ABSTRACT Federated reinforcement learning for multi‐UAV trajectory optimisation often suffers from slow convergence and high communication overhead. This paper proposes Geo‐FedKD, a geometry‐aware framework that integrates elliptical trajectory priors with proxy action‐output knowledge distillation. Exploiting the observation that dual‐connectivity requirements naturally favour elliptical UAV trajectories with foci at the base station and the user centroid, Geo‐FedKD guides policy exploration towards feasible regions instead of relying solely on trial‐and‐error learning. Exchanging compact proxy action outputs rather than high‐dimensional model parameters reduces communication cost by 93%. Simulation results show that Geo‐FedKD achieves a dual‐connectivity ratio of 73%, improves connectivity by 15% over existing baselines and converges in 15 training rounds compared with 25–30 rounds for conventional federated reinforcement‐learning methods. These results demonstrate significant improvements in learning efficiency and network performance for UAV‐assisted wireless networks.
Karimi et al. (Thu,) studied this question.