Deep Reinforcement Learning Models for Traffic Flow Optimization in SDN Architectures

Abstract

Deep reinforcement learning (DRL) has emerged as a promising approach to dynamic traffic engineering in software-defined networks (SDN). In this work, we evaluate three popular DRL agents—Deep Q-Network (DQN), Asynchronous Advantage Actor-Critic (A3C), and Proximal Policy Optimization (PPO)—on simulated SDN routing tasks. Using a Mininet emulated network with a Ryu controller and TensorFlow-based agents, we compare DRL models against traditional baselines (shortest-path routing and equal-cost multi-path (ECMP)). The DRL agents learn to select routes based on observed link loads and flow queues, with rewards reflecting combined throughput, latency, and packet loss. Our simulations show that all DRL methods significantly outperform fixed routing baselines: for example, a PPO-based agent reduced average flow latency by ≈20% and packet loss by ≈25% relative to shortest-path routing. PPO and A3C converged faster and to higher rewards than DQN, likely due to their on-policy and parallel learning designs. We provide a detailed comparison of algorithm characteristics, training stability, and network metric outcomes. The results highlight each model’s strengths: PPO’s stability and sample efficiency, A3C’s parallelism and multi-agent potential, and DQN’s simplicity. We critically discuss limitations such as training overhead and convergence variance. Finally, we outline future directions for improving real-world SDN traffic control with DRL, including transfer learning across topologies, online continual learning, and multi-agent coordination.