Fuzzy rules generation from data through fuzzy evaluation of fuzzy rules

This paper deals with generating fuzzy rules from numerical data. The rule evaluation which is necessary in order to determine rule fitness with respect to given data is often based on mere heuristics and usually lacks in semantic interpretation. The method proposed in this paper is based on two central ideas which face these shortcomings. Fuzzy rules are evaluated through fuzzy quality values which are similar to fuzzy truth values introduced by Zadeh (1987). They get a semantics by comparison with predefined standard fuzzy quality values. Thus, they are easy to understand even for laymen. The other idea refers to the computation of fuzzy quality values. It is based on frequency distributions of the underlying data which are transformed to fuzzy sets using the theory of mass assignments and a voting model semantics. Furthermore, the fuzzy quality values are involved in the fuzzy inference process. Altogether, this method is well based, uses less heuristics and is more transparent than other approaches