RADEL: Resilient and Adaptive Distributed Edge Learning in dynamic environments

Wang, Qiyuan (2025) RADEL: Resilient and Adaptive Distributed Edge Learning in dynamic environments. PhD thesis, University of Glasgow.

Abstract

The rapid evolution of edge computing has fundamentally transformed distributed machine learning by enabling intelligence at the network periphery, where data originates. This paradigm shift has facilitated real-time analytics and responsive services across decentralized environments. However, deploying effective machine learning systems at the edge introduces substantial challenges: unpredictable node failures compromise service continuity; evolving data distributions trigger concept drift that degrades model performance; irrelevant data inclusion reduces prediction accuracy; client mobility undermines traditional learning assumptions; and communication inefficiencies limit scalability in resource-constrained settings.

This thesis presents RADEL, a comprehensive framework for Resilient and Adaptive Distributed Edge Learning that systematically addresses these challenges through five interconnected contributions. First, we introduce a novel resilience mechanism that maintains service continuity during node failures by enabling surrogate nodes to effectively serve prediction requests of failing counterparts. Our approach leverages statistical signatures of neighboring nodes’ data to build enhanced local models through various information extraction strategies, demonstrating significant improvements over traditional replication-based methods while requiring minimal inter-node data transfer.

Second, we develop maintenance strategies to preserve model resilience under concept drift, a critical challenge in dynamic edge environments where data distributions evolve over time. By analyzing how different types of drift affect enhanced models, we propose efficient maintenance mechanisms that achieve optimal trade-offs between data transmission volume and adaptation effectiveness, maintaining performance across heterogeneous data sources while minimizing communication overhead.

Third, we propose a query-driven data-centric learning approach that progressively discovers relevant data regions from query patterns to optimize predictive performance. This mechanism integrates optimal stopping theory to determine when to conclude model refinement and incorporates adaptive update policies that respond to changes in both data and query distributions. Our experimental results demonstrate up to 63% improvement in predictive accuracy compared to traditional approaches that utilize all available data.

Fourth, we present DA-DPFL, a dynamic aggregation framework for decentralized personalized federated learning that addresses data heterogeneity while reducing communication and computational costs. By enabling clients to reuse previously trained models within the same communication round and employing a sparse-to-sparser pruning strategy based on model compressibility, DA-DPFL achieves superior test accuracy while reducing energy consumption by up to 5× compared to state-of-the-art approaches.

Finally, we introduce MOBILE, a mobility-aware framework that optimizes client selection and bandwidth allocation in federated learning environments with mobile clients. By formulating the problem as a regularized Mixed-Integer Quadratic Programming optimization and incorporating both historical and current mobility patterns, MOBILE increases successful client participation rates from 32% to 89% while reducing wasted bandwidth by 43% compared to mobility-agnostic approaches.

Through comprehensive theoretical analysis and extensive empirical evaluation on diverse real-world datasets, we demonstrate that RADEL significantly improves learning efficiency, reliability, and adaptability in dynamic edge environments. Our research establishes a robust foundation for deploying resilient machine learning services at the edge, advancing the state-of-the-art in distributed intelligence for next-generation computing systems. The frameworks and algorithms developed in this thesis provide practical solutions for the deployment of intelligent services in smart cities, industrial IoT, autonomous systems, and other edge-centric applications where reliability and adaptability are paramount.

• Item Type: Thesis (PhD)
• Qualification Level: Doctoral
• Subjects: T Technology > TK Electrical engineering. Electronics Nuclear engineering
• Colleges/Schools: College of Science and Engineering > School of Computing Science
• Supervisor's Name: Anagnostopoulos, Dr. Christos
• Date of Award: 2025
• Depositing User: Theses Team
• Unique ID: glathesis:2025-85641
• Copyright: Copyright of this thesis is held by the author.
• Date Deposited: 10 Dec 2025 16:19
• Last Modified: 12 Dec 2025 09:27
• Thesis DOI: 10.5525/gla.thesis.85641
• URI: https://theses.gla.ac.uk/id/eprint/85641

Related URLs

• Article DOI
• Article DOI
• Article DOI
• Enlighten Publications Record
• Enlighten Publications Record
• Conference proceeding
• Conference proceeding
• Enlighten Publications Record
• Enlighten Publications Record
• Pre-print