A neural predictor for blind equalization of digital communication systems: is it plausible?

In digital channel equalization, self-learning techniques are used in the cases where a training period is not available. Considering the transmitted sequence as composed of independent random variables, the equalization task can be done by means of prediction. In this work we propose artificial neural networks (ANN), instead of a linear prediction device, in order to obtain a better performance and analyse its performance and applicability. Linear and nonlinear prediction concepts are revisited and a new self-organized algorithm is proposed to update the first layer in the nonlinear predictor whose aim is to avoid local minimum points in the applied cost function. The second layer is updated by using a classical supervised algorithm based on prediction error. Simulation results are presented which illustrate the performance of this technique.