The University of British Columbia

Many equalizers based on neural networks have been proposed across the literature. Unfortunately, the complexity and the slow convergence still have to be overcome for neural equalizers to be implemented in real time. This paper presents neural equalizers suitable for multi-level QAM constellations, and trained using complex Extended Kalman and RLS algorithms, which makes them more robust against severely dispersive channels, like broadband outdoor or indoor mobile communication channels. The activation function is optimized to obtain good performance for large size signal constellations (i.e. up to 256-QAM). Extensive simulations show the benefits and limitations of these neural equalizers over traditional decision-feedback equalizers.