TopoFormer: An Efficient Link-Set Prediction Architecture for Ad Hoc Network Topology Generation

Félix Marcoccia, Victor Fagoo, Gilles Monzat de Saint Julien, Cédric Adjih, Thomas Watteyne, Paul Mühlethaler

Keywords: networks, topology, attention, transformer, encodings, graphs

Abstract: In this paper, we present TopoFormer, a powerful architecture for predicting links between communication nodes in mobile networks. The goal is to imitate, in real time, the results of a costly combinatorial algorithm that generates topologies for networks with directional antennas. These antennas offer excellent performance but require complex, interdependent steering decisions in real time. Our Transformer-based architecture is enhanced with efficient components that add useful inductive biases, making it suitable for environments where scaling is limited. A key contribution is the introduction of directional density encodings, which help the attention mechanism better separate nodes in dense clusters. Equipped with our modules, a single Transformer block of dimension 12 achieves over 95 % accuracy, reducing the gap to optimality by half compared to a plain 1-block Transformer while requiring only 12 % more computation. Using two blocks, the model comes close to perfect accuracy.

Submission Type: Full paper proceedings track submission (max 9 main pages).

Submission Number: 50