Bayesian Equalizers and RBF Networks

daptive equalization has been an active area of research for many years. Even in 1985 there were a A plethora of available solutions [ 17 whose properties were well understood. Many of the techniques are firmly based on linear adaptive filter algorithms and exhibit the same much-lauded ‘learning’ property as neural networks. Alternatively, maximum likelihood strategies, which are usually based on the Viterbi algorithm (VA) [2] and its variants, have long been understood to provide the best performance of all equalization techniques. Why then is it worth considering the application of artificial neural networks (NN) to this problem? The answer comes in two parts. The first is one that has always driven science, and that is curiosity. How well will a neural network perform in this benchmark problem and how will it fare when compared with standard solutions? Initial work [ 3 ] demonstrated that multilayer perceptron (MLP) equalizers were superior to conventional transversal and decision feedback equalizers in terms of the usual measure of equalizer performance, which is bit error rate (BER). On the other hand, the work also highlighted several of the difficulties that are well known in the wider application of MLP’ s. These are the extreme length of training times; the indeterminate nature of the training times; the lack of a methodology for architecture selection. These problems are largely unsolved and severely restrict the practical application of MLP’s in this area.