Abstract
High-altitude balloons (HABs) equipped with mobile edge computing (MEC) capabilities provide a flexible solution for on-the-fly processing in ground-level networks. However, the dynamic allocation of these resources is computationally expensive and traditionally requires centralized data collection, leading to significant latency and energy overhead. This paper presents a Support Vector Machine (SVM)-based Federated Learning (FL) framework that enables HABs to collaboratively predict optimal user-to-balloon associations. By incorporating a structure matrix to model inter-HAB relationships and sharing only model weights, the proposed approach reduces total energy and time consumption by up to 16.1% compared to centralized methods, while maintaining near-optimal prediction accuracy.
Introduction
The deployment of tethered High-Altitude Balloons (HABs) in the stratosphere offers a promising infrastructure for extending Mobile Edge Computing (MEC) to ground users, such as autonomous vehicles and mobile devices. These "flying base stations" can process heavy computational tasks offloaded from the ground. However, the dynamic nature of user requests—varying in size and location—presents a complex optimization challenge.
To minimize network-wide latency and energy consumption, three critical decisions must be made in real-time:
- User Association: Determining which ground user connects to which HAB.
- Service Sequence: Establishing the order in which a balloon processes tasks from its connected users.
- Task Allocation: Deciding the fraction of a task to be processed locally versus offloaded to the HAB.
Traditional centralized optimization requires transmitting raw user data (e.g., task sizes, locations) to a central controller, creating a communication bottleneck and increasing energy drain.
Proposed Methodology
To address these challenges, the authors propose a proactive, decentralized approach using SVM-based Federated Learning.
SVM-based Federated Learning
Instead of centralizing raw data, each HAB trains a local SVM model to predict future user associations based on historical data. The balloons then share only their trained model weights. A key innovation is the introduction of a structure matrix ($\Omega_m$) during training. This matrix mathematically captures the correlations and differences between the association schemes of different HABs, allowing the network to build a more accurate global model than standard federated averaging.
Problem Decoupling
The joint optimization of association, sequence, and allocation is a difficult mixed-integer programming problem. The authors simplify this by:
- Using the FL model to predict the integer-based user association.
- Proving that the optimal service sequence follows a simple rule: serving users in ascending order of their required processing time.
- Solving the remaining task allocation as a standard convex optimization problem, which each HAB can perform independently and rapidly.
Performance Evaluation
The framework was evaluated using real-world city cellular traffic data from the OMNILab at Shanghai Jiao Tong University, serving as a proxy for computational task sizes.
Key Results
- Efficiency Gains: The SVM-based FL algorithm reduced the weighted sum of energy and time consumption by 16.1% compared to centralized SVM and 26.7% compared to local learning.
- High Accuracy: The federated approach achieved a 19.4% gain in prediction accuracy over local-only models. It reached within 3% of the accuracy of a fully centralized model that had access to all raw data, but without the associated communication overhead.
- Convergence Trade-offs: While effective, the exchange of parameters for the relationship matrix increased the number of communication rounds, resulting in a 16.7% slower convergence speed compared to local learning.
Conclusion
This research demonstrates that proactive machine learning can effectively replace slow, iterative optimization in HAB-MEC networks. By leveraging federated learning with a structure matrix, the system achieves near-optimal resource allocation with significantly reduced energy and latency. Future work may explore adapting this framework for highly mobile Unmanned Aerial Vehicles (UAVs) or integrating deep reinforcement learning to improve convergence speeds in even more dynamic environments.
Summary updated