Deep Reinforcement Learning for Routing in Uncertain DTNs with Graph Neural Networks

We address the routing optimization problem in Delay/Disruption Tolerant Networks under uncertain contact plans, where contact time bounds are known but link availability is probabilistic and subject to failures. We propose GROGU (Graph-based Routing for Optimizing under GNN-driven Uncertain DTNs), a novel decentralized routing algorithm that models forwarding decisions as a partially observable Markov decision process and leverages Deep Reinforcement Learning with a Graph Neural Network architecture. GROGU learns a single, reusable routing policy deployable across all nodes, using Monte Carlo-driven simulated experiences to maximize the successful delivery probability while adapting to topology changes such as node or link failures without retraining. We evaluate GROGU on synthetic and realistic satellite network topologies, comparing it against state-of-the-art methods including RUCoP, Lightweight Scheduler Sampling, Q-learning, and Contact Graph Routing. Results show that GROGU achieves competitive delivery performance, robust adaptability to unforeseen link outages, and fast convergence, albeit at higher training costs.