一种基于行为空间的回声状态网络参数优化方法  

An Echo State Network Parameter Optimization Method Based on Behavior Space

作　　者：张昭昭 朱应钦 乔俊飞 余文 ZHANG Zhaozhao;ZHU Yingqin;QIAO Junfei;YU Wen(College of Computer Science and Technology,Xi'an University of Science and Technology,Xi'an 710054,China;Faculty of Information Technology,Beijing University of Technology,Beijing 100124,China;Department of Control Automatic,Cinvestav-IPN,Mexico City 07360,Mexico)

机构地区：[1]西安科技大学计算机科学与技术学院,陕西西安710054 [2]北京工业大学信息学部,北京100124 [3]墨西哥国立理工学院高级研究中心自动控制中心,墨西哥墨西哥城07360

出　　处：《信息与控制》2021年第5期556-565,共10页Information and Control

基　　金：陕西省自然科学基础研究计划陕煤联合基金资助项目(2019JLZ-08);陕西省自然科学基础研究计划资助项目(2020JM-522,2021JM-396);国家重点研发计划(2018YFC1900800-5,2018YFC1900801)。

摘　　要：针对回声状态网络参数难以选择的问题,提出一种基于行为空间优化回声状态网络参数的方法.其实质是通过泛化等级、核心等级、记忆容量构建回声状态网络行为空间.优化算法采用新颖搜索遗传算法(NSGA),该算法结合K近邻个体距离和NMSE,通过建立行为空间最低配置筛选基因来限定遗传算法的遗传方向,提高优化效率,进而找到影响网络性能的因素.该方法克服了传统回声状态网络(ESN)参数选择困难、遗传算法优化时间长且无合适理论阐明储层性能对任务的影响等缺陷,提升了优化效率和网络学习性能.实验结果表明,本文所提NSGA-ESN方法优化ESN参数基本上接近最佳网络结构,学习性能优于增长回声状态网络,且可以通过行为空间解释影响ESN网络性能的原因.Aiming to solve the problem of selecting echo state network parameters,a method for optimizing echo state network(ESN)parameters based on the behavior space is proposed.The essence is to construct ESN behavior space through generalization rank,kernel rank,and memory capacity.The optimization algorithm adopted the novel search genetic algorithm(NSGA),which combines the K-nearest neighbor individual distance and normalized mean squared error,limits the genetic direction of the genetic algorithm by establishing the minimum configuration of the behavior space to screen genes,therefore improves the optimization efficiency.Next,the factors that affect network performance were determined.This method overcomes traditional ESN parameter selection difficulties,the long optimization time of the genetic algorithm,and no suitable theory to clarify the impact of reservoir performance on tasks.As a result,it improved the optimization efficiency and network learning performance.The experimental results showed that the proposed NSGA-ESN method optimizes ESN parameters close to the best network structure,with a better learning performance compared to the growing echo state network.Furthermore,the factors that affect the performance of the ESN network can be explained through the behavior space.

关 键 词：回声状态网络 泛化等级 核心等级 记忆容量 行为空间 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]