A Neural Network Ensemble Method With New Definition Of Diversity Based On Output Error Curve

Neural network ensemble can significantly improve system's generalization ability by training several networks and combining their outputs. However, because of the "Black Box" characteristic of neural network, defining diversity on inner structure has certain blindness. Moreover, it limits the development of neural network ensemble in some degree. Based on this point, a certain method of defining diversity based on the outputs of neural networks is proposed. It only considers the outputs of each network, defining diversity without considering inner structure of neural networks. Furthermore, it extends the thoughts of the development of future neural network ensemble.This paper refers and improves some algorithm from text match, considers every network output curve as a vector and calculates diversity. This algorithm avoids the discussion on network inner structure as most traditional algorithms do. Results show this algorithm is highly-effective.Because original algorithm only considers the diversity on the value and the diversity on the space is neglected, paper uses clustering algorithm to improve original algorithm. Thus, algorithm considers both value and space diversity and better results are achieved.For testing, there UCI standard data sets are used. Compared with traditional bagging algorithm, the algorithm proposed in this paper can get much better results.PTA is one important component in chemical industry, in the last part, paper uses actual industrial dataset from PTA factory for testing. By modeling and testing, it shows algorithm in this paper has high stability and priority.