Cross-city generalization of machine learning surrogates for road capacity reductions via transfer learning

Agent-based transport models (ABM) can model traffic policies with high fidelity, but their computational cost makes systematic exploration of large intervention spaces infeasible. Recent work shows that machine learning surrogates can approximate ABM outcomes at link level in milliseconds, enabling rapid evaluation and optimization over large intervention spaces. However, a key practical barrier remains: accurate surrogates typically require thousands of computationally expensive ABM simulations for training. We investigate whether cross-city transfer learning can reduce this requirement for interventions that reduce road capacity by 50% on selected major roads. We further assess whether this enables a reusable multi-city “surrogate database” that can be fine-tuned to new target cities with minimal additional simulation effort. We propose a transferable, intervention-conditioned surrogate framework with a hybrid GNN-Transformer architecture, pre-trained on multiple source cities and fine-tuned in a few-shot manner on a target city. Using MATSim simulations across several small- and mid-sized cities, we show that few-shot fine-tuning consistently outperforms training from scratch across all target cities and target-city data budgets. Performance improves further with the number and diversity of source cities, and we analyze factors driving strong transfer. We also evaluate decision-relevant performance in the intended workflow: ranking candidate interventions under objective and feasibility constraints, then validating only a small elite subset with the ABM. The surrogate closely reproduces the ABM’s top-ranked scenarios, is orders of magnitude faster, and reduces simulation core-hour requirements by up to 99% relative to training a separate surrogate per city.