Research And Application Of A New Selective Neural Network Ensemble Method Based On Error Vectorization

Neural network ensemble is an effective method to enhance theaccuracy of modeling，which trains several sub-networks and simplycombines the results. Increasing the diversity among sub-networks is themost efficient way to improve the accuracy of ensemble. However thereis no unified understanding of diversity, since there are many factors thatcan impact neural network. Traditional method trains sub-networks allindependently, which is lack of interaction among them. Thus thediversity among them is not large enough.This paper vectorizes the output errors of networks and analyses thefactor which impact the accuracy of ensemble in the view of vector, andthen obtain the method to calculate the diversity based on errorvectorization. After that the EVSNE (Error Vectorization based SelectiveNetwork Ensemble) method is proposed. This method trainssub-networks which can offset each other by altering the optimizationobject of training networks. The diversity of sub-networks trained byEVSNE is much larger than that trained by Bagging. EVSNE can combine with the traditional subsets regenerationmethod. The traditional method can be looked on as the datapreprocessing method of EVSNE. And then EVSNE combines withBagging named B-EVSNE and EVSNE combines with clustering namedC-EVSNE are proposed.Six classical datasets are utilized to validation the method. Themethod is compared with the traditional method Bagging. Experimentsand comparisons demonstrate that this method can get better results thanBagging. Moreover, the B-EVSNE and C-EVSNE perform better thanEVSNE. At last the methods are used in the HDPE (high-densitypolyethylene) producing process. The results manifest EVSNE andimproved EVSNE have high modeling accuracy and stability.