摘要： 针对测试训练期间变化的信道环境,提出一种新的滑动窗近似线性依赖稀疏的核递推最小二乘算法。该算法核矩阵的尺寸只与滑动窗口宽度有关。选择字典表中最近的L个数据测试近似线性依赖准则,减少系统开销并降低系统实现的复杂度,克服ALDKRLS算法核矩阵随字典表线性增长的缺陷。当训练序列的自相关矩阵特征根谱大于40时,较SWKRLS均方误差性能有近3 dB的改善,且具有更小的稳态失调特性。仿真结果表明,与ALD-KRLS算法和KRLS算法相比,该算法具有更快的收敛速度和较好的均方误差性能。

关键词: 核递推最小二乘算法, 稀疏表示, 近似线性依赖, 滑动窗, 核矩阵, 高斯核函数

Abstract: Aiming at varied channel environment during test and training,this paper proposes a new Kernel Recursive Leastsquares(KRLS) algorithm with sliding-window approximately linear dependence sparsification.Kernel matrix size is only relative to the width of sliding window.The nearest L data in the dictionary are selected to test approximately linear dependence criterion.System consumption and complexity are redued to overcome the shortcomings of ALD-KRLS whose kernel matrix is linearly increasing with the size of the dictionary.When eigenvalue spectum of autocorrelation matrix of training data is more than 40,the proposed SWALD-KRLS outperforms the SW-KRLS with 3 dB of improvement in Mean Squared Error(MSE) performance,and obtains less misadjustment in stationary enviroment.Simulation results show that the proposed algorithm achieves faster convergence and obtains better MSE performance against ALD-KRLS and KRLS when the channel changes during the training phase.

Key words: Kernel Recursive Least-squares(KRLS) algorithm, sparse representation, Approximately Linear Dependence(ALD), sliding-window, kernel matrix, Gauss kernel function

中图分类号: 

• TP391