Investigating the Diversity of Evolved Cognitive Models in Delayed Match to Sample Experiments

Abstract

The construction of programs to model cognitive behaviour can be semi-automated, using program synthesis techniques such as genetic programming. However, these techniques can produce many candidate solutions of good fitness making it difficult to find the important similarities and differences which can explain model behaviour and lead to theoretical insights. Here we show how the diversity of generated models can be assessed using the Silhouette Index measure of cluster quality; the quantitative measure helps guide theorists to important qualitative differences in the models. Such differences can help inspire novel cognitive theories.