This is a hefty tome with 23 papers from members of the IRIS Group, each one a substantial contribution of 24 pages on average. The acronym IRIS expands as “Integration of Reasoning, Informing, and Serving” and is explained in the Preface by the primary author. The members of the group are 38 in number, of whom 28 are in the USA, and one other besides the primary author in Yugoslavia, three each in Japan and Korea, and one each in Austria and Belgium.
It is argued in the preface and in Chapter 1, both of them by the primary author, that “reasoning information systems”, including living ones, have the three functional divisions of Reasoning, Informing and Serving. The illustration of the idea in the preface, by listing topics under the three headings, is not very convincing, partly because general topics like “robotics” and “artificial life” are forced under a particular heading although they must contain elements of all three. The illustration in the first chapter is clearer, with reference to a person, and with “Reasoning” attributed to the brain and mind, “Informing” to memory and sensors, and “Serving” to body and actions.
The assertion in the preface is that powerful artificial systems can be constructed from modules classifiable under the three headings, and that the modules are already available, at least in laboratories if not yet on suppliers’ shelves. The modules are visualised as embodying the many advanced techniques of neural nets, genetic algorithms, expert systems, fuzzy and chaotic programming, holographic memory and so on. (The term “chaotic programming” is used both here and in the title of another book with the same authorship, where it is associated with Sixth Generation Computing. It is not explained in the book under review.) The main outstanding problem is seen as that of effectively and efficiently integrating the modules into overall systems, and this is the concern of the IRIS Group.
Rather a lot hinges on the meaning to be attached to “integration”. The author stresses that it is not to be equated with “interfacing” with its implication of application‐specific hardware or software, and certainly not with the “SYSGEN” approach in which the final system is globally compiled. The author visualises ready and unrestricted assembly of the modules and refers to two existing successful schemes allowing such assembly of software for robotic systems. The use of expert or knowledge‐based methods to plan large systems is part of the picture and a specific scheme is described in Chapter 19.
The primary author’s introductory material is enthusiastic, even ebullient, expressing confidence that the approach shows the way ahead. He claims that the book is to be seen not only as a review of the state of the art but as a compendium of current knowledge that will serve as a teaching text.
It is natural to question such a stance and my own feeling is that it is questionable in at least two respects. One is at the basic level at which the idea of modularity seems contrary to what we know about visual and auditory perception, where recognition is aided by feedback from all levels of interpretation including the semantic. It also seems contrary to a recent trend in robotics to eliminate the gap between high‐level planning and execution. Much hinges on the meaning to be attached to “integration” of modules, but if integration implies richness of interconnection and interaction it would seem to nullify modularity.
There is clearly a major problem here, since modularity is almost essential if overall system design is to be manageable by humans or expert systems, but on the other hand the study of living systems is often argued to require a holistic, or essentially non‐modular, view. It is not, of course, necessary to model every aspect of biological information processing to produce powerful commercially‐viable systems, and the IRIS approach may indeed be the way ahead for these. An interesting commercial aspect that receives attention towards the end of the first chapter is software protection to ensure that the originators of modules receive royalties according to the use of the product, even when the product can be duplicated without restriction, a situation referred to as “superdistribution”.
The other respect in which the stance seems questionable is in its over‐optimistic assessment of the achievements of work on artificial neural nets and genetic algorithms and even of mainstream AI. There is a change of tone between the introductory material and the chapters that follow. The latter are sound and valuable contributions, but realistic in acknowledging areas needing further investigation, and some of them reporting studies that are far removed from practical applications in the short term. The final chapter, by M.F. Peschl, is in fact critical of the whole field of mainstream AI, arguing that what is modelled is only the tip of the iceberg of the thought process, and advocating attention to the still‐elusive “common sense” aspect of intelligence.
The book is divided into three parts. The first, with 11 chapters (though effectively ten since the introductory first chapter is counted), has the heading: “Neural, genetic and intelligent algorithms and computing elements”. There is here a wealth of ideas and research findings. Chapter 2 is a description of a model of a real neuron, much more complex than the Hodgkin‐Huxley version that was a breakthrough in its day, and accounting for an impressive range of neural phenomena. The next is on concept learning by neural nets, and after a general review the author introduces his own special contribution according to which sensory data is subjected to learned classification that is of the unsupervised or “without a teacher” kind, except that it is influenced by feedback from the symbolic processing stages that it feeds. The next chapter after that is on fast algorithms for training multilayer perceptions, embodying the principle of the Kalman filter and hence matrix inversion and regression.
These valuable contributions are followed by Chapter 5 in which simulated annealing is used to train a net in a particular recognition task, by altering connectivity instead of adjusting weights. The fact that all weights are plus or minus unity allows fast, multiplication‐free, operation of the net. Chapter 6 reports work on nets of units more complex than the usual threshold or sigmoid kind, having discrete internal states. The results are linked to the later Chapter 14 on a neurally‐inspired model of rule‐based reasoning.
The next chapter in this section refers to yet another means of achieving fast learning, which results in nets with a remarkable ability to extract the structure of text. The next again is on the forming of effective nets by simulated evolution, and the remaining three in this section are about parallel computer architectures allowing efficient realisation of artificial neural nets.
All of these contributions are “meaty” and stimulating, and the same is true of the rest of the contents, amounting to five papers under the heading of “Integrated neural‐knowledge‐fuzzy hybrids” and seven under “Integrated reasoning, informing and serving systems”. Several chapters under the last heading refer to robotics.
The primary author’s enthusiastic depiction of current trends includes references to fifth and sixth‐generation systems, and it is not clear (to me, at least) exactly what these imply. Chapter 18 of the book is a progress report, probably now somewhat out‐of‐date, on the Japanese fifth‐generation initiative. Near the end of it the authors acknowledge that there is some doubt as to whether the research activity is best represented as a progression from one generation to another. Perhaps the imposition of such structure can only realistically be done retrospectively.
The set of contributions brought together in this book is of exceptionally high quality, with a greater diversity of approaches and more speculative ideas than the introductory comments might lead one to expect. This aspect makes it specially interesting to research workers, and certainly useful as collateral reading for students, though hardly as a primary teaching text as seems to be intended by the primary author. It is a very valuable addition to the literature.
