Deep Neural Network Based Reduced-Complexity Detector for Grassmann Constellation

In this paper, we propose a reduced-complexity detector for non-coherent communications with the Grassmann constellation, which enables joint channel and data estimation. Here, we employ deep learning techniques with the aim of reducing computational complexity while maintaining nearly the same channel estimation accuracy as the conventional detector. The conventional maximum likelihood detection is a discrete optimization problem, with complexity increasing exponentially with respect to the transmission rate. In our approach, the Grassmann constellation is detected using a neural network model trained with random signal-to-noise ratios and the corresponding received signals as inputs, and the detected codeword is used for channel estimation. Our simulations demonstrate that the channel estimation accuracy can be maintained with lower complexity compared to the conventional detector, at the cost of a slightly degraded symbol error rate performance. It was also found that the detection complexity can be reduced when the number of receive antennas exceeds four, which is practically relevant.