Numerical performance of non-monotone line searches on unconstrained global optimization problems

Details

Supervisors

Nunes Grapiglia, Geovani

Faculty

Ecole polytechnique de Louvain

Degree label

Master [120] : ingénieur civil en mathématiques appliquées, à finalité spécialisée

Abstract

In this work, we investigate the use of non-monotone line search methods to approximately solve unconstrained global optimization problems. Firstly, we present an extensive numerical comparison between monotone and non-monotone line search methods. The results show that non-monotone line search methods outperform the monotone Armijo line search, providing significantly better approximate solutions, with iterates exhibiting the ability to escape from nearby non-global local minimizers. Secondly, we develop a new non-monotone line search for problems in which the optimal function value is known a priori. We find numerically that this new method outperforms existing monotone and non-monotone methods in instances of the Molecular Distance Geometry Problem.