On the neural defuzzification methods
Saman Kumara Halgamuge, Thomas A. Runkler, Manfred Glesner · Proceedings of IEEE 5th International Fuzzy Systems · 2002
If representative real world or artificial data sets exist, neural networks can be trained to approximate different defuzzification methods-explicitly known standard methods like center of gravity, extended parametric methods like customisable basic defuzzification distribution, and also black box defuzzification methods. From the neural network point of view this kind of defuzzification, is a multidimensional function approximation problem. In non black box adaptive solutions the analysing capability of the trained network is significant to understand the specificity of the application. Using random membership functions or a carefully selected variation of membership functions as training data, a black box defuzzification method with the lowest amplification known, is achieved. The application of the trainable transparent defuzzification to a real world problem is presented. Since the neural defuzzification is an integral part of many neuro-fuzzy systems, such an example is also described.