A Cellular Genetic Algorithm for training Recurrent Neural Networks

Recurrent neural networks (RNNs), with the capability of dealing with spatio-temporal relationship, are more complex than feed-forward neural networks. Training of RNNs by gradient descent methods becomes more dii-cult. Therefore, another training method, which uses cellular genetic algorithms, is proposed. In this paper, the performance of training by a gradient descent method is compared with that by a cellular genetic algorithm. Experimental results indicate that the cellular genetic algorithm successfully nds the global optimum, whereas the gradient descent method is trapped in the sub-optima frequently. Furthermore, it is found that the solutions obtained from the cellular genetic algorithm have better generalization performance than that of the gradient descent method.