Combining Competitive And Cooperative Coevolution For Training Cascade Neural Networks

Cooperative Coevolution (CC) has been shown to be effective in problems where certain architectural details of the solution are evolved. This is the case of cascade neural networks where the number of hidden units is not preestablished but rather emerges through learning. We take a step towards having coadapted subcomponents emerge rather than being hand designed by showing that competing populations (evolved by GAs with different mutation and crossover probabilities) can be successfully used in selecting the species that are subsequently coevolved in a cooperative model. Our experimental results indicate that retraining is an essential step in the cooperative coevolution model. Previous studies used evolutionary algorithms (EAs) to train connection weights and neuron thresholds in artificial neural networks (ANNs). We show that by also evolving the characteristics of the neurons themselves, the quality of the solution (in terms of number of hidden units) could be significantly improved.