Design autonomous vehicle behaviors in heterogeneous traffic flow

The benefits of autonomous vehicles (AVs) not only depend on the maturity of technologies, but also how AVs behave and interact with their peers and human-driven vehicles (HVs). Similar to many other systems, individual and collective dynamics of traffic flow are not always aligned with each other (for instance, aggressive driving may benefit an individual driver but disrupts the overall traffic). It is therefore imperative to consider behavior design for AVs such that the benefits of AVs can be realized at both individual and collective levels, even absent of centralized control. Behavior protocols for AVs will define their “driving styles”, in terms of information perception, utility, and opportunisticity. One possibility is letting all AVs be “human-like”, as did in existing literature. This research will explore more sophisticated behavior designs based on system principles and data. The research team will explore two approaches. The first approach is game-theoretic. In this approach, the team starts from defining agent utilities and casts interactions of heterogeneous agents as a spatial game. When a potential function for this game can be constructed, the team may prove the existence of its equilibria, derive conditions that lead to the desirable equilibria, and design AV behavior protocols based on these characterizations. From models of similar nature (known as Schelling’s models, which reproduce residential segregation), the team anticipates that with proper behavior protocols AVs can spontaneously form into platoons, even without centralized controls. The second approach is data-driven, leveraging deep reinforcement learning and big traffic data. In this approach, the team will train AVs as reinforcement learning (RL) agents from real-world trajectory data. Behavioral protocols are then obtained as the RL agents are endowed with reward functions of different structures. The team will identify the reward structures that best balance the individual and system goals and quantify the corresponding effects through simulations.

Language

  • English

Project

  • Status: Programmed
  • Funding: $113973
  • Contract Numbers:

    69A3551747119

  • Sponsor Organizations:

    Office of the Assistant Secretary for Research and Technology

    University Transportation Centers Program
    Department of Transportation
    Washington, DC  United States  20590
  • Managing Organizations:

    USDOT/OST-R

    Washington DC,   United States 
  • Project Managers:

    Kline, Robin

  • Performing Organizations:

    University of California, Davis

    1 Shields Ave
    Davis, California  United States  95616
  • Principal Investigators:

    Li, Jia

    Zhang, Michael

  • Start Date: 20210401
  • Expected Completion Date: 20220331
  • Actual Completion Date: 0
  • USDOT Program: University Transportation Centers

Subject/Index Terms

Filing Info

  • Accession Number: 01744087
  • Record Type: Research project
  • Source Agency: Center for Transportation, Environment, and Community Health
  • Contract Numbers: 69A3551747119
  • Files: UTC, RiP
  • Created Date: Jun 25 2020 12:36PM