For tuning fuzzy controllers, several parameter identificationtechniques are available, ranging from more robust descent methods tosophisticated optimization. However, from an application point ofview, it is not always clear that numerical sophistication wins overmore pragmatic approaches to tuning. Obviously, the data sets playcrucial roles in efforts to reach successful tuning. Especially datasets generated from real processes often contain not only noisy dataand conflicting subsets, but also the connected problem ofnon-conferring input spaces. In this paper we will compare severalparameter identification techniques w.r. t. different data sets. Wefocus on selections of learning rates and on defining trainingsequences related to subclasses of parameters.
展开▼