Project Grant 2610563

This three-year Project Grant of $299,601, awarded July 1, 2026, by the National Science Foundation's Division of Mathematical Sciences under the Mathematical and Physical Sciences program (CFDA 47.049), supports fundamental research in distributional reinforcement learning (DRL) for risk-sensitive sequential decision-making. The University of Miami will develop theoretical foundations and practical methods that enable artificial intelligence systems to learn the full distribution of potential outcomes rather than relying solely on expected values, thereby reducing exposure to rare but catastrophic tail events. Core deliverables include: establishing finite-time convergence guarantees for quantile temporal difference learning in both synchronous and asynchronous settings; developing statistical inference methods including online bootstrap procedures for confidence intervals and offline methods for return quantiles and conditional value at risk; and creating trustworthy DRL approaches for constrained decision-making that incorporate privacy-preserving mechanisms.

The research addresses critical applications in clinical decision support, financial risk management, and autonomous systems where traditional reinforcement learning approaches prove inadequate. Beyond technical contributions, the project supports education and workforce development by training students at the intersection of statistics, machine learning, optimization, and responsible artificial intelligence, thereby strengthening the scientific enterprise in these emerging interdisciplinary areas. The work directly aligns with the Mathematical and Physical Sciences program's emphasis on advancing scientific knowledge while promoting broader societal impacts through enhanced trustworthiness in sensitive domains including healthcare and finance.