Mamba: Bringing Multi-Dimensional ABR to WebRTC

• DOI: 10.48550/arxiv.2308.03643
• Publication Date: 2023-01-01
• Authors: Yueheng Li; Zicheng Zhang; Hao Chen; Zhan Ma
• Keywords: Codec; WebRTC; Constant bitrate; Frame rate

Contemporary real-time video communication systems, such as WebRTC, use an adaptive bitrate (ABR) algorithm to assure high-quality and low-delay services, e.g., promptly adjusting according the instantaneous network bandwidth. However, target decisions in control codec are typically incoordinated simply ignoring effect of inappropriate resolution frame rate settings also leads compromised results control, thus devastatingly deteriorating quality experience (QoE). To tackle these challenges, Mamba, end-to-end multi-dimensional ABR is proposed, which utilizes multi-agent reinforcement learning (MARL) maximize user's QoE by adaptively collaboratively encoding factors including quantization parameters (QP), resolution, based on observed states conditions complexity information a conferencing system. We introduce curriculum improve training efficiency MARL. Both in-lab real-world evaluation demonstrate remarkable efficacy Mamba.