Robust learning of neural networks ensemble for modeling

Neural networks ensemble (NNE) has recently attracted great interests because of their advantages over single neural networks (SNN) as the ability of universal approximate and generalization. However, the performance of NNE trained by least squares methods deteriorates when lots of outliers emerge in I/O data. In this paper, a robust learning algorithm of NNE is applied based on the theory of robust regression that may deal well with the problems of outliers. Initial empirical study demonstrates that the robust learning algorithm of NNE has better precision and generalization than both neural networks ensemble with least square function and single neural network with the same robust algorithm do, when trained under the data with outliers.