Neural Equalizer for time varying channel Using Gauss-Newton training algorithm

Artificial neural networks (ANN) techniques have become very common as equalization solutions in several types of communication channels. These neural networks are presented in many topologies. The suitable choice of a topology for equalization purpose depends on different criteria such as: convergence rate, bit error rate (BER), computational complexity, among many others. In this paper, it is investigated the behavior of a structure similar to a decision feedback equalizer (DFE) employed to equalize time varying channels. The structure, a single recurrent perceptron, is based on a simplified recurrent neural network (RNN). The Gauss-Newton algorithm has been used to estimate the synaptic weights of the perceptron during the training and testing phases. Despite the simplicity of implementation and low computational cost, it has been shown that the proposed topology presents some good comparative performance related to more complex structures based on recurrent neural networks (RNN) and multilayer perceptrons (MLP) using Kalman filters.