Learning to Harness Bandwidth With Multipath Congestion Control and Scheduling

Multipath TCP (MPTCP) has emerged as a facilitator for harnessing and pooling available bandwidth in wireless/wireline communication networks data centers. Existing implementations of MPTCP such as, Linked Increase Algorithm (LIA), Opportunistic LIA (OLIA) BAlanced LInked Adaptation (BALIA) include separate algorithms congestion control packet scheduling, with pre-selected parameters. We propose Deep Q-Learning (DQL) based framework joint scheduling MPTCP. At the heart solution is an intelligent agent interface, learning actuation, which learns from experience optimal mechanism using DQL techniques policy gradients. provide rigorous stability analysis system dynamics provides important practical design insights. In addition, proposed DQL-MPTCP algorithm utilizes ‘recurrent neural network’ integrates it ‘long short-term memory’ continuously i) dynamic behavior subflows (paths) ii) responding promptly to their prioritized replay. With extensive emulations, we show that DQL-based outperforms LIA, OLIA BALIA algorithms. Moreover, robust time-varying network characteristics, exploration exploitation paths.