Research And Application Of Learning Mechanism Of Echo State Network

With the wide application of artificial intelligence in daily life,the research of related technologies in the field of artificial intelligence has also attracted wide attention.Echo state network(ESN)is an important method in the field of artificial intelligence.As a non-linear adaptive dynamic system,it contains a large number of artificial neurons connected by synapses.By imitating the structure of human brain and the behavior characteristics of neural network,it can realize the function of distributed and parallel information processing.Because of its fast learning speed,it has been successfully applied to the prediction of time series and network traffic.It becomes one of the research hotspots.In the research of echo state network,there are always problems that the random connection between neurons in the reservoir and the full connection of output synapses reduce the prediction performance of the network.To solve the above problems,this dissertation mainly studies the echo state network from three aspects,which are output synapse,synapse and neuron connection structure in the reservoir.The specific work of this dissertation is as follows.First,this dissertation proposes an echo state network with sparse output synapses and applies it in the prediction of chaotic time series.This dissertation discovers and verifies the existence of redundant output synaptic connections in the echo state network,proposes a method for locating redundant output synapses based on energy efficiency,and analyzes the reasons for the existence of redundant output synapses.The proposed method of locating redundant output synapses is used to compare the weight states of other four different output synapses.In the prediction of large scale echo state networks,the network prediction performance of the improved model is analyzed.This dissertation gives numerical simulation experiments of different chaotic systems.In the prediction of chaotic time series,the echo state network with sparse output synapses has better prediction performance.Second,the dissertation proposes a memristor-based echo state network and conducts simulation experiments in chaotic time series prediction.A memristor is a device with a memory storage function of its own resistance state.This dissertation improves the random connectivity of neuronal synapses in the reservoir,and improves the memory performance and network prediction performance of the reservoir by introducing the characteristics of the memristor into the echo state network,the memristor replaces the synapses between neurons in the reservoir,and the value of the memristor represents the weight of the synapse.Memristor-based echo state network has been conducted simulation experiments in chaotic time series prediction,and we compare the memristor-based echo state network with the classical echo state network in three aspects.At the same time,it is compared with other echo state network variants.Memristor-based improves the prediction accuracy and prediction performance of the echo state network in the chaotic time series prediction.Third,the dissertation proposes a dropout-based echo state network(Dropout ESN)for network flow prediction.The randomness of the reservoir of the traditional echo state network is weakened,the connection degree of neurons in the reservoir is reduced by introducing the dropout method into the echo state network.This dissertation applies the Dropout ESN to the actual network flow prediction task,set the neurons in the reservoir to stop working with different probabilities.The prediction performance of the improved echo state network algorithm and the classical echo state network algorithm is compared from three aspects.At the same time,it is compared with other networks.The results of comparative experiments show that the method has better performance in predicting network flow.