What does fuzzy logic bring to AI?

Abstract

The term “fuzzy logic” often refers to a particular control-engineering methodology that exploits a numerical representation of commensense control rules in order to synthesize, via interpolation, a control law. This approach has many features in common with neural networks. It is currently concerned mainly with the efficient encoding and approximation of numerical functions and has less and less relationship to knowledge representation issues. This is, however, a very narrow view of fuzzy logic that has little to do with AI. Scanning the fuzzy set literature, one realizes that fuzzy logic may also refer to two other topics: multiplevalued Iogics and approximate reasoning. Although the multiple-valued logic stream is very mathematically oriented, the notion of approximate reasoning as imagined by Zadeh is much more closely related to the program of AI research: he wrote in 1979 that “the theory of approximate reasoning is concerned with the deduction of possibly imprecise conclusions from a set of imprecise premises.” In the following, we use the term “fuzzy logic” to mean any kind of fuzzy set-based method intended to be used in reasoning systems. Fuzzy logic is 30 years old and has a long-term misunderstanding with AI. As a consequence, fuzzy logic methods have not been considered to belong to mainstream AI tools until now, although an important part of fuzzy logic research concentrates on issues in approximate reasoning and reasoning under uncertainty. Some reasons for this situation may be found in the antagonism which existed for a long time between purely symbolic methods advocated by AI and the numerically oriented approaches that were involved in fuzzy rule-based systems. Besides, fuzzy sets were a new emerging approach not yet firmly settled, but apparently challenging the monopoly of probability theory on being the unique proper framework for handling uncertainty. In spite of the fact that fuzzy sets have received more recognition recently, there is still a lack of appreciation by AI researchers of what fuzzy logic really is, as, for instance, recently exemplified by Elkan [ 1994].