ABSTRACT
So far we have discussed two distinct methods for building controllers: fuzzy and neural. Often the choice of method is dictated by the data available on the plant involved. If the data are pairs of numbers, we may turn to a neural method, and if the data are rules, fuzzy methods may be appropriate. Neural methods provide learning capability, whereas fuzzy methods provide ßexible knowledge-representational capability. Integrating these two methodologies, in control in particular and in intelligent technologies in general, can lead to better technologies that take advantage of the strengths of each methodology and at the same time overcome some of the limitations of the individual techniques. In this chapter, we discuss methods for combining neural and fuzzy methods
to build controllers. There are many ways in which these methods can be combined. Complicated controllers can have different component problems, each of which may require different types of processing, but such complex situations are beyond the scope of this book. Within a single component, there are still basically two ways that fuzzy and neural technologies can be combined. In one direction, fuzzy logic can be introduced into neural networks to enhance knowledge representation capability of conventional neural networks. This can be done by introducing fuzzy concepts within neural networks – that is, at the levels of inputs, weights, aggregation operations, activation functions, and outputs. Standard mathematical models for neurons can, for example, be changed to “fuzzy neurons” with t-norms and t-conorms used to build aggregation operations. This leads to a fuzzy-neural system with which one can present fuzzy inputs and develop an analog of the conventional backpropagation algorithm for training. In the other direction, neural networks can be used in fuzzy modeling and
to a neural-fuzzy system – a fuzzy system represented as a modiÞed neural network, resulting in a fuzzy inference system that is enhanced by neural network capabilities. Fuzzy systems are generally more “user friendly” than neural systems because their behavior can be explained based on fuzzy rules fashioned after human reasoning. Although fuzzy logic can encode expert knowledge directly using rules with linguistic labels, it usually takes a lot of time to design and tune the membership functions that quantitatively represent these linguistic labels, and applications of pure fuzzy control systems are restricted mainly to those Þelds where expert knowledge is available and the number of input variables is small. Neural network learning techniques can automate this process and substantially reduce development time and cost while improving performance. Neural networks are also used to preprocess data and to extract fuzzy control rules from numerical data automatically, as well as to tune membership functions of fuzzy systems. In this chapter, we Þrst address some issues of fuzzyneural systems for control problems, and then look at neural-fuzzy systems. Our primary focus is on adaptive neuro-fuzzy systems for control.
