Abstract
This paper investigates the optimization of transmission scheduling for spatially scattered nodes in a collaborative Federated Learning (FL) framework. A drone acts as a mobile orchestrator, coordinating learning schedules and data transmissions over wireless links. By treating the drone's trajectory as the primary tool for controlling data rates, the study aims to minimize "staleness"—the discrepancy in learning progress among nodes. We introduce the Average-Anchored Staleness (AAS) metric and evaluate two path-planning methods: Successive Convex Programming (SCP) and Deep Reinforcement Learning (RL). Results demonstrate that drone-orchestrated FL improves performance by 57% to 87.7% compared to static deployments, with RL showing superior adaptability in complex environments.
Introduction: The Challenge of Distributed Learning
Federated Learning (FL) allows multiple devices to train a global machine learning model collaboratively without sharing raw data. However, deploying FL in the field faces significant hurdles:
- Spatial Distribution: Nodes are often located in remote or infrastructure-poor areas.
- System Heterogeneity: Variations in processing power and data volume create "stragglers," slowing down the entire network.
To address these, the authors propose using a Drone Small Cell (DSC) as a mobile orchestrator. Unlike fixed base stations, a drone can dynamically adjust its proximity to specific nodes to optimize communication links.
Methodology: Optimizing the Flight Path
The core challenge is path planning: determining how the drone should move to ensure all nodes contribute effectively within a deadline. The authors define a new optimization criterion, Average-Anchored Staleness (AAS), which balances the total number of learning epochs against the maximum difference in progress between any two nodes. This metric ensures that the drone does not simply favor the fastest nodes, which would lead to a biased global model.
Two approaches were compared:
- Successive Convex Programming (SCP): A deterministic method that approximates the non-convex trajectory problem as a series of convex problems. It is computationally efficient and serves as a reliable baseline.
- Deep Reinforcement Learning (RL): A Double Q-learning agent operating on a hexagonal lattice. This discretization allows the drone to choose from seven actions at each timestep. A "secondary experience buffer" was used to store high-performing paths, accelerating convergence. The hexagonal grid was chosen over a square grid for its superior packing and more natural movement options.
Key Insights and Results
The drone's trajectory serves as a fairness mechanism. By lingering near nodes with weaker processors or poor signals, the drone reduces their communication time, allowing them more time for local computation. The study utilized parameters inspired by the YOLOv3 object detection model to simulate realistic data loads.
Performance Gains
- Superiority over Static Setups: Drone-orchestrated FL achieved AAS improvements of 57% to 87.7% over stationary deployments.
- RL vs. SCP: In complex "Hidden" node scenarios, RL outperformed SCP (AAS of 7.8 vs 7.1). RL agents proved more adept at finding optimal communication "sweet spots" in non-linear signal environments, discovering non-obvious paths that the deterministic SCP missed.
- Flight Dynamics: RL-guided drones typically operated at maximum speed to reach optimal positions quickly, maximizing the time spent in high-throughput zones.
Conclusion and Future Directions
Using a drone as a mobile FL orchestrator significantly mitigates the straggler problem by utilizing mobility as a resource for fairness. While the results are promising, future research must address drone energy constraints, real-world signal stochasticity, and the scalability of RL agents to handle hundreds of nodes. The current model assumes mean fading coefficients, which may not capture the instantaneous signal drops found in real-world suburban environments.
Summary updated