Machine learning based congestion control methods: a survey

I. A. Stepanov^ab, M. V. Popov^cb, A. I. Get'man^cadb, M. K. Ikonnikova^b, A. A. Belevancev<sup>c

a</sup> Moscow Institute of Physics and Technology
^b Ivannikov Institute for System Programming of the RAS
^c Lomonosov Moscow State University
^d National Research University Higher School of Economics

Abstract: Congestion control is a key aspect of modern networks. The first congestion control algorithms, such as TCP Tahoe and TCP Reno, were developed in the late 20th century, and their core concepts remain relevant to this day. With the development of high-speed networks, specialized algorithms such as TCP BIC and TCP CUBIC were created, which are adapted to these conditions. However, classical algorithms with predefined rules are not always effective in all network environments, and with the rise of 4G, 5G, and satellite communications, the congestion control issue has become increasingly relevant. This has led to the emergence of numerous works on machine learning-based congestion control algorithms, particularly reinforcement learning, which can adapt to dynamically changing network conditions. This paper presents and reviews both classical congestion control algorithms and the most popular and recent machine learning-based algorithms, along with some implementations using multipath. Additionally, it highlights the most significant challenges of machine learning-based algorithms and discusses potential directions for future research in this field.

Keywords: network congestion control, reinforcement learning.

DOI: 10.15514/ISPRAS-2025-37(3)-18