Deep Reinforcement Learning for Traffic Engineering in SDN

Traffic Engineering (TE) in SDN requires fast and adaptive routing to handle dynamic traffic demands. While DRL has been applied to path selection and weight optimization, existing methods face inefficiency and poor multi-objective balance. We propose a DRL-based link weight optimization method with a multi-sample generation mechanism that augments the replay buffer with Pareto-optimal samples, accelerating convergence and improving generalization. To further enhance efficiency, we introduce critical link selection to reduce the action space and design a multi-objective reward to balance utilization, delay, and stability. Experiments show our method consistently outperforms baselines, reducing maximum link utilization and delay.