Stochastic perturbation of Frank-Wolfe method for nonconvex programming problems

Abstract

In this paper, we present a random perturbation of Frank-Wolfe method (a.k.a. Conditional gradient method ) for solving nonconvex differentiable programming under linear differentiable constraints. The perturbation avoids convergence to local minima. Theoretical results guarantee the convergence of the proposed method towards a global minimizer. Some numerical results of medium and large size problems are provided to show the effectiveness of our approach.