Gray-box local search with groups of step sizes

Local search methods play an important role in several approaches to solving complex optimization problems. However, black-box local search methods for continuous optimization problems tend to be excessively time-consuming and their performance typically deteriorates when dealing with large-scale problems. In this context, we propose the Gray-Box Local Search with Groups of Step Sizes (GBO-LSGSS) algorithm. GBO-LSGSS explores the problem structure through partial evaluations and subcomponents for improving its efficiency. Thus, this proposed method implements an orchestrator module to learn and select the most promissory subproblems. Overall, an experimental study revealed the competitive GBO-LSGSS performance even compared to some of the main algorithms for solving large-scale continuous problems. The comparison between GBO-LSGSS and its original version provides evidence that the proposed method can find better solutions and save up to 90% of processing time for fully and partially large-scale problems.