Ph.D. Dissertation Defense

Machine Learning Solutions for Transportation Networks

Tomas Singliar (CS Grad/Pitt)

Thursday, November 13, 2008
10:00am - SENSQ 5317 - 5th Floor Conference Room

Abstract

This thesis brings a collection of novel models and methods that result from a new look at practical problems in transportation through the prism of the newly available datasets. The models reflect the complexity of interactions in road networks and their inherently stochastic nature. However, they are not confined to the specific domain and contribute to machine learning topics as diverse as modeling continuous multivariate distributions, robust planning under uncertainty and Bayesian network inference.

We design a generative probabilistic graphical model to describe multivariate continuous densities such as observed traffic patterns. The model implements a multivariate normal distribution with covariance constrained in a natural way, using a number of parameters that is only linear in the dimensionality of the data. The primary use in transportation management we envision for such a model is to impute data missing due to sensor malfunctions that are frequent in practice.

We build a model of traffic flow rooted in macroscopic flow models. Unlike traditional such models, our model deals with uncertainty of measurement and unobservability of certain important quantities and incorporates on-the-fly observations more easily. Because the model does not lend itself easily to exact inference, we develop a particle filter for inference.

Two new optimization algorithms for the common task or vehicle routing are designed, using the traffic flow model as their probabilistic underpinning. Both are intimately tied to the probabilistic dynamic model of traffic flow.

Finally, we present a new method for detecting accidents and other adverse events. Data collected from highways enables us to bring supervised learning approaches to incident detection. We show that a support vector machine learner can outperform manually calibrated solutions. A major hurdle to performance of supervised learners is the quality of data which contains systematic biases varying from site to site. We build a dynamic Bayesian network framework that learns and rectifies these biases, leading to improved supervised detector performance with little need for manually tagged data.

Dissertation Adviser

Prof. Milos Hauskrecht, Department of Computer Science

Committee Members

Prof. Gregory Cooper, Department of Biomedical Informatics
Prof. Rebecca Hwa, Department of Computer Science
Prof. Liz Marai, Department of Computer Science
Prof. Geoff Gordon, Department of Machine Learning, CMU