Knowledge-Based Control of Decision Theoretic Planning - Adaptive Planning Model Selection

This paper proposes a new planning architecture for agents operating in uncertain and dynamic environments. Decisiontheoretic planning has been recognized as a useful tool for reasoning under uncertainty; it calculates an optimal plan using a given planning model (state set, action set, probability distributions over possible state transitions, and utility function). In a dynamic environment, however, the current situation may be di erent from what an agent expects and the current planning model may not be feasible. It is, therefore, important for an agent to continuously examine the situation and to use an appropriate planning model. For this purpose, we propose to employ a knowledge-based meta-level reasoner to on-line select an appropriate planning model for an object-level decision-theoretic planning, based on the given knowledge of classi cation of the situation. This architecture could also be e ective in reducing the computational cost of decision-theoretic planning by limiting the search space according to the situation. Two applications of the architecture to a dynamic robot planning and to a decision-making in highway driving show the generality and the usefulness of the architecture.