文献类型：期刊文献

中文题名：带一类自适应非线性特性的变步长BP学习算法

英文题名：BP Training Algorithm of Alterable Pace with a Class of Adaptive Nonlinear Property

作者：杨慧中[1];陶振麟[1];张素贞[1]

机构：[1]华东理工大学自动化研究所,上海200237

年份：2001

卷号：13

期号：Z1

起止页码：108

中文期刊名：系统仿真学报

外文期刊名：Journal of System Simulation

收录：CSTPCD;;Scopus;CSCD:【CSCD2011_2012】；

语种：中文

中文关键词：神经网络;BP学习算法;自适应非线性特性项;误差代价函数

外文关键词：neural networks; BP learning algorithm; adaptive nonlinear property terms; square error function

摘要：针对前馈神经网络的反向传播(BP)学习算法收敛速度慢、易陷入局部最小等缺点,本文提出了在BP搜索进入误差代价函数曲率较小、收敛速度较慢处时,在变步长BP学习算法的基础上,引入一个非线性特性项,并将该特性项的强度系数构造为具有升温、降温策略控制的自适应非线性函数。仿真结果表明,该算法的收敛稳定、快速,具有较好的效果。
Because there are some disadvantages of slow convergence and getting easily into local minima in training feed forward neural networks, a nonlinear property term is inducted based on alterable pace BP algorithm when the error抯 gradient is small and convergent speed is slow, and its intensity coefficient is constructed, which is an adaptive nonlinear function using temperature-raising and temperature-descent strategy. The simulation result shows stable and fast speed convergence, as well as the efficiency of such algorithm.