Evaluation and comparison of calibration techniques for urban mobility behaviour ABM

Abstract

Models need to be calibrated to represent human mobility behaviour correctly. The literature identifies challenges in calibration since there are several approaches to calibration, but primarily dependent on the model purpose. This study addresses the lack of understanding of what makes a calibration method suitable specifically for mobility behaviour ABM. A data-driven approach is used, calibrating the proof-of-concept model of the EXPANSE project. An Amsterdam case study was selected with the ODiN dataset. An appropriate experimental calibration framework is presented, analysing model characteristics delivering an objective function and hierarchical level in the field of human mobility. The work also compared cutting-edge parameters search optimisation algorithms. Results hinted at surrogate model-based methodologies. Their performances outperform the other solutions on both fitness and computational load aspects. In addition, a metrics-based framework of how to compare calibration techniques for human mobility ABM is presented. Finally, the investigation also proved that a parameter dimensionality reduction method based on grouping does not bring any benefit compared to calibrating the entire parameter set. This project is just the first step in this field. The results are valuable, not only to the transport choice field only but generic to the main idea of mode choice. The established framework will be a profitable tool for all the researchers to be used in urban developments for a sustainable and healthier future.