IRI-9509165 Peter Haddawy University of Wisconsin - Milwaukee $78,038 - 12 mos. Practical Decision-Theoretic Planning The desire to solve practical real-world planning problems places conflicting demands on domain-independent planning systems. The representation used by the planner must be rich enough to capture all the salient aspects of the application, but at the same time, it must permit the use of efficient algorithms that will scale up effectively to handle large problems. In planning under uncertainty, decision theory allows the consideration of tradeoffs among partially satisfiable objectives. But most of the planning algorithms have either made unrealistic assumptions or been too inefficient for application to large problems. The purpose of this research is to develop efficient methods by using abstraction techniques which permit a planner to focus on important aspects of a domain before considering details. Methods of automatically generating abstractions are being developed, along with methods of applying these to planning problems in dynamic environments. Decision-theoretic planners are being extended ion their representational capabilities with the goal of not sacrificing efficiency. Techniques for efficient reasoning with probabilistic domain theories are being developed. The tangible products are expected to be a theoretical framework for efficient decision-theoretic planning, ore or more implemented planning systems, a theoretical framework for probabilistic domain modeling using probability logic, and an implemented inference system for constructing Bayesian networks from knowledge bases of probability logic sentences.
