Graph-Localized Offline Federated Multi-Agent Reinforcement Learning for Wireless Networks

Wireless networks motivate offline multi-agent reinforcement learning (MARL) as online exploration degrades service, and operator logs are rarely poolable across deployments. Existing offline MARL relies on global coverage that scales exponentially in the number of agents. We exploit the interaction graph induced by interference and contention to replace global coverage with $\kappa$-hop neighborhood coverage: a client contributes to agent~$i$'s local estimator only if it observes $\mathcal{N}_i^\kappa$. We prove a localized offline policy guarantee whose error decomposes into a locality bias decaying exponentially in $\kappa$, a near-neighbor shift bias, and a federated estimation term shrinking with the pooled observability-valid sample size. Our algorithm, F-GLOFF, matches a centralized raw-pooled oracle on multi-AP user association without sharing transitions and outperforms the engineered baselines by 12%.