On partially observable MDPs and BDI models

Abstract

Decision theoretic planning in ai by means of solving Partially Observable Markov decision processes (pomdps) has been shown to be both powerful and versatile. However, such approaches are computationally hard and, from a design stance, are not necessarily intuitive for conceptualising many problems. We propose a novel method for solving pomdps, which provides a designer with a more intuitive means of specifying pomdp planning problems. In particular, we investigate the relationship between pomdp planning theory and belief-desire-intention (bdi) agent theory. The idea is to view a bdi agent as a specification of an pomdp problem. This view is to be supported by a correspondence between an pomdp problem and a bdi agent. In this paper, we outline such a correspondence between pomdp and bdi by explaining how to specify one in terms of the other. Additionally, we illustrate the significance of a correspondence by showing empirically that it yields satisfying results in complex domains.