Complex discriminative learning Bayesian neural equalizer

Traditional equalizers try to invert the global, linear or nonlinear, channel response. However, in digital links, where transmitted symbols belong to a discrete alphabet, the complete channel inversion is neither required, nor desirable. Actually, symbol demodulation can be recast as a classification problem in the received symbol space. Following this approach, in recent years, neural networks have been used as demodulators. In this paper, we propose a neural architecture, which resorts to a somewhat intermediate approach between the channel inversion and the Bayesian classification. A complex-valued discriminative learning, which attempts to minimize the error risk, is applied to a nonlinear decision-feedback network, resulting in fast convergence and low degree of complexity.