Learning to Trade Off Performance and Safety in Mixed Autonomy Traffic

Abstract

With the advent of autonomous vehicles (AVs), and with the slow but steady consumer adoption of AVs on road networks, there is a newfound need to study the interactions between efficient traffic flow and driving safety in mixed autonomy traffic. Extending from reinforcement learning methods in robotic control methods and from learning methods for location-based actuators like traffic lights, this thesis considers control strategies afforded by individual AVs, which have recently seen potential for direct optimization of singular system objectives, such as traffic smoothing and emission reduction, and introduces a reinforcement learning-based methodological framework to facilitate a study of the trade offs between performance and safety at a fleet level. This investigation automatically produces Pareto frontier curves for four diverse traffic scenarios based on established mixed traffic benchmarks. The results of this study will inform decision-makers regarding inherent trade-offs in traffic control systems, and this framework can be extended to study arbitrary objectives in complex control systems.