Artificial embryogeny for network structures and its application for a robot generation task

Abstract

Artificial embryogeny (AE) is a strategy of evolutionary computation inspired by the developmental process of natural organisms. Yet while there are a few successful examples of generating network structures, existing AE models are insufficient to generate a network structure. The issue is that the possible links are limited to those connecting nodes with their predefined neighbors. Our novel AE model is capable of generating links connected to predefined neighbors as well as those to non-neighbors. In order to accelerate the convergence to a high fitness value, our AE model incorporates a heterogeneous mutation mechanism. We conduct experiments to generate not only a typical 2D grid pattern but robots with network structures consisting of masses, springs and muscles. The robots are evolved in various environments. The results show that our AE model has better convergence property, sufficient to search a larger space, than conventional AE models bounded by local neighborhood relationships.