AdaBoost-TCP: A Machine Learning-Based Congestion Control Method for Satellite Networks

Abstract

In a highly dynamic satellite network, frequent switching of satellite links leads to in-stability of the connection, which greatly increases the occurrence of packet loss. Existing TCP senders cannot effectively distinguish the types of network packet loss, resulting in lower network utilization. This paper proposes a machine learning-based congestion control strategy AdaBoost-TCP. AdaBoost-TCP constructs an adaptive Boost recognition model that can effectively classify the packet loss type in the satellite network. The receiver uses the model to differentiate the type of lost packets and combines the ECN flag to transmit the result to the sender. The sender adopts adaptive congestion control measures according to the type of packet loss obtained. Ada-Boost-TCP can have better classification speed and higher efficiency without increasing network load. From the ns-2 simulation results, when the packet loss rate is be-tween10-5-10-4, the AdaBoost-TCP strategy can increase throughput by up to 10% compared to Hybla, CUBIC, and Westwood. What’s more, it can achieve good fair-ness compared to NewReno.