ENHANCING RESOURCE UTILIZATION IN EDGE COMPUTING USING DEEP Q-NETWORK
No Thumbnail Available
Date
2026-01-18
Authors
Jarrar, Yazan
Journal Title
Journal ISSN
Volume Title
Publisher
An-Najah National University
Abstract
This thesis focuses on designing and evaluating machine learning techniques to enhance the scheduling process in an edge computing environment, especially for IoT applications, with the goal of maximizing the number of tasks executed within a specified window limit. Maximizing task completion within deadlines increases throughput and service quality, while missed deadlines can degrade system performance and render results unusable.
This study investigates the two proposed algorithms, Simulated Annealing (SA) and Deep Q Network (DQN), to determine if they outperform the existing solutions for batch task scheduling in an edge computing environment. To validate the results, we used real world Augmented Reality (AR) and Internet of Vehicles data generated by the EdgeCloudSim simulator. The results clearly show that our proposed algorithms outperform others solutions, especially where the resources are strictly constrained.
The results show that our simulated annealing-based algorithm achieved up to a 71% reduction in task failure rate compared with the baseline algorithm when tested in static environments using real IoT and AR data. They also show that our Deep Q-Network based scheduler consistently achieved the lowest failure rates across dynamic scenarios, especially under higher constraints. Moreover, the Deep Q-Network model evaluated in a dynamic environment required substantially less overhead time, under one-third of the simulated annealing algorithm’s runtime, while both algorithms outperformed other baseline algorithms by up to 10% in task failure rate.
These findings support the thesis objective of maximizing the edge resource utilization while ensuring task batches complete within a specified window limit and highlight the efficiency of the Deep Q-Network based scheduling algorithm.