In recent years, machine learning has enabled tremendous advances in urban planning and traffic management. However, as transportation systems become increasingly complex, due to factors like increased traveler and vehicle connectivity and the evolution of new services (e.g., ride-sharing, car-sharing, on-demand transit), finding solutions continues to be difficult. To better understand these challenges, cities are developing high-resolution urban mobility simulators, called “digital twins”, that can provide detailed descriptions of congestion patterns. These systems incorporate a variety of factors that might influence traffic flow, such as available mobility services, including on-demand rider-to-vehicle matching for ride-sharing services; network supply operations, such as traffic-responsive tolling or signal control; and sets of diverse traveler behaviors that govern driving style (e.g., risk-averse vs. aggressive), route preferences, and travel mode choices.
These simulators tackle a variety of use cases, such as the deployment of electric-vehicle charging stations, post-event traffic mitigation, congestion pricing and tolling, sustainable traffic signal control, and public transportation expansions. However, it remains a challenge to estimate the inputs of these simulators, such as spatial and temporal distribution of travel demand, road attributes (e.g., number of lanes and geometry), prevailing traffic signal timings, etc., so that they can reliably replicate prevailing traffic patterns of congested, metropolitan-scale networks. The process of estimating these inputs is known as calibration.
The main goal of simulation calibration is to bridge the gap between simulated and observed traffic data. In other words, a well-calibrated simulator yields simulated congestion patterns that accurately reflect those observed in the field. Demand calibration (i.e., determining the demand for or popularity of a particular origin-to-destination trip) is the most important input to estimate, but also the most difficult. Traditionally, simulators have been calibrated using traffic sensors installed under the roadway. These sensors are present in most cities but costly to install and maintain. Also, their spatial sparsity limits the calibration quality because congestion patterns go largely unobserved. Moreover, most of the demand calibration work is based on single, typically small, road networks (e.g., an arterial).
In “Traffic Simulations: Multi-City Calibration of Metropolitan Highway Networks”, we showcase the ability to calibrate demand for the full metropolitan highway networks of six cities — Seattle, Denver, Philadelphia, Boston, Orlando, and Salt Lake City — for all congestion levels, from free-flowing to highly congested. To calibrate, we use non-sparse traffic data, namely aggregated and anonymized path travel times, yielding more accurate and reliable models. When compared to a standard benchmark, the proposed approach is able to replicate historical travel time data 44% better on average (and as much as 80% better in some cases).