Search Space Restriction of Neuro-evolution through Constrained Modularization of Neural Networks

Evolving recurrent neural networks for behavior control of robots equipped with larger sets of sensors and actuators is difficult due to the large search spaces that come with the larger number of input and output neurons. We propose constrained modularization as a novel technique to reduce the search space for such evolutions. Appropriate neural networks are divided manually into logically and spatially related neuro-modules based on domain knowledge of the targeted problem. Then constraint functions are applied to these neuro-modules to force the compliance of user defined restrictions and relations. For neuro-modules this will facilitate complex symmetries and other spatial relations, local processing of related sensors and actuators, the reuse of functional neuro-modules, fine control of synaptic connections, and a non-destructive crossover operator. With an implementation of this so called ICONE method several behaviors for nontrivial robots have already been evolved successfully.