METHOD OF HIGH-PRECISION NEURO-FUZZY SYSTEMS DEFUZZIFIERS’ AUTOMATIC CONSTRUCTION

Neuro-fuzzy approximation precision increasing, based on systems’ self-organisation, is considered. Neuro-fuzzy systems’ self-organisation is primarily used to reduce systems’ training time, but their precision is less, than precision of systems without self-organisation. Root mean square error (RMSE) and maximal error are used as precision metrics. Method of neuro fyzzy systems’ self-organisation, which is based on defuzzifier automatic construction and fuzzy sets search, which provide high approximation precision, is suggested. Given method uses algorithms of detecting the most fitting borders of fuzzy sets and type of formed equations in order to produce the neuro-fuzzy approximation system with the best precision for a given dataset. Since fuzzy sets and defuzzufier configuration both influence systems’ precision, interaction of these parts of neuro-fuzzy system during training for saving training time is described. The computing experiment, aimed at estimating the complex mathematical function approximation precision reduction by implementing presented method was carried out. Using proposed method can be used for neuro-fuzzy approximation precision increase. Both fuzzy sets search and fitting output function selection reduce approximation error, but combining these approaches produces higher precision. Presented method provided lowering of both RMSE and maximal error. Size reduction can be achieved by implementing presented method since in can reduce number of required neurons.