.png){kind=icon}
.png){kind=icon}
Return
Sparse Penalty Data-Reusing Minimum Kernel Risk-Sensitive Loss Algorithms
.png){kind=icon}
PRE
AI
DOI：10.1007/s00034-025-03345-w.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
In comparison to the loss based on correlation entropy, the kernel risk-sensitive loss (KRSL) defined in the kernel space demonstrates a superior performance surface, rendering it more robust in non-Gaussian signal processing. To mitigate the limitation of the MKRSL algorithm in processing correlated input signals, the data-reusing minimum kernel risk sensitive loss (DRMKRSL) algorithm is derived by using a data-reusing strategy to modify the weight updating equation of the MKRSL-based adaptive filtering algorithm. In order to address sparse systems in realistic scenarios, three sparse penalty terms are introduced into the DRMKRSL algorithm cost function, which proposes the zero-attraction DRMKRSL (DRMKRSLZA) algorithm, the reweighted zero-attraction DRMKRSL (DRMKRSLRZA) algorithm, and the logarithmic norm zero-attraction DRMKRSL (DRMKRSLLZA) algorithm. Furthermore, a theoretical analysis of the algorithm's performance is provided, along with a comparison of its computational complexity. Simulation experiments conducted in system identification and echo cancellation demonstrate that proposed algorithms not only strongly prove the robustness of impulsive noise, but also exhibit good performance in accurately estimating sparse systems.
Keywords:
Kernel risk-sensitive loss
Data-reusing
Zero attraction
Impulsive noise
Sparse system identification
Echo cancellation
Journal
C
IF:
2
Papers:
318
Citations:
0