QPSK receiver based on recurrent neural networks

This paper describes the performance of a complete receiver based on Recurrent Neural Networks (RNNs). The receiver uses a cascade of two RNNs : one for equalization and one for demodulation. The simulations have been done by considering two cases : a single AWGN channel and a two paths model with the direct and the indirect rays interfering at the receiver. Two different types of digital filters have been used to simulate the effect of the transmitter-channel-receiver chain. QPSK digital modulation has been analyzed. The neural system performs better than acoherent receiver (composed of a coherent demodulator and a Transversal equalizer) when the channel bandwidth is narrow, i.e. when intersymbol interference increases. This is due to the fact that the Neural Receiver dynamically changes its Decision Regions for each symbol. This dynamic behaviour has been investigated.