Work

Enhancing throughput and Fairness in Dense WLANs Using Graph Neural Network-driven Multi-agent Reinforcement Learning

Year: 2026
Type: article
Abstract: Wireless local area networks (WLANs) must be optimized for reliable and efficient communication in the presence of contention and interference from dense wireless deployments.We propose a novel cooperative graphbased multi-agent reinforcement learning framework for intelligent channel selection in IEEE 802.11 networks.The framework combines cooperative multi-agent reinforcement learning with graph neural networks (GNNs) to form a model that captures interference-aware interactions between access points (APs) for adaptive optimization in dynamic WLAN environments.The authors explicitly model all transmission parameters for each AP as independent agent systems that choose their transmissions based on local observations that use graph representations of interference.GNNs provide a model of dynamic interference between nearby APs, allowing implicit coordination through local observations of neighbours rather than through explicit messaging.Proximal Policy Optimization (PPO) is utilized as the learning rule to provide stable and scalable learning in multi-agent scenarios.The overall framework is implemented under a centralized training/decentralized execution (CTDE) paradigm, where global network information is available only during training, while each access point executes its policy independently using local observations during deployment.Our evaluation using a realistic WLAN emulator showed that our framework produced an increase in throughput stability and a decrease in interference and collision rates relative to the randompolicy and MAPPO-without-GNN baselines implemented in this study.A matched-setting comparison against QMIX, a representative value-decomposition MARL baseline, is also reported, showing comparable throughput with a consistent latency advantage for the proposed framework.Our empirical results demonstrate a 55.5% throughput improvement of the full GNN-MAPPO model over the MAPPO variant without graph-based processing under the same training and execution settings.This improvement is accompanied by a 33.0%increase in average latency, while fairness improves by 21.7% and the collision rate is reduced by 52.4%.These results demonstrate that cooperative, graph-based reinforcement learning provides an effective mechanism for optimizing WLANs operating under complex, interference-driven conditions, while highlighting an explicit throughput-latency trade-off.
Language: English
Open Access status: hybrid
OpenAlex ID: W7234764415
FWCI:
Cites: 34
Cited by:
Related to: 10