Engineering AI-enhanced edge computing with predictive intelligence

Alfahad, Saleh (2026) Engineering AI-enhanced edge computing with predictive intelligence. PhD thesis, University of Glasgow.

Abstract

Distributed computing devices, including smartphones and edge micro-servers, gather and analyse data near users, presenting significant opportunities for enhancing service responsiveness and system efficiency. The constrained processing capabilities of edge nodes and the fluctuating nature of service requests pose considerable challenges in the management and orchestration of distributed services. This thesis investigates intelligent strategies for optimising decision-making in edge computing (EC) settings to improve Quality of Service (QoS), minimise latency, and ensure system flexibility under resource limitations.

The first part of the thesis introduces a time-efficient decision-making framework grounded in Optimal Stopping Theory (OST), tackling the dilemma of whether edge nodes ought to replicate unavailable services (pull action) or delegate tasks to neighbouring nodes (push action). This system adjusts dynamically to variations in service request patterns and resource availability. The OST-based solution exhibits enhanced performance over conventional service management methods in synthetic experimental settings motivated by smart city applications.

The second part presents a cost-aware service orchestration model that enhances the OST method by including time-sensitive service demand patterns and system-level restrictions. This paradigm guarantees the provision of edge services in a resource-efficient and timely manner, reducing both over-replication and superfluous task offloading. It is assessed in several dynamic contexts to illustrate its practical efficacy in enhancing edge resource allocation.

The third part presents the initial study of Experts with Gradient Bandits (ExpGradBand) and addresses the important problem of node selection in Edge Computing settings. Through gradient-based decision-making and lightweight contextual experiments, it introduces ExpGrad-Band, a novel node-selection technique that extends the conventional Multi-Armed Bandit (MAB) framework. This part provides preliminary evidence of the algorithm’s effectiveness through controlled, small-scale experiments with sparse contextual data. ExpGradBand outperforms baseline techniques in terms of convergence speed and cumulative regret, demonstrating adaptive decision-making under constrained experimental settings. These findings provide initial evidence of the framework’s ability to support intelligent node selection in real time.

The fourth part builds on that initial research by providing a broader empirical validation of the same framework within a more comprehensive sequential learning setting that incorporates deep-learning expert models into the Multi-Armed Bandit (MAB) architecture. In uncertain and diverse edge contexts, this framework supports asynchronous, lost, and delayed feedback, enabling reliable and scalable decision-making. By adding expert-driven predictors, including Long Short-Term Memory (LSTM), Feedforward (FNN), and Recurrent (RNN) networks, the ExpGradBand framework improves node-selection accuracy and strengthens adaptation to large-scale, real-world scenarios. This part therefore provides a broader experimental and analytical assessment of ExpGradBand, demonstrating its cross-scenario generalisation, scalability, and resilience in dynamic edge computing systems.

The fifth part extends the thesis by addressing dynamic assignment problems under strict budget constraints through a pure-exploration sequential learning framework. This chapter introduces a Thompson Sampling-based Recursive Block Elimination approach that efficiently explores large or continuous action spaces without requiring exhaustive sampling. By discretising the action space and recursively eliminating sub-optimal blocks using confidence-driven criteria, the proposed method enables reliable identification of near-optimal actions within limited budgets. The framework is evaluated on dynamic pricing and federated learning pruning scenarios, demonstrating strong performance in terms of misidentification probability and simple regret, and highlighting its scalability and effectiveness in resource-constrained decision-making environments.

The thesis closes by underscoring the need for adaptive, feedback-oriented orchestration systems in edge computing. This study advances the development of self-optimizing, resilient edge computing infrastructures by integrating Optimal Stopping Theory, cost-aware service modelling, and intelligent node selection algorithms, enabling the provision of efficient services within real-world operational limitations.

• Item Type: Thesis (PhD)
• Qualification Level: Doctoral
• Subjects: Q Science > QA Mathematics > QA75 Electronic computers. Computer science; 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 and Puthiya Parambath, Dr. Shameem Ahamed
• Date of Award: 2026
• Depositing User: Theses Team
• Unique ID: glathesis:2026-86163
• Copyright: Copyright of this thesis is held by the author.
• Date Deposited: 13 Aug 2026 09:11
• Last Modified: 13 Aug 2026 15:50
• Thesis DOI: 10.5525/gla.thesis.86163
• URI: https://theses.gla.ac.uk/id/eprint/86163

Related URLs

• Article DOI
• Article DOI
• Conference proceeding
• Article DOI
• Article DOI