Federated Multi-Agent Deep Reinforcement Learning for Joint Channel Selection and Power Control in Cognitive Radio Networks

Cognitive radio networks (CRNs) face significant challenges in dynamic spectrum access due to the complex interactions among multiple secondary users, sparse reward signals, and poor cross-domain generalization. Existing approaches, ranging from traditional optimization to single-agent deep reinforcement learning (DRL), struggle to balance spectral efficiency, collision avoidance, and adaptability in heterogeneous wireless environments. In this paper, we propose FedMA-DRL, a federated multi-agent deep reinforcement learning framework that integrates centralized training with decentralized execution (CTDE), graph neural network (GNN)-augmented Q-value prediction, age-aware federated aggregation (FedAge), and attention-based domain adaptation for joint channel selection and power control in CRNs. The GNN module captures topological relationships among secondary users through attention-weighted message passing on the interference graph, while the FedAge strategy enables privacy-preserving knowledge sharing with staleness-aware weighting. Extensive experiments on a CRN testbed with 10 PU channels and 15 heterogeneous SUs demonstrate that FedMA-DRL achieves 14.87 Mbps SU throughput, 0.038 collision probability, 4.35 bits/Joule energy efficiency, and 6.23 bits/s/Hz spectrum efficiency, outperforming existing methods including R2D2 and C-DRL. Ablation studies and cross-domain evaluations further confirm the effectiveness of each proposed component.