Lyapunov-Based Optimization of Edge Resources for Energy-Efficient Adaptive Federated Learning

Claudio Battiloro, Paolo Di Lorenzo, Mattia Merluzzi, Sergio Barbarossa

TL;DR

This paper introduces a dynamic framework for optimizing Federated Learning (FL) at the wireless network edge by balancing the competing demands of energy consumption, latency, and model accuracy. By leveraging Lyapunov stochastic optimization, the authors create a system that adaptively manages both radio and computing resources to ensure efficient, continuous learning even in unpredictable, non-stationary environments.

What Problem Does This Paper Address?

Training machine learning models at the "edge" (on devices like phones, sensors, or vehicles) is notoriously difficult. Unlike cloud-based learning, edge devices have strictly limited battery life and processing power, and they must communicate over unreliable wireless links.

Most existing research treats Federated Learning as a static task: you train a model until it converges and then stop. However, real-world edge applications—like autonomous driving or industrial IoT—require adaptive learning, where models must continuously update and track changes in their environment. The core challenge is: How do we jointly optimize communication (bits, rates) and computation (CPU cycles) to meet strict latency and accuracy targets without draining the devices' batteries?

Methods

The authors propose a dynamic resource allocation strategy based on Lyapunov Stochastic Optimization. This mathematical framework allows them to turn long-term average constraints (like "average latency must be under 20ms") into a series of simpler, per-time-slot optimization problems using "virtual queues."

Key components of their approach include:

  • Joint Optimization: The system doesn't just pick a transmission rate; it simultaneously decides which devices should participate, how many bits to use for gradient quantization, the transmit power, and the CPU frequencies for both the devices and the edge server.
  • Two Operational Modes:
    1. Model-Based: Used for convex tasks (like Least Mean Squares regression) where mathematical formulas can predict how resource choices will affect learning performance.
  • Data-Driven: Designed for complex, non-convex tasks (like Deep Learning). Since there are no simple formulas for CNN accuracy, the system estimates performance online using validation data and "surrogate functions" to guide optimization.
  • Greedy Algorithmic Solution: To handle the complexity of picking the best combination of devices and settings (a "mixed-integer nonlinear" problem), they developed a low-complexity iterative greedy algorithm that is practical for real-time execution.

Datasets

The researchers evaluated their framework using two primary setups:

  • Synthetic Data: Used to test a Federated Least Mean Squares (LMS) adaptive filter for regression.
  • MNIST: The classic dataset of handwritten digits was used to train a Deep Convolutional Neural Network (CNN).
  • Non-stationary MNIST: To test the "adaptive" part of their system, they modified the MNIST task so that the data distribution changed over time (e.g., the model suddenly had to learn a different subset of digits), simulating a changing real-world environment.

Key Insights and Ideas

  • Quantization as a Control Knob: The number of bits used to compress gradients isn't just a bandwidth saver; it's a fundamental trade-off. More bits mean better accuracy and faster convergence but higher latency and energy use. The Lyapunov framework treats this as a dynamic variable.
  • Struggler Mitigation: The system naturally ignores "strugglers"—devices with poor signal or slow processors—if including them would violate the overall latency or energy goals.
  • Closing the Loop: By using virtual queues for accuracy and convergence rate, the system "pushes harder" (uses more power/compute) when the model is performing poorly and "relaxes" (saves energy) once the target performance is reached.
  • Surrogate Functions for Deep Learning: One of the most clever bits is using a simple logistic sigmoid function as a "surrogate" for CNN accuracy. This allows the optimizer to make smart guesses about how changing quantization bits will affect a complex neural network without needing a perfect mathematical model of the network itself.

Key Results

  • Guaranteed Performance: The framework successfully maintained average latency and accuracy within predefined bounds across all tests.
  • Significant Energy Savings: Compared to baseline strategies (like giving every device an equal data rate), the proposed method achieved a much better trade-off curve. For a given latency target, it could reduce total power consumption by roughly 30-50% in some scenarios.
  • Rapid Adaptation: In the non-stationary MNIST test, the system reacted almost instantly to changes in the data distribution and stricter accuracy requirements, re-allocating resources to bring the model back to the target performance level.

Limitations and Future Work

  • Model-Based Non-convexity: The "Model-Based" approach currently relies on the assumption of strong convexity. Developing theoretical performance metrics for non-convex deep learning that are simple enough for real-time optimization remains an open challenge.
  • Heuristic Optimality: The greedy algorithm is a heuristic. While it's computationally efficient, it doesn't strictly guarantee a global optimum for every single time slot, though the authors show it works very well in practice.
  • Scalability: While the greedy approach is much faster than an exhaustive search, its complexity still grows with the number of devices. Future work could look at even more scalable "clustering" approaches for massive-scale IoT networks.

Summary updated

Abstract

The aim of this paper is to propose a novel dynamic resource allocation strategy for energy-efficient adaptive federated learning at the wireless network edge, with latency and learning performance guarantees. We consider a set of devices collecting local data and uploading processed information to an edge server, which runs stochastic gradient-based algorithms to perform continuous learning and adaptation. Hinging on Lyapunov stochastic optimization tools, we dynamically optimize radio parameters (e.g., set of transmitting devices, transmit powers, bits, and rates) and computation resources (e.g., CPU cycles at devices and at server) in order to strike the best trade-off between power, latency, and performance of the federated learning task. The framework admits both a model-based implementation, where the learning performance metrics are available in closed-form, and a data-driven approach, which works with online estimates of the learning performance of interest. The method is then customized to the case of federated least mean squares (LMS) estimation, and federated training of deep convolutional neural networks. Numerical results illustrate the effectiveness of our strategy to perform energy-efficient, low-latency, adaptive federated learning at the wireless network edge.
BibTeX
@article{battiloro2022lyapunov,
  author = {Claudio Battiloro and Paolo Di Lorenzo and Mattia Merluzzi and Sergio Barbarossa},
  title = {{Lyapunov-Based Optimization of Edge Resources for Energy-Efficient Adaptive Federated Learning}},
  journal = {IEEE Transactions on Green Communications and Networking},
  year = {2022},
  doi = {10.1109/tgcn.2022.3186879},
}
Figure