Crisp Rule Extraction from Perceptron Network Classifiers

A method of extracting intuitive knowledge from neural network classifiers is presented in the paper. An algorithm which obtains crisp rules in the form of logical implications which approximately describe the neural network mapping is introduced. The number of extracted rules can be selected using an uncertainty margin parameter as well as by changing the precision level of the soft quantization of inputs. A fuzzy decision system based on the iris database has been developed using this approach to produce linguistic rules for flower classification.