Echo State Network Algorithm Improvement And Application

Many issues include slow convergence, easy to fall into the local optimal, too complex training algorithm and other issues seriously limit the application of traditional recursive network, which prompted a new recursive network-echo state network came into being. Echo state network only need to train the output weights, and use the the learning algorithm of inverse to overcome some of the inherent problems of the traditional recursive network, and gradually become one of the important ways to predict time series. However, there are some problems in echo state network, and the improvement of echo state network algorithm is studied in this pape.Due to the random generation reservoir in the echo state network is not related to the specific issues, and the parameters are difficult to determine. To solve this problem, an orthogonal matrix synthesis method is introduced to build reservoir, which can regulate the spectral radius and the degree of reservoir. The experiment shows that the optimal adjustment of the reservoir has a better forecasting effect than the traditional one.In view of the poor performance of echo state network for some time series which are both cyclical and trend. Based on kernel regression algorithm, this paper proposes a kernel regression echo state network. Unlike the least squares in the traditional echo state network, this network uses the weighted least squares to carry out the output weights training. The new method is tested by the experiment to predict the nonlinear time series with periodicity and trend.Combining Kalman filter and echo state network to construct an online predictor which can overcome the traditional recurrent neural network (RNN) need to collect a large number of samples to fit and predict the defects, in ensuring the prediction accuracy at the same time makes the algorithm applicable scope is expanded. The effectiveness of the proposed method is verified by an example.