Automatically constructed fuzzy controller from training data

The paper discusses a way of designing controllers for affine TSK fuzzy models directly from training data, which may contain outliers. In the approach, an agglomeration clustering algorithm instead of split clustering algorithm is employed to determine the parameters both in premise and in consequent parts in the coarse tuning phase, and then a robust learning algorithm is used to fine tune the obtained fuzzy model. In controller design, fuzzy controllers share the same premise parts with the considered fuzzy systems and controllers are directly design for affine fuzzy systems. Because the proposed controllers are fully compensated for each rule, the closed loop performance can be theoretically anticipated.