Fuzzy multivariate rule‐building expert systems: Minimal neural networks

Abstract A fuzzy multivariate rule‐building expert system (FuRES) has been devised which also functions as a minimal neural network. This system builds rules from training sets of data that use feature transformation in their antecedents. The rules are constructed using the ID3 algorithm with a fuzzy expression of classification entropy. The rules are optimal with respect to fuzziness and can accommodate overlapped and underlapped clusters of data. The FuRES algorithm combines the benefits obtained from simulated annealing and gradient optimization, which provide robustness and efficiency respectively. FuRES classification trees support OR logic in their inference. The system automatically generates meaningful and consistent certainty factors during rule construction. Unlike other neural networks, FuRES uses local processing which furnishes qualitative information in the rule structure of its classification trees and variable loadings of the weight vectors.