Ensuring Accuracy and Transparency of Mamdani Fuzzy Model in Learning by Experimental Data

Typical violations of transparency of the Mamdani fuzzy model, which arise as a side effect of learning by experimental data are revealed. We suggest a new learning scheme of the Mamdani fuzzy model, which differs from the known ones by the following: 1) expansion of bearers of fuzzy sets of output variables; 2) excluding coordinates of maxima of belonging functions of extreme terms from the list of parameters to be tuned; 3) introducing constraints for linear ordering of fuzzy sets within limits of one term-set. Computer simulations indicate that learning by the new scheme does not break transparency of a fuzzy model. Moreover, accuracy of fuzzy model is not worse than for the case of typical learning.