Efficient Training Scheme for Neural Network Based 4K-QAM Soft Demapper

Abstract

The practicality of the densely packed and spectrally efficient 4096-Quadrature Amplitude Modulation (4K-QAM) is obstructed by the ultra-high computational complexity of its soft demapper, which is essential for generating the soft outputs required by channel decoders. In this paper, we propose an efficient training scheme to build a highly effective neural network based 4K-QAM soft demapper that can offer significantly lower computational complexity. The results demonstrate that this alternative demapper can achieve comparable decoding performance in coded 4K-QAM systems, while reducing computational complexity by up to 20% compared with the well-known low-complexity max-logarithm of maximum a posteriori (log-MAP) demapper.