Midpoint-Validation Method of Neural Networks for Pattern Classification Problems

In this paper, we propose a midpoint-validation method, which improves the generalization of neural networks. The problem associated with the former cross validation method is that efficiency is affected due to the separation of training data into two or more set. As for the proposed method, it creates midpoint data from the known training data and calculates a set of criteria using the newly created midpoint data and the previous training data. The implementation is easy since there is no unnecessary processing involved in separating the data into two or more sets. The advantage of the proposed method is that the method becomes much more efficient compared to the former method due to the numerical simulation used. We compare its performance with those of the support vector machine (abbr. SVM), multilayer perceptron (abbr. MLP), radial basis function (abbr. RBF) and the proposed method was tested on several benchmark problems. The results obtained from the simulation carried out shows the effectiveness of the proposed method.