Deep Reinforcement Learning Based Bitrate and Redundance Ratio Adaption for Panoramic Video Transmission

Abstract

User experience of panoramic video is difficult to guarantee for the conflict between limited bandwidth resources and ultra high encoding bitrate, simultaneously, existing research neglects an important indicator called black area ratio in their quality of experience (QoE) models, and traditional bitrate adaptive (ABR) algorithms appear less intelligent in the complicated QoE optimization problem. In this paper, we first establish a novel QoE model taking black area ratio into account. Next, a bitrate and redundance ratio adaptive algorithm based on deep reinforcement learning (DRL) is proposed. Via continuous interaction with the underlying environment, the DRL agent can learn an effective policy to control video encoding bitrate and redundance ratio given current system state, which includes the predicted FoV, available wireless bandwidth, etc. With an open user viewpoint dataset, simulation results demonstrate that our proposal outperforms buffer-based (BB) and rate-based (RB) adaptive algorithms, and achieves competitive performance compared to model predictive control (MPC) that provides a performance upper bound in our case.