Um método de regularização proximal inexato para otimização irrestrita

Abstract

In this work, we study a regularized algorithm to solve optimization problems without restrictions when the objective function is two-fold differentiable. The algorithm was proposed in [1] and it is basically a Newtonian method appropriated to solve problems when the Hessian matrix is singular in an optimal local solution. This algorithm consists of two sub algorithms, named Algorithm 1 and Algorithm 2 and they are directly connected with the Proximal Point algorithm.We present a detailed proof of global convergence under the assumption that f is two-fold differentiable and lower bounded. We also highlight local convergence of the algorithm with super-linear rate with a local error margin condition in the gradient of f. Finnaly, we elaborate examples that allows one to glimpse the operation of the algorithm.