Machine learning surrogates for agent-based models in transportation policy analysis

Abstract

Effective traffic policies are crucial for managing congestion and reducing emissions. Agent-based transportation models (ABMs) offer a detailed analysis of how these policies affect travel behaviour at a granular level. However, computational constraints limit the number of scenarios that can be tested with ABMs and therefore their ability to find optimal policy settings.

In this proof-of-concept study, we propose a machine learning (ML)-based surrogate model to efficiently explore this vast solution space. By combining Graph Neural Networks (GNNs) with the attention mechanism from Transformers, the model predicts the effects of traffic policies on the road network at the link level.

We implement our approach in a large-scale MATSim simulation of Paris, France, covering over 30,000 road segments and 10,000 simulations, applying a policy involving capacity reduction on main roads. The ML surrogate achieves an overall $R^2$ of 0.91; on primary roads where the policy applies, it reaches an $R^2$ of 0.98. This study shows that the combination of GNNs and Transformer architectures can effectively serve as a surrogate for complex agent-based transportation models with the potential to enable large-scale policy optimization, helping urban planners explore a broader range of interventions more efficiently.