Non-linear equalizers that estimate error rates during reception

Neural Networks can be used to estimate the a posteriori probabilities of the transmitted symbols in digital communication systems. In this paper we apply this property to the on-line estimation of the bit error rate (BER) in the receiver, without using any reference signal. We discuss two different approaches to BER estimation: (1) computing the a posteriori symbol probabilities from estimates of the conditional distributions of the received data, and (2) estimating probabilities by gradient minimization of a special type of cost functions. We show that Importance Sampling (IS) techniques can be combined with the first approach to reduce drastically the variance of the probability estimates. Finally, we analyze the effect of channel variations during transmission.