Analyzing and Influencing Traffic Networks with Mixed Autonomy

As publicly available cars gain semi-autonomous capabilities and companies promise to deliver fully autonomous vehicles in the near future, it is important to understand the effects that these vehicles will have on traffic networks. If all vehicles on the road are autonomous and are properly coordinated, road capacity and intersection throughput can be greatly increased. However, in the medium-term future roads will exist in a state of mixed autonomy, meaning they will be shared between human-driven and autonomous vehicles. The properties of this regime are not yet well-established. Preliminary studies show that even if autonomous vehicles can increase road capacity, converting human-driven vehicles to autonomous vehicles can increase the overall delay experienced by travelers. This phenomenon stems from the nature of this sociotechnical system, containing humans who are making selfish routing choices rather than choices that are socially optimal. Accordingly, we must understand the nature of these choices in order to design efficient mixed autonomous traffic networks.This thesis theoretically characterizes areas of inefficiency and opportunity in mixed autonomy and offers control algorithms with theoretical guarantees or benchmarks. Specifically, it studies problems including 1) formulating a framework for understanding and bounding the inefficiency due to selfish routing, 2) establishing toll structures which can decrease or eliminate this inefficiency, including tolls which differentiate between the different types of vehicles to varying degrees, 3) actively learning human preferences for time/money tradeoffs to determine optimal pricing for an autonomous ride-hailing service in the presence of human drivers who minimize their travel time, 4) decongesting roads by using Reinforcement Learning to route autonomous vehicles in a way which leverages the reaction of the surrounding uncontrolled human-driven vehicles, and 5) designing a controller for the low-level actions of autonomous vehicles to form platoons in the presence of vehicles driven by humans. In all these settings, we provide theoretical guarantees or benchmarks so that a system designer can intelligently choose whether to and how to implement each suggested scheme. In this fashion, we can ensure that autonomous vehicles will not worsen, and will rather improve, the performance of transportation networks in the years to come