Intelligent control based on flexible neural networks. (English) Zbl 0938.93002
International Series on Microprocessor-Based and Intelligent Systems Engineering. 19. Dordrecht: Kluwer Academic Publishers. xiv, 235 p. Dfl 235.00; $ 140.00; £83.00 (1999).
The reviewed book deals with applications of artificial neural networks (NN) in control. To attract the attention, the authors use to speak instead about intelligent control based on flexible neural networks. So what is the book in fact about?
Chapter 1, named Introduction, presents the organization of the book. Then some ‘fundamentals of NN’ are sketched in Chapter 2 in a rather heterogeneous manner. As activation functions, the so-called flexible sigmoidal functions (FSF) are used; they are introduced by the authors in Chapter 3. Alas, this is a rather elementary generalization of standard sigmoidal functions (SF) exploited by others, too. The point is that the ‘height’ of the SF is not fixed and a steepness of the SF is just equal to it. Such FSF are then used as activation functions in NN constructed with such neurons for self-tuning PID control (Chapter 4) and for self-tuning torque control (Chapter 5 & 6). The FSF are also used as an inverse-dynamic model in Chapter 7. The self-organizing flexible NN is constructed in the sense of Kohonen self-organizing mappings (Chapter 8), which is used then as the controller. In this case, the Hebbian and Grossberg’s learning algorithms for such flexible NN are introduced. The final Chapter 9 summarizes the contents of the book.
One should say the book is written rather vaguely. It seems it was stitched up to wear this evening. Also the style of the book is not precise. There are many technical errors, even misleading or mathematically incorrect conclusions. E.g., the assertion (pp. 23) that ‘the USF (unimodal SF?) is attracting as an activation function because of its monotonicity and simple form’ is mathematically a pure nonsense. It is funny to classify SFs as bipolar SF, unipolar SF and radial basis functions (pp. 21). The authors use a non-typical terminology, which is commonly not appearing in the theory of NN, interchange SF with neurons, et cetera. The last sentence of the book, namely that ‘the concept of NN based on FSF can find a wide range of applicability to several problems in diverse fields such as engineering, science and humanities’ seems to be in fact non-scientific, as it is not true.
So on the basis of what has been said above one cannot recommend the book, especially for students and young researchers.
Chapter 1, named Introduction, presents the organization of the book. Then some ‘fundamentals of NN’ are sketched in Chapter 2 in a rather heterogeneous manner. As activation functions, the so-called flexible sigmoidal functions (FSF) are used; they are introduced by the authors in Chapter 3. Alas, this is a rather elementary generalization of standard sigmoidal functions (SF) exploited by others, too. The point is that the ‘height’ of the SF is not fixed and a steepness of the SF is just equal to it. Such FSF are then used as activation functions in NN constructed with such neurons for self-tuning PID control (Chapter 4) and for self-tuning torque control (Chapter 5 & 6). The FSF are also used as an inverse-dynamic model in Chapter 7. The self-organizing flexible NN is constructed in the sense of Kohonen self-organizing mappings (Chapter 8), which is used then as the controller. In this case, the Hebbian and Grossberg’s learning algorithms for such flexible NN are introduced. The final Chapter 9 summarizes the contents of the book.
One should say the book is written rather vaguely. It seems it was stitched up to wear this evening. Also the style of the book is not precise. There are many technical errors, even misleading or mathematically incorrect conclusions. E.g., the assertion (pp. 23) that ‘the USF (unimodal SF?) is attracting as an activation function because of its monotonicity and simple form’ is mathematically a pure nonsense. It is funny to classify SFs as bipolar SF, unipolar SF and radial basis functions (pp. 21). The authors use a non-typical terminology, which is commonly not appearing in the theory of NN, interchange SF with neurons, et cetera. The last sentence of the book, namely that ‘the concept of NN based on FSF can find a wide range of applicability to several problems in diverse fields such as engineering, science and humanities’ seems to be in fact non-scientific, as it is not true.
So on the basis of what has been said above one cannot recommend the book, especially for students and young researchers.
Reviewer: L.Andrey (Praha)
MSC:
| 93-02 | Research exposition (monographs, survey articles) pertaining to systems and control theory |
| 93B51 | Design techniques (robust design, computer-aided design, etc.) |
| 92B20 | Neural networks for/in biological studies, artificial life and related topics |
| 68T05 | Learning and adaptive systems in artificial intelligence |
