Modeling and optimizing network infrastructure for autonomous vehicles
MetadataShow full item record
Autonomous vehicle (AV) technology has matured sufficiently to be in testing on public roads. However, traffic models of AVs are still in development. Most previous work has studied AV technologies in micro-simulation. The purpose of this dissertation is to model and optimize AV technologies for large city networks to predict how AVs might affect city traffic patterns and travel behaviors. To accomplish these goals, we construct a dynamic network loading model for AVs, consisting of link and node models of AV technologies, which is used to calculate time-dependent travel times in dynamic traffic assignment. We then study several applications of the dynamic network loading to predict how AVs might affect travel demand and traffic congestion. AVs admit reduced perception-reaction times through technologies such as (cooperative) adaptive cruise control, which can reduce following headways and increase capacity. Previous work has studied these in micro-simulation, but we construct a mesoscopic simulation model for analyses on large networks. To study scenarios with both autonomous and conventional vehicles, we modify the kinematic wave theory to include multiple classes of flow. The flow-density relationship also changes in space and time with the class proportions. We present multiclass cell transmission model and prove that it is a Godunov approximation to the multiclass kinematic wave theory. We also develop a car-following model to predict the fundamental diagram at arbitrary proportions of AVs. Complete market penetration scenarios admit dynamic lane reversal -- changing lane direction at high frequencies to more optimally allocate road capacity. We develop a kinematic wave theory in which the number of lanes changes in space and time, and approximately solve it with a cell transmission model. We study two methods of determining lane direction. First, we present a mixed integer linear program for system optimal dynamic traffic assignment. Since this program is computationally difficult to solve, we also study dynamic lane reversal on a single link with deterministic and stochastic demands. The resulting policy is shown to significantly reduce travel times on a city network. AVs also admit reservation-based intersection control, which can make greater use of intersection capacity than traffic signals. AVs communicate with the intersection manager to reserve space-time paths through the intersection. We create a mesoscopic node model by starting with the conflict point variant of reservations and aggregating conflict points into capacity-constrained conflict regions. This model yields an integer program that can be adapted to arbitrary objective functions. To motivate optimization, we present several examples on theoretical and realistic networks demonstrating that naïve reservation policies can perform worse than traffic signals. These occur due to asymmetric intersections affecting optimal capacity allocation and/or user equilibrium route choice behavior. To improve reservations, we adapt the decentralized backpressure wireless packet routing and P0 traffic signal policies for reservations. Results show significant reductions in travel times on a city network. Having developed link and node models, we explore how AVs might affect travel demand and congestion. First, we study how capacity increases and reservations might affect freeway, arterial, and city networks. Capacity increases consistently reduced congestion on all networks, but reservations were not always beneficial. Then, we use dynamic traffic assignment within a four-step planning model, adding the mode choice of empty repositioning trips to avoid parking costs. Results show that allowing empty repositioning to encourage adoption of AVs could reduce congestion. Also, once all vehicles are AVs, congestion will still be significantly reduced. Finally, we present a framework to use the dynamic network loading model to study shared AVs. Results show that shared AVs could reduce congestion if used in certain ways, such as with dynamic ride-sharing. However, shared AVs also cause significant congestion. To summarize, this dissertation presents a complete mesoscopic simulation model of AVs that could be used for a variety of studies of AVs by planners and practitioners. This mesoscopic model includes new node and link technologies that significantly improve travel times over existing infrastructure. In addition, we motivate and present more optimal policies for these AV technologies. Finally, we study several travel behavior scenarios to provide insights about how AV technologies might affect future traffic congestion. The models in this dissertation will provide a basis for future network analyses of AV technologies.