Evolving Walking : The Anatomy of an Evolutionary Search

The evolution of a continuous time recurrent neural network central pattern generation for walking is characterized and found to proceed in two phases. The first phase spans the beginning of the search through the generation at which a " breakthrough " individual is discovered. The second phase proceeds from that generation forward. The first phase is most quickly completed if each succeeding population is as random as possible. Hence GA searches performed at lower mutation variances require more generations to discover a breakthrough individual than higher mutation variances searches. In the second phase the best fitness in the population most rapidly increases at low mutation variances. The role of parameter space structure in these trends will be examined.