Learning Decomposition Optimization for WAN Traffic Engineering With Graph Attention Networks

With the rapid expansion of Wide Area Networks (WANs), traffic engineering (TE) faces increasing challenges in optimizing network traffic to ensure high-bandwidth, low-latency services. Existing reinforcement learning (RL)-based TE methods improve computational efficiency by directly approximating traffic allocation solutions; however, their models tend to suffer from convergence difficulties as network scale increases. In this paper, we propose a decomposition-based optimization framework leveraging Graph Attention Networks (GAT2) to achieve scalable TE optimization in large-scale networks. GAT2 is trained using an RL algorithm to learn a decomposition policy that leverages network topology, traffic demands, and link constraints to partition a TE problem into multiple subproblems. These subproblems are then solved in parallel using linear programming solvers, and their solutions are combined to obtain the final TE solution. In the experiments, our method outperformed POP, NCFlow, ADMM, and TEAL on Internet Topology Zoo topologies, with improvements in satisfied demand ranging from 0.95% to 5.66%. In scenarios involving larger network topologies, link failures, and traffic demand fluctuations, our model outperformed advanced methods, highlighting its robustness and scalability in dynamically adapting traffic allocation to varying network environments.