Christian Blum, Filippo Bistaffa

This thesis addresses two major challenges in metaheuristics: the lack of information about the problem instance and the lack of transparency in understanding their results. 

To tackle the first issue, I propose integrating two Machine Learning techniques into metaheuristic algorithms: Graph Neural Networks and, more notably, Large Language Models. These techniques enhance the quality of solutions obtained for NP-hard combinatorial problems when compared to using the metaheuristic alone.

To address the second issue, the thesis introduces a visualization tool called STNWeb, designed to support researchers in the comparative analysis of metaheuristics by facilitating the understanding of how problem instances influence their behavior.

Thus, the thesis is structured in two parts: the first focuses on improving the solution quality of metaheuristics through the incorporation of information extracted via modern Machine Learning techniques; the second explores the use of visualization as a means to better understand the behavior of metaheuristics.