PURPOSE: To learn a membership function of fuzzy inference, which performs min. operation as the and operation of a rule antecedent by a method of descent by introducing a 0-1 approximate function which is increased in the gradient of a sigmoid function.

CONSTITUTION: An input is multiplied by the weight Wg of the coupling link between layers A and B, Wc is added in the layer B to determine one point on a straight line y=Wg.X+Wc, and the grade of the membership function is calculated in a layer C. In a layer D, respective congruency degrees are distributed to respective rules and in a layer E, respective min's are calculated to find an antecedent congruency degree; and the output of a layer f is multiplied by the weight Wa of the link between layers F and I and the antecedent congruency degree is weighted with Wa as a consequent constant value to obtain a rule output value. The sum of outputs of the layer E is calculated in a layer G, its reciprocal number is obtained in a layer H to obtain a normalization constant for the rule congruency degree, and the rule congruency degree is normalized in a layer H to calculate a normalization rule congruency degree. In a layer J, the sum of the normalization rule congruency degrees of the respective rules is calculated to obtain a fuzzy inference result.