A Study Of The Back Propagation Extreme Learning Machine And Its Application

The neural network is a kind of model with non-linearity,which has good generalization performance.However,traditional neural networks,such as Support Vector Machines,Back Propagation Neural Networks and Radial Basis Function Neural Networks,have some issues of low convergence rate,over-fitting,difficulty of selecting the optimal parameters of hidden node,and so on.To reduce network output error and raises its training speed,the Extreme Learning Machine(ELM)proposed by Huang et al.randomly selects the parameters of hidden node during training process and is widely used because of its high efficiency.However,the optimal parameters selection of hidden node and over-fitting limits development of the Extreme Learning Machine.In order to solve the above-mentioned problems,based on the basic ELM theory,further research work is conducted and three improved ELM algorithms are proposed in the thesis.The main contributions are as follows:Firstly,the Back Propagation Extreme Learning Machine(BP-ELM)is proposed to further improve the generalization performance of the Incremental Extreme Learning Machine(I-ELM).The original idea of BP-ELM is to discover better parameters of hidden nodes which are calculated via formulas.In the thesis,the concrete approximation order of BP-ELM to arbitrary continuous target function is calculated,and the major results are presented in the form of theorem.Simulation results show that BP-ELM has more compact neural network architecture and enhances the learning efficiency observably.Secondly,BP-ELM can reduce the residual error of neural network as soon as possible,this is done by discovering the better parameters of hidden nodes and the corrected output weights.However,BP-ELM does not recalculate the output weights of all the existing nodes when a new node is added.The Back Propagation Convex Extreme Learning Machine(BP-CELM)shows that while maintaining the same simplicity,the convergence rate of BP-ELM can be further improved by recalculating the output weights of the existing nodes derived from a convex optimization method when a new hidden node is randomly added.By theoretical analysis and simulation,it is verified that BP-CELM obtains a faster convergence rate and more compact neural network architecture while keeping the efficiency and simplicity of BP-ELM.Thirdly,BP-ELM and BP-CELM can greatly improve the learning efficiency by calculating the optimal parameters of hidden node.Even so,how to determine the optimal number of hidden nodes is still uncertain.Based on the Bidirectional Incremental Extreme Learning Machine,the Back Propagation Bidirectional Extreme Learning Machine(BPBELM)is proposed in the thesis.Unlike BP-ELM and BP-CELM,even-numbered hidden nodes parameters of BP-BELM are calculated by feedback residual error.BP-BELM is proved to have universal approximation ability and the relationship between residual error and network output weights,called error-output weights ellipse equation is given.By theoretical analysis and simulation,BP-BELM obtains better performance than BP-CELM and BP-ELM.Lastly,since temperature and humidity of solar greenhouse are neither steady nor linear,single traditional predicting method easily lead into low precision and robustness.To solve these problems,the thesis proposes a combined nonlinear prediction model BPBELM-EMD based on BP-BELM and Empirical Mode Decomposition(EMD).In modeling,EMD is adopted to decompose the sequence data of temperature and humidity in solar greenhouse into a series of Intrinsic Mode Function(IMF),predict each IMF component,and then combine the predicted values of each IMF component to get the predicted value of original solar greenhouse temperature and humidity.BP-BELM-EMD is tested through the experiment of predicting temperature and humidity of solar greenhouse using real-world data in the solar greenhouse.