A robust RBF neural net Bayesian estimator for channel equalization

Abstract

The characteristic (transfer function) of a dispersive M-ary channel equalizer designed through a Bayesian estimator, a performance indicator that is not trivial to obtain for M>2 due to intersymbol interference (ISI), is investigated. A set of curves is obtained and interpreted. Implementation through a radial basis function neural network is considered. It is shown that because network centers endure updating with different rates, the equalizer characteristic errates off the optimum, causing thus the symbol error rate and/or the training time to increase. A solution, based on incorporating an underlying symmetry in the channel response levels into the updating algorithm, brings the characteristic uniformly closer to the optimal one. It is also provided a strategy endowing the network with self-initializing. Simulation results are presented for a channel with sufficient ISI strength.