Are State-of-the-Art Fine-Tuning Algorithms Able to Detect a Dummy Parameter?

Currently, there exist several offline calibration techniques that can be used to fine-tune the parameters of a metaheuristic. Such techniques require, however, to perform a considerable number of independent runs of the metaheuristic in order to obtain meaningful information. Here, we are interested on the use of this information for assisting the algorithm designer to discard components of a metaheuristic (e.g., an evolutionary operator) that do not contribute to improving its performance (we call them "ineffective components"). In our study, we experimentally analyze the information obtained from three offline calibration techniques: F-Race, ParamILS and Revac. Our preliminary results indicate that these three calibration techniques provide different types of information, which makes it necessary to conduct a more in-depth analysis of the data obtained, in order to detect the ineffective components that are of our interest.