Institute of Transportation Studies at UC Berkeley

This project developed a quantile regression method for predicting future traffic flow at a signalized intersection by combining both historical and real-time data. The algorithm exploits nonlinear correlations in historical measurements and efficiently solves a quantile loss optimization problem using the Alternating Direction Method of Multipliers (ADMM). The resulting parameter vectors allow determining a probability distribution of upcoming traffic flow. These predictions establish an efficient, delay-minimizing control policy for the intersection. The approach is demonstrated on a case study with two years of high resolution flow measurements. It is emphasized that the results are applicable to any traffic intersection equipped with sensors that provide sufficiently high resolution of data acquisition. In particular, the data must have sufficient spatial resolution, e.g., measuring turning counts, and sufficient temporal resolution, e.g., measurements each 15 minutes. For example, numerous sites in California, including a large number of intersections in LA County, possess sensors that provide the required data to a central server.