Abstract
Federated Edge Learning (FEEL) enables collaborative model training across distributed devices while preserving data privacy. However, the practical deployment of FEEL is often hindered by the limited battery life of edge devices and constrained communication bandwidth. This paper presents an energy-aware dynamic scheduling algorithm specifically designed for Over-the-Air (OtA) FEEL systems. By integrating Lyapunov optimization with a novel gradient-norm estimation technique (EST-P), the proposed framework manages long-term device battery constraints while optimizing learning performance. Experimental results on MNIST and CIFAR-10 datasets demonstrate that this approach improves model accuracy by up to 4.9% compared to traditional myopic scheduling methods, while strictly adhering to energy budgets.
Introduction
In the era of the Internet of Things (IoT), Federated Learning (FL) has emerged as a key paradigm for training machine learning models on decentralized data. In wireless "Edge" environments, Over-the-Air (OtA) computation offers a highly efficient way to aggregate local updates by leveraging the superposition property of the wireless channel. However, OtA systems face a unique "causality dilemma": the energy required for transmission depends on the $l_2$-norm of the local gradient, which is unknown until after the computation is complete.
Existing scheduling policies often overlook the energy cost of local computation or assume communication costs are independent of the data. This paper addresses these gaps by treating device scheduling as a long-term constrained optimization problem, balancing the trade-off between learning progress and power consumption.
Methodology
The proposed framework is built upon three core components:
1. Convergence Analysis
The authors derive a theoretical upper bound on the model's loss function. This analysis quantifies how the number of participating devices in each round influences the convergence rate, providing a mathematical foundation for the scheduling logic.
2. Gradient Norm Estimation (EST-P)
To overcome the causality dilemma, the algorithm employs Estimation with Past information (EST-P). Recognizing that gradients in neural networks typically exhibit temporal correlation, the system uses the gradient norm from a device's most recent active round as a proxy for its current state. This method proved more effective and less computationally expensive than complex mini-batch sampling.
3. Lyapunov Optimization
The scheduling problem is formulated using Lyapunov optimization's "drift-plus-penalty" method. By maintaining a "virtual queue" for each device's energy consumption, the algorithm converts long-term battery constraints into a series of online, greedy decisions. Devices that have recently consumed significant energy are less likely to be scheduled, ensuring long-term sustainability.
The resulting algorithm decouples signal-to-noise ratio (SNR) requirements from scheduling, achieving an efficient $O(N \log N)$ complexity for device selection.
Experimental Evaluation
The algorithm was evaluated using the MNIST (hand-written digits) and CIFAR-10 (object recognition) benchmarks under both i.i.d. and non-i.i.d. data distributions.
Key Results
- Improved Accuracy: On the CIFAR-10 dataset with non-i.i.d. data, the proposed dynamic scheduling achieved a 4.9% accuracy boost over myopic benchmarks.
- Energy Management: The system successfully respected all device energy budgets. By allowing devices to "wait" for favorable channel conditions or more informative gradients, the Lyapunov approach utilized energy more intelligently than rigid policies.
- Robustness: The framework maintained high performance even with up to 30% error in channel state information (CSI), indicating its suitability for real-world wireless environments.
Discussion and Future Work
A significant insight from this work is that in OtA FEEL, "important" updates (those with larger gradients) are more energy-intensive. This creates a fundamental tension between learning speed and battery life. While the current model focuses on energy, future iterations could incorporate data diversity—prioritizing devices with rare data—and address training latency. Additionally, automating the tuning of hyper-parameters like the loss function's "smoothness" would enhance the system's plug-and-play capabilities.
Conclusion
This paper introduces a robust, energy-aware scheduling mechanism for OtA FEEL. By leveraging temporal correlations in gradients and Lyapunov optimization, the proposed method achieves superior learning performance while strictly managing the finite energy resources of edge devices.
Summary updated