A New Incentive Design Theory for Learning Agents in Federated Population Games

This NSF project aims to develop the scientific foundations for guiding the collective behavior of large groups of decision makers, such as drivers, consumers, and automated computer agents sharing roads, energy, supplies, and communication networks. When many participants each pursue their own goals, the overall outcome can be inefficient or unstable, as seen in traffic jams and power grid disruptions. Coordinators can promote better outcomes with appropriate incentives such as tolls, discounts, or rewards. Yet there is no systematic way to design incentives when the coordinator cannot directly measure how overall performance depends on participants' choices. The project will bring transformative change by showing how a coordinator can learn what it needs from incomplete or partial information participants voluntarily share as they change their choices and steer the group toward better outcomes with guarantees of stability, safety, and limited cost. This will be achieved by combining mathematical models of large populations with methods from control engineering, estimation, machine learning, and optimization. The intellectual merit includes new mathematical models, algorithms and rigorously validated approximations for large populations of learning agents, together with tractable gradient-inspired incentive mechanisms that improve performance when the performance measure must be learned rather than assumed known. This research will require a new multidisciplinary approach fusing methods and concepts from federated learning, population games, evolutionary dynamics, nonlinear control, optimization, and game theory. Broader impacts include potential improvements on transportation, energy, and other critical infrastructure, freely shared simulation software, undergraduate research experiences, and a new graduate course. The work will be disseminated in top venues reaching communities beyond engineering, including economics, applied mathematics, and biology. The technical framework of this research is built on principles of population games, which describe a large population by the fractions of participants selecting each choice, keeping the mathematics manageable regardless of population size. Existing incentive design methods assume that the coordinator already knows how performance depends on participants' choices. This project relaxes that assumption. Inspired by federated learning, which combines information from many participants without collecting private data, the team will use the information that participants share when changing their choices to estimate the direction of improvement for the whole system. The team will derive simple equations that track how the population, shared information, and estimates evolve together, and validate that their accuracy improves as the population grows. The team will then design incentives that provably steer the population toward suitable outcomes. 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.