Mutant - Learning Congestion Control from Existing Protocols via Online Reinforcement Learning

Learn about an innovative online reinforcement learning approach to network congestion control through this 17-minute conference presentation from NSDI '25. Discover how Mutant, a novel algorithm developed by researchers from Saint Louis University and Politecnico di Torino, addresses the longstanding challenge of congestion control by adapting to and learning from the behavior of existing high-performing protocols. Explore the key design challenges involved in determining optimal protocols for different network scenarios and creating an evolutionary system that can accommodate future protocols with minimal modifications. Understand how this approach differs from traditional machine learning solutions by reducing dependencies on extensive training and specific network configurations. Examine the comprehensive evaluation results demonstrating Mutant's superior performance in achieving lower delays and higher throughput compared to prior learning-based schemes while maintaining fairness and exhibiting negligible harm to competing flows across diverse and dynamic network conditions.