Prototype based granular neuro-fuzzy system for regression task

Abstract

Artificial intelligence is often inspired by biological solutions. Prototypes are among these inspirations. Humans often describe complex entities by comparing them to previously known items (prototypes) instead of providing a detailed description. In this study, we apply this approach to neuro-fuzzy systems. Neuro-fuzzy systems operate using fuzzy IF-THEN rules. In our approach, the premises of rules are represented by prototypes. A new item is compared to the prototypes (as wholes) in the rule premises and its similarities to the prototypes in the rules define the firing strengths of the rules. The similarity is assessed based on the Minkowski metric. Prototypes are elaborated by granulation of the input domain with clustering and by applying the principle of justifiable granularity. We tested the proposed method in numerical experiments based on the regression task for benchmark data sets in public data repositories.