Fusion-based multi-kernel learning filter with maximum correntropy criterion

Kernel learning filters have been effective tools for addressing nonlinear function fitting problem. In these filters with Gaussian kernels, the performance depends on the choice of the kernel width and an inappropriate kernel width might degrade the learning performance. To further improve the learning performance, a fusion-based multi-kernel learning filter with maximum correntropy criterion is proposed in this paper, in which multiple learning filters with different kernel widths run independently and the output estimates are fused by a set of weighting coefficients. The weighting coefficients are treated as the posterior probabilities of the kernel widths in effective and they are computed recursively by using the likelihood functions. Simulation results show that the proposed filter outperforms the existing single kernel learning filters and the multi-kernel learning filter with maximum mixture correntropy criterion.