Iwona Skalna Bogdan Rębiasz

Bartłomiej Gaweł Beata Basiura

•

Jerzy Duda Janusz Opiła

•

Tomasz Pełech-Pilichowski

Advances in
Fuzzy Decision Making
Theory and Practice

123
Iwona Skalna Jerzy Duda
AGH University of Science and Technology AGH University of Science and Technology
Kraków Kraków
Poland Poland

Bogdan Rębiasz Janusz Opiła

AGH University of Science and Technology AGH University of Science and Technology
Kraków Kraków
Poland Poland

Bartłomiej Gaweł Tomasz Pełech-Pilichowski

AGH University of Science and Technology AGH University of Science and Technology
Kraków Kraków
Poland Poland

Beata Basiura
AGH University of Science and Technology
Kraków
Poland

ISSN 1434-9922 ISSN 1860-0808 (electronic)

Studies in Fuzziness and Soft Computing
ISBN 978-3-319-26492-9 ISBN 978-3-319-26494-3 (eBook)
DOI 10.1007/978-3-319-26494-3

Library of Congress Control Number: 2015954593

© Springer International Publishing Switzerland 2015

Springer International Publishing AG Switzerland is part of Springer Science+Business Media

(www.springer.com)
 Preface

When dealing with real-world problems, one can rarely avoid uncertainty. Very
often in practice, it is not possible to obtain precise information about the values
of the parameters of a modelled system. For a long time, probability approaches
were dominant in the literature devoted to problems involving uncertainty.
However, the most common situation in practice is when some model parameters
are subject to aleatory uncertainty, whereas others are subject to epistemic
uncertainty. This situation is especially common in economic calculus, which stems
from the fact that economic data usually comes from various sources, but also
concerns areas such as human decision making, risk analysis or engineering
applications.
Aleatory uncertainty, also called variability or statistical uncertainty, is a
description of naturally random behaviour in a physical process or property.
Probability theory is a natural model for this type of uncertainty, and the quan-
tification for the aleatory uncertainty is usually performed using the Monte Carlo
techniques. On the other hand, epistemic uncertainty refers to limited knowledge
about the system (modelled or real) or lack of information. This type of uncertainty
can be reduced by, e.g. taking more measurements, conducting more tests, “buying”
more information, etc. Because of that, the epistemic uncertainty is also called
reducible uncertainty, incertitude or subjective uncertainty. Often, uncertainty
quantification intends to work toward reducing epistemic uncertainties to aleatory
uncertainties. However, epistemic uncertainty is not well characterised by proba-
bilistic approaches. To evaluate epistemic uncertainties, methods such as interval
analysis, fuzzy logic or evidence theory (Dempster–Shafer theory) are more
suitable.
Measurement errors are one of the possible sources of epistemic uncertainty. It is
well known that at the empirical level, uncertainty is an inseparable companion of
almost any measurement. The difference between the measured and the actual
values is called a measurement error. Since the absolute value of the measurement
error can usually be bounded, it is therefore guaranteed that the actual (unknown)
value of the desired quantity belongs to the interval (number) with the midpoint
being the measured value, and the radius being the upper bound for the (absolute
value) of the possible errors. Therefore imprecision, approximation, or uncertainty
in the knowledge of the exact values of physical, technical or economic parameters
can be modelled conveniently by intervals. By its nature, interval arithmetic yields
rigorous enclosures for the range of operations and functions. The results are
intervals in which the exact results must lie. In order that this inclusion remains true
in numerical computations, the problem of round-off errors must be taken into
account during the implementation of the interval computations. Proper handling of
outward rounding in numerical computations forces the result to be the interval
approximation of the correct real interval (that could be hypothetically obtained
assuming that infinite precision is available).
A finite number of increasingly precise measurements give a finite family of
usually increasingly narrower intervals (interval numbers). By assigning a respec-
tive level of possibility to each interval number from this family, a discrete
collection of the so-called α–cuts is obtained, which can be viewed as a finite
representation of a fuzzy number. The strict relation between interval and fuzzy
numbers is often emphasised in the literature devoted to fuzzy theory. Fuzzy
numbers enhance the expressive power of intervals, and therefore they are often
referred to as generalised intervals.
This book presents some advances in fuzzy decision making. It is organised in
eight chapters. Chapter 1 introduces some basic concepts from the fuzzy numbers
theory. The main part of this chapter concerns the problem of performing arithmetic
operations on fuzzy numbers. Linear operations such as addition and subtraction are
rather obvious, whereas nonlinear operations, such as multiplication or division,
pose a problem. Nonlinear operations usually result in fuzzy numbers of different
types than operands. Therefore, various, the so-called “shape preserving”,
approaches to multiplication and division of fuzzy numbers are proposed in the
literature. On the other hand, many researchers recommend to use the α–cuts based
approach to performing operations on fuzzy numbers, because this approach allows
fuzzy and interval techniques to be combine and used to effectively solve problems
involving both types of uncertainty. The remaining part of this chapter is devoted to
the problem of performing arithmetic operations on interactive fuzzy numbers. The
reason is that all of the above-mentioned approaches implicitly assume that there is
no dependency between fuzzy numbers involved in a computation. In practice this
assumption is rarely satisfied, especially when dealing with economic problems. To
cope with the dependency problem, stochastic simulation of fuzzy systems and
nonlinear programming approaches are proposed.
The problem of comparing and ordering fuzzy numbers is described in Chap. 2.
Theoretically, fuzzy numbers can only be partially ordered, and hence cannot be
compared. However, in practical applications, such as decision making, scheduling,
market analysis or optimisation with fuzzy uncertainties, the comparison of fuzzy
numbers becomes crucial. That is why several methods for comparing and ordering
fuzzy numbers have been proposed in the literature. They can be generally divided
into two groups. The first group consists of methods which enable two fuzzy
numbers to be compared. One can mention the probabilistic approach, centroid
point approach or radius of gyration approach. To order a set of fuzzy numbers
using these methods, some dedicated procedures are required. The second group
consists of methods, which assign to a fuzzy number a real value. These are, for
example Yager ranking index based approach, defuzzification approach or weighted
average. The methods from the second group can be directly used to order a set of
fuzzy numbers, by employing one of the several methods for ordering (sorting) real
numbers. All the described methods are compared using a simple example.
Chapter 3 presents the concept of a fuzzy random variable and the Dempster–
Shafer theory of evidence. Fuzzy random variable extends the classical definition of
a random variable and is one of the possible ways to jointly consider randomness
and imprecision. The simultaneous occurrence of randomness and imprecision is
often the case in real-world decision problems, because data in such problems
usually comes from various sources, such as historical datasets or experts opinions.
The theory of evidence (also called the theory of belief functions), on the other
hand, provides mathematical tools to process information, which is, at the same
time, of random and imprecise nature. It allows imprecision and variability to be
treated separately within a single framework. The evidence theory encompasses
both possibility and probability theories.
Chapter 4 discusses selected issues of the fuzzy multi-criteria decision making
(FMCDM). Generally, multi-attribute decision making (MADM) is concerned with
ranking alternatives with respect to multiple criteria. Two basic techniques of
multi-criteria decision making are analytical hierarchy process (AHP) and technique
for order of preference by similarity to ideal solution (TOPSIS). Both AHP and
TOPSIS were initially designed to deal only with crisp numbers. Later, fuzzy
variants of those methods were developed, because in the real world available data
are often imprecise and vague. The integrated fuzzy approach to solve
multi-attribute decision problems is proposed in this chapter. Its use is illustrated
using a real case from a steel industry.
Chapter 5 is devoted to a method which is able to process hybrid data, i.e. to
jointly handle both randomness and imprecision. Random variables are described
by probability distributions and imprecise values are modelled using possibility
distributions. The main advantage of the proposed method is that it takes into
account the dependencies between economic parameters. The correlation matrix is
used to model dependencies between stochastic parameters, whereas interval
regression is used to model both dependencies between fuzzy parameters and
between fuzzy and stochastic parameters. The proposed method combines tools
such as Monte Carlo simulation, interval regression and nonlinear programming. As
the result of hybrid data processing, a random fuzzy set is received. Assessment of
risk is obtained by computing the standard deviation and also by estimating the
upper and lower cumulative distribution functions of the analysed variable. The
method is verified through computing the operating profit for a metallurgical
industry enterprise.
Chapter 6 describes the application of fuzzy sets to planning and scheduling of
production in steel industry. Primarily the problem of steel grade assignment to
customers’ orders is analysed, which is the first stage of steel production planning.
Fuzzy sets are used to reduce the variety of potential steel grades and to describe the
characteristic of materials by decision makers. Next, the use of fuzzy logic systems
for steel production scheduling is examined. Parameters like timeliness, amount,
priority and the sequence length on casters are expressed using linguistic variables.
Whereas fuzzy rules are used to determine an initial schedule and to perform a
quick rescheduling. More advanced systems use a multi-agent approach. Each agent
may use its own fuzzy logic in order to satisfy the constraints related to the certain
level of steel production. Finally, an example of a fuzzy scheduling agent is pro-
vided. It uses a genetic algorithm to generate feasible and economically efficient
schedules for a continuous caster.
Chapter 7 presents a new method for forecasting the level and structure of
market demand for industrial goods. The method employs two data mining meth-
ods: k-means clustering and fuzzy decision trees. The k-means method serves to
separate groups with items of a similar consumption level and structure of the
analysed products (consumption patterns). Whereas fuzzy decision trees are used to
determine the dependencies between consumption patterns and predictors (param-
eters determining the level and structure of consumption). The proposed method is
verified using the extensive statistical material on the level and structure of steel
products consumption in selected countries during 1960–2010.
Chapter 8 discusses various techniques of visualisation of fuzzy numbers in one-,
two- and more-dimensional spaces. As canonical box-and-whiskers representation
is not suitable for visualisation of uncertainty in three-dimensional spaces, an
approach based on ScPovPlot3D templates for POVRay, which is a powerful
photorealistic renderer equipped with domain-specific programming language,
dubbed scene description language (SDL), is proposed. In order to show the use-
fulness of the proposed technique some examples of visualisation of fuzzy objects
are included, in one, two and three dimensions. For three-dimensional approach two
examples of application of fuzzy visualisation are presented. The first one shows a
surface defined by a function which assigns a fuzzy number to a point on the real
plane. In the second example, a lesion deposited in a segment of cardiac vessel is
depicted using “thick surface” approach.
 Contents

1 Fuzzy Numbers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1 Preliminary Theory. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.1.1 LR–type Fuzzy Numbers . . . . . . . . . . . . . . . . . . . . . . . 5
1.1.2 T-Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.2 Arithmetic of Fuzzy Numbers . . . . . . . . . . . . . . . . . . . . . . . . 9
1.2.1 Zadeh’s Extension Principle . . . . . . . . . . . . . . . . . . . . 9
1.2.2 α–cuts Based Operations. . . . . . . . . . . . . . . . . . . . . . . 10
1.2.3 Shape-Preserving Operations . . . . . . . . . . . . . . . . . . . . 12
1.2.4 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
1.3 Modelling Dependencies Between Fuzzy Numbers . . . . . . . . . . 17
1.3.1 Modelling Stochastic Dependencies . . . . . . . . . . . . . . . 18
1.3.2 Interval Regression. . . . . . . . . . . . . . . . . . . . . . . . . . . 20
1.4 Possibility and Credibility Theory. . . . . . . . . . . . . . . . . . . . . . 22
1.4.1 Non-linear Programming Approach . . . . . . . . . . . . . . . 23
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
2 Ordering of Fuzzy Numbers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
2.1 Probabilistic Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
2.2 Defuzzification Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
2.3 Centroid-Point Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
2.4 Yager Ranking Index Approach . . . . . . . . . . . . . . . . . . . . . . . 34
2.5 Degree of Possibility Approach . . . . . . . . . . . . . . . . . . . . . . . 34
2.6 Weighted Averaging Approach Based on α–cuts . . . . . . . . . . . 35
2.7 Two-Dimensional Radius of Gyration Approach. . . . . . . . . . . . 37
2.8 Fuzzy Maximising-Minimising Points Approach. . . . . . . . . . . . 40
2.9 Area Based Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
2.10 Left and Right Dominance Approach . . . . . . . . . . . . . . . . . . . 43
2.11 An α-weighted Valuations Approach . . . . . . . . . . . . . . . . . . . 44
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
3 Fuzzy Random Variable and the Dempster-Shafer Theory
of Evidence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
3.1 Fuzzy Random Variable . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
3.2 Dempster–Shafer (D–S) Theory of Evidence . . . . . . . . . . . . . . 50
3.3 The Hybrid Data Propagation Method. . . . . . . . . . . . . . . . . . . 52
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
4 Multi-attribute Decision Making Process and Its Application . . . . . 55
4.1 General Steps of a Multi-attribute Decision Making Process . . . 57
4.2 Uncertainty in MADM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
4.3 Fuzzy Analytical Hierarchical Process . . . . . . . . . . . . . . . . . . . 60
4.3.1 Example I—Comparison of AHP Methods . . . . . . . . . . 62
4.3.2 Fuzzy TOPSIS. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
4.4 Example—Investment Strategy of the Steel Company . . . . . . . . 65
4.4.1 Calculation of Fuzzy Ratings. . . . . . . . . . . . . . . . . . . . 65
4.4.2 Computation of Performance Index of Criteria . . . . . . . . 68
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
5 Risk Assessment in the Presence of Uncertainty. . . ............ 73
5.1 Description of Uncertainty in the Economic
Risk Assessment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
5.2 Measures of Risks in the Case of Hybrid Data . . . . . . . . . . . . . 78
5.3 Computing an Operation Profit . . . . . . . . . . . . . . . . . . . . . . . 79
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
6 Application of Fuzzy Theory in Steel Production Planning
and Scheduling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
6.1 Fuzzy Grade Assignment. . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
6.2 Scheduling of Production Processes Using Fuzzy Logic . . . . . . 99
6.3 Cooperation of Planning with Fuzzy Logic . . . . . . . . . . . . . . . 107
6.4 Continuous Caster Scheduling Based on Fuzzy Lot-Sizing. . . . . 111
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
7 Application of Fuzzy Decision Trees in Analog Forecasting ...... 119
7.1 Methods for Forecasting Apparent Consumption
of Steel Products . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
7.2 Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
7.2.1 The k-means Method . . . . . . . . . . . . . . . . . . . . . . . . . 122
7.2.2 Fuzzy Decision Trees . . . . . . . . . . . . . . . . . . . . . . . . . 123
7.3 The Proposed Forecasting Method . . . . . . . . . . . . . . . . . . . . . 126
7.4 Empirical Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128
7.5 The Evaluation of Obtained Results . . . . . . . . . . . . . . . . . . . . 132
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
8 Selected Issues of Visualisation of Fuzziness in Cardiac
Imaging Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135
8.1 Visualisation of Fuzzy Numbers . . . . . . . . . . . . . . . . . . . . . . . 135
8.1.1 Visualisation of a Fuzzy Number . . . . . . . . . . . . . . . . . 135
8.1.2 Visualisation of a Two Dimensional Fuzzy Point . . . . . . 137
8.1.3 Visualization of a Three-Dimensional Fuzzy Points . . . . 138
8.2 Scattered Fuzzy 3D Chart . . . . . . . . . . . . . . . . . . . . . . . . . . . 139
8.2.1 Fuzzy Surface Plot . . . . . . . . . . . . . . . . . . . . . . . . . . . 141
8.2.2 Imaging of Fuzziness of Cardiac Vessels Walls . . . . . . . 144
8.2.3 DICOM Format Versus STL . . . . . . . . . . . . . . . . . . . . 145
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150
Chapter 1
Fuzzy Numbers

Abstract This chapter introduces some basic concepts from the fuzzy numbers
theory. The main part of the chapter concerns the problem of performing arithmetic
operations on independent fuzzy numbers. The remaining part is devoted to the
problem of performing arithmetic operations on interactive fuzzy numbers. To cope
with the dependency problem, stochastic simulation of fuzzy systems and non-linear
programming approaches are proposed.

Fuzzy set theory was introduced in the 60s by Lofti Zadeh [1]. It was initially intended
to be an extension of a dual logic and/or classical set theory. However, for the last
decades, it has been developed as a powerful “fuzzy” mathematics [1–4] since almost
all mathematical objects can be described by sets (e.g., a function can be described
by an ordered set of points). Fuzzy sets provide a mathematical framework for the
precise and rigorous study of vague conceptual phenomena. Applications of fuzzy set
theory can be found, for example, in computer science, artificial intelligence, control
engineering, communication, medicine, decision theory, expert systems, logic, man-
agement science, pattern recognition, operations research, and robotics. The fuzzy
concepts of fuzzy set theory take into account the fact that all phenomena in the
physical universe have a degree of inherent uncertainty.
Real fuzzy numbers, which are of the main concern in this book, are a special
case of convex, normalised fuzzy sets of the real line . Fuzzy numbers play an
important role in human thinking and provide a natural way of dealing with real
world problems. This is because the way that people perceive the world is continually
changing and cannot always be defined in true or false statements. For example, it
is more natural to say “she is about eighteen” or “he is rather tall” rather than “she
is eighteen year old” or “he is 185 cm tall”. Fuzzy numbers are actively used in
various fields, such as artificial intelligence, computer science, medicine, control
engineering, decision theory, expert systems, logic, management science, operations
research, pattern recognition, robotics etc. They can also be considered as a modelling
language, well suited for situations involving fuzzy relations, criteria and phenomena.
2 1 Fuzzy Numbers

1.1 Preliminary Theory

In classical set theory each item either belongs to a set or does not belong to that set.
Whereas a fuzzy set is as a class of objects with continuum of grades of membership
(characteristic) function (see, e.g., [1, 3, 5]). The following example pictures the
difference between a classical and a fuzzy set.

Example 1.1 Consider the set of “tall men”. In classical set theory the statement
“people taller than or equal to 1.82 m are tall” can be represented graphically by
a step function as depicted in Fig. 1.1. This crisp membership function does not
seem to work well, since it makes no distinction between somebody who is 1.86 m
and someone who is 2.16 m, they are both simply tall. Moreover, the difference
between a 1.80 and 1.83 man is only 3 cm, however the membership function just
says one is tall and the other is not tall. The fuzzy set approach to the set of tall men
provides a much better representation of the tallness of a person. The set, shown in
Fig. 1.2, is defined by a continuously increasing function. So, a person with the grade
of membership equal to 0.95 is a really tall person, whereas a person with the grade
of membership equal to 0.3 is really not very tall at all.

Definition 1.1 A fuzzy set à on the domain X is a pair (X, μ à ), where μ à : X →

[0, 1] is called the grade of membership of x in (X, μ Ã ).

Fig. 1.1 Crisp membership function for the set of tall men

µ
µ

Fig. 1.2 Fuzzy membership function for the set of tall men
1.1 Preliminary Theory 3

A fuzzy set à can also be expressed as a set of ordered pairs. The following two
notations are used the most often:
 
à = (x, μ à (x)) | x ∈ X, μ à : X → [0, 1] , (1.1)
 
à = (x, μ à (x) | x ∈ X, μ à : X → [0, 1] . (1.2)

A fuzzy set (X, μ Ã ) with a finite domain is usually denoted by {μ Ã (x1 )/x1 , . . . ,
μ (x ) μ (x )
μ Ã (xn )/xn } or Ãx1 1 , . . . , Ãxn n .
A fuzzy number is a special case of a convex, normalised fuzzy set of the real line.
Complex fuzzy numbers can also be encountered in the literature, however, they are
out of the scope of this book.
Definition 1.2 A real fuzzy number à is a fuzzy subset of the real numbers 
characterised by a membership function μ Ã :  → [0, 1] that assigns to each x ∈ 
a grade of membership μ Ã (x). It is required that the membership function fulfils the
following four conditions:
1. Normality: there exists x0 ∈ , such that μ Ã (x0 ) = 1.
2. Convexity: ∀x, y ∈ , ∀λ ∈ 

μ Ã (λx + (1 − λ)y)  min{μ Ã (x), μ Ã (y)}. (1.3)

3. Upper semi-continuity (u.s.c., see Fig. 1.3a): for all x0 ∈  and all ε > 0 there
exist a neighbourhood V (x0 ), such that for all x ∈ V (x0 )

μ Ã (x0 )  μ Ã (x) + ε. (1.4)

4. Compactness: the support of a fuzzy number

supp( Ã) = cl{x ∈  | μ Ã (x) > 0} (1.5)

is bounded; cl(S) denotes the closure of a set S, i.e., the set S together with all
of its limit points.
The set of all real fuzzy numbers will be denoted throughout by F().
Definition 1.3 The core of a fuzzy number à (see Fig. 1.3a) is defined as a set of all
points for which the value of the membership function equals 1, i.e.,

core( Ã) = {x ∈  | μ Ã (x) = 1}. (1.6)

Definition 1.4 Let à ∈ F() and let α ∈ (0, 1]. Then, an α-cut of à is defined as
the following closed interval:

Ãα = {x ∈  | μ Ã (x)  α}. (1.7)

4 1 Fuzzy Numbers

The 0-cut is defined separately (otherwise it would give the entire real line) and
equals to the support of a fuzzy number, Ã0 = supp( Ã). The A1 –cut is simply the
core of a fuzzy number Ã. A strong (or strict) α-cut is an open interval defined by
Ãα = {x ∈  | μ Ã (x) > α}.
The well-known properties (consistency conditions) of α–cuts are the following:
for α, β ∈ [0, 1]
• Ãα ⊆ Ã β
 for α  β (monotonicity),
α
• Ã = Ãβ (continuity).
β<α

A fuzzy number à can be represented by an infinite family of nested α–cuts

(see Fig. 1.3b) 
à = (α, Ãα ). (1.8)
α∈[0,1]

This representation is referred to as parametric representation of a fuzzy number and

is extensively used in various methods for solving problems involving fuzzy numbers.
In practical computations it is recommended to select a finite subset of α–cuts with
relevant degrees of membership; they must also be semantically distinguishable. The
problem of selecting the best approximation of a fuzzy number by a finite family of
its α–cuts was considered, e.g., by Pedrycz [6].

Theorem 1.1 (Representation Theorem) Let à ∈ F(). Then,

 
μ Ã (x) = sup min(α, μ Ãα (x)) , (1.9)
α∈[0,1]

where 
1, x ∈ Ãα ,
μ Ãα (x) = (1.10)
0, otherwise.

The concept of α–cuts plays an essential role in the fuzzy numbers theory. The
α–cuts representation is often used to process fuzzy data by means of interval
(level-by-level) techniques. This allows to combine fuzzy and interval techniques
and to effectively solve problems involving both types of uncertainty.

(a) (b)

Fig. 1.3 General form of a fuzzy quantity with upper semi-continuous membership function (a);
nested α–cuts corresponding to the selected levels: α–level, β–level and γ–level (b)
1.1 Preliminary Theory 5

1.1.1 L R–type Fuzzy Numbers

LR-type fuzzy numbers are an important subclass of fuzzy numbers [7]. The solution
of fuzzy problems is somewhat easier if only L R–type fuzzy numbers are involved.

Definition 1.5 An L R-type fuzzy number à is described by the following member-

ship function [7]:
⎧ 
⎪
⎪ L Ã a−x
α
, x ∈ [a − α, a],
⎪
⎨ 1,
μ Ã (x) =   x ∈ [a, b], a  b, (1.11)
⎪
⎪ R Ã x−b
β
, x ∈ [b, b + β],
⎪
⎩
0, otherwise.

where a and b are, respectively, the lower and upper modal values, [a, b] is the
core of Ã, α > 0 is the left spread, β > 0 is the right spread, and L Ã , R Ã are the
so-called shape functions, L Ã , R Ã : [0, 1] → [0, 1], with L Ã (0) = R Ã (0) = 1 and
L Ã (1) = R Ã (1) = 0.

The L Ã and R Ã functions are non-increasing, continuous mappings [7]. The sup-
port of an LR-type fuzzy number supp( Ã) = [a − α, b + β]. An LR-type fuzzy num-
ber is usually denoted by à = (a, b, α, β) L R . However, if L à (x) = R à (x) = 1 − x,
then a simpler notation, Ã = (a, b, α, β), is usually used. The set of all L R-type
fuzzy numbers will be denoted throughout by F() L R .
Two types of LR-type fuzzy numbers are the most commonly used in practice.
These are triangular (TFN) (Fig. 1.4) and trapezoidal (TRFN) (Fig. 1.5) fuzzy num-
bers with both shape functions linear.

Definition 1.6 A triangular fuzzy number is defined by the following triangle-

shaped membership function:
⎧
⎪
⎪ 0, x  a,
⎨ x−a , a < x  b,
μ Ã (x) = b−a (1.12)
⎪ c−x
⎪ , b  x < c,
⎩ c−b
0, x  c.

Fig. 1.4 A triangular fuzzy number à = (1, 4, 5) with ker( Ã) = {4} and supp( Ã) = [1, 5]
6 1 Fuzzy Numbers

Fig. 1.5 A trapezoidal fuzzy number à = (2, 3, 4, 6) with ker( Ã) = [3, 4] and supp( Ã) = [2, 6]

The above definition differs from the general definition of LR-type fuzzy numbers.
This stems from that it is adopted to the notation of triangular fuzzy number used
in this book, which is the following: Ã = (a, b, c), where a = inf(supp( Ã)), c =
sup(supp( Ã)) and b is a modal (central) value. A similar notation will be used for
trapezoidal fuzzy numbers.
Trapezoidal fuzzy numbers are often called fuzzy intervals. They form the most
general class of fuzzy numbers with linear shape functions.

Definition 1.7 A trapezoidal fuzzy numbers is defined by the following trapeze-

shaped membership function:
⎧
⎪
⎪ 0, x  a,
⎪
⎪ x−a
,
a < x  b,
⎨ b−a
μ Ã (x) = 1, b < x < c, (1.13)
⎪
⎪
⎪
⎪
d−x
, c  x < d,
⎩ d−c
0, x ≥ d.

A trapezoidal fuzzy number will be denoted as à = (a, b, c, d), where a =

inf(supp( Ã)), d = sup(supp( Ã)), and b and c are, respectively, lower and upper
modal values. Note that trapezoidal fuzzy numbers are a generalisation of triangular
fuzzy numbers.
The α–cuts of a trapezoidal fuzzy number can be presented in the form of intervals
with endpoints being linear functions of α. Given a trapezoidal fuzzy number à =
(a, b, c, d), its α–cuts are given by:

Ãα = [(b − a)α + a, d − (d − c)α]. (1.14)

Note that this formula, with obvious modifications, applies to triangular fuzzy
numbers.

1.1.2 T-Operators

The triangular norms (t-norms), triangular conorms (t-conorms) an important role

in fuzzy set theory. There is a large amount of literature devoted to t-norm based
1.1 Preliminary Theory 7

operations on fuzzy numbers. The t-norms and t-conorms generalise, respectively,

the conjunctive (“AND”) and disjunctive (“OR”) operators, and hence can be used
to define the intersection and union operations in fuzzy logic and fuzzy inference
systems. This possibility was first noted by Höhle [8] and was later exploited by
Alsina et al. [9], Klement [10] and Dubois and Prade [11]. The properties of t-norms
for possible use in the development of intelligent systems were investigated, e.g., by
Bonissone [12]. T-operators has also been widely used in the design of fuzzy logic
controllers, in the modelling of other decision-making processes [13, 14] and in the
scheduling of production. The current notion of t-norms is due to Schweizer and
Sklar [15].

Definition 1.8 A function T : [0, 1] × [0, 1] → [0, 1] is a t-norm if and only if for
any x, y, z ∈ [0, 1]
(i) T (x, 1) = x (existence of a unit 1),
(ii) x  y ⇒ T (x, z)  T (y, z) (monotonicity),
(iii) T (x, y) = T (y, x) (commutativity),
(iv) T (x, T (y, z)) = T (T (x, y), z) (associativity).

A t-norm is Archimedean, iff:

(i) T (x, y) is continuous,
(ii) T (x, x) < x ∀x ∈ (0, 1).
An Archimedean t-norm is strict, iff

T (x , y ) < T (x, y), if x < x, y < y, ∀ x , y , x, y ∈ (0, 1).

The most prominent examples of t-norms are the following:

• Minimum t-norm (also called min-norm, Gödel t-norm or standard t-norm):

Tmin (x, y) = min{x, y}.

• Standard product or probabilistic t-norm (the ordinary product of real numbers)

T pr od (x, y) = x ∗ y,

• Lukasiewicz t-norm (also called bounded difference)

TLuk (x, y) = max{0, x + y − 1}.

• Drastic t-norm (also called weak t-norm)

⎧
⎨ y, if x = 1,
Td (x, y) = x, if y = 1,
⎩
0, otherwise.
8 1 Fuzzy Numbers

This is the only one t-norm which is not continuous.

• Hamacher t-norm
λx y
TH (x, y) = .
1 − (1 − λ)(x + y − x y)

The drastic t-norm is the weakest t-norm, and the minimum t-norm is the strongest
one, i.e., Td (x, y)  T (x, y)  Tmin (x, y), for any t-norm T and all x, y ∈ [0, 1].

Definition 1.9 A function ⊥ : [0, 1] × [0, 1] → [0, 1] is a t-conorm if and only if

for any x, y, z ∈ [0, 1]
(i) ⊥(x, 0) = x (existence of a zero 0),
(ii) x  y ⇒ ⊥(x, z)  ⊥(y, z) (monotonicity),
(iii) ⊥(x, y) = ⊥(y, x) (commutativity),
(iv) ⊥(x, ⊥(y, z)) = ⊥(⊥(x, y), z) (associativity).
A t-conorm is Archimedean, iff:
(i) ⊥(x, y) is continuous,
(ii) ⊥(x, x) > x ∀x ∈ (0, 1).
An Archimedean t-conorm is strict, iff

⊥(x , y ) < ⊥(x, y), if x < x, y < y, ∀ x , y , x, y ∈ (0, 1).

The most popular t-conorms are listed below.

• Maximum t-conorm (also called min-norm, Gödel t-norm or standard t-norm):

⊥max (x, y) = max{x, y},

• Probabilistic t-conorm

⊥ pr od (x, y) = x + y − x ∗ y,

• Lukasiewicz t-conorm

⊥ Luk (x, y) = min{1, x + y},

• Drastic t-conorm (also called a weak t-conorm)

⎧
⎨ y, if x = 0,
⊥d (x, y) = x, if y = 0,
⎩
0, otherwise.

This is the only one t-conorm which is not continuous.

1.1 Preliminary Theory 9

• Hamacher t-norm
λ(x + y) + x y(1 − 2λ)
⊥ H (x, y) = .
λ + x y(1 − λ)

Dually to t-norms, all t-conorms are bounded by the maximum and the drastic t-
conorm, i.e., ⊥max (x, y) ≤ ⊥(x, y) ≤ ⊥d (x, y), for any t-conorm ⊥ and all x, y ∈
[0, 1].

1.2 Arithmetic of Fuzzy Numbers

The fuzzy arithmetic allows mathematical operators such as addition, subtraction,

multiplication, and division, to be applied to the fuzzy domain, i.e., it is able to propa-
gate the fuzzy uncertainty through computations. Different approaches to performing
these operations, yielding sometimes different results, can be found in, e.g., Dubois
and Prade [16], Filev and Yager [17], Kreinovich and Pedrycz [18], Fodor and Bede
[19] and Tsao [20]. The comparison of four methods for multiplication of fuzzy
numbers was presented, e.g., by Zhang et al. [21] and Dutta et al. [22]. They recom-
mended the α–cut based approach for executing fuzzy arithmetic. The α–cut based
approach employs interval arithmetic which additionally emphasizes the relation
between fuzzy and interval numbers. Bansal [23] presented the basic mathematical
operations formulated on trapezoidal fuzzy numbers. Roy [24] described arithmetic
operations for general trapezoidal fuzzy numbers. Extension of algebraic mathemat-
ics for positive trapezoidal fuzzy numbers was described, e.g., by Vahidi and Rezvani
[25]. Kolesárová [26] defined the t-norm based multiplication of LR-type fuzzy num-
bers with positive supports. Dubois and Prade [27] proposed an approximate formula
for multiplications of fuzzy number with positive supports. Nevertheless, regardless
which approach is used, the algebraic system of fuzzy arithmetic is only an abelian
monoid under both addition and multiplication, additive and multiplicative inverses
exist only for fuzzy numbers with support of zero width. In general case, the dif-
ference between two same fuzzy numbers results in a fuzzy number with support
symmetrical around the modal value which is equal to zero (such fuzzy number is
sometimes considered as fuzzy zero). Several other algebraic laws, valid for real
numbers, are not preserved in fuzzy arithmetic.

1.2.1 Zadeh’s Extension Principle

The most general definition of operations on fuzzy numbers is based on the so-called
extension principle. It was first introduced by Zadeh [1], that is why it is usually called
Zadeh’s extension principle [1]. Later on, several modifications were suggested (see,
e.g., [28]).
10 1 Fuzzy Numbers

The most general form of extension principle is called the t-norm (triangular
norm) based extension principle [7].
Definition 1.10 (Extension principle) Let the mapping f : X × Y → Z , where
X, Y, Z ⊆ , be given . It can be extended to fuzzy numbers in the following
way:
μ f ( Ã, B̃) (z) = sup T (μ Ã (x), μ B̃ (y)), (1.15)
z= f (x,y)

where à ∈ F(X ), B̃ ∈ F(Y ) and T is a t-norm.

For a binary arithmetic operation, Zadeh’s extension principle can be defined as
follows.
Definition 1.11 Given two fuzzy numbers Ã, B̃ and an arithmetic operation ◦ ∈
{+, −, ∗, /} (in the case of division, it is assumed that 0 ∈
/ supp( B̃)), the result of an
operation C̃ = Ã ◦ B̃ is defined by the following membership function:

μC̃ (z) = sup T (μ Ã (x), μ B̃ (y)). (1.16)

z=x◦y

The most frequently used t-norm is the strongest min-norm, however, extension
principle with other norms can also be encountered in the literature. The min-norm
based Zadeh’s extension principle takes the form:

μ Ã◦ B̃ (z) = sup min(μ Ã (x), μ B̃ (y)). (1.17)

z=x◦y

Example 1.2 Consider two trapezoidal fuzzy numbers à = (−2, −1, 3, 8) and B̃ =
(−4, 1, 5, 6), and take the corresponding discrete fuzzy sets

à = {0/−2, 1/−1, 1/0, 1/1, 1/2, 1/3, 0.8/4, 0.6/5, 0.4/6, 0.2/7, 0/8},
B̃ = {0/−4, 0.2/−3, 0.4/−2, 0.6/−1, 0.8/0, 1/1, 1/2, 1/3, 1/4, 1/5, 0/6}.

The goal is to find the membership μ Ã + B̃ . Table 1.1 shows all possible results of
addition of the elements from the domains of à and B̃ . The corresponding operands
and their membership grades are given on the top and on the left side of the table. The
final result obtained using Zadeh’s extension principle (1.17) is depicted in Fig. 1.6.

1.2.2 α–cuts Based Operations

Zadeh’s extension principle, in spite of all its usefulness, has been regarded as a time-
consuming and expensive computational tool since its usage requires at each sample
the inverse function in addition to the effect induced by the discretization process.
That is why, the α–cuts based approach is often used instead.
1.2 Arithmetic of Fuzzy Numbers 11

Table 1.1 The sums of the elements from the domains of à and B̃
0 0.2 0.4 0.6 0.8 1 1 1 1 1 0
−4 −3 −2 −1 0 1 2 3 4 5 6
0 −2 −6 −5 −4 −3 −2 −1 0 1 2 3 4
1 −1 −5 −4 −3 −2 −1 0 1 2 3 4 5
1 0 −4 −3 −2 −1 0 1 2 3 4 5 6
1 1 −3 −2 −1 0 1 2 3 4 5 6 7
1 2 −2 −1 0 1 2 3 4 5 6 7 8
1 3 −1 0 1 2 3 4 5 6 7 8 9
0.8 4 0 1 2 3 4 5 6 7 8 9 10
0.6 5 1 2 3 4 5 6 7 8 9 10 11
0.4 6 2 3 4 5 6 7 8 9 10 11 12
0.2 7 3 4 5 6 7 8 9 10 11 12 13
0 8 4 5 6 7 8 9 10 11 12 13 14

Fig. 1.6 Discrete fuzzy

numbers à , B̃ and their
sum obtained using Zadeh’
extension principle

Definition 1.12 Given two fuzzy numbers Ã, B̃ and an arithmetic operation ◦ ∈
{+, −, ∗, /}, the result is defined by the following α–cuts:

( Ã ◦ B̃)α = Ãα ◦ B̃ α = {x ◦ y|x ∈ Ãα , y ∈ B̃ α }, α ∈ [0, 1]. (1.18)

 α
The result of an arithmetic operation ( Ã ◦ B̃) = α∈[0,1] α( Ã ◦ B̃) is always a fuzzy
number.
Using the α–cuts representation (1.14), an arithmetic operation on two trapezoidal
fuzzy numbers à = (a1 , a2 , a3 , a4 ) and B̃ = (b1 , b2 , b3 , b4 ) can be performed for
each α ∈ [0, 1] using the following formula:

( Ã ◦ B̃)α = [(a2 − a1 )α + a1 , a4 − (a4 − a3 )α] ◦ [(b2 − b1 )α + b1 , b4 − (b4 − b3 )α].

(1.19)

In order to calculate the resulting α-cut, ( Ã ◦ B̃)α , interval arithmetic [29] must be
employed.
12 1 Fuzzy Numbers

Linear α–cuts based operations on fuzzy numbers, such as addition and subtrac-
tion, are straightforward. For two trapezoidal fuzzy numbers à = (a1 , a2 , a3 , a4 ) and
B̃ = (b1 , b2 , b3 , b4 ), they are defined as follows:

à + B̃ = (a1 + b1 , a2 + b2 , a3 + b3 , a4 + b4 ), (1.20)

à − B̃ = (a1 − b4 , a2 − b3 , a3 − b2 , a4 − b1 ). (1.21)

However, in the case of nonlinear operations, such as multiplication and division,

the resulting fuzzy number in general case will be of a different type than operands.
For convenience and also to simplify the computation, the resulting fuzzy number
can be approximated by a fuzzy number of the respective type (see Example 1.3).
For triangular and trapezoidal fuzzy numbers the choice of the most appropriate
approximation is obvious. Otherwise, this choice could be non-trivial.
In the case when both operands are non-negative (i.e., have non-negative supports),
the α–cuts based multiplication takes the form:

α
( Ã ∗ B̃)α = [( Ã ∗ B̃)α , ( Ã ∗ B̃) ], (1.22)

where
α
( Ã ∗ B̃) = ((a2 − a1 )α + a1 ) ∗ ((b2 − b1 )α + b1 ),
α
( Ã ∗ B̃) = (a4 − (a4 − a3 )α) ∗ (b4 − (b4 − b3 )α).
Then, the membership function μ Ã∗ B̃ (z) can be easily calculated form (1.22), by
equating both endpoints of the above interval to x.
The α–cuts based division of fuzzy numbers, providing that 0 ∈ / supp( B̃), is
defined as follows:
( Ã/ B̃)α = Ãα ∗ (1/ B̃ α ), (1.23)

where 1/ B̃ is the reciprocal of a fuzzy number. In the case when B̃ non-negative, the
reciprocal takes the form:

(1/ B̃ α ) = [1/(b4 − (b4 − b3 )α), 1/((b2 − b1 )α + b1 )]. (1.24)

1.2.3 Shape-Preserving Operations

In the case of trapezoidal fuzzy numbers with positive supports, the trapezoidal result
of multiplication and division can be obtained using the following formulae:

( Ã ∗ B̃) = (a1 ∗ b1 , a2 ∗ b2 , a3 ∗ b3 , a4 ∗ b4 ), (1.25)

( Ã/ B̃) = (a1 /b4 , a2 /b3 , a3 /b2 , a4 /b1 ). (1.26)

1.2 Arithmetic of Fuzzy Numbers 13

It can be shown that the support and the core of the resulting fuzzy number coin-
cide with the support and the core of the fuzzy number obtained using the α–cuts
based approach. The operations (1.25) and (1.26) are called the shape preserving
operations. Another approaches to shape-preserving operations on fuzzy numbers
can be found, e.g., in [30]. Given two triangular fuzzy numbers à = (a1 , a2 , a3 ) and
B̃ = (b1 , b2 , b3 ) with positive central values (a2 > 0, b2 > 0), the shape-preserving
multiplication of triangular fuzzy numbers has the following form [30]:

Ã∗Td B̃ = (min(a1 b2 , a2 b1 ), a2 b2 , max(a3 b2 , a2 b3 )). (1.27)

The drastic t-norm is proved to be the norm which preserves the shape of fuzzy
numbers.
For triangular fuzzy numbers à = (a1 , a2 , a3 ) and B̃ = (b1 , b2 , b3 ) with positive
supports (a1 > 0, b1 > 0), the same result can be obtained using the formula proposed
by Kolesárová [26]:

à ∗Td B̃ = (a2 b2 (1 − d1 ), a2 b2 , a2 b2 (1 + d2 )), (1.28)

where
 
d1 = max((a2 − a1 )a2 , (b2 − b1 )b2 ),
d2 = max((a3 − a2 ) a2 , (b3 − b2 ) b2 ).
Bansal [23] presented formulae for selected arithmetic operations as well as basic
functions. For two arbitrary trapezoidal fuzzy numbers à = (a1 , a2 , a3 , a4 ) and B̃ =
(b1 , b2 , b3 , b4 ), the so-called extended multiplication rule has the form:

à ∗ B̃ = (d1 , d2 , d3 , d4 ), (1.29)

where
d1 = min{a1 b1 , a1 b4 , a4 b1 , a4 b4 },
d2 = min{a2 b2 , a3 b2 , a3 b3 , a2 b3 },
d3 = max{a2 b2 , a3 b2 , a3 b3 , a2 b3 },
d4 = max{a1 b1 , a1 b4 , a4 b1 , a4 b4 }.
Assuming that 0 ∈
/ [b1, b4], the division of two trapezoidal fuzzy numbers takes the
form [23, 31]:
Ã/ B̃ = (d1 , d2 , d3 , d4 ), (1.30)

where
d1 = min{a1 /b1 , a1 /b4 , a4 /b1 , a4 /b4 },
d2 = min{a2 /b2 , a2 /b3 , a3 /b2 , a3 /b3 },
d3 = max{a2 /b2 , a2 /b3 , a3 /b2 , a3 /b3 },
d4 = max{a1 /b1 , a1 /b4 , a4 /b1 , a4 /b4 }.
14 1 Fuzzy Numbers

Dubois and Prade [27] derived the following approximate formula for the Tmin -
multiplication of LR-type fuzzy numbers à = (a1 , a2 , a3 ) and B̃ = (b1 , b2 , b3 ) with
positive supports:

Ã∗TM B̃ ≈ (a2 b1 + a1 b2 − a2 b2 , a2 b2 , a2 b3 + a3 b2 − a2 b2 ). (1.31)

The above formula holds providing that a2a−a 2

1
+ b2b−b
2
1
 1. The approximation is
necessary to preserve the shape of the L , R functions, since unlike addition and
subtraction multiplication based on the strongest t-norm does is not preserving oper-
ation.

1.2.4 Examples

Example 1.3 Let the following two triangular fuzzy numbers be given
à = (1, 2, 4), B̃ = (1, 4, 6) (see Fig. 1.7a) and for clarity of presentation put C̃ =
à ∗ B̃.

The results obtained using formulae (1.22), (1.25), (1.27)–(1.29) and (1.31) are
presented in Fig. 1.7b. The solid black line represents the product obtained using the
α–cuts based approach (1.22). The dashed grey line represents the product obtained
using the formula (1.31), the dashed black line denotes the product obtained using the
formulae (1.25) and (1.29), whereas the grey solid line denotes the product obtained
using formulae (1.27) and (1.28). It can be seen that the obtained results have the
same modal value, but they differ in the width of the supports.
The results obtained by using the formulae (1.22), (1.25) and (1.29) have the same
support supp(C̃) = [1, 24], which is the widest among all the supports. However, the
shape functions of the fuzzy number produced by the formula (1.22) are no longer
linear, they are nonlinear functions of x:
 √
−2+ 3x+1
, 1  x < 8,
μC̃ (x) = √
3
(1.32)
5− x+1
2
, 8  x  24.

(a) (b)

Fig. 1.7 Two exemplary triangular fuzzy numbers à = (1, 2, 4) and B̃ = (1, 4, 6) (a); the results
of multiplication of à and B̃ using different multiplication formulae (b)
1.2 Arithmetic of Fuzzy Numbers 15

In order to remain within the class of triangular fuzzy numbers, the result of the for-
mula (1.22) can be approximated by the triangular fuzzy number (1, 8, 24) (obtained
by taking the three characteristic points: lower bound of the support, modal value
and upper bound of the support), which coincides with the result of the formulae
(1.25) and (1.29).
The formulae (1.27) and (1.28) give the triangular fuzzy number C̃ = (2, 8, 16)
with the narrower support supp(C̃) = [2, 18] being about 30 % narrower than the
result of formulae (1.22), (1.25) and (1.29).
Finally, the approximate formula (1.31) results in the fuzzy number C̃ = (−2,
8, 20) with partially negative support supp(C̃) = [−2, 20], which contradicts with
the rather obvious expectation that the multiplication of two positive quantities should
yield a positive quantity. This support is about 27 % wider than the result of the
formula (1.27) and about 4 % narrower than the results of the formulae (1.22), (1.25)
and (1.29). Crisp values obtained using the centroid defuzzification approach (Centre
of Area, COA) approach, which is described in Sect. 1.2, are presented in Table 1.2.
Summarising, the obtained results differ significantly. This concerns both fuzzy
and crisp results. Which formula to use in a specific problem remains an open ques-
tion. However, many authors are inclined to use the α-cut based operations on fuzzy
numbers, for the reasons described at the beginning of this chapter.
As already mentioned, the reciprocal of an arbitrary fuzzy number can be obtained
using the α–cut based approach, providing the support of this number does not contain
zero. When dealing with trapezoidal fuzzy number à = (a1 , a2 , a3 , a4 ) with positive
(or negative) support, the following formula can be used for computing the reciprocal:

1/ Ã = (1/a4 , 1/a3 , 1/a2 , 1/a1 ). (1.33)

It can be shown that the support and the core of the resulting trapezoidal fuzzy number
coincide with the support and the core of the fuzzy number obtained using the α–cuts
based approach.
In the case of triangular fuzzy numbers, the same result can be obtained using the
following formula for division of triangular fuzzy numbers:

(1, 1, 1)/ Ã = (α, β, γ), (1.34)

where
α = min(a1 /b1 , a1 /b3 , a3 /b1 , a3 /b3 ),
β = a2 /b2 ,
γ = max(a1 /b1 , a1 /b3 , a3 /b1 , a3 /b3 ).

Table 1.2 Crisp values Formula Value

obtained using the Centre of
Area (COA) defuzzification (1.22) 18.61
method (1.25), (1.29) 19.00
(1.27), (1.28) 16.67
(1.31) 16.67
16 1 Fuzzy Numbers

Example 1.4 Let the fuzzy number à = (3, 5, 10) be given (black solid line in
Fig. 1.8a). The reciprocal of Ã, obtained from the formula (1.34), is the fuzzy number
1/ Ã = (0.1, 0.2, 1/3), whereas the α–cuts based approach gives the fuzzy number
with the following membership function:

2 − 0.2/x, 0.1  x  0.2,
μ1 Ã (x) =
0.5/x − 1.5, 0.2  x  1/3,

which is a rational (non-linear) function of x. Both reciprocals are presented in

Fig. 1.8b. The dotted black line represents the α–cuts based reciprocal, and the dashed
black line denotes the reciprocal obtained from (1.34). The results are very similar in
that their supports and modal values coincide. The question is whether the approxi-
mation by a triangular fuzzy number does not imply a loss of information.

The division of fuzzy numbers was already defined using α–cuts based approach
(1.22) and using the inversion of characteristic points (1.34). In the example below,
the division of fuzzy numbers à and B̃ is performed using different methods of
multiplication of à and the reciprocal of B̃ (Fig. 1.8).
Example 1.5 Let the following two triangular fuzzy numbers be given à = (3, 5, 10),
B̃ = (1, 2, 8) (see Fig. 1.9a). The division is performed here by first computing the
reciprocal of B̃, and then by multiplying à and 1/ B̃ using different multiplication
formulae. The results are presented in Fig. 1.9b. The dashed grey line represents
the quotient obtained using the combination of formulae (1.31) and (1.33). The
dashed black line denotes the quotient obtained using the combination of formula
(1.33) and, respectively, formulae (1.25) and (1.29) (the quotient obtained using the
formula (1.26) coincides with this result). The grey solid line denotes the quotient
obtained using the combination of formulae (1.33) and, respectively, formulae (1.27)
and (1.29). The solid black line represents the quotient obtained using the α–cuts
based approach. The membership function, in this case, is a rational function and
has the following form:
 8x−3
2+6x
, 0.375  x < 2.5,
μC̃ (x) = (1.35)
10−x
5+x
, 2.5  x  10.

(a) (b)

Fig. 1.8 The fuzzy number à = (3, 5, 10) (a); reciprocal obtained using the inversion of charac-
teristic points (dashed black line) and using. α–cuts (dotted black line) (b)
1.2 Arithmetic of Fuzzy Numbers 17

(a) (b)

Fig. 1.9 Two exemplary triangular fuzzy numbers à = (3, 5, 10) and B̃ = (1, 2, 8) (a); the results
of computing Ã/ B̃ = Ã ∗ (1/ B̃) using different multiplication formulae (b)

Table 1.3 Crisp values obtained using the Centre of Area defuzzification method
Formula Crisp number
(1.22) 6.35
(1.25), (1.29) 6.79
(1.27), (1.28) 5.21
(1.31) 5.71

As it was expected, the resulting fuzzy numbers differ significantly. The corre-
sponding crisp numbers are summarised in Table 1.3. The difference between the
smallest and the largest crisp value is greater than 1. An attempt to verify whether
the choice of the respective approach to fuzzy arithmetic influences (or not) the final
result will be made in Chap. 5.

1.3 Modelling Dependencies Between Fuzzy Numbers

Arithmetic operations on fuzzy numbers, presented in the previous section, implicitly

assume that all combinations of implementation of fuzzy numbers involved are pos-
sible. In practical applications this assumption is rarely satisfied. For example, large
implementations of crude oil prices usually involve large implementations of petrol
prices. Combinations of small implementations of prices of one product and large
prices of the other one are not likely to appear in reality. Ignoring this dependency
may lead to significant overestimation of the exact result.
The dependency between variables (model parameters) can take different forms.
When the dependency between variables X 1 , X 2 , . . . , X n exists under the form of
a domain D ⊆ n , which acts as the actual range of (X 1 , X 2 , . . . , X n ), and if, more-
over, X 1 , X 2 , . . . , X n have respective fuzzy ranges Ṽ1 , Ṽ2 , . . . , Ṽn , then Zadeh’s
extension principle (1.15) must be adapted as follows [36]: for each z ∈ ,


μ f ( Ṽ ,Ṽ ,...Ṽ  (z) = sup {min(μṼ1 (x1 ), . . . , μṼn (xn ))}. (1.36)
1 2 n D)
z = f (x1 , . . . , xn )
(x1 , . . . , xn ) ∈ D
18 1 Fuzzy Numbers

The fuzzy set f (Ṽ1 , Ṽ2 , . . . , Ṽn | D̃) can be defined in terms of α–cuts of the fuzzy
numbers Ṽ1 , Ṽ2 , . . . , Ṽn [36] as follows: for each α ∈ [0, 1],
 
f (Ṽ1 , . . . , Ṽn | D̃) = { f (x1 , . . . , xn )| (x1 , . . . , xn ) ∈ (Ṽ1 )α × · · · × (Ṽn )α ∩ ( D̃)α }.
α
(1.37)

Calculating a fuzzy quantity f (Ṽ1 , . . . , Ṽn | D̃) using the Eq. (1.37) is difficult and
very often involves nonlinear programming methods. Therefore, various methods
for handling selected cases of interactivity between fuzzy variables have been devel-
oped. Dubois and Prade [27] solved the problem of arithmetic operations on fuzzy
variables when the set D̃ is specified by a linear equation. Enea and Piazza [32] dis-
cussed the problem of interactive fuzzy numbers in the context of Constrained Fuzzy
AHP (CFAHP). Klir [33] presented the results of arithmetic operations on trapezoidal
fuzzy numbers under requisite constraints. For example, he discusses the results of
arithmetic operations on two fuzzy variables under the constraint that one variable is
smaller than another and compares the results with the ones obtained without impos-
ing this constraint. Kuchta [31] introduced the so-called constrained subtraction,
which for à = (a1 , a2 , a3 , a4 ) and B̃ = (b1 , b2 , b3 , b4 ) with ai − bi  ai+1 − bi+1 ,
for i = 1, 2, 3, is defined as follows:

à − B̃ = (a1 − b1 , a2 − b2 , a3 − b3 , a4 − b4 ) . (1.38)

The constrained subtraction refers to the situations when large implementations of

one variable are always accompanied by large implementations of another variable
and small implementations of one variable by small implementations of another
variable. However, when analysing economical facts, the link between the variables
are not always that clear-cut.

1.3.1 Modelling Stochastic Dependencies

If the dependence between fuzzy variables is expressed in the form of the a correla-
tion, then operations on those variables can be performed using the fuzzy simulation.
The simulation of fuzzy systems was introduced by Liu and Iwamura [34]. It consists
in executing random tests on a model with fuzzy parameters.
Given an univalued function f and a vector of correlated variables X = (X 1 ,
X , . . . , X M ) ∈  M , for which values are limited by the respective elements of a fuzzy
vector Ṽ = (Ṽ1 , Ṽ2 , . . . , ṼM ), a computer fuzzy simulation presented in Algorithm 1
[35] can be used to specify the possibility distribution of f ( Ṽ ) as well as its expected
value E( f ( Ṽ )) and a standard semi-deviation S Dev( f ( Ṽ )).
The elements of the vector X = (x1 , x2 , . . . , x M ) are correlated random numbers
from the α–cuts of the respective fuzzy numbers Ṽ1 , Ṽ2 , . . . , ṼM . They are obtained
in the following way. First, random values z 1 , z 2 , . . . , z M are drawn from a uniform
1.3 Modelling Dependencies Between Fuzzy Numbers 19

Algorithm 1 Computing the value of a function for interactive fuzzy numbers

Input: Function f , vector of correlated fuzzy variables for which values
are limited by the respective elements of the vector of fuzzy numbers
Ṽ = (Ṽ1 , Ṽ2 , . . . , ṼM )
Output: Possibility distribution of f ( Ṽ )
...
1: Define α0 , ℘, n̂, n = (1/℘ + 1)n̂
2: α = α0
3: Define α-levels of the fuzzy numbers
4: n = 1
5: Generate a vector of correlated random values
(x1 , x2 , . . . , x M ) ∈ Ṽ1α × Ṽ2α × · · · × ṼM
α

6: f n = f (x1 , x2 , . . . , xm ), μ( f n ) = min {μṼi (xi )}

i=1,...,M
7: n = n + 1
8: If n  n̂ Then Go To Step 5
9: α = α + ℘
10: If α < 1 Then Go To Step 3
11: Define possibility distribution f ( Ṽ )
11.1: α = 1, k = 1
11.2: Lsk = min...{ f n | μ ( f n )  α}, μ(L sk ) = μ( f n )|L sk = f n
1 n  n
Rsk = max...{ f n | μ ( f n )  α}, μ(Rsk ) = μ( f n )|Rsk = f n
1 n  n
11.3: α = α − ℘, k = k + 1
11.4: Repeat While α > 0
Lsk = min...{ f n |μ ( f n )  α ∧ f n  Ls1 }, μ(L sk ) = μ( f n )|L sk = f n
1 n  n
Rsk = max...{ f n |μ ( f n )  α ∧ f n  Rs1 }, μ(Rsk ) = μ( f n )|Rsk = f n
1 n  n
α = α − ℘, k = k + 1
Loop
12: Define credibility distribution function for f ( Ṽ ) on the basis of
L sk , Rsk , μ(L sk ) and μ(Rsk )

distribution on the interval [0, 1). Put Z = (z 1 , z 2 , . . . , z M ) and let Ω be a correlation

matrix describing dependencies between the considered fuzzy variables. Then, Z =
L Z , where L is obtained from the Cholesky decomposition of the correlation matrix,
i.e., Ω = L L T . In order to obtain the vector X, the elements of the vector Z are
mapped onto α–cuts of the respective fuzzy numbers. Once the vector X is obtained,
the grade of membership of z = f (X) is calculated as follows:

μ f (Ṽ ) (z) = min {μṼi (xi )}. (1.39)

i=1,2,...,M

The method presented in Algorithm 1 allows to take into account the interactivity
between the fuzzy parameters while computing the possibility distribution of f ( Ṽ ).
The values Lsk , Rsk and μ(Lsk ), μ(Rsk ) specify the possibility distribution of the
function f .
20 1 Fuzzy Numbers

1.3.2 Interval Regression

An interval regression is identified when parameters of a regression equation are

expressed in the form of bounded intervals [36–38]. Numerous practical problems
[39] can be solved using interval and fuzzy regressions.
Until now, several methods were developed to estimate the parameters of the
interval regression equation. The best known method uses a linear programming for
this purpose [37, 40, 41]. However, this approach has many deficiencies. The major
ones include [36–38]:

• It is often the case that some of the estimated regression parameters tend to be
crisp; it even happens that the method produces only a few unexpectedly wide
interval parameters, while others are crisp (the drawback called unbalancedness).
This problem is generally considered to be the most serious and the most restrictive
drawback that limits the usefulness of this method.
• The method might produce interval regression parameters with centres that only
poorly fit the data with respect to traditional goodness-of-fit measures (such as
R-squared). In literature, this problem is referred to as non-centrality property.
• The method is highly sensitive to outliers.
Many authors present solutions eliminating these deficiencies. Often, quadratic pro-
gramming methods are combined with least squares methods [37, 40, 41]. Other
methods of interval regression are based on Minkowski distance [42] or multi-criteria
programming [43].
The abovementioned modified methods provide more balanced intervals repre-
senting the coefficients of interval regression equations, however, they require longer
computation time. Moreover, the estimates of weight coefficients are made by experts,
thus these methods become heuristic [39].
Another method for determining parameters of interval regression equations was
developed by Hládik and Černy [39]. It seems to be very promising. Below is
described a variant of this method which covers the case when input data (dependent
and independent variables) are presented as determined figures (crisp input-crisp
output). The proposed method is based on the sensitivity analysis of linear systems.
It consists of two stages:
• the estimation of the centres of the interval parameters by using standard estima-
tors,
• the estimation of the radii of the interval parameters.
Suppose that p observations are given:
⎡ ⎤ ⎡ ⎤
y1 x11 . . . x11
⎢ ⎥ ⎢ .. ⎥ ,
ŷ = ⎣ ... ⎦ , X = ⎣ ... . ⎦ (1.40)
yp x p1 · · · x pn
1.3 Modelling Dependencies Between Fuzzy Numbers 21

where X is an input matrix and ŷ is an output vector. The problem is to determine

the parameters of interval regression:
⎡ ⎤
a1
⎢ ⎥
â = ⎣ ... ⎦ (1.41)
an

that comprise all possibilities determined by the model and data. More formally, â
must be such that [39]:

y j ∈ x j1 a1 + x j2 a2 + · · · + x jn an , ∀ j = 1, . . . , p. (1.42)

It is natural to try to achieve several desirable properties of the interval parameters.

They should be as tight as possible, they should be balanced and they should respect
the central tendency. Easy handling with outliers is also an advantage.
Hladik and Černy [39] present interval regression coefficients in the form of inter-
vals â = [b̂ − ĉ, b̂ + ĉ]. Vector b̂ is estimated by means of the least squares method.
Width of intervals defined by the vector ĉ is expressed in form: ĉ = δ ĉΔ , where ĉΔ
is a non-negative vector of sensitivity coefficients and δ  0 is an unknown value.
The introduction of sensitivity coefficients enable the width of intervals representing
regression coefficients to be controlled.
Solution of the problem of estimation of interval regression parameters is reduced
to finding the minimum value δ  0, such that for the vector â = [b̂ − δ ĉΔ , b̂ + δ ĉΔ ]
the following is met:

∀ j ∈ {1, . . . , p} ∃ â ∈ [b̂ − δ ĉΔ , b̂ + δ ĉΔ ] : y j = X j∗ â . (1.43)

Hladik and Černy [39] present a simple formula for computing δ. Namely, when
there exists j ∈ {1, . . . , n}, such that |X | j∗ ĉΔ = 0 and simultaneously y j = X j∗ b̂,
then there doesn’t exist δ, which would meet the interrelation (1.43). In any other
case, δ may be calculated based on the following relation:
 
 
y j − X j∗ b̂
δ∗ = max . (1.44)
j : |X | j∗ >0 |X | j∗ ĉΔ

The value δ ∗ thus defined is a minimum value of δ. This relation constitutes an effec-
tive method for solving the problem of the interval regression. It eliminates the issue
of significant imbalance of intervals representing particular regression coefficients
and the problem of non-centric location of the estimated intervals.
Sensitivity coefficients are most often assumed as ci Δ = 1 or ci Δ = |bi | for i =
1, . . . , n. The first case is one of the most natural choices since it forces the interval
parameters to have a minimum sum of their radii. The second case is called relative
tolerances. In this case, minimum value of δ (δ ∗ ) gives the following information to
22 1 Fuzzy Numbers

a user: it is sufficient to perturb the regression parameters by no more than 100 %

δ ∗ in order that all observations are satisfied. Hence, it is an alternative measure of
goodness-of-fit.

1.4 Possibility and Credibility Theory

Possibility theory was first introduced by Zadeh [44] as an extension of theory of

fuzzy sets and fuzzy logic. According to Dubois and Prade [45], each fuzzy set
generates two functions: the measure of possibility Pos and the measure of necessity
Nec. They are defined for every classic set X ⊆  as follows.

Definition 1.13 Let X ⊆  and let à ∈ F(). The possibility measure Pos is
a mapping Pos : 2 X → [0, 1] defined by

Pos(X ) = sup{μ Ã (x)}, (1.45)

x∈X

where μ Ã is the membership function of a fuzzy number Ã.

Definition 1.14 Let X ⊆  and let à ∈ F(). The necessity measure Nec is a map-
ping N ec : 2 X → [0, 1] defined by:

N ec(X ) = inf {1 − μ Ã (x)}, (1.46)

x ∈X
/

where μ Ã is the membership function of a fuzzy number Ã.

These definitions are associated with the well-known interpretation of a fuzzy set
given by Zadeh [13]. They assume that a fuzzy set à is a fuzzy restriction of a certain
variable X , which takes values in space , and the only available information about
this variable is that “X is Ô. If X is a variable that takes values in  and à is a fuzzy
number characterised by a membership function μ à , then à is a fuzzy restriction on
X if à acts as an elastic constraint on the values that may be assigned to X . In other
words, the assignment of a value u, u ∈ , to X has the form X = u : μ Ã (u), where
X = u : μ Ã (u) is interpreted as the degree to which the constraint represented by Ã
is satisfied when u is assigned to X [44].
Based on the possibility distribution, it can be determined how possible is the
event that the value of X belongs to a non-fuzzy set A ⊆ , [1]:

π(X ∈ A) = sup{π X (u) : u ∈ A} = sup{μ Ã (u) : u ∈ A} = π(A). (1.47)

Such defined quantity does not have complementary characteristics, i.e., π(X ∈
A) does not have to be equal to 1 − π(X ∈ Ac ), where Ac is the absolute complement
of A. As a remedy, Liu [46] introduced the concept of the credibility measure.
1.4 Possibility and Credibility Theory 23

Definition 1.15 The set function Cr : 2 → [0, 1] is called a credibility measure if

it satisfies the following axioms:
1. Cr () = 1.
2. Cr is increasing, i.e., Cr (A)  Cr (B) whenever A ⊂ B.
3. Cr isself-dual,
 i.e., Cr (A) + Cr (Ac ) = 1 for any A ∈ 2 .
4. Cr i Ai ∧ 0.5 = sup Cr (Ai )for any Ai with Cr (Ai )  0.5.
i

Definition 1.16 A fuzzy variable ξ˜ is defined a function from a credibility space

(, 2 , Cr ) to the set of real numbers.

Definition 1.17 Let ξ˜ be a fuzzy variable and let A be a set of real numbers. The
degree of credibility that the value of ξ˜ belongs to A can be defined as follows:

1
Cr (ξ˜ ∈ A) = (π(ξ˜ ∈ A) + (1 − π(ξ˜ ∈ Ac )). (1.48)
2
Contrary to the possibility distribution, the degree of credibility has complementary
characteristics [46]. Moreover, when the grade of credibility reaches a value of 1,
there is a confidence that the fuzzy event will definitely occur. On the other hand,
when the degree of possibility reaches a value of 1, such a confidence does not exist.
In addition to the credibility measure, Liu [46] has defined the concept of the cred-
ibility distribution Φ(x), which is for a fuzzy variable what is probability distribution
for a random variable.

Definition 1.18 The credibility distribution Φ :  → [0, 1] of a fuzzy variable ξ˜ is

˜
defined by Φ(x) = Cr {u ∈ |ξ(u)  x}.

The value Φ(x) determines the grade of credibility that a fuzzy variable ξ˜ will have
˜
a value equal to or less than x. If μξ̃ is a membership function of fuzzy variable ξ,
then for each x in , the credibility distribution function Φ(x) can be expressed by:

1
Φ(x) = (sup μ (y) + 1 − sup μξ̃ (y)). (1.49)
2 yx ξ̃ y>x

1.4.1 Non-linear Programming Approach

The main problem addressed in this section is to determine possibility distribution of

f ( X̃). The vector X̃ = (X 1 , X , . . . , X m ) is a vector of the projected fuzzy variables
about membership functions μ1 μ2 , . . . , μm . Additionally, itis assumed that there may
be defined subsets X K of dependent variables X i , X K = X̃ i , i ∈ K , K ∈ K s . In
such case, K is the subset of dependent variable indices, and K s is the set of indices
of the selected subsets of dependent variables.
24 1 Fuzzy Numbers

The problem of calculating the fuzzy number characterising f ( X̃) can be solved
by using the concept of α-levels of fuzzy numbers. Here one can take advantage of
the Eq. (1.41). Upper bound (sup) and lower bound (inf) of an α-level of a fuzzy
number f ( X̃) may be determined by solving the following non-linear programming
tasks:
when searching for maximum, find:

f (x1 , x2 , . . . , xm ) → max (1.50)

when searching for minimum, find:

f (x1 , x2 , . . . , xm ) → min (1.51)

subject to the following constraints:

inf(Ṽi )α  xi  sup (Ṽi )α for i = 1, 2 . . . , m, (1.52)

xi  inf(a1i z )x z + inf(a2i z ) for i ∈ K , z ∈ K ; i = z; K ∈ K s , (1.53)

xi  sup(a1i z )x z + sup(a2i z ) for i ∈ K , z ∈ K ; i = z; K ∈ K s , (1.54)

The values a1i z , a2i z are the coefficients of interval regression equations. They deter-
mine a relation between variables X̂ i and X̂ z based on historical data.

References

1. Zadeh, L.A. 1965. Fuzzy sets. Information and Control 8(3): 338–353.
2. Wang, L.-X. 1997. A course in fuzzy systems and control. Prentice Hall.
3. Zadeh, L.A. 1971. Similarity relations and fuzzy orderings. Information Sciences 3(2): 177–
200.
4. Heilpern, S. 1995. Comparison of fuzzy numbers in decision making. Tatra Mountains Math-
ematical Publications 6: 47–53.
5. Zimmermann, H.-J. 1996. Fuzzy Set Theory and its Applications, 3rd ed. MA: Kluwer Academic
Publishers Norwell.
6. Pedrycz, A. 2006. An optimization of ac-cuts of fuzzy sets through particle swarm optimization.
Fuzzy information processing society. NAFIPS 2006. Annual meeting of the North American,
IEEE, 57–62.
7. Carlsson, C., and R. Fullèr. 2005. On additions of interactive fuzzy numbers. Acta Polytechnica
Hungarica 2: 59–73.
8. Höhle, U. 1978. Probabilistic uniformization of fuzzy topologies. Fuzzy Sets and Systems 1.
9. Alsina, C., E. Trillas, and L. Valverde. 1983. On some logical connectives for fuzzy set theory.
Journal of Mathematical Analysis & Applications 93: 15–26.
10. Klement, E.P. 1982. A theory of fuzzy measures: A survey. In Fuzzy information and decision
processes, ed. M. Gupta, and E. Sanchez, 59–66. Amsterdam: North Holland.
11. Dubois, D., and H. Prade. 1981. Triangular norms for fuzzy sets. In International symposium
of fuzzy sets, 39–68, Linz.
References 25

12. Bonissone, P.P. 1985. Selecting uncertainty calculi and granularity: An experiment in trading
off precision and complexity. In Proceedings of the first Workshop on Uncertainty in Artificial
Intelligence, 57–66, Los Angeles.
13. Zadeh, L.A. 1975. In Fuzzy sets and their applications to cognitive and decision processes, ed.
Zadeh, L.A., K.-S. Fu, K. Tanaka, and M. Shimura., Calculus of fuzzy restrictions.
14. Zadeh, L.A. 1973. Outline of a new approach to the analysis of complex systems and decision
processes. IEEE Transactions on Systems, Man, and Cybernetics 3: 28–44.
15. Schweizer, B., and A. Sklar. 1960. Statistical metric spaces. Pacific Journal of Mathematics
10: 313–334.
16. Dubois, D., and H. Prade. 1980. Fuzzy sets and systems: Theory and applications, 4th ed., Tom
144: Mathematics in science and engineering New York: Academic Press.
17. Filev, D., and R. Yager. 1997. Operations on fuzzy numbers via fuzzy reasoning. Fuzzy Sets
and Systems 91(2): 137–142.
18. Kreinovich, V., W. Pedrycz, and H.T. Nguyen. 2001. How to make sure that /appr ox100 in
fuzzy arithmetic: solution and its (inevitable) drawbacks. In Proceedings of the joint 9th world
congress of the international fuzzy systems association and 20th international conference of the
North American fuzzy information processing society IFSA/NAFIPS, 1653–1658, Vancouver,
Canada, July 25–28.
19. Fodor, J., and B. Bede. 2006. Recent advances in fuzzy arithmetics. International Journal of
Computers, Communications and Control 1(5): 199–207.
20. Tsao, C.T. 2002. An interval arithmetic approach for fuzzy multiple criteria group decision
making. Asia Pacific Management Review 7(1): 25–40.
21. Gao, S., Z. Zhang, and C. Cao. 2009. Multiplication operation on fuzzy numbers. Journal of
Software 4(4): 331–338.
22. Dutta, P., H. Boruah, and T. Ali. 2011. Fuzzy arithmetic with and without using /alpha-cut
method: A comparative study. International Journal of Latest Trends in Computing 2(1): 99–
107.
23. Bansal, A. 2011. Trapezoidal fuzzy numbers (a, b, c, d): Arithmetic behavior. International
Journal of Physical and Mathematical Sciences 2(1): 39–44.
24. Banerjee, S., and T.K. Roy. 2012. Arithmetic operations on generalized trapezoidal fuzzy
number and its applications. Turkish Journal of Fuzzy Systems 3(1): 16–44.
25. Vahidi, J., and S. Rezvani. 2013. Arithmetic operations on trapezoidal fuzzy numbers. Journal
Nonlinear Analysis and Application 2013: 1–8. www.ispacs.com/jnaa. Accessed 20 Feb 2015.
26. Kolesárová, A. A note on shape preserving t-norm-based multiplication of fuzzy numbers.
http://www.polytech.univ-savoie.fr/fileadmin/polytech_autres_sites/sites.
27. Dubois, D., and H. Prade. 1988. In Analysis of fuzzy information, 2nd ed, ed. Bezdek, J.C.,
3–39., Fuzzy numbers: An overview Boca Raton: CRC Press.
28. Dubois, D., and H. Prade. 1982. Towards fuzzy differential calculus part 1: Integration of fuzzy
mappings. Fuzzy Sets and Systems 8(1): 1–17.
29. Moore, R.E. 1966. Interval analysis., Prentice-Hall, Series in automatic computation New
Jersey: Prentice-Hall.
30. Hong, D.H. 2001. Shape preserving multiplications of fuzzy numbers. Fuzzy Sets and Systems
123(1): 81–84.
31. Kuchta, D. 2001. Mie˛kka matematyka w zarza̧dzaniu: Zastosowanie liczb przedziałowych
i rozmytych w rachunkowości zarza̧dczej. Wrocław: Oficyna Wydawnicza Politechniki
Wrocławskiej.
32. Enea, M., and T. Piazza. 2004. Project selection by constrained fuzzy ahp. Fuzzy Optimization
and Decision Making 3(1): 39–62.
33. Klir, G.J. 1997. Project selection by constrained fuzzy ahp. Fuzzy Sets and Systems - Special
issue: Fuzzy Arithmetic 91(2): 165–175.
34. Liu, B., and K. Iwamura. 2001. Fuzzy programming with fuzzy decisions and fuzzy simulation-
based genetic algorithm. Fuzzy Sets and Systems 122(2): 253–262.
35. Rȩbiasz, B. 2013. Selection of efficient portfoliosprobabilistic and fuzzy approach, comparative
study. Computers & Industrial Engineering 64(4): 1019–1032.
26 1 Fuzzy Numbers

36. Tanaka K., S. Uejima, and K. Asai. 1982. Linear regression analysis with fuzzy model. IEEE
Transactions on Systems, Man, and Cybernetics, SMC-12, 6:903–907.
37. Tanaka, H. 1987. Fuzzy data analysis by possibilistic linear models. Fuzzy Sets and Systems
24: 363–375.
38. Tanaka, H., and H. Lee. 1998. Interval regression analysis by quadratic programming approach.
IEEE Transactions on Fuzzy Systems 6: 473–481.
39. Hladík, M., and M. Černý. 2012. Interval regression by tolerance analysis approach. Fuzzy Sets
and Systems 193(1): 85–107.
40. Lee, H., and H. Tanaka. 1999. Lower and upper approximation models in interval regression
using regression quantile techniques. European Journal of Operational Research 116: 653–
666.
41. Tanaka, H., and J. Watada. 1988. Possibilistic linear systems and their application to the linear
regression model. Fuzzy Sets and Systems 27: 275–289.
42. Inuiguchi, M., H. Fujita, and T. Tanino. 2002. Robust interval regression analysis based on
minkowski difference. Proceedings of the 41st SICE annual conference SICE 2002, vol. 4,
2346–2351, Japan: Osaka.
43. Tran, L., and L. Duckstein. 2002. Multiobjectine fuzzy regression with central tendency and
possibilistic properties. Fuzzy Sets and Systems 130: 21–31.
44. Zadeh, L.A. 1978. Fuzzy sets as a basis for a theory of possibility. Fuzzy Sets and Systems 1(1):
3–28.
45. Dubois, D., and H. Prade. 1983. Ranking of fuzzy numbers in the setting of possibility theory.
Information Sciences 30(3): 183–224.
46. Liu, B. 2006. A survey of credibility theory. Fuzzy Optimization and Decision Making 5(4):
387–408.
Chapter 2
Ordering of Fuzzy Numbers

Abstract This chapter describes different methods for comparing and ordering
fuzzy numbers. Theoretically, fuzzy numbers can only be partially ordered, and hence
cannot be compared. However, in practical applications, such as decision making,
scheduling, market analysis or optimisation with fuzzy uncertainties, the comparison
of fuzzy numbers becomes crucial.

Theoretically, fuzzy numbers can only be partially ordered, and hence cannot be
compared. However, when they are used in practical applications, e.g., when a deci-
sion must be made among alternatives or an optimal value of an objective function
must be found, the comparison of fuzzy numbers becomes crucial.
There are numerous approaches to the ordering relation between fuzzy numbers
[1–6] qualitative, quantitative and based on α–cuts. Jain [7] and Dubois and Prade
[4] were the first who considered this problem. Some methods to rank fuzzy numbers
were reviewed by Bortolan and Degani [8]. Detyniecki and Yager [3] proposed the
α-weighted valuations of fuzzy numbers. Hong and Kim [9] proposed an easy way to
compute the min and max operation for fuzzy numbers. Asady and Zendehnam [10]
proposed the ranking fuzzy numbers by distance minimisation method. Comparison
of various ranking methods for fuzzy numbers with the possibility of ranking the crisp
numbers was described by Thorani et al. [11]. The problem of comparing of fuzzy
numbers was also considered by Allahviranloo et al. [12]. They proposed a method
based on the centroid point of a fuzzy number and its area. Sevastjanov and Róg
[13] developed a probability-based comparison of fuzzy numbers. The probabilistic
approach was also considered in [14]. The large number of fuzzy ordering methods
can be justified by the fact that different methods can be useful for different purposes.
For example, problems involving ranking, prioritising or choosing between large
number of alternatives will benefit from methods that assign to fuzzy numbers crisp
values thus reducing the fuzzy ordering problem to ordering of real numbers.
An overview of selected approaches to ordering (ranking) of fuzzy numbers is pre-
sented below. The presented approaches can be generally divided into two groups.
The first group consists of methods which enable two fuzzy numbers to be com-
pared. Included in this group are such methods as probabilistic approach, centroid
point approach or radius of gyration approach. To order a set of fuzzy numbers using
28 2 Ordering of Fuzzy Numbers

these methods, some dedicated procedures are required. The second group consists
of methods, which assign to a fuzzy number a crisp value. These are methods such as
Yager ranking index based approach, defuzzification approach or weighted average.
The methods from the second group can be directly used to order a set of fuzzy num-
bers, by employing one of the several methods for ordering (sorting) real numbers.
All the above mentioned methods are compared using an example of ordering four
triangular fuzzy numbers.

2.1 Probabilistic Approach

The probabilistic (also known as probability degree-based or probability-based)

approach to ordering fuzzy numbers is based on the α–cuts representation of fuzzy
numbers. The α–cuts based orderings are so attractive, because they can be used
regardless the type of the membership function. Moreover, each α-level is an interval,
so the powerful tools of interval arithmetic [15] can be employed to solve the problem
of fuzzy ordering [13].
Let a = [a1 , a2 ] and b = [b1 , b2 ] be two closed and compact intervals. The
possibility degree-based ranking method which is shown in Table 2.1 was proposed
by Jiang et al. [16]. The non-overlapping cases are omitted as they are obvious.
A similar, but slightly extended approach to ordering of intervals was proposed
in [13]. Let the real values a ∈ a and b ∈ b be given. They can be considered as two
independent uniform random variables. If a and b overlaps, then some disjoint subin-
tervals can be distinguished. The fall of random variables a and b in the subintervals
[a1 , b1 ], [b1 , a2 ], [a2 , b2 ] may be treated as a set of independent random events.
Let the events Hk : a ∈ ai , b ∈ b j be defined for k = 1, . . . , n, where  ai and
b j are certain subintervals of intervals a and b in accordance with a = ai and
 i
b = bi (n = 4 for the case depicted in Fig. 2.1) [13]. Let P(Hk ) be the probability
i
of event Hk , and P(b > a|Hk ) be the conditional probability of b > a given Hk .
Hence, the composite probability may be expressed as follows:


n
P(b > a) = P(Hk )P(b > a|Hk ). (2.1)
k=1

Table 2.1 Cases of interval comparison proposed by Jiang [16]

Case P(b  a)
a1 −b1 a2 −a1
1. b1  a1 ∧ b2  a2 ∧ b1  a2 b2 −b1 · b2 −b1
a1 −b1
2. a1  b1 ∧ a2  b2 b2 −b1 · 21 ab22 −a
−b1
1

a1 −b1 b2 −a1 a2 −b2 1 b2 −a1 b1 −a1

3. a1  b1 ∧ a2  b2 ∧ a1  b2 b2 −b1 + b2 −b1 · a2 −a1 + 2 b2 −b1 · a2 −a1
a2 −b2
4. b1  a1 ∧ b2  a2 a2 −a1 + 21 ab22 −b
−a1
1
2.1 Probabilistic Approach 29

Fig. 2.1 Example of

overlapping intervals

The resulting formula for the case of overlapping intervals is as follows:

(a2 −b1 )2
P(b > a) = 1 − 1
2 (a2 −a1 )(b2 −b1 )
. (2.2)

The results obtained in [13] for all possible interval overlapping are shown in
Table 2.2.
The above possibility degree-based methods have the following features:
1. 0  P(a  b)  1.
2. If P(b  a)  α, then P(a  b)  1 − α.
3. If P(b  a) = P(a  b), then a ≡ b.
It follows from 2 and 3 that if a ≡ b, then P(b  a) = P(a  b) = 0.5.
This approach can be treated as a framework for elaboration of constructive meth-
ods of interval comparison in various special situations. Some aspects of the interval
comparison and ordering group of intervals, based on this approach, is presented,
e.g., in [17].
Now, let à and B̃ be some arbitrary fuzzy numbers, and let Ãα = {x | μ A (x)  α}
and B̃ α = {x | μ B (x)  α} be their respective α–cuts. Since Ãα and B̃ α are intervals,
probability P α ( B̃ α > Ãα ) for each pair Ãα and B̃ α can be calculated in the way
described in the previous section. The set of probabilities P α , α ∈ (0, 1], may be
treated as the support of the fuzzy subset [13]:

P( Ã > B̃) = {α | P α ( B̃ α > Ãα )}, (2.3)

Table 2.2 Typical cases of interval comparison [13]

Case P(a > b) P(a = b)
b2 −a1
1. a1 > b1 ∧ a1 < b2 ∧ a1 = a2 b2 −b1 0
b1 −a1
2. b1 > a1 ∧ b1 < a2 ∧ b1 = b2 a2 −a1 0
b1 −a1 a2 −a1 b2 −b1
3. b1  a1 ∧ b2  a2 a2 −a1 + 1
2 b2 −b1 a2 −a1
b2 −a2 a2 −a1 a2 −a1
4. a1  b1 ∧ a2  b2 b2 −b1 + 1
2 b2 −b1 b2 −b1
(a2 −b1 )2 (a2 −b1 )2
5. b1  a1 ∧ b2  a2 ∧ b1  a2 1− 1
2 (a2 −a1 )(b2 −b1 ) (a2 −a1 )(b2 −b1 )
(b2 −a1 )2 (b2 −a1 )2
6. a1  b1 ∧ a2  b2 ∧ a1  b2 1− 1
2 (a2 −a1 )(b2 −b1 ) (a2 −a1 )(b2 −b1 )
30 2 Ordering of Fuzzy Numbers

where the values of α may be considered as grades of membership of the fuzzy

number P( Ã > B̃). In this way, the fuzzy subset P( Ã = B̃) may also be easily
created.
In the case of triangular or trapezoidal fuzzy number comparison, the obtained
results may be interpreted as a fuzzy number [13]. Nevertheless, in practice, real
number indices are needed for fuzzy numbers ordering. For this purpose, some char-
acteristic numbers of a fuzzy set [18] could be used. It seems, however, more natural
to substitute the obtained discrete set I of α-levels with a real number:
 
P( B̃ > Ã) = αP α ( B̃ α > Ãα )/ α. (2.4)
α∈I α∈I

The equation (2.4) emphasises that the contribution of the α-level to the overall
probability estimation is increasing with an increase in its number. Of course, as
proposed in [3], the set of complementary parametrised functions of α can be applied
in the equation (2.4) instead of α.

Example 2.1 Let the following four triangular fuzzy numbers, depicted in Fig. 2.2,
be given [19]:

Ã1 = (0.12, 0.19, 0.29), Ã2 = (0.22, 0.32, 0.48),

Ã3 = (0.11, 0.15, 0.23), Ã4 = (0.21, 0.33, 0.49).

The results of pairwise comparison of these numbers using the probability degree-
based approach are presented in Table 2.3. This gives the following order: Ã3 < Ã1 <
Ã2 < Ã4 .

Fig. 2.2 Exemplary triangular fuzzy numbers

2.2 Defuzzification Approach 31

Table 2.3 The results of Pair P Comparison

comparison of triangular
fuzzy numbers Ã1 , Ã2 0.999 Ã1 < Ã2
Ã1 , Ã2 , Ã3 , Ã4 Ã1 , Ã3 0.059 Ã1 > Ã3
Ã1 , Ã4 0.998 Ã1 < Ã4
Ã2 , Ã3 3.00E-08 Ã2 > Ã3
Ã2 , Ã4 0.6823 Ã2 < Ã4
Ã3 , Ã4 1 Ã3 < Ã4

2.2 Defuzzification Approach

Fuzzy numbers can also be ranked using the defuzzification methods. A defuzzifica-
tion is the process of producing a real (crisp) value corresponding to a fuzzy number.
In order to rank fuzzy numbers using the defuzzification approach, the fuzzy numbers
are first defuzzified and then, the obtained crisp numbers are ordered using the order
relation of real numbers. There are several defuzzification methods, among them:

• Centre of area (COA) or centre of gravity (COG):

 xmax
x xμ Ã (x)d x
C O A( Ã) = C OG( Ã) =  min
xmax .
xmin μ Ã (x)d x

• First of maxima (FOM):

F O M( Ã) = min ker( Ã).

• Middle of maxima (MOM):

min ker( Ã) + max ker( Ã)

M O M( Ã) = .
2
• Last of maxima (LOM):

L O M( Ã) = max ker( Ã).

In the case of triangular fuzzy numbers of the form à = (a, b, c):

F O M( Ã) = M O M( Ã) = L O M( Ã) = c.

The results of ordering of fuzzy numbers from Example 2.1 using the abovementioned
defuzzification methods are the same as the result from Example 2.1. The values of
FOM, MOM and LOM are obvious, whereas C OG( Ã1 ) = 0.2, C OG( Ã2 ) = 0.34,
C OG( Ã3 ) = 0.1633, C OG( Ã4 ) = 0.3452.
32 2 Ordering of Fuzzy Numbers

2.3 Centroid-Point Approach

A fuzzy number à can be identified with an ordered pair of continuous real functions
defined on the interval [0, 1], i.e., à = ( f à , g à ), where f à , g à : [0, 1] →  are
continuous functions. Functions f à and g à are called, respectively, the up and down-
parts of a fuzzy number Ã.
The continuity of the functions f and g implies that their images are bounded
intervals (see Fig. 2.3a) denoted, respectively, as U P and D O W N . If, additionally,
the f à and g à functions are monotone, and thus invertible, the following membership
can be defined:
⎧ −1
⎪
⎪ f (x), x ∈ [ f à (0), f à (1)] = [l à , 1−à ],
⎨ Ã
μ à (x) = g −1 (x), x ∈ [g à (1), g à (0)] = [1+à , p à ], (2.5)
⎪
⎪ Ã
⎩ 1, − +
x = [1 , 1 ],
à Ã

if f à is increasing and g à is decreasing, and f à  g à for all y ∈ [0, 1]. The

obtained membership function μ Ã (x), x ∈  represents a mathematical object which
resembles a convex fuzzy number in the classical sense.

Definition 2.1 Let à = (a, b, c, d). Then, the centroid (centre of gravity) point of
à is obtained as follows [20]:

C OG P( Ã) = (x 0 ( Ã), y 0 ( Ã)), (2.6)

Fig. 2.3 An ordered fuzzy number (a), an ordered fuzzy number presented as a fuzzy number in
a classical sense (b), and a simplified mark denoting the order of inverted functions (c)
2.3 Centroid-Point Approach 33

where
⎧  b −1 c d
⎪
⎪ x f à (x)d x + b xd x + c xg −1 (x)d x
⎪
⎪
a
x 0 ( Ã) =  b −1 c  d −1
Ã
,
⎪
⎨
a f à (x)d x + b d x + c g à (x)d x
1 1 (2.7)
⎪
⎪
⎪
⎪
⎪ 0 y f à (y)dy + 0 yg à (y)dy
⎩ 0
y ( Ã) = 1 1 .
0 f à (y)dy + 0 g à (y)dy

In the case of trapezoidal fuzzy numbers, the above formula takes the form:
⎧
⎪
⎨ x 0 ( Ã) = 1
a+b+c+d − cd−ab
,
3 (c+d)−(a+b)
(2.8)
⎪
⎩ y 0 ( Ã) = 1
3
1+ c−b
(c+d)−(a+b)
.

Based on a centroid point, two fuzzy numbers à and B̃ are compared using the
following rules [21]:

If x 0 ( Ã) > x 0 ( B̃), Then à > B̃.

If x 0 ( Ã) < x 0 ( B̃), Then à < B̃.
If x 0 ( Ã) = x 0 ( B̃), Then
(2.9)
If y 0 ( Ã) > y 0 ( B̃), Then à > B̃.
Else If y 0 ( Ã) < y 0 ( B̃), Then à < B̃.
Else à = B̃.

For the fuzzy numbers from Example 2.1 the following centroid points were
obtained (Table 2.4):
This gives the ordering: Ã3 < Ã1 < Ã2 < Ã4 . The method is rather simple, but
it requires a pairwise comparison of fuzzy numbers to be ordered.

Table 2.4 The considered Fuzzy number x0 y0

fuzzy numbers and their
centroid points Ã1 = (0.12, 0.19, 0.29) 0.343 0.333
Ã2 = (0.22, 0.32, 0.48) 0.34 0.333
Ã3 = (0.11, 0.15, 0.23) 0.2 0.333
Ã4 = (0.21, 0.33, 0.49) 0.163 0.333
34 2 Ordering of Fuzzy Numbers

Table 2.5 Yager index for Fuzzy number Yager index

the fuzzy numbers from
Example 2.1 Ã1 = (0.12, 0.19, 0.29) 0.1317
Ã2 = (0.22, 0.32, 0.48) 0.2233
Ã3 = (0.11, 0.15, 0.23) 0.1067
Ã4 = (0.21, 0.33, 0.49) 0.2267

2.4 Yager Ranking Index Approach

In [22], Yager proposed the following index to ordering fuzzy numbers:

Y ( Ã) = 1
2
( f à (y) + g à (y))dy.
0

For example, given two triangular fuzzy numbers Ã1 = (35, 50, 61) and Ã2 =
(30, 41, 49) the Yager’s values are Y ( Ã1 ) = 49 and Y ( Ã2 ) = 40.25. Thus, Ã2 is
smaller than Ã1 in the context of Yager index.
For the fuzzy numbers from Example 2.1, the Yager index takes the values pre-
sented in Table 2.5. They yield exactly the same order as the one obtained using the
possibilistic approach.

2.5 Degree of Possibility Approach

The ordering of fuzzy numbers using priority approach is based on the research
presented in [19].

Definition 2.2 Given two convex fuzzy numbers à and B̃ the degree of possibility
of à > B̃ is defined as

V ( Ã > B̃) = sup{min{μ Ã (x), μ B̃ (y)}}. (2.10)

x≥y

Thus, if there exists a pair (x, y) such that x  y and μ Ã (x) = μ B̃ (y) = 1, then
the degree of possibility V ( à > B̃) = 1. Since à and B̃ are convex fuzzy numbers,
the following holds [19]:

V ( Ã > B̃) = 1 iff sup ker( Ã)  inf ker( B̃),

(2.11)
V ( B̃  Ã) = hgt( Ã ∩ B̃) = μ Ã (d),

where d is the x-coordinate of the highest intersection point between μ Ã and μ B̃ .

When à = (a1 , a2 , a3 ) and b̃ = (b1 , b2 , b3 ), then
2.5 Degree of Possibility Approach 35

a1 − b3
hgt( Ã ∩ B̃) = . (2.12)
(b2 − b3 ) − (a2 − a1 )

To compare à and B̃ both values V ( à  B̃) and V ( B̃  Ã) are needed.

Now, the degree of possibility for a convex fuzzy number à to be greater than k
convex fuzzy numbers Ãi (i = 1, . . . , k) is given by

V ( Ã ≥ Ã1 , . . . , Ãk ) = V [( Ã ≥ Ã1 ) ∧ ( Ã ≥ Ã2 ) ∧ · · · ∧ ( Ã  Ãk ) ]

(2.13)
= min V ( Ã  Ãi ), i = 1, . . . , k.

For the triangular fuzzy numbers from Example 2.1, the following values of the
respective degrees of possibility are obtained.

V ( Ã1  Ã2 , Ã3 , Ã4 ) = 0.35,

V ( Ã2  Ã1 , Ã3 , Ã4 ) = 0.96,
V ( Ã3  Ã1 , Ã2 , Ã4 ) = 0.06,
V ( Ã4  Ã1 , Ã2 , Ã3 ) = 1.

This gives exactly the same order as the one obtained using the probabilistic approach
and Yager index.

2.6 Weighted Averaging Approach Based on α–cuts

This section describes the ordering of LR-type fuzzy numbers associated with
defuzzification of parametrically represented fuzzy numbers [23]. In the case of
LR-type fuzzy numbers, the parametric representation (1.8) can be written in the
following form:
à = (α, [L −1
Ã
(α), R −1
Ã
(α)]), (2.14)
α∈[0,1]

where L −1
Ã
, R −1
Ã
: [0, 1] →  are inverse functions of the respective shape functions
of an LR-type fuzzy number Ã.

Definition 2.3 ([24]) Let à ∈ F() L R . The weighted averaging based on α–cuts
representation of a fuzzy number à is defined by:

I ( Ã) = (c L L −1
Ã
(α) + c R R −1
Ã
(α)) p(α)dα, (2.15)
0

where c L , c R are, respectively, the optimism and pessimism parameters, p(α) is

a distribution function of the importance of the α–cuts.
36 2 Ordering of Fuzzy Numbers

The c L , c R parameters and the function p(α) satisfy the conditions:

c L > 0, c R > 0, c L + c R = 1,
1
p : [0, 1] → + , p(α)dα = 1.
0

The function p(α) is also called the weighted averaging parameter. Following [23],
it is assumed that
p(α) = (k + 1)αk ,

where k > 0 is a parameter.

Theorem 2.1 ([24]) Let à = (a, b, α, β) L R and assume that the distribution of the
function of the importance of the degrees have the form of relation (2.15). Then, the
following formula is valid for weighted averaging:
 
k+1 k+1
I ( Ã) = c L β− (β − α) + c R a − (b − a) . (2.16)
k+2 k+2

The value I ( Ã) is a crisp value used to rank fuzzy numbers. The greater this value
is, the greater is the fuzzy number. Moreover, I ( Ã) = I ( Ã) = I ( B̃) if and only if
à = B̃.

Example 2.2 In this example it is assumed that p(α) = 2 (k = 1), and the “opti-
mism/pessimism” coefficients are 0.5. Now, let the following three sets of trapezoidal
fuzzy numbers and a set of triangular fuzzy numbers be given (see Fig. 2.4):
Set 1: Ã1 = (0.5, 0.5, 0.1, 0.5), Ã2 = (0.7, 0.7, 0.3, 0.3), Ã3 = (0.9, 0.9, 0.5, 0.1);
Set 2: Ã1 = (0.4, 0.7, 0.4, 0.1), Ã2 = (0.5, 0.5, 0.3, 0.4), Ã3 = (0.6, 0.6, 0.5, 0.2);
Set 3: Ã1 = (0.5, 0.5, 0.2, 0.2), Ã2 = (0.5, 0.8, 0.2, 0.1), Ã3 = (0.5, 0.5, 0.2, 0.4);
Set 4: Ã1 = (0.12, 0.19, 0.29), Ã2 = (0.22, 0.32, 0.48), Ã3 = (0.11, 0.15, 0.23),
Ã4 = (0.21, 0.33, 0.49).

The ranking index values obtained for the set 1 are I ( Ã1 ) = 0.37, I ( Ã2 ) = 0.50,
I ( Ã3 ) = 0.63. This gives the following order Ã1 < Ã2 < Ã3 . For the set 2, the
ranking index values are I ( Ã1 ) = 0.45, I ( Ã2 ) = 0.42, I ( Ã3 ) = 0.50, which gives
the order Ã2 < Ã1 < Ã3 . For the set 3, the ranking index values are I ( Ã1 ) = 0.35,
I ( Ã2 ) = 0.43, I ( Ã3 ) = 0.38, which gives the order Ã1 < Ã3 < Ã2 . Finally, for the
set 4 which is the same as in Example 2.1, the ranking index values are I ( Ã1 ) = 0.14,
I ( Ã2 ) = 0.22, I ( Ã3 ) = 0.10, I ( Ã4 ) = 0.23. This gives exactly the same order as
the one obtained using the previous approaches.
The α–cuts based approach is the most time consuming as it requires the pairwise
comparison of all fuzzy numbers to be ordered. Also, the procedure of computing
the possibility degree is more complicated that the computation of ranking indices
in the two latter approaches.
2.7 Two-Dimensional Radius of Gyration Approach 37

(a) (b)

(c) (d)

Fig. 2.4 Set 1 (a); Set 2 (b), Set 3 (c); Set 4 (d)

2.7 Two-Dimensional Radius of Gyration Approach

The two-dimensional radius of gyration (ROG) or gyradius is a concept in mechanics

[25]: 
rg = I /A (2.17)

where I is the second moment of area (see Fig. 2.5) and A is the total cross-sectional
area. The second moment of area of an arbitrary shape with respect to an arbitrary
axis Z is defined and is computed by:

IZ = r 2 d A, (2.18)
A

Fig. 2.5 A scheme of how

the second moment of area is
calculated for an arbitrary
shape with respect to the Z
axis; r is the radial distance
to the element dA, with
projections x and y on the
axes
38 2 Ordering of Fuzzy Numbers

where d A is a differential area of the arbitrary shape and r is a distance from the axis
Z to d A.
For example, when the desired reference axis is the X -axis, the second moment
of area, Ix can be computed in Cartesian coordinates as:

Ix = y 2 d xd y (2.19)
A

The ROG point (r xà , r yà ) for a fuzzy number à is provided as [25]:

r xà = Ix /A, (2.20)

r yà = I y /A, (2.21)

where Ix is the second moment of area with respect to x, and I y is the second moment
of area with respect to y. It is assumed that the mass density at each point of the area
equals 1.
In the case of a trapezoidal fuzzy number, the second moment of area can be
calculated in the following way. First, the trapezoidal area of a fuzzy number is
divided onto three areas A1 , A2 , A3 (see Fig. 2.6). It is known that for an area made
up of a number of simple shapes, the second moment of area is the sum of the second
moments of each of the individual areas about the desired axis [25]:

Ix = IxA1 + IxA2 + IxA3

(2.22)
I y = I yA1 + I yA2 + I yA3

The respective moments of inertia of the areas are given by [25]:

1
b−a
IxA1 = y2d A = y 2 (b − q)(1 − y)dy = (2.23)
A1 0 12

(b − a)3 (b − a)a 2 2(b − a)2 a

I yA1 = x 2d A = + + (2.24)
A1 4 2 3

Fig. 2.6 A division of a trapezoid into three parts A1 , A2 and A3

2.7 Two-Dimensional Radius of Gyration Approach 39

(c − b)
IxA2 = (2.25)
3

(c − b)3
I yA2 = + (c − b)b2 + (c − b)2 b (2.26)
3

(d − c)
IxA3 = (2.27)
12

(d − c)3 (d − c)c2 (d − c)2 c

I yA3 = + + (2.28)
12 2 3
Then, the ROG point of a trapezoidal fuzzy number is calculated as [25]:
 A A A
I x 1 +I x 2 +I x 3
r xà = ((c−b)+(d−a))/2
,
 (2.29)
A A A
I y 1 +I y 2 +I y 3
r yà = ((c−b)+(d−a))/2
.

For a crisp number a, the ROG point is defined by [25]:

√
r xa = 3/3, r ya = a. (2.30)

The ROG point (r xà , r yà ) is used to define the index [25]

S( Ã) = r xà · r yà (2.31)

which is used to compare fuzzy numbers. The larger is the index, the greater is a fuzzy
number. Thus, given two fuzzy numbers à and B̃, the following holds:

If S( Ã) > S( B̃) Then à > B̃,

If S( Ã) < S( B̃) Then à < B̃, (2.32)
If S( Ã) = S( B̃) Then à = B̃.

Example 2.3 The values of the index S obtained for the fuzzy numbers from
Example 2.1 are presented in Table 2.6.

This gives exactly the same ordering as those obtained using previously described
approaches.
40 2 Ordering of Fuzzy Numbers

Table 2.6 Yager index for Fuzzy number S

the fuzzy numbers from
Example 2.1 Ã1 = (0.12, 0.19, 0.29) 0.8167
Ã2 = (0.22, 0.32, 0.48) 1.2484
Ã3 = (0.11, 0.15, 0.23) 0.4410
Ã4 = (0.21, 0.33, 0.49) 0.4217

2.8 Fuzzy Maximising-Minimising Points Approach

The ordering of fuzzy numbers using fuzzy maximising-minimising points is based

on the centre of gravity point, defined in the previous section, left and right spreads
and the distance between fuzzy numbers.

Definition 2.4 The distance between two arbitrary fuzzy numbers à and B̃ is defined
by:
 1 1 
d( Ã, B̃) = ( f à (y) − f B̃ (y)) dy +
2
(g à (y) − g B̃ (y)) dy
2
(2.33)
0 0

The fuzzy minimising-maximising points are obtained using the method from
[26]. Let the fuzzy numbers Ãi , i = 1, 2, . . . , n be given and let M̃ denote the fuzzy
maximising point and m̃ the fuzzy minimising point. The COGP of the minimising
and maximising points are computed as follows:

C OG P( M̃) = max {x 0 ( Ãi )}, max {y 0 ( Ãi )}
i=1,2,...,n i=1,2,...n


C OG P(m̃) = min {x 0 ( Ãi )}, min {y 0 ( Ãi )}
i=1,2,...,n i=1,2,...n

The left and right spreads of M̃ and m̃ are computed analogously:

L M̃ = max {L Ãi }, R M̃ = max {R Ãi },

i=1,2,...,n i=1,2,...,n

L m̃ = min {L Ãi }, Rm̃ = min {R Ãi }.

i=1,2,...n i=1,2,...n

Now, given C OG P( M̃), L M̃ , R M̃ and C OG P(m̃) L m̃ , Rm̃ , the goal is to uniquely

determine, respectively, M̃ and m̃.
In general case, an unknown fuzzy number à can be uniquely determined based
on its centroid point (x 0 ( Ã), y 0 ( Ã)) and left L and right R spreads by solving the
following system of nonlinear equations [26]:
2.8 Fuzzy Maximising-Minimising Points Approach 41

⎧
⎪ b−a = L
⎪
⎪
⎨d − c = R
⎪
1
a + b + c + d − (c+d)−(a+b)
cd−ab
= x 0 ( Ã)
⎪
⎪
3
⎪
⎪
⎩1
3
1 + (c+d)−(a+b)
c−b
= y 0 ( Ã)

The fuzzy ranking using fuzzy minimising-maximising points uses the following
relative closeness coefficient:

D ÃL
D( Ã) = γ( Ã) · , (2.34)
1 + D ÃR

where
D ÃL = d( Ã, m̃),

D ÃR = d( Ã, M̃),

and  1
1, if 0 { f à (y) + g à (y)}dy  0,
γ( Ã) = 1
−1, if 0 { f à (y) + g à (y)}dy < 0.

The ranking rules for fuzzy numbers Ãi , i = 1, 2, . . . , n, are the following [26]:

Ai < A j iff D(Ai ) < D(A j )

Ai > A j iff D(Ai ) > D(A j )
Ai ≈ A j iff D(Ai ) = D(A j )

Example 2.4 Consider the fuzzy numbers from Example 2.1. The corresponding
fuzzy maximising and minimising points are depicted in Fig. 2.7.

Fig. 2.7 Fuzzy maximising M̃ and minimising m̃ points

42 2 Ordering of Fuzzy Numbers

The following values of the relative closeness coefficient were obtained: D(A1 ) =
0.0067, D(A2 ) = 0.0395, D(A3 ) = 0.0015, D(A4 ) = 0.0468, which gives exactly
the same ordering as those obtained using previously described approaches.
The method is quite complicated compared to other presented methods. It is also
time consuming and requires to use additional tools, such as numerical integration
and solving systems of nonlinear equations. It also requires the pairwise comparison
of fuzzy numbers to be ordered.

2.9 Area Based Approach

For a fuzzy number à = ( f à , g à ), the following values are defined [27]:

Ãl = m + 21 Hl ,
(2.35)
Ãu = m + 21 Hu

where m = 21 ( f Ã−1 (1) + g −1

Ã
(1)), and Hl , Hu are defined as follows:
1
0 f à (y)dy
Hl = 1 1 ,
f
0 à (y)dy+ 0 g à (y)dy
1 (2.36)
0 g à (y)dy
Hu = 1 1
0 f à (y)dy+ 0 g à (y)dy

Now, for given two fuzzy numbers à and B̃, the following values are defined [27]:

R( Ã, B̃) = Ãu − B̃u , R( Ã, B̃) = Ãl − B̃l (2.37)

They are used to determine the comparison rules:

R( B̃, Ã) > R( Ã, B̃) iff à < B̃,

(2.38)
R( B̃, Ã) = R( Ã, B̃) iff à ≈ B̃.

It follows from the definition of R( Ã, B̃) and R( B̃, Ã) that

R( Ã, B̃) = −R( B̃, Ã)

R( Ã, B̃) = −R( B̃, Ã)

Thus, the comparison rules (2.38) can be written in the following form:

−R( Ã, B̃) > R( Ã, B̃) iff à < B̃,

(2.39)
−R( Ã, B̃) = R( Ã, B̃) iff à = B̃.
2.9 Area Based Approach 43

Table 2.7 The values of R, R and the pairwise comparison results

R R Comparison
R( Ã1 , Ã2 ) = −0.07558 R( Ã1 , Ã2 ) = −0.05442 Ã1 < Ã2
R( Ã1 , Ã3 ) = 0.006155 R( Ã1 , Ã3 ) = 0.033845 Ã1 > Ã3
R( Ã1 , Ã4 ) = −0.074654 R( Ã1 , Ã4 ) = −0.065346 Ã1 < Ã4
R( Ã2 , Ã3 ) = 0.081735 R( Ã2 , Ã3 ) = 0.088265 Ã2 > Ã3
R( Ã2 , Ã4 ) = 0.000926 R( Ã2 , Ã4 ) = −0.010926 Ã2 > Ã4
R( Ã3 , Ã4 ) = −0.080809 R( Ã3 , Ã4 ) = −0.099191 Ã3 > Ã4

Example 2.5 Consider the fuzzy numbers from Example 2.1. The values of R, R
and the pairwise comparison results are presented in Table 2.7.

This gives exactly the same order as the one obtained using the methods described
so far.

2.10 Left and Right Dominance Approach

The ordering of fuzzy numbers based on the left and right dominance was proposed
in [28]. This approach uses left and right bounds of selected α–cuts of fuzzy numbers
to be compared.
Definition 2.5 The left Di,L j and right Di,R j dominance of a fuzzy number Ãi over
a fuzzy number à j is defined as the average difference of the left and right bounds
of Ãi and à j at some α-levels:

1 n
DiLj = (lik − l jk ) (2.40)
n+1 k=0

1 n
DiRj = (rik − r jk ) (2.41)
n+1 k=0

where n is the numbers of α–cuts, lik , rik are, respectively, left and right spreads of
a fuzzy number Ãi at the αk -level.
It is assumed that α-levels are spread uniformly, i.e., k/n, k = 1, 2, . . . , n. The
values Di,L j and Di,R j approximate the area difference of Ãi over à j according to the
membership axis to the, respectively, left and right membership function as n → ∞
[28]. The total dominance of Ãi over à j with the index of optimism β ∈ [0, 1] is
defined as follows.

Definition 2.6 The total dominance Di,T j of a fuzzy number Ãi over a fuzzy number
à j with optimism index β ∈ [0, 1] is defined as a convex combinations of left Di,L j
44 2 Ordering of Fuzzy Numbers

Table 2.8 The values of total Total dominance index Value

dominance T (0.5)
D12 −0.1375
T (0.5)
D13 0.0375
T (0.5)
D14 −0.1425
T (0.5)
D23 0.1750
T (0.5)
D24 −0.005
T (0.5)
D34 −0.1800

and right Di,R j dominance:

DiTj (β) = β DiLj + (1 − β)DiRj (2.42)

The index of optimism is used to reflect a decision maker’s degree of optimism [28].
The total dominance index is used to define the rules of comparison of two fuzzy
numbers. They are the following:

If DiTj < 0 Then Ãi < Ã j ,

If DiTj > 0 Then Ãi > Ã j ,
If DiTj = 0 Then Ãi = Ã j .

Example 2.6 Consider the fuzzy numbers from Example 2.1. The obtained values
of the total dominance with β = 0.5 and n = 5 are summarised in Table 2.8.
This gives the ordering Ã3 < Ã1 < Ã2 < Ã4 , which exactly the same as the one
obtained using the previous methods.

2.11 An α-weighted Valuations Approach

Approaches to the ranking of fuzzy numbers based upon the idea of associating
with a fuzzy number a scalar value, i.e., its valuation, was developed by Yager [22].
Later, Yager and Filev [29] improved this valuation method by the transformation of
a fuzzy subset into an associated probability distribution. They introduced a family
of parametric valuation functions. The problem of ranking fuzzy numbers using
valuation methods was also considered by Detyniecki and Yager [3].
A generalised formula for a class of valuation functions has the form:
1
Ave( Ãα ) f (α)dα
V al( Ã) = 0
1 (2.43)
0 f (α)dα
2.11 An α-weighted Valuations Approach 45

where f is a mapping f : [0, 1] → [0, 1]. In [29], Yager and Filev proposed two
complementary families of parametric valuation functions. The one is an increasing
family:
f : [0, 1] α → αq ∈ [0, 1], q > 0,

and the other one is decreasing

f : [0, 1] α → (1 − α)q ∈ [0, 1], q > 0.

Some interesting properties of this two families of functions can be found in [29].
One of them is that increasing family emphasises the higher α-levels, whereas the
decreasing family emphasises lower α-levels, which causes that these two families
can produce two opposite orderings.
In order to calculate the valuation for a given fuzzy number Ã, the value of

inf( Ãα ) + sup( Ãα )

Ave( Ãα ) =
2
must be first computed. In the case of trapezoidal fuzzy numbers

a + (b − a)α + d − (d − c)α b+c a+d

Ave( Ãα ) = = α+ (1 − α).
2 2 2
Then, the valuation formula (2.43) takes the form:
1
1
((b + c)α + (a + d)(1 − α)) f (α)dα
V al( Ã) = 2 0
1 . (2.44)
0 f (α)dα

which can be simplified to:

b+c a+d
V al( Ã) = w+ (1 − w),
2 2
where
1
α f (α)dα
w = 0 1 .
0 f (α)dα

For the increase case

q +1
w= ,
q +2

and for the decrease case

1
w= .
q +2
46 2 Ordering of Fuzzy Numbers

Table 2.9 The results of q V al( Ã1 ) V al( Ã2 ) Comparison

comparison of Ã1 and Ã2
using increasing valuation 0 6 7 <
function with different values 1 6.33 6.67 <
of the parameter q 2 6.5 6.5 =
3 6.6 6.4 >
∞ 7 6 >

Example 2.7 Consider the following triangular fuzzy numbers: Ã1 = (1, 7, 9), Ã2 =
(4, 6, 12), Ã3 = (5, 8, 9), Ã4 = (2, 9, 10). The results of comparison of Ã1 and Ã2 ,
obtained using an increasing functions with different values of q, are presented in
Table 2.9. The results of ordering obtained using increasing and decreasing functions
with q = 2 are given in Table 2.11. Finally, Table 2.10 presents the ordering of
fuzzy numbers from Example 2.1 obtained using decreasing and increasing valuation
functions with different value of the parameter q.

The results show that for q = 2, the ordering of fuzzy numbers is exactly the
same as the one obtained using other considered methods.

Table 2.10 The results of comparison of exemplary fuzzy numbers using increasing and decreasing
valuation functions with different values of the parameter q
V al(·) V al(·) Order Order
(increasing) (decreasing) (increasing) (decreasing)
q=0
Ã1 6.0 6.0 Ã4 = Ã3 > Ã2 > Ã1 Ã4 = Ã3 > Ã2 > Ã1
Ã2 7.0 7.0
Ã3 7.5 7.5
Ã4 7.5 7.5
q=1
Ã1 6.33 5.67 Ã4 > Ã3 > Ã2 > Ã1 Ã3 = Ã2 > Ã4 > Ã1
Ã2 6.67 7.33
Ã3 7.67 7.33
Ã4 8.00 7.00
q=2
Ã1 6.5 5.5 Ã4 > Ã3 > Ã2 = Ã1 Ã2 > Ã3 > Ã4 > Ã1
Ã2 6.5 7.5
Ã3 7.75 7.25
Ã4 8.25 6.75
q=∞
Ã1 7.0 5.0 Ã4 > Ã3 > Ã1 > Ã2 Ã2 > Ã3 > Ã4 > Ã1
Ã2 6.0 8.0
Ã3 8.0 7.0
Ã4 9.0 6.0
References 47

Table 2.11 The results of comparison of exemplary fuzzy numbers using increasing and decreasing
valuation functions with different values of the parameter q
V al(·) V al(·) Order Order
(increasing) (decreasing) (increasing) (decreasing)
q=0
Ã1 0.2 0.2 Ã4 = Ã2 > Ã1 > Ã3 Ã4 = Ã2 > Ã1 > Ã3
Ã2 0.34 0.34
Ã3 0.16 0.16
Ã4 0.34 0.34
q=1
Ã1 0.2 0.2 Ã4 > Ã2 > Ã1 > Ã3 Ã4 = Ã2 > Ã1 > Ã3
Ã2 0.33 0.34
Ã3 0.16 0.16
Ã4 0.34 0.34
q=2
Ã1 0.19 0.2 Ã4 > Ã2 > Ã1 > Ã3 Ã4 > Ã2 > Ã1 > Ã3
Ã2 0.33 0.34
Ã3 0.16 0.17
Ã4 0.34 0.35
q=∞
Ã1 0.19 0.2 Ã4 = Ã2 > Ã1 > Ã3 Ã4 = Ã2 > Ã1 > Ã3
Ã2 0.33 0.35
Ã3 0.15 0.17
Ã4 0.33 0.35

References

1. Baas, S.M., and H. Kwakernaak. 1977. Rating and ranking of multiple-aspect alternative using
fuzzy sets. Automatica 13(1): 47–58.
2. Chanas, S., and D. Kuchta. 1996. Multiobjective programming in optimization of interval
objective functions a generalized approach. European Journal of Operational Research 94(3):
594–598.
3. Detyniecki, M., and R.R. Yager. 2000. Ranking fuzzy numbers using α-weighted valuations.
International Journal of Uncertainty, Fuzziness and Knowledge-Based Systems 8(3): 573–591.
4. Dubois, D., and H. Prade. 1983. Ranking of fuzzy numbers in the setting of possibility theory.
Information Sciences 30(3): 183–224.
5. Sengupta, A., and T.P. Kumar. 2009. Fuzzy preference ordering of intervals. Fuzzy preference
ordering of interval numbers in decision problems, Studies in fuzziness and soft computing,
vol. 238, pp. 59–89. Berlin: Springer.
6. Yager, R.R., M. Detyniecki, and B. Bouchon-Meunier. 2001. A context-dependent method for
ordering fuzzy numbers using probabilities. Information Sciences 30(3): 237–255.
7. Jain, R. 1976. Decision-making in the presence of fuzzy variables. IEEE Transactions on
Systems, Man and Cybernetics 6(10): 698–703.
8. Bortolan, G., and R. Degani. 1985. A review of some methods for ranking fuzzy subsets. Fuzzy
Sets and Systems 15(1): 1–19.
48 2 Ordering of Fuzzy Numbers

9. Kim, K.T., and D.H. Hong. 2006. An easy computation of min and max operations for fuzzy
numbers. Fuzzy Sets and Systems 21(1–2): 555–561.
10. Asady, B., and A. Zendehnam. 2007. Ranking fuzzy numbers by distance minimization. Applied
Mathematical Modelling 31(11): 2589–2598.
11. Thorani, Y.L.P., P.P.B. Rao, and N.R. Shankar. 2012. Ordering generalized trapezoidal fuzzy
numbers. International Journal of Contemporary Mathematical Sciences 7(12): 555–573.
12. Allahviranloo, T., M.A. Jahantigh, and S. Hajighasemi. 2013. A new distance measure and
ranking method for generalized trapezoidal fuzzy numbers. Mathematical Problems in Engi-
neering 2013: 1–6.
13. Sevastjanov, P.V., and P. Rog. 2003. A probabilistic approach to fuzzy and crisp interval order-
ing. TASK Quarterly : Scientific bulletin of Academic Computer Centre in Gdansk 7(1): 147–
156.
14. Huynh, V.-N., Y. Nakamori, and J. Lawry. 2008. A probability-based approach to comparison
of fuzzy numbers and applications to target-oriented decision making. IEEE Transactions on
Fuzzy Systems 16(2): 371–387.
15. Moore, R.E. 1966. Interval analysis., Prentice-Hall, Series in automatic computation New
Jersey: Prentice-Hall.
16. Jiang, C., X. Han, G.R. Liu, and G.P. Liu. 2008. A nonlinear interval number program-
ming method for uncertain optimization problems. European Journal of Operational Research
188(1): 1–13.
17. Sevastianov, P., P. Rog, and K. Karczewski. 2002. A probabilistic method for ordering group
of intervals. Informatyka Teoretyczna i Stosowana R2(2): 45–53.
18. Wang, X., and E.E. Kerre. 2001. Reasonable properties for the ordering of fuzzy quantities (ii).
Fuzzy Sets and Systems 118(3): 387–405.
19. Chang, D.-Y. 1996. Applications of the extent analysis method on fuzzy AHP. European
Journal of Operational Research 95(3): 649–655.
20. Wang, Y.M., J.B. Yang, D.L. Xu, and K.S. Chin. 2006. On the centroids of fuzzy numbers.
Fuzzy Sets and Systems 157(7): 919–926.
21. Abdullah, L., and N.J. Jamal. 2010. Centroid-point of ranking fuzzy numbers and its application
to health related quality of life indicators. International Journal on Computer Science and
Engineering 2(8): 2773–2777.
22. Yager, R. 1981. A procedure for ordering fuzzy subsets of the unit interval. Information Sciences
24(2): 143–161.
23. Saneifard, R. 2009. Ranking L-R fuzzy numbers with weighted averaging based on levels.
International Journal of Industrial Mathematics 1(2): 163–173.
24. Nasibov, E.N., and A. Mert. 2007. On methods of defuzzification of parametrically represented
fuzzy numbers. Automatic Control and Computer Sciences 41(5): 163–173.
25. Deng, Y., Z. Zhenfu, and L. Qi. 2006. Ranking fuzzy numbers with an area method using radius
of gyration. Computers and Mathematics with Applications 51(6–7): 1127–1136.
26. Salahshour, S., S. Abbasbandy, and T. Allahviranloo. 2011. Ranking fuzzy numbers using fuzzy
maximizing-minimizing points. In 2011.115, Proceedings of the 7th conference of the European
Society for Fuzzy Logic and Technology (EUSFLAT-2011) and LFA-2011, ed. Galichet, S.,
J. Montero, and G. Mauris, Advances in intelligent systems research, vol. 1, pp. 763–769.
European Society for Fuzzy Logic and Technology (EUSFLAT-2011) and LFA.
27. Nasseri, S.H., F. Taleshian, Z. Alizadeh, and J. Vahidi. 2012. A new method for ordering LR
fuzzy numbern. The Journal of Mathematics and Computer Science 4(3): 283–294.
28. Chen, L.-H., and H.-W. Lu. 2001. An approximate approach for ranking fuzzy numbers based on
left and right dominance. Computers and Mathematics with Applications 41(12): 1589–1602.
29. Yager, R.R., and D. Filev. 1999. On ranking fuzzy numbers using valuations. International
Journal of Intelligent Systems 14(12): 1249–1268.
Chapter 3
Fuzzy Random Variable
and the Dempster-Shafer Theory
of Evidence

Abstract This chapter presents the concept of uncertainty propagation in real world
desicion problems, where some input parameters are stochastic while information
about others is partial and is represented by fuzzy random variable. It also introduces
fuzzy random variable and the Dempster-Shafer theory which provide mathematical
background for such propagation.

3.1 Fuzzy Random Variable

In many decision problems, it is necessary to jointly consider randomness and impre-

cision, because data coming from various sources can be subject to both types of
uncertainty. One of the possible approaches that enables to deal with this issue is
based on the concept of a fuzzy random variable, which extends the classical defi-
nition of a random variable. It was introduced by Féron [1] and later modified by,
e.g., Nahmias [2], Stein and Talati [3], Kwakernaak [4, 5], Puri and Ralescu [6],
Diamond and Kloeden [7] and Kruse [8]. Krätschmer [9] surveyed all of these def-
initions and proposed a unified approach. He defined a fuzzy random variable as
a function (mapping) that assigns a fuzzy subset to each of possible outputs of a
random experiment. The similar approach to a fuzzy random variable was proposed
by Liu and Liu [10–12]. Their approach is adopted in this book.

Definition 3.1 Let (Ω, Σ, P) be a probability space. A fuzzy random vector is

a mapping ξ = (ξ1 , ξ2 , . . . , ξn ) : Ω → F n () such that for any closed C ⊆ n

ξ ∗ (C)(ω) = Pos{ξ(ω) ∈ C} = sup μξ(ω) (t) (3.1)

t∈C

is a measurable function of ω, where μξ(ω) is defined as

μξ(ω) (t) = min μξi (ω) (ti ) (3.2)

1i

for any t = (t1 , t2 , . . . , tn ) ∈ n .

50 3 Fuzzy Random Variable and the Dempster-Shafer Theory of Evidence

Three kinds of common fuzzy random variables, triangular, trapezoidal, and

normal are used the most often. They are defined as follows:
• a fuzzy random variable ξ is said to be triangular if for each ω, ξ(ω) is a triangular
fuzzy variable given by (a(ω), b(ω), c(ω)), where a, b and c are random variables
defined on a probability space Ω,
• a fuzzy random variable ξ is said to be trapezoidal if for each ω, ξ(ω) is a trapezoidal
fuzzy variable given by (a(ω), b(ω), c(ω), d(ω)), where a, b, c and d are random
variables defined on a probability space Ω,
• a fuzzy random variable ξ is said to be normal if for each ω, ξ(ω) is a fuzzy variable
with a membership function given by
   
r − c(ω) 2
μ(r ) = exp −
w(ω)

where c(ω) and w(ω) are random variables defined on a probability space Ω.

3.2 Dempster–Shafer (D–S) Theory of Evidence

The theory of belief functions (also called evidence theory) was introduced by Shafer
[13]. It allows imprecision and variability to be treated separately within a single
framework. Indeed, belief functions provide mathematical tools to process infor-
mation, which is at the same time of random and imprecise nature. This kind of
knowledge is typically found in real decision processes, where some parameters
are described by probability distributions, whereas others are described by fuzzy
numbers. Contrary to probability theory, which assigns probability weights to atoms
(elements of the referential), the theory of evidence may assign such weights to any
subsets, called focal sets, with the understanding that portions of these weights may
move freely from one element of such subsets to another. Most often, a sample of
random intervals is obtained. In such case, information is presented in the form of
intervals [a i , a i ] (i = 1, 2, . . . , I ). To each interval is attached a probability vi . That
is, a mass distribution vi on intervals is obtained. The probability mass vi can be
freely re-allocated to points within interval [a i , a i ]. However, there is not enough
information to do it.
As in the possibility theory, the evidence theory provides two indicators, plausi-
bility Pl and belief Bel to qualify the validity of a proposition stating that the value
of variable X should lie within a set A (a certain interval for example). Plausibility
Pl and belief Bel measures are defined from the mass distribution.
Definition 3.2 Let P() be the power set of . Then, the mass distribution is a
mapping
v : P() → [0, 1], (3.3)

such that E∈P() v(E) = 1.
3.2 Dempster–Shafer (D–S) Theory of Evidence 51

Definition 3.3 The belief function is defined by:


Bel(A) = v(E). (3.4)
E, E⊆A

Definition 3.4 The plausibility function is defined by:


Pl(A) = v(E) = 1 − Bel(A). (3.5)
E, E∩A=ϕ

The element E is called a focal element of P() if v(E) > 0.

Bel(A) gathers the imprecise evidence that asserts A. Following Dempster [14],
this is the minimal amount of probability that can be assigned to A by sharing the
probability weights defined by the mass function among single values in the focal
sets. Whereas Pl(A) gathers the imprecise evidence that does not contradict A. This
is the maximal amount of probability that can be assigned to A in the same fashion.
Evidence theory encompasses possibility and probability theory because
[13, 15]:
• When focal elements are nested, the belief measure Bel is a necessity measure, i.e.,
Bel = N ec, and a plausibility measure Pl is a possibility measure, i.e., Pl = Pos.
• When focal elements are some disjoint intervals, the plausibility Pl and belief
Bel measures are both probability measures, i.e., Bel = P = Pl, for unions of
intervals.
Thus, all probability distributions and all possibility distributions may be interpreted
as mass functions. Hence, one may work in a common framework to treat information
of imprecise and random nature.
Below are presented the ways of creating the probability mass on the basis of the
probability distribution and the possibility distribution.
• Probability → Density function
Let X be a real random variable with a probability density p X . By discretizing X
into m disjoint intervals ([ai , ai+1 ])i=1,2,...,m being used to define focal elements,
the mass distribution (vi )i=1,2,...,m can be built as follows:

v([ai , ai+1 ]) = vi = P(X ∈ [ai , ai+1 ])

• Possibility → Belief function

Let Y be a possibilistic variable. The possibility distribution of Y is denoted
here by π, and π α denotes an α-cut of π. Focal elements for Y corresponding
to α-cuts are denoted by (π α j ) j=1,...,q with α0 = α1 = 1 > α2 > · · · αq >
αq+1 = 0. Finally, the mass distribution associated to (π α j ) j=1,...,q is denoted by
(v j = α j − α j+1 ) j=1,...,q .
Any pair of functions [Nec, Pos] or [Bel, Pl] can be interpreted as upper and
lower probabilities induced from specific probability families. As stated above, every
52 3 Fuzzy Random Variable and the Dempster-Shafer Theory of Evidence

possibility distribution π induces a pair of functions Nec and Pos. The probability
family is defined by

P(π) = {P, ∀A, N ec(A)  P(A)} = {P, ∀A, P(A)  N ec(A)}

In this case sup P∈P(π) P(A) = Pos(A) and inf P∈P(π) P(A) = N ec(A), and thus
P = Pos and P = N ec. Hence, upper F and lower F cumulative distribution
functions can be defined such that ∀x ∈  F(x)  F(x)  F(x) where:

F(x) = Pos(X ∈ [−∞, x]), (3.6)

F(x) = N ec(X ∈ [−∞, x]). (3.7)

Similarly a mass distribution v may encode probability family P(v) = {P, ∀A,
Bel(A)  P(A)} = {P, ∀A, P(A)  Pl(A)}. In this case P = Pl and P = Bel.
Hence, upper F and lower F can be defined as follows: ∀x

F(x) = Pl(X ∈ [−∞, x]), (3.8)

F(x) = Bel(X ∈ [−∞, x]). (3.9)

3.3 The Hybrid Data Propagation Method

This section describes a method for processing of hybrid data, i.e., data consisting of
both random and fuzzy variables. The method aims to determine the value of fˆ( X̂),
where X̂ = (X 1 , X 2 , . . . , X m ) is a vector of variables burdened with uncertainty. It
is assumed that there are k (k < m) random variables (X 1 , X 2 , . . . , X k ) and m − k
fuzzy variables X̃ k+1 , X̃ k+2 , . . . , X̃ m . Additionally, it is assumed that there may be
defined subsets X K of correlated variables X i ; X K = {X i | i ∈ K }, K ∈ K s . In such
case, K is the subset of correlated variable indices, and K s is the set of indices of the
selected subsets of correlated variables.
The proposed method (see Algorithm 1) for determining the value fˆ( X̂) combines
stochastic simulation with nonlinear programming. The latter is used to take into
account dependencies between fuzzy variables. The computational procedure per-
forms as follows. Values (x1 , x2 , . . . , xk ) of random variables are drawn using a pro-
cedure which accounts for correlation of the stochastic variables. The obtained values
and fuzzy variables X̃ k+1 , X̃ k+2 , . . . , X̃ m allow to determine fˆ(x1 , x2 , . . . , xk , X̃ k+1 ,
X̃ k+2 , . . . , X̃ m ) as a fuzzy number. This can be achieved using the concept of α-
cuts. Upper bound (sup) and lower bound (inf ) of an α-cut of a fuzzy number
fˆ(x1 , x2 , . . . , xk , X̃ k+1 , X̃ k+2 , . . . , X̃ m ) may be determined by solving the following
non-linear programming tasks.
3.3 The Hybrid Data Propagation Method 53

When searching for supremum, find:

f (x1 , x2 , . . . , xk , xk+1 , xk+2 , . . . , xm ) → max (3.10)

When searching for infimum, find:

f (x1 , x2 , . . . , xk , xk+1 , xk+2 , . . . , xm ) → min (3.11)

subject to the following constraints:

 
inf X iα  xi  sup X iα , i = k + 1, k + 2, . . . , m, (3.12)

xi  inf(a1i z )x z + inf(a2i z ) for i, z ∈ K ; i = z; i  k, z > k, K ∈ K s , (3.13)

xi  sup(a1i z )x z + sup(a2i z ) for i, z ∈ K ; i = z; i  k, z > k, K ∈ K s , (3.14)

xi  inf(a1i z )x z + inf(a2i z ) for i, z ∈ K ; i = z; i > k, z > k, K ∈ K s , (3.15)

xi  sup(a1i z )x z + sup(a2i z ) for i, z ∈ K ; i = z; i > k, z > k, K ∈ K s , (3.16)

The values a1i z , a2i z are the coefficients of interval regression equations deter-
mining a relation between variables X i and X z . Drawing values (x1 , x2 , . . . , xk ) and
...
determining fˆ(x1 , x2 , . . . , xk , X̃ k+1 , X̃ k+2 , . . . , X̃ m ) is repeated n times. The overall
... f f
procedure yields n fuzzy sets characterised by membership functions (μ1 , . . . , μ... n ).
ˆ
Thus, the value f ( X̂) is represented by a fuzzy random variable.

Algorithm 1 Hybrid Propagation Method

Input: Function f , vector of correlated variables X̂ = (X 1 , X 2 , . . . , X M )
Output: Random fuzzy set defining f ( X̂)
1: Define α0 , ℘, n̂ = 1
2: Generate a vector of correlated random values
(x1 , x2 , . . . , xk )
3: α = α0
4: Determine α-levels X̃ iα , i = k + 1, k + 2, . . . , m
5: Compute infimum (inf) and supremum (sup) of α-levels of a fuzzy number defining fˆ( X̂) by
solving the following two optimisation problems:
f (x1 , x2 , . . . , xk , xk+1 , xk+2 , . . . , xm ) → max
and f (x1 , x2 , . . . , xk , xk+1 , xk+2 , . . . , xm ) → min
under the problem constraints specified by inequalities (3.12)–(3.16)
6: α = α + ℘
7: If α  1 Then Go To Step 4
8: n = n + 1
...
9: If n  n Then Go To Step 2
f f
10: return the set of fuzzy numbers (μ1 , . . . , μ... n)
54 3 Fuzzy Random Variable and the Dempster-Shafer Theory of Evidence

References

1. Fèron, R. 1976. Ensembles aléatoires flous. Comptes Rendus de l’Académie des Sciences, Série
A (Sciences mathématiques) 282: 903–906.
2. Nahmias, S. 1978. Fuzzy variables. Fuzzy Sets and Systems 1(2): 97–110.
3. Stein, W.E., and K. Talati. 1978. Convex fuzzy random variables. Fuzzy Sets and Systems 6(3):
271–283.
4. Kwakernaak, H. 1978. Fuzzy random variables - I. definitions and theorems. Information.
Science 15(1): 1–29.
5. Kwakernaak, H. 1979. Fuzzy random variables - II. algorithms and examples for the discrete
case. Information. Science 17(3): 253–278.
6. Puri, M.L., and D. Ralescu. 1986. Fuzzy random variables. Journal of Mathematical Analysis
and Applications 114(2): 409–422.
7. Diamond, P., and P. Kloeden. 1994. Metric spaces of fuzzy sets: Theory and applications. World
Scientific.
8. Kruse, R. 1982. The strong law of large numbers for fuzzy random variables. Information
Science 28(3): 233–241.
9. Krätschmer, V. 2001. A unified approach to fuzzy random variables. Fuzzy Sets and Systems
123(1): 1–9.
10. Liu, Y.-K., and B. Liu. 2003. Fuzzy random variables: A scalar expected value operator. Fuzzy
Optimization and Decision Making 2(2): 143–160.
11. Liu, Y.-K., and B. Liu. 2003. Expected value operator of random fuzzy variable and random
fuzzy expected value models. International Journal of Uncertainty, Fuzziness and Knowledge-
Based Systems 11(2):195–215.
12. Liu, Y.-K., and B. Liu. 2002. Expected value of fuzzy variable and fuzzy expected value models.
IEEE Transactions on Fuzzy Systems 10(4): 445–450.
13. Shafer, G. 1976. A Mathematical Theory of Evidence. Princeton: Princeton University Press.
14. Dempster, A.P. 1967. Upper and lower probabilities induced by a multivalued mapping. The
Annals of Mathematical Statistics 38(2): 325–339.
15. Baudrit, C., D. Dubois, and D. Guyonet. 2006. Joint propagation and exploitation of proba-
bilistic and possibilistic information in risk assessment. IEEE Transaction on Fuzzy Systems
14(5): 593–607.
Chapter 4
Multi-attribute Decision Making Process
and Its Application

Abstract This chapter proposes the integrated fuzzy approach to solve Multi
Attribute Decision Problems. Fuzzy Analytical Hierarchy Process (FAHP) is used
to assign relative weights to criteria, and Technique for Order of Preference by Sim-
ilarity to Ideal Solution (TOPSIS) is employed to rank the alternatives. The use of
the proposed approach is illustrated using a real case from a steel industry.

Decisions regarding implementation of investment projects are crucial to the growth

and success of each company. The quality of these decisions impacts the long-term
effectiveness and market position of a company. However, financial and material
limitations cause that not all potential investment projects can be implemented. So,
almost every company sooner or later faces the problem of selecting a portfolio of
investment projects. This problem is particularly important nowadays, in the age of
omnipresent uncertainty in every business activity. Factors such as the increase in
variability of products prices, growing pressure on reduction of the cost of production,
increase of competitions, which can be observed in all branches of the economy, cause
that the choice of an effective portfolio becomes an increasingly complex decision
task. This motivates managers and decision makers (DMs) to use modern techniques
and tools for managing the allocation of a company’s capital, which in turn causes
that methods for selecting an effective portfolio continue to be an important research
area in the field of project management.
Selection of an effective portfolio should take into account not only financial
aspects but also environment, market, organisation, technology, human resources
and compliance with company’s objectives. This leads to a multi-attribute evaluation
of an investment. The main advantage of these approach is the ability to cover not
just the financial aspects of the attractiveness of the investment project. It is quite
obvious that the more parameters of investment projects subjected to analysis, the
more reliable the results of the selection process of the projects. Unfortunately, the
wider spectrum of criteria the higher the cost of analysis. Too many criteria may also
cause that interpretation of results of analysis is very difficult.
Multi-attribute evaluation of investments belongs to a wider class of problems
called Mutli-Attribute Decision Making (MADM) problems. A multi-attribute deci-
sion making (MADM) problem can be defined as:
56 4 Multi-attribute Decision Making Process and Its Application

C1 . . . Cn
A1 f 11 . . . f 1n
.. .. ..
. . .
Am f m1 . . . f mn

w = [w1 , . . . , wn ]

where A1 , . . . , An are possible alternatives, C1 , . . . , Cm are criteria, f i j is the rating

of the alternative Ai with respect to the criterion C j . The vector w is vector of weights,
where w j denotes the weight of the criterion C j [1]. The goal of the MADM is to
choose the best set of alternatives according to the defined criteria.
The MADM analysis requires that the selection be made among a predetermined,
limited number of decision alternatives described by multiple, and often conflicting,
criteria. The criteria are assigned weights relative to their importance, which then
help a decision maker to make possibly the best decision. So, from the decision
maker point of view, MADM analysis consists of methods that allow to combine
information coming from different sources with additional information coming from
a decision maker (e.g., preference) to rank available alternatives. The most important
features of the MADM analysis are the following:
• Limited number of decision alternatives,
• Finite set of attributes describing each alternative.
• Discrete number of preference points.
Despite its undeniable usefulness, there are many problems related to classical
MADM analysis, such as:
• Difficulties in the formulation of decision criteria.
• Competitiveness and inconsistency between criteria.
• Difficulties in the formulation of synthetic criterion based on sub-criteria.
• The presence of both quantitative and qualitative criteria.
• The uncertainty in parameter description.
The two latter problems are especially important. Usually, the values of decision
criteria cannot be evaluated exactly. Moreover, the values of qualitative criteria esti-
mated by different experts are difficult to compare.
The most common problems of the MAMD evaluation of projects in the presence
of uncertainty are the following:
• How to take into account uncertainty during the process of evaluation of the value
of criteria for each alternatives.
• How to establish appropriate fuzzy metrics for each qualitative criterion.
• How to appropriately propagate uncertainty through the multi-attribute decision
making process.
4.1 General Steps of a Multi-attribute Decision Making Process 57

4.1 General Steps of a Multi-attribute Decision Making

Process

A typical multi-attribute decision making problem consists of the following three

phases: identification, choice, ranking.
In the identification phase, a decision maker defines the decision context, i.e.,
alternatives and decision criteria. Alternatives represent available decisions or
projects, whereas decision criteria measure how well each alternatives meet objec-
tives. Each criterion must be measurable, or at least it must be possible to assess
how much an alternative meets a criterion. So, an appropriate evaluation scale must
be defined for each criterion. In this phase, projects are selected on the basis of
the threshold criteria, which are determined by decision-makers. The value of each
criterion for each alternative must strictly meet threshold criteria. Typical threshold
criteria include financial criteria (e.g., NPV > 0, IRR > threshold(IRR)), risk and
strategy. Only those projects that meet threshold criteria pass to the next stage of the
analysis.
In the choice phase all criteria are first arranged in a hierarchical order. Then,
relative weights are assigned to all criteria. When the number of criteria is small
enough, the weights can be assigned by experts. However, as the number of criteria
included in evaluation process increases, the assignment of weights becomes a diffi-
cult task. In such cases, usually the Analytical Hierarchy Process (AHP) is used. In
AHP, decision makers systematically compare criteria, two at a time, and determine
which criterion is more important using some comparison scale (see Table 4.1). The
comparison continues until all criteria have been compared. The comparison results
are stored in the form of a matrix, which is called a pairwise comparison matrix.
Then, the following steps are performed:
1. Calculate a priority vector in order to weight the elements of the comparison
matrix.
2. Calculate global priorities by aggregating all local priorities using a simple
weighted sum.
3. Use the eigenvalue in order to assess the strength of the consistency ratio of the
comparative matrix and determine whether to accept the information. If a com-
parison matrix is not consistent, its elements should be adjusted and a consistency
test should be carried out until the consistency ratio attains an acceptable value.

Table 4.1 Likert scale for a pairwise comparison

Definition Intensity of importance Fuzzy intensity of importance
Equally important 1 1
Moderately more important 3 (2, 3, 4)
Strongly more important 5 (4, 5, 6)
Very strongly more important 7 (6, 7, 8)
Extremely more important 9 (8, 9, 10)
Intermediate values 2, 4, 6, 8
58 4 Multi-attribute Decision Making Process and Its Application

In the ranking phase, synthetic values of importance are computed for each alter-
native. There are many methods to obtain those values. Probably, the most popular
is the Technique for Order of Preference by Similarity to Ideal Solution (TOPSIS)
method, which is based on the assumption that the chosen alternative should have
the shortest distance from the positive-ideal solution and the longest distance from
the negative-ideal solution, where the positive ideal solution is the solution that max-
imises the benefit criteria and minimises the cost criteria, whereas the negative ideal
solution is the solution that maximises the cost criteria and minimises the benefit
criteria.

4.2 Uncertainty in MADM

Application of the approach presented above to real-life problems faces some diffi-
culties. These are among others:
• The presence of quantitative and qualitative criteria.
• Description of some of evaluations using linguistic variables.
• Problems with comparison of criteria.
• Multiple and contradictory goals.
• Dependent projects.
• Uncertainty in data regarding specific criteria.
• A large number of feasible projects.
• Organisational requirements and controls.
To deal with those difficulties, the fuzzy approach is usually used instead of
the crisp one. Probably the most important reasons for using the fuzzy approach
to represent criteria and values of alternatives in MADM is the consistency with
available information. In most problems of multi-attribute evaluation of investment
projects, the subjective opinions on selected parameters are usually used. Neglecting
imprecision and subjectivity of decision makers in pairwise comparison matrix may
lead to errors in AHP methods.
The second problem is a determination of future values of the parameters deter-
mining the profitability of investment projects. However owing to the availability
and uncertainty of information, it is very difficult to obtain the exact assessment
data such as investment cost, gross income, expenses, depreciation, salvage value,
interest rate, flexibility, productivity, quality etc. All of these factors, no matter tan-
gible or intangible, are generally difficult to be quantified. In this situation most of
decision-maker tend to give assessment based on their knowledge, past experience
and subjective judgements. Linguistic terms such as “around 10 %”, “approximately
between $300 000 and 450 000”, “about $80 000”, “very low”, “medium”, “high”, a
frequently used to convey their estimations [2]. To deal with the vagueness of human
thought fuzzy set theory can play significant role in this kind of decision making
environment [3].
4.2 Uncertainty in MADM 59

Fig. 4.1 The schematic diagram of investment selection

Hence a fuzzy multi-attribute decision making method is used to integrate various

linguistic assessments and weights to determine the best investment project selection.
In Fig. 4.1 method of fuzzy multi-attribute evaluation of investment is presented.
60 4 Multi-attribute Decision Making Process and Its Application

Chan et al. [3] presented a proposal of methods for hierarchical multi-criteria

evaluation of investment projects.
The procedure for evaluation of investment projects in this case is as follows:
Stage 1. Form a committee of decision-maker who come from different manage-
rial levels of the company. Identify various available m investment projects
(A1 , A2 , . . . , Am ) under each of the k criteria (C1 , C2 , . . . , Ck ),
Stage 2. Arrange criteria hierarchically and classify into subjective and objective
criteria,
Stage 3. Choose proper linguistic scale and ask decision-makers to give their judge-
ment by either directly assigning weight using pairwise comparisons,
Stage 4. Calculate fuzzy ratings for alternatives.
Stage 5. Divide criteria into subjective and objective. Subjective criteria are char-
acterised by linguistic assessments. On the other hand, objective criterion, i.e.,
economic, is evaluated in monetary terms.
Stage 6. Convert the linguistic variables into fuzzy number scale and objective vari-
ables into possibilities distribution.
Stage 7. Calculate fuzzy performance index for each project.
Stage 8. Construct rating of the projects.
At each stage of the above algorithm different MADM methods may be used. The
two most populars methods for assigning fuzzy weights and ranking projects with
fuzzy weights are, respectively, fuzzy AHP and fuzzy TOPSIS.

4.3 Fuzzy Analytical Hierarchical Process

One of the most well-known MADM methods is the Analytical Hierarchy Process
(AHP) developed by Saaty [4]. The AHP integrates different measures into a single
overall score for ranking alternative decisions. It is based on a pairwise comparison
of experts’ judgements (Table 4.1). However, some researchers claim that it is better
to make that scale fuzzy due to imprecise of expert judgements. It’s more convenient
for decision makers to give an interval judgement instead of fixed value. Due to
cognitive biases, decisions may be deviated from a standard of rationality or a good
judgement. In [5] are described difficulties in translating linguistic terms (e.g., “very
strongly more important”) into numbers.
There are thousands of articles that apply fuzzy AHP (FAHP) to a wide range of
topics. The earliest work in fuzzy AHP appeared in van Larhooven and Pedrycz [6]
and Buckley [7]. In these articles, comparison ratios are described by triangular and
trapezoidal fuzzy numbers respectively. Chang [8] introduced a new approach for
handling fuzzy AHP with use of the extended analysis method for making pairwise
comparison. Kahraman et al. [9] use a fuzzy objective and subjective method obtain-
ing the weights from AHP and make a fuzzy weighted evaluation. In 00’s occurs
thousands of articles about fuzzy numbers. Most important articles on theories and
4.3 Fuzzy Analytical Hierarchical Process 61

application of fuzzy AHP has more than 1000 citations [6–8]. Thus, fuzzy AHP can
be considered as a well-established theory.

Algorithm 3 Buckley’s Fuzzy AHP

1: Decision maker build the pairwise comparison matrix. Each element of comparison matrix is a
trapezoidal fuzzy number t˜i j = (a, b, c, d),
2: For each row of pairwise comparison matrix the g index is calculated as a geometric mean of
 1/n
n
the row: gi = t˜i j
j=1
3: The weight of i criterion is obtained using arithmetic mean of gi
gi
wi = n
i=1 gi

Below, the most popular fuzzy approaches to AHP (Chang [8] and Buckley [7])
are presented. The Buckley’s approach is as follow:

Algorithm 4 Chang’s Fuzzy AHP approach

Input: The pairwise comparison matrix T = [ti j ] consists of I rows and I columns. Each row and
column represents a single criterion. The matrix is built by a decision maker.
1: Convert each element of the comparison matrix into a triangular fuzzy number
t˜i j = (a, b, c).
2: Calculate fuzzy synthetic value: n
t˜i j
w j = n i=1 n .
i=1
˜
i=1 ti j

3: For each pair of synthetic values calculate its degree of possibility:

 
V (wi ≥ wk ) = sup(min µwi (x) , µwk (y) .

For triangular fuzzy numbers this expression is equivalent to

⎧
⎨ 1, if bi > bk ,
V (wi ≥ wk ) = 0, if ck ≥ ai ,
⎩ ai −u k
(bk −ck )−(bi −ai ) , otherwise.

4: Weight of criterion i
ai = min (V (wi ≥ wk ))
k

5: Normalise the weights of criteria. AHP impact score of each criterion is measured by w index.
62 4 Multi-attribute Decision Making Process and Its Application

Table 4.2 Crisp evaluation matrix

Coffee Wine Tea Beer Sodas Milk Water
Coffee 1 9 5 2 1 1 0.5
Wine 0.11 1 0.5 0.11 0.11 0.11 0.11
Tea 0.2 2 1 0.33 0.25 0.33 0.11
Beer 0.5 9 3 1 0.5 1 0.5
Sodas 1 9 4 2 1 2 0.5
Milk 1 9 3 1 0.5 1 0.33
Water 2 9 9 3 2 3 1

4.3.1 Example I—Comparison of AHP Methods

An exemplary problem of relative consumption of drinks in the United States, which

was considered in [10], will be used to compare fuzzy and crisp AHP methods.
Tables 4.2 and 4.3 present crisp and fuzzy evaluation matrix, respectively.
The above fuzzy evaluation matrix was obtained using pairwise comparison scale
from Table 4.1. The weights of criteria was obtained using classical crisp algorithm,
Buckley algorithms and Chang approach respectively. Buckley fuzzy weight are also
defuzzified using formula:

aj + bj + cj + dj
wj = (4.1)
4
The comparison is presented in Table 4.4. In column “real consumption” the actual
consumption (from Statistical Abstract of the United States) are presented as a ref-
erence point.
The most important observations are as follows:
• The smallest difference is between real consumption and crisp approach, although
it doesn’t mean that crisp methods are the best. Fuzzy weight obtained from Buck-
ley’s approach represents also ignorance.
• Chang approach flatten weights of criteria.

4.3.2 Fuzzy TOPSIS

The second important problem in project portfolio selection is ranking of the projects.
One of the most popular methods is fuzzy TOPSIS. The Technique for Order of Pref-
erence by Similarity to Ideal Solution (TOPSIS) is a multi-criteria decision analysis
method. It was developed by Hwang and Yoon [11] with further developments by
Yoon [12] and Hwang et al. [13].
Like it was shown during AHP presentation, there are many approaches to Fuzzy
TOPSIS methods. Below, the method proposed by Chen and Hwang [14] is presented.
Table 4.3 Fuzzy evaluation matrix
Coffee Wine Tea Beer Sodas Milk Water
Coffee (1, 1, 1, 1) (8, 9, 9, 10) (4, 5, 5, 6) (1, 2, 2, 3) (1, 1, 1, 1) (1, 1, 1, 1) (0.33, 0.5, 0.5, 1)
4.3 Fuzzy Analytical Hierarchical Process

Wine (0.1, 0.11, 0.11, 0.125) (1, 1, 1, 1) (0.33, 0.5, 0.5, 1) (0.1, 0.11, 0.11, 0.125) (0.1, 0.11, 0.11, 0.125) (0.1, 0.11, 0.11, 0.125)
Tea (0.167, 0.2, 0.2, 0.25) (1, 2, 2, 3) (1, 1, 1, 1) (0.25, 0.33, 0.33, 0.5) (0.2, 0.25, 0.25, 0.33) (0.25, 0.33, 0.33, 0.5) (0.1, 0.11, 0.11, 0.125)
Beer (0.33, 0.5, 0.5, 1) (8, 9, 9, 10) (2, 3, 3, 4) (1, 1, 1, 1) (0.33, 0.5, 0.5, 1) (1, 1, 1, 1) (0.33, 0.5, 0.5, 1)
Sodas (1, 1, 1, 1) (8, 9, 9, 10) (3, 4, 4, 5) (1, 2, 2, 3) (1, 1, 1, 1) (1, 2, 2, 3) (0.33, 0.5, 0.5, 1)
Milk (1, 1, 1, 1) (8, 9, 9, 10) (2, 3, 3, 4) (1, 1, 1, 1) (0.33, 0.5, 0.5, 1) (1, 1, 1, 1) (0.33, 0.5, 0.5, 1)
Water (1, 2, 2, 3) (8, 9, 9, 10) (8, 9, 9, 10) (2, 3, 3, 4) (1, 2, 2, 3) (2, 3, 3, 4) (1, 1, 1, 1)
63
64 4 Multi-attribute Decision Making Process and Its Application

Algorithm 5 Fuzzy TOPSIS

Input: The input variable are decision matrix D and weights of criteria. Each row of D represent
j-th alternative and each column i-th criterion.
⎡ ⎤
x11 · · ·
⎢ ⎥
D = ⎣ ... . . . ... ⎦
· · · xmn

The value of xi j is represented by fuzzy number.

1: Normalisation using linear scale. Each element of the matrix D is transformed using linear
scale:   xi j
Dnor m = ri j mxn = 
i=1...m x i j
The normalisation is performed in order to eliminate anomalies with different measurement
units and scales.
2: Calculation of normalised decision matrix. For each element of Dnor m calculate the weighted
normalised decision matrix    
vi j mxn = w j ri j mxn ,
where w j is a weight of j-criterion obtained from FAHP methods.
3: Positive Ideal Solution/Negative Ideal Solution: Determine Positive Ideal Solution (PIS) and
Negative Ideal Solution (NIS). PIS and NIS are defined as

a P I S = [v1max , ..., vnmax ]

a N I S = [v1min , . . . , vnmin ]
where v max
j = max vi j and
i
v min
j = min vi j
i
To obtain maximal value of fuzzy data, some ranking procedures should been made. This example
shows different methods of ranking fuzzy numbers approaches.
4: Separation Measure - Obtain Separation Measures from PIS and NIS. Classically, this is
a Manhattan or Euclidean measure. In order to simply the approach the Manhattan Measure is
used
n 
 
 
SiP I S = vi j − v max
j 
i=1
n 
 
 
SiN I S = vi j − v min
j 
i=1
Difference between fuzzy numbers are described in Chap. 1.
5: Closeness – Compute Relative Closeness to Ideal

SiN I S
Ci =
SiP I S + SiN I S
4.3 Fuzzy Analytical Hierarchical Process 65

Table 4.4 Importance of the drinks: fuzzy weights, crisp weights and real consumption
Buckley fuzzy weights Buckley after Chang Crisp Real
deffuzification approach consumption
Water (0.18, 0.325, 0.325, 0.546) 0.344 0.200 0.323 0.240
Sodas (0.109, 0.189, 0.189, 0.332) 0.205 0.000 0.188 0.267
Coffee (0.114, 0.178, 0.178, 0.294) 0.191 0.000 0.177 0.133
Milk (0.089, 0.13, 0.13, 0.204) 0.138 0.835 0.129 0.129
Beer (0.076, 0.118, 0.118, 0.204) 0.129 0.235 0.125 0.173
Tea (0.024, 0.039, 0.039, 0.067) 0.042 0.072 0.039 0.040
Wine (0.014, 0.021, 0.021, 0.032) 0.022 0.410 0.020 0.014

As a result a vector C is obtained. The elements of the vector C represent the

indexes for each alternative. In the next step alternatives are ranked in descending
order. This is the final ranking for projects.

4.4 Example—Investment Strategy of the Steel Company

A board of directors wants to develop investment strategy of a steel company. They

take into consideration 10 potential investment and 23 attributes.

4.4.1 Calculation of Fuzzy Ratings

To solve such problem the analyst should construct hierarchy of the criteria. The
tree of the criteria is presented on Fig. 4.2. Tree consists of three layers—criteria,
sub-criteria and attribute. Factors belong to criteria layer are marked by Cx, sub-
criteria layer Cx.x, and attribute layer Cx.x.x. There are five main group of criteria
(first layers). There are five criteria—financial, market, technology and environment,
staff and compliance with the company’s strategic objective. Each of them is divided
into subcriteria. The following objective subcriteria were used:

• NPV—is defined as a sum of discounted net cash flows appearing in consecutive

years of the economic life of an investment project.
• IRR—is the annualised effective compounded return rate that makes the NPV from
a particular investment equal to zero. It can also be defined as the discount rate at
which the present value of all future cash flow is equal to the initial investment or
in other words the rate at which an investment breaks even.
• Payback period—refers to the period of time required to recoup the funds
expended in an investment.
66 4 Multi-attribute Decision Making Process and Its Application

Fig. 4.2 Hierarchy of factors

Table 4.5 Pair-wise comparison matrix of criteria (level 1)

C1 C2 C3 C4 C5
Financial criteria C1 1 5 9 9 9
Market C2 0.2 1 3 3 3
Technological and Environmental C3 0.11 0.33 1 1 1
Staff C4 0.11 0.33 1 1 1
Compatibility with the strategical objective C5 0.11 0.33 1 1 1

After determination of criteria hierarchy, the pairwise comparison matrix for each
layer was prepared. The exemplary matrix for first level of criteria is presented in
Table 4.5.
In this example six pairwise comparison matrix were prepared—one for criteria
level, five for sub criteria level (with exception for C5.x), and one for attribute level.
4.4 Example—Investment Strategy of the Steel Company 67

Table 4.6 Summary of Consistency ratio

consistency check
Criteria 0.04
Subcriteria C1.x 0
Subcriteria C2.x 0.07
Subcriteria C3.x N/A
Subcriteria C4.x 0.05
Attributes 0.03

Simultaneously, definitions of scales for each decision factors were prepared.

There are two type of scales—numerical and linguistic. Only three factors—NPV,
IRR and Payback period was described by fuzzy numbers, the rest of subcriteria was
subjective. The value of linguistic attributes are presented as a leaf in grey boxes, the
rest of factors were numerical and described by fuzzy numbers.
Building investment strategy has started from calculation of weight factors using
AHP method. Each expert prepared pairwise comparison matrices in terms of fuzzy
numbers. Then the consistency on each of the matrices was verified to ensure that all
consistency requirements was satisfied. Consistency ratio of each matrix is presented
in Table 4.6.
Next, all comparisons are converted into fuzzy scale and weights are calculated
using Buckley’s approach. Values of the weight of criteria are presented in Table 4.7.

Table 4.7 Importance weights of individual requirements

Weight Weight
C1 (0.429, 0.633, 0.917) C3.1 (0.039, 0.056, 0.124)
C2 (0.073, 0.175, 0.372) C3.2 (0.006, 0.008, 0.018)
C3 (0.051, 0.064, 0.122) C4.1 (0.003, 0.007, 0.039)
C4 (0.051, 0.064, 0.122) C4.2 (0.013, 0.036, 0.145)
C5 (0.051, 0.064, 0.122) C4.3 (0.008, 0.019, 0.085)
C1.1 (0.181, 0.292, 0.471) C4.4 (0.001, 0.003, 0.013)
C1.2 (0.181, 0.292, 0.471) C5.1 (0.051, 0.064, 0.122)
C1.3 (0.025, 0.042, 0.072) C2.3.1 (0.002, 0.027, 0.313)
C2.1 (0.006, 0.05, 0.307) C2.3.2 (0.001, 0.02, 0.264)
C2.2 (0.018, 0.061, 0.204) C2.3.3 (0.002, 0.015, 0.127)
C2.3 (0.011, 0.063, 0.31)
68 4 Multi-attribute Decision Making Process and Its Application

4.4.2 Computation of Performance Index of Criteria

To compute performance index the fuzzy evaluation matrix was created. Boards of
directors took into consideration 10 alternatives:

• P1—Modernisation of the heavy section mill,

• P2—Increase of the capacity of the hot rolling mill,
• P3—Construction of the cold rolling mill with the capacity of 1000 thousand
t/year,
• P4—Construction of the cold rolling mill with the capacity of 1500 thousand
t/year,
• P5—Construction of the hot dip galvanising line with the capacity: 300 thousand
t/year,
• P6—Construction of the hot dip galvanising line with the capacity: 400 thousand
t/year,
• P7—Construction of the organic coating line with the capacity of 200 thousand
t/year,
• P8—Construction of the organic coating line with the capacity of 300 thousand
t/year,
• P9—Construction of the tinning plant with the capacity of 100 thousand t/year,
• P10—Construction of the wire drawing plant.

For each investment project the objective criteria were characterised by fuzzy
numbers which resulted from the hybrid simulation described in Sect. 1.3.1. The
levels of subjective criteria are specified by experts. Values of subjective criteria were
translated into fuzzy numbers using procedure as follows. In the presented example,
there are two kinds of subjective attributes—some of them describe patterns, and
some of them judgements. For example, market size criterion C2.1 and prospects
for market growth criterion C2.2 belong to the first group. They describe the belief
of decision maker that market for alternatives will behave in accordance with some
pattern. For example, pattern had been stable means the dynamic of the market
growth which may be described by the fuzzy number (−1.02, 0, 1.02). The second
group that is subjective criteria represents judgements of experts. Therefore, they are
treated as ordinal fuzzy variables. Description of project alternatives is presented in
Fig. 4.3.
Exemplary translation of subjective criteria for one factor is presented in Table 4.8
and coding of one alternative is presented in Table 4.9.
The translation table is used to create the so-called performance matrix. Rows in
the latter matrix describe alternatives and column correspond to criteria. The entries,
described by fuzzy numbers, asses how well each option performs with respect to
each of the criteria. Then, using FTOPSIS, the ranking of criteria was constructed
(Table 4.10).
4.4 Example—Investment Strategy of the Steel Company 69

Fig. 4.3 Description of the projects alternatives

 70

Table 4.8 Coding of exemplary alternative

Projects P1 value of attributes P1 numerical code of attributes
Capital investment 000’s PLN (135, 150, 165) (216, 240, 264)
C1 C1.1 (−66083.9, 14376.1, 127533.0, 225122.1) (−66083.9, 14376.1, 127533.0, 225122.1)
C1.2 (−5.1, 4.5, 9.3, 12.1) (−5.1, 4.5, 9.3, 12.1)
C1.3 2.1 2.1
C2 C2.1 Middle (1100, 1300, 1400, 1600)
C2.2 Stability (−1.02, 1.0, 1.02)
C2.3.1 Average (0, 0.1, 1)
C.2.3.2 Average (0.95, 1, 1.05)
C.2.3.3 External (0, 0.5, 1)
C3 C3.1 Widely used (0, 0.1, 0.2)
C3.2 Neutral (0, 0.5, 1)
C4 C4.1 Without growth (0, 0.1, 1)
C4.2 No impact (0.5, 1, 1)
C4.3 Positive (0, 0, 0.5)
C4.4 Available (0, 0.33, 0.66)
C5 C5.1 One (0, 0.2, 0.4)
4 Multi-attribute Decision Making Process and Its Application
References 71

Table 4.9 Exemplary Linguistic value Fuzzy value

transformation of subjective
criteria Not important (0, 0, 0.33)
Neutral (0, 0.33, 0.66)
Important (0.33, 0.66, 1)
Very important (0.66, 1, 1)

Table 4.10 Final ranking of Project Rank

projects
P9 0.7154
P10 0.7101
P1 0.7095
P4 0.7011
P3 0.6817
P7 0.6789
P8 0.6782
P5 0.6770
P6 0.6714
P2 0.6615

References

1. Yu, P.-L. 2013. Multiple-criteria decision making: concepts, techniques, and extensions, vol.
30. Springer Science and Business Media.
2. Ward T.L. 1985. Discounted fuzzy cash flow analysis. In Fall industrial engineering, 476–481.
3. Chan, F.T.S., M.H. Chan, and N.K.H. Tang. 2000. Evaluation methodologies for technology
selection. Journal of Materials Processing Technology 107(1–3): 330–337.
4. Saaty, T.L. 1980. The analytic hierarchy process: planning, priority setting resource allocation.,
Advanced Book Program New YorK: McGraw-Hill.
5. Bocklisch, F., S.F. Bocklisch, and J.F. Krems. 2010. How to translate words into numbers?
Fuzzy approach for the numerical translation of verbal probabilities, vol. 6178., Lecture Notes
in Computer Science Berlin: Springer.
6. van Laarhoven, P.J.M., and W. Pedrycz. 1983. A fuzzy extension of saaty’s priority theory.
Fuzzy Sets and Systems 11(1–3): 199–227.
7. Buckley, J.J. 1985. Fuzzy hierarchical analysis. Fuzzy Sets and Systems 17(3): 233–247.
8. Chang, D.-Y. 1996. Applications of the extent analysis method on fuzzy AHP. European
Journal of Operational Research 95(3): 649–655.
9. Kahraman, C., D. Ruan, and E. Tolga. 2002. Capital budgeting techniques using discounted
fuzzy versus probabilistic cash flows. Information Science 142(1–4): 57–76.
10. Saaty, T.L., and L.T. Tran. 2007. On the invalidity of fuzzifying numerical judgments in the
analytic hierarchy process. Mathematical and Computer Modelling 46(7–8): 962–975.
11. Hwang, C.-L., and K. Yoon. 1981. Multiple attribute decision making: methods and appli-
cations a state-of-the-art survey, vol. 186., Lecture Notes in Economics and Mathematical
Systems Berlin: Springer.
12. Yoon, K. 1987. A reconciliation among discrete compromise solutions. The Journal of Oper-
ational Research Society 38(3): 277–286.
72 4 Multi-attribute Decision Making Process and Its Application

13. Hwang, C.-L., Y.-J. Lai, and T.-Y. Liu. 1993. A new approach for multiple objective decision
makings. Computers and Operational Research 20(8): 889–899.
14. Chen, S.-J.J., and C.L. Hwang. 1992. Fuzzy multiple attribute decision making: methods and
applications, vol. 375., Lecture Notes in Economics and Mathematical Systems New York:
Springer.
Chapter 5
Risk Assessment in the Presence
of Uncertainty

Abstract This chapter is devoted to a method which is able to process hybrid

data, i.e., to jointly handle both randomness and imprecision. Random variables
are described by probability distributions and imprecise values are modelled using
possibility distributions. The main advantage of the proposed method is that it takes
into account the dependencies between economic parameters.

At present, methods for risk assessment constitute a fundamental tool supporting

decision-making processes. The risk associated with decisions arises from the fact
that economic parameters are usually burdened with uncertainty.

5.1 Description of Uncertainty in the Economic

Risk Assessment

Risk and uncertainty are defined in various ways in the world literature. The definition
that will be adopted in this book says that risk results from actions taken by humans
and depends functionally on uncertainty, whereas uncertainty concerns objective
states of nature, that is the inability to predict and accurately identify future business.
The uncertainty can be connected with external (e.g., income tax rates, raw material
and energy prices, forecasted market size, interest rates, exchange rates etc.) and
internal (e.g., product prices, material consumption indicators, labour consumption
indicators etc.) conditions of company’s activity. Therefore, the economic risk is
most frequently determined as a possibility of occurrence of unfavourable values
of the measure of the efficiency of the business activity. It is also identified with
variableness of the measure of the efficiency of the business activity [1].
Quantification of risk is one of the most difficult tasks in economic risk man-
agement. A chief problem in this phenomena is not only to develop methods for
estimation of the economic risk, but also to improve methods for data gathering and
processing for a formal description of uncertainty of parameters of the risk calculus.
An adequate description of uncertainty of those parameters has a decisive meaning
74 5 Risk Assessment in the Presence of Uncertainty

in estimation of the risk. It is a condition of effective application of quantification

methods of economic risk assessment in practice.
For many years, the only tool that allowed to express uncertainty in mathematical
language was probability calculus. In fact, it still remains the most common tool
used in practice and prevails in the literature concerning the economic risk. The
probabilistic approach to risk analysis employs stochastic simulation to evaluate
the risk. However, high cost of data preparation and difficulties with determining
probability distributions of economic parameters significantly limit the usage of this
approach. Furthermore the last decades have shown that the number and complexity
of dependencies both inside and outside a company makes it difficult to use the
probability theory to represent all kinds of the uncertainty appearing in case of the
economic risk assessment. That is why, apart from quantitative methods, qualitative
methods are also utilised to predict the values of economic parameters and to assess
the risk. Experts’ opinions and subjective probability distributions are then used.
Qualitative methods are more important, since in numerous decision-making sit-
uations uncertainty of economic parameters is not probabilistic in nature, but rather
results from insufficient or vague information and is epistemologically indeterminate
[2]. Sometimes, as pointed by Gupta [3], it happens that uncertainty is probabilistic,
but the available information is rather fuzzy. In practice, quite often it is not possible
to determine probability distribution, because of no (and there is no option to get)
sufficient volume of data facilitating execution of statistical tests. On the other hand,
assumption of “no data available at all” is also not true. In general, there is always
some information available. These could be estimates of unknown values made by
experts. A good example is the problem of assessment of profitability and risk of
project investments. The main difficulty in this problem stems from the uniqueness
of each investment project and the time interval between the moment of studies on
a project and its realisation and exploitation. Because of the uniqueness, it is usually
difficult to predict all possible values that the parameters of efficiency calculus can
take and to determine the probability of their realisation [4]. Pluta and Jajuga [5] say
that when assessing the risks attached to investment projects, usually only experts’
opinions and subjective probability distributions of the possible values of parame-
ters can be used. The estimation of net present value (NPV ) expected values from
historical data is much more difficult because of the specificity and uniqueness of
investment projects. Namely, it is not possible to obtain perfectly reliable information
regarding similar past projects. Each company is unique in its business model, invest-
ment needs, production capabilities, liaisons with the environment. Each company
goes through different stages of development during its lifecycle.
The estimate of subjective probability is based on the experience of a person which
determines the probabilities of occurrence of individual events. These probabilities
represent the level of conviction of an expert that the event will occur [4]. Dittmann
[6] interprets the subjective probability assigned to experts’ forecasts as a chance that
the predicted value is equal approximately the actual value of the analysed variable.
Subjective probability can also be interpreted as a possibility of occurrence of an
event [4, 6]. The level of subjectivity depends on the way of estimation and the
knowledge on other similar event. In practice, an arbitrary probability distribution
between minimal and maximal estimation is usually adopted.
5.1 Description of Uncertainty in the Economic Risk Assessment 75

The usage of subjective probability distributions in the economic risk assessment

causes many problems associated with estimation of those distributions. Choobineh
and Behrens [1] and Kuchta [4, 7] point out these difficulties. Kuchta says that some-
times a decision-maker does not know how to answer the question on the probability
of the unique, unrepeatable event. The question about frequency has not much sense.
Decision-maker can have, however, a point of view on the degree of possibility of
occurrence of respective values [4]. Moreover, the subjective probability distribution
must have the same properties as any probability distribution. For example, the sum
of probabilities of all elementary events must sum up to 1 and the probability of
the simultaneous occurrences of two independent events is the product of the prob-
abilities of each event. It is extremely difficult to maintain these properties in expert
judgements about subjective probability of future values [4]. This problem can be
partially solved by modelling uncertainties using fuzzy numbers. When uncertainty
is described using fuzzy numbers, a decision-maker can give arbitrary values of
possibility degrees according to own feelings. Fuzzy approach does not impose the
form of expression of subjective opinions as much as probabilistic approach does
[4]. Mohammed and McCowan [8] argue that for most practitioners triangular and
trapezoidal fuzzy numbers are much easier to understand and to apply than prob-
ability distributions. People hardly think in probabilistic terms, fuzzy sets notation
or linguistic description of uncertainty seems to be more natural and much closer to
human thinking. The construction of a triangular fuzzy number based on the best, the
worst and average values is closer to the possibility theory than to the probability the-
ory [4]. Moreover, many authors questions the legitimacy of modelling the absolute
lack of knowledge about selected parameter using uniform probability distribution
[2, 9].
In most of the existing approaches, different ways of uncertainty representation
are usually unified in a single modelling framework. In order to perform the unifi-
cation, it is necessary to be able to transform one form of uncertainty into another.
Obviously, such transformation is not without problems. For example, transforma-
tion of a probability distribution into a possibility distribution causes the loss of
information, whereas the opposite one requires additional information to be intro-
duced. This leads eventually to systematic errors in risk assessment, i.e., overesti-
mation or underestimation of the risk, depending on the direction of transformation.
The most appropriate approach to risk assessment is to develop and use methods
which allow hybrid representation of uncertainty, i.e., expressed by probability dis-
tributions, fuzzy numbers to be processed according to their nature and only finally
combine them into a synthetic easy-to-interpret risk measure. The research of the
subject [2, 9, 10] shows that models with hybrid data can be successfully used to
support decision-making in economic business.
Taking into consideration the above mentioned problems, most of the real-world
evaluations of risk contain a mixture of quantitative and qualitative data. Two meth-
ods of description of the uncertainty of the economic calculus parameters (probabil-
ity distribution, fuzzy numbers) are used usually as alternatives. The most common
situation in practice is when for some parameters it is possible to determine probabil-
ity distributions, while for some, information is available in form of fuzzy numbers.
In case of economic calculus, data which is available, is usually heterogeneous,
76 5 Risk Assessment in the Presence of Uncertainty

uncertain and imprecise, and it is usually coming from various sources. These are
both statistical data as well as subjective assessments of phenomena made by experts
[2, 3, 9–15].
Ward [16] was the first to utilise fuzzy numbers in financial analysis. Ward pre-
sented cash flows by means of trapezoidal fuzzy numbers. Buckley [17] uses fuzzy
numbers for calculation of net present values of investment projects. Calzi [18] pre-
sented principles of expanding financial mathematics for fuzzy numbers. Choobineh
and Behrens [1] present the use of possibility distributions in economic analysis. Chiu
and Park [11] apply fuzzy numbers in the calculation of the efficiency of investment
projects. The authors introduce the methods of the choice of one project, from the
set of mutually exclusive projects. Esogbue and Hearnes [19] utilise fuzzy numbers
in problems concerning the replacement of fixed assets. Their aim was to describe
the economic life-cycle of fixed assets. Kahraman et al. [20] present methods calcu-
lating selected investment project effectiveness ratios on the assumption that certain
parameters are presented in the form of fuzzy numbers. Kuchta presents the use of
fuzzy numbers in capital budgeting [7]. In [13] the results of the evaluation of the
profitability and the risk of investment projects in case when the uncertainty of para-
meters is presented in the form of probability distributions and fuzzy numbers are
compared. Rȩbiasz [12, 21–23] presents the usage of fuzzy numbers and probability
distribution for the evaluation of projects and selection of the most profitable project
from the steel industry.
Opinions of authors on usefulness of fuzzy and probabilistic approach in decision-
making analysis vary. Majority of authors argue that fuzzy and probabilistic
approaches are supplementary to each other, and in each case it must be decided
which approach will be more adequate. The selection of an appropriate approach
should be conditioned mostly on the degree of subjectivity of the available informa-
tion. On the other hand, Gupta [3] and Smets [24] claim that in decision-making,
probabilistic description of uncertainty is more effective than description using pos-
sibility distribution. They say, moreover, that a decision-maker is not interested in
“what is possible” but in “what is probable”. Thus, they suggest purposefulness of
transformation of possibility distribution into probability distribution. Kuchta [4], in
turn, says that selection of the method of uncertainty representation depends mainly
on the experience and habits of the decision-maker. Choobineh and Behrens [1]
claim that maintaining the probabilistic approach stems more from tradition then
from conscious selection.
In practice, the most common situation is when for some parameters it is possible
to determine probability distributions, while for others information is available in
form of fuzzy numbers. It must be underlined that probabilistic approach and fuzzy
approach are not contradictory and can be applied simultaneously. The probability
theory and the possibility theory emphasize different aspects of uncertainty. The
probability theory offers quantitative model of randomness, whereas the possibility
theory describes qualitative model of incomplete knowledge [2–4, 25, 26]. Baudrit
et al. [2] argue that randomness and imprecise or missing information are two rea-
sons of uncertainty, which have an impact on risk. Thus, it is necessary to include
these two approaches to description of uncertainty in the process of risk assessment.
The usage of both probability and possibility distributions allows to reflect more
5.1 Description of Uncertainty in the Economic Risk Assessment 77

properly the knowledge on economic parameters. However, for the time being there
are few studies which are devoted to the use of hybrid representation of uncertainty
[2, 26]. As suggested by Ferson and Ginzburg [14] distinct methods are needed to
adequately represent random variability, often referred to as “objective uncertainty”,
and imprecision often referred to as “subjective uncertainty”. In risk assessment, no
distinction is traditionally made between these two types of uncertainty, both being
represented by means of a single probability distribution [2–4, 25, 26]. In case of
partial ignorance, the use of a single probability measure introduces information
that is in fact not available. This may seriously bias the outcome of risk analysis in
a non-conservative manner. Kaufman and Gupta [27] introduced hybrid numbers,
which simultaneously express inaccuracy and randomness. Guyonnet et al. [26] pro-
posed a method of risk estimation in the case when both probability and possibility
distributions are used to represent uncertainty. This method is a modification of the
method previously developed by Cooper et al. [15]. It combines stochastic simulation
with arithmetic of fuzzy numbers. The result of processing such data is given in the
form of two cumulative distribution functions: optimistic and pessimistic. Similarly,
Baudrit et al. [2] use probability and possibility distributions in risk analysis. Their
procedure also combines stochastic simulation with arithmetic on fuzzy numbers,
but they give the result in the form of a fuzzy random variable, which characterises
the examined phenomenon.
The use of hybrid data causes many problems. They involve primarily the prob-
lem of dependency between model parameters. Economic problems often involve
parameters that are mutually correlated. For example, there is a correlation between
enterprise product prices and raw material prices, also the volumes of sales of differ-
ent assortments are correlated. Usually, when processing such data, the independence
of the parameters is assumed [2, 26]. However, the omission of this dependency leads
to systematic errors in risk quantification, usually results in a large overestimation of
the actual risk. Another problem connected with the use of hybrid data is the lack of
universally accepted and easy interpretable measures that synthetically express the
risk of an economic activity. So far in the literature there is no solution to these prob-
lems. Construction of methods which will be able to take into account dependencies
(correlations) between uncertain parameters would enable the risk to be assessed far
more accurately. Solution of the dependence problem can potentially broaden the
scope of application of such methods to a wide range of decision problems.
The second weakness is the lack of synthetic risk measures (indexes) readable
for decision-makers. A method for risk appraisal can be accepted by economic life
practitioners only if it describes the risk in the form of synthetic indexes readable
and comprehensible for them. In the papers mentioned above risk was expressed by
optimistic and pessimistic cumulative distribution functions of the examined financial
ratio. Such cumulative distributions are difficult to interpret and often unintelligible
for applied decision-makers. Moreover, in the literature there is no consistent opinions
on how to designate optimistic and pessimistic cumulative distribution [2, 12].
Several numerical and symbolic methods have been proposed for handling uncer-
tain information. Three of the most common frameworks for representing and rea-
soning with uncertain knowledge are [1, 2, 4]:
78 5 Risk Assessment in the Presence of Uncertainty

• Bayesian probability theory.

• Dempster–Shafer (D–S) theory of evidence.
• Fuzzy set theory.
Each of these frameworks is aimed at a special application environment and has its
own features.

5.2 Measures of Risks in the Case of Hybrid Data

In order to effectively use fuzzy random variables in risk assessment, it is necessary

to define the expected value and the variance for these variables. Many authors define
the expected value for such variables in various ways. Most frequently, it is defined
in the form of a fuzzy set [28]. However, in decision-making problems, the expected
value is desired in a scalar form [29–31]. This facilitates interpretation of results.
Methods using such values are easily accepted by practitioners.
In case of the application of hybrid data the risk can be measured using standard
semi-deviation of analysed indicator calculated on the basis the fuzzy number or the
fuzzy random set or the upper and lower cumulative distribution function calculated
on the basis fuzzy random set.
The concept of the expected value is introduced for fuzzy variables. Liu and Liu
[29–31] define the expected value E(ξ) of a fuzzy variable ξ as follows:

+∞
E(ξ) = xdΦ(x), (5.1)
−∞

where Φ is a credibility distribution, if conditions lim Φ(x) = 0 and lim

x→ −∞ x→ +∞
Φ(x) = 1 are satisfied.
Standard deviation S Dev(ξ) of fuzzy variable ξ, which has a finite expected value
e, is defined by: 
S Dev(ξ) = E((ξ − e)2 ). (5.2)

Liu and Liu [29–31] proposed a new method for calculating the expected value
and the variance of a fuzzy random variable. The basic concept of these values is
based on the credibility of a fuzzy event [29–31].

Definition 5.1 Let ξ be a normal random fuzzy variable defined on a possibility

space (Θ, P(Θ), Pos). The expected value, E(ξ), of ξ is defined by:
⎡ ⎤
 ∞ ∞
E(ξ) = ⎣ Cr {ξ(ω)  x} d x− Cr {ξ(ω)  x} d x ⎦ P(dω). (5.3)
Ω
0 0
5.2 Measures of Risks in the Case of Hybrid Data 79

Definition 5.2 Let ξ be a normal random fuzzy variable defined on a possibility

space (Θ, P(Θ), Pos) and assume that E(ξ) < ∞. Then, the variance V ar (ξ) is
defined as the expected value of a random fuzzy variable (ξ − E(ξ))2 , i.e.,

V ar (ξ) = E (ξ − E (ξ))2 . (5.4)

Based on a fuzzy random variable, upper and lower distribution functions may
be estimated. These functions characterise uncertainty of the analysed variable. It is
known, that each fuzzy variable Z̃ with possibility distribution π induces a random
set [32]. Let α-levels of the said variable be marked with π α . Focal elements of
a random set generated by a fuzzy variable Z̃ are α-levels (π α j ) j=1,2,...,q , where
α0 = α1 = 1 > α2 > · · · αq > αq+1 = 0. Values (v j = α j − α j+1 ) j=1,...,q
constitute probability mass of the generated random set. Every random set induces
Plausibility (Pl) and Belief functions (Bel). Based on the defined functions Pl and
Bel, one may determine the upper F(x) and lower F(x) distribution function in
compliance with formulae (3.8) and (3.9).

5.3 Computing an Operation Profit

In order to illustrate the effectiveness of the proposed methods, below are presented
the results of calculations made for a simple model problem. The methods were
verified on the example of calculation of operating profit for a metallurgical industry
enterprise. The calculation was performed for the production system presented in
Fig. 5.1.

Fig. 5.1 Diagram of the analysed production system

80 5 Risk Assessment in the Presence of Uncertainty

Operating profit can be expressed by the formulae below.

Z O = cbg G bg + cbz G bz + cbo G bo + cbp G bp

− ksp Prsp − ksu Prsu − kst Prst
(5.5)
− kbg Prbg − kbz Prbz − kbo Prbo − kbp Prbp
− Z u zl czl − Z u gr cgr − Z u r u cr u − k f,

Prsp = Prsu m sp , (5.6)

Prsu = Prst m su , (5.7)

Prst = Prbg m pl , (5.8)

Prbg = Prbz m bg + G bg , (5.9)

Prbz = Prbo m bz + G bz , (5.10)

Prbo = Prbp m bo + G bo , (5.11)

Prbp = G bp , (5.12)

G bg = J Z bg u bg , (5.13)

G bz = J Z bz zu bz , (5.14)

G bo = J Z bo zu bo , (5.15)

G bp = J Z bp zu bp , (5.16)

Z u zl = Prst m zl , (5.17)

Z u gr = Prsu m gr , (5.18)

Z u r u = Prsp m r u , (5.19)

where:
G bg , G bz , G bo , G bp the sale of, respectively, hot rolled strip, cold rolled
sheets, hot dip galvanised strip and sheets, organic
coated sheets,
J Z bg , J Z bz , J Z bo , J Z bp the apparent consumption of, respectively, hot
rolled strip, cold rolled sheets, hot dip galvanised
strip and sheets, organic coated sheets,
5.3 Computing an Operation Profit 81

u bg , u bz , u bo , u bp the market share of, respectively, hot rolled strip,

cold rolled sheets, hot dip galvanised strip and
sheets, organic coated sheets,
cr u , cgt , czl , cbg ,cbz , cbo , cbp the price of, respectively, iron ore, pellets, steel
scrap, hot rolled strip, cold rolled sheets, hot dip
galvanised strip and sheets, organic coated sheets,
m sp , m su , m pl ,m bg , m bz , m bo respectively, the sinter consumption ratio per tonne
of pig-iron, pig-iron consumption ratio per tonne
of continuous casting stands, continuous casting
stands consumption ratio per tonne of hot rolled
sheets, hot rolled sheets consumption ratio per
tonne of cold rolled sheets, cold rolled sheets con-
sumption ratio per tonne of galvanised sheets, gal-
vanised sheets consumption ratio of per tonne of
organic coated sheets,
m zl , m gr , m r u the ratio of, respectively, scrap consumption per
tonne of steel, ratio of pellets consumption per
tonne of pig-iron, iron ore consumption ratio per
tonne of sinter,
Prsp , Prsu , Prst , Prbg , the production of, respectively, sinter, pig iron,
Prbz , Prbo , Prbp continuous casting stands, hot rolled strip, cold
rolled sheets, hot dip galvanised strip and sheets,
organic coated sheets,
ksp , ksu , kst , kbg , kbz , kbo , kbp the properly adjusted unit variable cost1 of, respec-
tively, sinter, pig iron, continuous casting stands,
hot rolled strip, cold rolled sheets, hot dip gal-
vanised strip and sheets, organic coated sheets,
kf company’s fixed costs,
Z u zl , Z u gr , Z u r u scrap, pellets and iron ore consumption, respec-
tively.
Figure 5.2 shows prices of metallurgical products manufactured by analysed com-
pany and prices of iron ore, pellets and steel scrap in 1992–2011. Apparent con-
sumption of metallurgical products manufactured by analysed company is depicted
in Fig. 5.3.
Trapezoidal fuzzy numbers specifying forecast of parameters for calculating the
operating profit for the analysed company are presented in Table 5.1. The relations
between prices of the analysed ranges of steel products, prices of iron ore, pellets and
scrap as well as the apparent consumption of particular product ranges were expressed
by means of the interval regression model. The coefficients of the regression equations
were estimated using the method described in Sect. 1.3.2. Table 5.2 presents the

1 This cost does not account for the values of used steel products manufactured in previous stages
of the cycle as well as value of used raw materials, corrections are done in order to avoid multiple
calculation of the same cost components during calculation of profit, according to formula (4.1).
82 5 Risk Assessment in the Presence of Uncertainty

Fig. 5.2 Prices of metallurgical products manufactured by analysed company and prices of iron
ore, pellets and steel scrap in 1992–2011

Fig. 5.3 Apparent consumption of metallurgical products manufactured by analysed company

exemplary coefficients of the interval regression equations characterising relations

between prices of products manufactured by analysed producer and prices of iron
ore, pellets and scrap.
The value of the fixed cost was adopted at the level of USD 315 090 thou/year.
Adjusted unit variable processing cost for particular product ranges were also adopted
at the levels given below:

Product Sinter Pig iron Continuous Hot rolled Cold rolled Hot dip Organic
casting strip sheets galvanised coated
stands strip and sheets
sheets
Adjusted 16.6 153.8 25.4 28.4 28 116.7 175.3
unit variable
processing
cost, USD/t

The following market share values of particular product ranges were adopted in
the calculations:
5.3 Computing an Operation Profit 83

Hot rolled strip Cold rolled sheets Hot dip galvanised strip and sheets Organic coated sheets
42.50 % 40.00 % 46.00 % 45.00 %

Table 5.1 Trapezoidal fuzzy numbers representing forecasts of products and raw material prices,
material consumption indicators and apparent consumption of metallurgical products
Price Trapezoidal fuzzy numbers, USD/tonne
Iron ore (111.7, 120.0, 133.3, 141.7)
Pellets (125.0, 133.3, 146.7, 156.7)
Steel strap (313.3, 320.0, 336.7, 345.0)
Hot rolled strip (666.7, 680.0, 711.7, 728.3)
Cold rolled sheets (715.0, 730.0, 763.3, 781.7)
Hot dip galvanised strip and sheets (805.0, 821.7, 860.0, 880.0)
Organic coated sheets (1 080.0, 1 101.7, 1 153.3, 1 175.0)
Material consumption indicators Trapezoidal fuzzy numbers, tonne/tonne
Iron ore—sinter (0.918, 0.920, 0.920, 0, 922)
Sinter—pig iron (1.352, 1.354, 1.359, 1.362)
Pellets—pig iron (0.338, 0.339, 0.340, 0.341)
Scrap—continuous casting stands (0.269, 0.276, 0.279, 0.288)
Pig iron—continuous casting stands (0.855, 0.860, 0.870, 0.875)
Continuous casting stands—hot rolled strip (1.058, 1.064, 1.075, 1.078)
Hot rolled strip—cold rolled sheets (1.105, 1.111, 1.124, 1.130)
Cold rolled sheets—hot dip galvanised strip (1.010, 1.020, 1.026, 1.031)
and sheets
Hot dip galvanised strip and sheets—organic (0.998, 0.999, 1.000, 1.001)
coated sheets
Apparent consumption Trapezoidal fuzzy numbers, thousands of
tonnes
Hot rolled strip (2 405.4, 2 704.0, 2 704.0, 3 033.6)
Cold rolled sheets (1 018.2, 1 162.3, 1 162.3, 1 293.0)
Hot dip galvanised strip and sheets (1 008.4, 1 147.9, 1 147.9, 1 287.4)
Organic coated sheets (625.9, 708.4, 708.4, 791.2)

The operating profit was calculated using different methods of mathematical oper-
ations. The obtained results are presented in Figs. 5.4, 5.5, 5.6, 5.7, 5.8 and 5.9. Profits
labelled with numbers I–IV have been calculated using various methods of implemen-
tation of arithmetic operations on fuzzy numbers. Operating profits V and VI were
calculated using the hybrid propagation method. At the calculation of this profits the
apparent steel consumption of individual metallurgic product mixes one expressed
in the form of probability distributions. Table 5.3 presents probability density func-
tion parameters presenting apparent steel consumption for products produced by the
company. Remaining parameters burdened an uncertainty in case of these calcula-
tions were expressed in the form of numbers fuzzy presented in Table 5.1. Table 5.4
Table 5.2 Interval regression equations coefficients depicting interrelations between prices of particular product ranges produced by the manufacturer in
84

question prices of raw materials

Independent variable Dependent variable
Iron ore Pellets Steel scrap Hot rolled Cold rolled Hot dip Organic
strip sheets galvanised coated sheets
strip and
sheets
Iron ore a1 [0.80, [0.20, [0.17, [0.10, [0.06, 0.17] [0.07, 0.16]
0.89] 0.34] 0.26] 0.20]
a2 [6.87, [13.00, [−46.54, [−18.17, [−7.70, [−48.12,
7.69] 22.39] −30.09] −9.18] −2.80] −21.79]
Pellets a1 [1.11, 1.25] [0.24, 0.42] [0.19, 0.30] [0.11, 0.24] [0.07, 0.21] [0.08, 0.19]
a2 [−8.30, [8.46, [−63.98, [−32.08, [−23.86, [−67.82,
−7.37] 14.91] −39.89] −15.30] −7.89] −28.57]
Steel strap a1 [1.46, 4.07] [1.32, 3.42] [0.40, 0.87] [0.21, 0.71] [0.32, 0.60] [0.13, 0.56]
a2 [5.16, [14.65, [−172.16, [−86.59, [−73.93, [−178.41,
14.37] 38.10] −80.04] −26.03] −38.69] 42.87]
Hot rolled strip a1 [3.48, 4.51] [2.90, 3.80] [1.01, 1.36] [0.60, 0.79] [0.48, 0.66] [0.39, 0.66]
a2 [211.97, [237.92, [247.72, [112.16, [112.61, [31.68,
275.29] 311.40] 334.56] 148.39] 155.43] 53.32]
Cold rolled sheets a1 [3.55, 5.80] [3.04, 4.83] [1.12, 1.70] [0.87, 1.41] [0.75, 0.92] [0.57, 0.96]
a2 [191.11, [221.35, [239.84, [−17.12, [−2.65, [−159.73,
312.26] 351.44] 364.71] −10.59] −2.18] −95.30]
Hot dip galvanised strip and sheets a1 [3.59, 5.76] [3.14, 4.89] [1.19, 1.78] [0.96, 1.50] [0.98, 1.22] [0.60, 1.07]
a2 [292.01, [318.79, [331.01, [45.67, [57.88, [−87.59,
468.72] 497.23] 492.93] 71.29] 72.08] −48.89]
Organic coated sheets a1 [3.96, 6.14] [3.28, 5.20] [1.16, 1.79] [0.35, 0.87] [0.81, 1.31] [0.63, 1.12]
a2 [458.75, [482.39, [511.55, [385.57, [248.18, [236.97,
5 Risk Assessment in the Presence of Uncertainty

712.03] 765.86] 790.07] 941.97] 403.48] 420.89]

5.3 Computing an Operation Profit 85

Fig. 5.4 Fuzzy numbers representing the operating profit I

Fig. 5.5 Fuzzy numbers representing the operating profit I–IV

presents a matrix of correlation of apparent consumption of particular product ranges

manufactured by a producer.
The value of all coefficients of correlation was significant at testing by means of
the t-Student statistics for the significance level α = 0.05.
The operating profit I was calculated using simulation of fuzzy systems (method
described in Sect. 1.3.1) for execution of arithmetic operations on interactive fuzzy
numbers. The operating profit II one calculated using non-linear programming
(method described in Sect. 1.4.1) for execution of arithmetic operations on inter-
active fuzzy numbers. Additionally, for comparison one calculated the operating
profit III and IV. The operating profit III one calculated on the assumption that did
not exist dependencies among prices of individual product mixes and prices of raw
materials and dependencies among the apparent consumption of individual metallur-
86 5 Risk Assessment in the Presence of Uncertainty

Fig. 5.6 Cumulative credibility distribution functions for operating profit I–IV

Fig. 5.7 Example fuzzy numbers depicting operating profit calculated in the selected iterations of
computer simulation

gic product mixes. The operating profit III was calculated as follows: the production
cost of each product, from organic coated sheets to pig iron, was calculated using
the Eqs. (1.29) and (1.30), the revenue was calculated using Eqs. (1.29) and (1.30),
and the operating profit was calculated by subtracting the production costs from the
revenue, according to the Eq. (1.38). At this stage one applied the constrained sub-
traction. The use of the Eq. (1.38) would mark, the subtraction of costs calculated for
the greatest possible production from the income calculated for least sales and vice
versa.
The operating profit IV one calculated on the assumption that existed strong,
functional dependencies among prices of individual product mixes and prices of raw
5.3 Computing an Operation Profit 87

Fig. 5.8 Credibility distribution functions depicting operating profit calculated in selected iterations
of computer simulation

Fig. 5.9 Upper and lower distribution functions depicting operating profit

materials and dependencies among the apparent consumption of individual metal-

lurgic product mixes. The operating profit IV one calculated on the assumption that
the highest possible prices of the various product ranges correspond to the highest
prices of metallurgical raw materials and the lowest possible prices of individual
assortments of steel products correspond to the lowest prices of raw materials. Sim-
ilarly one assumed that the apparent consumption of all metallurgic product mixes
accepts highest possible values or lowest possible values. To implement the cal-
88 5 Risk Assessment in the Presence of Uncertainty

Table 5.3 Probability density function parameters presenting apparent steel consumption for prod-
ucts produced by a company
Product (Average, 000’ tonnes, standard deviation, 000’
tonnes)
Hot rolled sheets (2704.0, 117.5)
Cold rolled sheets (1162.3, 51.4)
Hot dip galvanised sheets (1147.9, 52.4)
Organic coated sheets (708.4, 30.8)

Table 5.4 Correlation matrix for the apparent consumption of metallurgical products manufactured
by analysed company
Hot rolled strip Cold rolled sheets Hot dip Organic coated
galvanised strip sheets
and sheets
Hot rolled strip 1 0.878 0.911 0.863
Cold rolled sheets 0.878 1 0.915 0.888
Hot dip 0.911 0.915 1 0.966
galvanised strip
and sheets
Organic coated 0.863 0.888 0.966 1
sheets

culation under these assumptions we used the method described in Sect. 1.4.1. To
take into account the above assumptions, the system constrains were modified. The
constraints (1.53) and (1.54) and the constraint (1.52) were removed for parameters
corresponding to the price, whereas the apparent consumption was modified to the
form xi = sup (X iα ) at the search of solution with the criterion function (1.50) and
to the form xi = inf (X iα ) at the search of solution with the criterion function (1.51).
Figure 5.5 presents the comparison of fuzzy numbers representing the operating
profit I–IV and Fig. 5.6 shows the cumulative credibility distribution functions for
these profits.
The comparison of trapezoidal fuzzy numbers defining the operating profit I is
presented in Fig. 5.4. The figure presents data obtained from the simulation and the
data determined thanks to the appraisal value of parameters of the equation:
⎧
⎪
⎪ 1− b−x
for a ≤ x ≤ b
⎨ b−a
1 for b ≤ x ≤ c
μ (x) = ,
⎪
⎪ 1− x−c
for c ≤ x ≤ d
⎩ c−b
0 otherwise

where a, b, c, d—parameters of the equation.

5.3 Computing an Operation Profit 89

Figures 5.7 and 5.8 depict example results of calculations of the operating profit
with the use of the hybrid method (operating profit V). Figure 5.7 presents four fuzzy
numbers characterising operating profit calculated in the selected iterations of com-
puter simulation and Fig. 5.8 presents respective credibility distribution functions.
The average value of operating profit was USD 361333.6 thou. and standard semi-
deviation USD 94303.6 thou. Figure 5.9 depicts pessimistic and optimistic cumula-
tive distribution functions resulting from the said computations.
Additionally, calculations were effected without consideration of the correlation
of economic parameters (operating profit VI). In this case, the average of operating
profit obtained was 315488.9 and standard semi-deviation was 308475.5. Figure 5.10
presents the resulting optimistic and pessimistic cumulative distribution functions.
The average and standard semi-deviation of the operating profit for different vari-
ants of calculations one presented in Table 5.5.
Fuzzy numbers and cumulative credibility distribution functions presented in
Figs. 5.4, 5.5 and 5.6 show that the application of different variants of arithmetic
operations give different results. The assumption adopted for the calculation of the
operating profit III and IV results in too narrow or too broad compartments for α-
levels of fuzzy numbers representing the result of arithmetic operations. They are
unreal and impossible to obtain in reality. This results in a suitably large or small
values of the standard semi-deviation of operating profit. The application of the
method for performing arithmetic operations on fuzzy numbers that was described
in Sects. 1.3.1 and 1.4.1 eliminates the deficiencies of arithmetic operations per-
formed according to the Eqs. (1.29), (1.31) and (1.38). Thanks to this methods, it
is possible to take into account all interactivities between the analysed quantities.
However, in the case of typical links appearing between economical parameters in

Fig. 5.10 Upper and lower distribution functions depicting operating profit obtained in the variant
of calculations not taking into account correlation of parameters of the economic calculation
90 5 Risk Assessment in the Presence of Uncertainty

Table 5.5 The average and standard semi-deviation of the operating profit for different variants of
calculations
Average, 000’ USD Std. semi-dev., 000’ USD
Operating profit I 307 918.00 217 917.70
Operating profit II 305 070.40 214 469.40
Operating profit III 330 511.30 348 214.80
Operating profit IV 309 084.40 123 603.30
Operating profit V 331333.6 94 303.60
Operating profit VI 315488.9 308475.5

a company, calculating the values of arithmetic expressions characterising financial

indices requires the application of nonlinear programming methods, indeed, this
complicates the process of calculation. Moreover, in case of application of the meth-
ods of non-linear programming (the method described in Sect. 1.3.1), estimation of
a large number of equations of fuzzy linear regression is necessary. With a larger
number of interactive parameters, it may complicate the calculation process. This
defect is not present with the method using simulations of fuzzy systems. However,
in case of a large number of analysed parameters, the correspondingly large number
of replications is required, which extends the calculation process.
Proposed hybrid method facilitates processing of hybrid data, taking into account
the correlation between economic parameters. The procedure for processing such
data combines stochastic simulation with Zadeh’s extension principle (the method for
execution of arithmetic operations on fuzzy numbers). Non-linear programming was
used in execution of arithmetic operations on interactive fuzzy numbers. Relations
between economic parameters were expressed in the form of interval regression.
The proposed method may be used in the case of fuzzy numbers with arbitrary
membership functions, because arithmetic operations are executed on α-levels. This
method is an universal and flexible tool for processing hybrid data, which enable the
existing correlation to be taken into account.
The concerned Figs. 5.4, 5.5, 5.6, 5.9 and 5.10 and Table 5.5 presented results of
computations indicate that correlation of economic parameters has a considerable
impact on estimates of operating profit. Estimation of risk without consideration of
these interrelations would bear a considerable, systematic error.

References

1. Choobineh, F., and A. Behrens. 1992. Use of intervals and possibility distribution in economic
analysis. Journal of Operations Research Society 43(9): 907–918.
2. Baudrit, C., D. Dubois, and D. Guyonet. 2006. Joint propagation and exploitation of proba-
bilistic and possibilistic information in risk assessment. IEEE Transaction on Fuzzy Systems
14(5): 593–607.
References 91

3. Gupta, C.P. 1993. A note on the transformation of possibilistic information into probabilistic
information for investment decisions. Fuzzy Sets and Systems 56(2): 175–182.
4. Kuchta, D. 2001. Mie˛kka matematyka w zarza̧dzaniu: zastosowanie liczb przedziałowych
i rozmytych w rachunkowości zarza̧dczej. Wrocław: Oficyna Wydawnicza Politechniki
Wrocławskiej.
5. Pluta, W., and T. Jajuga. 1995. Inwestycje: Capital Budgeting—Budżetowanie Kapitałowe.
Fundacja Rozwoju Rachunkowości w Polsce.
6. Dittmann, P. 2008. Prognozowanie w przedsiȩbiorstwie: metody i ich zastosowanie. Wolters
Kluwer.
7. Kuchta, D. 2000. Fuzzy capital budgeting. Fuzzy Sets and Systems 111(3): 367–385.
8. Mohamed, S., and A.K. McCowan. 2001. Modelling project investment decisions under uncer-
tainty using possibility theory. International Journal of Project Management 19(4): 231–241.
9. Ferson, S. 1996. What Monte Carlo method cannot do. Human and Ecological Risk Assessment:
An International Journal 2(4): 990–1007.
10. Georgescu, I., and J. Kinnunen. 2011. Modeling the risk by credibility theory. In Proceedings
of the 3rd international IEEE conference on advanced management science (ICAMS 2011),
Kuala Lumpur, Malaysia, International proceedings of economics development and research,
vol. 19, 15–19.
11. Chiu, C.-Y., and S.C. Park. 1994. Fuzzy cash flow analysis using present worth criterion. The
Engineering Economist 39(2): 113–138.
12. Rȩbiasz, B. 2013. Selection of efficient portfolios-probabilistic and fuzzy approach, compara-
tive study. Computers and Industrial Engineering 64(4): 1019–1032.
13. Rȩbiasz, B. 2007. Fuzziness and randomness in investment project risk appraisal. Computers
and Operations Research 34(1): 199–210.
14. Ferson, S., and L.R. Ginzburg. 1996. Difference method are needed to propagate ignorance
and variability. Reliability Engineering and System Safety 54(2–3): 133–144.
15. Cooper, J.A., S. Ferson, and L. Ginzburg. 1996. Hybrid processing of stochastic and subjective
uncertainty data. Risk Analysis 16(6): 785–791.
16. Ward, T.L. 1985. Discounted fuzzy cash flow analysis. In 1985 Fall Industrial engineering
476–481.
17. Buckley, J.J. 1992. Solving fuzzy equations in economics and finance. Fuzzy Sets and Systems
48(3): 289–296.
18. Calzi, M.L. 1990. Toward a general setting for the fuzzy mathematics of finance. Fuzzy Sets
and Systems 35(3): 265–280.
19. Esogbue, A.O., and W.E. Hearnes II. 1998. On replacement models via a fuzzy set theoretic
framework. IEEE Transactions on Systems Manufacturing and Cybernetics—Part C: Applica-
tions and Reviews 28(4): 549–560.
20. Kahraman, C., D. Ruan, and E. Tolga. 2002. Capital budgeting techniques using discounted
fuzzy versus probabilistic cash flows. Information Science 142(1–4): 57–76.
21. Rȩbiasz, B., B. Gaweł, and I. Skalna. 2014. Capital budgeting of interdependent projects with
fuzziness and randomness. In Information systems architecture and technology, ed. Zofia Wil-
imowska et al, 125–135.
22. Rȩbiasz, B., B. Gaweł, and I. Skalna. 2015. Fuzzy multi-attribute evaluation of investments.
In Advances in ICT for business, industry and public sector: ABICT’13 (4th international
workshop on Advances in business ICT), eds. Tomasz Pełech-Pilichowski Maria Mach-Król,
and Celina M. Olszak, 141–156.
23. Rȩbiasz, B., B. Gawel, and I. Skalna. 2014. Hybrid framework for investment project portfolio
selection. In Federated conference on computer science and information systems 1123–1228.
24. Smets, P. 1990. Constructing the pignistic probability function in a context of uncertainty. In
Proceedings of the fifth annual conference on uncertainty in artificial intelligence, ed. Henrion
M., E.D. Schachter, L.N. Kanal, and J.F. Lemmer. UAI’89, 29–39. Amsterdam: North-Holland
Publishing Co.
25. Shafer, G. 1976. A mathematical theory of evidence. Princeton: Princeton University Press.
92 5 Risk Assessment in the Presence of Uncertainty

26. Guyonnet, D., B. Bourgine, D. Dubois, H. Fargier, B. Côme, and P.J. Chilès. 2003. Hybrid
approach for addressing uncertainty in risk assessment. Journal of Environmental Engineering
129(1): 68–76.
27. Kaufmann, A., and M.M. Gupta. 1985. Introduction to fuzzy arithmetic: Theory and applica-
tion., Electrical-computer science and engineering series New York: van Nostrand Reinhold
Company.
28. Puri, M.L., and D.A. Ralescu. 1985. The concept of normality for fuzzy random variables.
Annals of Probability 13(4): 1371–1379.
29. Liu, Y.-K., and B. Liu. 2003. Fuzzy random variables: A scalar expected value operator. Fuzzy
Optimization and Decision Making 2(2): 143–160.
30. Liu, Y.-K., and B. Liu. 2003. Expected value operator of random fuzzy variable and random
fuzzy expected value models. International Journal of Uncertainty, Fuzziness and Knowledge-
Based Systems 11(2): 195–215.
31. Liu, Y.-K., and B. Liu. 2002. Expected value of fuzzy variable and fuzzy expected value models.
IEEE Transactions on Fuzzy Systems 10(4): 445–450.
32. Zadeh, L.A. 1978. Fuzzy sets as a basis for a theory of possibility. Fuzzy Sets and Systems 1(1):
3–28.
Chapter 6
Application of Fuzzy Theory in Steel
Production Planning and Scheduling

Abstract This chapter describes the application of fuzzy sets to planning and
scheduling of production in the steel industry. Primarily, the problem of steel grade
assignment to customer’ orders is analysed. Fuzzy sets are used to reduce the vari-
ety of potential steel grades and to describe characteristic of materials by decision
makers. Next, fuzzy logic systems for steel production scheduling are examined.
Fuzzy parameters and fuzzy constraints are used to describe some aspects of the
steel production process, with a special respect to the continuous casting. Finally,
the cooperation of steel production planning between different shops using a multi-
agent approach and fuzzy sets is discussed and the practical example of a genetic
algorithm applied to solve a fuzzy lot-sizing problem for a continuous casting plan-
ning agent is presented.

Typical steelworks consists of a steel melting shop (furnace), at least one continuous
casting line, a hot strip mill and, optionally, further processing facilities, such as cold
rolling mill or pickling line. The steel is molten in furnaces and transferred directly
to continuous casters. Caster shapes molten steel into slabs with different widths and
chemical composition (grades). Finally, the slabs are rolled in the hot strip mill to
produce steel coils.
Planning and scheduling of steel production is a very demanding challenge due
to the complex nature of a steel making process. Many parameters of this process
are expressed by normative values that may differ from the real values occurring in
a given production circumstances. Decision makers must take into account some level
of uncertainty of parameters and adjust the process according to their own knowledge
and experience. The most important aspects of the steel production process that must
be reflected in the production schedules are the following [1]:
• At the steel-making stage: orders change, machines failures, smelting time exceed-
ing the prescribed limit etc.
• At the continuous caster stage: information concerning the true weight of the ladle
contents, heats (a fixed tonnage of molten steel) arrive randomly and dynamically,
heats of steel having the wrong chemistry, steel leak appears in the continuous
caster, machine failure etc.
94 6 Application of Fuzzy Theory …

• At the hot strip mill stage: slab quality is not up to standard, slabs are delayed,
slabs are backed up in the hot strip mill, a new high-priority order is introduced,
an order is cancelled, machine failure etc.
The uncertainty concerns not only the production process itself, but also the cus-
tomer demand for specific types and grades of steel products, quality issues, and
logistics (procurement, transport between successive stages, delivery to the final
customer).
Zarandi and Ahmadpour [2] enumerate general characteristic of the steel industry
in the context of planning and scheduling tasks:
• Steel production is a multi-stage process, logically and geographically distributed,
involving a variety of production processes
• There are different problems and different problem solving methods for different
steel making stages.
• Manual techniques used in steel making processes are based on the know-how
and the professional experience of expert people who have worked in the plant for
years.
• The output of some stages is the input for some other stages, so the integrated
process is necessary.
• Making contracts requires negotiation between buyer and supplier.
All the above factors cause that steelworks usually do not use a fully computerised
planning and scheduling system, especially at the shop floor level. Nevertheless,
some individual solutions supporting planning and scheduling of the production
process at various stages of decision making has been successfully developed in
many steelworks. Tang et al. [1] have classified the methods used in such solutions
into the following groups:
• Operations research methods in which linear programming or mixed integer pro-
gramming model is built and then some exact or heuristic method is used in ordered
to solve it.
• Artificial intelligence methods including expert systems, computational intelli-
gence and constraint satisfaction approach.
• Visualisation techniques in human-machine coordination systems.
• Multi-agent approach.
In both artificial intelligence and multi-agent approaches fuzzy sets or fuzzy logic is
often applied in order to describe uncertain parameters of a steel production process.
The following sections present the application of fuzzy theory and fuzzy logic to the
planning and scheduling process, beginning with the problem of grade assignment,
job scheduling taking into account fuzzy due dates, and coordination of the plans
by exchanging messages between agents. Finally a detailed optimisation model with
fuzzy parameters for the continuous caster agent has been shown and a genetic
algorithm able to solve it in an efficient way has been proposed.
6.1 Fuzzy Grade Assignment 95

6.1 Fuzzy Grade Assignment

Customer orders for steel products are characterised by two major parameters:
dimension and steel grade. Steel grades are usually standardised according to either
intended use and mechanical properties of the steel or its chemical composition.
In such standards many of the mechanical parameters and chemical characteristics
are given by their lower and upper bounds (or one of them). For example the steel
graded identified as SAE 304 gives the following ranges of the chemical components:
Cr 23–26%, Ni 19–22 %, C 0.25 %, Mn 2 %, Si 1.5–3 %, P 0.045 % and S 0.03 %.
Due to inevitable uncertainties in the steel production process (especially at the
steelmaking level), it is hardly possible to produce steel with a precisely prede-
termined quality and to maintain the same quality for all orders placed by a cus-
tomer. Therefore the quality of the steel assigned to the product is usually better than
requested by the customer. This is done not only to protect the quality requirements
of the customer, but also to improve the efficiency of the steel making process, e.g.,
to consume all the steel that was melted in the furnace, or to achieve a maximal
throughput of the continuous casters.
The uncertainty concerning chemical composition results from the fact, that the
steelmaking process for different grades is performed on the same units and in par-
ticular from [3]:
• residuals in an electric arc furnace remain in the wall and will be assimilated by
the subsequent heat,
• two heats with different steel grades casted one after another are mixed in the
tundish.
To avoid the above problems Vasko et al. [4] and later Woodyatt et al. [5] proposed a
method based on the concept of fuzzy sets that reduced the variety of potential steel
grades. The idea was that instead of a single grade specified in a customer order,
it is assigned a set of possible grades that would meet customer’s requirements.
Figure 6.1 illustrates that requirements for carbon and manganese can be satisfied by
five different steel grades.
The method works as follows. First, a set of potential grades that can be used
to satisfy customer’s requirements is assigned to the order and then the grade that
will be sufficient to produce all the orders is selected. The probability that the grade
will meet the specification of the customer’s order is described by fuzzy sets. The
selection of one grade for a collection of customer orders is performed by a set
covering approach.
The complete collection of N grades is identified as:

G = {G i }, ∀ j = 1, 2, . . . , N (6.1)

A collection of M customer orders to which a steel grade must be assigned:

S = {Si }, ∀i = 1, 2, . . . , M (6.2)
96 6 Application of Fuzzy Theory …

Fig. 6.1 Grade assignment on the basis of fuzzy carbon and manganese requirements [4]

Finally, a collection of fuzzy subsets of G, such as A(i) measures the ability of

a grade in the list G to meet the specification of customers order Si .

A = {A(i)}, ∀i = 1, 2, . . . , M (6.3)

The membership function f for each A(i)depends on three components:

• the ability of a grade i in the list G to meet all mechanical property requirements
of customer order Si
• the ability of a grade i in the list G to meet all chemical requirements of customer
order Si
• desirability of producing a grade i in the list G relative to other grades in the list.

and is defined as follows:

f A(i,K ) (G i ) = Prob[m(i, K ) ≤ X ( j, K ) ≤ M(i, K )]

, (6.4)
∀ j = 1, 2, . . . , N , ∀i = 1, 2, . . . , M, ∀K = 1, 2, . . . , U

where U is the number of mechanical properties, m(i, K ) and M(i, K ) are, respec-
tively, the minimum and maximum level of mechanical property K for order Si .
The solution proposed by Vasko et al. [4] has been implemented in Bethlehem
Steel (USA) as the MGAP module (metallurgical grade assignment program) within
the production planning and control (PPC) system. Limitation of steel grades pro-
duced by the steelworks contributed to better utilisation of continuous casting lines,
but also to reduction of molten steel necessary to satisfy customers’ orders.
In [6] Wang and Chang used fuzzy sets to solve a similar problem of assign-
ment tool steel to customer’s order. This time, however, the fuzzy sets were used to
express the importance of criteria for decision makers. The following characteristic
of materials were considered:
6.1 Fuzzy Grade Assignment 97

• non-deforming properties,
• safety in hardening,
• toughness,
• resistance to softening effect of heat,
• wear resistance,
• machinability.

All those criteria are weighted by n decision makers using linguistic variables: VL—
very low, L—low, M—medium, H —high and V H —very high. Opinions of decision
makers were then aggregated using the formula:

Wt = (1/n) ⊗ (Wt1 ⊕ Wt2 ⊕ . . . Wtn ), t = 1, 2, . . . , k (6.5)

where Wt j was the value of the weight for t-th criteria assigned by the j-th decision
maker. A linguistic ranking scale used for describing the weights is shown in Fig. 6.2.
Next, the performance ratings for the above material characteristics and the cost
for the given material are evaluated by the decision makers. This time the linguistic
scale is: worst (W ), poor (P), fair (F), good (G), and best (B). The membership
functions are analogous to the one shown in Fig. 6.2.
Again the ratings of all decision-makers are aggregated:

Rit = (1/n) ⊗ (Rit1 ⊕ Rit2 ⊕ . . . Ritn ), m = 1, 2, . . . , m, t = 1, 2, . . . , k (6.6)

where Rit j was the aggregated rating of alternative i under criterion t assigned by
the j-th decision maker.
After assignment of weights and ratings the final rating Fi for i-th alternative is
evaluated using the formula:

Fi = (1/k) ⊗ [(Ri1 ⊗ W1 ) ⊕ (Ri2 ⊗ W2 ) ⊕ . . . (Rik ⊗ Wk )] (6.7)

where Fi ≈ (Yi , Q i , Ri , Z i ) is the approximated fuzzy number of the fuzzy suitabil-

ity index of alternative i.

Fig. 6.2 Linguistic ranking scale for criteria weights [6]

98 6 Application of Fuzzy Theory …

The most suitable tool steel material for a given order in the proposed method is
chosen after ranking such fuzzy ratings basing on simple maximisation and minimi-
sation operators for fuzzy sets proposed by Chen [7]. Maximisation is defined as:

M = {(x, f M (x))|x ∈ R} (6.8)

with the membership function defined as:


(x − x1 )/(x1 − x2 ), x1  x  x2 ,
FM (x) = (6.9)
0, otherwise.

Analogously, the minimisation is defined as:

G = {(x, f G (x))|x ∈ R} (6.10)

with the membership function:


(x − x2 )/(x2 − x1 ), x1  x  x2 ,
FG (x) = (6.11)
0, otherwise,

where m
x1 = inf D, x2 = sup D, D = i=1 Di ,
(6.12)
Di = {x| f Fi (x) > 0}, i = 1, 2, . . . , m

The right utility function U M (Fi ) and the left utility function UG (Fi ) of each fuzzy
rating evaluation Fi are defined, respectively, by:

U M (Fi ) = sup( f Fi (x) ∧ f M (x)), (6.13)

x

UG (Fi ) = sup( f Fi (x) ∧ f G (x)). (6.14)

x

The final value of ranking for alternative i is calculated as:

U M (Fi ) + 1 − UG (Fi )
UT (Fi ) = (6.15)
2
The higher value of UT the higher is the ranking of alternative i.
Although such approach can provide very accurate assignments, complex calcula-
tions result in a high computational time for evaluation of the final ranking. Therefore
Chen in [8] proposed a simplified method of ranking calculation.
In this approach the aggregation of weights has been replaced by their deffuzifica-
tion. A simple formula of deffuzification for trapezoidal fuzzy numbers is used [8]:

a+b+c+d
e= (6.16)
2
6.1 Fuzzy Grade Assignment 99

each criterion t assigned by j-th decision maker is defuzzified as:

at j + bt j + ct j + dt j
wt j = (6.17)
4
where (at j , bt j , ct j , dt j ) is a quadruple characterising linguistic variable Wt j repre-
sented by a fuzzy number.
The aggregated weight for each criterion t can be calculated as:

wt1 + wt2 + · · · + wtn

T (Ct ) = (6.18)
n
Similarly the value of criterion t for potential candidate material Ai can be deffuzzi-
fied as:
pit + xit + yit + z it
rit = (6.19)
4
where ( pit , xit , yit , z it ) is a quadruple characterising linguistic variable Rit repre-
sented by a fuzzy number.
The rating for material Ai based on the weighted criterion t can be achieved as a
simple product of T (Ci )*rit except for the cost of the material. For this criterion a
converting function is used:

1
F(Ai ) = (6.20)
vi ∗ ( v11 + 1
v2
+ · · · v1m )

Finally the rating for material Ai is calculated as:

R(Ai ) = ri1 ∗ T (C1 ) + ri2 ∗ T (C2 ) + · · ·

+ ri(k−1) ∗ T (Ck−1 ) + F(Ai ) ∗ T (Ck ) (6.21)

Chen showed in a practical example that his method gave the same ranking as the
method proposed by Wang and Chang [9].

6.2 Scheduling of Production Processes Using Fuzzy Logic

The most effective way to control continuous casting process is to run it as fast as
possible, but with the respect to the metallurgical quality of the steel. The chemical
and mechanical properties of the steel primary depend on the temperatures reached
during the casting process. This process must be controlled in such a way that the
temperatures fit in a specific range. Those temperatures are described by technologists
in operation sheets according to the desired mechanical and chemical characteristic
of the steel. The casting process can be controlled by regulating the speed of casting
100 6 Application of Fuzzy Theory …

and cooling intensities for each coolant circuit. Necessary condition of the correct
setting is that the material has to be solidified before it leaves from the caster. This
property is so-called the metallurgical length—the length of liquid material from the
meniscus [10].
In the planning and scheduling systems for steel production described in the litera-
ture uncertain parameters of the process and customers’ requirements are principally
expressed in the form of IF-THEN rules based fuzzy system.
Fuzzy logic can be seen as an extension of traditional Boolean logic based on the
theory of fuzzy sets [11]. By introduction of the notion of degree in the verification of
a condition, fuzzy logic enables a condition to be in a state other than only true or false
(e.g. partially true that is between completely true and completely false). Thereby in
many cases fuzzy logic may be more flexible for reasoning about phenomena that
occur in real life, including the production processes that are not fully predictable
and stable like in metal or chemical industry. Conventional logic forces the users
to describe some inaccurate or uncertain parameters of the production process in a
rough or approximate way, whereas sometimes it is easier and overall more accurate
to express them in the form that is similar to the natural language (e.g., temperature
is adequate).
The fuzzy system that is most commonly used in decision making consists of the
set of rules in the following form:
IF pr emise1 and pr emise2 and pr emiseN THEN conclusion

where pr emise1 is a statement of the type “xi is of L i, j ”, xi represents a flexible

predicate naming j-th linguistic term of the corresponding i-th linguistic variable
and L i, j is given by a fuzzy set that represents the use of flexible predicate on the
domain of xi . The conclusion is also a fuzzy set representing a flexible predicate on
the output of the system if the premises are satisfied [12].
Inference based on a fuzzy system and modus ponens/modus tollens rules requires
the definition of a fuzzy equivalent to implication relationship R(x, y) : A(x) →
B(y). Many definitions such relationships have been provided in the literature, but
two of them are the ones most commonly used in practise:
• minimum operation: R(x, y) = min{A(x), B(y)} proposed by Mamdani [13]
• product operation: R(x, y) = A(x) · B(y) proposed by Larsen [14]
For compositional rule of inference min-max is used in Mamdani method and
product-sum in Larsen.
Dorn et al. [15] proposed a reactive scheduling system for steel making process
that uses fuzzy logic to create schedules that are robust with respect to changes due to
certain types of event and fuzzy constraints to minimise the changes in the schedules,
if rescheduling is required. The system has been built for a steel plant that produces
high-grade steel from crude steel in a form of slabs and ingots. The products are then
transported for subsequent processing like rolling and forging to the other plants. An
outline of the steelmaking process in the steelworks is shown in Fig. 6.3.
6.2 Scheduling of Production Processes Using Fuzzy Logic 101

Fig. 6.3 The production flow in the steelworks analysed by Dorn [15]

Crude steel is melted two electric arc furnaces (EAF1 and EAF2), then it processed
in the secondary metallurgy (SM1 and SM2) by pouring into a ladle furnace where
fine alloying takes place, then in a vacuum oxygen decarburation unit. After that the
steel is cast in a continuous caster (CC) to form slabs or into the moulds to form
ingots. Solidification of ingots moulds requires a space in a teeming bay (TB) and
for some large ingots, the Böhler-Electro-Slag-Topping (BEST) technology is used.
The authors enumerates three main kinds of fuzzy constraints (in the sense that
they are not 100 % crisp) that must be taken into account during generation of sched-
ules. First is the compatibility of the chemical properties of the steel between succes-
sive heats. As has been already described in the previous section, metal remaining
after one heat in the electric arc furnace can interfere chemical characteristic of the
next heat. In order to deal with this problem the engineers use as a rule of thumb that
3 % of a chemical element in a heat remain in the wall of the electric arc furnace, so
3 % of the difference of the elements in two consecutive heats should be assimilated
by the second heat. The second group of constraints is related to due dates. Some
steel forming processes require the cast to remain hot for subsequent treatments and
for those castings (about 10 % of the production) due dates should be met within
a tolerance of 2 h. Due to the nature of the casting process, it is also required the
jobs on continuous caster to be scheduled either continuously or with sufficiently
breaks enough for set-up or maintenance operations. The third group of constraints
concerns capacity restriction. Continuous casting is a key process that directly influ-
ences production throughput and production costs, so it must be scheduled efficiently
to reduce unnecessary breaks and reheatings. Contrary to it, the jobs for ingots that
needs to solidify using BEST technology has to be scheduled with the certain time
gaps between each other due to a long solidification time of such ingots.
The authors also postulate to describe processing times using fuzzy numbers
with trapezoidal membership function as all processes have “an inescapable element
of variability”. Assuming that some jobs may take a little longer than average is
especially important for electric furnaces, as the tardiness that occurs at this stage
will propagate on the successive stages and finally cause that desired due dates are
missing.
102 6 Application of Fuzzy Theory …

In the initial schedule that is generated all processing times are treated as crisp
numbers, representing the most probable values. The jobs are first sequenced basing
on their criticality, ignoring other constraints, so some constraints may be violated.
The criticality of the jobs is described on the basis of fuzzy logic rules. For example
a rule for the ingots the will be processed on BEST looks as follows [15]:
IF the proportion of BEST-jobs is greater than 0.3 AND
a job J is a BEST-job AND the importance of a job J is high
THEN the criticality of job J is very high
The importance of the jobs is expressed by linguistic variables on the basis of the
data reflecting the customers expectations and the market influence.
In the next step the schedule is evaluated for constraint violations. A schedule
cannot be feasible if the degree of satisfaction of any constraint is below some
critical threshold. The schedule is evaluated according to importance of the jobs and
satisfaction of the constraints that have different weights:

f (S)
  = Ji ∈ S ∧ importance  (Ji ) +    (6.22)
j C j ∈ S ∧ satis f action C j · weight t ype C j

In this step the crisp processing times and desired due dates are replaced by fuzzy
numbers.
Figure 6.4 shows that if the due date x is met within the interval of [dd −1, dd +1],
the membership function will evaluate to A(x) = 1, whereas A(x) = 0.6 for a due
date violation of ±2 h, which can be linguistically described as early/late, and finally
it will evaluate to A(x) = 0.3 for a violation of ±3 h, which can be described as
very early/very late. Outside the fuzzy set it will evaluate to A(x) = 0 for a due
date violation larger then ±4 h. The membership function values in between can
be linguistically described in terms of, e.g., almost very good for 0.7 and so on.
In order to find fuzzy completion times for the jobs with assigned fuzzy due dates
(e.g. for the BEST unit) processing times of all the jobs that precede such jobs must
be fuzziffied. Exemplary definitions of the processing times for electric furnaces,
secondary metallurgy, continuous caster and finally the BEST unit are shown in
Fig. 6.5.
Despite the final schedule is generated in a fairly precautionary manner, some
disturbances in the production process cannot be envisaged, for example if some job
took longer to complete that it was expected, what may result in due date violation.
Such tread can be evaluated and, if necessary, a repair mechanism can be applied.

Fig. 6.4 Membership function evaluation for a due date

6.2 Scheduling of Production Processes Using Fuzzy Logic 103

Fig. 6.5 Fuzzy definition of processing times on different production units

Tabu search algorithm is used to find which jobs should be exchanged to repair the
schedule. In order to limit the number of necessary re-scheduling thresholds are
defined to measure significance of adverse events and to assess their impact on the
final schedule evaluation (i.e. whether some constraints may remain violated).
More complex production process system of the highly specialised steel in the
LD3 plant in Linz was analysed by Dorn in [16]. The overall flow of the steel
production is shown in Fig. 6.6. The blast furnace (BF) delivers the pig iron either
to the desulfurization unit (DS), to the mixer (MX), or directly to one of the three
converters (CV7, CV8, CV9), however only two of them are usually used. The mixer
is a 2000t container that plays a role of a stock between BF and the converters. There
are two typical production routes: the steel from CV7 is poured into a ladle and
delivered to the ladle furnace (LF) and later to the single-stranded continuous caster
CC3, and the steel from CV9 is delivered to the conditioning stand (CS) and to the
two-stranded caster CC4. On both routes also the vacuum-degassing unit (VAC) can
be involved. About 90 % of the cast slabs are delivered either hot to the hot rolling
mill (HRM) or stored in the slab stock (SS). The remaining jobs are produced for
the forge and the foundry [16].
Like in the previous studies, the steel grade compatibility constitutes the main
group of constraints for the blast furnace and continues casters. The author noticed,
that there are two kind of such constraints: hard constraints saying that two grades

Fig. 6.6 Steel production process analysed by Dorn in Linz LD3 steelworks [16]
104 6 Application of Fuzzy Theory …

may not be sequenced, and soft rules that say that it is not so good to sequence two
grades. Even if the rules are hard, some additional operation may be performed to
process virtually any sequence of jobs that is scheduled. However, such operations
(inserting a plate into the strand, changing of the tundishes, re-set-up of the caster)
are very costly and/or time consuming. Hot rolling technology requires also the strict
coordination with the steelmaking process, as the slabs cannot cool down below a
certain temperature, as no reheating of slabs is assumed. The vacuum-degassing
unit has also some important constrains, as it is used in both technological routes and
requires frequent maintenance (usually after four jobs). Like in the previous solutions
the constraints are represented using fuzzy sets. Timeliness is expressed as linguistic
variables and takes values like very early, in time, etc. The same is with the quality
description (e.g. bad, very high), lower limit (e.g. zero, few, too many), amount (e.g.
few, sufficient, too much), priority (e.g. normal, very important) and the sequence
length on the casters (e.g. very short, too long). Some values of such variables are
not allowed in a valid schedule if the constraint is treated as a hard one. For example
if a group casting jobs last too long this violates the necessary maintenance of the
tundish and cannot be accepted. In order to better deal with soft constraints when
searching for a valid schedule Dorn introduced weights for each type of constraints.
For example the most costly set-up of continuous caster has weight of 1.0, while due
dates or compatibility violations are weighted as 0.6 and undesirable format changes
only as 0.2.
The schedule is evaluated on the basis of weighted constraints satisfaction and
importance of jobs, which is also expressed by fuzzy sets (similar to the system that
has been described in the beginning of this section). The finale value of schedule
estimation is always between 0 and 1, where 1 means the optimal value. Schedules
can be rated as good if the evaluation is above 0.9.
If the schedule needs to be repaired due to some unpredicted disturbances in
the production process, again Dorn suggest using tabu search algorithm and repair
heuristic as he found them a very effective one if the problem is not too complex.
However, if the constraints cannot be easily satisfied due to more restrictive tech-
nological and organisational dependencies and/or more complex production routes
(in this case taking into account three converters instead of two) the performance of
the system may significantly decrease [16]. To face this problem Dorn propose to
take advantage of a case-based reasoning system. The overview of such system is
depicted in Fig. 6.7.
The core of the system is the database with valuable cases that occurred in the past
and can be used to solve similar, not necessary identical, current problem encountered
in the scheduling process. Similar old problems found in the database need to be
adopted to the current problem and after the knowledge gained from them (usually
in the form of rules) can be applied to reduce the complexity of the current scheduling
problem. No matter whether it will bring success or failure, the case is also stored in
the database for the future reasoning. The cases are usually described as the triple:
problem, situation (context) and solution. For example, a case in which converter
CV7 was replaced by the converter CV8 was described as:
6.2 Scheduling of Production Processes Using Fuzzy Logic 105

Fig. 6.7 Case reasoning system for fuzzy reactive scheduling [16]

problem: break down of CV7 for some hours

situation: many jobs on CV7, free CV8
solution: exchange CV7 with CV8 in schedule
Note that in description of the cases fuzzy sets are used to soften the constraints. This
can be better seen in the description of the following case respecting the shortage of
pig iron:
problem: very few pig iron
situation: medium set-ups on CC3, few set-ups on CC4,
many quality separations on CC3, medium quality separations on CC4,
few tundish changes on CC3, medium tundish changes on CC4,
some important jobs on CC4
solution: increase weight set-ups CC3, increase weight important jobs
strong decrease weight set-ups CC4, optimize schedule
The above case shows also an interesting approach to solve the problem. Instead of
finding a new sequence of jobs on casters that requires less steel that may be very
time consuming, only the weights for those casters are updated and the schedule is
re-optimised.
A crucial stage in the inference process is to find cases that are similar to the
current problem, as it is necessary to use some mechanism to express and measure
the similarity. If a problem is described by a set of events E and a set of propositions
P, E i and Pi represent the set of events and the set of propositions for the case Ci ,
the following formula is used to determine the similarity of the problem to the case:

S(Ci ) = α (1 − |δ(e j , E i )Θδ(e j , E)|)
 ∀e j ∈(E∩Ei ) 
⊕β (1 − δ(e j , E)) ⊕ χ (1 − δ(e j , E i )) (6.23)
∀e j ∈(E i −E) ∀e j ∈(E−E i )

⊕γ (1 − |δ( p j , Pi )Θδ( p j , P)|)
∀ p j ∈Pi

where δ is the function that tells if the event or proposition j exists in the cur-
rent problem or the case (1 for crisp relationships, fuzzy number, otherwise), α is
106 6 Application of Fuzzy Theory …

the weight of the cases that events are contained in both the case and the problem
E ∩ E i , β is the weight of the events that are in the old case, but not in the current
problem E i − E (rather not desirable, so β should be less than 0), χ is the weight of
the events that are in the current problem, but not in the old case E − E i , and finally
γ is representing the weight of the propositions existing in both the problem and the
case. The calculation of final value of the similarity expressed as fuzzy numbers, so
fuzzy operators: addition ⊕ and subtraction  are used.
Despite the idea of the system seems to be very promising only a prototype of the
system has been implemented in practise.
Another system basing on knowledge base and fuzzy logic was proposed by
Adenso-Diaz et al. [9]. The proposed system was dealing with planning in a roll
shop and was implemented in one of the Spanish steelworks. Roll shops serve for
storing and maintenance of rolls used in hot rolling and cold rolling mills. Rolls
depend on their planned location in the mill (type of the mill, a stand in the mill) and
the type of the product that can be rolled using them. There was about 150 different
roll types in the analysed steelworks. The authors developed a fuzzy system to plan
the work in the roll shop by determining the priority of the rolls. The system contains
over 50 variables that can be divided into following four groups:
• Frequency of incidents: An undetermined number of rolls suffer certain processing
problems and never get to the stand; other rolls have a lamination time lower than
that accepted as normal.
• Mill/Stand importance: Some stands or mills may be more conflictive or key than
others, due to the frequency of incidents or the importance/urgency of the orders.
• Estimation deviation: Deviation between the estimation of rolls to be sent to the
mills and the actual rolls sent.
• Optimum deviation: The most important variable for priority evaluation. This is
the roll shop time needed to get the optimum number of rolls with the current use
of resources.
The proposed system has two main modules. The goal of the first is to gather data
from the database and to transform it into the four types of inputs presented above.
The second is the fuzzy system itself, which provides priority of processing in the
roll shop for every roll type. The inputs are fuzzified and 39 fuzzy rules are used
to produce the final output that after centroid defuzzification is presented as a crisp
number. In membership functions linguistic variables are used. Exemplary rule looks
as follows
IF Optimum_deviation is medium AND Mill/stand_importance is high
THEN Priority is high
The system has been implemented in practise and, according to the authors, three
main benefits from its application can be enumerated:
• decrease in costs, mainly stock and warehousing costs,
• increase in security, because the system attenuates the information handicap,
• general improvement in quality.
6.3 Cooperation of Planning with Fuzzy Logic 107

6.3 Cooperation of Planning with Fuzzy Logic

Dorn and Kerr [17] proposed to apply cooperative scheduling to optimise schedules
in a steelmaking plant and a rolling mill for VA Stahl steelworks in Linz (Austria).
In this approach fuzzy logic was used to perform so called fuzzy communication
in order to agree two separate schedules in different plants that may have different
technological and organisational constraints. As it was already mentioned, the main
group of constraints for the steelmaking plant is related to the compatibility: the
steel grade of consecutive heats must be similar, castings width must stay within
a given range and the degassing procedure in the secondary metallurgy must be
consistent. Also for the rolling mill, subsequent jobs must have similar steel quality,
but different constraints on width and thickness of rolled goods must be considered.
Another problem is that not all steel qualities can be rolled warm, so some slabs must
cool down before they can be processed in the rolling mill.
Scheduling process works in the proposed system works in such a way that the
schedule generated for one plant is evaluated and sent to the second plant. Instead of
firm constraints, fuzzy sets and so called fuzzy sequences are used. For example the
schedule generated for the rolling mill consists of a list of jobs, a number of fuzzy
sets describing windows, when casting can be performed, and a set of fuzzy sequence
constraints that constrain the sequence of jobs. Figure 6.8 shows that job j1 should
be done first, job j12 should be rolled in the end, as it has the lowest weight and job
j2 should be performed sometime between job j1 and j12 .
The subsystem in the steelmaking plant adds its own constraints, evaluates the
schedule and repairs it to achieve better evaluation using a tabu search procedure.
After that, the schedule is sent back to the rolling mill plant. If the evaluation of the
new schedule is worse than before, the subsystem in the rolling mill plant tries to
repair the schedule. A repaired version of the schedule is sent again to the steelmaking
plant and if its evaluation is better than it was earlier, it is accepted, otherwise it is
rejected.
According to Dorn despite that the proposed system was not tested in practise,
fuzzy communication process enables for sharing more information than in the tradi-
tional crisp data communication. The prototype of the system itself has not been built
according to multi-agent architecture, and communication is done sequentially, what
may limit the potential benefits from application of fuzzy sets and fuzzy systems for
production scheduling of several plants working in a common supply chain.

Fig. 6.8 Fuzzy sequence of rolling mill jobs

108 6 Application of Fuzzy Theory …

Fig. 6.9 A fuzzy multi-agent system the planning and scheduling in steelworks

Fazel Zarandi and Ahmadpour [2] proposed a much more complex solution for
the cooperative planning and scheduling in the steel industry. Their system is based
on multi-agent architecture and adaptive neuro-fuzzy networks. An overview of the
system is shown in Fig. 6.9.
The system consists of six agents. Agents are built around several modules: user
interface, communication interface, reasoning module that uses knowledge base and
learning module. Customer agent is responsible transforming customers’ orders into
production orders after negotiations with a User agent. The User agent deals with
customers’ orders and communicates with other agents to process them. The role of
the Ingot Casting (IC) agent is to supply ingots produced in ingot moulds to the User
agent. The Vacuum Degassing (VD) agent is responsible for determining parameters
of degassing process. The Ladle Furnace (LF) agent is responsible for determining
parameters of molten steel refining process according to the customer’s order. Finally,
the Electric Arc Furnace (EAF) agent supplies molten steel to the LF agent.
The User agent has to assess the order sent from the Customer agent whether it
is a feasible one, taking into account cost of necessary additives and cost of delivery
date differences. If the order receives a status of “infeasible” status the User and the
Customer agents may negotiate different delivery date and/or required properties.
The IC agent determines the amount of molten steel that is necessary to produce the
ingots ordered by the customer and provides the estimation of total casting process-
ing time for the order. Both VD and LF agents have to deal with uncertain parameters
of the production process. The VD agent must determine such parameters as vacuum
pressure, the amount of neutral gas (Argon) and processing time. For this purpose, an
adaptive neuro-fuzzy inference system (ANFIS) has been used [18]. The input data
includes the amount of hydrogen (H ), input temperature (T 1) and output tempera-
ture (T 2). The output data includes all necessary parameters. After training stage,
fuzzy rules are saved for each of the output parameters. Exemplary rules for vacuum
pressure (out of total 11) look as follows:
IF H is very low AND T1 is very high AND T2 is rarely low
THEN Pv is ( − 0.1848 ∗ H − 0.6485 ∗ T 1 + 1.8305 ∗ T 2 − 178.1581)
6.3 Cooperation of Planning with Fuzzy Logic 109

IF H is rarely low AND T1 is very strongly low AND T2 is moderate

THEN Pv is (3.1852 ∗ H − 4.1816 ∗ T 1 + 4.2968 ∗ T 2 + 7.7689)
IF H is rarely high AND T1 is moderate AND T2 is very high
THEN Pv is (0.5049 ∗ H + 0.9166 ∗ T 1 + 0.4582 ∗ T 2 − 78.2941)
After learning phase total 54 rules have been stored in the knowledge base.
The Ladle Furnace agent must determine such parameters of the refining process
as the amounts of additives and the processing time. Also in this case ANFIS system
is used to collect rules in the knowledge base. The input data includes the chemical
properties of input molten steel (S2, Cr2 and Ni2), temperature and required chemical
properties of output molten steel. The output data includes the amounts of needed
additives to change the chemical properties of molten steel, and processing time.
Exemplary rules for processing time (out of 12) are shown below:
IF S2 is rarely low AND Cr2 is moderate AND Ni2 is high
AND Temperature is rarely high THEN Time is
(0.5138 ∗ S2 − 0.0357 ∗ Cr 2 − 1.2426 ∗ N i2 + 0.0843 ∗ T emperatur e −
106.6996)
IF S2 is moderate AND Cr2 is very high AND Ni2 is rarely low
AND Temperature is rarely low THEN Time is
(0.1389 ∗ S2 + 1.6269 ∗ Cr 2 − 1.4204 ∗ N i2 + 2.4173 ∗ T emperatur e −
4145.3457)
IF S2 is low AND Cr2 is very low AND Ni2 is strongly high
AND Temperature is very low THEN Time is
(− 13.0052 ∗ S2 + 61.592 ∗ Cr 2 − 296.1599 ∗ N i2 − 0.7135 ∗ T emperatur e +
307.2859)
This time 111 rules have been generated after learning phase.
The system was implemented in MATLAB7. A single agent was used for every
process, however in real application more agents can work for a single process. The
authors have not provided any information about the plans on implementation of the
proposed system in real steelworks.
Later Zarandi and Azad [18] presented a multi-agent system based on Type-2
fuzzy sets. The authors enumerate four sources of uncertainty about membership
functions used in classical fuzzy sets:

• different meaning of the same word to different people,

• experts opinions may have histogram of values, especially if they vary,
• measurements activating a type 1 fuzzy logic system may be noisy,
• the data used to tune the parameters of a type 1 fuzzy logic system may also be
noisy.

In the Type-2 fuzzy logic system rules look as follows [19]:

IF x1 is of type F̃1 AND … AND x p is of type F̃ p THEN y is of
type Ỹ
110 6 Application of Fuzzy Theory …

The scheduling system was built using a combination of fuzzy programming and
fuzzy contract net protocol proposed by Li et al. in [20]. In such approach an Order
agent transfers an order to a Scheduling agent and when the order arrives, the Schedul-
ing agent models a scheduling programming problem and then selects an algorithm
to solve the problem (e.g., some heuristic like genetic algorithm). The Scheduling
agent must also adjust the initial schedule to a new condition that occurred during
the production process, like machine breakdown or appearance of a new task. In the
fuzzy programing stage a Type-2 fuzzy flow shop problem formulation is used to find
the sequence of the n jobs maximising the total satisfactory degree. Type-2 fuzzy
due date is represented by the degree of satisfaction with respect to the completion
time c j , and denoted by a doublet:

D̃ j (d Lj , d Uj ) (6.24)

where d Uj is not fixed, since penalties for the delay of the jobs are defined differently
by the experts as:
d Uj = d Lj + ρk (6.25)

in which penalty
ρk ∈ [ρk1 , ρk2 ].

Figure 6.10 shows the fuzzy due date and the membership function of satisfactory
degree μ j (c j ). The objective function is then defined as maximising total satisfactory:
n
max f sum = μ∗j (c j ) (6.26)
j=1

where μ∗j is the final crisp value of the interval set of a due date set calculated as the
average of upper and lower bound.
When an emergency event occurs fuzzy, a contract protocol is used. First, the
Scheduling agent builds some Task agents, each for one or more jobs. The Task
agent announces the job (or group of jobs) to the Resource agents inviting them to
bide for it. After receiving the answers, the Task agent selects the best Resource
agent by evaluating the received bids and it is removed from the system [20]. Each
agent has a fuzzy module that includes fuzzifier, rule base, fuzzy inference engine,
and output processor (type-reducer and defuzzifier). A detailed structure of the fuzzy
module is shown in Fig. 6.11.

Fig. 6.10 Fuzzy due date

and the membership function
of satisfactory degree
6.3 Cooperation of Planning with Fuzzy Logic 111

Fig. 6.11 Fuzzy due date and the membership function of satisfactory degree

Fig. 6.12 Aggregation of

two fuzzy experts’ opinions

The fuzzifier changes the crisp values of the parameters into three linguistic vari-
ables (e.g., for completion time: early, medium, late; for priority: high, medium,
low; for due date: long, medium, short). Those linguistic variables can be defined as
Type-2 membership function, shape of which is obtained basing on experts’ opinions,
which are written as Type-1 fuzzy numbers. Aggregation of two experts’ opinions
into Type-2 membership function is shown in Fig. 6.12.
The contract fuzzy protocol uses fuzzy rules. For example in announcement stage
following rules are used:
IF due date is short THEN priority of announcement is high
IF due date is medium THEN priority of announcement is medium
IF due date is long THEN priority of announcement is low

6.4 Continuous Caster Scheduling Based on Fuzzy

Lot-Sizing

A scheduling agent that uses a fuzzy version of a capacitated lot-sizing problem

(CLSP) is presented in this section. Mathematical integer programing (MIP) models
basing on capacitated lot-sizing model and its variations [21] is one of the approaches
used for planning and scheduling in steel industry. Wang and Tang [22] presented
112 6 Application of Fuzzy Theory …

a model for hot rolling production scheduling. The objective function in this model
includes three kinds of costs:
• the cost of coils changing,
• the earliness and tardiness of delivering coils,
• and the unused capacity of the rolling mill.
The authors proposed a tabu search algorithm to solve the scheduling problem and
tested its performance using data from Baosteel plant in China. More complex model
has been proposed by As’ad and Demirli [23]. The objective functions includes
various types of costs, such as:
• ordering and purchasing of raw materials,
• inventory holding costs,
• setup costs,
• regular and overtime production costs,
• holding costs of finished products and backorder costs.
The authors proposed to use some approximations schemes in order to simplify the
original problem. Finally Mattik [24] presented some models for integrated schedul-
ing of continuous casters and hot strip mills.
A CLSP model presented by the author for continuous casting can be the basis
for development of a fuzzy model used by the Scheduling agent. A modified version
of the model is defined as follows:


min Z̃ it = λ1 ( I˜mt
c−
+ I˜mt
c+
)
m∈M t∈T
    
+ λ2 c c
stlm xlmt + λ3 ( csmc Z mt
c
/Ctc ) (6.27)
l∈L m∈M|l=m t∈T m∈M t∈T

constraints:

I˜m,t−1
c+
− I˜m,t−1
c−
+ Z mt
c
− I˜mt
c+
+ I˜mt
c−
= d̃mt
c
, m ∈ M, t ∈ T (6.28)

  
c
Z mt csm + c c
stlm xlmt ≤ Ctc , t ∈ T (6.29)
m∈M l∈L m∈M|l=m

csmc Z mt
c
≤ Ctc ymt
c
, m ∈ M, t ∈ T (6.30)
c
xlmt ≥ yl,t−1
c
+ ymt
c
− 1, l ∈ M, m ∈ M|l = m, t ∈ T (6.31)


c
ymt = 1, t ∈ T (6.32)
m∈M
6.4 Continuous Caster Scheduling Based on Fuzzy Lot-Sizing 113

c
Z mt ≥ ml c ymt
c
, m ∈ M, t ∈ T (6.33)
c
xlmt ∈ {0, 1}, l ∈ M, m ∈ M|l = m, t ∈ T (6.34)
c
ymt ∈ {0, 1}, m ∈ M, t ∈ T (6.35)
c
Z mt ≥ 0 m ∈ M, t ∈ T (6.36)

Indices/sets:

t∈T time period;

m∈M steel grades;

Parameters:

d̃mt demand for steel grade m in period t;

csi capacity consumption for item I ;
gik = 1 if item i is produced from steel grade k, otherwise 0;
st k setup cost for steel grade k;
Ct capacity a heat in period t;
Variables:

I˜it+ , I˜it− items i delayed (–) and stored (+) at the end of period t;
z tk = 1 if there is a setup (resulting from a change in grade) in period t;
ytk = 1 if steel grade k is produced in periodt, otherwise 0;
xit number of items i produced in period t.

It should be noted that the above model does not include many technological
constraints such as, e.g., adequate slabs width in a sequence or their temperature.
Such model should then be considered as a first stage of the continuous casting
production planning process, in which the quantities of particular steel grades are
determined. In the seconds stage jobs on continuous casters has to be sequenced with
respect to the technological constraints. As it was shown in the previous sections,
such constraints may also include some fuzzy rules, due to uncertainty of the casting
process. The same concerns the steel grade, despite the model assumes predetermined
steel grades for the slabs, this can be easily change with the use of fuzzy logic, saying
which steel grade maybe assigned to satisfy customers’ requirements.
Lot-sizing with fuzzy parameters is rather rarely studied in the literature, espe-
cially when compared to fuzzy shop scheduling problems. Yan et al. [25] analysed lot-
sizing production planning problem with profits, customer demands and production
capacity characterised by fuzzy variables with trapezoidal membership functions. To
solve the problem the authors proposed to apply a standard genetic algorithm per-
forming fuzzy simulation. Rezaei and Davoodi [26] studied a lot-sizing problem with
supplier selection under fuzzy demand and costs (price, transaction cost and hold-
ing cost) with triangular membership functions. Also in this case a standard genetic
algorithm was used to determine upper and lower bound for production quantities.
114 6 Application of Fuzzy Theory …

Most recently Sahebjamnia and Torabi [27] considered a multi-level capacitated lot-
sizing problem with uncertain setup, holding, and backorder costs expressed as a
fuzzy numbers with trapezoidal membership functions. They proposed a heuristic in
which uncertain constraints as well as imprecise objective functions are converted
into the crisp values by using the expected interval and value of the ill-known para-
meters, respectively. Then the authors solved such a problem using a standard MIP
solver.
When planning continuous casting process, the demand for a given period of
time is usually taken as a strict constraint. However strict deadlines are necessary
mainly for the cases when a minimal temperature of slabs is required in further
processing (hot rolling). It is especially true for direct hot charge and hot charge
rolling technologies (see [1] for the classification of hot rolling technologies). On
the other hand the demand for a given steel grade depends on the delivery dates that
were confirmed to the customer. Thus in our solution we introduced rules telling if
the desired term for steel grade (demand in a give period) may vary within some
limits. The exemplary rules are as follows:
IF rolling temperature constraint is obligatory THEN demand for grade
is strict
IF due date is short THEN demand for grade is strict
IF rolling temperature constraint is strong THEN demand for grade is
tight
IF due date is medium THEN demand for grade is tight
IF rolling temperature constraint is weak THEN demand for grade is
normal
IF due date is long THEN demand for grade is normal
A linguistic variable demand for grade is mapped for a fuzzy demand with a triangular
membership function. For the strict demand membership function is defined as:
(dmt *0.99, dmt , 0), for tight demand it is defined as: (dmt *0.95, dmt , dmt *1.05), and
for normal demand (dmt *0.9, dmt , dmt *1.1).
To solve the fuzzy capacitated lot-sizing problem for continuous casting schedul-
ing a dedicated genetic algorithm was used. Contrary to the problems presented in
the literature, in the computational experiment we consider a scheduling problem
of an industrial-size, so a standard genetic algorithm cannot be applied effectively
to receive acceptable solutions in a reasonable time. Instead, we applied a genetic
algorithm with two dedicated repair algorithms, described in [28] and adopted to the
fuzzy scheduling of continuous caster. A chromosome representing the solution con-
sists of the values of xit variables. A non-standard crossover operator, copying one
period from a parent’s chromosome to a child’s chromosome, and irregular mutation
[29] were used as recombination operators.
6.4 Continuous Caster Scheduling Based on Fuzzy Lot-Sizing 115

The following parameters of GA where used:

• crossover rate c R = 0.8,
• mutation rate m1 R = 0.05,
• population size popsi ze = 50 solutions.
The algorithm has been tested on a problem that includes 50 different products,
10 different steel grades and a time period with 48 discrete time slots (hours). The
minimum size of a cast was set to 10 ton. The other parameters have been generated
from a uniform distribution:

• demand for steel in period t − dmt ∈ [100, 300] tons,

• capacity consumption csmt ∈ [0.1, 0.3] min/ton,
• setup time for changing steel grade stlm ∈ [30, 60] min.

As it has been already explained, the demand for the particular steel grade is fuzzified
according to the values of linguistic variables in two variants. In the tight variant 30 %
of the steel grades had a strict demand, 30 % had a tight demand and the remaining
ones had normal demand. In the loose variant 5 % of the steel grades had a strict
demand, 25 % had a tight demand and the remaining ones had normal demand.
The genetic algorithm has been run for 20 times for the same variant of the demand
and the average results for the best solution in the population after 20,000 generations
(ca. 1 min for a single run) have been collected. The results are shown in Table 6.1.
The capacity utilisation of continuous caster is provided together with the produc-
tion costs, including the holding costs and the penalty for tardiness and setup costs.
As it was expected, loosening the demand constraints brings a significant decrease
in production costs (up to 50 %), but also a noticeable increase in capacity utilisation
(up to 12 %). Even though in the real production systems such advantages cannot be
achieved due to technological and organisational constraints that have been omitted
in the proposed lot-sizing model, it has been shown that the application of fuzzy set
theory to planning and scheduling of steel production processes may bring significant
savings.

Table 6.1 The average values of capacity utilisation and cost function for different variants of CC
scheduling
Capacity utilisation [%] Cost function [penalty points]
Crisp scheduling 0.75 14694
Fuzzy scheduling with thigh 0.81 10390
variant
Fuzzy scheduling with loose 0.84 9833
variant
116 6 Application of Fuzzy Theory …

References

1. Tang, L., J. Liu, A. Rong, and Z. Yang. 2001. A review of planning and scheduling systems and
methods for integrated steel production. European Journal of Operational Research 133(1):
1–20.
2. Zarandi, M.H.F., and P. Ahmadpour. 2009. Fuzzy agent-based expert system for steel making
process. Expert Systems with Applications: An International Journal 36(5): 9539–9547.
3. Dorn, J. 1996. Expert systems in the steel industry. IEEE Expert 11(1): 18–23.
4. Vasko, F.J., K.L. Stott, F.E. Wolf, and L.R. Woodyatt. 1989. A fuzzy approach to optimal
metallurgical grade assignment. In Applications of fuzzy set methodologies in industrial engi-
neering, ed. G.W. Evans, W. Karwowski, and M.R. Wilhelm, 285–298., Advances in Industrial
Engineering Amsterdam: Elsevier Science Publishing.
5. Woodyatt, L.R., K. Stott, F.E. Wolf, and F.J. Vasko. 1992. Using fuzzy sets to assign metallur-
gical grades to steel. JOM, The Journal of The Minerals, Metals & Materials Society (TMS)
44(2): 28–31.
6. Wang, M.-J.J., and T.-C. Chang. 1995. Tool steel materials selection under fuzzy environment.
Fuzzy Sets and Systems 72(3): 263–270.
7. Chen, S.-H. 1985. Ranking fuzzy numbers with maximizing set and minimizing set. Fuzzy Sets
and Systems 17(2): 113–129.
8. Chen, S.-M. 1997. A new method for tool steel materials selection under fuzzy environment.
Fuzzy Sets and Systems 92(3): 265–274.
9. Adenso-Díaz, B., I. González, and J. Tuya. 2004. Incorporating fuzzy approaches for production
planning in complex industrial environments: The roll shop case. Engineering Applications of
Artificial Intelligence 17(1): 73–81.
10. Mauder T., J. Štětina, and M. Masarik. 2013. On-line fuzzy regulator for continuous casting
process. In In:Proceedings of 22nd International Conference on Metallurgy and Materials
METAL, 23–38, Brno, Czech Republic, 15–17 May 2013.
11. Zengchang, Qin, and Yongchuan, Tang. 2014. Uncertainty modeling for data mining: A label
semantics approach. Springer.
12. Moraga, C. 2005. Introduction to fuzzy logic. Facta universitatis - series: Electronics and
Energetics 18(2): 319–328.
13. Mamdani, E.H., and S. Assilian. 1975. An experiment in linguistic synthesis with a fuzzy logic
controller. International Journal of Man-Machine Studies 7(1): 1–13.
14. Larsen, M.P. 1980. Industrial applications of fuzzy logic control. International Journal of
Man-Machine Studies 12(1): 3–10.
15. Dorn, J., R.M. Kerr, and G. Thalhammer. 1995. Reactive scheduling: Improving the robust-
ness of schedules and restricting the effects of shop floor disturbances by fuzzy reasoning.
International Journal of Human-Computer Studies 42(6): 687–704.
16. Dorn, J. 1995. Case-based reactive scheduling. In Artificial intelligence in reactive scheduling,
ed. E. Szelke, and R.M. Kerr, 32–50. Kluwer Academic Publishers.
17. Dorn, J. and R.M. Kerr 1994. Co-operating scheduling systems communicating through fuzzy
sets. In Preprints of the 2nd IFAC/IFIP/IFORS-Workshop on Intelligent Manufacturing Systems
IMS94, 367–373, Vienna.
18. Zarandi, F.M.H., and K.F. Azad. 2013. A type 2 fuzzy multi agent based system for schedul-
ing of steel production. In IEEE Joint IFSA World Congress and NAFIPS Annual Meeting,
IFSA/NAFIPS, 992–996, Edmonton, Alberta, Canada, 24–28 June 2013.
19. Castillo, O., and P. Melin. 2008. 5 Design of Intelligent Systems with Interval Type-2 Fuzzy
Logic, vol. 223, Studies in Fuzziness and Soft Computing Berlin: Springer.
20. Li, Y., J.-Q. Zheng, and S.-L. Yang. 2010. Multi-agent-based fuzzy scheduling for shop floor.
The International Journal of Advanced Manufacturing Technology 49(5–8): 689–695.
21. Helber, S. 1995. Lot sizing in capacitated production planning and control systems. Operations-
Research-Spektrum, 17(1): 5–18.
References 117

22. Wang, X., and L. Tang. 2008. Integration of batching and scheduling for hot rolling production
in the steel industry. The International Journal of Advanced Manufacturing Technology 36(5–
6): 431–441.
23. Asad, R., and K. Demirli. 2010. Production scheduling in steel rolling mills with demand
substitution: Rolling horizon implementation and approximations. International Journal of
Production Economics 126(2): 361–369.
24. Mattik, I. 2014. Integrated scheduling of continuous casters and hot strip mills: A block plan-
ning application for the steel industry. Produktion und Logistik: Gabler Verlag.
25. Yan, W., J. Zhao, and Z. Cao. 2005. Fuzzy programming model for lot sizing production
planning problem. In Fuzzy systems and knowledge discovery, ed. L. Wang, and Y. Jin, 285–
294, Lecture Notes in Computer Science Berlin: Springer.
26. Rezaei, J., and M. Davoodi. 2006. Genetic algorithm for inventory lot-sizing with supplier
selection under fuzzy demand and costs. In Advances in applied artificial intelligence, ed. M.
Ali, and R. Dapoigny, 1100–1110, Lecture Notes in Computer Science Berlin: Springer.
27. Sahebjamnia, N., and S.A. Torabi. 2014. A fuzzy stochastic programming approach for multi-
level capacitated lot-sizing problem under uncertainty. In Recent developments and new direc-
tions in soft computing, volume 317 of Studies in fuzziness and soft computing, ed. L.A. Zadeh,
A.M. Abbasov, R.R. Yager, S.N. Shahbazova, and M.Z. Reformat, 393–407. Springer Interna-
tional Publishing.
28. Duda, J. 2005. Lot-sizing in a foundry using genetic algorithm and repair functions. In Evolu-
tionary computation in combinatorial optimization, ed. G.R. Raidl, and J. Gottlieb, 101–111,
Lecture Notes in Computer Science Berlin: Springer.
29. Michalewicz, Z., and C.Z. Janikow. 1991. Genetic algorithms for numerical optimization.
Statistics and Computing 1(2): 75–91.
Chapter 7
Application of Fuzzy Decision Trees
in Analog Forecasting

Abstract This chapter presents a new method for forecasting the level and struc-
ture of market demand for industrial goods. The method employs two data mining
methods: k-means clustering and fuzzy decision tree learning. The k-means method
serves to separate groups with items of a similar consumption level and structure
of the analysed products. Whereas, fuzzy decision tree learning are used to deter-
mine the dependencies between consumption patterns and predictors. The proposed
method is verified using the extensive statistical material on the level and structure
of steel products consumption in selected countries in the years 1960–2010.

Analog forecasting is a method of forecasting a given variable by using information

about the behaviour of another variable whose changes over time are similar, but not
simultaneous. In this chapter a new method for forecasting the level and structure of
demand based on the concept of analog forecasting is proposed. The method uses the
concepts underlying historical and geographical analogies. It involves the estimation
of the level and structure of demand based on the analysis of historical data and the
use of data clustering and decision trees. The k-means method is used to separate
groups of items having similar consumption level and similar structure of the analysed
products (consumption patterns), whereas the fuzzy ID3 algorithm is used to build
a fuzzy decision tree, i.e., to determine dependencies between consumption patterns
and predictors (parameters determining the level and structure of consumption).
The proposed method can primarily be used to forecast the level and structure
of demand for products that are traded on the industrial market. The proposed con-
cept can, therefore, be used to forecast demand for products such as: steel industry
products, non-ferrous metals industry products, certain chemical industry products,
casts, construction materials industry products, energy carriers.
The investigations focus on the long-term forecasting method of demand for steel
products, in which case forecasts for the level and structure of apparent consumption
of steel products are produced. Apparent consumption is calculated as the difference
between domestic production of the analysed range of steel products and the con-
sumption of these products for further processing in the steel industry. This calculated
difference is adjusted by the balance of foreign trade for these products [1].
120 7 Application of Fuzzy Decision Trees in Analog Forecasting

The presented concept, which has been verified on the example of steel prod-
ucts, can be applied, after modifications, to forecasting the demand for the above-
mentioned products, however, certain conditions must be met:
• historical data regarding the level and structure of consumption must be available,
• data regarding the value of predictors, which define the level and structure of
demand for a given product, must be determined and available.

7.1 Methods for Forecasting Apparent Consumption

of Steel Products

The following methods are used for forecasting apparent consumption of steel prod-
ucts:
• econometric models,
• sectoral analysis,
• trend estimation models,
• analog methods.
In econometric models, the level of apparent consumption of steel products is
a function of selected macroeconomic parameters. Typically, gross domestic product
(GDP), GDP composition, the value of investment outlays, and the level of industrial
production are used as exogenous variables [1–6]. And often in econometric models
GDP steel intensity is used as the endogenous variable. The term GDP steel inten-
sity was defined by Malenbaum [7] as the amount of steel products consumed per
one unit of GDP. The dependence of GDP steel intensity on selected macroeconomic
parameters was used to forecast steel consumption in selected countries [1, 2]. Evans
analysed GDP steel intensity of Britain’s GDP in the postwar period. He accepted that
GDP steel intensity depends on the share of industry and construction in making up
the GDP, high technology industries of the GDP produced by the industry as a whole,
the scope of use of material-saving technologies in industry and construction, and
the substitution of steel with other materials.
Tilton [8] applied the sectoral analysis to forecast steel products consumption.
It involves forecasting steel intensity indicators for selected sectors of the economy
and the GDP generated in those sectors. Crompton [9] used the method proposed by
Tilton for the analysis and forecast of steel consumption in Japan. Tilton’s method
was also used for forecasting steel consumption in the world [6], in the U.S. [5] and
in Poland [1].
Apparent consumption of steel products was also forecasted by use of trend esti-
mation models [10]. The forecasts produced by these methods are usually short-term
forecasts.
Analog methods are also used for forecasting apparent consumption of steel prod-
ucts. They are based on the assumption that the indicators characterising the level
and structure of the apparent consumption of steel products in the country for which
7.1 Methods for Forecasting Apparent Consumption of Steel Products 121

the forecast is drawn up tend to attain the indicators characterising selected countries
that serve as comparator. The indicators most often compared include consumption
of steel products per capita, GDP steel intensity, and the assortment structure of
apparent consumption [1].

7.2 Methodology

During the last decade, data mining methods have become very popular tools sup-
porting decision-making processes. These methods can be used to search for patterns
in large data sets (data clustering) or for assigning new objects to existing patterns
(data classification). As a result, they can be used to characterise the dependencies
between the level and structure of demand (consumption patterns) and selected para-
meters (numerical or qualitative), which in this case perform the role of predictors.
Such dependencies, if existing, can be used to forecast the level and structure of
demand.
There are various data mining methods that can be used for building classification
models. Among statistical data mining methods, the linear discriminant function
model has been widely used in solving many practical problems [11]. However, the
effectiveness of this method deteriorates when the dependencies between forecasted
values and exogenous variables are very complex and/or non-linear [11]. Situations
of this kind are often encountered in practice. In such cases the machine learning
methods are more appropriate. Examples include algorithms for generating decision
trees, neural networks, Bayesian networks, and genetic algorithms [10]. Each of these
methods leads to different knowledge representation.
Among the referred to methods for building classification models, neural net-
works and algorithms for building decision trees are most commonly used [12–14].
The strength of neural networks lies in their many possibilities for generalising infor-
mation contained in the analysed data sets. Decision trees, however, surpass neural
networks in regards to ease of interpretation of obtained results. Rules generated by
decision trees that assign objects to different classes are easy to interpret even for
users unfamiliar with the problems of data mining [15, 16]. This characteristic is very
important, because decision-makers in the industry prefer tools that operate based
on algorithms understandable to all participants in the decision-making process and
provide easily interpretable information.
There are many methods for generating data clusters. In addition to traditional
statistical methods, among which the most popular are hierarchical methods and
methods for optimising the initial division of sets, Kohonen’s Self Organizing Maps
(SOM) and genetic algorithms are also used. Among the methods for optimising the
initial division of sets the most popular is k-means method. The k-means is one of
the simplest unsupervised learning algorithms that solve the well known clustering
problem. The procedure follows a simple and easy way to classify a given data
set through a certain number of clusters (assume k clusters) fixed a priori.
122 7 Application of Fuzzy Decision Trees in Analog Forecasting

Taking of above under the attention, from the very large number of data mining
methods available, two were used in the proposed forecasting method: the k-means
method, used for data clustering, and the Fuzzy Interactive Dichotomizer 3 (FID3),
used to build decision trees (data classification rules). The k-means method creates
consumption patterns, while the FID3 algorithm builds a decision tree that assigns
specific patterns of consumption to exogenous variables (predictors).

7.2.1 The k-means Method

The k-means method is one of the most popular methods for cluster analysis. Given
a set of observations (x1 , x2 , . . . , xn ), where each observation is a d-dimensional real
vector, k-means clustering aims to partition the those observations into k (k  n)
sets S = S1 , S2 , . . . , Sk . The k-means method can also be used to optimise an initial
division of items. The most common k-means algorithm uses an iterative refinement
technique:
1. Randomly, arbitrarily, or using a different criterion select k cluster centres.
2. Calculate the distance between each item and cluster centres.
3. Assign each item to the closest cluster.
4. Determine the centroids of the newly formed clusters.
5. If the stopping criterion is met Stop, otherwise Go To 3.
Subsequent iterations are characterised by the error function of items separation
to individual clusters (SES) defined by the formula:


k 
ni
SES = d 2j Si , (7.1)
i=1 j=1

where:
d 2j Si —the distance of the j-th item from the centroid of the i-th cluster,
n i —the number of items belonging to i-th cluster.
Once the values of SES in subsequent iterations do not show significant changes
(i.e., changes are less than the prescribed value) or when the maximum change of
centroids (the distance of the new centroids from the previous ones) does not exceed
the prescribed value, the procedure will stop (the stopping criterion is met). The
stopping criterion can also be defined as the maximum number of iterations after
which the process is stopped. The distance between centroids or between items is
usually calculated using the Euclidean distance, which for xi and y j being two n-
dimensional vectors, is defined as

 n
  2
d E (xi , y j ) =  xik − y jk . (7.2)
k=1
7.2 Methodology 123

Besides the Euclidean distance, the following two distance measures can also be
used in k-means clustering:
• the root mean square distance (or average geometric distance):

 n
1  2
d R M S (xi , y j ) =  xik − y jk . (7.3)
n k=1

• Minkowski distance, which is a generalisation of Euclidean distance:


 p
  q
d M (xi , y j ) = q  xik − v jk , (7.4)
k=1

where q is a positive integer. For q = 1 the Manhattan distance is obtained and

for q = 2 the Euclidean distance.

7.2.2 Fuzzy Decision Trees

In the literature,several different algorithms of construction of crisp and fuzzy deci-

sion trees can be found. Fuzzy decision trees is the method, which combines fuzzy
sets theory and fuzzy logic.
The most popular method for decision tree learning is ID3 algorithm which was
initially proposed by Quinlan [17] and which was meant for making crisp decision
trees. ID3 algorithm applies to a set of data and generates a decision tree for clas-
sifying the data. It uses the minimal entropy as a criterion at each node. The ID3
algorithm played a large role in the development of algorithms for building decision
trees. The general idea of the tree induction algorithm is as follows:
• A tree starts with a single node representing the entire set of items.
• If all items belong to one decision class, the node becomes a leaf and is labelled
with an appropriate decision.
• Otherwise the algorithm uses the information gain as a criterion for selecting the
attribute that best divides the set of items.
• For the selected attribute a branch is created and the items are divided into the new
nodes (subtrees).
• The algorithm further works recursively for the sets of items assigned to subtrees.
• The algorithm terminates when the stopping criterion is met.
The statistical quantity entropy is applied to define the information gain, to choose
the best attribute from the candidate attributes. The definition of entropy is as follows:
124 7 Application of Fuzzy Decision Trees in Analog Forecasting


N
H (S) = −Pi log2 Pi (7.5)
i=1

where Pi is the ratio of each clusters in set S (the number of elements in the cluster
(subset) i divided by the number of all elements of the set S).
One of a disadvantage of decision tree is its instability. Decision tree is recognised
as highly unstable classifier with respect to minor perturbations in the training data
[18]. The structure of the decision tree may be entirely different if some things change
in the data set. To overcome this problem, some scholars have suggested Fuzzy
Decision Tree [19] by utilising the fuzzy set theory to describe the connected degree
of attribute values, which can precisely distinguish the deference of subordinate
relations between different examples and every attribute values [20].
The simple generalisation of ID3 algorithm on fuzzy numbers is so-called Fuzzy
Interactive Dichotomizer 3 (FID3). The FID3 algorithm, is extended to apply to
a fuzzy set of data (several data with membership grades) and generates a fuzzy
decision tree using fuzzy sets defined by a user for all attributes. A fuzzy decision
tree consists of nodes for testing attributes, edges for branching by test values of
fuzzy sets defined by a user and leaves for deciding class names with certainties.
Such an algorithm uses minimal fuzzy entropy or Information Gain as a criterion
for decision-making. In the beginning, Fuzzy ID3 is only an extension of the ID3
algorithm achieved by applying fuzzy sets. It generates a fuzzy decision tree using
fuzzy sets defined by a user for all attributes and utilises minimal fuzzy entropy to
select expanded attributes. However, sometimes the result of this Fuzzy ID3 is poor
in learning accuracy [19].
There are several variants of Fuzzy ID3 in the literature. For example one of
them uses minimal classification ambiguity instead of minimal fuzzy entropy [21].
Other algorithms, which uses cumulative information estimations, were proposed
by Levashenko and Zaiteva [22]. In these estimations we obtained Fuzzy Decision
Trees with different properties (unordered, ordered, stable etc.).
Algorithms, which are quite different from ID3 were proposed for example by
Wang et al. [23]. There are presented optimisation principles of fuzzy decision trees
based on minimising the total number and average depth of leaves. Dong and Kothari
[24] considered another non-Fuzzy ID3 algorithm where is used look-ahead method.
Its goal is to evaluate so-called classifiability of instances that are split along branches
of a given node.
Below is described the algorithm of FID3 used in the forecasting procedure. This
algorithm is very similar to ID3, except ID3 selects the test attribute based on the
information gain which is computed by the probability of ordinary data but this
algorithm by the probability of membership values for data [19].
Assume that we have a set of data D, where each data has p numerical values for
attributes A1 , A2 , . . . , A p and one classified class C = {C1 , C2 , . . . , Cn } and fuzzy
sets F̃i1 , F̃i2 , . . . , F̃il for the attribute Ai (the value of l varies on every attribute). Let
D̃ Ck be a fuzzy subset in D whose class is Ck and let |D| be the sum of the membership
values in a fuzzy set of data D. Then an algorithm for generating a fuzzy decision
7.2 Methodology 125

Algorithm 6 ID3 algorithm

1: Generate the root node that has a set of all data, i.e., a fuzzy set of all data with the membership
value 1
2: If a node t with a fuzzy set of data D̃ satisfies the following conditions:
– the proportion of a data set of a class Ck is greater than or equal to a threshold Θr that is:
| D̃ Ck |
 Θr ,
| D̃|
– the number of a data set is less than a threshold Θn that is | D̃|  Θn ,
– there are no attributes for more classification,
then t is a leaf node and assigned by the class name (more detailed method is described below)
3: If a node t does not satisfy the above conditions, it is not a leaf and the test node is generated as
follows:
3.1. For each Ai , (i = 1, 2, . . . , p), calculate the information gain G(Ai , D) (to be described
below) and select the test attribute Amax with the maximum gain.
3.2. Divide D into fuzzy subsets D̃1 , D̃2 , . . . , D̃l according to Amax , where the membership
value of the data in D̃ j is the product of the membership value in D and the value of Fmax, j
of the value of Amax , in D.
3.3. Generate new nodes tl , t2 , . . . , tl for fuzzy subsets D̃1 , D̃2 , . . . , D̃l , and label the fuzzy sets
Fmax, j to edges that connect the nodes t j and t.
3.4. Replace D̃ by D̃ j ( j = 1, 2, . . . , l) and repeat from recursively from 2.

tree is the following: The information gain G(Ai , D̃) for the attribute Ai by a fuzzy
set of data D is defined by

G(Ai , D̃) = I ( D̃) − E(Ai , D̃),

where:

n
| D̃ Ck | | D̃ Ck |
I ( D̃) = − log2 (7.6)
k=1
D D̃

| D̃ Fi j |
E(Ai , D̃) = m I ( D̃ Fi j ) (7.7)
j=1 D̃ Fi j

As for assigning the class name to the leaf node, are proposed three methods as
follows:
1. The node is assigned by the class name that has the greatest membership value,
that is, other than the selected data are ignored.
2. If the condition (a) in step 2 in the algorithm holds. Do the same as the method
(a). If not, the node is considered to be empty, that is, the data are ignored.
3. The node is assigned by all class names with their membership values, that is, all
data are taken into account.
126 7 Application of Fuzzy Decision Trees in Analog Forecasting

7.3 The Proposed Forecasting Method

The overview of forecasting methods conducted in Sect. 7.1 shows that the structure
and level of apparent consumption of steel products or GDP steel intensity depend
on selected macroeconomic parameters characterising a country’s economy. The fol-
lowing were most often used as exogenous variables: the value of GDP per capita
and the sectoral composition of the GDP. The sectoral composition is characterised
by the contribution of industry and construction to GDP. Additionally, in the case of
industry, the share of sectors determining the level of steel consumption (so-called
steel intensity industries) in the industry is significant to the total GDP. Such sectors
are: manufacture of finished metal products excluding machinery and equipment,
manufacture of electrical equipment, manufacture of machinery and equipment not
classified elsewhere, manufacture of motor vehicles and trailers, excluding motor-
cycles and the manufacture of other transport equipment.
The proposed method is used to forecast the following (endogenous variables in
the model):
• GDP steel intensity (St),
• share of various ranges of products in consumption
– share of long products (Ud),
– share of flat products (Up),
– share of pipes and hollow sections (Ur)
– relation of the consumption of organic coated sheets to the consumption of
metallurgic products altogether (Uo).
The exogenous variables in the model (predictors) are:
• value of GDP per capita (GDP),
• share of generated gross value in industry and construction in making up GDP
(UPB),
• share of steel intensity industries in making up gross value of industry (UPS).
Endogenous variables create a consumption profile for each country in a particular
year. Besides a consumption profile, the historical data include a summary of values
of predictors (exogenous variables) for each country and year. The historical data
covering consumption profiles and values of predictors are from different countries
and different periods. For certain countries, only values of predictors were available
in the forecasted period. Consumption profiles are forecasted on the basis of these
values.
An example of historical data for a single country in one year is presented in
Fig. 7.1.
The proposed forecasting procedure is as follows. First, consumption patterns are
created. A consumption pattern is understood as GDP steel intensity and structure
of consumption expressed by the share of individual ranges of steel products in total
consumption. Such patterns are defined using data on consumption profiles of dif-
ferent countries over many years. The k-means method is used for this purpose. The
7.3 The Proposed Forecasting Method 127

Fig. 7.1 An example of historical data used in the forecasting process

centroids of identified clusters make up the consumption patterns. After defining

clusters and their centroids, a fuzzy decision tree is built based on which consump-
tion patterns (the value of GDP steel intensity and consumption structure) can be
assigned to certain values of the predictors (the value of GDP per capita and GDP
composition).
In the proposed method historical data are used to build a fuzzy decision tree distin-
guishing easily interpretable links between predictors and pre-defined consumption
patterns. The fuzzy decision tree allows for combining a vector of values of predictors

Fig. 7.2 The proposed forecasting procedure

128 7 Application of Fuzzy Decision Trees in Analog Forecasting

with a specific consumption profile. This provides the possibility for forecasting the
level and structure of steel products consumption while knowing the forecasted value
of GDP and GDP composition. The discussed forecasting procedure is presented in
Fig. 7.2.

7.4 Empirical Results

Historical data used to generate the fuzzy decision tree was derived from the follow-
ing countries: Austria, Belgium, Denmark, Finland, France, Spain, Holland, Japan,
Lithuania, Norway, Czech Republic, Russia, Slovakia, Slovenia, Sweden, United
Kingdom, Hungary, Italy, and the USA. Data for Japan, the USA, and Western Euro-
pean countries came from the period 1960–2010, and data for the remaining countries
came from the period 1993–2010. It was not possible to gather data on the structure
of consumption for all counties in all years of the periods indicated above. Taking
into account the gaps in the data for some of the years, there was a total of 730
items collected. The data was divided into two sets: a set of 667 items, which served
to build a classifier and a set of 73 items, which were used to test the classifier and
compare it with other method. The division of the set was conducted at random. Data
on GDP for each country was expressed in dollars according to prices from 2007.1
The values of GDP steel intensity and demand structure defined by consumption
patterns are used in determining the forecast in the proposed method. These patterns
are defined on the basis of historical data describing the consumption profiles of
various countries and periods. Between the time which the data characterising the
consumption profiles comes from and the time for which the forecast is made, there
are production technology changes and structural changes of the products in sectors
utilising steel products. Changes also occur in the parameters of steel products in
terms of their durability characteristics, technological characteristics (for example,
bonding), and usability characteristics (for example, the type and quality of the
surface). These changes affect the reduction of sectoral indicators of steel intensity.
This results in a reduction of GDP steel intensity not directly associated with changes
in the GDP and its sectoral composition. Research conducted by EUROSTAT allows
assessing the extent to which the above-mentioned processes change GDP steel
intensity indicators. According to these studies GDP steel intensity indicators in
the years 1980–2000 in individual countries of the European Union (15) decreased
as a result of these processes an average of about 0.5 % per year. Accepting this
coefficient, steel intensity was adjusted in the profiles of the individual countries
for the various years. Adjustments (decreasing steel intensity) were made by the
percentage value determined by the following formula: (2007—the year the profile
originates from) 0.5 %.

1 The sources of data were corresponding yearly publications: The Steel Market, Annual Interna-
tional Statistics, and annual statistics of individual countries.
7.4 Empirical Results 129

In the analysed example, the share of gross value produced by steel intensity indus-
tries as part of the gross value produced by the industry in general was accepted as
one of the exogenous variables. This assumption is justified when the profiles take
into account the structure of consumption by major groups of ranges of steel products
(long products, flat products, pipes and organic coated sheet). Deeper disaggrega-
tion of consumption is likely to require determining GDP composition divided by
individual steel intensity industries. This is the subject of further research. The main
problem is obtaining reliable data on the consumption of particular ranges of steel
products in different countries.
In the first step one performed the clustering of steel intensity and demand structure
for the purpose of the obtainment of patterns of the consumption. In the process of
the clustering one accepted 9 classes. Results of the clustering are presented in the
Table 7.1 (values after the normalisation).
Table 7.2 presents the characteristics of individual clusters.
To build FID3, the fuzzification of all independent variables was carried out.
Fuzzification relied on the division of every attribute into classes of the same size.
The affiliation to the classes is described by fuzzy number. To the fuzzification
one used the following algorithm. The attribute A is a sequence J + 1 observation
A = {x j } j=0,1,...,J such that x j+1  x j . Observations are divided into l disjoint

Table 7.1 Result of the clustering of structures of the consumption

St Ud Up Ur Uo
1 0.184083 0.343596 0.651336 0.431067 0.507877
2 0.53204 0.442453 0.546201 0.502227 0.226074
3 0.39037 0.53588 0.550116 0.152016 0.314489
4 0.737196 0.514734 0.438294 0.665591 0.177773
5 0.31663 0.567211 0.46416 0.38451 0.072881
6 0.334618 0.549177 0.498734 0.311504 0.494481
7 0.313672 0.673825 0.360576 0.420971 0.33778
8 0.113727 0.268793 0.752175 0.303377 0.68626
9 0.529184 0.788603 0.261072 0.407442 0.393627

Table 7.2 The characteristics of individual classes

Cluster 1 2 3 4 5 6 7 8 9
Number of items 118 61 73 56 85 105 76 90 66
Average
St, kg/1000USD 15.702 35.826 27.632 47.691 23.368 24.408 23.197 11.633 35.66
Ud 0.396 0.418 0.439 0.435 0.447 0.442 0.471 0.379 0.497
Up 0.504 0.477 0.478 0.449 0.456 0.465 0.429 0.53 0.404
Ur 0.1 0.105 0.083 0.115 0.097 0.093 0.1 0.092 0.099
Uo 0.2 0.129 0.152 0.117 0.091 0.196 0.157 0.244 0.171
130 7 Application of Fuzzy Decision Trees in Analog Forecasting

subsets (Sk , k = 1, 2, . . . , l), where l = 5 + 3.3 ∗ log(J ) + 1 (in the example pre-
sented here l = 11). Then, a trapezoidal membership function H̃k = (ak , bk , ck , dk )
is determined for each subset Sk , where values ak , bk , ck , dk are designated in the
following way:

⎧
⎪
⎪ a k = gk
⎨
bk = gk
k=1 H̃k =
⎪
⎪ ck = gk+1
⎩
⎧ dk = gk+1 (1 + m)
⎪
⎪ a = gk (1 − m)
⎨ k
bk = gk
1<k <l +1 H̃k =
⎪
⎪ c k = gk+1
⎩ d = g (1 + m)
⎧ k k+1
⎪
⎪ a = gk (1 − m)
⎨ k
bk = gk
k =l +1 H̃k =
⎪
⎪ ck = gk+1
⎩ d =g
k k+1


where gk = x jk , jk = (k − 1) J +1
l
, k = 1, . . . , l + 1 and m ∈ [0, 1].
The following values of m were considered:

m (%)
GDP 5
UPB 1
UPS 4

Further, specific class names were used. Class names for variables are: the class
name(i), where i is number of the class, e.g., GDP1 (the first class for variable GDP).
Then, the decision tree of the type CART was built using the algorithm FID3
described in Sect. 7.2.2.
In the process of construction of a fuzzy decision tree, the following stopping
criterion was employed:

Threshold (%)
Θr 3.00
Θn 90.00

The final decision tree consists of 216 leaves and 100 nodes. The GDP attribute
is the most rarely tested. This means that it is the least important in determining the
level and structure of steel products consumption. Sectoral composition of the GDP
is more relevant in this case.
7.4 Empirical Results 131

Table 7.3 The example-case GDP BUD STAL

for prediction
GDP7 GDP8 BUD2 STAL7 STAL8
0.22 0.78 1 0.54 0.46

Inference in a ordinary decision tree is executed by starting from the root node
and repeating to test the attribute at the node and branch to an edge by its value until
reaching at a leaf node, a class attached to the leaf being as the result. The difference
between the ordinary and fuzzy tree relies on this that the given case is not credited
only to one branch, but to many with some degree of the membership.
For example, Table 7.3 presents the example-case. On the basis this case will
be elaborated the forecast. For simplicity, only non-zero values of the membership
function are shown below.
To elaborate the forecast the following three operations must be executed. First, for
every branch one ought to designate the indicator F—this is the value of membership
function with which the attribute of the case for which we elaborate the forecast
satisfies the rule by which the branch is based. Figure 7.3 presents a part of the
obtained fuzzy decision tree, where values of indicators F for individual branches are
non-zero. Values of indicators F are placed in the grey rectangles near the respective
branches. In the second step is designated membership function for analysed case to
individual clusters. Below one showed suitable calculations.
In the third step on the basis of such a defined membership function for the
analysed case is designated forecast.

Fig. 7.3 Part of the obtained fuzzy decision tree

132 7 Application of Fuzzy Decision Trees in Analog Forecasting

0.45 0.09 0.00 0.00 0.06 G3

0.220.54 0.35 + 0.220.46 0.85 + 0.780.54 0.34 + 0.780.54 0.34 = 0.90 G6
0.20 0.06 0.00 0.00 0.04 G7

G3 G6 G7 Forecast
St, kg/1000USD 27.632 24.408 23.197 24.553
Ud 0.439 0.442 0.471 0.443
Up 0.478 0.465 0.429 0.464
Ur 0.083 0.093 0.1 0.093
Uo 0.152 0.196 0.157 0.192

7.5 The Evaluation of Obtained Results

The quality of the classifier built in the form of a fuzzy decision tree can be assessed
by various indicators. In the analysed example, the values of continuous variables are
forecasted. As a measure of accuracy, the mean absolute percentage error (MAPE)
is used, defined by the formula:
n 
 
 yi − ŷi 
MAPE = 1
n  yi  (7.8)
t=1

where yi is the actual value and ŷi is the forecasted value.

Because in the analysed example the value of vector variables was forecasted,
MAPE was calculated according to the formula (7.9).
 n   n  
  Sti − Ŝt i    U di −Û d i 
MAPE = 1
5
1
n  Sti  + 1
n  U di 
i=1 i=1
n   n 
 
 U pi −Û pi   Uri −Ûr i 
+ n1  U pi  + 1
n  Uri 
i=1 i=1
n  
 U oi −Û oi 
+ n1  U oi  (7.9)
i=1

where n is the number of items for which the forecast was made.
It is difficult to compare the proposed method with conventional econometric
models, since a vector characterising the level and structure of consumption is fore-
casted. The vector components must meet certain conditions. The sum of the fore-
casted shares of the three main product groups (long products, flat products, pipes)
must be 1. In order to compare the effectiveness of the proposed fuzzy method an
additional classifiers were also tested based on traditional (crisp) C&RT method
which used the Gini coefficient to build the tree.
7.5 The Evaluation of Obtained Results 133

Table 7.4 Comparison of the mean absolute percentage error between the 2 tested models
Mean absolute percentage error
The proposed method (fuzzy decision tree) 0.028
The CART crisp method 0.076

A comparison of the quality of the forecasts was made using 67 selected items.
The quality of the forecasts was rated by calculating MAPE according to the formula
(7.9). The results of the tests are presented in Table 7.4.
The data in Table 7.4 indicate that the proposed fuzzy algorithm provides the
more accurate forecasts of the level and structure of consumption. This algorithm
provided the smaller value of the mean absolute percentage error. In this case, the
value of the mean absolute percentage error constitutes 36.8 % values of the error in
the crisp CART method using the Gini coefficient to select attributes based on which
the training set is divided.
The presented concept provides good results, which allow it to be recommended as
one of the methods for long-term forecasting of demand for selected products traded
on the industrial market. This method can be classified into the group of analog
methods. In such methods a forecast is formulated on the basis of comparator. Most
comparators are defined by experts. The proposed method allows for objectifying
the selection of comparators by making the selection dependent on the values of
chosen predictors. The method makes it possible to forecast the level and structure
of demand.

References

1. Rȩbiasz, B. 2006. Polish steel consumption 1974–2008. Resources Policy 31(1): 37–49.
2. Evans, M. 1996. Modeling steel demand in the UK. Ironmaking and Steelmaking 23(1): 19–23.
3. Ghosh, S. 2006. Steel consumption and economic growth: Evidence from India. Resources
Policy 31(1): 7–11.
4. Labson, B.S. 1997. Changing patterns of trade in the world iron ore and steel market: An
econometric analysis. Journal of Policy Modeling 19(3): 237–251.
5. Roberts, M.C. 1990. Predicting metal consumption: The case of US steel. Resources Policy
16(1): 56–73.
6. Roberts, M.C. 1996. Metal use and the world economy. Resources Policy 22(3): 183–196.
7. Malenbaum, W. 1975. World demand for raw materials in 1985 and 2000. New York: McGrow-
Hill.
8. Tilton, J.E. 1986. Beyond intensity of use. Materials and Society 10(3): 1–14.
9. Crompton, P. 2000. Future trends in Japanese steel consumption. Resources Policy 26(2): 1–14.
10. Thomassey, S., and A. Fiordaliso. 2006. A hybrid forecasting system based on clustering and
decision trees. Decision Support Systems 42(1): 408–421.
11. Altman, E.I., G. Macro, and F. Varetto. 1994. Corporate distress diagnosis: Comparison using
linear discriminant analysis and neural networks (the Italian experience). Journal Banking and
Finance 18(3): 505–529.
134 7 Application of Fuzzy Decision Trees in Analog Forecasting

12. Andrès, J., M. Landajo, and P. Lorca. 2005. Forecasting business profitability by using clas-
sification technique: A comparative analysis based on Spanish case. European Journal of
Operational Research 167(2): 518–542.
13. Rastogi, R., and K. Shim. 2000. Public: A decision tree classifier that integrates building and
pruning. Data Mining and Knowledge Discovery 4(4): 315–344.
14. Tsujimo, K., and S. Nishida. 1995. Implementation and refinement of decision tree using
nneural network for hybrid knowledge acquisition. Artificial Intelligence in Engineering 9(4):
265–275. Selected papers from the 1994 Japan/Korea joint conference on expert systems.
15. Hansen, L.K., and P. Salamon. 1990. Neural network ensembles. IEEE Transaction on Pattern
Analysis and Machine Intelligence 12(10): 993–1001.
16. Zhou, Z.-H., and Y. Jiang. 2004. Nec4.5: Neural ensemble based c4.5. IEEE Transaction on
Knowledge and Data Engineering 16(6): 770–773.
17. Quinlan, J.R. 1986. Induction of decision trees. Machine Learning 1(1): 81–106.
18. Olaru, C., and L. Wehenkel. 2003. A complete fuzzy decision tree technique. Fuzzy Sets and
Systems 138(2): 221–254.
19. Umanol, M., H. Okamoto, I. Hatono, H. Tamura, F. Kawachi, S. Umedzu, and J. Kinoshita.
1994. Fuzzy decision trees by fuzzy ID3 algorithm and its application to diagnosis systems. In
proceedings of the third IEEE conference on Fuzzy systems, IEEE world congress on compu-
tational intelligence, vol. 3, 2113–2118, Orlando, FL, June 26–29. IEEE World Congress on
Computational Intelligence.
20. Chandra, B., and P.P. Varghese. 2007. On improving efficiency of SLIQ decision tree algorithm.
In International joint conference on neural networks, IJCNN 2007, 66–71.
21. Yuan, Y., and M.J. Shaw. 1995. Induction of fuzzy decision trees. Fuzzy Sets and Systems 69(2):
125–139.
22. Zaitseva, E., V. Levashenko, K. Pancerz, and J. Gomuła. 2014. Fuzzy decision tree based
classification of psychometric data. In Position papers of the 2014 federated conference on
computer science and information systems, ACSIS, vol. 3, 37–41.
23. Wei, J.-M., M.-Y. Wang, J.-P. You, S.-Q. Wang, and D.-Y. Liu. 2007. VPRSM based decision
tree classifier. Computing and Informatics 26(6): 663–677.
24. Dong, M., and R. Kothari. 2001. Look-ahead based fuzzy decision tree induction. IEEE Trans-
actions on Fuzzy System 9(3): 461–468.
Chapter 8
Selected Issues of Visualisation of Fuzziness
in Cardiac Imaging Data

Abstract This chapter proposes an approach to visualisation of fuzzy numbers

in one-, two- and more-dimensional spaces. The proposed approach is based on
ScPovPlot3D templates for POVRay.

In the recent years technological development enabled elaboration of numerous

outstanding data gathering methods [1] which produces large amounts of data in
both numeric and raster image format. Amongst most productive disciplines one
can mention astronomy, engineering, economics, econometrics, automatics, medical
sciences including medical imaging. Reliable analysis of obtained data more often
than not requires the usage of computer processing and then data visualisation.
Accordingly to the subject of this book, processed data are assumed to be fuzzy,
so special techniques of visualisation are required for both input and processed
(output) data.

8.1 Visualisation of Fuzzy Numbers

Chapter 1 presents several pictures illustrating main concepts of interval and fuzzy
numbers. Such form, while being communicative, is highly impractical regarding
visualisation of non-trivial sets of fuzzy numbers, such as one and multi-dimensional
series of measurements, solutions of differential equations with fuzzy parameters or
fuzzy boundary conditions or computed or measured fuzzy surfaces [2].

8.1.1 Visualisation of a Fuzzy Number

One dimensional fuzzy point is simply a fuzzy number. It can be visualised in the
well-known form of interval (Fig. 8.1) with whiskers and if membership function is
to be presented, as a Box and Whiskers symbol (Fig. 8.2). However, according to the
interval or fuzzy value paradigm no uniquely defined “central” or “most important”
136 8 Selected Issues of Visualisation of Fuzziness in Cardiac Imaging Data

Fig. 8.1 Traditional

visualisation of one
dimensional fuzziness of
data. In this case, ‘x’
coordinate is known exactly
(it may be just a
measurement number) but
‘y’ coordinate is determined
with limited accuracy.
Usually, central point,
represented here by symbol
of square, points at
hypothetical accurate ‘y’
value while whiskers shows
level of uncertainty. Source
own

Fig. 8.2 Traditional

visualisation of one
dimensional fuzziness of
data including internal
distribution or membership
function. The box represents
for example range of
standard deviation while
whiskers shows lower and
upper boundary of the fuzzy
number

point exists. As stated in the first and second chapters, this is the main reason why
relational operators cannot be uniquely defined.
Fuzzy symbols shown in the both pictures may be modified in several ways.
For example one may add additional whiskers showing end points of consecutive
α-cuts of triangular fuzzy number (see Sect. 1.1, Definition 1.5, Fig. 1.5) or core and
support values of trapezoidal fuzzy number (see Sect. 1.1, Definition 1.6, Fig. 1.7).
Thus shape and properties of membership function can be effectively visualised,
even on scattered plots composed of several fuzzy values.
In this context Japanese Candlesticks, commonly used for technical analysis of
trends on stock markets (see [3]), may be assumed as visualisation technique for time
dependent fuzzy numbers (Fig. 8.3).
8.1 Visualisation of Fuzzy Numbers 137

Fig. 8.3 Example of Japanese Candlestick chart (one session). In the picture white candlestick is
shown. In this case central box illustrates opening (B) and closing (C) prices while ends of vertical
line segments depicts session maximum (D) and minimum (A), however price dynamic during
session cannot be shown in this manner. Several candlesticks may create specific formations which,
if spotted, might be helpful for share traders at buy-keep-sell decision making. Source own

8.1.2 Visualisation of a Two Dimensional Fuzzy Point

Visualisation of a two-dimensional fuzzy point may be done by simple extension of

one dimensional case. As shown in Fig. 8.4, both, vertical and horizontal uncertainty
may be shown together by means of “Box and Whiskers” style applied to both
dimensions respectively. The size of the internal rectangle is usually connected with
the standard deviation, but in this case it may be interpreted as the “core” of a fuzzy
number, whereas the whiskers may depict the “support” of a fuzzy number. As in
the one-dimensional case, marking of “central” makes no sense.
The described solution seems to be obvious and quite descriptive. However there
is a serious flow in such approach. Box and whiskers mark exact values of kernel
and support around “central lines” only. Support and kernel values measured along
intermediate directions are expected to be located on some kind of curve rather than
on the straight line. Thus, by analogy to multidimensional error analysis, ellipses, as
in Fig. 8.5 are proposed as approximate boundaries, i.e., support and kernel curves,1
respectively. More detailed analysis may suggest other curves family in the future.
In this example fuzzy ellipse is aligned to XY axes, but in general other orientations
cannot be excluded á priori.
Usually, measurements or calculations produce series of data, which in this book
are assumed to be fuzzy points. In order to visualise relations between them scatter

1 In statistics error ellipse is connected with covariance matrix of two stochastic, potentially
correlated variables. While this problem is out of scope of this book, it is thoroughly discussed
in relevant literature. Excellent explanation is given on the page [4].
138 8 Selected Issues of Visualisation of Fuzziness in Cardiac Imaging Data

Fig. 8.4 Example of two

dimensional fuzzy point.
Small central square visually
pinpoints symbol to the
chart. Larger rectangle gives
clue on uncertainty
distribution along horizontal
and vertical axes or just
shape of a membership
function. Endings of
horizontal and vertical line
segment may show support
of relevant fuzzy number

Fig. 8.5 Example of two

dimensional fuzzy point.
Small central square visually
pinpoints symbol to the
chart. Ellipses marked as
“Support” and “Kernel”
represents support and core
boundaries of trapezoidal
fuzzy point respectively

data plots may be employed. However there are no software packages supporting
explicitly drawing charts with elliptic two-dimensional fuzzy points, but this may be
easily changed in the future. Some of the packages may only roughly simulate this
behaviour using bitmap symbols. In the latter case, unfortunately, all symbols ought
to have the same size, what may not be relevant in every case (Fig. 8.6).

8.1.3 Visualization of a Three-Dimensional Fuzzy Points

In three dimensions Box and Whiskers style is no longer suitable as there is no

simple extension of line-art into 3D drawing. However ellipse style can be easily
extended into more dimensional spaces as in 3-dimensional space it just takes shape
of 3 axis ellipsoid (Fig. 8.7). Assuming that all components of the point in three
8.1 Visualisation of Fuzzy Numbers 139

Fig. 8.6 Simulated fuzzy

scattered data plot,
employing symbols filled
with radial gradient (Excel
2007)

dimensional space may be fuzzy numbers, real 3D chart is needed to visualise results
of measurements or calculations. In such case, point data may be imaged as scattered
fuzzy 3D chart, and surface data as fuzzy surface plot.
As mentioned above a dedicated software is necessary in order to ease drawing
of fuzzy 3D plots. One of them2 is a developer version of ScPovPlot3D package3
built on top of POVRay,4 well known ray-tracing program. ScPovPlot3D employs
a domain specific language of POVRay called SDL (Scene Description Language),
which makes it flexible and extensible solution. While thorough discussion of the
package is out of the scope of this book, some examples are shown below.

8.2 Scattered Fuzzy 3D Chart

One of common cases where 3D fuzzy points may be found is a measurement series
of three variables. Such series consist of ordered (for example by readout time5 ) or
unordered list of 3D points indexed by natural numbers (Fig. 8.8).
In the presented example a short series has been computed using simple linear, but
stochastically perturbed formula Pi (x, y, z) = [2 + δ; yi , ayi + b + ε] (see Eq. (8.1))
⎧
⎨ 2 + δ,
Pi (x, y, z) = yi , (8.1)
⎩
ayi + b + ε .

2 Authors have no information on availability of other packages.

3 www.scpovplot3d.sf.net, accessed 2014-09-30.
4 www.povray.org, accessed 2014-09-30.
5 Such a case may be considered 4D example with time as an independent variable.
140 8 Selected Issues of Visualisation of Fuzziness in Cardiac Imaging Data

Fig. 8.7 Visualisation of 3D

fuzzy trapezoidal
point—consecutive surfaces
represent outwards: kernel
surface, selected α-cut
surface and support surface.
More α-cuts may be
displayed as well however it
may impair legibility of the
chart if too many α-cuts are
selected. Developed using
ScPovPlot3D package

where δ and ε are random perturbations, a and b are coefficients, yi is an independent

variable of a measurement series and i = 1, . . . , N . For testing purposes it has been
assumed, that every measurement is in the form of 3D trapezoidal fuzzy number
described above and length of semiaxis is equal to half of support value for all
components. For x, y and z axis values 0.6, 0.2 and 0.8 respectively were adopted.
Resulting picture is shown in Fig. 8.8.
Fuzziness of given data point may result from inaccuracy embedded into laboratory
setup, assessment method, especially in social sciences, manufacturing inaccuracy
(engineering, metallurgy) and sometimes from Heisenberg indeterminacy
(uncertainty) principle (quantum phenomena).
Apart from visualisation of fuzzy numbers this technique may be applied to
visualisation of dispersion of rigid numbers obtained in the course of repetitive
measurements of single data point, for example diameter of a cylinder. In this case,
consecutive layers of 3D fuzzy number may represent for example percentiles or
multiples of standard deviation range. Furthermore, in case of correlation between
measured x, y, z variables, a point-ellipsoid may be adequately rotated. In the latter
case a covariance error ellipse procedure extended into three dimensions might be
useful as well as solutions based on Mahalanobis distance (refer to [5, 6]).
In order to enhance visual message of the picture presented in the Fig. 8.8, right
projections of fuzzy points may be added on base planes of coordinate system.
8.2 Scattered Fuzzy 3D Chart 141

Fig. 8.8 Visualization of 3D fuzzy trapezoidal points—consecutive surfaces represent outwards:

core surface, selected α-cut surface and support surface. Lengths of semiaxes of point-ellipsoids
are proportional to their fuzziness: 0.6, 0.2, 0.8 measured along x, y and z axis respectively. Points
are distributed due to the stochastically perturbated formula: Pi (x, y, z) = [2 + δ; yi , ayi + b + ε].
Developed using ScPovPlot3D package

8.2.1 Fuzzy Surface Plot

Problem of fuzzy surfaces visualisation, has been primarily addressed in the paper [2].
In the range of problems uncertainty is distributed over surface z = f(x, y),6 usually
defined over rectangular domain, but in fact, there is no such restriction in general.
One of the sciences widely employing surface visualisation is geology which uses
it to present distribution of mineral resources, principally ore deposits, for example
coal or oil. Such a distribution may be probed using wide range of exploratory
geophysical techniques like borehole logging, electrical resistivity tomography or
gravimetry measurements. Unfortunately usually such mapping could not be done
over rectangular grid required to display surface visualisation thus measurement
points are freely distributed over the survey field. By applying numerical technique
known as kriging (refer to [2, 7]) a regular grid may be evaluated, but values
computed in every node are subject to uncertainty, resulting both from uncertainty
of measured values and numerical, for example round-off, errors. As stated before,
such uncertainty may be expressed in numerous ways, including standard deviation
measure and of course fuzzy number paradigm. In the latter case first solution is to
draw “thick surface”, in contrast to normal, “infinitely thin surface” commonly used.
Package ScPovPlot3D cannot explicitly implement “thick” or “fat” surfaces but this
quality can be simulated utilising two features of POVRay: prism object and media7
statement.
Media statement introduces fog resembling interior of an object with controlled
density. Thick facets may be rendered with or without media statement, and
accordingly, with opaque or transparent surface, fully or partially, so there is a lot

6 Or more generally f(x, y, z) = 0.

7 Media is a specific to POVRay implementation of particles.
142 8 Selected Issues of Visualisation of Fuzziness in Cardiac Imaging Data

Fig. 8.9 Surface computed

by kriging based on
measurement points (larger
spheres). Rectangular grid
(smallers balls) has been
computed using simple
kriging algorithm. Source
[2], own by ScPovPlot3D

Fig. 8.10 Example of fuzzy

surface. Support range is
represented by two separate
surfaces, lower end
displayed as grayscale
mapped surface and the
upper as unicolour but
semitransparent one.
Unfortunately implicit
functions can not be
represented this way. Source
[2], visualisation created in
ScPovPlot3D

of different arrangements and one of them may be suitable for a specific situation
(Figs. 8.9 and 8.10).
The common practise in 3D computer graphics is splitting objects into meshes
composed of myriads of tiny triangles, named facets. Specially crafted shadow
functions are responsible for experience of smooth tonal transitions, as well as,
simulating “smooth” look of the surface. By replacing every facet by prism with
base aligned with the facet one can obtain “thick” surface elements comprising
whole thick surface object. Of course information on the thickness of every element
ought to be delivered for every facet separately. Some variants of thick facets are
shown in Fig. 8.11 with relevant explanation.
In the next two sections another two examples of application of fuzzy surfaces
will be presented. First one shows surface defined by fuzzy relationship z̃ = z̃ (x, y),
8.2 Scattered Fuzzy 3D Chart 143

Fig. 8.11 Examples of fuzzy surface facets. Different settings has been applied to the same
geometry. a Texture only and central triangle, b media only and central triangle, c media and
texture without central triangle, d all components. Support range is represented by distance between
triangular bases of the prism. Image produced using ScPovPlot3D package

where z̃ denotes fuzzy number. The second example shows lesions deteriorating
blood flow deposited in a segment of cardiac vessel.

Example 1: Visualisation of z̃ = z̃(x, y) Surface

Implementation of thick surfaces using media (or particles) requires generation of

complex polymeshes. This can be fully done using applications employing mesh
as a generic object. Some of such applications can be listed: Blender, AutoCAD
or 3DStudio. In every case efficient automated evaluation of the mesh requires
sophisticated coding.8 However, resulting object will render perfectly smooth.
In order to illustrate “thick surface” drawing style, the POVRay, popular raytracing
program has been used. In this case, thick surface consists of a set of individual prism
objects. As there is no obvious method of deformation of the prism to follow curvature
of the surface (outer base should be wider than inner, and rarely symmetrically),

8 For Blender efficient user space API written in Python is available.

144 8 Selected Issues of Visualisation of Fuzziness in Cardiac Imaging Data

Fig. 8.12 Example of fuzzy

surface. Support range is
represented by two separate
planes upper and lower
(dilute and monochromatic),
while predicted “central”
plane is presented in between
as color mapped surface

number of artifacts is produced if the “thick surface” gets really thick, thus only “not
so fat surfaces” can be represented this way.
This problem can be resolved using three supplementary surfaces—upper and
lower, representing range of support of fuzzy surface and the middle, standard surface,
representing visually expected position of central (modal) point, if triangular fuzzy
number is employed.
In Fig. 8.12 such a rendering is presented. Upper and lower surfaces are rendered
using “fat facets” filled with semitransparent media. Vertical faces of prisms tend to
form artifacts and due to curvature of the surface, some cracks can be observed as
well. However, despite these defects, the picture clearly represents fuzzy nature of
related problem while artifact, as side effect, increase its readability.
In the end another problem should be addressed. Usage of media interior, rapidly
increases compute time of the scene. Thus powerful multicore computer is required,
possibly employing CUDA technology, happily a range of plugins to popular 3D
modellers is available.9

8.2.2 Imaging of Fuzziness of Cardiac Vessels Walls

Angiography is a diagnostic method suitable for visualisation of geometry of

coronary blood vessels and assessment of arteries walls. The method requires
intravenous administration of radiocontrast agents and subsequent registration of
X-Ray images in a classical or Computed Tomography (CT) setup [1, 9, 10]. The
variant of angiography is a coronary angiography (CAG). In this case, the image of a
vessel is obtained using catheter introduced into coronary vessel, which enables the
localisation and assessment of extent of lesions and severity of coronary stenoses [11].

9 Forexample, FurryBall GPU plugin renderer is available for Autodesk Maya, 3DS Max and
Cinema 4D [8], nice comparison to another plugins is available at http://furryball.aaa-studio.eu/
aboutFurryBall/compare.html.
8.2 Scattered Fuzzy 3D Chart 145

This diagnostic method is invasive and dangerous for a patient. Therefore it should
be performed only if really necessary.
Raw angiography does not allow for collecting detailed information on extent
and severity of development of arteriosclerotic vascular disease, for example in the
vicinity of the wall of a blood vessel [1, 10]. However, it is usually used for rough
assessment of stenosis of coronary vessels, changes in walls of coronary arteries and
anomalies in their placement, as well as, orifices [9]. Angiography, as opposed to
coronary angiography (CAG) [12], is assumed to be safe, even for high risk patients
[13]. However the risk connected with intravascular administration of radiocontrast
agent still remains in patients with kidney disorder or failure.
Coronary angiography enhanced by computer analysis of image of coronary
arteries (QCA, Quantitative Coronary Angiography) allows for increase in a
reliability of state assessment of vessel wall and profile of cross section [14]. Increase
in credibility of coronary angiography (CAG) can also be obtained by suitable use
of Computed Tomographic Angiography (CTA) [1].
Main disadvantage of the method is acquiring different images depending on
angle of angiographic projection, what is the cause of uncertainty in interpretation
of imaging, in particular in case of presence of extensive stenoses of cardiac arteries,
where possible contours of cross section through given stenosis may exhibit severe
eccentricity (usually they are assumed to be irregular, in fact) as well as heavy
reduction of the lumen of the vessel.
Unfortunately, above mentioned methods of visualisation of coronary arteries
presents approximate only path and cross section of the vessel, involving some
level of fuzziness of resulting assessment. Thus additional visualisation techniques
ought to be developed and employed in order to enable cardiologists to evaluate and
visualise possible deviations from calculated vessel surface and thickness of its wall
and atheromatous plaques. Eventually this leads to estimation of uncertainty (upper
and lower boundary) of FFR value (Fractional Flow Reserve [1]).

8.2.3 DICOM Format Versus STL

In order to satisfy the need of recording of medical imaging data dedicated standards
has been established allowing for interoperability between applications and systems
as well as enabling or facilitate data interpretation and visualisation also for
telemedicine.10
In the first place DICOM (Digital Imaging and Communications in Medicine)
standard should be mentioned, which has been elaborated in 1983 commonly
by American College of Radiology (ACR) and National Electrical Manufacturers
Association (NEMA) organisations [15]).11 Popularity of DICOM (in almost all
branches of medicine) emerges from its simplicity and conformance with

10 Ex. cardiac, stomatologic, radiologic, orthopaedic consultations.

11 Published as ACR-NEMA Standards Publication No. 300–1985.
146 8 Selected Issues of Visualisation of Fuzziness in Cardiac Imaging Data

requirements of other standardising agencies, ao. CEN, JIRA, IEEE, HL7 and ANSI,
what without doubts positively influenced coherency and integrity of the standard.
DICOM standard ensures high quality visualisation of medical data both static
and dynamic. Besides imaging data, supplementary data are recorded in relevant
attributes (exemplary parameters ranges), what enables extensive postprocessing and
unique identification of the patient. Simultaneously open12 structure of STL format
allows for processing data by third party applications, systems or devices.
These remarks are also valid for integration of information produced by dedicated
applications of diverse kinds (various hardware/software environments), delivering
data for processing in heterogeneous systems dedicated for storage and processing
of medical data (EHR systems, Electronic Health Record). DICOM specification
defines also requirements for data transfer, storage and accessibility by various EHR
systems.
Despite many advantages,13 visible ao. during transfer of patient’s data between
various medical centres equipped with imaging systems delivered by diverse
manufacturers, and, what is related, diverse applications for processing and data
archiving, as a main drawback of the standard, possibility of supplying redundant
or useless data is mentioned. At the same time definitions of graphical objects
may remain incomplete, due to filling improper fields or supplying non relevant
or erroneous data or even omitting crucial fields. Another bunch of problems
is connected with data processing by apparatus calibrated with different sets of
parameters (ex. amplitude ranges) what deteriorates quality of images, for example
by affecting of contrast or gamma profiles [16].
Nowadays, read-out of images recorded using DICOM standard is possible not
only by apparatus (mainly medical) used to produce them, but also, due to openness of
the specification, by third party applications14 or by general use packages employing
suitable plugins.15 However many popular applications which are able to deal with
DICOM files, converts them on import to their own working format, what can lead
to difficulties during analysis including loss of information or quality.16
It has to be explicitly stressed that DICOM files stores spatial information mainly
in raster format, it is, they comprise several, and often numerous, 2-dimensional
grayscale raster images (layers, slices) produced by a variety of imaging systems like
MRI, CT, CAT, CTA, CAG and supplementary information on their mutual layout.
By the rule, these layers are parallel and laid one over other, separated by constant
distance resulting from scanning resolution used (see Fig. 8.13). This is why raw
DICOM files, ensuring important diagnostic capabilities, simultaneously are useless

12 As opposed to closed usually proprietary specifications.

13 Versatilityi.e. support for all medical disciplines, high quality of data, interoperability.
14 Ex. MatLab (functions: dicomread, dicomwrite), Photoshop Extended.
15 Plugin is a small program or library dynamically linked on runtime extending functionality of

main application (host).

16 A good example is GIMP graphics editor which can import images from DICOM files but at the

same time drops crucial data on their geometry and topology.

8.2 Scattered Fuzzy 3D Chart 147

Fig. 8.13 Subsequent layers of scan of a human head (Source [17])

for automated, computer analysis of spatial relations, especially for assessment of

state of coronary artery.
Initially STL (STereoLithography) format [18] has been developed years ago as
input format for 3D printing systems [19].17 Presently, this format is supported by
many software packages,18 proprietary as well as Free Software (FLOSS—Free
Libre/Open Source Software) and is used extensively ao. for cardiology diagnostics.
STL standard can be classified as a 3D vector graphics format. The surface of
the object is defined in a 3D space as plain set of triangles (tessellation). The STL
definition comprises both textual (ASCII) as well as binary variant [20]. The surface
is composed from triangular facets19 coded separately (thus coding is redundant).
Definition of every single facet includes vector normal to the surface of the triangle,
directed “outside” of the object, and three additional vectors defining three vertices
of the triangle ordered counterclockwise, when watching in the direction opposite to
the normal vector.20

17 For example implants and prosthetic devices or its parts can be manufactured this way.
18 Ex. Blender, application MeshLab (http://meshlab.sourceforge.net/ revealed 16.11.2013) and
many other including CAD related software (Computer Aided Design).
19 This term is adopted from jewellery and means one of the several flat polished surfaces cut on

a gemstone (exemplary diamond, brilliant cut), or occurring naturally on a crystal. Facets usually
are flat polygons. In computer graphics all surfaces, even very complex, are built from of triangle
facets (sometimes counted in myriads) in order to minimise rendering times.
20 As an vector graphics format, STL allows for detailed visualisation of focused parts of an analysed

object.
148 8 Selected Issues of Visualisation of Fuzziness in Cardiac Imaging Data

Specification of ASCII-STL file is presented in Listings 8.1 and 8.2. Key words
like “facet” have to be written with small letters. The first item in STL file is a key
word “solid” followed by name of elaborated object. This name may, but does not
have to be used later. The last item in the file is key word “endsolid”.

Listing 8.1 Specification of ASCII-STL file format, the text including and following
“#” character comprises author’s comment and is not part of STL file. Source: [18]
solid Description # file header
facet normal n1, n2, n3 # declaration of the first facet
outer loop # begin of declaration of vertices of facet
vertex v11, v12, v13
vertex v21, v22, v23
vertex v31, v32, v33
endloop # end of declaration of vertices
endfacet # end of declaration of facet
facet normal n1, n2, n3 # declaration of the second facet
outer loop # begin of declaration of vertices of facet
vertex v11, v12, v13
vertex v21, v22, v23
vertex v31, v32, v33
endloop # end of declaration of vertices
endfacet # end of declaration of facet
facet normal n1, n2, n3 # declaration of the next facet
outer loop # begin of declaration of vertices of facet
vertex v11, v12, v13
vertex v21, v22, v23
vertex v31, v32, v33
endloop # end of declaration of vertices
endfacet # end of declaration of facet
# all remaining facets follows
endsolid # End of file

Next keyword—facet—begins declaration of the first triangle which is finished

by keyword endfacet. Number of facets in the STL file may easily reach and even
exceed hundreds of thousands. facet keyword is followed by components of the
normal vector. Then there is a definition of three vertices in the section beginning
with words outer loop and terminated by words end loop. Declaration of components
of every given vertex is open by keyword vertex, then follows three floats, and end
of line character (EOL in short, in fact it is represented on Windows systems by
two characters, CR and LF21 ) is expected. While original definition of the standard
forbids the use of negative values of vertices components, it is not always obeyed. It is
worth to say, that this restriction is artificial and is not derived from any programming
principle.

21 CR-carriage return (ASCII ‘13’), LF—line feed (ASCII ‘10’). On Unix systems there is always

one character, namely CR.

8.2 Scattered Fuzzy 3D Chart 149

Listing 8.2 Example of ASCII-STL file. Words formatted in italics can be edited.
solid ASCIIFile
facet normal 0.940 −0.039 0.337
outer loop
vertex −38.000 −18.549 −8.320
vertex −37.970 −18.496 −8.399
vertex −37.990 −18.367 −8.326
endloop
endfacet
facet normal 0.941 −0.040 0.333
outer loop
vertex −38.000 −18.549 −8.320
vertex −37.990 −18.367 −8.326
vertex −38.020 −18.420 −8.247
endloop
endfacet
facet normal 0.921 0.0503 0.385
outer loop
vertex −38.02 −18.42 −8.24
vertex −37.99 −18.36 −8.32
vertex −38.02 −18.23 −8.25
endloop
endfacet
…
…
endsolid

In “outer loop” section at least three vertices have to be present. Though STL
format specification allows for more than three vertices, usually only three (i.e.
triangles) are used, what enhances efficiency of calculations. Subsequent numbers
(floats with decimal points) are separated by spaces (at least one space, ASCII code
‘32’), keywords are put without quotations around. There is no scale definition in
the file and units are relative. Program performing tessellation (generating facets)
has to define them in such a way that vertices of one facet are not on the edge of
another. So, if there are two triangle neighbours, they have exactly one common
vertex or one whole edge. Due to round up errors a couple of vertices belonging to
the neighbouring triangles may not coincide, thus a tiny crack may be spotted in the
tessellated surface.
As the above description suggests, structure of STL file resembles XML file
structure, however tags in STL file are not enclosed in acute parentheses. It should
be noted, that STL standard had been elaborated long ago XML draft emerged.
As stated in the previous paragraph, STL format defines object geometry only
and is relatively ascetic—contains three even tags (a pair of opening and closing tag)
and one odd tag (opening only). Simplicity of the definition is clearly advantage of
the STL format. Further, this format is suitable for processing on contemporary, fast
graphics cards, designed for rendering hundreds of thousands of triangles per second.
150 8 Selected Issues of Visualisation of Fuzziness in Cardiac Imaging Data

Fig. 8.14 Element of cardiac vessel rendered using “thick facets”. On the internal surface visible is
pattern created by individual prisms. External surface is patterned using marble like texture and by
no means reflects real look of blood arteries. Source visualisation produced using STL.inc module
from ScPovPlot3D package

In particular after few adjustments can be supplied as input data file to thick surface
procedures, implemented in STL.inc library from mentioned above ScPovPlot3D
package. Example of resulting image of cut-out of coronary vessel is given in the
Fig. 8.14.
Despite that STL format has severe shortcomings, it is satisfactory as a base for
visualisation of 3D fuzzy surface. However, coronary vessel can be visualised only
in tiny slices in order to avoid screening of one part by another. Whole vessel may
be then presented using animation created in automated or interactive manner. The
latter requires additional presentation software, which is beyond of the scope of this
book.

References

1. Taylor, C.A., T.A. Fonte, and J.K. Min. 2013. Computational fluid dynamics applied to cardiac
computed tomography for noninvasive quantification of fractional flow reserve: Scientific basis.
Journal of the American College of Cardiology 61(22): 2233–2241.
2. Opiła J., I. Skalna, and T. Pełech-Pilichowski. Interval arithmetic for irregularly distributed
data visualisation by Kriging. In XIV International scientific conference, editor, Corporate
Governance—Theory and Practice, 1–14, Krakow, 22–23 Nov 2012. WZ AGH.
3. StockCharts.com, Inc. Japanese candlesticks, introduction to candlesticks. http://stockcharts.
com/school/doku.php?id=chart_school:chart_analysis:introduction_to_candlesticks.
Accessed 25 Sept 2014.
4. Spruyt, V. How to draw a covariance error ellipse? http://www.visiondummy.com/2014/04/
draw-error-ellipse-representing-covariance-matrix/. Accessed 29 Oct 2014.
5. Jenness, J., and L. Engelman. Jennes enterprises, ArcView. http://www.jennessent.com/
arcview/mahalanobis_description.htm. Accessed 29 Oct 2014.
6. Mahalanobis, P.C. 1936. On the generalised distance in statistics. In Proceedings of the National
Institute of Sciences of India 2: 49–55.
7. Krige, D.G. 1951. A statistical approach to some basic mine valuation problems on the
witwatersrand. Journal of the Chemical, Metallurgical and Mining Society 52: 119–139.
References 151

8. Art and animation studio 2015, GPU vs CPU. GPU rendering vs software CPU rendering.
http://furryball.aaa-studio.eu/aboutFurryBall/compare.html. Accessed 25 Nov 2014.
9. Achenbach, S., J. Walecki, M. Zawadzki, and M. Witulski. 2006. Clinical applications of
multislice computed tomography in cardiology. Postȩpy w Kardiologii Interwencyjnej 2(2):
160–168.
10. Legutko, J., J. Ja̧kała, B. Mrevlje, S. Bartuś, and D. Dudek. 2011. Fractional flow reserve-guided
myocardial revascularization. Postȩpy w Kardiologii Interwencyjnej 7(3 (25)): 228–241.
11. Bartoszek, B., M. Mościński, T. Niklewski, B. Szyguła-Jurkiewicz, and A. Lekston. 2011.
Wirtualna histologia—nowoczesna metoda oceny tȩtnic wieńcowych. Folia Cardiologica
Excerpta 6(3): 203–209.
12. Kubica, J., R. Gil, and P. Pienia̧żek. 2005. Wytyczne dotycza̧ce koronarografii. Kardiologia
Polska 63(5(3)): 491–500.
13. Paluszek, P., P. Pienia̧żek, P. Musiałek, T. Przewłocki, A. Kabłak-Ziembicka, R. Motyl, Ł.
Tekieli, A. Leśniak-Sobelga, K. Żmudka, and W. Tracz. 2009. Symptomatic vertebral artery
stenting with the use of bare metal and drug eluting stents. Kardiologia Polska 5(1(15)): 1–6.
14. Barańska-Kosakowska, A., M. Hawranek, M. Ga̧sior, A. Spatuszko, R. Przybylski, M.
Zakliczyński, and M. Zembala. 2010. Serial quantitative coronary angiography (QCA) in the
assessment of transplant coronary artery disease (TxCAD). Kardiochirurgia i Torakochirurgia
Polska 7(3): 319–324.
15. Digital imaging and communications in medicine. Strategic document (last revised
23.10.2013). http://medical.nema.org/dicom/geninfo/Strategy.pdf. Accessed 14 Nov 2013.
16. Mustra, M., K. Delac, and M. Grgic. 2008. Overview of the DICOM standard. 50th international
symposium ELMAR-2008, vol. 1, 39–44. Croatia.
17. Haeggstroem, M. 2008. Computed tomography of human brain. Uppsala University Hospital.
http://en.wikipedia.org/wiki/File:Computed_tomography_of_human_brain_-_large.png.
Accessed 10 May 2015.
18. Burns, M. The STL format. Standard data format for fabbers. reprinted from section 6.5 of
automated fabrication, Ennex corporation. http://www.ennex.com/~fabbers/StL.asp. Accessed
14 Nov 2013.
19. Palermo, E. What is stereolithography? Live science. http://www.livescience.com/38190-
stereolithography.html. Accessed 16 Nov 2013.
20. Burns, M. 1993. Automated fabrication: Improving productivity in manufacturing. PTR
Prenrice Hall.