Asynchronous federated reinforcement learning for scalable load balancing in software-defined networks
Apratim K Jha, Anand Mahendran, Ibrahim Ghafir, Mohamed Hamada · Discover Computing · 2026
Abstract Software-Defined Networks face increasing load-balancing demands under heterogeneous and time-varying traffic. This work proposes an asynchronous federated reinforcement learning framework with a hierarchical controller topology and selective parameter sharing tailored for SDN. Using PPO with asynchronous aggregation, the approach reduces end-to-end latency and communication overhead versus centralized and synchronous baselines, and maintains high throughput across most traffic regimes. Performance under medium and very high traffic shows degradation in load-imbalance for certain settings, indicating robustness limitations that motivate adaptive staleness bounds, traffic-aware participation, and reward reweighting. The hierarchy supports regional-to-local coordination and preserves data locality for multi-domain deployments.