Project Grant 2224213

This three-year Project Grant from the National Science Foundation's Division of Computing and Communication Foundations under the Computer and Information Science and Engineering program will support the development of new optimization approaches for machine learning problems. The $600,000 award to the University of Wisconsin-Madison beginning October 1, 2022 will advance optimization algorithms and analysis techniques for convex-concave minimax problems incorporating sparsity or regularity. New theoretically grounded algorithms will be developed for solving nonlinear programs with nonconvex functions and stochastic oracles, motivated by constrained neural networks, fairness constraints, and distributionally robust optimization. Research will also focus on applying optimization techniques like primal-dual methods to reinforcement learning and extending policy gradient method theory to account for inexactness in practical implementations. The work builds on classical optimization foundations and recent developments in optimization, control, learning theory, and statistics to provide principled optimization approaches addressing growing complexity in machine learning problem formulations.