Evolution of Development and Heterochrony in Artificial Neural Networks
Artur Matos, Reiji Suzuki, Takaya Arita · 2005
Recently, evolutionary algorithms coupled with simulated developmental processes have been used successfully for generating designs ranging from neural networks to artificial creatures. Although several of these models do exist, their evolutionary dynamics, and more specifically how the ontogenic models themselves interact with artificial evolution are still poorly understood. One of these specific interactions, and of particular importance in biological systems is heterochrony — the change in timing and rate of developmental events by evolution. In this paper, we analyze heterochronic change in one artificial developmental model- the cellular encoding model first described by Gruau [1]. For this purpose, we apply the framework and methods defined by Alberch et al [2] for biological systems to neural networks evolved for the odd-3-parity problem. Preliminary results show that: 1) All heterochronic changes occur with significant frequency; 2) The combined effects of predisplacement, hypermorphosis, and neoteny was the most common heterochronic change; 3) Pure recapitulation (isomorphosis) is prevalent.