中央研究院 資訊科學研究所

This presentation introduces advanced Bias Compensation for Kernel Adaptive Filtering Algorithms to address the critical challenge of nonlinear system identification in "dual-threat" environments characterized by noisy regressors and impulsive output noise. To counteract the parameter estimation bias and system instability caused by these non-ideal effects, we propose two novel approaches: the Bias-Compensated Kernel Least Mean Square (BC-KLMS) algorithm, which explicitly corrects for input noise bias , and the Bias-Compensated Kernel Maximum Correntropy (BC-KMC) algorithm, which utilizes a modified cost function to simultaneously neutralize input noise bias and mitigate impulsive output outliers. Validated through complex time-series predictions, such as sunspot and