Federated Learning in UAV-Enhanced Networks: Joint Coverage and Convergence Time Optimization

Abstract:
Federated learning (FL) has recently emerged as a promising learning paradigm, in which several devices participate in training a shared model without transferring their local data. Besides preserving privacy, such an approach reduces the communications overhead. Therefore, it is a promising method for learning in wireless communication networks with scarce communication resources, especially for devices such as UAVs that have limited access to power resources. Nevertheless, implementing FL in UAV-enabled networks is challenging as the network architecture and management play crucial roles in the convergence of the FL algorithm. In particular, the conventional UAV placement methods that maximize coverage often result in prohibitive FL delays. In addition, the uncertainty and lack of a priori information about, for example, the active sensors’ density and channel quality, exacerbate the problem. In this paper, we first analyze the statistical characteristics of a UAVenabled wireless sensor network with energy harvesting. We then apply multi-objective multi-armed bandit theory to maximize the network coverage while minimizing the FL delay. Numerical results show the effectiveness of our approach.