A Machine Learning Approach for Simultaneous Demapping of QAM and APSK Constellations

TL;DR

The paper tackles the challenge of flexible yet accurate demapping by introducing a probabilistic, multi-constellation DNN framework. It employs mapping independence and hierarchical representations to enable a single neural demapper to handle both $4^n$-QAM and APSK constellations, reducing output requirements without sacrificing BER. Empirical results show the model approaches the AWGN optimal demodulation bound across tested constellations and demonstrates effective transfer and sharing of learned representations. This approach promises practical, scalable neural receivers capable of supporting families of constellations with reduced training and complexity.

Abstract

As telecommunication systems evolve to meet increasing demands, integrating deep neural networks (DNNs) has shown promise in enhancing performance. However, the trade-off between accuracy and flexibility remains challenging when replacing traditional receivers with DNNs. This paper introduces a novel probabilistic framework that allows a single DNN demapper to demap multiple QAM and APSK constellations simultaneously. We also demonstrate that our framework allows exploiting hierarchical relationships in families of constellations. The consequence is that we need fewer neural network outputs to encode the same function without an increase in Bit Error Rate (BER). Our simulation results confirm that our approach approaches the optimal demodulation error bound under an Additive White Gaussian Noise (AWGN) channel for multiple constellations. Thereby, we address multiple important issues in making DNNs flexible enough for practical use as receivers.

Figures (7)

• Figure 1: An example of QPSK and 16QAM overlaid. The representation is determined by counting from left to right and bottom to top in Gray code. Afterwards, interleave the in-phase and quadrature components to get a hierarchy. The representation values match the position in the representation by colour.
• Figure 2: A QPSK (blue) and offset 8-PSK (red) constellation overlaid. The latter constellation hierarchically relates to the former, as by adding one more bit (red) to the two bits (blue), we can split the quadrants in two and move from QPSK to offset $8$-PSK.
• Figure 3: The system model used to evaluate our approach. For symbol demapping, we use a hard-decision demapper as a baseline and compare it to our DNN.
• Figure 4: The BER of a DNN trained on $256$-QAM compared to the optimal hard decision bound in our setup. The (trained) and (transferred) tags imply whether the model saw the constellation during training or not.
• Figure 5: BER of a DNN trained on all included $4^n$-QAM and APSK constellations using our framework. Shown are the first half of the constellations trained.