Using big data and machine learning to rank traffic signals in Tennessee

Abstract

This paper discusses low-cost approaches capable of ranking traffic intersections for the purpose of signal re-timing. We extracted intersections that are comprised of multiple roads, defined by alphanumeric traffic message channel segment codes per international classification standards. Each of these road segments includes a variety of metrics, including congestion, planning time index, and bottleneck ranking information provided by the Regional Integrated Transportation Information System. Our first approach was to use a ranking formula to calculate intersection rankings using a score between 0 and 10 by considering data for different times of the day and different days of the week, weighting weekdays more heavily than weekends and morning and evening commute times more heavily than other times of day. The second method was to utilize unsupervised machine learning algorithms, primarily k-means clustering, to accomplish the intersection ranking task. We first approach this by checking the performance of basic k-means clustering on our data set. We then explore the ranking problem further by utilizing data provided by traffic professionals in the state of Tennessee. This exploration involves using MATLAB to minimize the mean-squared error of intersection rankings to determine the optimum weights in the ranking formula based on a city's professional data. We then attempted an optimization of our weights via a brute-force search approach to minimize the distance from ranking formula results to the clustering results. All the ranking information was aggregated into an online SQL database hosted by Amazon web services that utilized the PHP scripting language.