9211512 Russell This is a REU supplemental award to IRI-9211512. The objective of the RALPH (Rational Agents with Limited Performance Hardware) project is to understand the effect of limited computational resources on the design of optimal decision-making systems. Work to date has specified a normative theory of metareasoning to control computation, and has applied it to several combinatorial problem classes; in all cases significant performance improvements over standard algorithms have been record. The methodology for applying the theory has been developed in detail, and has been used by other researchers to solve complex problems. The PI will continue to work on more general decision-theoretic metalevel architectures; probabilistic planning and scheduling systems; and composition of real-time systems, for which preliminary results have been obtained. A software environment will be developed that allows complex real-time agents to be constructed automatically from elementary soft-real-time components. A final element of the proposal is the construction of provably optimal bounded agents. Working initially in a class of simple rile-based machines, it will be shown that a learning algorithm can converge quickly to an approximately optimal rule set. These results will be extended to find the scope and limits of adaptive methods for achieving bounded rationality, thereby contributing to the foundation for a science of artificial intelligence.
