Dynamic Routing in Challenged Networks with Graph Neural Networks

Abstract

Challenged networks are characterized by a time-varying operational environment that constrains the optimality of standard routing algorithms. This work investigates a deep learning method that tackles the bundle routing problem by taking advantage of the available network metrics and performance data through a graph neural network (GNN). A cognitive routing decision unit is formulated by defining a GNN structure that accepts both edge and node input features, and that is trained with reinforcement learning. The GNN allows the inputs to be permutation invariant and independent of the network size and connectivity. Simulation results demonstrate that the proposed cognitive routing method is able to learn how to optimize the next-hop for each data bundle of a flow to achieve lower end-to-end delivery latency than the standard Contact Graph Routing algorithm. The GNN achieves the optimization by detecting and avoiding the extended wait times caused by both butter congestion and the stall times for the next contact when long link disruptions occur.