Abstract

The Algorithm Configuration Problem aims to find suitable parameter values for metaheuristics. irace is a well-known tuner successfully applied in the literature to obtain proper parameter values for several algorithms. Understanding the search process of irace (and tuners) can be complex due to several factors, including its hyperparameters and the inherent stochasticity of both the tuner and the target algorithm. In this work, we analyze irace's search process through Search Trajectory Networks (STNs), focusing on the role of elite configurations, network structure, and the impact of the soft-restart component. Using ACOTSP (11 parameters) as the target algorithm, our results show that elite configurations represent only 4% of all visited configurations. Additionally, the soft-restart component enhances exploration, leading to more diverse and structurally distinct trajectories.