A metasemantics to refine fuzzy if-then rules

Fuzzy if-then rules are used to represent fuzzy models. Real data is later used to tune the model. Usually this forces a modification of the initial linguistic terms of the linguistic variables used for the model. Such modifications may lead to a loss in interpretability of the rules. In this paper, we suggest using a form of multiresolution to tune the rules, by introducing more linguistic terms in the regions of the universe of discourse where this is needed. Starting with triangular shaped linguistic terms and recalling that symmetric triangles are first degree splines, a form of multiresolution is readily obtained, supporting a better accuracy of the model. It is shown that by using the linguistic modifiers, "very" and "more or less" as well as the metasemantic modifiers "between" and "from-to" the interpretability of the original rule may be preserved.