Tanaka, Shuntaro (The University of Tokyo), Fukuda, Daisuke (The University of Tokyo)

Development of a Deep Learning-Based Hybrid Calibration Method for Large-Scale Activity-Based Travel Simulator Using Passive Aggregate Population Data

Scheduled for presentation during the Regular Session "S39c-Data-Driven Optimization in Intelligent Transportation Systems" (FR-LA-T39), Friday, November 21, 2025, 16:00−16:20, Coolangata 3

2025 IEEE 28th International Conference on Intelligent Transportation Systems (ITSC), November 18-21, 2025, Gold Coast, Australia

This information is tentative and subject to change. Compiled on October 18, 2025

Abstract

Activity-based models (ABMs) have been applied to analyzing large-scale and detailed human mobility patterns. However, previous studies have highlighted that the ABM simulator faces challenges in accurately reproducing real-world conditions. In this study, we integrated a deep learning model, a variational autoencoder, with an ABM simulator to perform large-scale parameter calibration in a single step, without modifying the aggregate population data. This integration enabled the development of a calibration framework that updates specific model parameters. To allow learning using the backpropagation method in deep learning, we employed a differentiable sampling method from a multinomial distribution using the Straight-Through Gumbel Softmax trick. We implemented a selection behavior process based on the utility functions of each option within the simulator. We conducted a case study using a submodel of an existing ABM simulator for the Tokyo Metropolitan Area and real-world aggregated population statistics, and showed that it can be applied to calibrate large-scale ABM simulators.