The National Science Foundation Division of Computing and Communication Foundations awarded the University of Southern California $871,225 on August 1, 2026, under the Computer and Information Science and Engineering program (CFDA 47.070) to develop theoretical foundations for multi-agent learning systems in artificial intelligence.
The project advances the theory of online optimization and learning within multi-agent systems through three research thrusts. The first thrust characterizes optimal regret guarantees and convergence rates in normal-form games under different feedback models and game structures, including zero-sum and general-sum settings. The second thrust explains the practical success of the counterfactual regret minimization algorithm for solving extensive-form games. The third thrust extends these theoretical advances to modern AI systems built with deep neural networks. The work addresses fundamental questions about when online learning methods with long-term performance guarantees work reliably, why they perform well in practice, and how they scale to applications in transportation, communications, automated trading, cybersecurity, robotics, and strategic planning.
The award supports open-source software and educational resources, graduate and undergraduate student training, high school research experiences, and outreach activities to broaden participation in computing. Performance occurs in Los Angeles, California through July 31, 2030.