Toward Boundedly Rational Analysis

Thomas IcardCarnegie Mellon

Abstract

The Bayesian program in cognitive science has been subject to criticism, due in part to puzzles about the role of rationality and approximation. While somewhat sympathetic with these concerns, I propose that a thoroughgoing boundedly rational analysis strategy can answer to some of them. Through simulation results I illustrate the method by showing how one can retrodict recently reported results about particle filter models of categorization (Sanborn et al., 2010). I also introduce new obstacles that surface once we take bounded rationality seriously. Specifically, again through simulation, I show that the analysis of optimal sampling from Vul et al. (2014) is interestingly complicated by the introduction of agents capable of metareasoning. Under broad conditions, such agents outperform all uniform k-sampling agents. This motivates the computational study of boundedly rational metareasoning in its own right.

Files

Toward Boundedly Rational Analysis (127 KB)



Back to Table of Contents