Federated Reinforcement Learning for Distributed MAC Optimization in IEEE 802.11bn Networks

IEEE 802.11bn (Wi-Fi 8) introduces Multi-AP Coordination (MAPC) to meet ultra-reliable low-latency communication (URLLC) demands in dense wireless deployments. While centralized MAC-layer scheduling improves coordination, it introduces overhead, privacy risks, and scalability challenges. In this paper, propose a decentralized federated reinforcement learning (FRL) framework for MAC scheduling across distributed access points (APs). Each AP independently learns optimal transmission policies using deep Q-learning, while periodically synchronizing model updates through a lightweight federated server. This approach preserves local traffic privacy and reduces control latency, without sacrificing performance or adaptability. The system dynamically adjusts to varying interference, heterogeneous traffic loads, and dynamic topology changes in real-time. Evaluate the proposed FRL-based scheduler under diverse STA densities, mobility scenarios, and stochastic channel conditions using extensive custom simulations. Results show that our model achieves up to 29% lower latency, 22% higher fairness, and 17% reduction in signaling overhead compared to centralized RL and OFDMA-based MAC methods. The proposed solution offers a scalable, privacy-preserving, and resilient path toward intelligent MAC optimization in next-generation Wi-Fi networks, paving the way for mission-critical industrial and latency-sensitive applications.