Download as pdf or txt
Download as pdf or txt
You are on page 1of 3

Artificial

Intelligence
ELSEVIER Artificial Intelligence in Medicine 6 (1994) 533-535
in Medicine

Book review

Frank Puppe, Systematic Introduction to Expert Systems: Knowledge Representa-


tions and Problem-Solving Methods 352 + XII pp., 129 figs., hardcover, DM 78.00,
Springer-Verlag, Berlin, 1993.

The author argues in the visionary sections of this book that in the not too
distant future expert systems will become a variety of modern knowledge media.
According to him, this promising goal motivates the current expert systems
research to develop domain-specific tools which will enable human experts to
produce expert systems without any other help. A necessary step towards this goal
is the identification and classification of problem types and problem-solving
methods. From the viewpoint of computer science the book tries to provide a
comprehensive survey of knowledge representation techniques and of problem-
solving algorithms for the programming of such tools and so-called expert systems.
The substance of the book is a slightly abridged version of Chapter II and some
comments on general topics such as the architecture and usability of expert
systems, translated from a German book of the author, Introduction to Expert
Systems, by Springer, 1988, and the translation into English of his German book
Problem Solving Methods in Expert Systems, Springer, 1990. Unfortunately, a
reference to these sources has been omitted.
Let us first consider the second of the two parts of the book, Puppe’s discussion
of problem-solving. He discusses problem-solving on the basis of a tripartition into
classification, construction, and simulation. He calls these categories ‘problem-
solving types’. He then identifies subclasses for problem-solving types, i.e. certain,
heuristic, model-based, statistical, case-based classification. These problem-solving
types are decomposed into problem types (e.g. fault finding, assessment, object
identification in the case of classification problems), and knowledge acquisition.
The author adds examples of computer programs which prototypically realize these
aspects.
Problem-solving types (classification, construction, simulation) are realized in
expert systems with the help of special algorithms which Puppe refers to as
problem-solving methods. He defines: “The term problem-solving method is used
for an algorithm which determines how domain-specific knowledge is used for

0933-3657/94/%07.00 0 1994 Elsevier Science B.V. All rights reserved


SSDZ 0933-3657(94)00018-2
534 Book review /Artificial Intelligence in Medicine 6 (1994) 533-535

solving problems” (p. 101). According to him, problem-solving methods for classifi-
cation problems are forward chaining, backward chaining, the establish-refine
strategy, and the hypothesize-and-test strategy (p. 103). For construction tasks the
basic problem-solving methods are the generate-and-test strategy and the propose-
and-revise strategy (pp. 213-214). Simulation is modelled by forward chaining (p.
267). The author points out that the results of research on problem-solving
methods for simulation are still fragmentary.
Again and again, these useful differentiations get lost throughout the text. The
chapters are hard to read due to the confusing use of notions. In Chapter 12.1 on
‘Survey of problem-solving methods’ and the corresponding Figure 12.1 the author
characterizes the problem-solving type classification with its subclasses as
problem-solving methods. In Figure 13.3 he characterizes these problem-solving
methods as “different types of knowledge” and relates them to problem-solving
methods in the sense of algorithms. On page 295 classification is “the best-under-
stood problem-solving method”, and on page 296 he says that classification
problems “in principle can be solved with any of the problem-solving methods
alone”. However, the blame for this may not be laid on the author alone. The
terminological confusion possibly reflects the state of the art in expert systems
research. A certain notational consistency is restored by the appended glossary.
After characterizing problem-solving methods the author goes back to a distinc-
tion between weak and strong probIem-solving methods. “Weak (or basic) prob-
lem-solving methods are open to a great variety of knowledge representation
and/or functions” (p. 101). “In strong problem-solving methods the representation
and function of the domain-specific knowledge are more or less fixed” (p. 102).
The algorithms are tuned to be more efficient by restricting the representation and
use of the domain knowledge. Complex problem classes have to be transformed
into simple problem classes. These reduction relations would have been explained
in a theory of strong problem-solving methods the author talks about in the
preface of his book. However, the outlines of this theory remain blurry. The text
does not answer the question of whether the distinction between strong and weak
problem-solving methods is of a sufficient empirical significance to form the
foundation of a theory of knowledge processing. The author is content with
preliminary phrasings such as “In the ‘classification’ group the subdivision is well
elaborated and the methods are fairly strong” (p. 103). And further, “Under
‘construction’, problem-solving methods have been given only for the subgroup
‘heuristic construction’, and these, particularly the least-commitment strategy, are
relatively weak” (p. 103).
Finally, let us turn to the first part of the book. In his introductory chapters to
problem-solving methods, Puppe necessarily touches techniques of knowledge
representation. This may have prompted the author to start his book with some
chapters about ‘Basic techniques of knowledge representation’. The result is that
the reader is faced with two introductions. A very general, to some extent
rhapsodic (e.g. the presentation of logic) and hurriedly presented introduction to
the representation of object-oriented, conceptual and procedural knowledge is
followed by an introduction to the specialized and more elaborated second part
Book review /Artificial Intelligence in Medicine 6 (1994) 533-535 535

about problem-solving. Thus a lot of repetitions and overlappings arise. To some


extent the chapters get their substance from the references to classical papers.

Ulrich Miiller-Kolck
Bielefeld, Germany

You might also like