An Improved Bagging Neural Network Ensemble Algorithm and Its Application

For aggregation to be effective the component artificial neural networks (ANNs) must be as accurate and diverse as possible, an improved Bagging neural network ensemble algorithm is proposed to cope with this problem. The Euclidean distances between two arbitrary samples of the original training set are analyzed, the training subsets of component ANNs are distilled from this set then. The subsets elements have good properties of ergodicity and representativeness in sample space. The outputs of component ANNs are combined via weighted averaging and the optimal weights are determined by particle swarm optimization. Experimental studies on four typical regression datasets show that this approach has improved the quality of training subsets. Thus, the ensemble generalization ability is improved. Finally the improved algorithm is applied to construct an ANN-based soft sensor model for real-time measuring the ethylene yield. Application results show that this model has high measuring precision as well as good generalization ability.