Evolving Understandable Cognitive Models

Abstract

Cognitive models for explaining and predicting human performance in experimental settings are often challenging to develop and verify. We describe a process to automatically generate the programs for cognitive models from a user-supplied specification, using genetic programming (GP). We first construct a suitable fitness function, taking into account observed error and reaction times. Then we introduce post-processing techniques to transform the large number of candidate models produced by GP into a smaller set of models, whose diversity can be depicted graphically and can be individually studied through pseudo-code. These techniques are demonstrated on a typical neuro-scientific task, the Delayed Match to Sample Task, with the final set of symbolic models separated into two types, each employing a different attentional strategy.