A New Perspective on Algorithms for Optimizing Policies under Uncertainty

The paper takes a fresh look at algorithms for maximizing expected utility over a set of policies, that is, a set of possible ways of reacting to observations about an uncertain state of the world. Using the bucketelimination framework, we characterize the complexity of this optimization task by graph-based parameters, and devise an improved variant of existing algorithms. The improvement is shown to yield a dramatic gain in complexity when the probabilistic subgraph of the inuence diagram is sparse, regardless of the complexity introduced by its utility subgraph.

• Type: Conference
• Year: 2000
• Where: AIPS
• Authors: Rina Dechter
• Keywords: AIPS 2000, Artificial Intelligence, Expected Utility, Fresh Look, Graph-based Parameters