Experimental Sensitivity Analysis of Grid-Based Parameter Adaptation Method

Grid-based parameter adaptation method has been recently proposed as a general-purpose approach for online parameter adaptation in metaheuristics. The method is independent of the specific algorithm technicalities. It operates directly in the parameter domain, which is properly discretized forming multiple grids. Short runs of the algorithm are conducted to estimate its behavior under different parameter configurations. Thus, it differs from relevant methods that usually incorporate ad hoc procedures designed for specific metaheuristics. The method has been demonstrated on two popular population-based metaheuristics with promising results. Similarly to other parameter tuning and control methods, the grid-based approach has three decision parameters that control granularity of the grids and length of algorithm runs. The present study extends a preliminary analysis on the impact of each parameter, based on experimental statistical analysis. The differential evolution algorithm is used as the targeted metaheuristic, and the established CEC 2013 test suite offers the experimental testbed. The obtained results and analysis verify previous evidence on the method’s parameter tolerance, offering also an insightful view on the parameters interplay.