| Title |
Adaptive Kernel-width Estimation for Learning Algorithms based on Maximum Zero-Error Probability |
| Authors |
김남용(Namyong Kim) ; 권기현(Kihyeon Kwon) |
| DOI |
https://doi.org/10.5573/ieie.2021.58.10.3 |
| Keywords |
Adaptive kernel-width; Zero-error probability; Error probability density; ITL; Impulsive noise |
| Abstract |
The existing adaptive methods for optimum kernel-width estimation are to make the estimated error-distribution most similar to its real error-distribution. These methods are inappropriate to the weight adjustment of MZEP (maximum zero error probability) algorithm which is a type of the information theoretic learning (ITL). In this paper, a new adaptive estimation method for kernel-width for the effective adjustment of weights of the MZEP algorithm is proposed. Two separate MZEP algorithms operate in parallel for calculation of two different averaged error powers. The kernel-width for the main MZEP is updated in order for the error power to decrease and the one for the sub-MZEP is defined as a small increased value. The proposed method produces stable and fast convergence in the experiment on an impulsive-noise inflicted communication system both for very small initial kernel-widths that can cause instability and for very large initial kernel-widths that can induce very slow convergence. These results confirm that the proposed adaptive kernel-width estimation method can significantly enhance stability and performance of ITL learning systems employing the MZEP algorithm. |