Boris Kerkez

Case-based plan recognition with novel input

Our research investigates a case-based approach to plan recognition using incomplete incrementally learned plan libraries. To learn plan libraries, one must be able to process novel input. Retrieval based on similarities among concrete planning situations rather than among planning actions enables recognition despite the occurrence of newly observed planning actions and states. In addition, we explore the benefits of predictions using a measure that we call abstract similarity. Abstract similarity is used when a concrete state maps to no known abstract state. Instead a search is performed for nearby abstract states based on a nearest neighbour technique. Such a retrieval scheme enables accurate prediction in light of extremely novel observed situations. The properties of retrieval in abstract state-spaces are investigated in three standard planning domains. We first determine optimal radii to use that determines a spherical sub-hyperspace that limits the search. Experimental results then show that significant improvements in the recognition process are obtained using abstract similarity.

Incremental Case-Based Plan Recognition Using State Indices

We describe a case-based approach to the keyhole plan-recognition task where the observed agent is a state-space planner whose world states can be monitored. Case-based approach provides means for automatically constructing the plan library from observations, minimizing the number of extraneous plans in the library. We show that the knowledge about the states of the observed agents world can be effectively used to recognize agents plans and goals, given no direct knowledge about the planners internal decision cycle. Cases (plans) containing state knowledge enable the recognizer to cope with novel situations for which no plans exist in the plan library, and to further assist in effective discrimination among competing plan hypothesis.

INCREMENTAL CASE-BASED PLAN RECOGNITION WITH LOCAL PREDICTIONS

We present a novel case-based plan recognition method that interprets observations of plan behavior using an incrementally constructed case library of past observations. The technique is novel in several ways. It combines plan recognition with case-based reasoning and leverages the strengths of both. The representation of a plan is a sequence of action-state pairs rather than only the actions. The technique compensates for the additional complexity with a unique abstraction scheme augmented by pseudo-isomorphic similarity relations to represent indices into the case base. Past cases are used to predict subsequent actions by adapting old actions and their arguments. Moreover, the technique makes predictions despite observations of unknown actions. This paper evaluates the algorithms and their implementation both analytically and empirically. The evaluation criteria include prediction accuracy at both an abstract and a concrete level and across multiple domains with and without case-adaptation. In each domain the system starts with an empty case base that grows to include thousands of past observations. Results demonstrate that this new method is accurate, robust, scalable, and general across domains.

Local Predictions for Case-Based Plan Recognition

This paper presents a novel case-based plan recognition system that interprets observations of plan behavior using a case library of past observations. The system is novel in that it represents a plan as a sequence of action-state pairs rather than a sequence of actions preceded by some initial state and followed by some final goal state. The system utilizes a unique abstraction scheme to represent indices into the case base. The paper examines and evaluates three different methods for prediction. The first method is prediction without adaptation; the second is predication with adaptation, and the third is prediction with heuristics. We show that the first method is better than a baseline random prediction, that the second method is an improvement over the first, and that the second and the third methods combined are the best overall strategy.