Simulation and Learning for Urban Mobility: City-scale Traffic Reconstruction and Autonomous Driving

Traffic congestion has become one of the most critical issues worldwide. The costs due to traffic gridlock and jams are approximately $160 billion in the United States, more than £13 billion in the United Kingdom, and over one trillion dollars across the globe annually. As more metropolitan areas will experience increasingly severe traffic conditions, the ability to analyze, understand, and improve traffic dynamics becomes critical. This dissertation is an effort towards achieving such an ability. I propose various techniques combining simulation and machine learning to tackle the problem of traffic from two perspectives: city-scale traffic reconstruction and autonomous driving. Traffic, by its definition, appears in an aggregate form. In order to study it, we have to take a holistic approach. I address the problem of efficient and accurate estimation and reconstruction of city-scale traffic. The reconstructed traffic can be used to analyze congestion causes, identify network bottlenecks, and experiment with novel transport policies. City-scale traffic estimation and reconstruction have proven to be challenging for two particular reasons: first, traffic conditions that depend on individual drivers are intrinsically stochastic; second, the availability and quality of traffic data are limited. Traditional traffic monitoring systems that exist on highways and major roads can not produce sufficient data to recover traffic at scale. GPS data, in contrast, provide much broader coverage of a city thus are more promising sources for traffic estimation and reconstruction. However, GPS data are limited by their spatial-temporal sparsity in practice. I develop a framework to statically estimate and dynamically reconstruct traffic over a city-scale road network by addressing the limitations of GPS data. Traffic is also formed of individual vehicles propagating through space and time. If we can improve the efficiency of them, collectively, we can improve traffic dynamics as a whole. Recent advancements in automation and its implication for improving the safety and efficiency of the traffic system have prompted widespread research of autonomous driving. While exciting, autonomous driving is a complex task, consider the dynamics of an environment and the lack of accurate descriptions of a desired driving behavior. Learning a robust control policy for driving remains challenging as it requires an effective policy architecture, an efficient learning mechanism, and substantial training data covering a variety of scenarios, including rare cases such as accidents. I develop a framework, named ADAPS (Autonomous Driving via Principled Simulations), for producing robust control policies for autonomous driving. ADAPS consists of two simulation platforms which are used to generate and analyze simulated accidents while automatically generating labeled training data, and a hierarchical control policy which takes into account the features of driving behaviors and road conditions. ADAPS also represents a more efficient online learning mechanism compared to previous techniques, in which the number of iterations required to learn a robust control policy is reduced.Doctor of Philosoph