MRLRO: Meta-Reinforcement Learning Based Segment Routing Optimization

Segment Routing over IPv6 (SRv6) has been considered as a promising and efficient traffic engineering (TE) technology. However, traditional SR-TE solutions primarily fo-cus on improving the performance of long-term optimization strategies, while neglecting adaptability to dynamic environments. Some solutions place greater emphasis on the ability to adjust routing strategies in real-time for practical applications, but they suffer from slow adjustment speeds and high data col-lection costs. To address these challenges, this paper proposes a meta-reinforcement learning based SRv6 routing optimization algorithm (MRLRO). During the pre-training phase, MRLRO leverages multi-task training and Kullback-Leibler divergence to enhance the ability of agent to quickly learn and adapt to new tasks. With the fine-tuning operations of the agent, MRLRO can interact rapidly with the environment and adjust routing strategies under traffic bursts. Once the network stabilizes, the agent utilizes the experience accumulated during the pre-training phase to make routing strategies. Additionally, MRLRO involves the generalization of graph neural networks to improve its robust-ness. Simulation results demonstrate that MRLRO significantly outperforms existing reinforcement learning algorithms under dynamic traffic matrices with traffic bursts.