1. Decorrelation algorithm based on the information theoretic learning.
- Author
-
Hou, Xinyan, Zhao, Haiquan, and Long, Xiaoqiang
- Subjects
- *
ADAPTIVE filters , *SYSTEM identification , *PARAMETER estimation , *ALGORITHMS , *COMPUTER simulation - Abstract
Recently, the Bayesian decorrelation algorithms have gained attention because they can effectively avoid trade-off between learning rate and estimation accuracy by circumventing the problems associated with constant step-size and significantly enhance the convergence speed of adaptive filtering algorithms by decorrelation of signals. However, the present Bayesian decorrelation algorithms are constructed on the basis of the minimum-mean-square-error (MMSE) criterion, leading to deterioration in performance when confronted with non-Gaussian noise environments. Therefore, this paper introduces the maximum correntropy criterion in the decorrelation algorithm and develops a decorrelation recursive maximum correntropy criterion (DRMCC) algorithm based on the observation model to achieve robust performance. The proposed algorithm has a similar iterative equations as the Bayesian decorrelation algorithm. In addition, we present the convergence condition and analyze the transient behavior of the DRMCC algorithm. When the free parameters of the proposed algorithm are unknown, a common method of parameter estimation is provided. Finally, numerical simulations corroborate that the designed DRMCC algorithm can realize outstanding performance under system identification and acoustic echo cancellation (AEC) applications in non-Gaussian noise environments and the derived theoretical model can arrive at good agreement with simulation results for both stationary and nonstationary scenarios. • The decorrelation recursive maximum correntropy criterion algorithm is proposed. • The convergence condition and transient behavior of the DRMCC algorithm are analyzed. • The proposed algorithm has excellent performance in acoustic echo cancellation. [ABSTRACT FROM AUTHOR]
- Published
- 2025
- Full Text
- View/download PDF