Machine Learning for Plan Recognition

action subsuming make spaghetti and make fettucini (make pesto and make marinara). Edge labels like h=; 1; 2i represent between two actions a and b represent the fact that the rst argument of a is identical to the second argument of b. Computing the join of G1 and G2 then consists of nding action nodes sharing a common abstraction and identifying temporal and strcutural relations common to both action graphs (the details of this algorithm can be found e.g. in [Bau98a]). As a result the two alternative generalizations|so-called valid joins|of G1 and G2 depicted on the right-hand side of Figure 1 are obtained. Each of them represents an abstract, partially ordered plan. The choice of which one to actually include in the plan library depends on the intended use of the plan recognition result. The basic \join" procedure sketched above is complemented by a clustering algorithm that is used to identify sets of \similar" action sequences in order to generate alternative plan decompositions in cases where a number of strongly diverging ways exist to achieve a particular goal (this is actually the normal case). Additionally this algorithm can be used to group unlabeled action sequences (i.e. not marked with their associated domain goal), compute an abstract plan representing their essential aspects, and have them labeled by a domain expert to produce a useful plan library (for details see [Bau99]). Empirical tests showed that the plan libraries produced using the above approach allow very high recognition rates, even in cases when no additional domain knowledge is used. This is probably due to the fact that these libraries re ect actual users' behaviors, not idealized concepts of a knowledge engineer. 3 User Models Most plan recognition systems only maintain a list of plan hypotheses all of which proved compatible with the observations made so far and are considered equally plausible. In many cases, however, it is not su cient to simply state that the observed agent is pursuing one of the plans contained in this list. Whenever reactivity is expected, e.g. in intelligent help systems, a decision has to be made as to which of these hypotheses is the \best" one. The required quality criterion can be based on a numerical measure that re ects the \suitability" of a hypothesis to describe the current situation, i.e. the plan currently being pursued. This property directly depends on the likelihood that a hypothesis correctly represents the observed agent's plan which obviously depends on this particular agent's typical behavior. Intelligent help systems, for example, have to take into account both the expertise and the personal preferences of the user of an application system because both factors crucially in uence the way she tries to reach her goals. Not knowing about the user's preferences hinders both the plan recognition process|especially if this user's behavior deviates from what is typically to be expected in a domain|and the generation of an adequate system response like suggesting an action sequence to accomplish a certain goal. Knowledge about a user's peculiarities helps focusing on plans likely to this user and giving support tailored to her particular needs. In [Bau96] an approach was presented that combines inductive machine 3 > 50Store_messages< 50ManagerCo-worker(#=5) [Manager, *]:(#=2) [Co-Worker, 50+]:(#=8) [Co-Worker, 50+]:Read_delete_msgs(#=48) [Co-Worker, 1-49]:Read_delete_msgsStore_messageslengthsender Figure 2: The decision tree for action read mail.learning with the numerical formalism of the Dempster-Shafer Theory (DST).During a training phase the user's interaction with a system is logged. Thelog then consists a set of actions, a representation of the context in which eachaction took place, and nally the domain goal this particular action sequenceachieved. The \context" of an action is represented as a vector describing im-portant aspects of the current state of the world in terms of a set of prede nedfeatures.At the end of the training phase those entries of the so constructed database containing the same action are grouped and forwarded to an inductivemachine learning algorithm (ID3 and C4.5 were used in the experiments). Sta-tistical information representing the relative frequencies of certain actions withina particular type of situation are used to label the leaves of the decision treesso produced.Whenever a new action is observed, the current situation is encoded in termsof the same features as used during the training phase and classi ed using thedecision tree associated with this action. The classi cation result is transformedinto a so-called basic probability assignment from DST which is used to computea numerical assessment of the set of plan hypotheses and eventually select the\best" one or even decide that more observations are required to arrive at a safedecision.Example: The decision tree depicted in Figure 2 represents the fact that|in some email application|the observed user always deleted a message afterreading it when the send was her manager, but always stored it when it wasfrom a co-worker and shorter than 50 lines. Longer message from co-workerswere sometimes stored (twice) and sometimes deleted (8 times).References[Bau96] M. Bauer. Acquisition of User Preferences for Plan Recognition. In D. Chin,editor, UM96, pages 105{112, Kailua-Kona, HI, USA, 1996.[Bau98a] M. Bauer. Acquisition of Abstract Plan Descriptions for Plan Recognition.In AAAI'98, pages 936{941, 1998.[Bau99] M. Bauer. From Interaction Data to Plan Libraries: A Clustering Approach.In IJCAI '99, Stockholm, Sweden, 1999. Morgan Kaufmann. to appear.[LE96] N. Lesh and O. Etzioni. Scaling up goal recognition. In J. Doyle, L.C. Aiello,and S.C. Shapiro, editors, KR '96, pages 178{189, Cambridge, MA, 1996.[Moo88] R.J. Mooney. A General Explanation-Based Learning Mechanism and itsApplication to Narrative Understanding. PhD thesis, University of Illinois,Urbana, 1988.4