Tuning a fuzzy logic controller

In this paper several tuning methods for Sugeno's fuzzy systems wiu be discussed. &t 4he first case the fuzzy controller is identified off-line based on training data. Following this approach, first the structure of the controller is identified by means of a clustering algorithm. A Kohonen se4f-organizing neural network performs this task. TheE the parameters of the fuzzy controlier (ie. membership functions) are tuned by using a gradient descent algorithm. This approach shows analogies with training a Radial basis function neural network. In the second case the fuzzy controller will learn to control the system in an on-line situation. The controller parameters are adapted on a supervised manner by using a gradient descent method. To make the parameter adaptation possible we need to know the sens&ivity functions of the system or at least their sign. If this knowledge is availabk the specialised learning technique becomes possible. Otherwise we need a preceding learning stage, the identification of the system. It is easily shown that when a multilayer perceptrons neural network emulates the system the sensitivity functions of the system can be derived by a mechanism of back propagation (applying gradient descent) no more on the weights but on the input of the emulator.