Learning Accurate Network Dynamics for Enhanced Adaptive Video Streaming
Jiaoyang Yin, Hao Chen, Yiling Xu, Zhan Ma, Xiaozhong Xu
Abstract
The adaptive bitrate (ABR) algorithm plays a crucial role in ensuring satisfactory quality of experience (QoE) in video streaming applications. Most existing approaches, either rule-based or learning-driven, tend to conduct ABR decisions based on limited network statistics, e.g., mean/standard deviation of recent throughput measurements. However, all of them lack a good understanding of network dynamics given the varying network conditions from time to time, leading to compromised performance, especially when the network condition changes significantly. In this paper, we propose a framework named <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">ANT</i> that aims to enhance adaptive video streaming by accurately learning network dynamics. <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">ANT</i> represents and detects specific network conditions by characterizing the entire spectrum of network fluctuations. It further trains multiple dedicated ABR models for each condition using deep reinforcement learning. During inference, a dynamic switching mechanism is devised to activate the appropriate ABR model based on real-time network condition sensing, enabling <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">ANT</i> to automatically adjust its control policies to different network conditions. Extensive experimental results demonstrate that our proposed <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">ANT</i> achieves a significant improvement in user QoE of 20.8%-41.2% in the video-on-demand scenario and 67.4%-134.5% in the live-streaming scenario compared to state-of-the-art methods, across a wide range of network conditions.