MUABR: Multi-user Adaptive Bitrate Algorithm based Multi-agent Deep Reinforcement Learning

Adaptive bitrate (ABR) algorithms have been considered as efficient ways to improve the performance of video streaming by adaptively selecting appropriate video transmission bitrate. However, traditional ABR algorithms show poor adaptability to complex and dynamic networks, thus limit user’s quality of experience (QoE). In the case of multi-user streaming, multiple users simultaneously requesting video from the server often compete for a bottleneck link. The bandwidth sharing problem should also be carefully considered, otherwise it will result in poor fairness and insufficient bandwidth utilization. To address these issues, this paper proposes a multi-user adaptive bitrate algorithm (MUABR) based on multi-agent reinforcement learning. MUABR takes the overall QoE of multi-user as the optimization goal, designs neural network structure and training method by reinforcement learning, so as to solve the problem of bitrate selection and poor adaptability when performing video streaming. Furthermore, the bandwidth allocation strategy adopted in MUABR ensures fairness to a certain extent and improves bandwidth utilization. Comprehensive experimental results show that MUABR has reached the desired goal. Compared with Pensieve and Festive algorithms, MUABR can improve the average QoE of users by about 7.5%-24.6%.