Loading...
Content-Aware Adaptive Video Streaming Using Actor-Critic Deep Reinforcement Learning
Amer, Hala Nagi
Amer, Hala Nagi
Description
A Master of Science thesis in Electrical Engineering by Hala Nagi Amer entitled, “Content-Aware Adaptive Video Streaming Using Actor-Critic Deep Reinforcement Learning”, submitted in November 2024. Thesis advisor is Dr. Mahmoud Ibrahim and thesis co-advisor is Dr. Mohamed Hassan. Soft copy is available (Thesis, Completion Certificate, Approval Signatures, and AUS Archives Consent Form).
Abstract
Adaptive streaming over HTTP aims to maximize user Quality-of-Experience (QoE) through video quality adaptation. Conventional adaptation schemes measure the video quality for variable bitrate (VBR) videos in terms of the average bitrate. However, video bitrate alone is not an accurate measure of perceptual quality. Alternative quality measures, such as the Video Multi-method Assessment Fusion (VMAF), can be used to better represent the quality perceived by the human eye. Studying the VMAF of video chunks across the same quality level shows that the perceived quality depends not only on the overall video bitrate, but also on the content complexity. More complex video content has a more noticeable impact on the viewer’s QoE than static content because it affects the quality perceived by the human eye more significantly. As a result, dealing with all types of content in the same way can lead to bandwidth wastage and reduced perceptual quality. This thesis proposes four deep reinforcement learning adaptation algorithms, which are the Complexity-Aware Bitrate Selection (CABS), Complexity-Aware Resource-aware Bitrate Selection (CARBS), Complexity-Aware Resource-aware Bitrate Selection with SR (CARBS-SR), and Complexity-Aware Resource-aware Bitrate Selection with Binary SR (CARBS-BSR) algorithms. Each algorithm accounts for content complexity by prioritizing complex video chunks during bitrate selection. The CABS algorithm prioritizes only VMAF, while CARBS prioritizes only data saving. The next two algorithms, CARBS-SR and CARBS-BSR, both attempt to strike a balance between the two by using super-resolution to compensate for the VMAF loss. Simulation results show the effectiveness of the proposed complexity-aware algorithms. First, CABS achieves up to a 6.4% VMAF improvement compared to the baseline algorithms at the cost of up to a 2.7% increase in bandwidth. On the other hand, the CARBS algorithm achieves up to 20% in bandwidth savings at the cost of 31% VMAF loss. CARBS-SR and CARBS-BSR achieve up to 9.6% and 18% bandwidth savings, respectively, at the cost of approximately 15% and 28% drop in VMAF, respectively.