A quasi-Newton algorithm for nonconvex, nonsmooth optimization with global convergence guarantees (Q903922)

The authors propose a new algorithm for minimizing a locally Lipschitz function which is continuously differentiable in an open dense subset of a real vector space. The algorithm requires only the first derivative of the cost function to be optimized and is based on the method initiated previously by Broyden, Fletcher, Goldfarb and Shanno. The algorithm is first described, then the authors proceed to a randomized analysis of its global convergence and conclude with the implementation of the method together with numerical experiments to illustrate the matter. The algorithm is supported mainly by line search technique.