Comparative Study on Five Defuzzification Methods in Chaotic Time Series Prediction
Yanyan Chen · Jisuanji fangzhen · 2013
During the construction of a Mamdani fuzzy inference system,there are five defuzzification methods commonly employed,and those methods have significant impacts on the prediction quality when predicting chaotic time series.Therefore,to quantitatively analyze the impacts for those defuzzification methods will help determine an appropriate defuzzification mechanism when modeling and provide valuable information for the improvement of defuzzification method.This paper deliberatively selected three one-dimensional chaotic maps and traffic flow data,in order to perform comparative study on those defuzzification methods through a series of simulated experiments.The experimental results indicate that the prediction qualities are largely influenced by the defuzzification methods,and the first attempt should be made on the defuzzification method of mean value of maximum to construct a Mamdani fuzzy inference system.Meanwhile,the experiments demonstrated that a relatively good generalization was achieved,implying the Mamdani fuzzy inference system is capable of predicting chaotic time series.In addition,the results also imply that the prediction quality is negatively correlated with the Lyapunov exponent of time series,in general,for all the defuzzification methods evaluated.