Reinforcement Learning-Based Data Scheduling for MPTCP in Streaming Media Delivery

In this paper, an MPTCP scheduling method based on DQN is proposed, which aims to solve the dynamic adaptive problem in the process of video streaming media transmission by introducing the chunk manager. To address the problem of performance degradation caused by targeting only a single performance metric in the DASH environment, a reward function is proposed that uses two metrics, throughput and chunk download time, and aims to help the intelligence to better learn the video chunk delivery policy. The proposed use of a chunk manager helps the DQN to understand the chunk information, which is used to accurately detect the transmission status of video chunks and pass the chunk information to the scheduler. Finally the use of online training and offline decision making can make the DQN more efficient and flexible, enabling it to better adapt to complex environments and tasks. It is experimentally verified that the DQN-based data scheduling algorithm proposed in this paper has some performance improvement over the classical MPTCP data scheduling algorithm, and the DQN-based scheduling algorithm keeps the download time distribution of chunks at a low and stable level.