Investigation of a GNN approach to mitigate congestion in a realistic MANET scenario

Mobile Ad-hoc Networks (MANETs) can be modelled as time-varying graphs as their topology and traffic demands change. Optimizing routing in MANETs by proactively adapting the routes is a challenge, even with a fixed topology and user demand Omniscient Dijkstra Routing (ODRb) is one of the best known approaches, which computes alternative paths but has limitations to mitigate congestions. In this paper, we investigate Graph Neural Networks (GNNs) for routing optimization in MANETs. Our contribution is inspired by the centralized GNN-based Data Driven Routing (GDDR) framework developed by Hope [1]. GDDR was developed for optical fibre networks to support time varying user demands. After failing to obtain good results using the GDDR approach on the tactical Anglova MANET scenario, we adapted GDDR to minimize the maximum number of traversals. Our GNN-t proposal is able to find alternative longer paths mitigating congestion on central nodes. Considering a challenging static topology, the first second of the 24-node Anglova scenario: GNN-t achieves a Completion Ratio of CR=99% for a traffic of acked-messages averaging 1msg/s/node and CR=77% when the traffic is doubled (2msg/s/node). For the challenging first 300s of Anglova CP1, similar performance is reported for ODRb and GNN-t: CR=81% without fading and CR=54% with fading.