Growth of Structured Artificial Neural Networks by Virtual Embryogenesis

In the work at hand, a bio-inspired approach to robot controller evolution is described. By using concepts found in biological embryogenesis we developed a system of virtual embryogenesis, that can be used to shape artificial neural networks. The described virtual embryogenesis has the ability to structure a network, regarding the number of nodes, the degree of connectivity between the nodes and the amount and structure of sub-networks. Furthermore, it allows the development of inhomogeneous neural networks by cellular differentiation processes by the evolution predispositions of cells to different learning-paradigms or functionalities. The main goal of the described method is the evolution of a logical structure (e.g., artificial neural networks), that is able to control an artificial agent (e.g., robot). The method of developing, extracting and consolidation of an neural network from a virtual embryo is described. The work at hand demonstrates the ability of the described system to produce functional neural patterns, even after mutations have taken place in the genome.