摘要
基于最小二乘(LMS)统计算法的自适应线性元件(Adaline)神经网络是非线性分类的重要工具之一。从计算机仿真的角度研究随机逼近LMS学习方法的特点,从步长设置、收敛性、收敛速度、算法抗噪性、判断的准确率等多个参量评估随机逼近法的性能。仿真结果表明,对于不同的初始步长设置,神经元完成学习任务的训练时间不同;在保证学习收敛性的前提下,步长越大,收敛速度越快,但收敛的稳定性变差。权矢量的初始值设置对学习的收敛性没有影响。
The neural network of adaptive linear element （Adaline）based on least mean square （LMS）algorithm is one of the important tools of nonlinear classification.The characteristics of stochastic approximation LMS algorithm are researched by using computer simulation in this paper,the performance of the algorithm is evaluated by several parameters including setting of recursive step, convergence, time cost of convergence, robust of the algorithm and accuracy of classification. The simulation result shows that the training time needed of the neural network is different for different initial step setting, un- der assurance the study convergence,the bigger recursive step is,the faster convergence reaches,but the steady of the convergence is becoming worse. The initial value setting of weighting vector has no effect on study convergence.
出处
《电子设计工程》
2009年第9期88-90,共3页
Electronic Design Engineering
作者简介
张帆（1973-），女，陕西泾阳人，工程师。研究方向：智能计算及其在软件工程中的应用，软件测评与安防工程。
