Answers are grounded in this paper and cite their sources.
Increasing the precision of neuro-fuzzy approximation systems is considered. Modification of defuzzifier, which allows producing multiple types of approximating functions is presented. RMSE (Root Mean Square Error), maximal error, training time and models’s size are used to evaluate the efficiency of proposed method. Post-processing functions were also considered as a tool to increase approximation precision. Common solution to this problem is a modification of model’s fuzzy sets (membership functions and their number), which significantly increases model’s size and complexity. Given approach uses search algorithms to find the configuration of defuzzification layer, which produces the best approximation precision for a given function and fuzzy sets configuration, since all possible configurations are checked. It is possible to model classic approximation methods and combine them. Possible modifications of presented search algorithms application are described. Search depth can be limited in order to reduce the size of produced model. The computing experiment, aimed at finding best approximation precision was carried out. Presented method provided lowering of RMSE without significant increasing of model’s size. Using Fourier series components in resulting expression can be used for precise periodic functions approximation or for general precision improvement. Using post-processing functions provided significant improvement in precision for multiplication approximation, which proves, that it can be used as another tool to be used to reduce errors in some specific cases.
No takes yet. Share an insight, caveat, or question.
S. M. Morozov (2024) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: