Bandwidth Estimation and Bitrate Allocation for Enhancing QoS in Video Conferencing Using Deep Reinforcement Learning

Abstract

The proportion of video traffic within the total internet traffic is steadily increas- ing and then video traffic already accounts for over half of the internet traffic. After COVID-19, there has been a rapid increase in the use of video conferencing platforms, leading to a significant rise of the traffic volume in video conferencing. Because video conferencing platforms are attracting a growing number of users, the importance of maintaining high quality of service (QoS) for these platforms is becoming more crucial. However, compared to other video-related services such as video streaming and one-to-one video telephony, video conferencing operates in a more complex environment, making the optimization of its performance challenging. Furthermore, the rule-based bitrate control algorithms used in video conferencing systems often fail to make optimal bitrate control decisions across diverse network conditions, leading to issues of underutilization and overutilization of network bandwidth. This thesis aims to enhance the overall QoS in video conferencing. To address the mentioned issues, this thesis proposes a two-stage Deep Reinforcement Learning (DRL)-based bitrate control approach. The proposed method applies DRL to bandwidth estimation and bitrate allocation, enabling DRL models to learn from various network situations and adaptively control bitrate according to the user’s network status. The DRL-based bitrate controller consists of the DRL-based bandwidth estimator and DRL-based bitrate allocator. The DRL-based bandwidth estimator enables more accurate bandwidth estimations, contributing to the overall improvement of QoS. The DRL-based bitrate allocator adjusts the bitrate allocation decisions made by the existing allocator, delivering a higher quality video stream and ensuring stable bitrate allocation. The results of various experiments validate that the DRL-based bandwidth estimator improves QoS performance compared to the existing method. When the DRL-based bitrate allocator is used in conjunction with the DRL-based bandwidth estimator, average throughput is further enhanced, delivering a higher-quality video stream. Moreover, the DRL-based bitrate controller reduces the stalling rate and improves bandwidth utilization.