Modeling and Predicting Agent Trajectory in Urban Road Networks

Julien Baudru, Hugues Bersini

Track: Type C (Demonstration Abstracts)

Keywords: Urban Mobility, Next Location Prediction, Road Network Simulation, Deep Learning, Graph Neural Networks, Transformer Models, LSTM, Agent-Based Simulation.

Abstract: Accurate next-node prediction in road networks supports efficient carpooling, routing, and traffic management in urban systems. We developed an agent-based simulator of the Brussels road network to generate synthetic mobility data using five source-target selection patterns. Using this data, we evaluate LSTM, Transformer, and GNN models for next-node prediction. Results show that models perform best on Activity and Hub-and-Spoke datasets (around 94\% accuracy), while the Zone dataset poses greater challenges, especially for GNNs, highlighting the importance of dataset complexity over size in prediction accuracy.

Serve As Reviewer: ~Julien_Baudru1

Submission Number: 2