Coupling weight elimination with genetic algorithms to reduce network size and preserve generalization

Recent theoretical results support that decreasing the number of free parameters in a neural network (i.e., weights) can improve generalization. These results have triggered the development of many approaches which try to determine an ''appropriate'' network size for a given problem. The main goal has been to find a network size just large enough to capture the general class properties of the data. In some cases, however, network size is not reduced significantly or the reduction is satisfactory but generalization is affected. In this paper, we propose the coupling of genetic algorithms with weight elimination. Our objective is not only to significantly reduce network size, by pruning larger size networks, but also to preserve generalization, that is, to come up with pruned networks which generalize as good or even better than their unpruned counterparts. The innovation of our work relies on a fitness function which uses an adaptive parameter to encourage reproduction of networks having small size and good generalization. The proposed approach has been tested using both artificial and real databases demonstrating good performance.