Abstract

The question of conceptual representation has received considerable attention in philosophy, neuroscience and embodied evolved agents. Numerous theories on the interpretation of the term `representation' exist, and many arguments have been made for and against the existence of representations in animate and animat agents. Our work studies this question in evolved artificial embodied agents in a quantitatively rigorous manner, for the first time. We develop two measures, based on information theory, to account for representations. These measures are studied by applying them to evolved agents performing a visual categorization, generalized XOR task. Our results show that having quantitative measures still leaves one with arbitrary "threshold values" decisions which permit wide freedom in determining the existence of representations. However, and more importantly, our results show that information-theoretic measures can still be used e#ciently to identify discriminative neural patterns and internal structures that characterize a representation, if the latter is formed.

Citations