基于扩展卡尔曼滤波的回声状态网络在线训练算法  

An on-line training algorithm for echo state networks with extended kalman filter

作　　者：王建民[1] 彭宇[1] 彭喜元[1] 王红[1] 

机构地区：[1]哈尔滨工业大学自动化测试与控制研究所,哈尔滨150001

出　　处：《仪器仪表学报》2011年第7期1514-1520,共7页Chinese Journal of Scientific Instrument

基　　金：教育部新世纪优秀人才支持计划(No.NCET-10-0062);教育部高等学校博士学科点专项科研基金(No.20092302110013);装备预研重点基金(No.9140A17040409HT01)资助

摘　　要：针对在线应用中回声状态网络(echo state network,ESN)的储备池适应性和训练算法效率问题,文中提出一种基于扩展卡尔曼滤波(extended kalman filter,EKF)的ESN在线训练算法。该算法以ESN的储备池参数以及输出连接权矩阵为目标参数,利用EKF对其进行联合训练提高储备池适应性,并能够有效地克服交叉验证参数选择导致的ESN训练效率下降问题。Lorenz混沌时间序列以及移动通信话务量时间序列预测实验证明,新方法可显著提升ESN算法的总体计算效率。A novel on-line training algorithm base on Extended Kalman Filter（EKF） is proposed for the adaptation of reservoir and training efficiencies Echo State Network（ESN）.The output weight-matrix and parameters of reservoir are selected as objective parameters and trained simultaneously by EKF.The adaptive property of reservior is improved by simultaneously training the objective parameters,which also can overcome the poor training efficiencies resulted from cross validation for parameter selection.The simulation experiments on benchmark data set Lorenz and mobile traffic time series show that the method proposed in this paper can improve the training efficiencies of the ESN when the parameters selection of reservoir are considered.

关 键 词：时间序列预测 递归神经网络 回声状态网络 扩展卡尔曼滤波 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]