Project Grant 2438392

This Project Grant award from the National Science Foundation's Engineering program (CFDA 47.041) provides $520,000 in funding to Arizona State University to advance the fundamental algorithmic and theoretical limits of distributionally robust reinforcement learning (RL). The research aims to develop provably convergent, efficient, and minimax optimal robust RL algorithms that can perform well even when the training and deployment environments differ due to factors like adversarial perturbations, modeling errors, or non-stationary conditions. The project's three thrusts will: (I) establish theoretical and algorithmic foundations for distributionally robust RL under the long-term average-reward criterion, (II) create a unified framework for learning robust policies from offline datasets, and (III) explore constructive approaches and limits of robust RL under constraints. The research outcomes are expected to significantly improve the robustness of RL algorithms, with applications in domains like healthcare, critical infrastructure, transportation, and smart cities. The award also integrates the research into educational and outreach activities targeting K-12 educators, graduate/undergraduate students, and underrepresented groups. The project period runs from September 1, 2024 to August 31, 2029.