A few decades ago, the word “hybrid” would have been understood to indicate a combination of digital and analog techniques. It is used here to denote various combinations out of the three choices of the knowledge‐based, or expert‐system approach (assumed from the start to embody fuzziness) and artificial neural nets and genetic algorithms. The Editors start by observing that humans are hybrid information processing systems, and in many applications it is counter‐productive to restrict attention to one technique. They have assembled this interesting collection of accounts of successful hybrid schemes.
The schemes fall into three categories, and the descriptions are placed correspondingly in the first three parts of the book, with then a fourth part consisting of a single chapter on tools and environments for hybrid systems. The three categories are function‐replacing hybrids, intercommunicating hybrids, and polymorphic hybrids, and there are four papers under each of these headings.
By function‐replacing hybrids are meant schemes depending essentially on a single technique, but with some principal function within it replaced by a different technique. The first paper in this section is applied to process control and is of triple rather than hybrid parentage since it uses fuzzy logic, neural nets and genetic algorithms. The neural net is used to form a process model, which is then used to adjust the fuzzy logic controller by a method using genetic algorithms. Another paper in this first part is again applied to process control, using as a sample application the well‐known pole‐balancing problem. Here again a genetic algorithm provides an effective way of adjusting the fuzzy logic controller.
Of the two remaining papers under this heading, one refers to the use of a neural network within an expert system, useful when the patterns or situations to be recognised are unsuitable to be represented by sets of IF ... THEN rules. The other is on the use of genetic algorithms to adjust the weights of a neural network, resulting in a method that outperformed backpropagation in a series of test problems.
The topic of the second part of the book is intercommunicating hybrids, in which the distinct techniques can be seen as residing in separate, but communicating, modules. The first paper in this section is on the combination of genetic algorithms with well‐established techniques of numerical optimisation in engineering design. Another is on a hybrid system called RECON for “data mining”, or the intelligent search of databases. RECON brings to bear a battery of statistical techniques and possibilities of visual display, and can also embody a neural net. Another paper is on the use of fuzzy pre‐processing for neural nets in connection with chemical process diagnostic problems and another is on the integration of neural networks with expert systems.
The third part treats what are termed polymorphic hybrids, or those that are chameleon‐like in showing multi‐functionality within one system. Again there are schemes in which genetic algorithms are applied to the training of neural nets. The first paper in this part deserves special attention, as it is a thoughtful and detailed comparison of symbol‐processing systems and connectionist networks. The conclusion is that both offer the possibility of general‐purpose processing and that they have more in common than is usually supposed. Some tasks are more readily performed by one and some by the other. General rules are given for combining the two in hybrid systems. This paper is particularly noteworthy because one of its authors is Leonard Uhr, one of the very early workers in this area, with the distinction of appearing (Uhr and Vossler, 1960) in the famous anthology of Feigenbaum and Feldman.
The book under review is valuable from a number of points of view. For one thing, it presents a challenge to neural net theorists, who will certainly want to consider carefully how it comes about that the established training schemes can be outperformed by methods, particularly genetic algorithms, imported from quite different application areas. All of the presentations have a reassuring air of practicality, and in fact most of them describe successful working systems. Applications of the techniques to financial management are mentioned at several points, including the editors’ biographical notes. The reader’s confidence must be boosted by the observation that the editors are prepared to back their expertise with real money.
A paper that would seem relevant, but does not appear among the references in the book, is the account by Fogel, Fogel and Porto (1990) of their work on adjustment of neural nets using the principles of simulated evolution developed much earlier by Fogel, Owens and Walsh (1966).
