Urban traffic simulators are used by cities worldwide to inform their transportation planning and operations. We discuss the main challenges in the calibration of inputs for stochastic microscopic urban traffic simulators. In particular, we focus on the calibration of travel demand or origin-destination matrices (ODs). The problem is underdetermined, we discuss the implications of this for transportation practice and show how the use of higher-order moment information from field data can help mitigate the level of underdetermination. We propose a deep learning generative adversarial network (GAN) approach to this problem. The method readily extends to various types of traffic data and to the calibration of other input models and parameters, such as other travel behavioral models. We benchmark the approach on a Salt Lake City case study, considering vehicular count data. The analysis shows the added value of using higher-moment information for OD calibration.
Speaker: Dr. Carolina Osorio
Osorio is a Staff Research Scientist at Google Research. Osorio is also a faculty at HEC Montreal, where she holds the SCALE AI Research Chair in Artificial Intelligence for Urban Mobility and Logistics. Osorio has consulted for Alphabet’s Sidewalk Labs, and has collaborated with top private and public sector, transportation and supply chain stakeholders, including Zipcar, Ford Motor Company, New York City Department of Transportation, and SANDAG. Her work focuses on the design of AI and simulation-based optimization algorithms to tackle high-dimensional optimization problems. Osorio was recognized as an outstanding early-career engineer in the US by the National Academy of Engineering’s EU-US Frontiers of Engineering Symposium, is the recipient of an MIT Technology Review EmTech Colombia TR35 Award, and was on the GOOD 100 list of innovators in the category of “Minds That Are Hacking Our Surroundings for the Better”.
