GRNN与BPNN的函数逼近性能对比研究
2014-04-18丁硕常晓恒巫庆辉
丁硕+常晓恒+巫庆辉
摘 要: 为了研究GRNN和BPNN非线性函数的逼近能力,从数学角度详细阐述了GRNN和基于LM优化算法改进的BPNN的学习过程,编程建立了GRNN和BPNN,并分别用两种神经网络对指定的非线性函数进行逼近实验。仿真结果表明,在训练样本数量相等且中小规模网络的条件下,相对于BPNN而言,GRNN的逼近精度更高、收敛速度更快,具有很好的逼近能力,为解决非线性函数的逼近问题提供了良好的解决手段。
关键词: 广义回归神经网络; 反向传播神经网络; 函数逼近; 逼近能力对比; 仿真
中图分类号: TN711?34;TP183 文献标识码: A 文章编号: 1004?373X(2014)07?0114?04
Comparative study on function approximation performances of GRNN and BPNN
DING Shuo, CHANG Xiao?heng, WU Qing?hui
(College of Engineering, Bohai University, Jinzhou 121013, China)
Abstract: To study the nonlinear function approximation performances of GRNN and BPNN, the learning processes of GRNN and BPNN based on LM optimization algorithm improvement are illustrated mathematically in this paper. Then GRNN and BPNN were established with computer programming. A given nonlinear function was approximated by the two neural networks respectively. The simulation results indicate that when the numbers of training samples are the same and the networks are small or medium?sized, GRNN has higher precision, faster convergence speed, and better approximation ability than BPNN. Thus GRNN is a good method to solve the problem of nonlinear function approximation.
Keywords: GRNN; BPNN; function approximation; approximation capability comparison; simulation
0 引 言
数值逼近是指给定一组数据,用数学分析的方法来分析这组数据,常用的数学分析方法有多项式拟合和插值运算。由于人工神经元网络(Artificial Neural Networks, ANN)具有很强的非线性映射能力、自学习性和容错性,近些年来,采用ANN对于非线性函数进行逼近成为数值逼近领域的一个研究热点。目前,国内外学者绝大部分使用的ANN模型是BP神经网络,但是,传统的BP网络收敛速度慢,训练时间长和目标函数存在局部最小等缺点,所以,很多学者提出了许多改进算法。广义回归神经网络(General Regression Neural Networks, GRNN)具有很强的非线性映射能力和柔性网络结构以及高度的容错性和鲁棒性,很适合于解决非线性函数的逼近问题[1?5]。笔者以标准BP(Back Propagation)算法为基础,利用收敛速度相对较快、拟合精度较高且性能稳定的LM(Levenberg?Marquart)算法来构建LMBP神经网络,同时构建了GRNN,分别对指定的非线性函数进行逼近实验,并对结果进行比较分析。……
