In Fuzzy systems in computer science, Kruse R, Gebhardt J, Palm R, eds,

155-169, Vieweg, Braunschweig/Wiesbaden 1994.

Fuzzy Systems in AI
Christian Freksa
Abstract
This paper reviews motivations for introducing fuzzy sets and
fuzzy logic to knowledge representation in artificial intelligence.
First we consider some areas of successful application of conventional approaches to system analysis. We then discuss limitations
of these approaches and the reasons behind these limitations.
We introduce different levels of representation for complex
systems and discuss issues of granularity and fuzziness in connection with these representation levels. We make a distinction
between decomposable and integrated complex systems and discuss
the relevance of this distinction for knowledge representation and
reasoning. We also distinguish fuzzy relations between quantities
of different granularity within one domain from fuzzy relations
between two different domains and discuss the need of considering
both in artificial intelligence.
We distinguish methods for describing natural, artificial, and
abstract systems and contrast the modeling of system function with
the modeling of system behavior in connection with the representation of fuzziness. The paper briefly discusses recent criticism of
the fuzzy system approach and concludes with a prospect on soft
computing in AI.

Why do we Need Fuzzy Sets and Fuzzy

Logic in AI?

The notion of a fuzzy set [Zadeh 1965] and the development of fuzzy set
theory and fuzzy logic were motivated by the severe difficulties to adequately
characterize complex systems by conventional approaches of system analysis.
Adequate means, for example, that insignificant variations on the
component level of a system should not add up to significant changes on the
system level. This criterion is an absolute requirement for understanding
complex systems in terms of their components.
Conventional approaches represent complex systems in a reductionist
manner by specifying well-defined components and their individual interactions. We will investigate the question why these approaches are of limited
use in artificial intelligence and cognitive science.

C. Freksa: Fuzzy Systems in AI

The success o f conventional approaches i n well-defined s y s t e m s

Until fairly recently, most researchers in artificial intelligence believed, AI
systems could benefit from the same virtues that made much of modern
science and technology a success: precise measurements, complete knowledge
of the domain, and rigorous tools for dealing with them. Before discussing
this belief I will briefly investigate under which conditions dealing with
precise, crisp, and complete measurements has been successful.
A fundamental prerequisite for succeeding with high-precision characterizations of complex systems is the availability of basic entities and relations
suitable for capturing everything that is relevant in the system under consideration. Under this condition, we can decompose complex systems into
less complex subsystems and/or basic components and we can explain their
role in terms of the basic components and their interactions.
Obviously, we can correctly describe complex systems in terms of their
components when (1) the components conform with their definitions, (2) the
interactions conform with their specifications, and (3) no other interactions
interfere. These conditions certainly hold in abstract systems that are
specified according to the three conditions given above. Examples are
complex games defined by simple rules like the board games chess and go.
The three conditions also can be assumed to hold in technical systems
whose components can be studied in isolation and whose interactions can be
restricted to controlled local exchanges. Examples are closed chemical,
mechanical, and electronic systems. Complex computers and computer
networks constitute excellent examples that the bottom-up approach to
characterizing systems on the basis of crisp notions can be highly successful.
However, the more complex these systems get, the more obvious
becomes the need for developing high-level views and languages to better
capture the essential aspects including actions at the relevant system level.
Thus, the granularity of a description can make a big difference even if fine
and coarse descriptions of crisp systems can be considered equivalent, in a
formal sense.
The problem with ill-defined systems
The great success in representing complex systems in terms of their
simple components that could be achieved in closed technical systems may be
responsible for a blind belief that the same approach could be applied to other

C. Freksa: Fuzzy Systems in AI

complex systems equally well. Let us consider a class of non-technical

complex systems whose representation in terms of components has caused
great difficulties: systems of economics or climate systems may serve as
examples. Why are such systems different?
In case of an economic system, the three above requirements seem
fulfilled, at first glance: (1) monetary units define the economic value of any
item in basic terms, (2) mathematical rules precisely derive the effects of
economic transactions, and (3) no other interactions besides transactions
determine the monetary values on the component level. In fact, the strict
obedience of the mathematical laws by economic transactions make it
possible to precisely analyze in quantitative terms why someone became rich,
became poor, went bankrupt, etc.
The problem is, that it is not very interesting to have models that only
can be used for post hoc analysis. In the case of climate systems even the
post hoc analysis does not carry very far. We would like to build models to
make predictions! Why is it possible to make predictions on the component
level for board games but not for economic systems?
The role of system complexity
The fact is that practically speaking it is not even possible for board
games like chess or go to make reliable predictions on the component level of
description. The complexity of the chess game consisting of only 32 pieces
on an eight by eight checkers board already is too big for analyzing all legal
moves. Nevertheless, it is possible to build useful higher-level descriptions
of chess constellations from the basic entities and use these descriptions to
generate reasonable predictions.
In systems of economy, in contrast, this approach has not proven
practicable; condition (3) is severely violated: individual transactions in
economy systems can not be predicted on the basis of local transactions; they
are determined by complex global patterns and their dynamics involving
psychological and other factors which cannot be captured in terms of the
elementary configurations.
A similar situation is given in the case of climate systems and weather
prediction: our knowledge about the preconditions in terms of fundamental
facts will never be complete enough and our knowledge about the physical
laws on the local level does not suffice for making useful predictions about
global or local weather conditions. Thus, system complexity is only one

C. Freksa: Fuzzy Systems in AI

aspect that must be considered; the availability of knowledge and the

structure of that knowledge also play an import role for the representation we
can use.
How does fuzziness come into the picture?
If we compare systems that are well-defined (from the bottom up) with
systems that we know primarily from their global properties, we find different
mapping relations between the low-level and the high-level notions. The
high-level notions built up from low-level primitives typically are found in a
crisp relation to the synthesized complex notions while the low-level notions
postulated from the high-level concepts are found in a fuzzy relation.
For example, a taxonomy of plants and/or animals based on low-level
primitives will yield a crisp classification of creatures as found in biology
textbooks. On the other hand, the identification of low-level features for the
definition of the high-level everyday notion living animal yields fuzzy
relations between the high-level notion and the low-level features, as it is
impossible to precisely capture the everyday notion universally by composition of low-level features. In this sense, living animal is a fuzzy concept,
when related to low-level primitives while on the high level on which we
typically use the notion it would not be considered fuzzy at all.

The level of representation

Systems about whose properties we learn from global behavior can be

described meaningfully on a global level, for example in the case of an economic system we might know when the interest rate goes up, the money flow
decreases. Such a rule implicitly can take into account complex interaction
regularities which cannot be captured on the component level. Although the
total money flow results from individual monetary transactions, the rule does
not hold for each individual transaction. Therefore it is difficult to give
precise definitions of the global notions in terms of local transactions; it is
much easier to identify the global effect as the net effect of local transactions
and to control economy on a global level (e.g. by manipulating the interest
rate) than through local transactions.
In the case of weather prediction we encounter a similar situation: on a
coarse level (which may be relevant for agriculture, for example), predictions
may be quite reliable, while they may be useless on the level of description

C. Freksa: Fuzzy Systems in AI

on which local measurements of weather indicators (like rainfall per square

meter) are made.
As a consequence, by representing complex systems not on the level of
the most primitive notions but on a coarser level, these systems may become
tractable. Numerous complex interactions on the local level simply can be
ignored! Lotfi Zadeh recognized the importance of abstraction from low-level
properties at an early stage of the artificial intelligence enterprise. He characterized intelligent systems by their ability to summarize complex descriptions
by abstracting from details. In the case of everyday non-synthetic systems
this requires taking into account the fuzzy relations between the high-level
and the low-level features.
Integrated complex systems
There is a class of complex systems in which fuzzy relations play a
particularly important role.
This class consists of systems whose
components cannot be studied in isolation or whose components cannot be
studied in all relevant conditions. Most natural complex systems belong to
this class. We can observe global behavior under varied condition patterns;
from these we infer local influences.
Examples are biological systems which we describe in terms of presumed
local functions and observed effects. Neither the description of the global
effects nor the description of the local functions are suitable to capture all
possible situations and to crisply represent the system. The reason is, that
we are bound to use concepts which have a meaning outside of these systems
(like living animal) since elementary local components are not available and
there is no way to guarantee that these concepts precisely match the
components of the described system.
Fuzzy relations between different domains
Fuzzy pattern recognition, fuzzy control, and most other successful
applications of fuzzy set theory have focused on the fuzzy relation between
the fine and coarse levels of representation of the artifacts involved. But from
the inception of fuzzy set theory, its inventor Zadeh also suggested to
represent fuzzy relations between real entities like physical objects and mental
entities like concepts which are manifested in natural language expressions.
In artificial intelligence, this type of relations is of particular interest.

C. Freksa: Fuzzy Systems in AI

Classical artificial intelligence had been focusing its attention on the

representation structures within the medium computer. Little attention had
been paid to ontological questions and the actual representation problem, i.e.
the mapping between what is represented and what is representing [c.f. Palmer
1978] and to the problem of building representation structures from existing
knowledge. Many AI-researchers did not consider this issue a problem; they
believed, concepts could be used like nuts and bolts to screw together
intelligent systems and the role of each concept would be clearly defined.
But when AI-systems grew up and moved outside their purely synthetic
laboratory environments it became evident that there was a serious matching
problem between natural concepts derived from the use and the behavior of
systems and artificial concepts synthesized from low-level components. It
became clear that we could not simply view natural concepts as imperfect
entities whose objectives would be much better served by artificial substitutes.
Instead, expert system research and research in cognitive science investigated structures and properties of human knowledge in order to exploit its
potential and to understand more of its function. In this process, fuzzy sets
have played an important role in characterizing the relation between human
concepts and natural or artificial entities and in mimicking their interactions
(c.f. [Zimmermann 1992], [Dubois et al. 1993], [Kruse et al. 1994]). In this
way, many of the properties of human thought and natural language,
specifically with regard to their modification and their combination, could be
simulated.

Natural systems vs. artifacts

As argued in the previous sections, the development of fuzzy set theory

and fuzzy logic were influenced by properties of natural human concepts and
their relation to entities in the real world. Specifically, Zadeh suggested to
use fuzzy sets to represent notions like tall and beautiful in natural language
phrases like The tall dwarf is more beautiful than the small giant. However,
the present success in the application of fuzzy set theory is not so much in
artificial intelligence e.g. in the representation of natural language expressions or human concepts but rather in control engineering, e.g. due to the
improvement of purely artificial systems like household appliances, photo
equipment, trains, and helicopters [Munakata, Jani 1994]. Why have fuzzy

C. Freksa: Fuzzy Systems in AI

sets not caught on in artificial intelligence to the same degree as in control

engineering?
The utility of fuzzy sets in artificial systems is not only due to reduced
complexity as a consequence of the coarser representation of systems, but also
due to properties of human concepts which support the engineering process of
these systems. Fuzzy sets serve as a knowledge transfer vehicle, as a way of
getting the judgment of engineers into complex systems.
They are
particularly suited for this task since they allow for reasonable representation
relations even if the representation system is not yet completely understood.
In this sense, human concepts are involved in the synthetic products of the
engineers. But does this mean that fuzzy sets represent human concepts?
The characteristic function which defines a fuzzy set characterizes the
relationship between real world entities and specific concepts (or labels
typically associated with concepts); however, it does not model or explain
how this fuzzy relationship comes about. For certain domains or types of
tasks, a characterization of the fuzzy relationship between a concept and given
instances in reality is sufficient (for some tasks even this relationship is not
required) - but there is an important class of problems which requires going
beyond the characteristic function.
Shallow vs. deep modeling
I will argue that classical fuzzy sets represent human concepts on a rather
shallow level, on the level of denoting entities within a well-defined framework, a framework in which the relevant dimensions are known, but precise
values within these dimensions are missing. Human concepts, however, are
not merely underdetermined physical values. They form a strong system on
their own which have a meaning and make sense independent of real-world
instances even though during the knowledge acquisition process they may
have been derived from such instances.
In Figure 1, I present a classification of different epistemic qualities of
knowledge associated with methods appropriate for processing them. On the
basis of this classification I will suggest ways of extending the fuzzy set philosophy towards representing human concepts on a deeper level. The goal is
to adapt notions developed in fuzzy set theory and fuzzy logic to make them
better suitable for processing human knowledge and knowledge representation
in artificial intelligence.

C. Freksa: Fuzzy Systems in AI

KNOWLEDGE TO BE REPRESENTED
Knowledge available
through formal
methods

Cognitively
available
knowledge

Uncertain

Certain

complete
(closed world)
Horn clause
logic

incomplete
(open world)
predicate
logic

statistical
model
statistics

quantifiable
uncertainty

comparable
uncertainty

functional
model

qualitative
model

formal
derivation

empirical
distribution
fuzzy
uncertain
knowledge

crisp
uncertain
knowledge

(technical
domains)

disjunctions
intervals

ABSTRACT
WORLDS

Figure 1: Classification
accessibility criteria

fuzzy fuzzy uncertain

uncertain knowledge
knowledge
(abstract
(natural physical cognitive
domains)
world)

fuzzy logic

conceptual
neighborhood

ARTIFICIAL
REAL WORLD

NATURAL
REAL WORLD

of knowledge types on the basis o f

C. Freksa: Fuzzy Systems in AI

A classification of knowledge types

The classification of knowledge types presented in Figure 1 distinguishes
different types of situations in which knowledge can be acquired. I will first
present the distinctions made in this classification and subsequently I will
associate types of approaches which appear suitable for dealing with the
respective situations and knowledge.
The top-level distinction in this classification is made between certain
knowledge and uncertain knowledge. Certain knowledge is only available in
abstract domains where facts and rules can be postulated to be true. Knowledge obtained from concrete domains are subjected to uncertainty, due to limitations in modeling and knowledge acquisition. Certain knowledge may be
further classified into complete knowledge in which the non-existence of a
true fact is equivalent to the existence of the negated fact (closed world
assumption) and incomplete knowledge in which this assumption is not
generally valid.
Uncertain knowledge can be further classified into knowledge with quantifiable uncertainty, i.e. knowledge with which we can associate an absolute
degree of uncertainty, and comparable uncertainty, i.e. knowledge that we
merely can rank according to relative degrees of uncertainty.
Uncertainty may be due to statistical effects, due to incomplete knowledge, or due to fuzziness; accordingly uncertainty can be quantified on the
basis of statistical models or on the basis of functional models. Statistical
models can be classified according to the source of the probability distribution: are the probability distributions derived from a formal model on the
basis of assumptions about the processes involved (e.g. each face of a die is
equally likely to be thrown) or from empirical observations.
In the class of functional models of uncertainty we can distinguish
between those yielding crisp sets of possible values and those yielding fuzzy
sets. The crisp sets may be either intervals (in the case of continuous
domains) or disjunctions (in the case of discrete domains). Fuzzy sets can be
obtained when the functional dependencies are clearly enough defined to allow
for a quantitative characterization of the elasticity of constraints. This is the
case in synthetic domains in which fuzzy membership values can be derived
from the system specifications. For an interesting illustration of the
advantages of studying synthetic systems rather than natural systems, see
Braitenbergs experiments in synthetic psychology [Braitenberg 1984].

C. Freksa: Fuzzy Systems in AI

In the domain of natural language or other natural domains, quantitative

membership values generally are not available; but frequently membership
values can be compared like in I would rather say John is tall than John is
very tall. Such comparisons lead to qualitative models of uncertainty. We
can further distinguish between fuzzy uncertain knowledge about the natural
physical world where we may have empirical sensory evidence for comparative uncertainty judgments and fuzzy uncertain knowledge about abstract
cognitive domains where it is not possible to identify objective correlates to
the comparative judgments.
In summary, this classification yields three different major types of
knowledge processing situations: 1) The situation of abstract domains in
which we work on the basis of formal specifications; here we either deal with
complete certainty or with well-controlled uncertainty.
2) Real-world
situations which are well-controlled in the sense that we know all relevant
factors and the possible extreme situations; this is the case in many technical
domains whose function in principle could be completely described on the
basic level by exhaustive analysis; such artificial domains are ideal for the
application of classical fuzzy logic. 3) In artificial intelligence, we are frequently confronted with situations in the natural world in which the available
knowledge does not justify a quantitative fuzzy set approach; in particular,
we must deal with open worlds in which representations in terms of numerical membership functions do not make sense; instead, qualitative knowledge
may be available. For this type of situation, more adequate representational
approaches are needed.
I would like to suggest that fuzzy reasoning may become as successful as
fuzzy control and will be useful for natural language processing and other AI
applications when it is integrated with deep conceptual representations of the
domain and of the language describing the domain. In particular, the horizontal dimension connecting cooperating and competing concepts must be
exploited in addition to the vertical dimension which connects concepts with
their definitions. This horizontal dimension is required for determining which
concepts are to be used in the first place. Depending on this selection, fuzziness may or may not play a role for the given task.
A possibility for extending the scope of fuzzy reasoning on the basis of
the notion of conceptual neighborhood is currently being explored by the
author. The notion of conceptual neighborhood was introduced in the context
of qualitative temporal and spatial reasoning [Freksa 1992]. It addresses the
logical structure of the represented domain under the operations that can take

C. Freksa: Fuzzy Systems in AI

10

place in the domain. Specifically, two relations are conceptual neighbors

when there is an operation in the represented domain which can transform one
of the two relations into the other. Conceptual neighborhood can be used to
represent horizontal relationships between concepts and to form coarse concepts from fine concepts and vice versa.

Criticism of the fuzzy systems approach

In a prize-winning article, Charles Elkan expressed his surprise about the

success of fuzzy logic [Elkan 1993]. Specifically, Elkan attempts to show
formally that fuzzy logic collapses to classical two-valued logic and he argues
that it is not adequate for reasoning about uncertain evidence in expert
systems from an empirical perspective. Elkans overall judgment is that
fuzzy logic is fundamentally wrong and will cause serious problems in more
challenging applications.
Like Zeno's famous ancient paradoxes [c.f. Vlastos 1967], Elkans paradox
appears rather convincing at first glance, when the reader submits himself to
the formal framework used by the author. However, a more careful analysis
of Elkans argument reveals that like Zeno he presents a restricted view.
He does this by forcing a notion of equivalence valid for two-valued logic on
the analysis of fuzzy logic. In this way, Elkan addresses only special cases in
which two-valued and fuzzy logic in fact are equivalent [cf. Shastri, to
appear]. It is easy to provide numerical counterexamples to Elkans equivalence assertion using standard notions of fuzzy sets for intermediate truth
values.
Limitations of formal analysis
I will take the occasion of this attack on fuzzy systems on the basis of
purely formal arguments to draw your attention to an important issue of
knowledge representation systems which in principle cannot be resolved by
formal analysis. Formal analysis helps us to understand systems which are
entirely formal. However, in representing knowledge about the real world,
one part of the system is the body of knowledge to be represented, another
part is the representing formal structure, and a third part establishes the relations between the body of knowledge and the formal structure [c.f. Palmer
1978].
Only the second part, the formal structure, can be rigorously analyzed
formally. The first part, the body of knowledge is not accessible with formal

C. Freksa: Fuzzy Systems in AI

11

tools directly; human perception and/or intuition present the knowledge to be

represented by the formalism. The representation of the knowledge can be
only as good as our understanding of the structure of the knowledge itself!
For example, if we take a natural language statement like John is tall
merely as a different way of writing the predicate logic statement tall (John) is
true, then we will never be able to reach aspects of the original statement
which are not coded in the predicate logic equivalent. We easily can become
victims of the same kind of fallacy Charles Elkan was subjected to when he
viewed the world of fuzzy logic through the glasses of two-valued logic.
In reasoning about the real world, making formally sound inferences is
only one aspect. Equally important but much more difficult than widely
believed is to adequately formalize real world knowledge in the first place.
The paradoxical success of the fuzzy logic approach in restricted domains
may be considered as an indication that clearer perception or sharper intuition
about the relation between the domain knowledge and the domain states have
been involved in the knowledge formalization process. Of course, having
found a new representation structure, we must develop appropriate reasoning
methods to go along with.
It may well be impossible to find methods which will both fit the more
adequate representation structure for real world phenomena and satisfy the
classical criteria of formal analysis such as logical equivalence. Nevertheless,
the resulting inferences may be more useful than formally correct inferences
on the basis of less adequate knowledge structures. The notion of representational adequacy is not yet sufficiently understood.

Soft computing

Fuzziness is one of several aspects of our knowledge about the real world
which must be taken into account in knowledge representation and processing. In general, we must deal with imprecision, uncertainty, and partial
truth. The human mind can be viewed as a working realization of a system
which rather successfully deals with all of these aspects simultaneously. In
contrast to conventional (hard) computing approaches, systems that are
tolerant of these aspects of everyday knowledge are united by the label soft
computing.
The guiding principle of soft computing is: Exploit the tolerance for
imprecision, uncertainty, and partial truth to achieve tractability, robustness,
and low cost solutions [Zadeh 1994]. The basic ideas underlying soft

C. Freksa: Fuzzy Systems in AI

12

computing have links to many early influences of fuzzy set theory, including
Zadehs original publication on fuzzy sets [Zadeh 1965], his paper on the
analysis of complex systems and decision processes [Zadeh 1973], and his
paper on possibility theory and soft data analysis [Zadeh 1981].
Besides fuzzy logic, probabilistic approaches for reasoning under uncertainty and related models for belief maintenance and revision play an important role. In artificial intelligence, Pearls probabilistic reasoning in Bayesian
networks, Nilssons probabilistic logic, the certainty factor model used in the
MYCIN expert system for medical diagnosis, Dempster-Shafers theory of
evidence have attracted much attention (c.f. [Lpez de Mntaras 1990], [Kruse
et al. 1991]). In the mid 1980s, neural network theory also joined into the
soft computing effort.
Combining different approaches to soft computing
It has become evident during the past ten or twenty years, that no single
approach to the study of cognitive or intelligent processes will succeed in
understanding the interactions of cognitive agents with complex environments
and no single approach to representing complex knowledge will fulfill all our
requirements. Successful AI approaches must take into account effectiveness,
efficiency, timeliness, robustness, adequacy, and cost of the solutions.
Classical requirements like provability of correctness and completeness of the
solution can be expected as little from computer systems reasoning about
complex situations as from humans in the same situation.
After an era of increasing specialization in almost all areas of research and
technology, we have now entered an era in which the interaction of
approaches is of particular importance and concern. This is true for numerous
areas, but interdisciplinary efforts like cognitive science and multi-approach
efforts like the Berkeley Initiative in Soft Computing (BISC) might serve as
examples. Such efforts require a considerable amount of re-orientation, as we
have to recognize that the former competitors must become partners.
Although fuzzy set theory and fuzzy logic have faced strong opposition
from conventionally oriented theoreticians in artificial intelligence and logic
during the past 30 years, the rapidly growing number of successful applications developed mainly in Japan have shifted the focus of interest from local
formalistic concerns to global system considerations. As in the case of the
Fifth Generation Computing Project in the early 80s it required the Japanese
challenge before European and American opposition was matched by a

C. Freksa: Fuzzy Systems in AI

13

growing interest in the industry and an increased willingness by theoreticians

to understand the principles of soft computing.
In Europe, we now find an increasing interest in the theory and applications of soft computing techniques in artificial intelligence. This is evident
from the growing number of fuzzy logic workshops and soft computing
contributions to artificial intelligence conferences, from the establishment of
special interest groups in fuzzy logic and soft computing (e.g. within the
German computer science society) and from the growing number of tutorials
offered both by the industry and by academic institutions.
Three papers on specific topics of fuzzy reasoning
In the remainder of this chapter you find three articles dealing with soft
computing for artificial intelligence.
The article by Sascha Dierkes, Bernd Reusch, and Karl-Heinz Temme
presents a tool for supporting the representation of fuzzy knowledge and for
fuzzy reasoning in an experts system shell.
The article by Jrg Gebhardt uses the possibilistic interpretation of fuzzy
sets in the context of model-based reasoning. The approach described in the
paper allows for evidential reasoning in multidimensional numerical
hypothesis spaces under imprecision and uncertainty.
Jochen Heinsohn presents a language to extend the taxonomic knowledge
representation approach of terminological logics by a probabilistic knowledge
representation component. In this way uncertain knowledge can be included
in the reasoning process.

Acknowledgments
I thank Jrg Gebhardt, Jochen Heinsohn, and Ramon Lpez de Mntaras
for valuable comments on an earlier version of this paper.

References
[Braitenberg 1984] V. Braitenberg Vehicles. MIT-Press, Cambridge 1984.
[Dubois et al. 1993] D. Dubois, H. Prade, R.R. Yager (eds.) Readings in
Fuzzy sets for intelligent systems. Morgan Kaufmann Publishers, San
Mateo 1993.

C. Freksa: Fuzzy Systems in AI

14

[Elkan 1993] C. Elkan: The paradoxical success of fuzzy logic.

AAAI-93. 698-703, AAAI Press/MIT Press, Menlo Park 1993.

Proc.

[Freksa 1992] C. Freksa: Temporal reasoning based on semi-intervals, Artificial Intelligence 54 (1992) 199-227.
[Kruse et al. 1991] R. Kruse, E. Schwecke, J. Heinsohn: Uncertainty and
vagueness in knowledge based systems: numerical methods. Series
Artificial Intelligence, Springer, Heidelberg 1991.
[Kruse et al. 1994] R. Kruse, J. Gebhardt, F. Klawonn: Foundations of
Fuzzy Systems. John Wiley and Sons, Chichester, 1994.
[Lpez de Mntaras 1990] R. Lpez de Mntaras: Approximate reasoning
models. Ellis Horwood, Chichester 1990.
[Munakata, Jani 1994] T. Munakata, Y. Jani: Fuzzy Systems:
view. Communications of the ACM vol. 37, 3, 69-76, 1994.

An over-

[Palmer 1978] S.E. Palmer: Fundamental aspects of cognitive representation. In: Rosch, E., Lloyd, B. (eds.), Cognition and categorization,
Erlbaum, Hillsdale 1978.
[Shastri, to appear] L. Shastri (ed.): Symposium on Fuzzy Logic, IEEE
Expert (to appear).
[Vlastos 1967] G. Vlastos: Zeno of Elea. In: P. Edwards (ed.) The Encyclopedia of Philosophy vol. 8, 369-379. New York: Macmillan 1967.
[Zadeh 1965] L.A. Zadeh: Fuzzy sets. Information and Control 8, 338353, 1965.
[Zadeh 1973] L.A. Zadeh: Outline of a new approach to the analysis of
complex systems and decision processes. IEEE Trans. SMC 3, 1, 28-44,
1973.
[Zadeh 1981] L.A. Zadeh: Possibility theory and soft data analysis.
Mathematical Frontiers of the Social and Policy Sciences, L. Cobb and
R.M. Thrall (eds.), 69-129. Westview Press, Boulder 1981.
[Zadeh 1994] L.A. Zadeh: Fuzzy logic, neural networks, and soft computing. Communications of the ACM vol. 37, 3, 77-84, 1994.
[Zimmermann 1992] H.-J. Zimmermann. Fuzzy set theory and its applications. Kluwer, Boston 1992.

C. Freksa: Fuzzy Systems in AI

15