Improvement And Application Of Back Propagation Network Based On Partial Least- Squares Algorithm

The partial least square method is a multivariate statistical analysis method. This method has its popular utility in terms of chemical plant, financial analysis, market analysis, medicament analysis, computation chemical, industry design etc. This method settles preferably the problem which doesn't been settle by the common multivariate linear regression. This method applies multivariate regression, canonical correlational analysis and principal component analysis to data analysis. It realizes manifold data analysis method.The back-propagation algorithm has been widely recognized as an effective method for training feed-forward neural networks. This method has its popular utility in terms of pattern recognition, control engineering, signal processing and economic prediction etc. However,there exists several essential defects in back-propagation algorithm as follows: slow convergence,trap in local minima , worse tolerant capacity and there is not the uniform theory which the weights initialization of input floor and output floor are set , the member of hidden nodes are set and the member of hidden floors are set . This paper improve on the defect which there is not the uniform theory which the weights initialization of input floor and output floor are set , the member of hidden nodes are set and the member of hidden floors are set in BP network.Firstly, we propound a novel BP network model based on nonlinear iterative partial least-squares algorithm which can fit nonlinear data after we expound partial least-squares algorithm systematically in this paper and we validate this new model and compare to the BP, PLS models for instance. The result showed that the new model had better fitting and forecast than BP, PLS models. Thus ,the defect which there is not the uniform theory which the weights initialization of input floor and output floor are set , the member of hidden nodes are set and the member of hidden floors are set in BP network is settled partly. This model can reduce iterative step number and advance learning efficiency a certain extent, also advance availability the performance of back-propagation algorithm.Secondly, we base on the same idea. We use the orthogonal signal correction method to settle the defect which there is not the uniform theory which the weights initialization of input floor and output floor are set , the member of hidden nodes are set and the member of hidden floors are set in BP network. We propound a novel BP network model based on the orthogonal signal correction method and we validate this new model and compare to the BP, PLS models for instance. The result showed that the new model had better fitting and forecast than BP, PLS models.