Knowledge transfer for labeling unknown fuzzy sets using Grammar-Guided Genetic Algorithms

Abstract

Fuzzy Logic Systems have been popular in fields such as control, knowledge representation, and modeling. They involve linguistic information to represent a human perception in the form of Fuzzy Sets. Those systems are advantageous in cases where it is desired to have some interpretability. However, there are some processes, such as optimization, where the membership function related to a fuzzy set is modified. Therefore the linguistic label associated with them is unattached to the original meaning. This proposal aims to make linguistically-interpretable fuzzy sets from previously established ones but with modified parameters, thus maintaining the context of the knowledge of the domain application and the knowledge designer. The presented approach relies in the use of linguistic modifiers, which are operations linked to linguistic terms that apply on fuzzy sets, such as very, above, more-or-less, etc. A methodology using Grammar-Guided Genetic Algorithms is proposed to find the best linguistic modifier candidates to approximate the fuzzy set with modified parameters to guarantee a coherent and readable grammar structure at diverse levels of interpretability and accuracy. The obtained results support the feasibility of the proposed methodology with convex membership functions such as Gaussian, Triangular, and Trapezoidal.