Towards Collaborative Intelligence: An Integrated Learning Framework for Communication and Decision-Making

This NSF project aims to advance collaborative intelligence in networked multi-agent systems where many autonomous decision-makers must learn to cooperate using partial local information. Human collaboration thrives on adaptive exchange of information as learning and decision-making needs evolve. Inspired by this observation, the project will bring transformative change by treating communication and decision making as a coupled learning problem, enabling agents to learn not only how to coordinate their actions, but also what information to share, when to share it, and with whom. This will be achieved by developing an integrated learning framework that jointly optimizes communication and control to improve cooperation and scalability, with validation in smart grid applications. The intellectual merit of the project includes: (i) game-theoretic formulations that characterize how communication and decision-making jointly shape learning and coordination; (ii) analytical tools and metrics that provide a principled understanding of when and how communication improves learning; and (iii) decentralized learning algorithms that integrate event-triggered mechanisms to enable cooperation under imperfect communication. The broader impacts of the project include: (i) societal and economic benefits from advancing decentralized artificial intelligence solutions that enhance the resilience and efficiency of critical infrastructure; (ii) education and workforce development in artificial intelligence and power systems through graduate training, undergraduate research opportunities, and updated course modules; and (iii) broadening participation in STEM through mentoring and K–12 outreach activities. The proposed research will develop new multi-agent reinforcement learning theory and algorithms to address the tight coupling between communication and decision-making in networked multi-agent systems — what agents share shapes what they learn to do, and evolving decision policies in turn change what information is useful to share. This project will introduce a Coupled Partially Observable Games framework to characterize how information exchange influences decentralized policy updates and to establish sufficient-information conditions for joint policy improvement. Additionally, it will develop fully decentralized learning algorithms that combine trust-region methods for stable policy updates, event-triggered communication to share messages only when needed, and distributed optimization to enable coordination without centralized training. For validation, the methods will be evaluated in smart grid tasks including distributed Volt-Var control and networked microgrid coordination using Power-Hardware-In-the-Loop experiments under realistic grid dynamics and communication constraints. 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. NSF Award ID: 2529132 | Program: 01002627DB NSF RESEARCH & RELATED ACTIVIT | Principal Investigator: Hepeng Li | Institution: University of Maine, ORONO, ME | Award Amount: $249,959 View on NSF Award Search: https://www.nsf.gov/awardsearch/show-award/?AWD_ID=2529132 View on Research.gov: https://www.research.gov/awardapi-service/v1/awards/2529132.html