The expected increase in urbanisation in 21st century, coupled with the socio-economic motivation for increasing mobility, is going to push the transport infrastructure well beyond its current capacity. In response, more stringent and intelligent control mechanisms are required to better monitor, exploit, and react to unforeseen conditions. With an increasing attention towards autonomous vehicles which can assist or even replace human drivers it is important to consider a number of challenges ranging from technological issues to road safety issues. Especially with increasing traffic intensity, especially in urban areas, maintaining high standards for the road safety for all users (e.g., traffic, bikes, pedestrians) is of the utmost importance along with minimisation of congestion and travel times and management of pollution.
Fortunately, a wide variety of data is becoming available to support novel innovations that improve urban mobility and transport. For example, data from traffic cameras, cell phone GPS data, ride-hailing services and even drones has significantly improved the quality and quantity of information available for intelligent real-time traffic solutions. Furthermore, modern (semi-)autonomous vehicles contain a plethora of sensing modalities as well as vehicle to vehicle as well as infrastructure communication abilities.
The aim of this workshop is to bring together researchers who leverage various AI techniques (e.g., Machine Learning, Automated Planning) for innovation in areas of Urban Mobility and Transportation.
Abstract: The recent rise of artificial intelligence and machine learning methods offers new alternatives to tackle the elusive problem of controlling congested urban networks. The talk discusses the case of controlling large signalized networks and shows that i) near-optimal policies can be obtained using random search, supervised learning and with deep reinforcement learning (DRL) methods trained with free-flow data, and ii) DRL methods cannot learn under congestion. We show that these results are a consequence of the "congested property of urban networks", whereby the network flow tends to be independent of the signal control policy under congestion; i.e. there is nothing to learn under congestion. Our findings imply that it is advisable for current DRL methods in the literature to discard any congested data when training, and that doing this will improve their performance under all traffic conditions. we also show the similar results can be obtained using an augmented state space, which adds a congestion tree to track congestion spillback. Our findings also suggest that future control methods based on machine learning techniques will have to be adapted to cope with the macroscopic properties of urban networks, which are only beginning to be understood.
Bio: Jorge Laval is a Professor at the School of Civil and Environmental Engineering since 2006. After obtaining his B.S. in Civil and Industrial Engineering from Universidad Catolica de Chile in 1995, Dr. Laval worked as a transportation engineer for 5 years at the Chilean Ministry of Public Works in Santiago, Chile. He received his Ph.D. in Civil Engineering from the University of California, Berkeley in 2004. Prior to joining Georgia Tech, Dr. Laval held two consecutive one-year postdoctoral positions at the Institute of Transportation Studies at UC Berkeley, and at the French National Institute for Safety and Transportation Research (INRETS/ENTPE). Professor Laval's main research thrust is in the area of traffic flow theory, modeling and simulation, focusing in understanding congestion in urban networks and how to manage it. He has made important contributions towards understanding the capacity of freeways, the connection between driver behavior and stop-and-go waves, freeway ramp-metering strategies, dynamic traffic assignment, congestion pricing and machine learning models for congestion control.
Abstract: In this presentation we will give an overview of the work done at the Delft Center for Systems and Control in the field of model-based traffic management and control for urban traffic networks. The main focus is on how model-based predictive control (MPC) can be used to obtain a balanced trade-off between reduction of total time spent, emissions, and fuel consumption in large-scale road traffic networks. We address several methods to deal with the computational complexity issues arising in model-based control of large-scale road traffic networks, such as choosing appropriate traffic flow models, using parametrized control, and adopting a multi-agent or multi-level control framework. We consider in particular on multi-level, multi-agent traffic control with coordination within and across all control levels. We explain how model predictive control can be used at several levels of the control hierarchy. The proposed multi-level architecture provides a scalable approach for control of large-scale traffic networks where at different levels of the hierarchy different temporal and spatial scales are taken into account.
Abstract: This talk investigates Lagrangian (mobile) control of traffic flow at local scale (vehicular level). The question of how will self-driving vehicles will change traffic flow patterns is investigated. We describe approaches based on deep reinforcement learning presented in the context of enabling mixed-autonomy mobility. The talk explores the gradual and complex integration of automated vehicles into the existing traffic system. We present the potential impact of a small fraction of automated vehicles on low-level traffic flow dynamics, using novel techniques in model-free deep reinforcement learning, in which the automated vehicles act as mobile (Lagrangian) controllers to traffic flow. Illustrative examples will be presented in the context of a new open-source computational platform called FLOW, which integrates state of the art microsimulation tools with deep-RL libraries on AWS EC2. Interesting behavior of mixed autonomy traffic will be revealed in the context of emergent behavior of traffic: https://flow-project.github.io/
Bio: Alexandre Bayen is the Liao-Cho Professor of Engineering at UC Berkeley; a Professor of both Electrical Engineering and Computer Science and Civil and Environmental Engineering. He is currently the Director of the Institute of Transportation Studies (ITS), and a Faculty Scientist in Mechanical Engineering, at the Lawrence Berkeley National Laboratory (LBNL). He received the Engineering Degree in applied mathematics from the Ecole Polytechnique, France, in 1998, the M.S. and Ph.D. in aeronautics and astronautics from Stanford University in 1999 and 2004, respectively. He was a Visiting Researcher at NASA Ames Research Center from 2000 to 2003, and January 2004 through December 2004, he worked as the Research Director of the Autonomous Navigation Laboratory at the Laboratoire de Recherches Balistiques et Aerodynamiques, (Ministere de la Defense, Vernon, France), where he holds the rank of Major.
Abstract: Rapid "urbanization" (more than 50% of worlds' population now resides in cities) coupled with the natural lack of coordination in usage of common resources (ex: bikes, ambulances, taxis, traffic personnel, attractions) has a detrimental effect on a wide variety of response (ex: waiting times, response time for emergency needs) and coverage metrics (ex: predictability of traffic/security patrols) in cities of today. Motivated by the need to improve response and coverage metrics in urban environments, I will talk about our research on building intelligent agent systems that make sequential decisions to continuously match available supply of resources to an uncertain demand for resources. Our broad approach to generating these sequential decision strategies is through a combination of machine learning and sequential optimization methods, while exploiting structure (anonymity, limited influence, abstraction etc.) in problems. I will specifically delve into our recent research on developing Deep RL methods that can handle operational constraints and show some of our recent results on real world transportation and emergency response datasets.