Artificial Intelligence in Medicine 12 (1998) 287-289

                                    Book review

This book is about how information systems can be made to manage information which is in some sense imperfect (erroneous, imprecise, or uncertain). The underlying idea is that it may be advantageous to include in an information system all relevant information, even if it is in some sense imperfect.

The book's subject is at the intersection of two research areas, namely information systems, principally aimed at the design of practical systems for storing and retrieving information, and reasoning with uncertainty, a rather theoretical research area in artificial intelligence (AI) which investigates how to represent and reason with uncertainty.

There is an obvious tension between the practically oriented information systems research on the one hand and the more theoretical uncertainty in AI research on the other hand. To get satisfactory (real-time) performance of systems, the data and algorithms for manipulating the data should be kept as simple as possible, whereas to adequately handle imperfect information quite complex representations and reasoning method are required.

The (15) chapters of the book are based on papers presented at two workshops aimed at bringing together experts in both fields. However, the book is not a collection of research papers. Each chapter is written as an introduction to or survey of some aspect(s) of the topic, and is therefore accessible to nonexperts.

The book can roughly be divided into two parts. In the first part, researchers who work in information systems describe the issues, and the state of the art of dealing with these issues in the area of imperfect information in information systems. The following kinds of information systems are discussed: relational databases, intelligent databases, expert systems, and information-retrieval systems. The second part contains chapters authored by researchers in uncertainty in AI and describes the principal theories developed in this area and shows how these theories can be adapted to information systems. The discussed theories are probability theory, fuzzy sets and possibility theory, (Dempster-Shafer) theory of belief functions, and mathematical logic.

Throughout the book a great deal of attention is paid to clearly distinguishing the different kinds of imperfect information. The basic kinds are error (false and possibly inconsistent information), imprecision (including disjunctive, negative, and range information), and uncertainty (information with qualified confidence).

The information that the age of a particular employee is somewhere between 25 and 40 (without specifying the exact age) is an example of' imprecise (range) information. Storing this imprecise information may support more informative answers to queries (such as asking for a list of all employees younger than 50) than simply disregarding the information. Similarly, if the age of a particular employee is given as 35, but only with a confidence of 0.8 (on a scale of 0-1). then it may be more informative to represent this uncertainty of the information. Also, if it is given that in 5% of the cases the addresses of employees in the database are not correct then it may be advantageous to include such information about the quality of the information and the seriousness of possible errors.

Many of the chapters include an explanation of the different kinds of imperfection information. This results in a considerable overlap between different chapters, but it makes each individual chapter more self-contained. However, since both editors each authored an introductory chapter containing a lengthy discussion of the varieties of imperfection, these repeated explanations are in danger of becoming too much of a good thing.

The chapters on the theories proposed to handle imperfect information together cover a wide range of these theories. In some cases, the authors concentrate on merely explaining the theories (such as paraconsistent logic for inconsistent information, default logic for default information, and the transferable belief interpretation of Dempster-Shafer theory for uncertain information). In other cases. the authors pay detailed attention to practical issues when applying these theories (such as probability theory for uncertain information and possibility theory for imprecise information) to information systems.

The final chapter of the book contains an extensive bibliography on uncertainty management in information systems. This bibliography is divided into ten subfields and contains a total of 436 items. The book also includes a less extensive, but usable subject index.

On the whole, the book gives the impression of a very sound and accessible treatment of the subject mentioned in its title. Still, it is not complete. The items that I missed most are data fusion and a thorough discussion of Bayesian (belief) networks, which are used in one of the most popular techniques for handling uncertainty in AI.

The subtitle of the book suggests that it may contain solutions to problems with imperfect information in information systems. This is not really the case. Reading the book may clarify the problems and it may help in understanding relevant formalisms. However. in chapter 14, Mamdani discusses and argues against the possibility of finding a classification matrix showing under which abstract features of a problem involving the management of imperfect information each of the discussed (or related) theories are applicable.

In practice, deciding how much of the possible imperfection of information should be represented (and how) is a very complex problem. However. to solve it one needs to clearly understand the issues involved and to have a reasonable overview of the available techniques. The present book can be extremely useful for obtaining such an understanding and and overview.

Frans Voorbraak

Department of Mathematics, Computer Science, Physics and Astronomy
University of Amsterdam
Plantage Muidergracht 24
1018 TV Amsterdam
The Netherlands