A distributed genetic algorithm improving the generalization behavior of neural networks

Abstract

Artificial neural networks sometimes generalize poorly to unknown inputs, if they have been trained perfectly on relatively small training sets using standard learning algorithms like e.g. backpropagation. In this paper a distributed genetic algorithm is designed and used to improve the network's generalization capabilities by reducing the number of different weights in the neural network.