Autonomous Congestion Control in High-Speed Networks Via Multiagent Deep Q Learning

Abstract

The fast communication networks are crucial to support the modern digital services like cloud computing, massive data transmissions and real time multimedia applications. Since network traffic is constantly increasing exponentially, a proper approach to managing congestion is required to ensure the delivery of information is stable, minimize delays, and efficiently use bandwidth. Conventional congestion control mechanisms tend to use systems that are based on fixed rules and thresholds, and may be unable to be flexible in highly dynamic network situations. A graphical congestion control model is intelligent based on a multi-agent Deep Q-Learning model in which the distributed agents are tasked with monitoring network conditions such as queue length, delay, packet loss, and available bandwidth. The agents are taught the best acting policies in traffic regulation by means of interaction with their network environment and dynamically change their rates of transmission to reduce congestion. Learning organization is decentralized and enhances adaptability and scalability within large network systems. In comparison to traditional methods that attained a throughput of 780-910 Mbps, 2.9-5.8% packet loss, and 84-120 ms end-toend delay, performance evaluation has shown to achieve better network performance of 960 Mbps throughput, 1.8 end-to-end delay, and 0.8% end-to end packet loss. These enhancements underscore the success of smart use of reinforcement learning methods in adaptive congestion control in high-speed networking settings.