Deep Reinforcement Learning-Based Congestion Control for File Transfer over QUIC

Congestion control is one of the key mechanisms of communication in QUIC protocol which controls how much data and at which rate can be send to an endpoint at particular moment of time for better use of shared network resources and avoids moving into congestive collapse state. In this work we tackle the problem of congestion control for file transfer over QUIC. We propose Reinforcement Learning Soft Actor Critique based congestion control with monitor window limit (SAC-MWL) and address the challenges it may have in a real network. We show how these challenges can block RL-based congestion control from effective learning of best policy and suggested a way to solve it. Our experiments are conducted in three different domains: pure virtual environment, lab-controlled network and real network where end points are spread all over the world. We compared performance of our approach with classical congestion controls CUBIC and BBR in a various network conditions and achieved up to 66.9 % reduction in a file transfer time.