Frank Wolfe

The Frank-Wolfe algorithm is a projection-free method for solving constrained optimization problems, particularly valuable when projections are computationally expensive. Current research focuses on improving its convergence rates, especially for stochastic and large-scale settings, through techniques like variance reduction, multistep methods, and adaptive step sizes. These advancements are enhancing the algorithm's applicability in machine learning, particularly for problems with structured constraints like those arising in empirical risk minimization and robust covariance estimation, leading to more efficient and accurate solutions.