Adaptive neural network ensemble algorithm

Different individual neural networks in an ensemble that learn different samples have different performance for the same input data. The weights of conventional ensemble method is fixed, it may decrease the performance of some individual neural networks which can have better performance and lower weights, so it can influences performance of whole ensemble. An adaptive neural network ensemble (ANNE) algorithm is proposed, which dynamically adjusts weights of an ensemble based on clustering analysis. The algorithm use clustering analysis to classify the training samples in different classes which is used to train different individual neural networks. The weights of an ensemble are adjusted by the correlation of input data and the center of different sample classes. ANNE can increases the diversity of different individual NNs and decreases generalization error of ensemble. ANNE is a algorithm of not only weights assignment, but also training individual NNs. The paper provides both analytical and experimental evidence that support the novel algorithm.