CHANNEL EQUALISATION USING A SOFT BACK-PROPAGATION LEARNING ALGORITHM

José D. Martín-Guerrero - Luis Gómez-Chova - Gustavo Camps-Valls - Antonio Serrano-López - Joan Vila-Francés - Javier Calpe-Maravilla - Emilio Soria-Olivas

   This paper proposes some useful modifications to the Expanded Range Approximation (ERA) learning algorithm. In channel equalisation, it is a common practise to recover the original signal using an artificial neural network. The ERA algorithm is an alternative to the usual backpropagation algorithm that mitigates the effect of local minima during the training process. In the basis of ERA, we propose a soft-based profiling of the homotopy parameter in order to avoid local minima in the error surface more efficiently than in ERA. We use three specific membership functions that allow to control the smoothness of the learning process, yielding better results and lower computational cost.

Keywords: artificial neural networks, learning algorithms, soft computing, channel equalisation

[full-paper]