Scalable Deep Reinforcement Learning-Based User Pairing for Cell-Free MIMO Network

Wireless networks continuously evolve to support higher data rates and accommodate a growing number of wireless devices. However, conventional scheduling methods either incur high computational complexity and excessive latency, or suffer from significant performance degradation under imperfect system conditions. In this work, we employ a deep reinforcement learning (DRL) framework to address the NP-hard user pairing problem in cell-free multiple-input multiple-output (MIMO) networks. The proposed algorithm operates across multiple resource-block-groups (RBGs) in parallel, leveraging a pointer network that observes RBG-specific, transmit-power-scaled channel state information (CSI) to select users of this RBG. Then, RBG-specific uplink throughput serves as a feedback to optimize the user pairing policy without requiring labeled data. Simulation results show that the proposed approach is highly scalable across different deployment scenarios with varying network configurations. Meanwhile, the proposed DRL-based scheme achieves a favorable trade-off between computational complexity and throughput performance. Furthermore, DRL consistently outperforms all baselines particularly in the presence of CSI estimation errors and unknown noise, owing to its feedback-driven learning mechanism.