Franklin Abodo

Florida International University Knight Foundation School of Computing and Information Sciences

Lecture Information:
  • June 24, 2022
  • 10:00 AM
  • Zoom

Speaker Bio

Franklin Abodo is a master’s student of computer science at Florida International University (FIU)’s Knight Foundation School of Computing and Information Sciences under the advisement of Dr. Leonardo Bobadilla. Franklin also works full-time as a Research Engineer at STR, where he has worked on algorithm design, software development, and machine learning model development in the areas of artificial intelligence for large-scale graph analytics, machine learning for anomaly detection, and multi-agent reinforcement learning for wargame simulation. Franklin previously interned as a Student Trainee Computer Scientist at the U.S. Department of Transportation’s Office of Research and Technology where he applied deep computer vision to a variety of video analysis tasks centered on improving the safety and operational efficiency of transportation systems. He has shared the products of his work with the community through publications and open-sourced code. Franklin earned a bachelor’s degree in computer science from FIU in 2016.


Traffic simulation software is used by transportation researchers and engineers to design, and evaluate the efficacy of, changes intended to improve roadway networks. Underlying these simulators are mathematical models of microscopic driver behavior from which macroscopic measures of flow and congestion can be recovered. Many models are intended to be applicable to only a subset of possible traffic scenarios and roadway configurations, while others do not have any explicit constraint on their applicability. Some scenarios/configurations exist for which no model invented to date has been shown to accurately reproduce realistic driving behavior. Of specific concern is the inability to produce useful simulations of work zones on highways, which makes it difficult to optimize for safety and other metrics when designing a work zone.

The Federal Highway Administration (FHWA) has commissioned the Volpe National Transportation Systems Center (Volpe) to develop a new car-following model for use in microscopic simulators that captures and reproduces driver behavior equally well within and outside of work zones; the FHWA Work Zone Driver Model (WZDM). Toward this end, Volpe performed a naturalistic driving study (NDS) in which it collected telematic data from vehicles being driven along routes that included work zones on highways and urban roads. This telematic data included variables relevant to the car-following model’s prediction task: given a simulator’s state at time t, predict the driver’s choice of acceleration at time t + 1. This data would be used to calibrate and validate the proposed model.

During model development, Volpe researchers observed difficulties in calibrating their model, leaving them to question whether there existed flaws in their model, in the data, or in the procedure used to calibrate the model using the data. In this thesis, I use Bayesian methods for data analysis and parameter estimation to explore and, where possible, address these questions.

First, I use Bayesian inference to measure the sufficiency of the size of the NDS data set by exploiting the fact that as a data set’s size increases, the influence that the choice of prior for a parameter has on inference decreases. Second, I compare the procedure and results of the genetic algorithm-based calibration performed by the Volpe researchers with those of calibration based on Bayesian inference. Difficulties encountered when using a genetic algorithm to automate calibration forced the researchers to revert to guess-and-checking parameter values. Third, I explore the benefits of modeling car-following hierarchically, with the total population of car-following instances being grouped by driver and assigned driver-specific parameters whose means and variances are drawn from common global distributions. Finally, I apply what was learned in the first three phases, which are conducted using a previously invented car-following model that is already widely used, to the probabilistic modeling, calibration and validation of the WZDM. Validation is performed using information criteria as an estimate of predictive accuracy, and also using observations of vehicle movements in simulations that use the parameter values resulting from calibration. The results of both forms of validation are directly compared to those of a third model, Wiedemann ’99, which is implemented in the same simulation software as the WZDM.