Percentage-based hybrid pattern training with neural network specific cross over

In this paper, a new weight-setting method is proposed to improve the training time and generalization accuracy of feed-forward neural networks. This method introduces a percentage-based hybrid pattern training (PHP) scheme and aims to provide a solution to the problem dependency of other Genetic Algorithm (GA)-based Neural Network weight-setting methods. A neural network is trained using a neural network specific GA until a certain percentage of the training patterns is learned. The weights thus obtained are used as the initial weights for backpropagation (BP) training, which is then applied to complete the network training. Further improvement to the method was looked into and the use of a distributed GA in the weight-setting phase was investigated. The final approach derived was tested on four neural network problems - we observed that as the number of patterns trained using GA approaches 50% of the total number of training patterns, the proposed method is more effective in pulling the networks out of local minima. Additionally, the networks trained using this method showed as much as 75% improvement in training time and 15% improvement in generalization accuracy.