Rational Metareasoning and Compilation for Optimizing Decisions Under Bounded Resources

Eric Horvitz

Access pdf file.


Several years ago, our research group initiated the Protos project to investigate the use of decision theory as a framework for reasoning about the design and operation of ideal agents under bounded resources. We have studied metareasoning, reasoning, and compilation within the framework of decision theory. Our model of rationality centers on the use of design-time and tractable run-time decision-theoretic analyses to control the detail and completeness of problem-level decision making. Unlike straightforward decision analyses, we apply the principles of decision theory to enriched models that include not only distinctions and outcomes in the world, but also distinctions and outcomes about cognition. In this paper, we shall review some earlier work on rational metareasoning and describe the benefits of integrating deliberative models of decision-theoretic reasoning and metareasoning with several classes of precomputed or compiled actions.

Keywords: Bounded optimality, rationality, decision-theoretic control, metareasoning, expected value of metareasoning, compilation of action.

E. Horvitz. Rational Metareasoning and Compilation for Optimizing Decisions under Bounded Resources Reference: Horvitz, E. J. Rational Metareasoning and Compilation for Optimizing Decisions under Bounded Resources. Proceedings of Computational Intelligence '89. Association of Computing Machinery, Milan, Italy, September, 1989.

Also, available as Stanford University Technical Report KSL-89-81, Departments of Computer Science and Medicine, 1987.

Author Email: horvitz@microsoft.com

Back to Eric Horvitz's home page.