Developing memory-based models of ACT-R within a statistical framework

Abstract

Listen

Translate article

The ACT-R cognitive architecture is a computational framework for developing, simulating and testing comprehensive theories of cognition. By far, the most common method of evaluating ACT-R models involves generating predictions through Monte Carlo simulation and comparing those predictions to aggregated human data. This approach has several limitations, including computational inefficiency, the potential for averaging artifacts, and difficulty representing uncertainty in parameter estimates. In this paper, we demonstrate the fundamentals of developing models of ACT-R within a Bayesian framework. Instantiating ACT-R in a Bayesian framework has many advantages, including the ability to use modern parameter estimation and model comparison techniques, the ability to compare ACT-R to other closed-form models, increased computational efficiency, and the ability to perform a deeper mathematical analysis of model properties. We develop model variants of the classic fan experiment, beginning with a simple baseline model of ACT-R’s declarative memory system and progressing through increasingly complex variants until reaching a moderately complex general model. Our hope is that this will highlight connections between computational and mathematical approaches to formal modeling and facilitate new and exciting research.