A Generalized <i>t</i>-Distribution-Based Kernel Adaptive Filtering Algorithm
Huchuan Tang, Hongyu Han, Sheng Zhang, Wenting Feng
Abstract
In this brief, utilizing generalized t distribution and maximum correntropy (MC) criterion, a new kernel adaptive filtering algorithm in reproducing kernel Hilbert space (RKHS) is designed for robust learning. As compared to existing methods, our algorithm is better suited to non-Gaussian impulse noise environments due to its ability to depict heavy tail characteristics more accurately. To restrain the scale growth of the neural network and reduce computing cost in the proposed algorithm, we also implement a simple vector quantization algorithm, called Gt-QKRGMC. Finally, superiority of the proposed algorithm is verified by tracking the Mackey-Glass (MG) time series prediction in the context of non-Gaussian noise interference.