Stabilizing Deep Reinforcement Learning Model Training for Video Conferencing

Abstract

While many studies have been conducted to apply reinforcement learning (RL) to real world problems beyond games such as Atari, video conferencing is also one of real world applications. In video conferencing, reinforcement learning is used to control the bitrate to improve the user's quality of experience (QoE). However, real world problems such as video conferencing have different characteristics compared to electronic games. Usually the rewards in real world problems are not clear or abstract, and this makes it difficult to design the RL model and training process of the model to maximize the cumulative reward. Therefore, in this paper, we present the method for stabilizing the training of the models that apply reinforcement learning to video conferencing. In addition, we established a simulation environment that can train deep RL models in 1-to-1 video conferencing. An evaluation is performed to analyze the difference between the baseline model and the models generated using the stabilization method in the simulation environment.