Overfitting Problem in a Virtual Sensor Obtained with W–M Method
Przemysław Andrzej Wałęga · Procedia Computer Science · 2014
In the paper we will analyze how a virtual sensor may be obtained, by means of Wang and Mendel method of generating fuzzy Rule Base. In particular, we will analyze how the number of fuzzy sets influences on the method's performance. We will state that increasing number of fuzzy sets leads to the overfitting effect, which will be compared to the overlearning effect known from Arti- ficial Neural Networks. Afterwards, we will introduce an algorithm for overcoming overfitting problem in Wang–Mendel method. Finally, we will present a virtual sensor based on real industrial data and discuss its quality.