Efficient congestion control in communications using novel weighted ensemble deep reinforcement learning

In this paper, we introduce Deep Reinforcement Learning (DRL) for congestion control in the Transmission Control Protocol/Internet Protocol (TCP/IP) networks. We propose a weighted ensemble DRL model that combines four DRL models Deep Q-Learning (DQN), Proximal Policy Optimisation (PPO), Deep Deterministic Policy Gradient (DDPG), and Twin Delay DDPG (rlTD3). These four models are designed with varying action spaces for efficient congestion control in an Active Queue Management (AQM) system. The proposed model outperformed single DRL models and established congestion control algorithms like Random Early Detection (RED) in normal and stress testing. The proposed model improved throughput by 4% and delays by 10.52% compared to DQN and 2.92% compared to DDPG. The proposed model has shown promising results in managing congestion in dynamic network environments and handling high-traffic loads.