Beyond bounded rationality : towards a computationally rational theory of motor control

Abstract

Limitations in human information processing have been well documented, though remain less understood in terms of how cognitive systems adapt to information processing constraints. To address this challenge, a growing body of literature offers theories concerning intelligent systems (both biological and artificial) from the perspective of computational rationality (Gershman et al.,2015)—the maximization of performance subject to constraints on information processing. Thisdissertation aims to add to these theoretical and computational approaches by offering a noveland principled way of studying this problem. The central approach advocated in this thesis, uses the formal mathematics of rate-distortion theory, a branch of information theory (Shannon, 1948) which concerns the optimal solution to the problem of optimal, but ‘lossy’ data compression. As a mathematical framework, rate-distortion theory is ideally situated to predict and explain cognitive processing as the (rational) maximization of utility subject to information processing constraints. First demonstrating the applicability of this approach in the domain of visual working memory in human systems, this dissertation extends this form of rational analysis to the reinforcement learning framework and domain of motor control. This thesis demonstrates, through the combination of three empirical studies, and computational cognitive modeling, that rate-distortion theory provides a natural framework for the development of computationally rational learning agents and a promising normative framework to analyze human motor performance.