Constructive Ensemble of RBF Neural Networks and Its Application to Earthquake Prediction

Neural networks ensemble is a hot topic in machine learning community, which can significantly improve the generalization ability of single neural networks. However, the design of ensemble architecture still relies on either a tedious trial-and-error process or the experts' experience. This paper proposes a novel method called CERNN (Constructive Ensemble of RBF Neural Networks), in which the number of individuals, the number of hidden nodes and training epoch of each individual are determined automatically. The generalization performance of CERNN can be improved by using different training subsets and individuals with different architectures. Experiments on UCI datasets demonstrate that CERNN is effective to release the user from the tedious trial-and-error process, so is it when applied to earthquake prediction.