Abstract
Training machine learning models on edge devices is critical for low-latency applications, yet most research focuses on single-orchestrator scenarios. This paper introduces a framework for Mobile Edge Learning (MEL) that supports multiple independent orchestrators competing for a shared pool of resource-constrained devices. By jointly optimizing learner association and data allocation, the proposed framework significantly reduces energy consumption while preserving model accuracy. We explore centralized and decentralized solution paths, finding that decentralized heuristics offer a scalable and efficient alternative for large-scale IoT deployments.
Introduction
The proliferation of IoT devices has driven the need for Mobile Edge Learning (MEL), where training occurs directly on local sensors to ensure low latency and data privacy. However, real-world environments like smart cities often involve multiple independent orchestrators—such as different municipal departments or private companies—operating simultaneously. This creates a "straggler’s dilemma," where the slowest device in a distributed training process dictates the overall speed. Managing resources in this multi-task environment requires balancing energy efficiency, processing speed, and model convergence across a heterogeneous pool of learners.
Methodology
The authors formulate the resource management problem as a Multi-Objective Optimization Problem (MOP) aimed at minimizing total system energy consumption while maximizing learning accuracy. The framework follows a structured workflow to manage the complexity of multi-orchestrator environments.
Three primary solution strategies are proposed:
- Centralized Optimization (COPT): Uses convex relaxation and a Branch and Bound (BnB) algorithm to find the optimal trade-off. While precise, it requires global network knowledge and is computationally expensive.
- Associate-Allocate-Train (AAT): A heuristic that decomposes the problem into sequential steps: associating devices with tasks, allocating data volumes, and optimizing training cycles.
- Factor-Based Association (FBA & L-FBA): A lightweight approach using an "Association Factor" (AF) based on a device's CPU speed and proximity to the orchestrator. The Learner-driven version (L-FBA) allows devices to autonomously select orchestrators, enabling a fully decentralized system.
Experimental Evaluation
The framework was evaluated using three computer vision benchmarks: MNIST (handwritten digits), Fashion-MNIST (clothing items), and CIFAR-10 (complex color images).
Key Results
- Energy Efficiency: The proposed methods significantly outperformed energy-unaware baselines by explicitly considering the battery costs of both computation and transmission.
- Accuracy and Convergence: Even with the decentralized L-FBA heuristic, models achieved accuracy within 2-3% of the centralized benchmark. For simpler datasets like MNIST, convergence was reached in as few as four global cycles.
- Scalability: Unlike the centralized COPT approach, which scales exponentially, the FBA and L-FBA heuristics scale linearly with the number of devices, making them suitable for massive IoT networks.
Discussion and Conclusion
A key insight from this work is the viability of decentralization. The L-FBA approach demonstrates that edge networks can self-organize efficiently using simple local metrics like the Association Factor. This balances raw processing power against the energy costs of data transmission, naturally avoiding inefficient pairings.
While the framework shows promise, challenges remain regarding Non-IID (Independent and Identically Distributed) data, where variations in local data distributions can degrade accuracy. Future work will focus on integrating robust data-sharing techniques and automating parameter tuning to enhance system resilience in dynamic environments.
Summary updated