The objective of the proposed research is to understand the effect of limited computational resources on the design of optimal decision making sytems, so as to be able to build agents that are robust in the face of complexity, and capable of acting in real time. The theoretical phase involves the completion of a normative theory for reasoning about computations. The approach is to treat computations as actions, with outcomes of varying probabilities and utilities using decision theory to choose between them. The theory will be incorporated in a multilevel decision theoretic architecture, which extends the idea of explicit metalevel reasoning to cover recursively all aspects of the decision making process. The approach has been applied successfully to game playing systems, and will be modified to apply to single agent problems such as robot motion planning. However, metalevel reasoning to often expensive, and the cost can only be amortized over future problem solving if a compilation mechanism is available. The investigator will therefore study methods for compilation of reasoning processes occuring anywhere in the hierarchy. He will extend explanation based learning techniques to cover the case of choosing best actions, where probabilistic time accuracy tradeoffs play a significant role.
