Exploiting diversity of neural ensembles with speciated evolution

We evolve artificial neural networks (ANNs) with speciation and combine them with several methods. In general, an evolving system produces one optimal solution for a given problem. However we argue that many other solutions exist in the final population, which can improve the overall performance. We propose a method of evolving multiple speciated neural networks by fitness sharing that helps to optimize multi-objective functions with genetic algorithms, and several combination methods to construct ensembles of ANNs. Experiments with the UCI benchmark datasets show that the proposed methods can produce more speciated ANNs and, thus, improve the performance by combining representative individuals with combination methods