In this post, we are sharing the poster presentation we carried out at TRB, one of the biggest conferences in the transportation domain. We presented the work done on automating the calibration of hundreds of behavioral parameters for an activity-based model, through Bayesian Optimization. By approximating the predictions of the Random Forest surrogate model as a parametric normal probability distribution and by applying the limited memory BFGS algorithm with box constraints, the proposed methodology is able to calibrate 477 parameters while avoiding local optima and reaching satisfactory results. The daily activity schedules and trip chains modeled with the calibrated SimMobility model replicate all the important patterns we selected as benchmarks.
The case study is the city of Tallinn (https://www.finestcentre.eu/post/openly-available-dataset-synthetic-population-of-the-city-of-tallinn) and both the model and the code are openly available on the existing github folder (please refer to the link).
Authors: Serio Agriesti, Vladimir Kuzmanovski, Jaakko Hollmén, Claudio Roncoli and Bat-hen Nahmias-Biran