DACC: Data Augmentation for Learning-based Congestion Control

Machine learning (ML) stands as a powerful tool for advancing congestion control algorithms (CCAs). However, the contradiction between the substantial number of samples required for training and the cost of collecting samples from real networks, hurts the effectiveness of training models, especially for deep reinforcement learning (DRL). Through experimental measurements, we reveal that it is difficult for existing DRL-based CCAs to train high-quality models with insufficient samples. To address this problem, we propose a Data Augmentation framework for learning-based Congestion Control (DACC). To augment samples in a lightweight yet effective way, DACC selects high-quality policies to generate actions and employs a backward-forward state model to compute the next state. An extensive set of experiments in both real-world Internet and emulated networks demonstrate that DACC reduces latency by 24.41%, and decreases packet loss rate by 73.29% compared to the state-of-the-art learning-based CCA.