This event has passed.
Monday, November 20, 2023 – 12:00PM to 1:00PM
Katherine Flanigan, Assistant Professor in the Department of Civil and Environmental Engineering, Carnegie Mellon University
Cyber-physical systems (CPSs) have radically transformed engineering solutions over the last decade and garnered considerable attention, improving infrastructure performance through the combination of sensing, computing, and control. CPSs have even expanded to include human-in-the-loop control, where humans serve as operators or supervisors. While these paradigms have been wildly successful for the design and operation of physical systems decoupled from-or weakly coupled to-human social contexts, there are entirely unexplored human benefits derived from infrastructure that have yet to be scientifically understood and exploited. This is based on the hypothesis that the design and management of infrastructure plays a role in shaping human behavior, which has been more qualitatively explored in urban design and social science settings. For instance, social infrastructure (i.e., infrastructure supporting social interaction) is not social capital itself, but rather the physical space and infrastructure that determines whether social capital develops. There is an urgent need to reimagine current CPS theory, tools, and frameworks to even make it possible to address systems where human behavior is central. This talk proposes that the existing CPS paradigm be radically altered such that physical infrastructure is controlled to meet social, or human-centered objectives. This work deviates from existing-and highly limiting-agent-based modeling (ABM) approaches by proposing the use of data-driven, intelligent agents to mimic human behavior, including human-human and human-infrastructure interactions. We advocate for a hybrid approach, hierarchical imitation and reinforcement learning (HILRL), that better reflects human decision-making processes. HILRL leverages reinforcement learning's capacity to mirror human decision-making behavior while benefiting from imitation learning's capacity to incorporate real-world data. We demonstrate the effectiveness of this hybrid approach through a simulated conference room case study, illustrating the cooperative interactions of human agents in a social setting and offering insights into their collaborative behaviors.