Stochastic optimization methods often rely on noisy gradient information obtained from sampled parameters. We consider a class of methods that reduce this noise by reusing previously sampled information to construct increasingly accurate approximations of expected gradients. We discuss the underlying algorithmic ideas, convergence properties and convergence rates. Applications from structural optimization, in particular topology and shape optimization under uncertainty, illustrate the practical use of these methods.

For further information about the seminar, please visit this webpage.