On partially observable MDPs and BDI models

Abstract. Decision theoretic planning in ai bymeans of solving Partially Observ-ableMarkov 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 pro-pose a novel method for solving pomdps, which provides a designer with a more intuitive means of specifying pomdp planning problems. In particular, we investi-gate 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. Addition-ally, we illustrate the significance of a correspondence by showing empirically that it yields satisfying results in complex domains.