AI Seminar

Explorations in Computational Rationality: Explaining Human Behavior via Bounded Utility Maximization

Richard LewisProfessor of Psychology and LinguisticsUniversity of Michigan

A useful approach in multiple areas of cognitive science has been to observe human behavior on tasks thought to reveal limits of the underlying system. Detailed behavioral patterns that are departures from what would be expected from an unbounded machine or a normative analysis are then used to motivate theories of the mechanisms that shape behavior. Some of the most celebrated phenomena and theories in the field take this approach, but there is a danger of mistaking strategic adaptations for fixed, domain-specific constraints. What is needed is an analytic framework for deriving the implications for adaptive behavior of both posited computational constraints and task/goal structure. We present a candidate framework that is an application of bounded optimality to psychology, and formulates problems of adaptation as optimal control problems. We provide a summary overview of analyses of three classic empirical domains that have previously been used to motivate a variety of specific mechanisms or bounds: contextual choice and preference reversals (thought to reveal suboptimal/irrational mechanisms of choice); dual-task costs in overlapping stimulus-response tasks (thought to reveal a central decision bottleneck); and distractor-ratio/pop-out effects in visual search (thought to reveal limited mechanisms for feature-binding and attention). In each case, details of behavior can be explained by assuming a rational adaptation to task demands and general constraints concerning noisy representations and process durations. The analyses yield interesting new predictions (some of which we have tested experimentally), and call into question the need to posit specific mechanisms or resource limits to explain the phenomena. We explore the relation of this framework to related approaches for taking into account bounds in rational analysis
Richard Lewis is a cognitive scientist at the University of Michigan, where he is Professor of Psychology and Linguistics. His research interests include sentence processing, eye-movements, short-term memory, cognitive architecture, reinforcement learning, and bounded optimization approaches to modeling human behavior. He received his PhD in computer Science at Carnegie Mellon with Allen Newell, followed by a McDonnell Fellowship in Psychology at Princeton and a position as Assistant Professor of Computer Science at Ohio State. He was elected a Fellow of the Association for Psychological Science in 2010.

Sponsored by