A novel data reduction method for Takagi-Sugeno fuzzy system design based on statistical design of experiment

This paper introduces a simple, systematic and effective method for designing Takagi-Sugeno (T-S) fuzzy systems utilizing a significantly smaller training data set versus existing methods. Creating proper training data is usually not an easy task and requires spending considerable time and resources. The proposed method first uses a three-level factorial design to partition the output space. Next the least square technique is used to estimate each of the partitioned output spaces. The membership functions are introduced with only three variables (min, max and number of membership functions). Fuzzy rules are generated with respect to the partitioned output surfaces and the membership functions. The proposed method is applied to two benchmark problems, controlling an inverted pendulum as well as modeling a nonlinear function. In the case of the inverted pendulum simulation results demonstrate significant improvement. In the case of nonlinear function modeling we demonstrated sufficient accuracy with only 9 training data, which represents 98% reduction in the number of training data compare to other method. Additionally, the proposed method offers extremely low computation time allowing it to be used with adaptive type systems.