UC Berkeley

Automated driving systems are expected to play a critical role in the future of transportation. With fast reaction times, vehicle-to-vehicle communication, and the potential for socially optimal driving behaviors, automated vehicles may serve a central role in improving driving conditions within existing road networks, reducing the prevalence of traffic congestion and enabling fast and energy-efficient driving. Generating behaviors that produce such effects in real world settings, however, is no trivial task. In particular, when coupled with human drivers in mixed-autonomy settings, coordination between human-driven and automated vehicles becomes increasingly delicate, and motivates the need for new and advanced tools for solving these tasks.

Through the research outlined in this document, we aim to identify efficient methods for learning congestion-mitigating control strategies that can be employed by automated vehicles in partially automated road networks. Recent advances in deep reinforcement learning have highlighted the potential of said techniques in producing control strategies that match or outperform classical approaches on a variety of decision making and control tasks. The applicability of similar approaches to mixed-autonomy traffic control, however, is hindered by a number of challenges. For one, exploration in these settings is difficult, as individual actions may not influence the flow of traffic until multiple timesteps in the future. In addition, the process of modeling and executing simulations of realistic traffic flow networks at the level of individual vehicles is a difficult and computationally costly endeavor. Through techniques such as hierarchical learning, imitation from experts, and robust learning in simplified tasks, we hope to design data-efficient methods for generating control strategies for automated vehicles that are transferable to the real world.