SCAIM Seminar: Steve Wright (Wisconsin)

Algorithms for constrained and nonsmooth optimization sometimes have the property that they identify a manifold on which the solution lies, before actually reaching the solution. This property is useful, particularly when the manifold can be parametrized in many fewer dimensions than the full space, as it opens the possibility of switching to an alternative algorithmic strategy with faster local convergence, once a approximation to the optimal manifold has been found. The talk discusses this issue in several contexts, including composite nonsmooth optimization and regularized optimization using stochastic gradient methods. It includes joint work with Adrian Lewis and Sangkyun Lee.