Metacognitive Behavior in Adaptive Agents

This paper presents a novel multidisciplinary architecture which unites recent research in cognitive science and connectionism to develop an integrated solution, and computational model, for the acquisition and performance of recognitional and metacognitive skills by human subjects and intelligent agents. This work builds on three key components: (a) the Recognition / Metacognition model of human decision making; (b) the SHRUTI model of reflexive reasoning in a connectionist system; and (c) Adaptive Critics, a connectionist model of behavior learning. We are attempting to develop a hybrid computational realization of the Recognition/Metacognition model as a basis for the design of adaptive intelligent agents. Introduction: Intelligent Agents (IA) seem to be getting all the press these days, but few of the issues posed by the agents paradigm are new, and fewer still are easy. These are the early days for intelligent agents, and even the term remains ill-defined. The most intriguing and challenging definitions carry the connotations of personality: information-gathering, decision-making, communication, and autonomous action. In this paper, we build on psychological evidence of the decision-making capabilities of humans to develop a computer architecture suitable for further exploring human decision making skills and building artificial, intelligent agents that model those skills. We draw, in the course of this paper, on a series of interviews with Naval tactical decision makers, and present an example of how an intelligent agent would implement behavior such as observed in a human decision maker. Current Models of Decision-Making: Two quite different starting points have characterized recent discussions of how people learn and use decision-making skills. The analytical model (e.g., Raiffa, 1968, Keeney and Raiffa, 1976) specifies that decision makers: learn to generate options, outcomes, and goals; assess the probability of outcomes and the value of goals; aggregate the probabilities and values into a summary score for each option; and choose among the options. The observed effects of options serve as feedback with which the decision maker rationally updates her future assessments. The recognition model (Chase and Simon, 1973; Klein, 1993) supposes that experienced decision makers match perceptual cues and salient information to learned patterns, which in turn trigger specific responses. The success or failure of these responses alters the stored relations between cues, patterns and responses. According to a large corpus of research, such models do not account for human behavior in critical decision-making tasks. Experienced decision makers do not engage in the exhaustive generation of outcomes or comparison of alternatives required by the analytical model. Examination of interviews with Naval officers conducted for the TADMUS project reveals few instances of analytical processes, and a wealth of instances in which officers attempted to weave coherent stories from events that described the intent and future actions of a radar contact. Where the analytical process simply aggregates concordant and conflicting data, we observed that officers treat the two types of data quite differently conflicting