Tuning metaheuristics: A data mining based approach for particle swarm optimization

The paper is concerned with practices for tuning the parameters of metaheuristics. Settings such as, e.g., the cooling factor in simulated annealing, may greatly affect a metaheuristic’s efficiency as well as effectiveness in solving a given decision problem. However, procedures for organizing parameter calibration are scarce and commonly limited to particular metaheuristics. We argue that the parameter selection task can appropriately be addressed by means of a data mining based approach. In particular, a hybrid system is devised, which employs regression models to learn suitable parameter values from past moves of a metaheuristic in an online fashion. In order to identify a suitable regression method and, more generally, to demonstrate the feasibility of the proposed approach, a case study of particle swarm optimization is conducted. Empirical results suggest that characteristics of the decision problem as well as search history data indeed embody information that allows suitable parameter values to be determined, and that this type of information can successfully be extracted by means of nonlinear regression models.

Research highlights
► We devise a novel approach to determine effective parameter settings for metaheuristics by means of advanced regression methodology.
► The approach extracts useful information from data associated with the metaheuristic’s search history and characteristics of the underlying optimization problem.
► Empirical results indicate that the relationship between effective parameter settings and these types of information is sufficiently strong to be exploited for parameter tuning.
► Random Forest Regression is found to be particularly appropriate for the focal application and is recommended as prediction model for an automated parameter tuning system.