Neural networks training using genetic algorithms

Presents a genetic algorithm based system for evolving neural networks. New genetic operators, which combine a heuristic approach and pseudo gradient information, are designed to enhance the performance of genetic algorithms. In this way, the extension or contraction of search region can be more adaptive to the characteristics of the neural network's output error surface. The proposed methods are tested on the n-bit parity problem. By applying these methods, we have been able to find single layer networks in solving 2-, 3-, and 4-bit parity problems. Moreover, we attempt to incorporate GAs into the cascade correlation algorithm in optimizing the network architecture. Because of the complementary properties of exploration capability of genetic algorithms and local search of the derivative-base approach, the hybrid method is expected to outperform either method alone. Experimental results have demonstrated the effectiveness of our methods in terms of the average number of hidden nodes.