Interference-Aware Super-Constellation Design for NOMA

TL;DR

This work addresses interference in downlink NOMA with finite-alphabet inputs by designing interference-aware super-constellations with an SIC-free autoencoder. By jointly training a transmitter encoder and two receivers, the approach yields constellations that maximize minimum distances under varying channel gains, using an adaptive, SIC-free loss and power-shaping sub-network. Empirical results demonstrate significant BER improvements over traditional SIC-based designs and robust performance across fixed and randomly varying channels, with enhanced fairness between users. The findings suggest that interference-aware end-to-end learning can enable practical NOMA deployments with reduced decoding complexity and better performance in resource-constrained devices.

Abstract

Non-orthogonal multiple access (NOMA) has gained significant attention as a potential next-generation multiple access technique. However, its implementation with finite-alphabet inputs faces challenges. Particularly, due to inter-user interference, superimposed constellations may have overlapping symbols leading to high bit error rates when successive interference cancellation (SIC) is applied. To tackle the issue, this paper employs autoencoders to design interference-aware super-constellations. Unlike conventional methods where superimposed constellation may have overlapping symbols, the proposed autoencoder-based NOMA (AE-NOMA) is trained to design super-constellations with distinguishable symbols at receivers, regardless of channel gains. The proposed architecture removes the need for SIC, allowing maximum likelihood-based approaches to be used instead. The paper presents the conceptual architecture, loss functions, and training strategies for AE-NOMA. Various test results are provided to demonstrate the effectiveness of interference-aware constellations in improving the bit error rate, indicating the adaptability of AE-NOMA to different channel scenarios and its promising potential for implementing NOMA systems

Figures (7)

• Figure 1: The architecture of the implemented AE-NOMA for two-users. Each FCNN[$\ell$] represents a fully connected layer with $\ell$ nodes. The main structure which consists five hidden layers each with 32 neurons is replicated at $\rm Tx$, $\rm Rx1$ and $\rm Rx2$. The $\rm Tx$, however, has a side block (Sub-Network 2) which is used to adjust power allocation for each symbol.
• Figure 2: BERs versus SNR$_1$ for $h_1=1$, $h_2=2$. and $w =10$ for different number of epochs.
• Figure 3: AE-NOMA constellations at UE 1 and UE 2 when $h_2=2h_1 = 2$ were fixed during training and test (Test Case 1), and the number of epochs is 150K. The '+' signs represent noiseless super-constellation symbols ($\rm Tx$ output), whereas the dots represent noisy versions. Each color represents two bits.
• Figure 4: BERs versus SNR$_1$ for $h_2=2h_1=2$ for AE-NOMA compared to the theoretical BER for when both users use a QPSK constellation, and BER of a 16-QAM constellation.
• Figure 5: Constellations at both users for $h_2=2h_1=2$. Here, during training $h_2\sim {\rm Unif}[1,3]$ (Test Case 2).