Nonlinear adaptive filtering using annealed neural networks

A technique for nonlinear adaptive filtering which combines annealed neural networks (ANNs) and conventional recursive-least-squares/fast-transversal-filter (RLS/FTF) adaptive filtering is presented. The technique can efficiently handle large-order nonlinearities with or without memory. The nonlinear channel is divided into a nonlinearity followed by a dispersive linear system. ANN is applied to obtain the coefficients of the nonlinearity, and RLS/FTF is applied to determine the weights of the dispersive linear systems. The proposed technique has better convergence properties than gradient-descent-based methods, such as those based on steepest-descent or Newton´s method. The proposed technique is used to solve for the true minimum mean-squared-error estimate, as opposed to solving for an estimate which satisfies suboptimal equations. A simulation example is given which shows the superior performance of the technique compared to that of ordinary RLS/FTF and gradient-descent