A new selective neural network ensemble with negative correlation

An ensemble of neural networks exhibits higher generalization performance compared to a single neural network. In this paper, a new design method for a neural network ensemble is proposed. The hierarchical pair competition-based parallel genetic algorithm (HFC-PGA) is employed to train the neural networks forming the ensemble. The aim of the HFC-PGA is to achieve not only the best neural network, but also a diversity of potential neural networks. A set of component neural networks is selected to build an ensemble such that the generalization error is minimized and the negative correlation is maximized. Finally, some experiments are carried out using several data sets to illustrate and quantify the performance of the proposed method.