From: AAAI Technical Report FS-94-01. Compilation copyright © 1994, AAAI (www.aaai.org). All rights reserved. Case-Based Plan Recognition in Dynamic Domains Dolores Cafiamero Yves Kodratoff CNRS-LRI Bat. 490, Universit6 Paris-Sud 91405 Orsay Cedex, France { lola, yk}@lri.lri.fr Jean-Fram;ois Delannoy Michel BarBs Dept. of ComputerScience DRET-SDR University of Ottawa 4,medelaPorte d’Issy, Ontario K1N6N5, Canada 00460ParisArmies, dclannoy @csi.uottawa.ca France command, not to issue commands. In the context of Abstract Inthe context ofa Command, Control, Communication and electronic war, commanders find it more and moredifftcult Intelligence (C31) system wehavebuilt a CBRmodule for to simply process the huge amountsof data they are made planrecognition. Itisbased onknowledge representation aware of. The main role of C3I is thus to perform data structures called XPlans, inspired inpartby Schank’s fusion as intelligently as possible. Nevertheless, the same Explanation Patterns. Theuncertainty inherent toan data mayhave very different interpretations depending on uncontrolled flow ofinput andthepresence oflacunary data the context. Evenfor "simple" data fusion, determiningthe make difficult theretrieval ofcases. This ledustodevelop context is equivalent to guessing what the enemy’s analgorithm forpartial andprogressive matching ofthe intentions and plans are. This is why a high-level target caseontosomeofthesource cases. Thismatching interpretation layer is necessaryevenfor a C3I that does not amounts inpractice to a credit assignment mechanism, included inthealgorithm associated witheachXPlan. This pretend more than helping a commander avoid being method hasbeendesigned tomeettherequirements of a bogged under huge amounts of information. l project DRET tobuild a decision support module fora C3I Our system is "real-world" as it has to deal with real system--the MATIS project. Itstasks istointerpret and data--an evolving situation described at a level which is complete theresults ofanintelligent "pattern-recognitionlower than that of self-contained actions, and which and-data-fusion" module inorder tomaketheintentions consists of punctual events. Our data have been partially underlying therecognized situation explicit tothedecisioncollected by Intelligence, and we must adapt to their maker. Thisadvice isgiven asa causal explanation ofan knowledgerepresentation for our input/output (our other agent’s behavior from low-level information. information sources are the 3 lower levels of MATIS,see next section). Thesedata present six undesirable features: Introduction 1. Measures are taken in a fixed way. This means that incoming data may not contain the information In thecontext of a Command, Control Communication and characterizing a situation. For example, in order to Intelligence (C3I)systemwe havebuilta Case-Based recognize that a car stopped at a traffic light, one usually Reasoning (CBR)module forplanrecognition. It is based assumesthat_ the car is recordedwhile stoppedat the light. on knowledge representation structures calledXPlans, This maynot be true for our data, which mayonly report of inspired in partby Schank’s Explanation Patterns (Schank, a car slowing downbefore a light, and the measurementof 1986;Schankand Kass,1990),but whichhavebeen the car at nullspeed maybemissing. designed torecognize an agent’s plansandmotivations in 2. Since measurements performed at somepointmaytake dynamicdomainsandfromlow-level data.XPlansare some time to reach the interpretation center, thetemporai basically prototypical casesofplanscontaining both, the ordering of input data may not be the same as the ordering standard behavior ofan agentinsomegivencircumstances that occurred in the field. Consequently, a hypothesis must (planningknowledge), and knowledgeabouthow neverbe discarded on thegrounds of missing dataat a interpretthe behaviorof an agent in that same given time, sinceconfwming datamayarrive later. circumstances (recognition knowledge). Plausible plans 3.There aredifferent levels ofgranularity inthedata. For reflecting thestandard behavior in thecircumstances in whichtheobserved agentcurrently findshim/herself are example, someinformation may be veryrelevant for a combat section andnotatallfora company, etc. usedto interpret theactual agent’s behavior. Theyare progressively confu’mcd or rejected depending on the 4. Forevery level of granularity, theunitsareperforming compliance oftheobserved withthestandard behavior. tasksat whichtheynormally collaborate. Eachagentis Ourplanrecognition system isa module tobeintegrated intelligent initsway,butitisnotfreetothreaten inany ina C31system. Current C3Isystems areintended to help waythegeneral plan(ona battlefield, "backtracking" usually implies heavylosses). Freedom is strongly limited, i DRET is a service of the French Minist~re de la D#fense butnever totally absent. Nationale. 5.Dataarelacunary andcanbenoisy. 16 6. Datacontain informationthat represents actions which take place at specific locations and at specific points in time. Aseventhrequirementof a different nature mustalso be met. Since a C3I systemis intended to help a commander whohas heavyresponsibifities, andwhois held personally responsibleof his/her decisions,it is not possibleto make decisions on data fusion in a waywhichis not directly understandable to the commander.Answershave to be presented to the commander in his/her ownlanguage.This explanatoryfacultyis veryhigh-level,but it is also another requirement of real-life: a C3I system which does not explainits actionswill at best be reluctantlyacceptedunder heavyconstraints. This seventhrequirementis especially relevant for decision supportsystemswhichare to be used in critical domains. Strong specifications as the seven above cannot be exactlymetin the presentstate of the art, that saystittle about recognitionfromevents, or fromnoisy andlacunary input, althoughthese questionsare classic in statistical signal recognition and in the machinelearning community (Kodratoff et al., 1990). Thedifferent approachesonly providepartial answersto these problems.Plan recognition is generallyconsideredin an ideal context, wherethe input is supposedto be 100%correct (violating requirement5) and consists of actions, that is, chunks of activity representingwell-definedtransitions in the si01~tion(thus violating requirement1). Suchideally clean contexts can be described in strong-theory domains, and makeit possible the application of formal techniques such as abduction (Charniak and McDermott,1985; Kautz, 1991; Allen et at, 1991;van BeckandCohen,1991)or syntactic parsing(Vilain, 1990).Lacunarityis addressedby (Pelavin, 1991) whorefines on Kautz by incorporating partial descriptions of the world(correspondingto our lacunary input) andexternal events(as in our notionof an evolving situation). However,he does not consider false data. Uncertaintyhas been well addressedby probabilistic and spreading action approaches, whichhave in common the use of a confidencevalueattachedto the plans retained as hypotheses. (Charniak and Goldman,1991) propose probabilistic model of plan recognition based on a Bayesiannetwork,wherethe input is low-level, and the output as well, whereashigh-level output is one of our requirements.Modelsof spreadingactivation (Quast, 1993; Langeand Wharton,1993) maybe moreflexible as they allowa moreversatile specificationof the behaviorof the library nodesin the recognitionprocess.Withinthe framebasedapproachto plan recognition, (Schankand Riesbeck, 1991) have a powerfuluse of conunonsense knowledgeand of explanation questions, but cannot interpret evolving situations. In the narrative understandingsystemGENESIS (Mooney,1990a&b)combinescase-based recognition and explanation-based learning for automatic schema acquisition. However,this approach do not address problemsof lacunarity, uncertainty, and low-levelinput. (Laskowskiand Hofmann,1987) have efficient belief and 17 time-handling mechanisms,but their approach is again basedon actions, not on events. In order to cope at least partially with all these requirements, we have developed a form of knowledge representation called here XPlans. These knowledge representationstructuresare essentially cases or prototypes of plans whichare progressivelyrecognizedas the indices of the field are gathered. Theygive an interpretation of low-levelreal data in terms of a plan applied by an agent and of the causes and motivations that explain this behavior. The MATIS Project Ourplan recognition systemis to be integrated in MATIS (Mod~le d’Aide d la ddcision par le Traitement de l’Information Symbolique--"modelof decision-aid by symbolic information processing"). The MATIS modelis designedto providesupport for military decision (Bar~s, 1989)in critical or difficult situations,withinthe contextof C3I. However,it can obviouslybe of use in all civilian contexts in whichdecision makingbecomesdifficult. It coversa comprehensive range of tasks, fromthe processing of basic raw data (generally numeric)to the handling "knowledge objects", i.e., elementswhichcarry somedirect relevancefor the decision-maker. MATISArchitecture Somehypotheses are used to somehowanticipate the situation, or to enumeratepossiblebehaviorsof the enemy, by taking into account some knowledgestored in the memory of the system. This knowledgeconsists of facts already knownabout the situation. Facts are confirmedby the doctrine (see Figure 1). There is no need for these hypothesesto be exact at fast, since they are mainlyused to initialize the model,andthey will be changedaccording to the situation as perceivedby the upper layers. These hypotheses can be seen as a layer previous to MATIS interpretationlayers (Figure1): ¯ Thefast layer consists of a perceptualsystemmadeof a numberof sensor families whichtrack various physical parameters.It performsdata fusion on basic data observed by the sensors---observations.Thenumericoutput fromthe sensor families is convertedto symbolicformatby a fast interpretationtreatment,called perceptionin Figure1. ¯ Thesecond layer identifies elementsrelevant to the action, e.g., a car, the orderof battle, etc. It attributes a nameto eachrelevant elementor entity. ¯ Thethird layer groupsentities into significantcollections of objects---association. Thesethree layers progressivelyenrich the information flowing through the reasoning system of MATIS, mainly by increasing the semantic content of each knowledge object. Our plan recognition system can be seen as a fourth layer. It handlesthe semanticallyrichest informationof the third layer. Thefourth layer doesnot bear just on objects, but rather on their behavior--using XPlans, it interprets this behavior in terms of predefinedplans and goals. incremental evidence assignment. This robust approach is adopted in our system. In addition to the requirements imposed by the nature of data, several characteristics follow from this robust approach: ¯ The input consists of an imposedsequence of events (not of actions): (a) they are snapshots of the observedactivity (they have no duration), situated at whatever point of the behavior of a car; on the contrary, an action spans some amount of time; Co) they ate measured at any time, not necessarily at the time of a qualitative transition. In fact, qualitative transitions have to be reconstructed from events---for example, the system must find that the car is turning left, based on the position and speed of the car at various instants. ¯ The system must be able to process massive input, and to smoothout the incidence of false data. ¯ The output must include a list of plausible plans, the confidence rate of which must be incrementally confn’med. ¯ The system must be able to interpret low-level data in terms of high-level concepts so that the advice provided can be easily understood by the user. This advice must also be satisfactorily justified. 0 Field dire (eo.radar 0 [~¢ture s. I nft’lred pl ¢1~Jres, et¢~ m Percaptlon Red wodd - Symbolic world Figure 1: MATISarchitecture and knowledgeprocessing. Our plan recognition system can be seen as a fourth layer. Feedbackoccurs only through factual knowledge. Finally, a presentation layer displays the semantically enriched data in different strata depending on the user’s needs. These strata correspond to the following levels in the objects perceived: Shapes, like elements of a vehicle, objects, like vehicles, collections of objects, e.g., armored companies, relations amongobjects and their collections, elementary moves, e.g., tank # 32 moves forward, or company# 3 moves Eastwards. In MATIS there is no feedback from the semantic levels onto the numericdata, contrary to the present tendency to require high-level understanding to help correct potential errors in the numericdata wheneverthey induce high-level contradictions. In fact, there is a quite strict sequential process of relevance enrichment, i.e., while compressing information, the understandability of knowledgeobjects is progressively improved. Similarly, there is no feedback from a higher semantic level to a lower one, and the informationcontent must strictly increase fromthe lower to the higher levels. The only exception to this rule is the possibility to add new facts into the factual knowledge base. Addedfacts can come either from the user, or from the plan recognition module. A consequenceof this "almost no-feedback" policy was a stringent requirement on the robustness of the interpretation modules, that must inherently be able to handle noisy and irregular data. Constraints on Plan Recognition In real conditions of operation C3I systems are unavoidably faced with massive amountsof low-level data, which are not altogether reliable, and maylack at crucial moments. Such a context imposes weak-theory modeling and 18 The Problem Situation The domainchosen for the initial modelinghad to present the features listed above. In a car driving domain, our system must face the following type of problems. A car-the agent--is moving.Its driver is supposedto be a rational agent with goals or motivations (e.g., saving time, saving gas, etc.) that will lead higher to drive in a certain way rather than another in order to satisfy a particular goal in a specific situation. The observer--the recognition system with cooperation from the decision maker--cannot see the car directly. In principle, it knowsnothing about the agent. The recognition situation is a case of keyhole recognition (Cohen et al., 1981), i.e., the agent is unawareof being observed, and the observer cannot attribute to the actor any intention to cooperate in the recognition process. The only information the observer has available are sets of low-level data such as speeds, positions, and distances, whicharrive periodically. As additional assumption, the observer considers this agent "at the knowledgelevel". That is, it considers this agent, which is embedded in a task environment, as composedof bodies of knowledge(about the world), someof whichconstitute its goals, and someof which constitute the knowledge needed to reach these goals---the agent processes its knowledgeto determine the actions to take. Its behavioral law is the principle of rationalitym"If an agent has knowledgethat one of its actions will lead to one of its goals, then the agent will select that action" (Newell, 1982)--but in its more pragmatic form of the twO-step rationality theory (van de Velde, 1993). In this form, the practical application of the principle of rationality is constrained by the boundariesof a model of the problem-situation (Catiamero, 1994), which takes into account the specific circumstances in which the task is to be solved--’task features’. This way, the assumptionis madethat the behaviorof the agent is both rational and practical. Basedon this assumption,and from the availablelow-leveldata, the observerfiaust interpret the car driver’s behavior. In particular, it mustexplain: (a) whatthe driver seemsto be doing,e.g., overtakea car; Co) hows/he doesit, e.g., faster than it wouldbe safe; and(c) whys/he acts this wayrather than another--whathis/her goalsor motivationsare (savingtime, etc.). Plan Recognition Based on XPlans plan recognitioncan be definedas the interpretation of a changing situation. It can be seen as a process of understandinginverseto that of planning0Vilensky,1981, 1983):while planninginvolvesthe constructionof a plan whoseexecutionwill bring abouta desiredstate (goal), plan recognitionthe observerhas to followthe goals and plans of actors in a situation in order to makeinferences. This process implies making sense out of a set of observations by relating them to previous knowledgeso that the observer can integrate those observationsin an overall picture of the situation. In other words, plan recognitionhere is consideredas understandinga problem situation by relating someobservations to previous knowledge about an agent’s goals and plans. Accordingto Schank (Schank, 1986) this form of understanding closelyrelated to twoother cognitiveprocesses: ¯ Explanation.Theactions of others can be explainedby understandinghowthey fit into a broader plan. "Saying that an action is a step on a coherentplan towardsa goal, explainsthat action."(p. 70). ¯ Learning.Learningis usually definedas adaptationto a newsituation, but for Schankit also means"finding or creating a newstructure that will render a phenomenon understandable." (p. 78). In this line, weuse a set of patterns to understandand explain the behavior of an agent. Our approach is to interpret a coherent behavior in terms of schemataor frames, with a double presupposition concerning the observedagent: ¯ Intentionality. The agent has somepurpose. This excludesinexplicablebehavior. ¯ Rationality. Theagenthas the faculty of pursuinga goal by setting him/herself some subgoals. This excludes randombehavior. These assumptions allow us to explain and make predictions about the future behaviorof the agent on the basis of previousknowledge.Previousknowledgeconsists of a library of prototypicalplans indicatingwhata standard (ideally rational) agentshoulddo in different situations obtaincertain goals. Structure and Organization of XPlans XPlans are inspired in part by Schank’s Explanation Patterns. However, they present some substantial differences, since the contextweare workingin is entirely 19 different fromSchank’s.Wehaveto consider an evolving situation in whichmassive,uncertain and lacunar), input correspondingto events arrives in an uncontrollableflow. Therefore, we need a moreelaborated domainknowledge in orderto give a logic andunderstandable interpretationto row data that do not have a meaningby themselves. Additionally,weconsidertwo points of view: (a) that of standard agent, whois supposedto behavein a rational way,i.e., planningsequencesof actions in order to attain some goaismplanning knowledge; and (b) that of observerwhotries to understandwhata real agentis doing by comparing the information s/he gets to his/her knowledgeabout what a standard agent should do in that circumstances---recognitionknowledge.Weneedthus two types of semanticallydifferent rules: planningrules and recognition rules, whichare based on the planningones. Bothtypes of rules are integrated in complexstructures~ schemata---including,amongother information, the rule application context. Recognition knowledgeis highly contextual,since oneaction can respondto different causes in various contexts, and hence must be given different interpretations. For example,excessive speed wouldbe associatedto different interpretations in the contextof a plan for overtakinga car or in the context of a plan for approachinga red light. Among other information, XPlans include the followingelements: ¯ Parameters:List and description of the variables used, includingconfidencevalue. ¯ Preconditions:The basic facts that musthold for the plan to be considered.Whenthey hold, the plan is said to be active andits recognitionalgorithmis applied. ¯ Planningalgorithm(PA): Descriptionof the behaviorof a standard(ideal) driver in the situation describedby the XPlan.It consists of the sequenceof phasesthat the agent should follow to accomplishthe plan. Every phase has associated twodifferent types of rules: activation rules, whichcall anotherplan (and optionallydeactivatethe plan) whenpreconditions do not hold anymore, and behavior rules, whichresult in the modificationby the agentof the value of one or moreparameters. ¯ Recognitionrules (RR): Theycontrol the modificationof the confidence value of the plan, and assign an interpretation to the observed behavior, according to behaviorrules in the PA. ¯ Postconditions: Theydescribe sufficient conditions-elementsof the worlddescription after completionof the plan--for deactivatingthe plan, both in the case the plan 2. has succeeded andin the case it has failed 2 Failure conditions are considered because something external to the agent mayhappenthat prevents it from accomplishingthe plan (e.g., car2 suddenlystopped and cart collides into care). This is becausein this domainthe situation changesnot only as a consequence of the agent’s behavior,but also independently of it. ¯ Motivations: Theyrepresent the goal pursued by the agentwhenadoptinga specific strategy rather than another to accomplish a ta~k. XPlansare organizedin the formof a hierarchy(Figure 2) whichconstitutesa library of prototypicalcases. "apgoach "a~coach "ad ~--ed green oa ~t" ~ ~ ~¢h red Ight" -’~" ............... , u~u. "~.do~,~.~o~" "p~.s rea llgh~ "plum ~’een Bgh~ .~~~ bellmdmoth~rcar" "slow anoth~ downbdoce bdm~d ~, ~h~ Figure 2: A subpart ~.~~lJ~et down bd~ ¢~,red~ l~l~ beh~ of the hierarchy of XPlans relative to driving. There are two maintypes of links in this taxonomy.The first type of link (dark arrows)expressesthe Spec~ii7~tion of the context (preconditions) in whichthe XPlantakes place. The XPlanthat this type of link gives rise to correspondsto a task that the agent must perform. The other type of link (gray arrows)expressesthat the child XPlan is a consequenceof a choice of the driver---a prototypicalstrategy chosento performa task accordingto a certain motivation. deactivated in four ways: (a) If their postconditionsare satisfied; (b) Whencalling another, a plan maydeactivate itself if it becomes useless; (c) Below a certain threshold the confidence value (0.1), a plan is automatically deactivated. A low confidence value can be reached throughapplication of recognition rules with a negative effect on the confidence--reflectingdiscrepancybetween the driver’s courseof action andthe standardbehavior--or whenfor several events no recognition rule will apply (attrition mechanism); (d) Whena plan has been successfullycompleted,its concurrentsare deactivated. The general loop for the processing of an event is as follows: Initially, only the root node is active. The postconditionsof active plans are examined first; the plans whosepostconditionshold are deactivated,as well as their concurrents.Thena depth-first search on the preconditions of childrenplans is started. If the preconditionsof a plan hold, it is activatedandthe parent planis deactivated.At this point, a recognition algorithmis applied to every single active plan. If the recognitionrules of a plan cannot be applied,its confidencevalueis slightly decreased.If the confidence value of a plan is smaller than a certain threshold (fixed to 0.1), that plan is automatically deactivated(its confidencevalueis forceddownto 0). The recognition algorithmapplied on each plan is the following.For everyneweventin the input, the activation rules are checkedand triggered if applicable; other plans are activatedas specifiedin that rules, andthe plan itself can be deactivated. If the plan is still active, for each applicable behaviorrule, the ideal value of the parameter used as a reference is calculated and comparedto the measuredactual value. Then,basedon their difference, the confidencevalueof the plan is adjusted--it is increasedif the differenceis smallor nufl, decreasedelse. Figure3 showsthe evolutionof the confidencevaluesof three plans: at the beginning,the three of themcould be plausiblehypothesesto interpret the driver’s behavior,but only one of themwill be successfullycompletedand hence confmned. The Recognition Process The plan recognition process consists basically of the retrieval of cases (prototypical plans) from the XPlan hierarchyin order to interpret andexplainthe behaviorof the agent fromthe low-leveldata feeding the system.The uncertainty inherent to an uncontrolledflow of input and the presenceof lacunarydammakedifficult the retrieval of cases. This led us to developan algorithmfor partial and confidence value progressivematchingof the target case onto someof the successfully sourcecases stored in the library. This matchingamounts initial values of the 1 ~pleted in practice to a credit assignmentmechanism, includedin a recognitionalgorithmassociatedwith eachXPlan. In orderto recognizethe plan that the agentis applying, wetry to followhis/her steps by producinginterpretative concurrent to hypotheses, as specific as data allow, and prudently ~. 0..~ ~ successful plan confu’mingor disqualifying themas newdata arrive. The ~ow dropped mechanism we propose is somehowcomparable to spreadingactivation: plans activate oneanother downthe under a minimumthreshold: "~.~_ hierarchical structure of the plan library, or across the abandoned hierarchy. Severalconcurrenthypothesescan be plausible interpretations of the agent’s behavior. This meansthat succession of "events" (input information) time multipleplans can be simultaneouslyactive. Plans can be activated in two ways:(a) Forwardactivation of the most Figure 3: Evolutionof the confidencevalues of several specific set of preconditions.If a plan is active, it checks plans. the preconditionsof its subtree. If the preconditionsof a planin the subtreehold, it is activated;(b) Activationof plan by another through activation rules. They can be 2O Whenall the plans in a certain branch have been deactivated, the root node is made active again. This corresponds to the case where a sequence has come to an end, e.g., a car has driven past the traffic lights, and a new situation has to be considered. knowledgethat is not necessary for the recognition process but useful for explanation is then responsible for providing them. In this way, the performance of the recognition systemis not affected. Conclusion Providing Advice for Decision Support Explanations provided to the user play a essential role in our system(Caflameroet al., 1992), since it is intended to help a decision maker in critical domains. Our user has heavy responsibilities and s/he is held personally responsible of his/her decisions; therefore, explanations must be provided to him/her in his/her own language. Indeed, we can define the overall objectives of the system as: (1) relating the observedbehavior described in terms low-level data to high-level structures that describe the tactics or strategy applied by a rational agent in a given situation; (2) interpreting this behavior in terms of the agent’s motivations; and (3) providing satisfactory explanations to user’s questions. Advice is given as a causal interpretation--explanationmfrom the observer’s point of view, of lowlevel input data in terms of the agent’s plausible plan, strategy, and motivation in a specific situation. A causal account of the agent’s behavior seemsto be the right level of interpretation in our case (Delannoyet al., 1992), given that we must deal with intentional systems (Dennett, 1978; Seek 1990)--agents that have desires and motivations, and act accordingly. In addition, we consider these agents "at the knowledge level". This interpretation--and more generally, the explanations provided by the system--is the result of a trade-off amongthree factors: (1) the constraints imposedby the robust approach to plan recognition, (2) the purpose of the recognition task--decision support in critical domains---and(3) the user’s needs for information. Three main types of needs--and hence of explanations--are distinguished: 1. A reliable source of information. The interpretation provided is aimedat understanding a situation in the world by relating it to somepredefined high-level concept. This explanation is dynamically constructed during the problem solving process. 2. A means to test the user’s own hypotheses. They are considered and, if plausible, addedto the list of possible explanations. 3. Justification of different elementsof the interpretation, to ensure that the system’s reasoning and/or conclusion can be trusted. Onlyexplanations of type 1 and 2 take part in the plan recognition problem-solving process. Explanations of type 3 are considered only upon user’s request. They ate constructed after plan recognition problem solving has taken place, and its generation constitutes a complex problem-solving process itself. They require additional knowledgethat is not present in the recognition system, for the sake of efficiency. A separate modulecontaining the 21 This paper describes a wayto performplan recognition in a dynamic domain from low-level data. Plausible plans reflecting the standard behavior of a rational agent in a given situation arc used to interpret the actual agent’s behavior. Monitoring the difference between both behaviors allows to progressively confnan or reject the different plausible hypothe~s. Weaddress the problematic of plan recognition by tackling several real-world characteristics that_ are not well accommodatedby current techniques, such as low-level, noisy, and lacunary data, a dynamic environment where changes occur not only as a consequence of the agent’s behavior. Frominput consisting of parameter values, which give a partial description of the current situation, the system must produce high-level output. The recognition system we propose is based on a library of prototypical cases or plans structured as frames. It uses a mechanism comparable to spreading activation through which plans activate one another downthe hierarchical structure of the plan library, or across the hierarchy. In other words, we try to walk on the steps of the agent by producing interpretative hypotheses, as specific as data allow, and prudently confirming or disqualifying them. A CLOSprototype has been developed. Its graphical interface allows the users to select input, analysis mode (automatic or stepwise), display a bar chart of current confidencevalues of active plans and a historic curve of the confidence values, and to select several types of explanations. In addition to improvingexplanations, several interesting problemscan be the object of future work. Among others, a better treatment of temporalrelations, and a deeper study of the correspondence between the planning algorithm (domain and strategic knowledge) and the recognition rules. In this sense, a knowledge-level modelof the plan recognition task has been proposed (Caflamero, 1994), which aims at achieving a better understanding of the rationale behind this problem-solvingbehavior, and offers a guide for the acquisition of plan recognition knowledge. Acknowledgments.This research is partially funded by DRETunder contract 91-358. References Allen, J.F., Kautz, H.A., Pelavin, R.N., and Tenenberg, LD. 1991. Reasoning about Plans, Morgan Kaufmann. Bar~s, M. 1989. Aide h la d6cision darts les syst~mes informatis6s de commandement:r61e de l’information symbolique. DRETReport,July 1989, Pads. Caflamero, D. 1994. Modeling Plan Recognition for Decision Support. In Proceedings of the 8th European Workshop on Knowledge Acquisition (EKAW’94), Hoegaarden, Belgium, September 26-29. SpringerVerlag, LNAIseries. Forthcoming. Caflamero, D., Delannoy, L-F., and Kodratoff, Y. 1992. Building Explanations in a Plan Recognition System for Decision Support. Proceedings of the ECAIWorkshop on Improving the Use of Knowledge-Based Systems with Explanations, Vienna, Austria, August 1992; Technical Report LAFORIA 92/21, Universit~ Paris VIL France, pp. 35-45. Charniak, E. and Goldman, R.P. 1991. A Probabilistic Modelof Plan Recognition. Proc. Ninth AAAI, pp. 160165. Charniak, E. and McDermott, D. 1985. Introduction to Artificial Intelligence, Addison-Wesley. Cohen, P.R., Perrault, C.R., and Allen, J.F. 1981. Beyond question answering. In W. Lenhert and M. Ringle (eds.), Strategies for Natural Language Understanding, Lawrence Erlbaum, pp. 245--274. Delannoy, J.-F., Cagamero, D., and Kodratoff, Y. 1992. Causal Interpretation from Events in a Robust Plan Recognition System for Decision Support. Proc. Fifth International Symposium on Knowledge Engineering, pp. 179-189,Sevilla, Spain, October5-9. Dennett, D.C. 1978. Brainstorms: Philosophical Essays on Mind and Psychology, Bradford Books. Kautz, H.A. 1991. A Formal Theory of Plan Recognition and its Implementation. In Reasoningabout Plans, eds. J.F. Allen, H.A. Kautz, and J.D. Tenenberg, pp. 69-125, Morgan Kaufmann. Kodratoff, Y., Rouveirol, C., Tecuci, G., and Dural, B. 1990. Symbolicapproachesto uncertainty. In Intelligent Systems: State of the Art and Future Directions, eds. Z.W. Ras and M. Zermanova, pp. 226-255, Ellis HorwoodLtd. Lange, T.E. and Wharton, C.M. 1993. DynamicMemories: Analysis of an Integrated Comprehensionand Episodic MemoryRetrieval Model. Proc. 13th lntenational Joint Conferenceon Artificial Intelligence (IJCAI), pp. 208213, Chambery, France, August 28-September 3, Morgan Kaufmann. Laskowski, S.J. and Hofman, E.J. 1987. Script-based reasoning for Situation Monitoring. Proc. AAA187, pp. 819-823, Morgan Kaufmann. Mooney, R.J. 1990a. A General Explanation-Based Learning Mechanismand its Application to Narrative Understanding, Morgan Kaufmann. Mooney, R.J. 1990b. Learning Plan Schemata From Observation: Explanation-Based Learning for Plan Recognition. Cognitive Science, 14: 483-509. Newell, A. 1982. The Knowledge Level. Artificial Intelligence, 18: 87-127. Pelavin, R.N. 1991. Planning With Simultaneous Actions and External Events. In Reasoningabout Plans, eds. J.F. Allen, H.A. Kautz, and J.D. Tenenberg, pp. 127-211, Morgan Kaufmann. 22 Quast, K.J. 1993. Plan Recognition for Context-Sensitive Help. Proc. ACMConference on Intelligent User Interfaces, pp. 89-96. Schank, R.C. 1986. Explanation Patterns: Understanding Mechanically and Creatively, Lawrence Erlbaum Associates. Schank, R.C. and Kass, A. 1990. Explanations, Machine learning, and Creativity. In Machine Learning, An Artificial Intelligence Approach, Volume III, eds. Y. Kodratoff, R.S. Michalski, pp. 31-48, Morgan Kaufmann. Schank, R.C. and Riesbeck, C. (eds.) 1991. Inside CaseBased Reasoning, Lawrence Erlbanm Associates. Seel, N. 1990. From Here to Agent Theory. AISBQ72. Spring: 15-25. van Beck, P. and Cohen, R. 1991. Resolving plan ambiguity for cooperative response generation. Proc.Tenth AAAI, pp. 938-944. van de Velde, W. 1993. Issues in Knowledge-Level Modeling. In J.-M. David, J.-P. Krivine, and R. Simmons(eds.), Second Generation Expert Systems. Springer-Verlag, pp. 211--231. Vilain, M. 1990. Getting serious about parsing plans: A grammatical analysis of plan recognition. Proc.Ninth AAA/,pp. 190-197, Boston, MA. Wilensky, R.W. 1981. Meta-Planning: Representing and Using Knowledge about Planning in Problem-Solving and Natural Language Understanding.. Cognitive Science, 5: 197-233. Wilensky, R.W. 1983. Planning and Understanding: A Computational Approach to Human Reasoning, Addison-Wesley.