Assessing the differences in accuracy between GFSs with bootstrap tests

The study of the balance between linguistic interpretability and numerical accuracy in genetic fuzzy systems is an active area of research, and a rich set of procedures for comparing the understandability of fuzzy rule bases is available. Nevertheless, comparing the numerical accuracy of two GFS is relegated to a second plane, or assumed solved, as most of the researchers in this area use classical, parametric hypotheses based, statistical tests. With this paper, we intend to show that the straight use of classical tests to compare the accuracy of different machine learning algorithms may produce misleading results, and propose to substitute them by bootstrap tests in the experimental designs of the cross-validation kind.