Siqian Shen
$145,305
University of California-Davis
California
Engineering (ENG)
This EArly-Concept Grants for Exploratory Research (EAGER) project will investigate the emergence, mechanisms, and applications of collective rationality (CR) among self-interested agents in the design of mixed autonomy networks and infrastructure systems. In many natural and engineering systems, various collective phenomena, such as spontaneous cooperation, spatial segregation, and behavior evolution and formation of social norms, can emerge at system level when the decisions and maneuvers of self-interested agents interlace with each other. Strategic agent behaviors play a key role in this process. This observation suggests that one may obtain a system with desired properties by carefully designing behaviors of its agents. We explore this idea and put forward the concept of “collective rationality” of mixed traffic towards explaining the formation of cooperation among self-interested driving agents in mixed autonomy transportation systems, to reduce travel cost, uncertainties, fuel emission, as well as to enhance equity among all road users. Broader applications include autonomous vehicle behavior design, emergency evacuation, and mitigation of pandemic spreading. The research will be further disseminated through curriculum design, K-12 education, and collaboration with practitioners, local government, and industry partners. <br/><br/>This project will explore and rigorously define the concept of collective rationality in mixed traffic and explore its application in designing strategic behaviors of autonomous driving agents in mixed autonomy environments. Our core hypothesis is that collective rationality can emerge in broad scenarios even if the involved agents are self-interested. We will leverage game theory and reinforcement learning to verify this hypothesis theoretically and computationally. To establish theoretical models of collective rationality in mixed traffic, we will develop two classes of models with different levels of agent behavior details, respectively focusing on the one-shot interaction of n-class driving agents, and dynamic inter- and intra-class interactions and an analytical Fokker-Planck approximation to the corresponding evolution dynamics. To develop frameworks for collective rationality-informed autonomous vehicle behavior design, we consider two autonomous vehicle behavior design frameworks using reinforcement learning, which incorporate collective rationality in reward design and employ a bi-level pricing structure to equitably fine-tune the benefit of cooperation among agents. The research team will also expand and explore the CR concept in other application contexts such disaster evacuations.<br/><br/>This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.