Fuzzy and neuro-fuzzy intelligent systems (Q1567476)

This book is an introduction to different intelligent system techniques. Humans can easily solve many problems that computers cannot. It is therefore desirable to formalize human expertise in a computer-understandable form. One of the major problems facing such formalization is that we humans often do not describe our knowledge and our decision rules in precise form, we use statements like ``if the speed is large, then decrease \(y\) slightly''. The problem with terms like ``large'' is that for many actual values, experts are unsure whether the given value is large or not. A natural way to formalize ``fuzzy'' terms like ``large'' is to describe, for each possible value \(x\) of, e.g., speed, the degree to which an expert believes this value to be large. This degree can be, e.g., obtained by marking on a scale. For convenience, different scales are usually mapped into the interval \([0,1]\), so each term can be described as a function \(U\to [0,1]\), where \(U\) is the set of possible values of the described quantity. Such functions are called fuzzy sets. Techniques are known for describing the degrees to which logical combinations of such terms are true; the corresponding ``and'' and ``or'' operations are called \(t\)-norms and \(t\)-conorms. The basic techniques of fuzzy set theory and fuzzy logic are described in Chapters 1, 2, and 5. Instead of waiting for a human to come up with a decision, we may take the raw data and simulate the way the human neurons process this data into appropriate decisions. The corresponding artificial neural networks are described in Chapters 3 and 4, with the brief description of traditional and intelligent optimization techniques used to ``tune'' neural networks such as genetic algorithms, simulated annealing, etc. Neural networks come up with ``rules'' which are described in terms of numbers. Alas, these numbers are rather meaningless for a human expert: e.g., if a necessity comes to slightly update these rules, human decision makers cannot do anything but start anew. To make the rules easier to maintain, it is reasonable, instead of generating a function from input to output (as traditional neural networks do), to generate appropriate fuzzy rules. The corresponding systems are called neuro-fuzzy. Such systems are described in Chapter 6, and their applications to time prediction, data compression, control, clustering, etc., are described in Chapter 7. Most of the neuro-fuzzy algorithms and applications are new. In addition to describing these algorithms, the authors provide Matlab code for them. Minor pedagogical drawback: In Chapter 1, the authors describe different binary operations on the interval \([0,1]\), including \(t\)-norms, \(t\)-conorms, and averaging operators. As a basis for their joint description, the authors take the \(L^1\)-distance between the corresponding functions. The idea is great, but the actual description is somewhat confusing: first, the term ``products'' for t-norms is unusual; in the 1-D figure, operations are placed based on their distance from ``drastic product'', but the false impression may be created that their relative location on this 1-D axis reflects the difference between them. A similar 1-D figure in Chapter 2 is, in contrast, very clear and unambiguously explained.