Reference: Horvitz, E. J. Some Fundamental Problems and Opportunities From the Standpoint of Rational Agency. KSL, October, 1990.
Abstract: Growing enthusiasm about the application of autonomous reasoning in high-stakes domains like medicine and aerospace has stimulated interest in systems that behave in accordance with a coherent theory of rationality. Within such domains, the losses associated with suboptimal decisions tend to render simple satisfying approaches inadequate and to provide incentive for attempting to optimize the utility of computational activity. In pursuing research on rational agency over the last several years, a number of problems have come to be highlighted as rich areas for future research. In this paper, we review promising prospects for future study. The research topics will be motivated by recent research and preliminary theoretical and empirical results. At the heart of intelligent behavior is the pursuit of maximal utility by reasoners with relatively limited representational and inferential abilities. Constraints on an agent's reasoning and representation resources lead to inescapable uncertainties about the problems that may be faced and about the performance of alternative reasoning strategies in solving the problems. Almost any interesting work on intelligent problem solving centers on the development of techniques that simple artificats immersed in complex environments can depend on for successful competition with environment challenges and with threats by other intelligent competitors for precious commodities. Uncertainty runs deep at several levels of intelligent foundations in the complexity of the problems. However, an intelligent reasoner must turn inward and reason about the deep uncertainties about the structure of utility, about the representations that it must construct, and about the nature of solving inference problems that are posed by the models. As background, one focus of activity in our research lab has been the exploration of decision-theoretic reasoning strategies in medicine. In fact, our initial foray into research on rationality under resource constraints was motivated by problems with the effective application of decision theory in expert systems and in intelligent monitors for medicine. Current research forcuses on the control of decision-theoretic inference itself with decision-theory. However, our research has also dwelled on the decision-theoretic control of other types of problems, and on foundational problems with computational approaches to rationality.