Bayes Network Abstraction for Decision-Theoretic Planning

This is a preliminary report on research towards the development of an abstraction-based framework for decision-theoretic planning using Bayesian networks. We discuss two problems: the representation of sets of probability distributions for Bayesian networks, and the issue of representing and abstracting actions using Bayesian networks. For the first problem, we propose the use of cc-trees to represent sets of probability distributions and show how propagation in Bayesian networks can be performed without loss of information in this representation. The cc-tree representation provides an intuitive and flexible way to make tradeoffs between precision and computational cost. For the second problem, we identify a class of planning problems where a simple abstraction technique can be used to abstract a set of actions and to reduce the cost of plan evaluation.