构建复杂回响状态网络的新方法  

A New Method to Construct Complex Echo State Networks

作　　者：宋青松[1] 冯祖仁[1] 

机构地区：[1]西安交通大学系统工程研究所,西安710049

出　　处：《西安交通大学学报》2009年第4期1-4,共4页Journal of Xi'an Jiaotong University

基　　金：国家重点基础研究发展规划资助项目(2007CB311006);国家高技术研究发展计划资助项目(2006AA04Z222)

摘　　要：研究了人工神经网络混沌时间序列预测应用中的拓扑选择问题,受大脑皮层生长发育过程启发,提出一种类皮层网络——复杂回响状态网络(CESN)的构建方法.将影响皮层网络生长发育过程的依赖距离和依赖时间窗口的2个生长机制中的规模因子、距离敏感因子、密度敏感因子、种子神经元个数,以及时间窗宽度等调控因子,类比地定义为CESN构建过程中的构造参数.实验发现,CESN的拓扑结构能够被这些构造参数惟一地确定,并且当种子神经元个数取值约为网络规模的10%时,构建出的CESN几乎具有最优的预测精度.根据这个发现,仿真证实了CESN能够被更省时地构建,表明所提方法是一种高效的网络构建方法.The issue of topology choosing in artificial neural networks for chaotic time series forecasting is studied. Motivated by the growth and development process of cerebral cortex, an algorithm for constructing cortex-like neural network, complex echo state network （CESN）, is proposed. Some factors in both the distance-dependent growth mechanism and time-window-dependent growth mechanism, which have effects on the process of the growth and development of the cortex networks, are analogously defined as construction parameters that controls the process in creating a CESN, such as the size factor, the distance-sensitive factor, the density-sensitive factor, the seed-neuron number and the time-window width, etc. It is found from experiments that the topological characters of a CESN can be uniquely determined by these parameters, and that when the seed neuron number is set to be about 10% of the whole network size, the created CESN can have almost the best forecasting performance. Simulation results show that CESNs can be created with less time-consumption and better forecasting performance. It can be concluded that the proposed algorithm is a more efficient network constructing method.

关 键 词：混沌时间序列预测 复杂度 回响状态网络 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]