A Survey of Deep Reinforcement Learning for Network Congestion Control

With the rapid development of internet technologies and the diversification of network applications, congestion control (CC) has become crucial in maintaining network stability and improving transmission efficiency. Traditional TCP congestion control algorithms (CCAs) are mainly based on heuristic designs and rely on hardwired mapping control logic, making them difficult to adapt to the complex and dynamically changing network environment. Deep reinforcement learning (DRL) as an emerging technology has demonstrated remarkable success in various fields, and its potential in network congestion control has attracted the attention of researchers. DRL enables the agent to learn optimal decisions through interactions with the environment, allowing it to dynamically adjust rate control policies and achieve key performance goals such as high throughput, low latency, and minimal packet loss. This paper briefly introduces the progress of DRL in the field of congestion control, covering the core issues of CC, the concept of DRL, the feasibility of combining DRL and CC, and recent research on DRL-based CCAs. Finally, the paper summarizes existing studies and discusses the future development directions of this field.