Experimental Evaluation on Defuzzification of TSK-type-based Interval Type-2 Fuzzy Inference Systems

With A 2- C 0 and A 2- C 1 models, this study conducted a series of experimental evaluation on four kinds of type reduction and defuzzification algorithms, namely Karnik-Mendel iterative procedure (KM), Nie-Tan method (NT), q factor method (QF) and modified q factor method (MQ), to show which algorithm exhibits better performance. In the experimental part, the accuracy, computational cost and stability of these methods are compared, and the accuracy between different defuzzification methods are analyzed statistically with the Wilcoxon Nonparametric Test. It is shown that for both A 2- C 0 and A 2- C 1 models, the NT, QF and MQ method could save around 80% development time than KM. The associated improvement for the accuracy and stability caused by KM are mainly concluded as: 1) for the A 2- C 0 model, less than 5% and around 10% than NT, about 0% and less than 5% than QF and around 20% and about 40% than MQ; 2) for the A 2- C 1 model, more than 60% and over 70% than NT and less than 20% and more than 30% for both QF and MQ. It becomes clear that for the A 2- C 0 model, the NT and QF method are better choices; while for the A 2- C 1, KM is a better choice when high accuracy and stability are the prerequisites but if the accuracy, stability and cost of development time are considered synchronously, QF and MQ are better. During the design of the A 2- C 0 and A 2- C 1 structures, the reported results provide us a map for the most suitable selection of the defuzzification approach among these four kinds of methods.