Optimization of Neural Networks Weights and Architecture: A Multimodal Methodology

This paper describes a multimodal methodology for evolutionary optimization of neural networks. In this approach, we use Differential Evolution with parallel subpopulations to simultaneously train a neural network and find an efficient architecture. The results in three classification problems have shown that the neural network resulting from this method has low complexity and high capability of generalization when compared with other methods found in literature. Furthermore, two regularization techniques, weight decay and weight elimination, are investigated and results are presented.