Fairness Enhancement of TCP Congestion Control Using Reinforcement Learning
Sang-Jin Seo, You-Ze Cho · 2022
In TCP congestion control research, the use of machine learning to solve the issue of unused link bandwidth and to improve performance, such as maximizing link utilization or minimizing latency, is steadily increasing. Among such approaches, the Deep Q Network (DQN)-based TCP congestion control algorithm improves the link utilization but suffers from performance degradation when a specific link bandwidth is exceeded. In addition, inter-protocol fairness with other TCP congestion control algorithms has not been verified. In this paper, on a NS3 simulator, we conducted the experiments to enhance the improvement of the DQN-based TCP congestion control algorithm v2 in single flow and an inter-protocol fairness when several flows share the same bottleneck link. Our results confirmed that the average throughput was improved, and our approach is fairer than existing congestion control algorithms.