Orthogonal Maximum Correntropy Learning

One-hidden-layer networks are very efficient to solve structure-unknown system identification/modeling problems. But it is difficult and often redundant to choose a basis without prior knowledge. We present in this paper a new algorithm named orthogonal maximum correntropy (OMC), with which a robust and parsimonious system-identification/model-selection process is reached. Moreover, the proposed method is capable of suppressing non-Gaussian, especially heavy-tailed noises that may be encountered in the datasets, benefiting from the maximum correntropy criterion. Experiments on some common models such as polynomial NARx, RBF networks, and wavelet networks, demonstrate the effectiveness of the proposed method.