Modeling Probabilistic Actions for Practical Decision-Theoretic Planning

Proceedings Of The Third Artificial Intelligence Planning Systems Conference

Volume

Proceedings Of The Third Artificial Intelligence Planning Systems Conference

Contents

Most existing decision-theoretic planners represent uncertainty about the state of the world with a precisely specified probability distribution over world states. This representation is not expressive enough to model many interesting classes of practical planning problems, and renders inapplicable some abstraction based planning approaches. In this paper we propose as a remedy a more general world and action model with a well-founded semantics based on probability intervals. We introduce the concept of interval mass assigment . Unlike mass assignments, which assign a probability mass to each set of states, interval mass assignments assign a probability interval to each set of states and are more expressive. Interval mass as signments are interpreted as representing sets of probability distributions and are used in our framework to represent the uncertainty about the state of the world. Within this representation, we present a projection rule and a method for computing a plan’s expected utility. We compare our approach with existing probability-interval approaches. We provide complexity results and empirical evidence which suggest evaluating plans (projecting plans and computing the expected utility) in our framework is efficient, and the action model is applicable in real-world planning domains.