Extreme Learning Machine And Its Application In Channel Blind Equalization

Extreme learning machine is a novel single hidden layer feedforward neural net-work whose structure is simple and hidden layer nodes parameters can be randomly generated,then the output layer transforms the training of the network into a linear problem,tending to obtain the smallest norm of weights,thus,ELM runs extremely fast and has the good generalization performance,therefore,it has been widely used in func-tion fitting and classification and other fields.The blind equalization achieves higher spectrum utilization since no training sequence is required.Due to the powerful non-linear mapping and modeling capabilities,neural networks have become an effective method for solving blind equalization problems.Through the further analysis of the theory of Extreme Learning Machine,corre-sponding improvements have been made for the generation method of hidden layer's parameters and training method of output weight in this paper,and successfully applied to the blind equalization of communication channels.The specific research content are:(1)Although the random generation of the parameters of the hidden layer in ELM accelerates the speed of training,it limits the full exploration of the data features,es-pecially facing with big data or complex data,"curse of dimensionality" will result in deterioration of its performance.To solve this problem,the Restricted boltzmann ma-chine(RBM)is used to generate parameters of hidden layer in ELM,simultaneously,a grouping scheme is introduced to improve the training speed of RBM and avoid high computational complexity.Thus,two new model structures,RBM-H-ELM and FRBM-H-ELM,have been constructed to better extract feature information of high-dimensional large data.The performance of the new models were tested through the classification and regression benchmark experiments.The results show that while ensuring the train-ing speed is comparable to the traditional ELM,the classification accuracy and regres-sion performance are significantly improved,especially in high-dimensional data.The linear solution of the output weight in ELM simplifies the training process,but it loses some high-order statistical characteristics of the data information obtained through hidden layer mapping.To solve the problem,this paper adopts a nonlinear solution for output weight through the Volterra filter combined with the principal com-ponent analysis(PCA)technology,and we apply it flexibly to extreme learning machine and reduced kernel extreme learning machine(RKELM)to construct two novel models,PVELM and PVRKELM,making it possible to extract high-order information of the signals mapped by the hidden layer.The performance of the new models were tested through many classification dataset.The results show that compared to the traditional ELM,whether it is a binary classification or multiple-classification problem,classifi-cation accuracy is improved 2%-10%while the training time is almost the same.(3)Based on the prediction principle,the ELM,RKELM,and the proposed PVELM and PVRKELM are used to solve the problem of blind equalization of communication channels,and simulation experiments were performed through many different channel-s.The results show that compared with support vector machine(SVM),the proposed algorithms can achieve better equalization performance with less training time.Com-pared with ELM and RKELM,PVELM and PVRKELM can achieve a sharp drop in equalization error,and greatly improve the equalization performance.