Enhancing multiuser scheduling in massive MIMO mobile channels

Massive Multiple-Input Multiple-Output (MIMO) is a key enabler of 5G and beyond mobile networks, significantly improving spectral efficiency through multiuser beamforming. However, in massive MIMO systems, the multiuser scheduling problem, selecting which users to serve concurrently on the same time-frequency resources, remains a critical challenge. Due to potential channel correlation among users, suboptimal multiuser scheduling can lead to inter-symbol interference and throughput degradation. Additionally, the scheduler must balance the achieved spectral efficiency with user fairness. While the Optimal Proportional Fair (Opt-PF) scheduler seeks to achieve this balance, applying it to the massive MIMO scheduling problem leads to an NP-hard optimization problem. Although existing approximation algorithms can reduce the computational complexity of the Opt-PF multiuser scheduler, they often fail to provide adequate fairness or adapt to fast varying channels, making them impractical for real-world deployment. As an alternative, Machine Learning (ML)-based methods, particularly Deep Reinforcement Learning (DRL) models, have shown promise in addressing this problem. To further foster innovation in this area, the International Telecommunication Union (ITU) AI/ML in 5G Challenge hosted a competition focused on enhancing the performance of a DRL-based multiuser scheduler. The provided baseline scheduler employed a user-grouping algorithm to cluster users with low channel correlation and a Soft Actor-Critic (SAC) DRL framework for user selection. This paper presents the winning solution to the ITU competition, which proposes two approaches to enhance the performance of the baseline scheduler.