A. Chambolle : On acceleration and smoothing for a class of saddle-points methods for optimization in imaging
We will revisit some smoothing strategies (used in particular by
Nesterov, Teboulle) for accelerating first order schemes, and show how
a restarting strategy, based upon the primal-dual first order scheme
that we have studied a few years ago, yields again an optimal rate of
convergence for half-smooth convex problems. This is joint
(ongoing) work with Thomas Pock (TU Graz).
|