A National Parallelized Hybrid Activity/Agent-Based Demand Model to Characterize the Mobility of the United States

Abstract

The rise of emerging passenger-oriented mobility services, ranging from public multi-modal transportation systems to privately operated fleets of driverless vehicles, have the potential to completely revolutionize the current state of transportation throughout the world. To effectively and efficiently develop and implement these systems necessitates understanding the demand for transportation from both spatial and temporal lenses. Significant progress has been made in developing an activity-based model that generates a synthetic population complete with transportation needs, both spatial and temporal, disaggregated down to the individual level to provide a complete view of what a population’s daily transportation needs might look like on a typical day.

However, computational issues continue to impede the development and extension of this model. Modelling the billion trips taken by over 300 million citizens on a typical workday requires significant computational power and by extension, processing time. Most of the model's processing time is due to the large number of sequential modelling algorithms throughout the model. These sequential algorithms use a single processor to process a queue of tasks, one at a time, in a specified order. However, in many cases throughout the model, the ordering of this queue of tasks does not matter. Thus, each task can be run independently of the other tasks in the queue. By using parallel algorithms, whereby the queue of tasks is split up and run simultaneously, a significant reduction in processing time can be accomplished by simply increasing the number of processors working on the queue of tasks. This work seeks to analyze and understand the implications of implementing parallelized versions of the existing sequential algorithms within the model, especially with regards to the overall processing time.