Learning-based local routing decisions in sparse aeronautical communication networks

Geographic greedy routing provides the scalability required for future L-band Digital Aeronautical Communications System (LDACS) Air-to-Air (A2A) networks. Yet, its topology blindness becomes a critical limitation in sparse airspaces. This blindness induces a path-dependent process, recently characterized as the “memory effect,” where suboptimal local choices steer packets toward dead ends multiple hops away. These failures are not arbitrary: they follow learnable patterns embedded in the local network structure. We propose a decentralized, densityadaptive routing scheme that replaces standard geographic progress toward the destination with learned neighbor ranking. The model is trained using the Topological Advance (TA) metric as a ground-truth label to indicate whether a chosen next-hop reduces the shortest-path hop distance to the destination. By predicting TA from three-hop local features, nodes improve routing reliability. Evaluations in realistic French airspace scenarios show a success ratio exceeding 0.93 over topologically connected nodes in sparse networks, a 33% improvement over traditional greedy routing, while maintaining a near-optimal hop stretch of 1.03.