An evolvability-enhanced artificial embryogeny for generating network structures

Existing Artificial Embryogeny (AE) models are insufficient to generate a network structure because the possible links are limited to those connecting nodes with their predefined neighbors. We propose a novel network generating AE model capable of generating links connected to predefined neighbors as well those to non-neighbors. This mechanism provides additional flexibility in phenotypes than existing AE models. Our AE model also incorporates a heterogeneous mutation mechanism to accelerate the convergence to a high fitness value or enhance the evolvability. We conduct experiments to generate a typical 2D grid pattern as well as a robot with a network structure consisting of masses, springs and muscles. In both tasks, results show that our AE model has higher evolvability, sufficient to search a larger space than that of conventional AE models bounded by local neighborhood relationships.