Deep Autoencoder-Based Constellation Design in Multiple Access Channels

TL;DR

The paper addresses constellation design for finite-alphabet multi-user channels, specifically a two-user Gaussian MAC, where inter-user interference degrades performance of conventional constellations. It proposes a deep autoencoder (DAE) framework that jointly optimizes transmitters and the receiver in an end-to-end manner to mitigate interference, targeting the constellation-constrained CC-sum-rate, defined as $I(\sqrt{2-\alpha}\, x_1, \sqrt{\alpha}\, x_2; y)$. Across two- and three-user scenarios and varying $\alpha$, the DAE-designed constellations consistently match or exceed analytically derived designs, and adapt to different system parameters more robustly than MD-based or rotation-based methods. A key limitation is that a separate model must be trained for each parameter set $(K,\alpha,k_1,k_2)$, pointing to future work on generalizable training that outputs optimal constellations given system inputs. Overall, the work demonstrates that learning-based constellation design can achieve near-optimal performance in interference-limited MACs and extend to higher-order/multi-user settings with strong adaptability.

Abstract

In multiple access channels (MAC), multiple users share a transmission medium to communicate with a common receiver. Traditional constellations like quadrature amplitude modulation are optimized for point-to-point systems and lack mechanisms to mitigate inter-user interference, leading to suboptimal performance in MAC environments. To address this, we propose a novel framework for constellation design in MAC that employs deep autoencoder (DAE)-based communication systems. This approach intelligently creates flexible constellations aware of inter-user interference, reducing symbol error rate and enhancing the constellation-constrained sum capacity of the channel. Comparisons against analytically derived constellations demonstrate that DAE-designed constellations consistently perform best or equal to the best across various system parameters. Furthermore, we apply the DAE to scenarios where no analytical solutions have been developed, such as with more than two users, demonstrating the adaptability of the model.

Figures (10)

• Figure 1: Two-user MAC model with the assumed additive Gaussian noise.
• Figure 2: Model describing DAE layers. FC denotes a fully connected layer, with the activation function and the corresponding number of neurons indicated in brackets.
• Figure 3: DAE-designed constellations with (top) $k_1 = k_2 = 2$ and (bottom) $k_1 = 2$, $k_2 = 1$, with $\alpha=1$.
• Figure 4: CC-achievable sum rate for $\alpha = 1$, comparing the DAE-designed constellation with various approaches for (top) $k_1 = k_2 = 2$ and (bottom) $k_1 = 2$, $k_2 = 1$.
• Figure 5: Symbol error rate comparison for different constellations with $\alpha = 1$ and (top) $k_1 = k_2 = 2$, (bottom) $k_1 = 2$, $k_2 = 1$.