Dynamic Adaptive Streaming based on Deep Reinforcement Learning

Abstract

Dynamic adaptive streaming over HTTP (DASH) is the dominant technology of multimedia delivery over the Internet. In DASH system, adaptive bitrate (ABR) algorithms running on client-side video player are the key to improve user quality of experience (QoE). However, most existing ABR algorithms employ fixed control rules to make bitrate decisions based on throughput, playback buffer size, or a combination of the two. As a result, their performance in the complicated and fluctuant network environment is incompetent. In this paper, we propose QRL, a bitrate adaptation approach based on deep reinforcement learning. QRL uses double Q-Learning, an enhanced Q-Learning method. After training the neural network model, the algorithm can select proper bitrates for future video segments based on all the information collected by client during the video playback process. Simulation results show that QRL achieves better performance than other algorithms.