摘要
对于自适应自然梯度算法,选择步长参数以达到好的分离性能是非常必要的。提出了一种步长自适应自然梯度算法。由于该算法中的步长基于分离状态,其学习速率由信号的分离程度自适应选取,因而能很好地解决收敛速度与稳态误差之间的矛盾。计算机模拟试验结果显示,该算法优于传统的自然梯度算法。
Careful selection of step-size parameters is often necessary to obtain good performance from adaptive natural gradient algorithm. An adaptive step-size natural gradient algorithm is proposed. Because the step-size of the algorithm is based on separating state, the learning ratio is chosen adaptively according to separating degree, and it can improve convergence speed and reduce the misadjnstment error in the steady state simultaneously. The simulation results demonstrate the performance of the proposed algorithm is superior to the usual natural gradient one.
出处
《电声技术》
2009年第1期63-64,71,共3页
Audio Engineering
关键词
自然梯度算法
步长自适应
收敛速度
稳态误差
natural gradient algorithm
adaptive step-size
convergence speed
misadjustment error
