Unsupervised learning for channel equalization using a neural network data receiver

This paper addresses the problem of channel equalization in digital communications, whereby a transmitted message, distorted by a linear channel, can be recovered at the receiver. A radial basis function (RBF) network is used as a blind equalizer, that is, a system that removes the intersymbol interference caused by the channel when a reference sequence is not available at the receiver. Various algorithms have been proposed for blind equalization, based on a variety of criteria, but they are either too slow or too computationally intensive for real time operation. In a recent work, a recurrent neural network (RNN) has been proposed as a blind equalizer with encouraging results, featuring small size and a high convergence speed. The RBF network described in this work has several desirable characteristics for blind equalization, such as a reduced sensitivity to weight initialization, and a structure where unsupervised learning appears naturally. Additionally the solutions obtained with the RBF network can be qualitatively checked and improved with heuristic reasoning