Deductive Approaches to Plan Generation and Plan Recognition

Terminological Plan Recognition Diane Litman, AT&T Bell Laboratories, USA Description logics are widely used in AI to construct concept taxonomies based on subsumption inferences. However, current description logics are unable to handle complex compositions of concepts, for example constraint networks where each node is described by an associated concept. Plans can be thought of as constraint networks, collections of actions and states related by a rich variety of temporal and other constraints. We have developed the T-REX system, which integrates description logics with constraint network reasoning, to classify plans into an abstraction taxonomy. T-REX also introduces a new terminological view of plan recognition, which dynamically partitions the plan library by modalities (necessary, optional + impossible) while actions are observed. Plan recognition isperformed by computing subsumption + compatibily relations from the taxonomy. Plan Recognition in a Modal Temporal Logic Gabriele Paul, DFKI Saarbrücken, Germany It is natural to regard plan recognition as an abductive task: Given some description T of the world and a set of observations (e.g., actions performed by an agent), try to find a plan P that»--when added to Tallows to explain the observations. In a logical framework this explanation relation is based on the notion of entailment. The classical approach to logic-based abduction with predicate logic is to generate one ore more suitable elements contained in a prede ned set of so-called abducibles and to perform appropriate instantiations in order to obtain ground formulas as explanations. As, however, the (plan) hypotheses which are available to the plan recognizer in this approach are formulated in the very expressive modal temporal logic LLP (Logic Language for Planing), things become more complicated. Here, intuitively valid hypotheses do not satisfy the correctness criterion of classical abduction. This problem is caused by the fact that the hypotheses themselves contain a certain temporal extension. So, a weaker notion of explanations is introduced and characterized semantically. The basic idea of this new form of so called temporal abduction is to re ne the abstract hypotheses by stepwise incorporating the observations into their temporal and logical structure. This implies that the hypotheses can only be ground up to the current point in time. The validity of this approach is demonstrated by a prototypical implementation in the framework of an intelligent help system. Planning, Plan Recognition and Situation Semantics Wayne Wobcke, University of Sydney, Australia The frame, rami cat-ion and quali cation problems are three well known epistemological problems. facing any agent reasoning about action. _.Many approaches to addressing these problems suppose certain semantic principles, e.,g. the frame problem is addressed by aprinciple of minimal change, stating that as little asnecessary changes in the world as the result of performing an action. However, an agent does not reason directly with such a principle: agents are supposed to use an epistemic principle of minimal chage, stating thatas little as necessary changes in an agent s description of the world as the result of performing an action. .We claim that much current researchdoes not bridge the gap between the agent s epistemic principles and the theorist s intended semantics principles, and that part of the problem is the conception of agents as functions over complete world states. We propose an alternative conception of actions as primitive semantic objects occurring in the situations of situation semantics. We also argue that constraints as captured in a hierarchy of types of situations can form the basis of an agent s reasoning about action, and that the necessary constraints can be represented using a standard hierarchy of planning Schemas. We present a conditional logic of constraints and show how both planning and plan recognition can be characterized as inference in the logic. We claim that our approach to formalizing action models the practice of existing planning systems more closely than alternative approaches. Probabilistic Methods in Plan Recognition Mathias Bauer, DFKI Saarbrücken, Germany Plan recognition systems usually can only infer a disjunction of possible plans each of which is equally plausible. If, however, the system is forced to come up with a decision for 913 alternative-i e.g.,s to produce a cer tai'n type of cooperativebehaviour like supporting the user of a complex systemthere must be a criterion to judge the "quality of these hypotheses. Certainly, a -formalism» like probability theory might serve as the basis to define such a selection criterion. In this talk, however, it is argued that Dempster-Shafer Theory has many advantages over classical probability theory, the most important being the fact that also ignorance about the agent s-preferences can be represented explicitly and taken into account during the computations. On this basis a rule-based; approach to plan recognition is proposed which can utilize various forms of statistical information describing the agent s typical behaviour. The resulting numerical values are shown to possess a proper semantics in terms of probability theory.-and thus form a sound foundation to define a variety of criteria with which hypotheses can be assessed. The availability of such criteria enables the system to produce a kind of anytime behaviour in the sense that the best rhyipothesiscan be selected whenever this is required.