What role does metareasoning play in models of bounded rationality? We examine the various existing computational approaches to bounded rationality and divide them into three classes. Only one of these classes significantly relies on a metareasoning component. We explore the characteristics of this class of models and argue that it offers desirable properties. In fact, many of the effective approaches to bounded rationality that have been developed since the early 1980's match this particular paradigm. We conclude with some open research problems and challenges.
Computational models of bounded rationalityIn the pursuit of building decision-making machines, artificial intelligence researchers often turn to theories of "rationality" in decision theory and economics. Rationality is a desired property of intelligent agents since it provides well-defined normative evaluation criteria and since it establishes formal frameworks to analyze agents (Doyle 1990;Russell and Wefald 1991). But in general, rationality requires making optimal choices with respect to one's desires and goals. As early as 1947, Herbert Simon observed that optimal decision making is impractical in complex domains since it requires one to perform intractable computations within a limited amount of time (Simon 1947;1982). Moreover, the vast computational resources required to select optimal actions often reduce the utility of the result. Simon suggested that some criterion must be used to determine that an adequate, or satisfactory, decision has been found. He used the Scottish word "satisficing," which means satisfying, to denote decision making that searches until an alternative is found that is satisfactory by the agent's aspiration level criterion.Simon's notion of satisficing has inspired much work within the social sciences and within artificial intelligence in the areas of problem solving, planning and search. In the social sciences, much of the work has focused on developing descriptive theories of human decision making (Gigerenzer 2000). These theories attempt to explain how people make decisions in the real-world, coping with complex situations, uncertainty, and limited amount of time. The answer is often based on a variety of heuristic methods that are used by people to operate effectively in these situations. Work within the AI community-which is the focus of this paper-has produced a variety of computational models that can take into account the computational cost of decision making (Dean and Boddy 1988;Horvitz 1987;Russell et al. 1993; Wellman 1990; Zilberstein 1993). The idea that the cost of decision making must be taken into account was introduced by Simon and later by the statistician Irving Good who used the term Type II Rationality to describe it (Good 1971). Good said that "when the expected time and effort taken to think and do calculations is allowed for in the costs, then one is using the principle of rationality of type II." But neither Simon nor Good presented any effective computational framework to implement "satis...