Objective-Oriented Resource Pooling in MPTCP: A Deep Reinforcement Learning Approach

The most important benefit introduced by multi-path TCP (MPTCP) is its ability of resource pooling. The congestion control algorithm and the packet scheduler in MPTCP work together to consume the pooled network resource of different sub-paths. However, the objectives of a MPTCP congestion control algorithm and a packet scheduler are not always in agreement with each other, which hinders the enhancement of the applications’ performance. In this paper, we propose Partner, a distributed Deep Reinforcement Learning (DRL)-based congestion control algorithm in MPTCP. By setting the reward function which is in agreement with the corresponding packet scheduler, Partner can accurately shape the decision space utilized by the packet scheduler, thus unleashing its full power. Our results show that Partner significantly outperforms the state-of-the-art congestion control algorithms in terms of meeting various applications’ requirements.