Approche par apprentissage profond pour la sélection d’algorithmes dans l’optimisation à boîte noire (BBOP)

Abstract

Selecting the most suitable algorithm to solve a continuous black-box optimization problem is a challenging task. With a wide range of optimization algorithms evaluated annually and easier access to black-box optimization functions—primarily due to the efforts of the COCO platform—an automated method for algorithm selection in single-objective black-box optimization problems has become necessary. In this thesis, we everage Deep Learning technology combined with Exploratory Landscape Analysis techniques to predict the optimal solver (SBS) for any given set of problems. The performance of the proposed Deep Learning-based model surpasses that of a model based on traditional machine learning algorithms when applied to a set of continuous black-box optimization problems, thereby demonstrating its effectiveness.