Negatively Correlated Neural Network Ensemble with Multi-population Particle Swarm Optimization

Multi-population particle swarm optimization (MPPSO) algorithm is proposed to train negatively correlated neural network ensemble. Each sub-swarm is responsible for training a component network. The architecture of each component neural network in the ensemble is automatically configured. The component networks are trained simultaneously and successively exchange information among them. Because our approach automatically determines the number of hidden units for the component neural networks, the rational architectures of the component network are achieved and hence the performance of ensemble is improved. The experimental results show that MPPSO for negatively correlated ensemble is an effective and practical method.