A comparative study of ensemble back-propagation neural network for the regression problems

Abstract

This paper proposes a comparative analysis of the ensemble back-propagation neural network for the regression tasks. The ensemble technique is objectively used to improve the accuracy of single back-propagation neural network. Such technique alleviates uncertain generalization of the trained network due to its random initial weight and bias values and noisy data. This comparison includes linear regression, back-propagation neural networks, support vector machine, k-nearest neighbor, ensemble voting and bagging techniques. Seven benchmark regression datasets were used for evaluation. The experimental results indicated that the voting and bagging ensemble techniques provided considerable improvement. In addition, continued from the previous work, this paper also applied ensemble techniques to predict monthly rainfall time series data and compared to the back-propagation neural network optimized by the genetic algorithm. The results showed that voting and bagging ensemble techniques as well as genetic algorithm outstandingly improved the performance of single back-propagation neural network.