Оптимізація одношляхової маршрутизації OSPF за допомогою ші-моделей

Abstract

This study examines intelligent single-path routing in OSPF, evolving from MLP to GNN (GAT) and DRL. Formulated as a Boolean programming problem, routing was solved with an improved MLP, reaching 90%+ accuracy. To address topological limits, a hybrid GAT+MLP achieved 91,4% accuracy and zero-shot transfer to new networks. A pre-trained DRL agent showed promise for autonomous traffic management, raising rare-route recall to 80%. Despite GNN and DRL advances, MLP remains optimal for local routers due to microsecond decision speed. Integrating deep learning into OSPF enables dynamic interface cost optimization and congestion prevention.