Abstract

The search for proper parameter values is a key process for applying metaheuristic algorithms to solving complex optimization problems. Several specialized tuning methods have been proposed in the literature. One of the main difficulties when tuning parameters is the stochastic nature of metaheuristic algorithms and their requirement to solve problem instances with different features. In this work, we are interested in understanding different tuning process features using the Search Trajectory Networks approach. Here, a network of search processes can be constructed based on the solutions visited and the sequences of visits performed. Here, we extend the definitions of Search Trajectory Networks to tuning processes using two tuning methods from the literature: ParamILS and Evoca. We analyze the differences between the parameter tuning processes they perform and the incidence of their main hyper-parameters in these processes. From our results, we conclude the relevance of the number of pairs seed/instance for the search performed by ParamILS but not for Evoca regarding the number of visited configurations and the network's connectivity. Moreover, the evolutionary nature of Evoca promotes an exploratory behavior, traversing trajectories with fewer nodes in common compared to ParamILS.