See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/258484987

Fuzzy Control Systems

Chapter · January 2004

CITATIONS READS

19 12,414

3 authors, including:

Jens Jaekel Ralf Mikut

Leipzig University of Applied Sciences Karlsruhe Institute of Technology
106 PUBLICATIONS 1,093 CITATIONS 577 PUBLICATIONS 7,127 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Ralf Mikut on 23 May 2014.

The user has requested enhancement of the downloaded file.

FUZZY CONTROL SYSTEMS

Jens Jäkel, Ralf Mikut, and Georg Bretthauer

Institute of Applied Computer Science, Forschungszentrum Karlsruhe GmbH, Germany

Keywords: Adaptation, defuzzification, expert knowledge, fuzzification, fuzzy control, fuzzy con-
troller design, fuzzy control structures, fuzzy logic, fuzzy model, fuzzy operator, fuzzy Petri net, fuzzy
set, fuzzy system, fuzzy system analysis, implementation, inference, intelligent control, knowledge-
based system, linear system, linguistic term, linguistic variable, Mamdani-type fuzzy system, mem-
bership function, nonlinear system, rule base, Takagi-Sugeno-type fuzzy system, supervision

Contents

1. Introduction
2. Fuzzy Control - A Simple Example
2.1. Example
2.2. Fuzzy Sets, Linguistic Variables and Fuzzy IF-THEN Rules
2.3. Fuzzification - From Measurements to a Fuzzy Representation of the Input Situation
2.4. Inference - From a Fuzzy Input Representation to a Fuzzy Decision
2.5. Defuzzification - From a Fuzzy Decision to a Real Decision
3. Fuzzy Logic-related Issues in Fuzzy Control
3.1 Fuzzy Sets and Operations
3.2. Types of Rule-based Fuzzy Systems
3.3. Information Processing in Fuzzy Systems
4. Control Issues in Fuzzy Control
4.1. Structures in Fuzzy Control
4.2. Design of Fuzzy Control Systems
4.3. Analysis of Fuzzy Control Systems
4.4. Applications
4.5. Hardware and Software for Fuzzy Control Systems
4.6. Further Concepts Based on Fuzzy Sets Applied to Control Systems
5. Conclusions

Summary:
This chapter presents a perspective of fuzzy control systems. Fuzzy control is a form of intelligent
control characterized by the use of expert knowledge on the control strategy and/or the behavior of
the controlled plant. This expert knowledge is represented by means of IF-THEN rules and linguistic
variables. Attributes or values of these linguistic variables are linguistic terms associated with fuzzy
sets, a generalization of ordinary (“crisp”) sets. Fuzzy set theory is the theoretical basis underlying
information processing in fuzzy control systems. From the systems theory’s view, a fuzzy controller
is a static nonlinear transfer element incorporated into a control loop. This gives rise to methods
for analysis and systematic design. Fuzzy systems may perform different tasks within an automatic
control system leading to different structural schemes. Nowadays, fuzzy control systems are success-
fully applied in many technical and non-technical fields. The application of fuzzy control systems is
supported by numerous hardware and software solutions.

1
1 Introduction

Fuzzy control has been a new paradigm of automatic control since the introduction of fuzzy sets by
L. A. Zadeh in 1965. Its rationale can be summarized by the statement of Zadeh “As complexity rises,
precise statements lose meaning and meaningful statements lose precision.” Thus, fuzzy control is an
attempt to meet the challenges of increasing complexity of the processes to be controlled and of the
tasks to be solved by automatic control systems.
To be more concrete, fuzzy control may be an advantageous alternative to conventional control tech-
niques if

• the process to be controlled exhibits a pronounced nonlinear behavior,

• no mathematical model of the process is available because the modeling effort is unacceptably
high or the process is not well understood,

• expert knowledge plays a key role in controlling the process and should be acquired and used
for automatic control, or

• a multidimensional nonlinear relationship (e. g. a control law) should be represented such that
it can be understood and modified easily.

Fuzzy control systems may be considered under various aspects: A fuzzy controller may be seen as a
nonlinear controller described by linguistic rules rather than differential equations. Or a fuzzy control
system may be seen as the implementation of the control strategy of a human expert. Understanding
the functioning of fuzzy control systems, i. e. the information processing taking place within the fuzzy
control system and its interaction with the plant and other components of the automatic control system
requires knowledge of fuzzy logic and control theory.
The aim of this chapter, therefore, is

• to introduce the basic ideas of fuzzy control by means of a simple example (Section 2),

• to provide the essential theoretical bases of fuzzy systems (Section 3), and

• to discuss the control issues of fuzzy control (Section 4).

2 Fuzzy Control - A Simple Example

2.1 Example

In the following section, a simple and illustrative example will be used to explain information pro-
cessing in fuzzy systems:

Example 1 (Control of room temperature) The temperature of a room equipped with a hot water
heating should be controlled by adjusting the position of the valve at the radiator (see Fig. 1). A
human being would use meta-rules, such as
If things are not OK but change in the right direction then maintain present settings
or more specifically
If the temperature is too warm but decreases, then leave valve position unchanged or
If the temperature is too cold and decreases, then increase the valve opening significantly.

2
Starting from these meta-rules, an experienced user would develop a set of control rules which are
more specific regarding the linguistic description of the values of temperature, temperature change,
and change of valve position.

valve room
position temperature temperature
controller

control
strategy

Figure 1: Schematic representation of the control for Example 1

Fuzzy systems provide a means to represent and process expert knowledge as stated in the example
above. By treating them as knowledge-based systems, the separation of knowledge representation and
information processing is realized. Knowledge is represented in form of rules and the meaning of the
expressions or symbols appearing in them. In Example 1, the knowledge consists of the control rules
and the meaning of the linguistic labels describing the values of temperature, temperature change, and
change of valve position (valve change for short). The formal concepts for knowledge representation
in fuzzy systems are linguistic variables and fuzzy IF-THEN rules presented in the next subsection.
The inference engine forms the core of the information processing components. As part of a control
system, a fuzzy system/controller usually processes numerical inputs to numerical outputs. Therefore,
fuzzification and defuzzification supplement inference:

• Fuzzification: Transformation of numerical values, e. g. measurements, into a fuzzy represen-

tation of the input situation,

• Inference: Transformation of the fuzzy input representation into a fuzzy decision, and

• Defuzzification: Transformation of the fuzzy decision into a real decision, e. g. a real value of
a manipulated variable.

Fig. 2 shows a schematic representation of a fuzzy system whose components will be described in the
following using Example 1 as a basis.

2.2 Fuzzy Sets, Linguistic Variables and Fuzzy IF-THEN Rules

By means of Example 1, it will be shown first how the formal concepts of a linguistic variable with
their linguistic terms and membership functions and of a fuzzy rule are used to represent the available
knowledge. The notion of a linguistic variable formalizes the practices of many domains to describe
the values of certain variables in terms of natural language. For Example 1, a linguistic variable
is the temperature deviation expressing the discrepancy between the desired and the actual room
temperature, denoted as TDev = Tdesired − Troom , with the linguistic terms TOO WARM, OK and TOO
COLD. A second one is the temperature change ∆T with the terms {INCR, EQUAL, DECR} where
INCR stands for INCREASING and DECR for DECREASING. Temperature change is the difference

3
 Rule base
x1 y1
IF
.... Condition

THEN
xs Conclusion yv

Fuzzification Inference Defuzzification

Numerical Membership values Membership values Numerical

inputs to linguistic terms to linguistic terms outputs
of input variables of output variables

Figure 2: Structure of a fuzzy system with numerical inputs x 1 , . . . , xs and numerical outputs y1 , . . . , yv

between the current room temperature and the room temperature at the last time instant. Assuming a
constant sampling time the temperature difference is proportional to the temperature trend. It should
be noted that the expression of temperature in the example has two different ‘meanings’. Specification
is done by using different linguistic terms: TOO WARM specifies a difference between the desired
and the room temperature and INCR a temperature change. In contrast to this, the term WARM could
characterize the room temperature itself.

A third linguistic variable is the valve change y of the radiator, the output of the fuzzy controller.
Valve change is an incremental variable, such that the actual valve opening results from the previous
opening plus valve change both expressed in percent. The term set {NB, NS, ZE, PS, PB} uses
standardized names with the typical abbreviations B (big), M (medium), N (negative), P (positive),
S (small), ZE (zero). These abbreviations will be combined in names as NB (negative big) and so
on. In practical applications, a linguistic variable usually has between two and seven linguistic terms.
This corresponds to the result of psychological investigations stating that human beings differentiate
a maximum of five to seven objects at the same time.

The next problem is to define the ‘meaning’ of each linguistic term. In many real-world problems,
the decision, whether a given x (e. g. a temperature deviation) satisfies a certain property A (e. g. TOO
COLD) or not is impossible or not reasonable. In the example above, a human being would not
consider a small temperature deviation (e. g. TDev = 0.01 K) as TOO COLD, whereas deviations of
2 K or 5 K probably would be felt as TOO COLD, to a certain extent at least. In other words, the
membership of x in the subset A should be a matter of degree as x satisfies the property up to a certain
degree.

According to set theory, each ordinary subset A of the universe of discourse X is determined by its
characteristic function µA : X → {0, 1}. This means that µA (x) = 1 when x is an element of A and
zero when it is not. The value of µA (x) can be interpreted as the truth value of a proposition ‘x is an
element of A’ relating set theory with logic.

4
Zadeh proposed to introduce a fuzzy set as a generalization of ordinary (“crisp”) sets. This means that
the proposition of ‘x is an element of A’ is no longer true or false, but may be true with a certain degree
(fuzzy truth value). The characteristic function µA (x) or membership function of the fuzzy (sub-) set
A is allowed to assume real values between 0 and 1:

µA : X → [0, 1]. (1)

Such fuzzy sets associated with the linguistic terms in Example 1 are depicted in Fig. 3 and Fig. 4,
respectively.

TOO WARM OK TOO COLD DECR EQUAL INCR

1 1

0.8 0.8
µ

µ
0.6 0.6

0.4 0.4

0.2 0.2

0 0
−10 −5 0 5 10 −2 −1 0 1 2
TEMP−DEVIATION TDev [K] TEMP−CHANGE ∆ T [K]

Figure 3: Membership functions and fuzzification for temperature deviation (left) and temperature
change (right)

Statements as, for example, ‘TDev is T OO COLD’ are called fuzzy propositions as the truth value of
such a statement is a matter of degree. It is determined by the membership degree of TDev in the fuzzy
set labeled TOO COLD. Using connectives such as AND and OR, compound fuzzy propositions can
be formed.
A fuzzy IF-THEN rule or fuzzy conditional statement is expressed as

IF <fuzzy proposition> THEN <fuzzy proposition>

where <fuzzy proposition> is a simple or compound fuzzy proposition. For Example 1 a fuzzy rule
of a controller might be ‘IF temperature deviation is TOO COLD AND temperature change is INCR
THEN valve change is ZE’. The IF part is called premise, condition or antecedent, the THEN part
conclusion or consequence.
In Example 1, the rule premises contain the two linguistic input variables temperature deviation TDev
and temperature change ∆T , while the conclusions contain the linguistic output variable valve change
y. The complete rule base is shown in Table 1.

2.3 Fuzzification - From Measurements to a Fuzzy Representation of the Input Situation

The inputs of a fuzzy system, especially a fuzzy controller, are (crisp) values of some variables,
e. g. measurement signals. The vector of these values characterizes an input situation which may
be the system state, for example. Likewise, rule premises specify such input situations, but this
specification uses linguistic terms for the values of the input variables. In order to determine the

5
 Table 1: Rule base of Example 1
TDev
y T OO WARM OK T OO COLD
INCR R1 : NB R2 : NS R3 : ZE
∆T EQUAL R4 : NS R5 : ZE R6 : PS
DECR R7 : ZE R8 : PS R9 : PB

degree of fulfillment of the (compound) fuzzy proposition in the premise the truth values of the simple
propositions have to be known. It is the task of fuzzification to provide these values.
If the input variables assume the values of TDev = 8 K and ∆T = 0.2 K, fuzzification in Exam-
ple 1results in

temperature deviation TDev : µT OO WARM (8) = 0.0, µOK (8) = 0.2, µT OO COLD (8) = 0.8,
temperature change ∆T : µINCR (0.2) = 0.1, µEQUAL (0.2) = 0.9, µDECR (0.2) = 0.0.

A graphic explanation is depicted in Fig. 3.

This transformation is unique, but not one to one in general. In the example, only ∆T ≥ 2 can be
deduced for a given µINCR (∆T ) = 1, µEQUAL (∆T ) = 0.

2.4 Inference - From a Fuzzy Input Representation to a Fuzzy Decision

Inference essentially consists of three steps, namely,

• aggregation,

• activation, and

• accumulation.

Aggregation A rule premise in general is a compound fuzzy proposition (e. g. an AND connection
of two propositions with TDev and ∆T ). Its degree of fulfillment µPk results from the aggregation of
the truth values of the simple propositions given by fuzzification. The operations have to be chosen
in accordance with the connectives (AND, OR) between simple propositions. The connective AND
is related to the intersection, OR to the union of two (fuzzy) sets.

Table 2: Intersection (left) and union (right) of crisp sets defined by their characteristic functions

µB (x) = 0 µB (x) = 1 µB (x) = 0 µB (x) = 1

µA (x) = 0 µA∩B (x) = 0 µA∩B (x) = 0 µA (x) = 0 µA∪B (x) = 0 µA∪B (x) = 1
µA (x) = 1 µA∩B (x) = 0 µA∩B (x) = 1 µA (x) = 1 µA∪B (x) = 1 µA∪B (x) = 1

The definitions in Table 2, valid for crisp sets, have to be generalized for fuzzy sets. According to the
original proposal of Zadeh, the intersection A ∩ B and the union A ∪ B can be defined point-wise using

6
the respective membership degrees

µA∩B (x) = min{µA (x), µB (x)}, (2)

µA∪B (x) = max{µA (x), µB (x)}. (3)

For Example 1, the results of aggregation using a minimum for the AND connective are given in
Table 3.

Table 3: Aggregation for Example 1 with the minimum operation (TDev = 8 K, ∆T = 0.2 K)
TDev
µT OO WARM (TDev ) = 0.0 µOK (TDev ) = 0.2 µT OO COLD (TDev ) = 0.8
µINCR (∆T ) = 0.1 µP1 = 0.0 µP2 = 0.1 µP3 = 0.1
∆T µEQUAL (∆T ) = 0.9 µP4 = 0.0 µP5 = 0.2 µP6 = 0.8
µDECR (∆T ) = 0.0 µP7 = 0.0 µP8 = 0.0 µP9 = 0.0

Activation A fuzzy IF-THEN rule is a connection of two (compound) fuzzy propositions. Hence,
this connective has to be interpreted within the framework of set theoretic or logical operators. The
simplest interpretation is that of the conjunction of premise and conclusion, such that the appropriate
operation is the minimum. Thus, the result of activation µCk of a rule k is the minimum of the degree
of fulfillment µPk and the fuzzy set in the conclusion. In other words, the fuzzy set in the conclusion
is clipped to µPk .

In Example 1, activation leads to nonempty fuzzy sets as depicted in Fig. 4a for rules 2,3,5, and 6,
only.

a) Activation b) Accumulation c) Defuzzification

NB NS ZE PS PB NB NS ZE PS PB NB NS ZE PS PB
1 1 1

0.8 0.8 0.8

0.6 0.6 0.6

µ

0.4 0.4 0.4

0.2 0.2 0.2

0 0 0
−20 0 20 −20 0 20 5.83
VALVE CHANGE y [%] VALVE CHANGE y [%] VALVE CHANGE y [%]

Figure 4: Results of a) activation (only positive for rules R 2 , R3 , R5 , R6 ), b) accumulation, and c)

defuzzification for Example 1

7
Accumulation Usually, a rule base is interpreted as a disjunction of rules i. e. rules are seen as inde-
pendent “experts”. Accumulation has the task to combine the individual “expert statements”, which
actually are fuzzy sets of recommended output values. Consequently, an appropriate accumulation
operation is the maximum. The maximum of the activated rule conclusions maxY {µCk (y)} is a fuzzy
set µ(y) over the domain Y of the output variable. The membership degree µ(y) can be interpreted as
the degree to which the value y is suggested by the “expert committee” to be the real output value.
In Example 1, accumulation results in a fuzzy set µ(y) as shown in Fig. 4b.
To summarize, inference yields the output fuzzy set as the maximum of all clipped fuzzy sets of the
linguistic terms of the output variable. The clipping results from applying the minimum operation
between each output fuzzy set referred to in the rule conclusion and the degree of fulfillment of the
rule premise. This degree of fulfillment is calculated from the results of fuzzification by applying the
minimum for the AND and maximum for the OR connective. This inference scheme, using minimum
and maximum operations is called max-min inference.

2.5 Defuzzification - From a Fuzzy Decision to a Real Decision

As inference results in a fuzzy set, the task of defuzzification is to find the numerical value which
“best” comprehends the information contained in this fuzzy set. A frequently used method is the
so-called Center-of-Gravity defuzzification (CoG; also called Center-of-Area defuzzification COA):
R
µ(y)y dy
Y
y= R , (4)
µ(y)dy
Y

which chooses the y-coordinate of the center of gravity of the area below the graph µ(y). This de-
fuzzification can be interpreted as a weighted mean, i. e. each value y is weighted with µ(y) and the
integral in the denominator serves for normalization. In Example 1, defuzzification of µ(y) using
Center-of-Gravity defuzzification yields a value of 5.83 %, as shown in Fig. 4c.
The characteristic surface of the fuzzy controller or control surface, that is the graphic representa-
tion of the function y(TDev , ∆T ), is depicted in Fig. 5. Here, the tasks of fuzzification, inference,
and defuzzification have been performed for all possible combinations of TDev , ∆T in the universe of
discourse (with some reasonable discretization).

3 Fuzzy Logic-related Issues in Fuzzy Control

In the previous section, only the simplest possible type of a fuzzy system has been discussed. In
this section the theoretical fundamentals of fuzzy systems are introduced. This introduction goes as
far as necessary for the understanding of key concepts of knowledge representation and information
processing in fuzzy systems. Furthermore, it is intended to reveal alternatives to choices made in the
previous section and again uses Example 1 for illustration.

3.1 Fuzzy Sets and Operations

A fuzzy set A with membership function µA (x) is completely characterized by the set of pairs A =
{(x, µA (x)) | x ∈ X}. For finite universes of discourse, this notation of fuzzy sets may be appropriate. In
fuzzy control, where the universes of discourse are continuous or contain a large number of elements

8
 30

20

VALVE CHANGE y
10

−10

−20

−30

2
1 10
0 5
−1 0
−5
−2 −10
TEMP−CHANGE ∆ T TEMP−DEVIATION TDev

Figure 5: Control surface of the fuzzy controller in Example 1

(due to discretization), fuzzy sets are more reasonably expressed in the parametric form. Common
types of membership functions include triangular, trapezoidal, or Gaussian functions. In addition, a
parametric form is a prerequisite for parameter optimization in fuzzy systems (see Optimization of
Fuzzy Systems).

Parametric membership functions Trapezoidal membership functions (see Fig. 6c) are given by


 0 if x ≤ m1 − a


 1

 1 + (x − m1 ) if m1 − a < x ≤ m1

 a
µ(x) = 1 if m1 < x ≤ m2 (5)

 1
1 − (x − m2 ) if m2 < x ≤ m2 + b





 b
0 if x > m2 + b.

Special cases of trapezoidal membership functions are rectangular (a = b = 0, Fig. 6a), triangular
(m = m1 = m2 , Fig. 6b), and singleton (m = m1 = m2 , a = b = 0, Fig. 6d) membership functions.
Examples of smooth membership functions are Gaussian

(x − m)2
−
µ(x) = e 2a2 (6)

(Fig. 6e) and cos2 membership functions (Fig. 6f):


0 if x < m − a


π
µ(x) = cos2 ( (u − m)) if m − a ≤ x ≤ m + a (7)

 2a
0 if x > m + a.

Properties of fuzzy sets In analogy with ordinary set theory, fuzzy set theory defines certain prop-
erties of fuzzy sets. A fuzzy set is called convex, if

∀x1 , x2 ∈ X ∀λ ∈ [0, 1] : µA (λx1 + (1 − λ)x2 ) ≥ min{µA (x1 ), µA (x2 )}. (8)

9
 a) Rectangular MBF (crisp) b) Triangular MBF (fuzzy) c) Trapezoidal MBF (fuzzy)

1 1 1
µ(x)

µ(x)

µ(x)
0 0 0
m1m2 m−a m m+b m1−a m1 m2 m2+b
x x x
d) Singleton MBF (crisp) e) Gaussian MBF (fuzzy) f) cos2−Function MBF (fuzzy)

1 1 1
µ(x)

µ(x)

µ(x)
0 0 0
m m−a m m+a m−a m m+a
x x x

Figure 6: Examples of parametric membership functions

The support of a fuzzy set A is the ordinary set that contains all elements of the domain of A with a
non-zero membership degree in A and will be denoted as S(A), i. e. S(A) = {x ∈ X | µ A (x) > 0}. For
convex fuzzy sets, the support is a convex set, e. g. an interval for scalar, real x. The ordinary set
consisting of all elements of the domain of A with a membership degree in A of at least α is called
α-cut and denoted as Aα , i. e. Aα = {x ∈ X | µA (x) ≥ α}. The least upper bound of the membership
function is called the height of the fuzzy set, i. e. hgt(A) = sup x∈X µA (x). A fuzzy set with hgt(A) = 1
is
Psaid to be normal. In the case of a discrete universe, the cardinality of a fuzzy set A is |A| =
n
R i=1 A (xi ) (n number of elements of the universe). In the case of a continuous universe, it is |A| =
µ
X µA (x)dx.

Operations on fuzzy sets Notions like equality and inclusion of two fuzzy sets are directly derived
from ordinary set theory. In contrast to this, different definitions are possible for connectives like
intersection, union, or complement, which only coincide for the special case of ordinary sets. This
may be an advantage, as it allows for a flexible and context-dependent choice of them, but it is also a
challenge for the user to choose the appropriate ones for an application at hand.

A fuzzy set A is a subset of a fuzzy set B, i. e. A ⊆ B, if for all x of the universe µ A (x) ≤ µB (x). Two
fuzzy sets A and B are equal, if A ⊆ B and B ⊆ A hold for all x in X.

To define fuzzy set connectives, the definitions of connectives for ordinary set are restated first using
the concept of a characteristic function. The intersection of two ordinary sets A, B is defined as the set
of all elements contained in both sets

A ∩ B = {x | x ∈ X and µA∩B (x) = 1} = {x | x ∈ X and x ∈ A and x ∈ B}

(9)
= {x | x ∈ X ∧ µA (x) = 1 ∧ µB (x) = 1}.

The union of A, B is defined as the set of those elements which are contained in one of both sets at

10
least
A ∪ B = {x | x ∈ X and µA∪B (x) = 1} = {x | x ∈ X and (x ∈ A or x ∈ B)}
(10)
= {x | x ∈ X ∧ (µA (x) = 1 ∨ µB (x) = 1)}.

The complement of ordinary set A is defined as the set of all elements not contained in A

Ac = {x | x ∈ X and µAc (x) = 1} = {x | x ∈ X and x ∈

/ A} = {x | x ∈ X and µA (x) = 0}. (11)

The definitions of intersection and union of fuzzy sets (2), (3) as given in Section 2 obviously gener-
alize the operations defined for ordinary sets as well as the definition of the complement of a fuzzy
set Ac according to
µAc (x) = 1 − µA (x). (12)

As in ordinary set theory, where the pair of a set and its complement forms a partition of the universe,
the pair of fuzzy sets (A, Ac ) is called a fuzzy partition of X, since µA (x) + µAc (x) = 1. More generally,
a m-tuple (A1 , . . . , Am ) of non-zero fuzzy sets with
m
X
∀x ∈ X : µAi (x) = 1 (13)
i=1

is called a fuzzy partition of X. In fuzzy control or modeling it is often advisable to define the fuzzy
sets of the linguistic terms of a linguistic variable such that they form a fuzzy partition.
In ordinary set theory (P (X), ∩, ∪, ()c ), where P (X) denotes the power set of X (i. e. the set of all
subsets of X), represents a Boolean algebra (i. e. a complemented distributive lattice). The set of all
fuzzy sets on X, F (X), together with the operations (2), (3), and (12) shares most of the properties of
a Boolean algebra, i. e.

Commutativity: A ∪ B = B ∪ A, A ∩ B = B ∩ A
Associativity: A ∪ (B ∪C) = (A ∪ B) ∪C, A ∩ (B ∩C) = (A ∩ B) ∩C
Idempotence: A ∪ A = A, A ∩ A = A
Distributivity: A ∪ (B ∩C) = (A ∪ B) ∩ (A ∪C), A ∩ (B ∪C) = (A ∩ B) ∪ (A ∩C)
De-Morgans Laws: (A ∪ B)c = Ac ∩ Bc , (A ∩ B)c = Ac ∪ Bc

The only Boolean law of ordinary set theory, which is no longer true is the excluded-middle law, is

A ∩ Ac 6= 0,
/ A ∪ Ac 6= X.

Hence, (F (X), min, max, 1 − ()) has the properties of a distributive lattice. Therefore, the operations
in (2), (3), and (12) play a special role in fuzzy set theory. From a practical point of view, they lead
to very simple algorithms. However, they are often too simplistic models of fuzzy set operations (see
Example 2) and are, as mentioned above, not the only possible extensions of ordinary set operations.
Below, a second example will be considered to illustrate differences between definitions of the above
operations on fuzzy sets.

Example 2 (Selection of an optimal route) To get from point A to B by car different routes can be
chosen. A fuzzy system containing rules like
IF the distance between A and B is SHORT AND the driving speed is FAST
THEN the route is GOOD
should make a decision for finding the best of several routes. From an optimization point of view, the

11
fuzzy system selects a route which is optimal for both sub-conditions distance and driving speed (op-
timal solution of a multi-criteria optimization problem) or an appropriate compromise route (optimal
solution of a single-criterion optimization problem).
Assume, there are two candidate routes with their respective membership degrees in the fuzzy sets
SHORT and FAST: µSHORT (route 1) = 0.7, µSHORT (route 2) = 0.5, µFAST (route 1) = 0.5, and
µFAST (route 2) = 0.5. Which is the best route? Intuitively, route 1 would be preferred. But with
minimum for the intersection there is no preference for a route because both routes belong to the
fuzzy set of SHORT AND FAST routes with degree 0.5.

Possible definitions for intersection operations can be derived within the framework of triangular
norms. A function > : [0, 1] × [0, 1] → [0, 1] is called t-norm, if for all u, v, w ∈ [0, 1] the following
conditions hold:

i. >(u, 1) = u, >(u, 0) = 0 (boundary conditions),

ii. u ≤ v ⇒ >(u, w) ≤ >(v, w) (monotonicity),
iii. >(u, v) = >(v, u) (commutativity),
¡ ¢ ¡ ¢
iv. > u, >(v, w) = > >(u, v), w (associativity).

Applying De Morgan’s Law and the negation (12), a t-conorm ⊥ with ⊥(u, v) = 1 − >(u, v) can be
defined as respective union operation for any t-norm. The neutral element of a t-conorm is zero. Also,
the conditions ii.–iv. hold.
Generally, t-norms and t-conorms observe certain of the Boolean laws only, but never all. This may
restrict the possibility of transforming compound expressions into simpler ones. In fuzzy control,
however, this plays a minor role.
Examples of particular t-norm/t-conorm pairs are the minimum/maximum > m /⊥m as in (2), (3), the
algebraic product/algebraic sum >a /⊥a , the bounded difference/bounded sum >b /⊥b and the drastic
product/drastic sum >d /⊥d given by:
def def
>m (u, v) = min{u, v} ⊥m = max{u, v} (14a)
def def
>a (u, v) = uv ⊥a (u, v) = u + v − uv (14b)
def def
>b (u, v) = max{0, u + v − 1} ⊥b (u, v) = min{1, u + v} (14c)
 
u, v = 1
 u, v = 0

def def
>d (u, v) = v, u = 1 ⊥d (u, v) = v, u = 0 (14d)
 
0, otherwise 1, otherwise
 

Considering two fuzzy sets A, B, a t-norm > induces a connective for the intersection A ∩ > B by the
point-wise definition of
µA∩> B (x) = > (µA (x), µB (x)) . (15)
Analogously, the connective for the union A ∪⊥ B is induced by a t-conorm ⊥:

µA∪⊥ B (x) = ⊥ (µA (x), µB (x)) . (16)

In Fig. 7 the intersection and union of crisp and fuzzy sets are compared with the operations for
fuzzy sets exemplarily being the minimum and algebraic product for intersection and maximum and
algebraic sum for union.

12
 (crisp) set A fuzzy set A fuzzy set A
1 1 1

µA

A
µ

µ
0 0 0
(crisp) set B fuzzy set B fuzzy set B
1 1 1
µB

B
µ

µ
0 0 0
intersection A ∩ B min intersection A ∩ B alg. prod. intersection A ∩ B
1 1 1

B
µA ∩ µB

µ ∩µ
µ ∩µ

A
A

0 0 0
union A ∪ B max union A ∪ B alg. sum union A ∪ B
1 1 1

B
µA ∪ µB

µ ∪µ
µ ∪µ

A
A

0 0 0

Figure 7: Intersection and union of ordinary and fuzzy sets

Generally, the following relations hold for every t-norm > and t-conorm ⊥, respectively:

>d (u, v) ≤ >(u, v) ≤ >m (u, v) ≤ ⊥m (u, v) ≤ ⊥(u, v) ≤ ⊥d (u, v). (17)

Hence, the drastic product gives the lower bound and the minimum the upper bound for every t-norm,
while the maximum represents the lower and the drastic sum the upper bound for every t-conorm.
The following example is a continuation of Example 2.

Example 3 Using the algebraic product for intersection, route 1 and route 2 belong to the fuzzy set
of SHORT AND FAST routes with a degree of 0.35 and 0.25, respectively. Obviously, the alge-
braic product operator leads to the preference suggested by intuition. Assume, there is a third route
with µSHORT (route 3) = 1 and µFAST (route 3) = 0.35, then this route also belongs to the fuzzy set of
SHORT AND FAST routes with a degree of 0.35. Routes 1 and 3 are not distinguishable, which again
may contradict intuition as route 3 is much shorter, but only a little slower.

In practical problems, it may be important to have other than the given t-norms and t-conorms, which
are more appropriate to the situation under consideration. Here, parametric families of t-norms allow
us to adapt the intersection and union operations. An example is the Yager family with a parameter p:
n o n o
def p p 1/p def p p 1/p
> p (u, v) = 1 − min [(1 − u) + (1 − v) ] , 1 , ⊥ p (u, v) = min [u + v ] , 1 . (18)

> p increases from >0 = >d to >∞ = >m , including >1 = >b . Analogously, the dual ⊥ p decreases
from ⊥0 = ⊥d to ⊥∞ = ⊥m . Hence, the Yager family covers the whole range of t-norms and t-
conorms, respectively, in the sense of the inequalities (17).
In ordinary set theory complementation together with intersection or union suffice to define all further
connectives for two sets (as two operations suffice to define all operations of Boolean logic). In

13
contrast to this, there are no unique definitions for these operations in fuzzy set theory, and, in addition,
further possibilities of aggregating fuzzy sets exists. Of these general aggregation operations, the so-
called averaging operators play an important role in decision making or fuzzy control.
An averaging operator N is a function N : [0, 1] × [0, 1] → [0, 1] satisfying for all u, v, w ∈ [0, 1]

i. N(0, 0) = 0, N(1, 1) = 1 (boundary conditions),

ii. u ≤ v ⇒ N(u, w) ≤ N(v, w) (monotonicity),
iii. N(u, v) = N(v, u) (commutativity),
iv. N(u, u) = u (idempotency),
v. N is continuous.

Note, that averaging operators are not associative. For every averaging operator N the relation
>m (u, v) ≤ N(u, v) ≤ ⊥m (u, v) holds. Examples of averaging operators are the arithmetic mean, the
harmonic mean, the generalized p-mean, or, more generally, the ordered weighted averaging (OWA)
operators introduced by Yager. An OWA operator of dimension n is a mapping f : R n → R that has
T
an associated vector P w = (w1 , . . . , wn ) such that wi ∈ [0, 1] and w1 + · · · + wn = 1. Furthermore,
f (u1 , . . . , un ) = j w j b j where b j is the j-th largest of the ui . Special cases are the maximum, the
arithmetic mean, and the minimum:

f ∗ : w = w∗ = (1, 0, . . . , 0)T and f ∗ (u1 , . . . , un ) = max{u1 , . . . , un },

X
favg : w = wavg = (1/n, . . . , 1/n)T and favg (u1 , . . . , un ) = 1/n ui ,
i
T
f∗ : w = w∗ = (0, 0, . . . , 1) and f∗ (u1 , . . . , un ) = min{u1 , . . . , un }.

This is to continue Example 3.

Example 4 If the fuzzy set of SHORT AND FAST routes is created using the arithmetic mean f avg ,
then route 1, route 2, and route 3 belong to it with a degree of 0.6, 0.675, and 0.5, respectively. These
degrees reflect the preference according to intuition as sketched above.

3.2 Types of Rule-based Fuzzy Systems

Rule-based fuzzy systems can be differentiated according to the types of fuzzy IF-THEN rules. There
are three types of practical relevance:

1. rules of the type ‘IF x1 = A1,i1 AND ... AND xs = As,is THEN y = Bi ’ leading to linguistic or
Mamdani-type fuzzy systems as discussed in Section 2,
2. rules of the type ‘IF x1 = A1,i1 AND ... AND xs = As,is THEN y = fi (x1 , . . . , xs )’ used in the
so-called Takagi-Sugeno-type fuzzy systems (note, that the arguments of f i which in most cases
is a linear function are not necessarily the same as in the premise.), and
3. rules of the type ‘IF x1 = A1,i1 AND ... AND xs = As,iS THEN y = ci ’, where ci is a constant,
being a special case of the other two rule types. The resulting fuzzy systems are called singleton
fuzzy systems.

Here, the A j,i j and Bi denote linguistic terms of the linguistic variables x j and y, respectively. In
addition, a fuzzy rule may be assigned an individual weight or confidence value (from the interval
[0,1]) according to the degree of certainty or confidence.

14
For Example 1, as alternative to the Mamdani-type fuzzy controller presented in Section 2 (Exam-
ple 1a)

• Example 1b: a Mamdani-type fuzzy controller with the rule base in Table 1, the algebraic
product as intersection, the bounded sum as union, and the output membership functions as
shown in Fig. 4,
• Example 1c: a singleton fuzzy controller where the constants in the conclusion are labeled
according to Fig. 9b (c1 = −20, c2 = c4 = −10, c3 = c5 = c7 = 0, c6 = c8 = 10, c9 = 20), with
the rule base in Table 1, and the algebraic product as intersection,
• Example 1d: a Takagi-Sugeno-type fuzzy controller with linear functions f i (x1 , x2 ) = ai,0 +
ai,1 x1 + ai,2 x2 in the conclusions, the rule base in Table 4, and the algebraic product as intersec-
tion,

will be considered in the following subsection. All controllers posses input membership functions
as shown in Fig. 3. These versions cover the most important types of fuzzy controllers in practical
applications.

Table 4: Rule base for a Takagi-Sugeno-type fuzzy controller (Example 1d)

TDev
ai = (ai,0 , ai,1 , ai,2 )T TOO WARM OK TOO COLD
INCR a1 = (−10, 0.5, 2.5)T a2 = (−5, 1.0, 2.5)T a3 = (0, 0.5, 2.5)T
∆T EQUAL a4 = (−5, 0.5, 5)T a5 = (0, 1.0, 5)T a6 = (5, 0.5, 5)T
DECR a7 = (0, 0.5, 2.5)T a8 = (5, 1.0, 2.5)T a9 = (10, 0.5, 2.5)T

3.3 Information Processing in Fuzzy Systems

In Section 2, information processing in fuzzy systems is explained by means of Example 1. In order

to make the descriptions simple, no alternatives to the applied operations have been presented. But
this description and Section 3.1 imply that there are many alternatives. These alternatives concern
mainly the interpretation of connectives in fuzzy propositions and of fuzzy rules leading to different
inference schemes and strategies for defuzzification. Practically important alternatives will be briefly
described below.
Several inference mechanisms have been developed within the framework of approximate reasoning.
Mainly, they base either on the compositional rule of inference (using fuzzy set theory considerations)
or on the generalized modus ponens (analogous to the classical modus ponens). These inference
mechanisms differ in the interpretation of rules (logical implication vs. conjunction), the operations
used for logical or set connectives, the order of inference steps (composition based vs. individual-rule
based) etc. However, most fuzzy control algorithms used in practice employ an inference mechanism
based on individual-rule inference where rules are interpreted as conjunctions.

Aggregation In aggregation, different t-norms or t-conorms can be applied to model the connectives
between fuzzy propositions. However, in fuzzy control, where the premises usually contain AND
connectives only, the algebraic product is the most popular choice. For Example 1 aggregation with
the algebraic product yields the results shown in Table 5. They differ significantly from the minimum
aggregation in Table 3.

15
 Table 5: Aggregation for Example 1b-d with the algebraic product operation

TDev
µT OO WARM (TDev ) = 0.00 µOK (TDev ) = 0.20 µT OO COLD (TDev ) = 0.80
µINCR (∆T ) = 0.1 µP1 = 0.00 µP2 = 0.02 µP3 = 0.08
∆T µEQUAL (∆T ) = 0.9 µP4 = 0.00 µP5 = 0.18 µP6 = 0.72
µDECR (∆T ) = 0.0 µP7 = 0.00 µP8 = 0.00 µP9 = 0.00

Activation As fuzzy IF-THEN rules in fuzzy control are usually interpreted as a conjunction of the
premise and the conclusion, a t-norm is the appropriate operator. Mainly, the algebraic product is used
resulting in a scaling of the fuzzy set in the conclusion by the degree of premise fulfillment. Results
for the Example 1b are depicted in Fig. 8a. They also differ from Example 1a (Fig. 4a).

a) Activation b) Accumulation c) Defuzzification

NB NS ZE PS PB NB NS ZE PS PB NB NS ZE PS PB
1 1 1

0.8 0.8 0.8

0.6 0.6 0.6

µ

0.4 0.4 0.4

0.2 0.2 0.2

0 0 0
−20 0 20 −20 0 20 7.00
VALVE CHANGE y [%] VALVE CHANGE y [%] VALVE CHANGE y [%]

Figure 8: Results of inference and defuzzification using the product operator and bounded sum oper-
ator: a) aggregation, b) accumulation, c) defuzzification (Example 1b)

In the case of rules with additional confidence values these values are connected with the degree of
premise fulfillment by a t-norm, usually the algebraic product.

Accumulation If a rule base is interpreted as a disjunction of rules, the appropriate operator is a

t-conorm. In this case popular alternatives are the maximum and the bounded sum operators. An
alternative, motivated by the “expert” metaphor, is to combine the “expert opinions” by an averaging
operator, e. g. an ordered weighted averaging (OWA) operator. A possible choice is a weighted mean
of the activated fuzzy sets in the rule conclusions where the weights may be equal (no preference of
experts) or the confidence values.

The application of the bounded sum in Example 1b gives a fuzzy set with µ(y) as shown in Fig. 8b.

16
To summarize, inference yields the following output fuzzy set in case of t-conorm accumulation
µ ¶
r r ³ ´ r s
µ(y) = ⊥ µCk (y) = ⊥ > µPk , µBi(k) (y) = ⊥ > > µA j,i (k) (x j ), µBi(k) (y) , (19)
j=1 j
|k=1 {z } k=1 | {z } k=1 | {z }
accumulation activation aggregation
r s
where ⊥ denotes the r-ary extension of a t-conorm, > the s-ary extension of a t-norm (r number of
k=1 j=1
rules, s number of inputs). In case of an OWA accumulation, the output fuzzy set is given by
µ µ ¶ µ ¶¶
s s
µ(y) = f > > µA j,i (k) (x j ), µBi(k) (y) , . . . , > > µA j,i (k) (x j ), µBi(k) (y) , (20)
j=1 j j=1 j

with f (·) being an OWA operator.

Several inference schemes, named after the used operations, are especially popular in control ap-
plications, namely sum-prod inference, max-prod inference, and max-min inference. Aggregation
is mostly performed with the same operator as activation. For differences between these inference
schemes and the practical consequences see Analysis and Stability of Fuzzy Systems.

Defuzzification The understanding of what value “best” comprehends the meaning of the output
fuzzy set depends on the application. Therefore, several methods are employed. For applications
whit a continuously changing control signal being needed, the Center-of-Gravity (CoG) or Bisector-
of-Area (BOA) method are used. For non-convex fuzzy sets, these defuzzification methods may lead
to output values which have a very small or even a zero degree of membership in the inferred fuzzy
set, i. e. which are only a little or not supported by any of the rules. Maximum methods which select
only output values with a maximum membership degree prevent this effect. As several values may
have a maximum membership degree, a preference for example leads to the Leftmost-Maximum or
Rightmost-Maximum methods. The so-called Mean-of-Maximum method, however, which takes the
mean of all output values with maximum membership degree, may also lead to little supported values
in case of a non-convex fuzzy set. In comparison to CoG or BoA, maximum methods are computa-
tionally less complex. Their disadvantage is the generally discontinuous control signal produced. In
Table 6 the formulae of the mentioned defuzzification methods are summarized.
Defuzzification results for different methods applied to the inference results of Example 1a (Fig. 4b,
min-max inference) are shown in Fig. 9a. In general, the different defuzzification methods yield
different output values that are considered to best represent the inferred fuzzy set with µ(y).
A concept which generalizes these methods is the inference filter introduced by H. Kiendl. It allows a
continuous transition between the defuzzification methods mentioned above. The idea is to first filter
the membership function µ(y) to obtain an attractiveness function µ̂(y), and to secondly choose the
maximum value y = arg max µ̂(y), that is the most “attractive” value, as output value. The impulse
response h(y) of the inference filter is a parameterized function symmetric to zero. This function
changes its shape from a δ-function via a triangular function to a rectangular function of a specified
width by adjusting two parameters. The resulting output value changes from the value obtained with
the Mean-of-Maximum method via the value obtained by the Center-of-Gravity method to the mean
of the support of µ(y) by changing the shape parameter of h(y).
For Example 1b, the control surface of the Mamdani-type fuzzy controller with sum-prod inference
and Center-of-Gravity defuzzification is shown in Fig. 10a. Here, CoG defuzzification is a reasonable
choice. For bounded sum-prod inference, it leads to a value of 7 % as shown in Fig. 8c, which differs
significantly from 5.83% in Section 2 (Fig. 4c).

17
 Table 6: Defuzzification methods (s. t.: such that, sup: supremum, inf: infimum)
Method Formula
R
µ(y)ydy
Y
Center-of-Gravity (CoG) y= R
µ(y)dy
Y

Ry yR
max
Bisector-of-Area (BoA) y s. t. µ(y)dy = µ(y)dy
ymin y
R
ydy
y0 :µ(y0 )=supY µ(y)
Mean-of-Maxima (MoM) y= R
dy
y0 :µ(y0 )=supY µ(y)

Leftmost-Maximum (LM) y = inf{y0 ∈ Y | µ(y0 ) = sup µ(y)}

Rightmost-Maximum (RM) y = sup{y0 ∈ Y | µ(y0 ) = sup µ(y)}

For singleton and Takagi-Sugeno-type fuzzy controllers, defuzzification is integrated in activation

and accumulation. The output is calculated as a weighted mean of the singleton or function values,
respectively, where the weights are the degrees of rule fulfillment, i. e.
Pr
µPk ck
y = Pk=1r (singleton fuzzy system) (21)
k=1 µPk
Pr
µ f (x)
y = k=1Pr Pk k (Takagi-Sugeno-type fuzzy system). (22)
k=1 µPk

The control surface of the singleton fuzzy controller in Example 1c is given by

P9
µP (TDev , ∆T )ck
y(TDev , ∆T ) = k=1 Pk9 . (23)
k=1 µPk

The result is y(8, 0.2) = 7 with algebraic product aggregation (Table 5) (terms with non-zero elements
of µPk (8, 0.2) are c2 µP2 = −0.2, c3 µP3 = 0, c5 µP5 = 0, c6 µP6 = 7.2; 9k=1 µPk = 1) and defuzzification
P
is shown in Fig 9b.
For the Takagi-Sugeno-type fuzzy controller in Example 1d the control surface is given by
P9
µP (ak,0 + ak,1 TDev + ak,2 ∆T )
y(TDev , ∆T ) = k=1 k P9
k=1 µPk
= ã0 (TDev , ∆T ) + ã1 (TDev , ∆T )TDev + ã2 (TDev , ∆T )∆T (24)

(Fig. 10b). This control surface is similar to the one of the Mamdani-type fuzzy controller (Ex-
ample 1b) in Fig 10a. The main difference lies in the extrapolation behavior outside the region
[−10 K, 10 K] × [−2 K, 2 K]. Whereas the Mamdani-type controller extrapolates with constant val-
ues, the Takagi-Sugeno type controller extrapolates linearly. For the values TDev = 8 and ∆T = 0.2,
ã0 (8, 0.2) = 3.5, ã1 (8, 0.2) = 0.6, ã2 (8, 0.2) = 4.75 and y(8, 0.2) = 9.25 are obtained.
Under the assumptions of

• sum-prod inference,

18
 a) Defuzzification Methods b) Singleton Defuzzification

NB NS ZE PS PB NB NS ZE PS PB
1 1

MoM
0.8 LM RM 0.8

Results:
0.6 5.83 − CoG 0.6
µ

µ
8.00 − LM
8.12 − BoA
0.4 10.00 − MoM 0.4
12.00 − RM

0.2 0.2

0 0
CoG BoA 7.00
VALVE CHANGE y [%] VALVE CHANGE y [%]

Figure 9: a) Comparison of different defuzzification methods from Table 6 for Example 1a; b) Results
of algebraic product aggregation and defuzzification for the singleton system (Example 1c)

• output fuzzy sets with equal cardinality, and

• CoG defuzzification,

a Mamdani-type fuzzy system is equivalent to a singleton fuzzy system, where the singletons result
from an individual CoG defuzzification of the output fuzzy sets. It explains the equivalence between
the control surfaces of Examples 1b and c (10a). These assumptions are very reasonable in many
practical applications resulting in a very simple implementation of the fuzzy controller. Moreover,
the resulting control surface is piece-wisely multi-linear, a convenient form for the analysis of the
control loop (see Analysis and Stability of Fuzzy Systems).

Two-way fuzzy controller and hyperinference The so-called two-way fuzzy controller also intro-
duced by Kiendl is another type of fuzzy system. The rule base of this fuzzy system contains positive
rules as stated above and negative rules expressing explicit warnings or vetoes.
For a two-way fuzzy controller the inference scheme contains an additional processing step – the so-
called hyperinference. Inference is performed separately for positive and negative rules, resulting in
two fuzzy sets µ+ (y) for positive and µ− (y) for negative rules, respectively. Hyperinference combines
both fuzzy sets by applying a certain strategy, namely, the strong, the weak, or the fuzzy veto.

4 Control Issues in Fuzzy Control

4.1 Structures in Fuzzy Control

In Example 1, the fuzzy controller is directly incorporated in the basic control loop. That is, its
inputs are the control error, called temperature deviation, and the discrete derivative of the controlled
variable, called temperature change, and its output is the change of the manipulated variable, called
valve change. Hence, the resulting control law is of PI type. The fuzzy controller needs external

19
 30 30

20 20
VALVE CHANGE y

VALVE CHANGE y
10 10

0 0

−10 −10

−20 −20

−30 −30
2 2
1 10 1 10
0 5 0 5
−1 0 0
−1
−5 −5
−2 −10 −2 −10
TEMP−CHANGE ∆ T TEMP−DEVIATION TDev TEMP−CHANGE ∆ T TEMP−DEVIATION TDev

Figure 10: Control surface of the fuzzy controller a) Mamdani-type with sum-prod inference and CoG
defuzzification (Example 1b), b) Takagi-Sugeno-type fuzzy controller (Example 1d)

elements to form the derivative (or difference), to integrate (or to sum) its output, and adapt inputs
and output to signal ranges (scaling, normalization, and denormalization). Additional components
for an implementation include an analog-digital and a digital-analog converter, possibly a filter to
suppress measurement noise, elements for an anti-windup scheme, etc. The components necessary
for pre- and post-processing depends on the application and function of the fuzzy controller.
Apart from direct fuzzy control as in Example 1, fuzzy controllers can perform certain additional
functions in control systems. An overview of typical structural schemes of control systems incorpo-
rating fuzzy controllers is given in Table 7.
In the classification of control structures with a fuzzy controller as given in Table 7, there is no clear
cut between the classes. For instance, a Mamdani-type fuzzy controller which adapts the parameters
of a linear controller as a gain scheduler according to a desired reference trajectory can be realized as
a Takagi-Sugeno-type fuzzy system which is directly incorporated in the control loop.

4.2 Design of Fuzzy Control Systems

The design of fuzzy control systems covers

• the determination of inputs and outputs,

• the definition of the respective linguistic variables including the linguistic terms and their fuzzy
sets,

• the compilation of the rule base, and

• the choice of the inference strategy and defuzzification method.

Whereas the latter task requires knowledge of fuzzy theory and can therefore hardly be realized auto-
matically, the first three tasks can be handled by different approaches including heuristic, systematic
and partially automatic methods. The differences between theses approaches will be explained for
fuzzy control systems (Table 8 and Fig. 11), but similar approaches also exist for non-control appli-
cations.

20
 Table 7: Structures of control systems with a fuzzy controller
Description Structural scheme
Disturbance
Direct fuzzy control:
Fuzzy controller replaces conventional con- Setpoint r e u y
Fuzzy
troller, e. g. a PID controller, or manual con- - Controller
Plant

trol by operator.

Disturbance

Feedforward fuzzy control: Fuzzy

Fuzzy controller modifies the manipulated Controller

variable to compensate a measurable distur- y

Setpoint r e Linear u
bance or a change of the setpoint. - Controller
Plant

Fuzzy parameter-adaptive control:

Fuzzy
Fuzzy controller adapts parameters of lin- Controller
Disturbance
ear controller to changing operating con- Controller
Parameters
ditions (gain scheduling), or changes con- Setpoint r
e Linear u y
troller structure (switching D part). The ba- - Controller
Plant

sic control loop with a linear controller, e. g.

PID controller, is left unchanged.
Disturbance

Fuzzy model-based control: Setpoint r

e y
The controller incorporates a fuzzy plant Inverse
Fuzzy Model
u
Plant
-
model. Two different schemes exist: In-
ternal model control (IMC) as shown, and ym
Fuzzy
model-predictive control (MPC) containing Model

additionally an optimizer. Feedback

Filter
Fuzzy Controller

Fuzzy supervisory control:

Supervisory control works on the level of Fuzzy Super-
visory Contr.
human operation. Its general task is to im- Disturbance
prove process performance (e. g. stable op- Controller n
Setpoint r
eration, maximize product yield, minimize e u
Plant
y

energy consumption) accomplished by e. g. -

Controller 1
the appropriate setting of set-points for basic
control loops and switching or reparameteri-
zation of basic controllers.

21
 Data-based controller design Data-based model design
model identification for the model identification for the
operator's behavior plant behavior

Human operator Plant

Reference Controlled
variable w variable y

Manipulating
variable u

Knowledge-based model and

controller design
Asking the operator about its control
behavior and/or the plant's behavior

Figure 11: Different data-based and knowledge-based strategies for model and controller design

In the early years of fuzzy control, the primary approach was to interview human process operators
and to formalize the knowledge gained in the form of rules and membership functions of a fuzzy sys-
tem (knowledge-based controller design). Here, knowledge subsumes all types of qualitative (rules,
structure, properties) and quantitative information. This was not always successful for certain rea-
sons. To begin with, human knowledge contains unconscious parts, which can hardly be formalized
by rules. Moreover, many rules will be context sensitive, which results in incomplete rule premises
with missing context information. Furthermore, the complexity of human strategies may be very high,
quantification in the form of membership functions is difficult, and so on. Alternatively, the control
engineer who develops the fuzzy controller knows or learns how to control the process and captures
his knowledge. This may be advantageous as it prevents the communication losses between operator
and control engineer. Moreover, the control engineer may be more conscious of the control strategy,
which makes it easy to determine important variables, formulate rules, etc.

An alternative to this knowledge-based approach is to record data of process measurements, the set-
points, and the actions of the human process operator (manipulated variable). In the so-called data-
based controller design, a model of the human control behavior is established by identification of a
fuzzy system with a similar input-output behavior. Existing a-priori knowledge will be often used
to obtain appropriate structures for fuzzy systems (inputs, outputs, possible rules). The structures
may result in Mamdani-type fuzzy controllers (especially if the conclusions are directly referred to
manipulating variables) or Takagi-Sugeno-type fuzzy controllers (especially if the conclusions are
switching operations between different control strategies). An comprehensive overview of data-based
techniques for model or controller design is given in Data-Based Fuzzy Modeling.

Both approaches discussed do not require any plant model. This seems to be advantageous for a
simpler design, but it restricts the possibilities to simulate, to optimize, and to supervise fuzzy systems
similar to human strategies. As an example, a human process operator certainly has some perceptions

22
Table 8: Different data-based and knowledge-based strategies for model and controller (contr.) design
√
( - information given for the task, ? - aim of the task)
Data Knowledge Model Behavior
Plant Contr. Plant Contr. Plant Contr. Closed loop
√
Knowledge-based ?
controller design
√ √
Data-based ( ) ?
controller design
√ √
Model-based ?
controller design
√
Knowledge-based ?
model design
√ √
Data-based ( ) ?
model design
√ √
Analysis of closed- ?
loop systems

about the expected plant behavior and is able to detect deviations from it.
In contrast to these model-free approaches, a model-based controller design does not formalize human
strategies of problem solution but is based on a given problem description and a description of goals.
In case of fuzzy control, a fuzzy controller will be developed by means of a given model of the
plant and a definition of the desired closed-loop behavior (e. g. stability, performance, robustness).
Different approaches to designing for a model-based controller are used in practical applications:
As an example, the design can be accomplished by offline parameter optimization as presented in de-
tail in Optimization of Fuzzy Systems. The desired closed-loop behavior may be a minimal difference
(described by an integral criterion) between a given setpoint and the controlled variable. In optimiza-
tion, simulation-based techniques, which simulate the closed control loop with a given plant model
and a fuzzy controller, are very popular. The closed-loop behavior desired is only used for evaluation
and the choice of the best controller.
An other model-based approach is model-predictive fuzzy control (see Model-based Predictive Con-
trol). Here, one or more future values of the manipulated variable are found by an online optimization
using the fuzzy plant model and the desired closed-loop behavior. In this case, an explicit controller
structure is not necessary.
If the plant is modeled by a Takagi-Sugeno-type fuzzy system with linear functions as rule conclu-
sions, a Takagi-Sugeno-type controller may be designed using techniques from linear control design
(see Control of Linear Multivariable Systems) as e. g. pole placement. Each rule describes a fuzzy
region where a linear model, for which a controller is designed, is valid. This approach is called
Parallel Distributed Compensation. As the stability of the local controller designs does not imply
global stability more advanced techniques start with a formalization of the desired behavior of the
closed-loop system (e. g. a Lyapunov function to ensure the stability of the system, see Lyapunov
Stability). Here, appropriate controller parameters are found by optimization (e. g. by solving Linear
Matrix Inequalities, see Controller Design by LMI Approach).
The plant model required for a model-based controller design can be obtained by knowledge–based
or data-based modeling. Here, human knowledge or data of the plant behavior are obtained instead of

23
the knowledge or data of the control behavior. From a mathematical point of view, the techniques for
modeling the plant are the same as for modeling the control behavior. Modern data-based approaches
are able to find structures (inputs, outputs, rules) and parameters. Further elements of the fuzzy
system (fuzzy operators, types of membership functions, defuzzification method) are often fixed.
Data-based approaches preferably use Takagi-Sugeno-type fuzzy systems with linear functions as rule
conclusions. This makes well-known parameter estimation methods applicable. The remaining task
is to find an appropriate fuzzy representation (rules and membership functions) for the rule premises
(see System Identification using Fuzzy Models).

4.3 Analysis of Fuzzy Control Systems

Most fuzzy systems are nonlinear systems of a high complexity making the analysis of behavior and
properties an important and nontrivial task. There are several methods to analyze fuzzy systems and
their interactions with other systems which are described in Analysis and Stability of Fuzzy Systems.
Here, only the basic ideas shall be presented.
The overall objective of the analysis is to prove whether the performance specifications are met. Issues
in the analysis of fuzzy control systems can be divided into fuzzy logic and control and system theory
aspects.
The fuzzy logic aspect takes into account properties of the fuzzy system itself, for example the con-
sistency and completeness of the rule base, correctness of inference. For this purpose, methods have
been developed on the basis e. g. the theories of approximate reasoning and fuzzy relational equations.
From the control and system theory points of view, properties of the overall system consisting of
the plant and the control system, including the fuzzy system, are of interest, e. g. stability, robustness,
control performance. Many analysis techniques transform a fuzzy system into an equivalent nonlinear
function to describe the input-output behavior (see Closed-loop behavior). After this transformation,
methods from nonlinear system theory can be readily applied (see Basic Nonlinear Control Systems,
Control of Nonlinear Systems). Here, the typical structures of fuzzy controllers result in representa-
tions common to control theory, for example in linear parameter-varying (LPV) systems or in systems
consisting of parallel linear and nonlinear subsystems.
Another class of methods is specific for fuzzy systems. These methods typically represent the overall
control system and the plant as a fuzzy system. Analysis is performed on the level of linguistic terms
or fuzzy sets and, thus, results have the form of fuzzy statements. For this reason, such methods are
not well accepted.
Provided that a sufficiently accurate plant model exists, extensive simulation studies, apart from the-
oretical analysis, are a means for assessing the performance of the control system. Thus, properties
which cannot be proved theoretically are still checked.
Analysis should accompany system design and for this purpose a plant model is needed. However, a
rational for using a fuzzy control system is the lack of such a model. In this situation, an online super-
vision and analysis is applicable only. The supervisor might contain formal mathematical knowledge
and rule-based expert knowledge to evaluate the control system.

4.4 Applications

Since the 1980s, many applications of fuzzy systems in automatic control have been reported. This
section highlights some principal aspects of fuzzy control system applications. An overview on fields
of applications and descriptions of examples are contained in Fuzzy System Applications.

24
Applications of fuzzy systems can be found in various technical and non-technical domains. The
number of reported applications is higher for technical domains where the processes are difficult to
model due to complexity, nonlinearity, time-variance or uncertainty, and where experienced operators
are an essential part of the process control. Examples include the production of raw materials (e. g.
iron and steel production, cement production), environmental engineering (e. g. waste incineration,
water treatment plants), biotechnological processes, power systems, autonomous robots, automotive
systems. When the objective of automatic control is to enhance functionality and user friendliness
without high performance requirements, e. g. consumer products and household appliances, fuzzy
systems have been applied as well.
In the non-technical domains the dominating fields of applications are medicine and biomedical en-
gineering, image processing, and pattern recognition. These domains also are characterized by a high
complexity, uncertainty regarding underlying dependencies and laws as well as measurements. In
addition, expert knowledge often plays an important role in the successful solution of the problems.
From the application point of view, two questions are of major importance:
1. Which type of tasks in an automatic control system can be reasonably solved by a fuzzy system/
controller?
On the basic level, the use of fuzzy systems for tasks such as direct control or signal processing may
be attractive, if only imprecise or qualitative measurements are available, the process is nonlinear,
human control has to be imitated, or the controller functionality should be enhanced.
If the process is equipped with a basic automation system, but it is time-variant, or operated in differ-
ent regimes, then fuzzy systems may be appropriate for tasks, such as parameter adaptation, quality
control, and supervisory control.
On higher levels, where tasks of scheduling, supervision or multi-objective control have to be per-
formed by operators processing quantitative and qualitative information, fuzzy systems may be an
essential part of an automation solution.
According to the main task to be accomplished by the fuzzy system, the following main classes of
applications can be distinguished:

• Control,

• diagnosis, classification, and pattern recognition,

• modeling and forecasting,

• decision support.

2. What are the specific prerequisites and advantages of the application of fuzzy systems/ controller?
The prerequisite for the successful implementation of any automatic control concept is knowledge
about the process and in control engineering. Furthermore, appropriate control hardware (sensors,
actuators, programmable logical controllers - PLC, process control systems - PCS, etc.) and soft-
ware are necessary. This statement holds for fuzzy systems as well. Regarding the knowledge, the
specificity lies in the form of knowledge representation (linguistic variables and rules). The present
status regarding hardware and software, as described in Section 4.5, allows an easy implementation
and integration of fuzzy control systems. Nevertheless, basic knowledge of fuzzy system theory and
the specific design and analysis methods is a necessary prerequisite.

25
The advantages of fuzzy system applications are manifold. First, fuzzy systems help to automate
processes which could not be automated before or to improve the performance of existing automa-
tion systems. This often leads to improved quality, reduction of energy or material consumption,
inventory or maintenance costs. Moreover, this may improve the operability of processes for human
operators. Secondly, by using dedicated software tools and/or function blocks of control hardware,
fuzzy systems can be implemented faster than comparable control algorithms implemented in C or
other computer languages. Thirdly, fuzzy systems are a transparent means for documenting con-
trol strategies, provided that single rule bases remain manageable. This makes them easy to modify
and debug at a later time. Furthermore, fuzzy systems may provoke users to iteratively optimize the
automation system by deriving process knowledge and operation experience.
To summarize, fuzzy systems have become a tool of the control engineer, which is readily available
through theoretical and technological developments. Fuzzy systems are not a panacea for all control
problems, but successful applications demonstrate the usefulness of fuzzy technology.

4.5 Hardware and Software for Fuzzy Control Systems

Different CAD (Computer Aided Design) tools for fuzzy systems are on the market, which support the
design of fuzzy systems. They are available for almost all types of hardware and software platforms.
In the following, only some typical features shall be mentioned.
Usually, there is a graphic user interface for the design of membership functions, the editing of rules
(as a list of rules or as a table), the choice of fuzzy operations, and defuzzification methods. The
analysis of fuzzy systems often is restricted to the plot of control surfaces or to protocol the reactions
of the fuzzy system to given inputs. Some CAD tools contain more or less basic simulation facilities
or an interface to other simulation tools. A few tools also support the design of fuzzy systems by
searching for rules or by means of a parameter optimization for membership functions.
Most fuzzy CAD tools support the export of the fuzzy systems designed to other hardware and soft-
ware platforms. The latter step is of great practical relevance to the acceptance of fuzzy systems.
Different strategies are possible:

• The fuzzy system will be automatically translated into the source code of programming lan-
guages (e. g. languages of the international standard IEC 61131-3 or specialized Fuzzy Control
Language of the IEC 61131-7 for process control systems, or ANSI C for personal computers
or workstations). Here, the procedures of fuzzification, inference, and defuzzification are step-
wisely implemented. This strategy allows for an easy integration into existing systems and it
does not require any additional hardware effort. The running time in routine use is not very fast.
This strategy is most important for practical solutions.
• The fuzzy system is transformed into a look-up table with characteristic points and (optionally)
interpolation between them. Here, an equivalent functional description of the control surface is
implemented instead of the procedures of fuzzification, inference, and defuzzification. As dis-
cussed before, the integration in existing systems is possible without any additional hardware
effort. The run time is very fast. The solution might require a lot of memory to store character-
istic points for multiple-input fuzzy systems. It is mainly used for smaller fuzzy systems and
for low-cost hardware solutions, e. g. microcontrollers.
• The fuzzy system will be exported as assembler code for a special structured Fuzzy ASIC (Ap-
plication Specific Integrated Circuit). Such ASICs are extremely fast, but restrictions regarding
the structure of the fuzzy systems may occur. This solution requires additional hardware effort
and is not commonly used.

26
4.6 Further Concepts Based on Fuzzy Sets Applied to Control Systems

The idea of fuzzy sets has promoted many fields of scientific research. New concepts and methods
have been developed, often extending existing ones, by allowing for gradual truth values or gradual
set membership, by employing fuzzy numbers or fuzzy relations, or by describing relationships by
IF-THEN rules.
As the kind of task performed by an automatic control system is manifold, more fuzzy set-based
concepts are being applied nowadays than fuzzy controls. As mentioned in the previous section,
fuzzy systems are used for classification and pattern recognition, modeling, and decision support.
In the following sections, two examples shall be presented, namely, fuzzy Petri nets, a concept for
modeling manufacturing and control systems, and fuzzy decision making, a concept useful in higher-
level control. The application of fuzzy systems to system identification is the subject of System
Identification using Fuzzy Models.
Fuzzy Petri nets result from combining Petri nets with fuzzy sets. Petri nets are used to model and
analyze systems which are characterized by parallelism, concurrency and sequencing of processes,
e. g. complex manufacturing systems, communication systems or control systems of such systems (see
Discrete Event Systems). Petri nets consist of places representing conditions or situations, transitions
representing events or activities and arcs connecting places and transitions. Places can contain tokens.
Their flow through the net is controlled by the switching of the transitions. A marking, the distribution
of tokens, represents the system state. Models of such complex systems are uncertain to some degree
due to simplifications or abstractions. In particular, the difficulty of making precise distinctions leads
to vagueness, which can be dealt with by fuzzy sets. Furthermore, impreciseness occurs in situations,
where the choice between alternatives is left unspecified. This impreciseness is described within the
framework of possibility distributions, and thus, is strongly related to fuzzy sets, too. Hence, it seems
natural to introduce fuzziness in the Petri net model.
Several concepts of fuzzy Petri nets have been proposed. An example is the fuzzy Petri net model on
the basis of Place/Transition (P/T) nets, which can be considered a combination of several P/T nets
valid for specific process situations. This type of fuzzy Petri net can be used to model the control of
complex manufacturing systems. Another example is the fuzzy Petri net model in which places are
interpreted as fuzzy propositions, transitions as fuzzy rules. This type can be applied as a model of a
fuzzy reasoning process, e. g. of fuzzy expert systems.
Fuzzy decision making refers to (multiple criteria) decision-making, where some of the constituents
of the decision problem are not known or not described precisely. A decision problem is characterized
by a set of alternatives, a set of constraints on the choice between alternatives, and a performance or
objective function (or a set thereof) associating the gain (or loss) to each alternative. The alternatives
can be described by discrete or continuous variables. In many real world problems the value of the
objective function for a certain alternative is not known precisely, e. g. this function is given by a set of
expert rules. Moreover, the constraints are not necessarily “hard” (ordinary sets), but more reasonably
modeled with fuzzy sets.
One way to solve such fuzzy decision problems is to exploit their symmetry with regard to objectives
and constraints. The basic idea can be described as follows. Both, objectives and constraints, can
be represented as fuzzy sets over the set of alternatives and, thus, be treated identically when finding
a decision. The fuzzy set to represent an objective function measures the degree with which the
alternatives fulfill this objective, i. e. it is the fuzzy set of “good solutions” for this objective. Assuming
multiple objectives, a “good compromise solution” is the alternative which is “as good as possible”
for all objectives. Thus, an intersection of the fuzzy sets of the objectives gives the appropriate fuzzy
set measuring the joint fulfillment of all objectives for each alternative. The same reasoning applies
to the constraints resulting in a fuzzy set measuring the degree to which the constraints are jointly met

27
by each of the alternatives. Combining both fuzzy sets by intersection and searching for the maximum
gives the optimal decision.
Fuzzy decision making has been successfully applied to various control problems, e. g. scheduling in
manufacturing systems, or selection of control strategies in traffic control systems, power plants, etc.
The same principle of extending existing methods by fuzzification is used in many fields, e. g. fuzzy
classification, fuzzy clustering, fuzzy optimization, or fuzzy expert systems.

5 Conclusions

Fuzzy control has emerged as a new paradigm in automatic control. It aims at controlling of systems
of high complexity, for which only vague or fuzzy knowledge on the structure and the parameters is
available. A sound theoretical background has been developed since the first paper was written by
L. A. Zadeh in 1965. The theoretical achievements cover fuzzy logic-related issues, namely, different
methods and algorithms for fuzzification, inference, and defuzzification, as well as control issues, in
particular design and analysis methods. In addition, several tools have been developed to support the
design and implementation of fuzzy control systems. Meanwhile, many successful applications of
fuzzy systems as part of the automatic control system of technical and non-technical systems have
been reported.
Nevertheless, there are some open problems, e. g.

• the evaluation of the efficiency of fuzzy control systems by appropriate criterions,

• the proof of global stability and expected performance of these systems,

• the acceptance of fuzzy control theory by control engineers in practice,

• the automatic extraction of expert knowledge for different processes.

These problems will be the subject of research activities in the next years.
Glossary
Accumulation: The third step of inference, where the results of rule activation, i. e. the resulting
fuzzy sets, are combined in the output fuzzy set.
Activation: The second step of inference, where the conclusions of the rule are activated individually
according to the degree of premise fulfillment resulting in a fuzzy set per rule.
Aggregation: The first step of inference, where the degree of premise fulfillment is calculated.
Control error: The difference between the values of the reference signal and the controlled variable
(symbol e, e(t) = r(t) − y(t)).
Control surface: Graphic representation of the multivariate nonlinear relationship described by a
fuzzy controller.
Controlled variable: The output signal of a controlled plant, which is to assume certain desired
values (symbol y).
Defuzzification: The conversion of a fuzzy set into a numerical quantity.
Fuzzification: The conversion of a numerical quantity into degrees of membership of linguistic terms
of a linguistic variable.
Fuzzy associative map: The (multidimensional) representation of a rule base showing the (weighted)
associations of a combination of linguistic terms of the inputs with a linguistic term of the out-
put (fuzzy rules).
Fuzzy control: A control concept on the basis of fuzzy systems theory.

28
Fuzzy IF-THEN rule: A rule with a premise (condition) and a conclusion (consequence), where at
least the premise must be a fuzzy proposition.
Fuzzy logic: In a narrow sense, a generalization of the Boolean logic, i. e. of the two-valued logic
to a multi-valued logic. In a broader sense, the theory of fuzzy sets in all its facets (logical,
set-theoretical, relational and epistemic).
Fuzzy model: A model on the basis of fuzzy systems theory.
Fuzzy Petri net: A modification of classical Petri nets to describe discrete-event processes, where
marks are described in terms of memberships ranging from zero to unity, and/or transitions are
described in terms of fuzzy rules.
Fuzzy proposition: A sentence ’v is A’ where v is a variable taking values from some universal set
V , and A is a linguistic term labeling a fuzzy set on V .
Fuzzy relation: A fuzzy set on a Cartesian product space, which generalizes the notion of the clas-
sical relation.
Fuzzy set: A set of ordered pairs (x, µ(x)), where x is an element of the universal set X and µ(x) is a
membership function, whose values lie in the interval [0,1].
Fuzzy system: A system using linguistic rules or fuzzy relations to map numerical input quantities
to numerical output quantities, the steps of this process being fuzzification, inference, and de-
fuzzification.
Inference: The procedure which determines the membership degrees of the linguistic terms of the
output variables from the membership degrees of the linguistic terms of the input variables
using the given set of rules.
Mamdani-type fuzzy system: A fuzzy system with IF-THEN rules, in which conditions and con-
clusions contain fuzzy propositions.
Manipulated variable (control signal): The output signal of a controller, which is applied to a con-
trolled plant in order to make it respond in a certain desired way (symbol u).
Linguistic term: A label denoting vague or imprecise concepts, such as ’small’, ’large’, etc. The
meaning of a linguistic term is defined by a fuzzy set.
Linguistic variable: A variable whose values are described by linguistic terms.
Membership function: The characteristic function of a fuzzy set A, µA (x) : X → [0, 1], indicating the
degree of membership of a (numerical) element x in the fuzzy set A.
Plant: The object (system) that is to be controlled.
Reference signal (Setpoint): The signal, expressed in the same unit as the controlled variable, that
sets the desired values of the controlled variable (symbol r).
Rule base: A set of rules describing the knowledge available for reaching specific objectives.
Singleton fuzzy system: A fuzzy system with IF-THEN rules, in which only the conditions contain
fuzzy propositions, whereas the conclusions contain constants.
Takagi-Sugeno-type fuzzy system: A fuzzy system with IF-THEN rules, in which only the condi-
tions contain fuzzy propositions, whereas the conclusions contain expressions which can be a
function (e. g. of the variables of the premise) or a conventional control law.

Bibliography
Babuška R. (1998). Fuzzy Modeling for Control, Boston: Kluwer. 260 pages. [This monograph is an
excellent survey of fuzzy modeling and the application of fuzzy models in control].
Cardoso J. and Camargo H. (1998). Fuzziness in Petri Nets, Heidelberg: Physica-Verlag. 318 pages.
[This book provides a comprehensive account of developments concerning the incorporation of fuzzy
capabilities in Petri net models.]
Driankov D., Hellendoorn H., and Reinfrank M. (1996). An introduction to fuzzy control, Berlin:
Springer, 2nd edition. 316 pages. [This book covers the theoretical fundamentals of fuzzy control as

29
 well as design methods].
Farinwata S. S., Filev D. P., and Langari R. (Eds.) (2000). Fuzzy Control: Synthesis and Analysis,
Chichester: Wiley. 272 pages. [This collection includes papers on fuzzy modeling and control on the
basis of Takagi-Sugeno-type fuzzy systems].
Fuller R. and Carlsson C. (2001). Fuzzy Reasoning in Decision Making and Optimization, Heidel-
berg: Physica-Verlag. 352 pages. [This book describes the fuzzy concepts in decision making with
applications to strategic planning, project management, and supply chain management.]
Jamshidi M., Vadiee N. and Rose T. J. (Eds.) (1998) Fuzzy Logic and Control: Software and Hard-
ware Applications, Pearson Education POD. 416 pages. [This collection presents software and hard-
ware tools for fuzzy control by means of examples.]
Kiendl H., Knicker R., and Niewels F. (1996). Two-way fuzzy controllers based on hyperinference
and inference filter, In Jamshidi M., Fathi M., and Pierrot F., (Eds.), Proc. World Automation Congr.
(WAC’96), May 28-30 1996, Montpellier, France, pages 381–388, Albuquerque, N. M.: TSI Press.
[This paper describes hyperinference and the inference filter].
Mamdani E. H. (1974). Application of fuzzy algorithms for control of a simple dynamic plant, Proc.
IEEE, 121:1585–1588. [This is the seminal paper on control application of fuzzy theory].
Nguyen H. T. and Sugeno M. (Eds.) (1998). Fuzzy Systems: Modeling and Control, Boston: Kluwer
Academic. 519 pages. [This collection is a comprehensive survey of the field, focusing methodolog-
ical aspects].
Passino K. M. and Yurkovich S. (1998). Fuzzy control, Menlo Park, Calif.: Addison-Wesley. 475
pages. [This is a tutorial introduction to fuzzy control containing case studies for automotive systems,
robotics, aircraft, etc.].
Takagi T. and Sugeno M. (1985). Fuzzy identification of systems and its applications to modeling and
control, IEEE Trans. on Systems, Man, and Cybernetics, 15(1):116–132. [This paper introduces the
Takagi-Sugeno-type fuzzy system].
Tanaka K. and Wang H. O. (2001). Fuzzy Control Systems Design and Analysis: A Linear Matrix
Inequality Approach, New York: Wiley. 320 pages. [This book covers the systematic framework for
the stability and design of nonlinear fuzzy control systems based on the Takagi-Sugeno-type fuzzy
system].
Wang L.-X. (1997). A course in fuzzy systems and control, Upper Saddle River, N. J.: Prentice
Hall. 424 pages. [This book provides a comprehensive, self-tutorial course in fuzzy logic and fuzzy
control].
Yager R. R., Ovchinnikov S., Tong R. M. and Nguyen H. T. (Eds.) (1987). Fuzzy sets and applica-
tions: selected papers by L. A. Zadeh, New York: Wiley. 684 pages. [This is a collection of the most
important papers of L. A. Zadeh covering many fields of fuzzy theory].
Zadeh L. A. (1965). Fuzzy sets, Information and Control, 8:338–353. [This is the seminal paper on
fuzzy sets].

30
View publication stats