Evolving Neural Networks with
Collaborative Species
Mitchell A. Potter
Kenneth A. De Jong
We present a coevolutionary architecture for solving decomposable problems and apply it to the evolution of artificial neural networks. Although this work is preliminary in nature it has a number of advantages over non-coevolutionary approaches. The coevolutionary approach utilizes a divide-and-conquer technique in which species representing simpler subtasks are evolved in separate instances of a genetic algorithm executing in parallel. Collaborations among the species are formed representing complete solutions. Species are created dynamically as needed. Results are presented in which the coevolutionary architecture produces higher quality solutions in fewer evolutionary trials when compared with an alternative non-coevolutionary approach on the problem of evolving cascade networks for parity computation.