Using recurrent neural networks for blind equalization of linear and nonlinear communications channels

A recurrent neural network (RNN) equalizer for blind equalization of linear and nonlinear communication channels is proposed. RNNs have the ability to learn dynamical mappings of arbitrary complexity and therefore present a natural choice for implementing equalizers for communication channels. In several cases the nonlinear nature of a communication channel is too severe to ignore, and at the same time no nonlinear channel model can account sufficiently for the nonlinearities that are inherently present in the channel. In those cases a neural network equalizer is preferable over a conventional one. The real-time recurrent learning (RTRL) algorithm is used to train an RNN, and its performance is compared with that of a conventional equalizer based on the constant-modulus algorithm