Mapreduce Based Selective Neural Network Ensembles Using GA

摘要

Neural Network Ensembles (NNE) is proved an efficient method in both regression and classification. NNE is suitable for parallel environment owing to its computing construction. A Mapreduce based Selective Neural Network Ensembles algorithm using GA (MSNNE-GA) is proposed in this paper. MSNNE-GA uses Mapreduce to train member neural networks and also implements the GA on Mapreduce framework to select member networks for NNE. Some experiments have been done on the two-spiral classification problem. The result shows the effect of decreasing the MSE and increasing the accuracy rate of NNE with MSNNE-GA. The time- consuming of the whole process can be reduced by MSNNE-GA with the linear increasing of speed-up rate. At last the ability and sensitivity of different algorithms for big scale NNE is discussed. Comparing Bagging, MSNNE-GA can more effectively make use of the great number of member neural network to optimize the result of NNE.