Abstraction for Decision-Theoretic Planning

ion for Decision-Theoretic Planning AnHai Doan and Peter Haddawy Department of EE & CS University of Wisconsin-Milwaukee PO Box 784 Milwaukee, WI 53201 fanhai, haddawyg@cs.uwm.edu 414 229-4955 Abstract We present a general theory of action abstraction for reducing the complexity of decision-theoretic planning. We discuss how to abstract sets of actions, action branches, and sequences of actions. We develop projection rules for abstract actions and prove our abstraction techniques correct with respect to projection. In developing the theory we introduce the notion of generalized mass assignments. Submitted to UAI95. This work was partially supported by NSF grant IRI-9207262. 1 Introduction It has become widely accepted in the AI planning community that planning schemes which strive to solve realistic problems must be able to model uncertainty as well as tradeoffs among goals and the resources consumed in achieving goals. This position has given rise to the decision-theoretic..