Title: Deep Learning Predictors for Traffic Flows

Abstract: We develop a deep learning predictor for modeling traffic flows. The challenge arises as traffic flows have sharp nonlinearities resulting from transitions from free flow to breakdown and then to congested flow. Our methodology uses a deep learning architecture to capture nonlinear spatio-temporal flow effects. We show how traffic flow data, from road sensors, can be predicted using deep learning. For comparison, we use a benchmark sparse $\ell_1$ trend filter and we illustrate our methodology on traffic data from Chicago's Interstate I-55. We focus on forecasting traffic flows during special events, such as a Chicago Bears football game and an extreme snowstorm, where the sharp traffic flow regime can occur very suddenly and hardly predictable from historical patterns. Finally, we discuss directions for future research.