Heterochrony and evolvability in neural network development

Abstract

Recent studies in evolutionary computation have focused on using developmental processes together with genetic algorithms in order to achieve more complex designs. Although several models have been proposed, their growth dynamics, and their interactions with evolutionary algorithms are still poorly understood. One particularly neglected concept in artificial developmental systems is heterochrony: how evolution affects development by changing the timing and rate of developmental events. In this article we attempt to address this issue by analyzing heterochronic changes in a well-known artificial developmental model, the cellular encoding model, by using a previously developed heterochrony framework. We have conducted experiments by evolving networks to solve a Boolean problem, and analyzed heterochronic changes in both successful and unsuccessful runs. Our findings show that owing to its properties, the cellular encoding model strongly affects the developmental dynamics and the heterochronic changes that occur during evolution. Our experiments also show that hypermorphic changes (a kind of heterochronic occurrence) lead to greater evolvability in successful runs.