基于S形函数的变步长LMS算法  

Variable Step-size LMS Algorithm Based on Sigmoid Function

作　　者：亢润东[1] 宋国强 

机构地区：[1]中国电子科技集团公司第五十四研究所,河北石家庄050081 [2]空军驻京石地区军事代表室,北京100078

出　　处：《无线电工程》2013年第11期29-32,共4页Radio Engineering

摘　　要：针对定步长LMS算法在收敛速度、时变跟踪能力和稳态失调噪声几个重要指标上不能兼顾的问题,引入人工神经网络中一种常用激励函数———S形函数,并将其应用于变步长LMS算法中。结合算法对收敛速度、精度及稳态失调噪声的要求,引入改变S形函数曲线曲率及收敛终值的2个参数:α和β。分析当α和β值固定时的情况,仿真结果表明定参数算法若其值选择不当会引起较大误差。按照步长值在时变阶段自适应增大,在稳态阶段步长很小的原则,构造了变参数α(n)和β(n),仿真结果表明变参数算法兼顾多个参数,整体表现较好。The conventional LMS algorithm has difficulty in gaining the fast convergence speed,time-varying tracking capability and low steady state noise at the same time. This paper describes a common activation function in artificial neural networks： S-shaped function,and applies it to variable-step size LMS algorithm. Taking into account the convergence speed,accuracy and steady state noise,this paper introduces two parameters-α and β to change the curvature of curve and convergence value of S-shaped function. When α andβ are fixed parameters,the simulation results show that the improper value can result in large error. According to the idea that the step size of the algorithm increases adaptively when the channel is varying with time and becomes smaller during the steady state,the function of α（ n） and β（ n） is constructed. The simulation results show that the algorithm outperforms the former algorithm.

关 键 词：S形函数 变步长LMS 误差信号 稳态失调噪声 

分 类 号：TN911[电子电信—通信与信息系统]