievabilitywithrespect to an unordered set of operationsis implicitin the absenceof precedenceconstraints betweenthe operations, whichmeansthat anyof the possibletotal orderingswill achievethe goal. Thesecondaspect of achievability, optimality, also imposescertain requirementson the domaintheoryusedto construct explanations.Theheuristics to be usedin making dispatching or schedulingdecisions mustbe built in to the domaintheory. Theseheuristics can then be used both for constructinga priori explanationsandmakingruntimedecisions. In explainingparticular decisionsmadein a training example,a systemcan then construct explanationsincorporating the heuristics and learn general completableplans whichwill employ the heuristics in future applications. Incorporating Completors Thefinal step in learning howto construct a completable scheduleis to incorporatecompletionsteps into the learned generalplan. Thereare four types of completorscorresponding to the fourdifferenttypesof deferreddecision.Thefirst, a monitor,findsa valueto replacea conjectured variable---i.e. it’ determines a specific parametersetting. Thesecond,a repeat loop,repeatedlyexecutesan actionuntil a particularexit condition,the deferredgoal, is reached.Thethird, a conditional, evaluatesthe currentstate anddetermines an appropriate actionbasedon whichconditionsare satisfied. Finally,the fourth, a dispatcher, determinesa completeorderingfor an unordered set of operations,basedona givenset of heuristics. Theachievabilityproofs constructedfor the deferred decisions addressedby these completors are incorporatedinto the explanationssupportingthe learnedplan. Thusthe achievability conditionsguaranteeingthe existenceof a completion are also in the learnedplan. Provided these conditions,along with other preconditions,are satisfied in future instances,a completion is guaranteedto be foundfor the incompleteplan yieldedby the learnedgeneralplan. 120 IMPLEMENTATION A simpleschedulingdomaintheory has beenconstructedto comparethe performanceofacompletableschedulingsystem withthat of a purelyclassical scheduling systemas well as a purely reactive scheduling system. Thedomaininvolves a single machine whichcan be set up in variousways,eachsetup of whichis capableof performingsomeset of tasks. The sametask maytake different processingtimes on different set-ups. Furthermore,there is a set-up cost involvedin changingset-ups. Ajob consists of apartially orderedset of tasks, anda schedulingprobleminvolvesa set of independent jobs. Initially, the onlyorderingcoustmints between tasks are based on deadlines. However,additional precedenceconstraints maybe imposed between the tasksof a job if the a priori planningmodule of the systemdeterminesthat onetask is neededto establish the preconditionsfor anothertask. Uncertainty enters into the picture throughan unknown initial state--i.e, the systemdoesnot knowa priori whichset-up will be onthe machine whenit starts executingits plan. Finally, the goodnessof a scheduleis measured by the length of time takenby the systemto finish a set of jobs. Preliminaryresults showthat a completable system’sability to adapttovarying initial states enablesit to constructmore efficient plans/schedulesthan a classical scheduling,which commits itself to specific set-upsandcompletetask orderings prior to execution. Furthermore,the completablesystem needsless time both to learn a generalcompletable schedule as well as constructa specific completable schedule,although it doesincur the additionalcost of runtimeplancompletion. Thecompletable systemis also able to constructmoreefficient plans than a reactive systembecauseit is morefocused in its searchfor an applicableaction, havingdeterminedas manyprecedenceconstraints betweentasks as it can prior to execution.Althoughboth use the sameheuristics for choosing between multipleapplicableactions, the reactive system has the additionalburdenof sorting out precedencerelations betweentasks during execution. Furthermore,althoughthe completable systeminitially needsto constructa completable schedule,the use of learninghelpsreducethe a priori planning cost it incurs over the reactive scheduler.Weare currently runningexperimentsto gather moredata aboutthe performanceof the three approaches givendifferentdistributionsand different machine/set-up/task-processing profiles. Theresuits are expectedto helpidentifyparticulardomain andproblemcharacteristics whichfavor the different approaches. SUMMARY AND CONCLUSIONS This workintegrates planning--thedeterminationof an orderedset of tasks--andscheduling--theassignmentof those tasks toresource---throughcompletableplans.Becausecompletable plans are incomplete,additionalplanningis necessary duringexecution,whenschedulinghas begunto dispatch the tasks. Thus,this workdiffers fromreactive approaches, suchas those discussedin [Ow88,Prosser89,Smith90,Zweben90],whereplanningis separatedfromscheduling,andthe mainapproachto uncertaintyin the environment is to replan whenthe constraintsof the original planare violated. While replanningis a valuabletool whichanyreal systemwill even- ceedings of The Eleventh International Joint Conferenceon Artificial Intelligence, Detroit, MI, August1989, pp. 590-595. [DeJong86] G. E DeJoag andR. J. Moonoy,"Explanation-Based Learning: An Alternative View," Machine Learning 1, 2 (April 1986), pp. 145-176. (Also appears as Technical Report UII.UENG-86-2208,AI Research Group, Coordinated Science Laboratory, University of Illinois at Urbana--Champaign.) [Drvmmond90] M. Drummondand J. Bresina, "Anytime Synthetic Projection: MaximlzlngtheProbabilityofGoal Satisfaction," Proceedingsof the Eighth National Conferenceon Artificial Intelligence, Boston, MA,August 1990, pp. 138-144. [Fikes72] R. E. Fikes, P. E. Hart and N. J. Nilsson,"learningand Executing Generalized Robot Plans," Artificial Intelligence 3, 4 (1972), pp. 251-288. [Firby87] R. J. Firby, "An Investigation into Reactive Plnnningin ComplexDomains,"Proceedings ofthc NationaI Conference onArtificiallntelligence, Seattle, WA,July 1987, pp. 202-206. [Fox84] M. S. Fox and S. F. Smith, "ISIS--a knowledge-based system for factory scheduling," Expert Systems 1, 1 (July 1984),. [Gervasio90a] M. T. Gervasio, "Learning General Complotable Reactive Plans," Proceedingsof the Eighth National Conferenceon Artificiallntelligence, Boston, MA,August 1990, pp. 1016-1021. [Gervasio90b] M. T. Gervasio, "learning Completable Reactive Plans Through Achievabillty Proofs," Technical Report UIUCDCS-R-90-1605, Department of Computer Science, University of Illinois, Urbana,IL, May1990. [Gervasio91] M.T. Gervasio and G. F. DeJong, "Learning Probably Completable Plans," Technical Report UIUCDCS--R-91-1686, Department of Computer Science, University of Illinois, Urbana,IL, April 1991. [Hammond86] K. H,rnmond, "CHEF: A Model of Case-Based Planning," Proceedingsof the National Conferenceon Artificial Intelligence, Philadelphia, PA, August1986, pp. 267-271. [Kaelbling88]L. P. Kaelbling, "Goals as Parallel ProgramSpecifications," Proceedingsof The SeventhNationaI Conferenceon Artificial Intelligence, Saint Paul, MN,August1988, pp. 60-65. [Knobloek90]C. Knobloek,"learning Abstraction Hierarchies for ProblemSolving,"Proceedings of the Eight National Conference on Artificial Intelligence, Boston, MA,1990, pp. 923-928. [Martin90] N.G. Martin and J. E Allen, "Combining Reactive and Strategic Planning through DecompositionAbstraction," Proceedings of the Workshopon Innovative Approachesto Planning, Scheduling and Control, San Diego, CA, November1990, pp. 137-143. [Minton85] S. Minton, "Selectively Generalizing Plans for Problem-Solving," Proceedingsof the Ninth International Joint Conference on Artificial Intelligence, Los Angeles, August1985, pp. 596-599. [MitcheU86]T.M. Mitchell, R. Keller and S. Kedar--Caballi,"Explanation-Based Generalization: A Unifying View," Machine Learning1, 1 (January 1986), pp. 47-80. [Muscettola90] N. Museettola and S. E Smith, "integrating Planning and Scheduling To Solve Space Mission Scheduling Problems," Proceedings of the Workshopon Innovative Approachesto Planning, Scheduling and Control, San Diego, CA, November 1990, pp. 220-230. [Ow88]E S. Ow,S. Smith and A. Thiriez, "Reactive Plan Revision,"Proceedings of the Seventh National Conferenceon Artificial Intelligence, St. Paul, MN,August1988, pp. 77-82. [Presser89] P. Presser, "A Reactive Scheduling Agent," Proceedings of the Eleventh lnternationalJoint Conferenceon Artificial Intelligence, Detroit, MI, August1989, pp. 1004-1009. [Sacerdoti74] E. Sacerdoti, "Planning in a Hierarchy of Abstraction Spaces," Artificiallntelligence 5, (1974), pp. 115-135. [Smith90] S. E Smith, P. S. Ow,N. Muscettola, J. Potvin and D. C. Matthys, "OPIS: An Integrated Frameworkfor Generating and Revising Factory Schedules," Proceedingsof the Workshopon lnnovativ e Approachesto Planning, Schedulingand Control, San Diego, CA, November1990, pp. 497-507. [Stefik81] M. Stefik, "Planning and Metaplanning (MOLGEN: Part 2),"ArtificiallnteUigence 16, 2 (1981), pp. 141-170. [Zweben90] M. Zweben, M. Deal and R. Gargan, "Anytime Rescheduling," Proceedings of the Workshop on Innovative Approaches to Planning, Scheduling and Control, San Diego, CA,November 1990, pp. 251-259. tnally need, our work first focuses on constructing plans which are as flexible as possible to minimizethe need for failure recovery. In this sense, it is similar to ideas presented in [Drummond90, Martin90]. Drummondand Bresina present an algorithm for maximizingthe probability of goal satisfaction in the case of actions with different possible outcomes, which is one of the problems the conditionals in completable scheduling address. Martin and Allen also prove the achierability of goals deferred to the reactive planner,but they do so using empirical methods, in contrast to the explanation-based methods we use. Completable scheduling may also be viewed as a shallow hierarchical planner, where runtime decisions are at the lowest level. However,unlike other hierarchical planners and schedulers, such as ABSTRIPS[Sacerdoti74], MOLGEN [Stefik81], and ISIS [Fox84], a completable scheduling system uses the achievability constraint to guarantee completability at lower levels. The ordered monotonic hierarchies of ALPINE[Irmoblock90] are a similar idea. The difference is that ALPINEperforms abstraction based on the deletion of literals, while in proving achievability eompletable scheduling uses explicitly more general or abstract knowledgeregarding the deferred goals and their properties. The idea of deferred decisions is not a novel one---the least commitmentprinciple is a basic foundation ofnonlinearplanning, for example. What completable scheduling does is extend the least commitmentprinciple to execution time and in doing so, achieving a well-founded integration of planning and scheduling. Unlike other reactive approaches, in which all decisions are subject to deferment, in completable scheduling only achievable decisions may be deferred. This has two main benefits. The first is that the cost of dynamicdecision-making is minimized, since only some goals must be planned for and scheduled during execution. The second is that the robustness and flexibility afforded by reactivity is gained without losing the goal-directedness and guarantees of success afforded by a prioriplanning. Additionally, the use of contingent EBLenables a completable scheduling system to improve its performance through experience. By learning general completable schedules from example, the system can amortize the cost of constructing a completable schedule over the number of times the learned general schedule is applied in future instances as well as reduce the planning cost incurred by the system’s a priori planning module. Acknowledgments. This research was supported by the Office of Naval Research under grants N-00014-86-K-0309 and N-00014-91-J-1563. We would also like tothank Scott Bennett, Steve Chien, Jon Gratch, and Dan Oblingerfor many interesting discussions, as well as Michael Shaw, for introducing us to the domain of process planning and scheduling. References [Agre87] P. Agre and D. Chapman,"Pengi: An Implementation of a Theoryof Activity," Proceedingsof the National Conferenceon Artificial Intelligence, Seattle, WA,July 1987, pp. 268-272. [Bennett90] S. Bennett, "Reducing Real-world Failures of Approximate Explanation-based Rules," Proceedings of the Seventh International Conferenceon MachineLearning, Austin, TX, 1990, pp. 226--234. [Chien89] S.A. Chien, "Using and Refuting Simplifications: Explanation-basedlearning of Plans in Intractable Domains,"Pro- 121