On Lipschitz optimization based on gray-box piecewise linearization
成果类型：
Article
署名作者：
Griewank, Andreas; Walther, Andrea; Fiege, Sabrina; Bosse, Torsten
署名单位：
Universidad Yachay Tech; University of Paderborn; United States Department of Energy (DOE); Argonne National Laboratory
刊物名称：
MATHEMATICAL PROGRAMMING
ISSN/ISSBN：
0025-5610
DOI：
10.1007/s10107-015-0934-x
发表日期：
2016
页码：
383-415
关键词：
摘要：
We address the problem of minimizing objectives from the class of piecewise differentiable functions whose nonsmoothness can be encapsulated in the absolute value function. They possess local piecewise linear approximations with a discrepancy that can be bounded by a quadratic proximal term. This overestimating local model is continuous but generally nonconvex. It can be generated in its abs-normal form by a minor extension of standard algorithmic differentiation tools. Here we demonstrate how the local model can be minimized by a bundle-type method, which benefits from the availability of additional gray-box information via the abs-normal form. In the convex case our algorithm realizes the consistent steepest descent trajectory for which finite convergence was established earlier, specifically covering counterexamples where steepest descent with exact line-search famously fails. The analysis of the abs-normal representation and the design of the optimization algorithm are geared toward the general case, whereas the convergence proof so far covers only the convex case.