Abstract
Federated Learning (FL) typically assumes reliable client availability, an assumption that fails in Internet of Things (IoT) networks powered by ambient energy harvesting (EH). This paper introduces the Energy Harvesting Federated Learning (EHFL) framework to address the challenges of unpredictable device availability and temporal correlation in battery states. By employing adaptive learning rates and a "myopic" scheduling policy that prioritizes devices with higher energy reserves, the authors demonstrate that EHFL can achieve convergence rates comparable to standard FL. Experimental results on the CIFAR-10 dataset confirm that the proposed myopic strategy significantly outperforms traditional greedy and round-robin approaches.
Introduction
Standard Federated Learning (FL) paradigms often rely on the assumption that clients are either consistently available or that their drop-outs are independent and identically distributed (i.i.d.) random events. However, in the burgeoning field of the Internet of Things (IoT), many edge devices are powered by Energy Harvesting (EH) sources such as solar, wind, or thermal energy.
This reliance on ambient energy introduces two primary challenges:
- Uncertainty: Device participation is contingent upon having sufficient battery levels, which fluctuate based on environmental conditions.
- Temporal Correlation: A device's participation in one round depletes its battery, making it less likely to be available in subsequent rounds. This "memory" effect violates the independence assumptions required for standard convergence proofs.
The EHFL framework seeks to bridge this gap, providing a robust approach to training global models across networks of unreliable, energy-constrained devices.
The EHFL Framework
The authors propose EHFL to model the energy queue (battery level) of each client as a stochastic process. The system architecture involves an edge server coordinating with multiple EH-powered devices that perform local computation on local data before syncing with the global model.
Adaptive Learning Rates
A core innovation of EHFL is the use of a time-varying learning rate ($\eta_t$). To account for the fluctuating number of active clients ($n_t$) in each round, the authors suggest setting the learning rate proportional to the square root of the active client count: $\eta_t \propto \sqrt{n_t}$. This normalization ensures that the global model updates remain stable despite significant swings in participation.
Scheduling Policies
The paper evaluates several scheduling strategies:
- Greedy: Uses every device as soon as it has enough energy to participate.
- Round Robin: A fixed rotation of devices regardless of current energy levels.
- Myopic: Prioritizes devices with the longest energy queues (fullest batteries). This policy aims to balance battery levels across the network, ensuring a more consistent number of participants over time.
Theoretical Convergence and Results
The authors provide a theoretical analysis for both Parallel SGD (single local step) and Local SGD (multiple local steps). They prove that EHFL achieves a convergence rate of $O(1/\sqrt{n_{min}T})$, where $n_{min}$ is the minimum number of clients participating in any given round and $T$ is the total number of rounds. This result matches the performance of standard FL with always-available clients, provided the adaptive learning rate is correctly applied.
The framework was validated using a Convolutional Neural Network (CNN) trained on the CIFAR-10 dataset across 10 simulated EH clients.
As shown in the figure above, the Myopic policy consistently outperforms both Greedy and Round Robin strategies. By maintaining a stable pool of participants through battery balancing, the Myopic approach achieves higher testing accuracy more rapidly. The Greedy approach, while initially competitive, suffers from high variance in participation, leading to slower and less stable convergence.
Conclusion and Future Directions
This work demonstrates that federated learning is viable even in networks of energy-harvesting devices, provided that scheduling and learning rates are adapted to the underlying energy dynamics. The "myopic" scheduler proves to be a simple yet effective way to maintain system stability.
Future research directions include extending the framework to handle non-IID data distributions across clients, moving beyond simple Bernoulli energy arrival models to more realistic patterns, and dynamically adjusting communication costs based on real-time energy availability.
Summary updated