GenRL-ACC Generative Data-Augmented Deep Reinforcement Learning for Adaptive Congestion Control

Real-time communication (RTC) applications require congestion control (CC) mechanisms capable of sustaining high throughput under stringent latency constraints. Although deep reinforcement learning (DRL) provides adaptive control superior to conventional heuristic and loss-based schemes, learned policies often generalize poorly when exposed to unseen network dynamics. To address this limitation without modifying agent architectures or reward formulations, this paper presents GenRL-ACC, a data-centric framework that enhances DRL robustness through generative bandwidth trace augmentation. Using the ReCoCo platform, Twin Delayed DDPG (TD3) and Soft Actor-Critic (SAC) agents are trained across varying synthetic-to-real mixing ratios generated by TTS-GAN and Diffusion-TS, and subsequently evaluated against real-data baselines and Google Congestion Control (GCC) on held-out empirical traces. Experimental results demonstrate that: (1) generative trace augmentation substantially improves downstream Quality of Experience (QoE), where TD3 trained with 50% TTS-GAN traces achieves an 82.65 QoE score-outperforming the real-only baseline by 12.05% and GCC by 8.32%; (2) optimal synthetic-to-real ratios depend fundamentally on the interplay between the generative architecture and the reinforcement learning algorithm; and (3) statistical distribution fidelity alone is insufficient for performance gains, as generative diversity provides a stronger regularizing benefit. Consequently, training data composition must be treated as a co-designed parameter evaluated directly on downstream operational metrics.