Study On The Echo State Network Based On Nonlinear Readout And Its Application

As a new recurrent neural network, Echo state network attracted wide concern in academia recently because of its unique network topology structure and excellent performance. The core structure of the echo state network is the "reservoir", was originally proposed in2001by Jaeger."Reservoir" was used to replace the hidden layer of the traditional recurrent neural network, it can be randomly generated recurrent hidden-layer, so the scheme of network structure becomes simple, and greatly reduced the complexity of the output-weight training algorithm, and also could overcome the many defects of traditional recurrent neural network, such as the large amount of computation, slow convergence speed and "Memory fading" etc.At present, the research hotspots of the echo state network mainly focused on the adaptability of the reservoir, analysis of the stability of the network, realization of the hardware, the improved method of readout mode about the reservoir information, and son on. This paper mainly studies the improved readout mode about the reservoir information, using the different nonlinear structure as the readout unit, and transform the output unit based on linear regression of the classical echo state network into the structure based on nonlinear readout.This paper describes the basic principle, topology structure, and the training algorithm for output weights of the classic echo state network, and compares the performance advantages of the echo state network with the Elman neural network through three groups of benchmark experiment:a) Lorenz chaotic time series prediction, b) NARMA model of the nonlinear system modeling, and c) Speech characteristic signal classification.This paper studied the nonlinear readout mode about the reservoir information systematically, and proposed a new echo state network model based on radial basis function neural network as the nonlinear readout unit inspired by the research results of Boccato et al. and Butcher et al., and compared the performance of this method with the method that used the Volterra filter for the readout unit by Boccato et al proposed and that used the extreme learning machine for the readout unit by Butcher et al proposed through the first two groups of benchmark experiments above. The experimental results show that the echo state networks based on nonlinear readout overcome the shortcoming on the acquisition capability of the classical echo state network for the nonlinear characteristics of statistical data, and can effectively improve the performance of the network. Moreover, the performance of the echo state network with nonlinear readout unit based on the radial basis function neural network can be compared with the other two nonlinear readout structures.These new network structures had been applied to solve the problem of channel equalization. This paper choose three groups of nonlinear channel to compare the performance of the classic echo state network and the echo state network based on nonlinear readout unit using the method of supervised equalization and blind equalization respectively. The experimental results show that, whether it is supervision equalization or blind equalization, the echo state networks based on nonlinear readout unit are showing more excellent performance than the classical echo state network, so they are more suitable for nonlinear channel equalization.