Abstract:

The performance of the multi-kernel learning algorithm possesses higher degree of freedom and uses more features of data, which outperforms the mono-kernel methods in online nonlinear adaptive filtering applications. The multi-kernel affine projection algorithm with the same dictionary is presented, which is an umbrella of multiple kernel learning and affine projection methods. As the particular case of multi-kernel affine projection building almost distinct dictionary corresponding to different Gaussian kernel bandwidths based on coherence sparsification critera, multi-kernel normalized leastmeansquare with adaptive  l₁ norm regularization for update of dictionary is also proposed to overcome the drawback that the obsolete kernel functions cannot be discarded in non-stationary environment. Simulation results show the effectiveness of the proposed methods.