Study On Echo State Network With The Simplest Reservoir And Its Application

The idea of reservoir computing was originally proposed by Jaeger in2001as the method of Echo State Network and Maass in2002by means of Liquid State Machine, respectively. Contrasting to the classical recurrent neural networks, the prominent features of reservoir computing are:randomly generated recurrent hidden-layer and simple weight training algorithm. The method of reservoir computing greatly reduces the computing burden of recurrent network training process, effectively overcomes the problem of memory fading and somewhat solves the issue of network structure. Hence, many researchers in related fields are attracted by this method. Echo State Network, as the simplest and the most representative model of this technique, has rapidly developed in recent years and achieved a wide range of successful applications in many areas.The topology structures and training algorithms of the classical Echo State Network have been intensively studied in this paper. Considering the problem that its reservoir is randomly generated, structurally massive and hardly implemented in hardware, based on the theory of delay differential equation, the author has constructed a simplest reservoir framework which only contains one nonlinear node, and detailedly presents its constitution theory, topology structure and realization method. Two groups of benchmark experiments which study on the relationship of several parameters that influence the performance of this new network and compares the capability of this novel model to that of the classical Echo State Network have been designed in this paper, the results suggest that this new network simplifies the topology structure of reservoir, and its properties are comparable to that of the classical one.This new network has been applied to the channel equalization of communication systems. Three groups of experiments have been designed in this paper:a) Repeats Jaeger’s channel equalization experiment, the results show that the equalization performance of this new model is comparable to that of Jaeger’s conclusion, b) Draws lessons from Boccato’s experimental design, completes equalization experiments of this new network. The results indicate that, both in linear and nonlinear channel, or in supervised and unsupervised equalization, its property is better than that of the Echo State Network with nonlinear readout device, c) Under the conditions of three typical channel models, completes supervised and unsupervised experiments of this novel model. The results imply that its capability exceeds that of the classical Echo State Network.