Study On The Approach Of Chaotic Time Series Prediction Based On The Reservoir Computing

Chaotic time series prediction is one important and challenging work in the analysis of chaotic time series. With the characteristics of good nonlinear mapping, self-learning ability and robustness, artificial neural network has been widely applied in the chaotic time series prediction. At present, many research achievements concerning chaotic time series prediction with feed forward or traditional recurrent neural network have been made. However, feed forward neural networks is inadequate to deal with problems concerning Time Sequence due to its one-way transmission, and recurrent neural network is hard for practical application due to its excessive complex training, uncertain net structure and hypomnesia. As a new type of recurrent neural network model, the reservoir computing tries to avoid these problem. RC skillfully separates information expression from weights training process and also greatly simplify the training in traditional recurrent neural network. In recent years, reservoir computing has been growing rapidly with many successful applications and has become a research hotspot in the field of machine learning. Besides. its bottleneck problems, such as that of adaptability, the analysis of network stabililization and approaches of hardware realization, have been open research subjects.In this thesis, the prediction approaches of chaotic time series, which is based on neural network research, and two classic RC model, echo state network and liquid state machine, have been introduced systematically. Based on the example of echo state network, the thesis focuses on the classic research of chaotic time series prediction and discusses five points as follow:(1)The introduction of classic RC model;(2) Systematical research on chaotic time series prediction, based on feed forward neural networks. And conduct simulation experiments on Lorenz and the Sunspot Chaotic time series by the application of multi-layer perceptron, GRNN neural network;(3) Aiming at the prediction research of Lorenz and Sunspot chaotic time series, the thesis makes a close study of estimated performance of traditional recurrent neural network Elman net;(4)Study of three types of simply deterministic topology reservoir, including DLR, DLRB and SCR, and analysis of their short-term memory ability to satisfy the hardware realization, reaching the conclusion that the three types of reservoir are as good as the classic reservoir as the simulation experiments of Lorenz chaotic time series proves;(5) As to the prediction of chaotic time series, the thesis compares the different qualities of RC approach, feed forward neural networks and the prediction approaches of traditional recurrent neural networks.