查看文章

TCP congestion control, although prevalently used in network technologies, is not optimal for video streaming applications. High throughput and low latency requirements of video streaming applications present a challenge for traditional congestion control protocols to provide satisfactory quality of experience to users. This limitation is mainly caused by the misidentification of non-congestion losses as congestion losses at the sender side causing the rule-based TCP congestion control algorithm to conservatively reduce the congestion window. To counter that, we first devise a transfer learning assisted reinforcement learning framework for TCP congestion control problem. The proposed framework, in an online manner, learns to operate at the optimal congestion window and distinguish non-congestion losses from congestion losses spawning high throughput and low end-to-end delay. Next, we analyze the behavior …