Hybrid Deep Reinforcement Learning and Particle Swarm Optimization for Accelerated Multipath Routing in Congested SDN Environments

Network congestion remains a critical challenge in dynamic communication environments, often degrading data delivery performance and Quality of Service (QoS). This study proposes a Hybrid Deep Reinforcement Learning with Particle Swarm Optimization (DRL–PSO) model to adaptively optimize routing paths and mitigate congestion in Software Defined Networking (SDN). The proposed approach integrates the exploration capability of Deep Reinforcement Learning with the fast convergence characteristics of Particle Swarm Optimization to select optimal routes based on real-time network conditions. Simulations were conducted using realistic network topologies under three traffic scenarios: normal, moderate, and congested conditions. The proposed model was compared with several baseline methods, including Pure DRL, DDPG, PPO, DQN, Multi-Agent DRL (MARL), PSO-only routing, Shortest Path First (SPF), and Equal Cost Multi-Path (ECMP). The results show that Hybrid DRL–PSO achieves the lowest latency values of 15.2 ms, 34.8 ms, and 55.3 ms, as well as the highest throughput values of 9.45 Mbps, 6.34 Mbps, and 4.27 Mbps across the three scenarios. In addition, the model maintains low packet loss rates of 0.05%, 1.2%, and 8.5%, and jitter values of 4.3 ms, 9.2 ms, and 16.6 ms, respectively. The main novelty of this work lies in integrating PSO as a pre-selection mechanism to generate K-best candidate paths, reducing the DRL action space and accelerating learning convergence for QoS-aware multipath routing. This hybrid approach also demonstrates the practical potential of combining learning-based intelligence and optimization techniques for adaptive traffic management in real-world SDN infrastructures.