Trajectory Regression for Travel-Time Prediction

Abstract

We propose a new method for predicting the travel-time along an arbitrary path between two locations on a map. Unlike traditional approaches, which focus only on particular links with heavy traffic, our method allows probabilistic prediction for arbitrary paths including links having no traffic sensors. We introduce two new ideas: to use string kernels for the similarity between paths, and to use Gaussian process regression for probabilistic prediction. We test our approach using traffic data generated by an agent-based traffic simulator.