A new neural network ensemble approach and its application on meteorological prediction

Ensembles of artificial neural networks (ANN) show robust generalization capabilities that outperforms single network. However, for aggregation to be advantageous, each network should be optimal. To achieve this, a particle swarm optimization (PSO) method is used to evolve neural network architecture and connection weights. The ensemble strategy is based on a quadratic programming which is employed to calculate optimal non-negative weights. This training method and ensemble scheme are favorably tested against other methods, producing a sensible improvement on meteorological prediction.