Narrowing-Down of Vehicles by Transformer-based Surrogate Model for Assignment Process in AI On-demand Transportation Service

AI ride-sharing services such as Smart Access Vehicle Service (SAVS) are expected to have high efficiencyin large-scale operations. However, computational complexity increases significantly under large-scale conditions,leading to performance issues. This paper proposes a method to improve scalability by using a Transformer-basedmodel to estimate evaluation values for vehicle selection in the allocation method used in SAVS. The model predictsthe delay for each vehicle when inserting a new demand into existing routes, and vehicle selection is performed basedon these predictions. Simulation experiments with up to 200 vehicles show that the proposed method reduces thecomputation time by approximately 36% while limiting the increase in average trip time to 0.9%, and that keepingthe extraction ratio at 20–30% of the fleet maintains dispatching quality. These results indicate that the proposedmethod provides a basis for scaling up real-time dispatching in AI on-demand transportation services.