skip to primary navigationskip to content

Paul Waddell: Abstract and background reading

Integrated urban models have made numerous compromises while seeking to remain practically applicable to support actual planning decisions.  Two common stress points in modeling are computational speed and accuracy, ideally measured longitudinally as well as spatially, and with data held out from the model building stage.  Computational performance is critical in order for models to be practically useful in planning contexts, and to allow running sufficient number of scenarios, or to run uncertainty analysis, or both.  Computational performance improvements are also necessary in order to increase the spatial resolution of of locations and networks within models to better support modeling of active modes and transit access.

This talk describes ongoing work to overcome existing barriers to improving computational performance and model accuracy within a suite of integrated models, leveraging machine learning and GPU computing.  The goals include enabling much more rapid construction and calibration of models that attain high quality longitudinal validation, and reducing run times of models dramatically, while increasing their spatial resolution. We leverage GPU processing to parallelize microsimulation of traffic flows on large scale networks at the metropolitan scale, achieving high computational performance along with promising validation results.  We have begun using machine learning to improve predictive accuracy of price models to seed the structural microsimulation models of demand and supply in housing markets. And we have been developing a differentiable programming approach to building microsimulation models that enable calibrating the models longitudinally using machine learning algorithms, achieving high quality longitudinal validation by tuning the model parameters.



Microsimulation Analysis for Network Traffic Assignment (MANTA) at Metropolitan-Scale for Agile Transportation Planning:

Architecture for modular microsimulation of real estate markets and transportation:

A Comparison of Statistical and Machine Learning Algorithms for Predicting Rents in the San Francisco Bay Area: