| Summary: | In this paper, we investigate the problem of scheduling transmissions for spatially scattered nodes that contribute to a collaborative federated learning (FL) algorithm via wireless links provided by a drone. In the considered system, the drone acts as an orchestrator, coordinating the transmissions and the learning schedule within a predefined deadline. The actual schedule is reflected in a planned path: as the drone traverses it, it controls the distance and thereby the data rate to each node. Hence, the model is structured such that the drone orchestrator uses the path (trajectory) as its only tool to achieve fairness in terms of learning <italic>staleness</italic>, which reflects the learning time discrepancy among the nodes. Using the number of learning epochs performed at each learner as a performance indicator, we combine the average number of epochs computed and staleness into a balanced optimization criterion that is agnostic to the underlying FL implementation. We consider two methods for solving the complex trajectory planning optimization problem for static nodes: (1) successive convex programming (SCP) and (2) deep reinforcement learning (RL). Considering the proposed criterion, both methods are compared in three specific scenarios with few nodes. The results show that drone-orchestrated FL outperforms an immobile deployment by providing improvements in the range of 57% to 87.7%. Additionally, RL-guided trajectories are generally superior to SCP provided ones for complex node arrangements.
|
|---|