Abstract

The search for suitable parameter values in metaheuristic algorithms is known as the Algorithm Configuration Problem (ACP), which is a challenging problem that directly impacts the performance of target algorithms. Parameter tuners have been proposed to effectively identify suitable values for a set of instances to solve the ACP. However, the search process performed by tuners is a complex and difficult-to-understand task, considering the stochasticity of the tuner, the stochasticity of target algorithms, the quality estimation and definition of decimal precision, the hyperparameters of the tuner and their impact on its performance, among other aspects. In this work, we examine and analyze the search process of the Evolutionary Calibrator (Evoca) tuner using Search Trajectory Networks (STNs). Our goal is to investigate the structure of the configuration networks explored by Evoca and to extract relevant features from the existing STN model. Additionally, we analyze the effect of the population size hyperparameter and the nearness between visited configurations. To study the STNs of Evoca, we utilize two algorithms: a genetic algorithm designed to address NK landscapes and the standard particle swarm optimization (named SPSO2011) for solving continuous optimization problems. Our results indicate that the importance of the population size hyperparameter depends on the size of the parameter search space and the difficulty of the instance set. Additionally, Evoca can visit structurally different configurations, demonstrating the real exploratory effect of their crossover and mutation operators.