Dual-Agent Reinforcement Learning for Adaptive Bandwidth Allocation and Congestion Control in SDN

Software-Defined Networking (SDN) is a modern network architecture that distinguishes the control plane from the data plane, enabling centralized and programmable network management through a centralized controller.Despite its advantages, SDN faces challenges similar to traditional networks, with congestion being a prominent issue.While most existing solutions primarily address congestion in the data plane, control plane congestion, often caused by the increasing number of networks devices and their orchestration demand, remains underexplored.In this paper, we propose a dual-agent reinforcement learning framework for adaptive bandwidth allocation, designed to address congestion in both the control and data planes of SDNs.Our approach employs Dueling Deep Q-Networks (DQN), with two specialized agents: one managing the control plane and the other the data plane.The evaluation is conducted in two phases.The first phase involves hyperparameter optimization, where we assess various optimizers, activation functions, and loss functions to identify the most effective configuration for congestion resilience.We found that RMSprop combined with Mean Squared Error (MSE) loss achieved the highest cumulative reward (over 600 across episodes) and the most stable decision-making behavior.The second phase presents a comparative performance analysis, where we benchmark Dueling DQN against other deep reinforcement learning models, Rainbow DQN, Advantage Actor-Critic (A2C), and Deep Deterministic Policy Gradient (DDPG), in both single-agent and dual-agent setups.Experimental results show that our dual-agent Dueling DQN achieves a throughput of ~30396 Mbps in the control plane and ~200 Mbps in the data plane, with packet loss as low as 0.014 and delay consistently around 0.101 ms.Compared to other models, it demonstrates superior reward convergence and throughput performance, while DDPG exhibits slightly lower packet loss but poorer reward stability.Our findings highlight the potential of dual-agent adaptive congestion control as a scalable and intelligent solution for traffic management in SDN environments.