HybridRTS: A Hybrid Congestion Control Framework with Rule and Reinforcement Learning for Low-Latency WebRTC Live Video Streaming

Kaizhe Zhang, Zhiwen Wang, Hansen Ma, Haipeng Du, Weizhan Zhang

Abstract

Web Real-Time Communication (WebRTC) technology meets real-time communication standards and is currently supported by mainstream browsers to meet low-latency live streaming requirements. The WebRTC Google Congestion Control (GCC) algorithm utilizes a one-way delay gradient signal for conservative bandwidth estimation. However, this makes it sensitive to jitter, resulting in a decrease in the estimated bandwidth and low video quality. In this paper, we propose HybridRTS, a framework for congestion control aiming to strategically switch between proposed rule-based and reinforcement learning (RL) based methods. HybridRTS is evaluated via trace-driven simulations and real-world experiments, outperforming GCC with improvements in average Quality of Service (QoS) scores of 11.57% - 18.57%. Additionally, HybridRTS outperforms standalone rule-based and RL-based methods with decreasing in average latency by 11.46% - 29.66%. According to extensive experiments, HybridRTS yields solid results on various network conditions.
BibTeX
@inproceedings{zhang2022hybridrts,
  author = {Kaizhe Zhang and Zhiwen Wang and Hansen Ma and Haipeng Du and Weizhan Zhang},
  title = {{HybridRTS: A Hybrid Congestion Control Framework with Rule and Reinforcement Learning for Low-Latency WebRTC Live Video Streaming}},
  booktitle = {HPCC/DSS/SmartCity/DependSys},
  year = {2022},
  doi = {10.1109/hpcc-dss-smartcity-dependsys57074.2022.00265},
}

Loading related content...

Figure