PhD Dissertation Final Oral Defense (May 2025)

Title of Dissertation: Dynamic Adaptive Video Streaming Over HTTP using Deep Learning
Name of Candidate: Maram Wahed Rashad Amin Helmy
Program: PhD in Engineering Systems Management
Supervisor(s): Dr. Mohamed S. Hassan, Dr. Usman Tariq, Dr. Mahmoud H. Ismail

Abstract:
Maintaining a high Quality of Experience (QoE) in adaptive video streaming remains a significant challenge in the presence of dynamic and heterogeneous network environments. Traditional Adaptive Bitrate (ABR) algorithms in Dynamic Adaptive Streaming over HTTP (DASH) often rely on basic throughput estimation methods that struggle to adapt to rapid network fluctuations caused by mobility and handoff events. These limitations lead to frequent playback interruptions, abrupt quality changes and overall QoE degradation. This dissertation proposes a novel deep learning-based framework to enhance ABR decision-making by predicting future throughput with greater accuracy and responsiveness. A transformer-based throughput prediction model is developed to capture the complex temporal dependencies in network dynamics. The predicted throughput feeds into two novel intelligent modules: a mobility-aware throughput prediction system (MATH-P) that combines mobility classification with dynamic bandwidth forecasting, and a handoff-aware prediction engine (HATH-P) that anticipates transitions between access networks (e.g., 4G and 5G) to ensure seamless adaptation during handoff events. In addition, three novel ABR algorithms are introduced using deep reinforcement learning (DRL): a throughput-aware algorithm (THA-P) that directly incorporates predicted throughput into the decision process, a mobility-aware algorithm (MATH-P ABR) that adapts to user movement patterns, and a handoff-aware algorithm (HATH-P ABR) that optimizes decisions during access network transitions. These DRL-based agents are trained and evaluated in realistic mobility and network scenarios. Extensive experiments demonstrate that the proposed methods significantly outperform both heuristic and learning-based ABR algorithms across key QoE metrics, achieving higher bitrate utility, reduced rebuffering durations, smoother bitrate transitions and improved overall playback quality. The contributions of this work offer a robust and intelligent approach to ABR in mobile and heterogeneous environments, paving the way for next-generation adaptive streaming solutions.

For more information, please contact Yosr Mohammadnoor | [email protected].