摘要
在应用MATLAB软件神经网络工具箱函数训练获取BP神经网络权值矩阵函数或非线性数学模型时,由于BP神经网络训练样本集中输入、输出(目标)样本参数的绝对值和离散性有时太大或过于集中,在网络权值矩阵误差函数逼近过程中,易产生局部误差最优或误差振荡等缺陷。本文介绍一种利用MATLAB神经网络工具箱prestd和poststd函数,通过对输入、输出(目标)样本参数进行正规化处理,使其不至于太小,可防止在网络权值矩阵函数训练时不会进入局部误差最优或误差振荡的缺陷区域。经实例分析计算证明,该方法在如何优化获得实用可靠的BP神经网络权值矩阵函数是有效可行的,对研究MATLAB语言用于网络系统控制和非线性函数或数学模型逼近研究工程技术人员具有一定的参考价值。
In applying MATLAB software network tool-box functions for the right matrix-function or nonlinear function of BP network trained, they are disperser too or over-centralized abstract value that input parameters and output parameters are applied for BP network trained, part min-gradients or deviation vibrations could be apt to occur in training the right matrix-function of BP network trained. In this paper, adopting the normalization of input and output parameters by applying the ＂prestd＂ and ＂poststd＂ functions from MATLAB network tool box, introduces a method of avoiding part min-gradient or deviation vibration deficiency brought by BP network training. By analyzing and calculating a example in the paper, it could account for the problems of part min-gradients or deviation vibrations to occur in training the right matrix-function of BP network trained, and is of reference to researchers in network system control and function approach by applying MATLAB software.
出处
《四川理工学院学报（自然科学版）》
CAS
2007年第3期98-100,共3页
Journal of Sichuan University of Science & Engineering(Natural Science Edition)
