Machine Learning-Enhanced Active Queue Management Algorithms: A Systematic Literature Review

Abstract

Active Queue Management (AQM) algorithms rely on static parameters, which limits their ability to adapt to the dynamic and heterogeneous nature of modern network traffic. This paper presents the first Systematic Literature Review (SLR) conducted in strict adherence to PRISMA guidelines, mapping the various ML techniques integrated into AQM, the performance metrics improved and the network environment tested. Our review shows that multiple ML methods have been implemented with reinforcement learning being the most dominant approach, delivering satisfying results across various network environments. Supervised learning proves effective in flow classification scenarios, while recent approaches such as federated learning, multi-agent RL and LLM distillation continue the push the boundaries of the field. We identify four open challenges: networking dataset unavailability, single-based reward function in RL techniques, ML vulnerabilities, and cross environment adaptation. This review gives an organized classification of ML-AQM approaches and a well-defined roadmap for future research.