Adaptive implementation of minimum-error-rate equalizers via backpropagation neural networks

The authors introduce a minimum-error-rate equalizer (MERE) operating on a finite dimensional observation vector at every instant, and discuss its adaptive implementation via backpropagation neural networks (BPNNs). It is shown that even in the most general case of a nonlinear channel model with colored noise background, the successfully trained BPNN still can exactly implement the MERE, if the network is complex enough so that arbitrarily continuous mappings can be formed. Extension to the case of a decision feedback equalizer is also considered. Computer simulations are presented to illustrate the performance advantages of the present MERE