A new neural network based sequence estimator in non-Gaussian noise environment

The application of neural network for sequence estimation in the presence of both impulsive noise and intersymbol interference is presented. In this estimator, a nonlinearity is embedded in the conventional steepest descent method for suppressing the impulse noise during the iteration and thus a dual nonlinear steepest descent algorithm is developed for estimating the symbol sequence. This algorithm can be implemented by a recurrent correlation neural network with highly parallel processing. To further improve the performance, a decision feedback technique is developed. It is shown in computer simulations that the new estimator outperforms the linear Viterbi algorithm particularly when there is impulse noise.