ALGORITMO RECURSIVO BASEADO EM UMA FUNÇÃO NÃO LINEAR DO ERRO.

Abstract

Many of the adaptive filters are based on the Mean Squared Error (Mean Square Error - MSE). The development of these filters guarantees us to recover only second-order information of the signals to be filtered, ie not can fully recover information from Gaussian signals. However, the natural or artificial signals are not necessarily Gaussian. In this way, the use of high order statistics as a way of extracting more information of signals, has been shown to be of great value in adaptive systems [7] [8] [9]. In this work, we present the development of an adaptive algorithm based on nonlinear functions inspired by the deduction of the Recursive Lest algorithm Square (RLS) [1]. Such development is based on the use of high order to obtain more information on the signals involved in the process, with the goal of improving the performance of an adaptive filter. We will call this new nonlinear recursive algorithm - RNL. We deduce equations, based on a nonlinear function, to obtain convergence criteria. We also study covariance of the steady-state weight vector and we determine equations that calculate the mismatch and the learning time of the adaptive process of the RNL algorithm. We present the non - linear recursive algorithm, which uses the function "n = MP j = 1 nP i = 1nnnii [ei] 2jo, where M and n are positive integers. Were made simulations with this algorithm to validate the presented theory and study the convergence behavior of the RNL algorithm. The result showed that the RNL algorithm has a rapid convergence for the same mismatch when compared with the RLS algorithm.