URLM: Utilising reinforcement learning to schedule subflows in MPTCP

摘要

Parallel transmission in multiple access networks using MultiPath TCP(MPTCP) greatly enhances the throughput. However, critical packet disorder iscommonly observed due to traffic fluctuation and path diversity. Although severalpredictive scheduling algorithms have been proposed to solve this problem,they cannot accommodate prediction accuracy and real-time adaptation simultaneouslyin a dynamic network environment. The time overhead in modifyingscheduling parameters to adapt to network changes leads to performance degradationin throughput and packet disorder. In this study, we propose a schedulingalgorithm called Utilising Reinforcement Learning to Schedule Subflows inMPTCP (URLM). We apply reinforcement learning to select an optimal schedulingparameter in real time, which brings significant time benefits for modifyingthe parameters. The simulation comparison experiments show that URLMreduces the average number of out-of-order packets and the time overhead inadapting to network changes while improving global throughput.