Discriminative learning strategy for efficient neural decision feedback equalizers

Neural networks have been successfully applied to the equalization of digital communication channels. Decision feedback is a common technique to enhance the performance of linear equalizers. The two concepts can be effectively merged, generating a wide set of possible architectures. In this work several decision-feedback (DF) neural equalizers (DFNE) are compared with classical DF equalizers and Viterbi demodulators. In particular, it is shown that the choice of a cost functional based on Discriminative Learning (DL), coupled with a fast training paradigm, can provide neural equalizers that outperform the standard DF equalizer (DFE) at a practical signal to noise ratio (SNR). Resulting architectures are competitive with the Viterbi solution as for cost-performance aspects.