几种用于非线性函数逼近的神经网络方法研究  

Research on Several Neural Network Methods Applied in Nonlinear Function Approximation

作　　者：黄忠明[1] 吴志红[1] 刘全喜[2] 

机构地区：[1]四川大学计算机学院,四川成都610065 [2]西南技术物理研究所,四川成都610064

出　　处：《兵工自动化》2009年第10期88-92,共5页Ordnance Industry Automation

摘　　要：采用一个较小的学习率,在都满足相同误差指标的情况下,通过应用几种比较有代表性的BP网络,RBF网络和Elman网络逼近一非线性函数,得出各种不同的网络结构和训练算法对逼近结果的影响。L-M算法训练所需时间少,逼近精度较高。弹性BP算法的前向网络能消除梯度幅度的不利影响。变梯度SCG算法不需在迭代中进行线性搜索,从而避免搜索方向计算的耗时问题。径向基函数网络对于输入信号具有很好的局部逼近能力,对反馈型Elman网络而言,虽然其逼近效果也能满足误差指标的要求,但其训练所需的步数和时间却很长。With same error index, a lesser learning rate is adopted, several typically familiar methods of the BP network, RBF network and the Elman network are applied to investigate the nonlinear function approximation. And the influence of the different network structure and training arithmetic on approximation result is obtained. L-M arithmetic time less of training is requires and higher approximation accuracy. The former network of the stretch BP arithmetic can eliminate the adverse effect of the gradient magnitude. The changing gradient SCG arithmetic without linear search in the iteration, hence avoid the time-consuming problem of the search direction calculation. Radial basis function network has a good local approximation ability of the input signal. The approximation effect of the feedback Elman network can meet the request of the error index, but its training requires long steps and time.

关 键 词：函数逼近 神经网络 建模与仿真 性能分析 

分 类 号：O174.41[理学—数学] TP182[理学—基础数学]