Machine Learning Approaches on a Travel Time Prediction Problem

This thesis concerns the prediction of travel times between two points on a map, based on a combination of link-scale road network data and historical trip-scale data. The main idea is that the predictions using the road network data can be improved by a correction factor estimated from historical trip data. The correction factor is estimated both using a Machine Learning approach, more specifically Gaussian Progress Regression, and a simple baseline method inspired by an approach in the literature. The Gaussian Progress Regression is performed using a string kernel and a squared exponential kernel. The Gaussian Process Regression using the string kernel outperforms both the baseline and the squared exponential kernel, and is hence the most promising approach on the considered problem.