CONGESTION MITIGATION IN TRANSMISSION CONTROL PROTOCOL/INTERNET PROTOCOL NETWORKS USING A DEEP REINFORCEMENT LEARNING APPROACH

Abstract

Random Early Detection (RED) algorithms require manual parameter tuning of the queue weight parameter, limiting effectiveness in dynamic network environments. This research develops a Deep Reinforcement Learning approach to dynamically adjust RED's queue weight for improved congestion control. Using NS-3.43 with OpenAI Gym and ZeroMQ integration, three agents — DQN, DDQN, and DDQNPER — were trained on a 200-node scenario and evaluated on unseen traffic scales of 20, 100, 300, and 400 nodes. All DRL models-maintained throughput identical to standard RED (4.885–4.936 Mbps) across all scenarios, confirming no loss of network capacity. DDQNPER achieved the best overall performance, reducing packet loss by approximately 1.8% under heavy traffic conditions (300 and 400 nodes), while marginally underperforming RED in packet loss at 20 nodes. Queuing delay was reduced across all scenarios, with the largest improvement of approximately 11% occurring at 20 nodes and moderate reductions of 3–5% under heavier traffic. These results demonstrate that DDQN enhanced with Prioritised Experience Replay can meaningfully improve active queue management in TCP/IP networks, with advantages most consistent under high-traffic congestion conditions.