Ph.D. Thesis Defense Announcement

Identification of Traffic Congestion Precursors using Machine Learning Approaches

By

Nishu Choudhary

Advisor(s):

Dr. Michael P. Hunter (CEE) & Dr. Angshuman Guin (CEE)

Committee Members:

Dr. Michael O. Rodgers (CEE), Dr. Randall L. Guensler (CEE), and Dr. Christos Alexopoulos (ISyE)

Date & Time: 

June 30 2023, 2:00 PM

Location: 

(Hybrid) SEB 122 & Zoom https://gatech.zoom.us/j/92423394875

Traffic congestion is an everyday reality for drivers and has been on an upward trend for the last couple of decades. A promising frontier in congestion mitigation is proactive traffic management, which involves the prediction of traffic conditions and events through emerging technologies such as Machine Learning (ML) and Deep Learning (DL). This study presents an approach for using ML algorithms to predict imminent traffic congestion. The work also explores how site characteristics can influence the transferability and applicability of the method for congestion prediction. 

To predict impending traffic state degradation, potential congestion precursors are treated as a binary classification problem. Under this formulation, traffic in an uncongested state, characterized by a set of feature vectors that reflect the lane dynamics and spatiotemporal state of the traffic, is labeled as belonging to one of two classes: free-flow or pre-congestion. This approach was tested using a set of classical ML classifiers on two different freeway sections within the Atlanta metro area. Both sites exhibit recurrent evening peak hour congestion, but differ in terms of bottleneck locations within each site. These location differences were shown to influence the frequency of false positives and precision scores for each site. This inference has important implications for future training of ML algorithms for congestion prediction by suggesting that the underlying mechanism for congestion generation (e.g., bottleneck activation or congestion propagation) must be considered in ML algorithm training. 

It was observed that traditional ML performance metrics were clearly inadequate for comprehensively assessing the ability of a method to predict future congestion. Therefore, a new performance index was developed to assess the usefulness of this category of algorithms in real-world applications. The proposed index correlates the density of observed precursors (i.e., traffic data classified by the trained ML as containing precursors to the onset of congestion) of traffic congestion in a traffic data stream with the likelihood of its onset. The index was developed and tested on traffic data from the Atlanta, Georgia area. The empirical results obtained demonstrate the index's potential for predicting traffic congestion.