Training RBFNN with reglarized correntropy criterion and its application to foreign exchange rate forecasting

In the paper, a regularized correntropy criterion (RCC) for radial basis function neural network (RBFNN) is proposed. In RCC, the Gaussian kernel function is used to replace the Eculidean norm of the sum-squared-error (SSE) criterion. Replacing SSE by RCC can improve the anti-noise ability of RBFNN. Moreover, the optimal weights and the optimal bias terms can be iteratively obtained by the half-quadratic optimization technique. The effectiveness of the proposed method is validated on the foreign exchange rate time series. In comparison with the RBFNN trained with the SSE criterion, the proposed method demonstrates better generalization ability.