Muno: Improved Bandwidth Estimation Scheme in Video Conferencing Using Deep Reinforcement Learning

Abstract

Many studies have used machine learning techniques for bitrate control to improve the quality of experience (QoE) of video streaming applications. However, most of these studies have focused on HTTP adaptive streaming with one-to-one connections. This research examines video conferencing applications that involve real-time, multiparty, and full-duplex communication among participants. In conventional video conferencing systems, a rule-based algorithm is typically employed to estimate the available bandwidth of each participant, and the outcomes are then used to control the video delivery rate to the participant. This paper proposes Muno, a bandwidth prediction framework based on deep reinforcement learning (DRL) for multiparty video conferencing systems. Muno aims to enhance the overall QoE by using DRL to improve bandwidth estimation for each connection. The experimental results indicate that Muno achieves a significantly higher video streaming rate, video resolution, and framerate while lowering delay in highly dynamic networks when compared to the state-of-the-art rule-based algorithms and roughly equivalent streaming rate and delay in stable networks. Moreover, Muno can generalize well to different network conditions which were not included in the training set. We also implemented a high-performance and scalable version of Muno for in-campus deployment.