Multi-Agent Deep Reinforcement Learning for Dynamic Routing in MANETs Using Graph Neural Networks

Mobile Ad Hoc Networks (MANETs) consist of decentralized wireless networks with dynamic topologies and frequent link failures which make achieving efficient routing extremely difficult. Unlike prior ML-based MANET routing approaches that rely on static heuristics or offline training, our method performs real-time GNN embedding updates and decentralized execution to adapt immediately to topology changes. The paper introduces a new routing protocol for MANETs that combines Multi-Agent Deep Reinforcement Learning (MADRL) with Graph Neural Networks (GNNs) to improve routing decision-making processes. Network nodes function independently as autonomous agents which apply GNN-based state representations for capturing both local and wide network topology data. Nodes use a centralized training and decentralized execution model to collectively optimize their routing strategies through real-time network condition assessment. The protocol underwent evaluation through comprehensive NS-3 simulations within different mobility settings together with varying network densities and traffic patterns. The proposed approach shows superior performance compared to standard routing protocols like AODV and DSR in essential metrics such as packet delivery ratio (90.1%–97.8%), end-to-end delay (42ms–80ms), as well as routing overhead (12–24 packets/s). Merging MADRL with GNNs gives MANETs the ability to learn and adapt dynamically while providing a scalable and efficient solution for real-world applications such as disaster response and military communication environments.