深度回声状态网络概述  

Deep Echo State Network Overview

作　　者：程国建 魏珺洁 CHENG Guojian;WEI Junjie(Department of Postgraduate,Xi＇an Shiyou University,Xi＇an 710065,China)

机构地区：[1]西安石油大学研究生院,陕西西安710065

出　　处：《电子科技》2018年第8期92-95,共4页Electronic Science and Technology

基　　金：国家自然科学基金(40872087)

摘　　要：作为一种递归神经网络,回声状态网络凭借其简单的训练过程和独特的储备池结构受到广泛关注,目前已经应用于时间序列预测、非线性系统识别、定量预测等领域。然而回声状态网络也存在一些不足,例如储备池的优化问题、共线性问题等。许多研究学者尝试着优化回声状态网络的结构和性能,文中介绍的深度回声状态网络增加了特征链接和编码器,并且多次交替使用储备池和编码器来进行投影编码,改进了回声状态网络的性能。论文结合众多学者的研究,对深度回声状态网络的学习过程、应用和特点进行了详细介绍。该网络结构中加入了可以进行高维投影的编码器,在浅层回声状态网络的基础上进行了较大的改动,是一种值得研究的新型网络结构。As a kind of recurrent neural network,the echo state network had received extensive attention due to its simple training process and unique structure of reserver. It had been applied in many fields such as time series prediction,nonlinear system identification,and quantitative forecasting. However,there were some shortcomings in the echo state network,such as the optimization problem of the reserver and the problem of collinearity. Many researchers had tried to optimize the structure and performance of the echo state network. The deep echo state network introduced in this paper had added feature links and encoders,and alternately using the reserver and encoder for projection encoding,improving the performance of echo state network. In this paper,the study process,application and characteristics of deep echo state network were introduced in detail with the research of many scholars. This network structure incorporated an encoder that can perform high-dimensional projection,and had been greatly changed on the basis of the shallow echo state network. It is a new network structure worthy of study.

关 键 词：递归神经网络 回声状态网络 深度回声状态网络 储备池 编码器 特征链接 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]