Augmenting End-to-End Congestion Control with Centralized Reinforcement Learning

Traditional TCP congestion controls operate in a fully distributed, end-to-end manner. In the absence of explicit information about the network state, independently operating TCP flows must gradually and cautiously converge to high and fair rates.
We propose Coordinated Congestion Control (C3), a hybrid congestion control approach that augments conventional end-to-end control with periodic guidance from a centralized reinforcement learning (RL) agent. Leveraging a global network view, C3 coordinates congestion window adjustments across flows, improving throughput, reducing latency, and accelerating convergence to fair throughput allocations.
Our ns-3-based evaluations across diverse network scenarios show that C3-guided flows substantially improve latency and fairness over CUBIC and BBRv1, while maintaining high throughput even under challenging conditions for congestion control, including reverse-path congestion and non-congestion-related loss. Even infrequent C3 guidance already delivers significant performance gains.