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

net/publication/241283939

Encyclopedia of Complexity and Systems Science

Book · January 2009

DOI: 10.1007/978-0-387-30440-3

CITATIONS READS

340 5,431

1 author:

Robert Meyers
Ramtechlimited
68 PUBLICATIONS 1,532 CITATIONS

SEE PROFILE

All content following this page was uploaded by Robert Meyers on 03 June 2015.

The user has requested enhancement of the downloaded file.

Robert A. Meyers (Ed.)

Encyclopedia of Complexity
and Systems Science

With 4300 Figures and 420 Tables

123
ROBERT A. MEYERS, Ph. D.
Editor-in-Chief
RAMTECH LIMITED
122 Escalle Lane
Larkspur, CA 94939
USA
[email protected]

Library of Congress Control Number: 2008933604

ISBN: 978-0-387-30440-3
This publication is available also as:
Print publication under ISBN: 978-0-387-75888-6 and
Print and electronic bundle under ISBN: 978-0-387-69572-3

© 2009 SpringerScience+Buisiness Media, LLC.

All rights reserved. This work may not be translated or copied in whole or in part without the written permission of the publisher (Springer
Science+Business Media, LLC., 233 Spring Street, New York, NY 10013, USA), except for brief excerpts in connection with reviews or
scholarly analysis. Use in connection with any form of information storage and retrieval, electronic adaptation, computer software, or
by similar or dissimilar methodology now known or hereafter developed is forbidden.
The use in this publication of trade names, trademarks, service marks, and similar terms, even if they are not identified as such, is not to
be taken as an expression of opinion as to whether or not they are subject to proprietary rights.
springer.com
Printed on acid free paper SPIN: 11560258 2109letex – 5 4 3 2 1 0
 Rough Sets in Decision Making R 7753

Two kinds of classification problems are distinguished:

Rough Sets in Decision Making
 taxonomy, when the value sets of attributes and the
ROMAN SŁOWI ŃSKI 1,2 , SALVATORE GRECO3 , predefined classes are not preference ordered,
BENEDETTO MATARAZZO3  ordinal classification (also known as multiple crite-
1 Institute of Computing Science, Poznan University
ria sorting), when the value sets of attributes and
of Technology, Poznan, Poland the predefined classes are preference ordered.
2 Systems Research Institute, Polish Academy of Sciences,

Warsaw, Poland Two kinds of choice problems are distinguished:

3 Faculty of Economics, University of Catania,
 discrete choice, when the set of objects is finite and
Catania, Italy reasonably small to be listed,
 multiple objective optimization, when the set of
objects is infinite and defined by constraints of
Article Outline a mathematical program.
Glossary If value sets of attributes are preference-ordered, they
Definition of the Subject are called criteria or objectives, otherwise they keep the
Introduction name of attributes.
Classical Rough Set Approach to Classification Problems Criterion is a real-valued function g i defined on A, re-
of Taxonomy Type flecting a worth of objects from a particular point of
Dominance-Based Rough Set Approach view, such that in order to compare any two objects
to Ordinal Classification Problems a; b 2 A from this point of view it is sufficient to com-
DRSA on a Pairwise Comparison Table pare two values: g i (a) and g i (b).
for Multiple Criteria Choice and Ranking Problems Dominance Object a is non-dominated in set A (Pareto-
DRSA for Decision Under Uncertainty optimal) if and only if there is no other object b in A
Multiple Criteria Decision Analysis such that b is not worse than a on all considered crite-
Using Association Rules ria, and strictly better on at least one criterion.
Interactive Multiobjective Optimization Preference model is a representation of a value system of
Using DRSA (IMO-DRSA) the decision maker on the set of objects with vector
Conclusions evaluations.
Future Directions Decision under uncertainty takes into account conse-
Bibliography quences of decisions that distribute over multiple
states of nature with given probability. The preference
order, characteristic for data describing multiple at-
Glossary tribute decision problems, concerns also decision un-
der uncertainty, where the objects correspond to acts,
Multiple attribute (or multiple criteria) decision sup-
attributes are outcomes (gain or loss) to be obtained
port aims at giving the decision maker (DM) a recom-
with a given probability, and the problem consists in
mendation concerning a set of objects A (also called
ordinal classification, choice, or ranking of the acts.
alternatives, actions, acts, solutions, options, candi-
Rough set in universe U is an approximation of a set
dates, . . . ) evaluated from multiple points of view
based on available information about objects of U. The
called attributes (also called features, variables, crite-
rough approximation is composed of two ordinary sets
ria, objectives, . . . ).
called lower and upper approximation. Lower approx-
Main categories of multiple attribute (or multiple crite-
imation is a maximal subset of objects which, accord-
ria) decision problems:
ing to the available information, certainly belong to the
approximated set, and upper approximation is a mini-
 classification, when the decision aims at assigning mal subset of objects which, according to the available
each object to one of predefined classes, information, possibly belong to the approximated set.
 choice, when the decision aims at selecting the best The difference between upper and lower approxima-
object, tion is called boundary.
 ranking, when the decision aims at ordering objects Decision rule is a logical statement of the type “if . . . ,
from the best to the worst. then . . . ”, kursiv where the premise (condition part)
7754 R Rough Sets in Decision Making

specifies values assumed by one or more condition at- of attributes are preference ordered, they are called crite-
tributes and the conclusion (decision part) specifies an ria, otherwise they keep the name of attributes. For exam-
overall judgment. ple, in a decision regarding the selection of a car, its price
is a criterion because, obviously, a low price is better than
a high price. Instead, the color of a car is not a criterion but
Definition of the Subject
simply an attribute, because red is not intrinsically better
Scientific analysis of decision problems aims at giving the than green. One can imagine, however, that also the color
decision maker (DM) a recommendation concerning a set of a car could become a criterion if, for example, a DM
of objects (also called alternatives, solutions, acts, actions, would consider red better than green.
options, candidates, . . . ) evaluated from multiple points
of view considered relevant for the problem at hand and
Introduction
called attributes (also called features, variables, criteria, ob-
jectives, . . . ). For example, a decision can concern: Scientific support of decisions makes use of a more or less
explicit model of the decision problem. The model relates
1) diagnosis of pathologies for a set of patients, where pa- the decision to the characteristics of the objects expressed
tients are objects of the decision, and symptoms and by the considered attributes. Building such a model re-
results of medical tests are the attributes, quires information about conditions and parameters of the
2) assignment of enterprises to classes of risk, where en- aggregation of multi-attribute characteristics of objects.
terprises are objects of the decision, and financial ra- The nature of this information depends on the adopted
tio indices and other economic indicators, such as the methodology: prices and interest rates for cost-benefit
market structure, the technology used by the enterprise analysis, cost coefficients in objectives and technological
and the quality of management, are the attributes, coefficients in constraints for mathematical programming,
3) selection of a car to be bought from among a given a training set of decision examples for neural networks and
set of cars, where cars are objects of the decision, and machine learning, substitution rates for a value function
maximum speed, acceleration, price, fuel consump- of multi-attribute utility theory, pairwise comparisons of
tion, comfort, color and so on, are the attributes, objects in terms of intensity of preference for the analytic
4) ordering of students applying for a scholarship, where hierarchy process, attribute weights and several thresholds
students are objects of the decision, and scores in dif- for ELECTRE methods, and so on (see the state-of-the-art
ferent subjects are the attributes. survey [4]). This information has to be provided by the
DM, possibly assisted by an analyst.
The following three main categories of decision problems Very often this information is not easily definable. For
are typically distinguished [44]: example, this is the case of the price of many immaterial
goods and of the interest rates in cost-benefit analysis, or
 classification, when the decision aims at assigning each the case of the coefficients of objectives and constraints in
object to one of predefined classes, mathematical programming models. Even if the required
 choice, when the decision aims at selecting the best ob- information is easily definable, like a training set of deci-
jects, sion examples for neural networks, it is often processed in
 ranking, when the decision aims at ordering objects a way which is not clear for the DM, such that (s)he cannot
from the best to the worst. see what are the exact relations between the provided in-
formation and the final recommendation. Consequently,
Looking at the above examples, one can say that 1) and very often the decision model is perceived by the DM as
2) are classification problems, 3) is a choice problem and a black box whose result has to be accepted because the
4) is a ranking problem. analyst’s authority guarantees that the result is “right”. In
The above categorization can be refined by distin- this context, the aspiration of the DM to find good rea-
guishing two kinds of classification problems: taxonomy, sons to make decision is frustrated and rises the need for
when the value sets of attributes and the predefined classes a more transparent methodology in which the relation be-
are not preference ordered, and ordinal classification (also tween the original information and the final recommen-
known as multiple criteria sorting), when the value sets dation is clearly shown. Such a transparent methodology
of attributes and the predefined classes are preference or- searched for has been called glass box [32]. Its typical rep-
dered [12]. In the above examples, 1) is a taxonomy prob- resentative is using a learning set of decision examples as
lem and 2) is an ordinal classification problem. If value sets the input preference information provided by the DM, and
 Rough Sets in Decision Making R 7755

it is expressing the decision model in terms of a set of “if account preference orders and monotonic relationships
. . . , then . . . ” decision rules induced from the input infor- between condition and decision attributes. This gener-
mation. From one side, the decision rules are explicitly re- alization, called Dominance-based Rough Set Approach
lated to the original information and, from the other side, (DRSA), has been proposed by Greco, Matarazzo and
they give understandable justifications for the decisions to Slowinski [12,14,15,18,21,49].
be made.
For example, in case of a medical diagnosis problem, Classical Rough Set Approach
the decision rule approach requires as input information to Classification Problems of Taxonomy Type
a set of examples of previous diagnoses, from which some
Information Table and Indiscernibility Relation
diagnostic rules are induced, such as “if there is symp-
tom ˛ and the test result is ˇ, then there is pathology ”. Information regarding classification examples is supplied
Each one of such rules is directly related to examples of in the form of an information table, whose separate rows
diagnoses in the input information, where there is symp- refer to distinct objects, and whose columns refer to differ-
tom ˛, test result ˇ and pathology . Moreover, the DM ent attributes considered. This means that each cell of this
can verify easily that in the input information there is no table indicates an evaluation (quantitative or qualitative)
example of diagnosis where there is symptom ˛, test re- of the object placed in the corresponding row by means of
sult ˇ but no pathology . the attribute in the corresponding column. Formally, an
The rules induced from the input information pro- information table is the 4-tuple S D hU; Q; V ; vi, where U
vided in terms of exemplary decisions represent a decision is a finite set of objects, called universe, Q D fq1 ; : : : ; q n g
model which is transparent for the DM and enables his is a finite set of attributes, V q is a value set of the at-
S
understanding of the reasons of his past decisions. The ac- tribute q; V D q2Q Vq , and v : U  Q ! V is a total
ceptance of the rules by the DM justifies, in turn, their use function such that v(x; q) 2 Vq for each q 2 Q; x 2 U,
for future decisions. called information function.
The induction of rules from examples is a typical ap- Therefore, each object x of U is described by a vec-
proach of artificial intelligence. This explains our inter- tor (string) DesQ (x) D [v(x; q1 ); : : : ; v(x; q n )], called de-
est in rough set theory [38,39] which proved to be a use- scription of x in terms of the evaluations on the attributes
ful tool for analysis of vague description of decision situ- from Q; it represents the available information about x.
ations [41,48]. The rough set analysis aims at explaining Obviously, x 2 U can be described in terms of any non-
the values of some decision attributes, playing the role of empty subset P  Q.
“dependent variables”, by means of the values of condition To every (non-empty) subset of attributes P  Q there
attributes, playing the role of “independent variables”. For is associated an indiscernibility relation on U, denoted
example, in the above diagnostic context, data about the by I P :
presence of a pathology are given by decision attributes, I P D f(x; y) 2 U  U : v(x; q) D v(y; q) ; 8q 2 Pg :
while data about symptoms and tests are given by condi-
tion attributes. An important advantage of the rough set If (x; y) 2 I P , it is said that the objects x and y are P-in-
approach is that it can deal with partly inconsistent ex- discernible. Clearly, the indiscernibility relation thus de-
amples, i. e. cases where the presence of different patholo- fined is an equivalence relation (reflexive, symmetric and
gies is associated with the presence of the same symptoms transitive). The family of all the equivalence classes of the
and test results. Moreover, it provides useful information relation I P is denoted by U/I P , and the equivalence class
about the role of particular attributes and their subsets, containing object x 2 U, by I P (x). The equivalence classes
and prepares the ground for representation of knowledge of the relation I P are called P-elementary sets.
hidden in the data by means of “if . . . , then . . . ” decision
rules. Approximations
Classical rough set approach proposed by Paw-
Let S be an information table, X a non-empty subset of U
lak [38,39] cannot deal, however, with preference order
and ; ¤ P  Q. The P-lower approximation and the P-
in the value sets of condition and decision attributes. For
upper approximation of X in S are defined, respectively,
this reason, the classical rough set approach can deal with
as:
only one of four decision problems listed above – clas-
sification of taxonomy type. To deal with ordinal classi- P (X) D fx 2 U : I P (x)  Xg ;
fication, choice and ranking, it is necessary to general-
ize the classical rough set approach, so as to take into P (X) D fx 2 U : I P (x) \ X ¤ ;g :
7756 R Rough Sets in Decision Making

The elements of P (X) are all and only those objects is called quality of the approximation of classification Y by
x 2 U which belong to the equivalence classes generated set of attributes P, or in short, quality of classification. It
by the indiscernibility relation I P , contained in X; the ele- expresses the ratio of all P-correctly classified objects to all
ments of P (X) are all and only those objects x 2 U which objects in the system.
belong to the equivalence classes generated by the indis-
cernibility relation I P , containing at least one object x be- Dependence and Reduction of Attributes
longing to X. In other words, P (X) is the largest union
An issue of great practical importance is the reduction of
of the P-elementary sets included in X, while P (X) is the
“superfluous” attributes in an information table. Superflu-
smallest union of the P-elementary sets containing X.
ous attributes can be eliminated without deteriorating the
The P-boundary of X in S, denoted by Bn P (X), is de-
information contained in the original table.
fined by
Let P  Q and p 2 P. It is said that attribute p
Bn P (X) D P (X)  P (X) : is superfluous in P with respect to classification Y if
P(Y) D (P  p)(Y); otherwise, p is indispensable in P. The
The following inclusion relation holds: subset of Q containing all the indispensable attributes is
P (X)  X  P (X) : known as the core.
Given classification Y, any minimal (with respect to in-
Thus, in view of information provided by P, if an ob- clusion) subset P  Q, such that P(Y) D Q(Y), is called
ject x belongs to P (X), then it certainly belongs to set X, reduct. It specifies a minimal subset P of Q which keeps
while if x belongs to P (X), then it possibly belongs to set X. the quality of classification at the same level as the whole
Bn P (X) constitutes the “doubtful region” of X: nothing set of attributes, i. e. P (Y) D Q (Y). In other words, the
can be said with certainty about the membership of its el- attributes that do not belong to the reduct are superfluous
ements to set X, using the subset of attributes P only. with respect to the classification Y of objects from U.
Moreover, the following complementarity relation is More than one reduct may exist in an information ta-
satisfied: ble and their intersection gives the core.

P (X) D U  P(U  X) :
Decision Table and Decision Rules
If the P-boundary of set X is empty, Bn P (X) D ;, In the information table describing examples of classi-
then X is an ordinary (exact) set with respect to P, that is, fication, the attributes of set Q are divided into condi-
it may be expressed as union of a certain number of P-el- tion attributes (set C ¤ ;) and decision attributes (set
ementary sets; otherwise, if Bn P (X) ¤ ;, set X is an ap- D ¤ ;); C [ D D Q and C \ D D ;. Such an informa-
proximate (rough) set with respect to P and may be char- tion table is called a decision table. The decision attributes
acterized by means of the approximations P (X) and P (X). induce a partition of U deduced from the indiscernibility
The family of all sets X  U having the same P-lower relation I D in a way that is independent of the condition
and P-upper approximations is called the rough set. attributes. D-elementary sets are called decision classes, de-
The quality of the approximation of set X by means of noted by Cl t ; t D 1; : : : ; m. The partition of U into de-
the attributes from P is defined as cision classes is called classification Cl D fCl1 ; : : : ; Cl m g.
jP (X)j There is a tendency to reduce the set C while keeping
P (X) D ; all important relationships between C and D, in order to
jXj
make decisions on the basis of a smaller amount of infor-
such that 0  P (X)  1. The quality P (X) represents the mation. When the set of condition attributes is replaced
relative frequency of the objects correctly classified using by one of its reducts, the quality of approximation of the
the attributes from P. classification induced by the decision attributes is not de-
The definition of approximations of a set X  U teriorating.
can be extended to a classification, i. e. a partition Since the aim is to underline the functional dependen-
Y D fY1 ; : : : ; Ym g of U. The P-lower and P-upper ap- cies between condition and decision attributes, a decision
proximations of Y in S are defined ˚ by sets P(Y) D table may also be seen as a set of decision rules. These are
fP(Y1 ); : : : ; P(Ym )g and P(Y) D P(Y1 ); : : : ; P(Ym ) , re- logical statements of the type “if . . . , then . . . ”, where the
spectively. The coefficient premise (condition part) specifies values assumed by one
Pm or more condition attributes (description of C-elementary
iD1 jP(Yi )j
P (Y) D sets) and the conclusion (decision part) specifies an assign-
jUj
 Rough Sets in Decision Making R 7757

ment to one or more decision classes. Therefore, the syn- (b) Generation of an exhaustive representation, i. e. all
tax of a rule is the following: rules for a given decision table,
(c) Generation of a characteristic representation, i. e. a set
“if v(x; q1 ) D r q1 and v(x; q2 ) D r q2 and . . . of rules covering relatively many objects, however, not
v(x; q p ) D r q p , then x belongs to decision class Cl j1 necessarily all objects from a decision table.
or Cl j2 or . . . Cl jk ”,

where fq1 ; q2 ; : : : ; q p g  C; (r q1 ; r q2 ; : : : ; r q p ) 2 Vq1 

Vq2   Vq p and Cl j1 ; Cl j2 ; : : : ; Cl jk are some deci- Explanation of the Classical Rough Set Approach
sion classes of the considered classification Cl. If the con- by an Example
sequence is univocal, i. e. k D 1, then the rule is univocal, Suppose that one wants to describe the classification of ba-
otherwise it is approximate. sic traffic signs to a novice. There are three main classes of
An object x 2 U supports decision rule r if its descrip- traffic signs corresponding to:
tion is matching both the condition part and the decision
part of the rule. The decision rule r covers object x if it  Warning (W),
matches the condition part of the rule. Each decision rule  Interdiction (I),
is characterized by its strength, defined as the number of  Order (O).
objects supporting the rule. In the case of approximate
Then, these classes may be distinguished by such at-
rules, the strength is calculated for each possible decision
tributes as the shape (S) and the principal color (PC) of
class separately.
the sign. Finally, one can consider a few examples of traffic
If a univocal rule is supported by objects from the
signs, like those shown in Table 1. These are:
lower approximation of the corresponding decision class
only, then the rule is called certain or deterministic. If, a) Sharp right turn,
however, a univocal rule is supported by objects from the b) Speed limit of 50 km/h,
upper approximation of the corresponding decision class c) No parking,
only, then the rule is called possible or probabilistic. Ap- d) Go ahead.
proximate rules are supported, in turn, only by objects
from the boundaries of the corresponding decision classes. The rough set approach is used here to build a model
Procedures for generation of decision rules from a de- of classification of traffic signs to classes W, I, O on the
cision table use an inductive learning principle. The ob- basis of attributes S and PC. This is a typical problem of
jects are considered as examples of classification. In or- taxonomy.
der to induce a decision rule with a univocal and cer- One can remark that the sets of signs indiscernible by
tain conclusion about assignment of an object to decision “Class” are:
class X, the examples belonging to the C-lower approxi-
mation of X are called positive and all the others negative. W D fagClass ; I D fb; cgClass ; O D fdgClass ;
Analogously, in case of a possible rule, the examples be-
longing to the C-upper approximation of X are positive Rough Sets in Decision Making, Table 1
and all the others negative. Possible rules are character- Examples of traffic signs described by S and PC
ized by a coefficient, called confidence, telling to what ex-
tent the rule is consistent, i. e. what is the ratio of the num- Traffic sign Shape (S) Primary color (PC) Class
ber of positive examples supporting the rule to the num-
ber of examples belonging to set X according to decision
a) triangle yellow W
attributes. Finally, in case of an approximate rule, the ex-
amples belonging to the C-boundary of X are positive and
all the others negative. A decision rule is called minimal if
b) circle white I
removing any attribute from the condition part gives a rule
covering also negative objects.
The existing induction algorithms use one of the fol- c) circle blue I
lowing strategies [55]:

(a) Generation of a minimal representation, i. e. minimal

set of rules covering all objects from a decision table, d) circle blue O
7758 R Rough Sets in Decision Making

Rough Sets in Decision Making, Table 2

and the sets of signs indiscernible by S and PC are as fol- Examples of traffic signs described by S, PC and SC
lows:
Traffic sign Shape (S) Primary Secondary Class
fagS;PC ; fbgS;PC ; fc; dgS;PC : color (PC) color (SC)

The above elementary sets are generated, on the one

hand, by decision attribute “Class” and, on the other hand, a) triangle yellow red W
by condition attributes S and PC. The elementary sets of
signs indiscernible by “Class” are denoted by f gClass and
those by S and PC are denoted by f gS;PC. Notice that b) circle white red I
W D fagClass is characterized precisely by fagS;PC. In or-
der to characterize I D fb; cgClass and O D fdgClass , one
needs fbgS;PC and fc; dgS;PC , however, only fbgS;PC is in- c) circle blue red I
cluded in I D fb; cgClass while fc; dgS;PC has a non-empty
intersection with both I D fb; cgClass and O D fdgClass . It
follows, from this characterization, that by using condition d) circle blue white O
attributes S and PC, one can characterize class W precisely,
while classes I and O can only be characterized approxi-
mately: One way to increase the quality of the approximation is
to add a new attribute so as to decrease the ambiguity. Let
 Class W includes sign a certainly and possibly no other us introduce the secondary color (SC) as a new condition
sign than a, attribute. The new situation is now shown in Table 2.
 Class I includes sign b certainly and possibly signs b; c As one can see, the sets of signs indiscernible by S, PC
and d, and SC, i. e. the elementary sets f gS;PC;SC, are now:
 Class O includes no sign certainly and possibly signs c
and d. fagS;PC;SC ; fbgS;PC;SC ; fcgS;PC;SC ; fdgS;PC;SC :
The terms “certainly” and “possibly” refer to the ab-
It is worth noting that the elementary sets are finer
sence or presence of ambiguity between the description of
than before and this enables the ambiguity to be elimi-
signs by S and PC from the one side, and by “Class”, from
nated. Consequently, the quality of approximation of the
the other side. In other words, using description of signs
classification by attributes S, PC and SC is now equal to 1.
by S and PC, one can say that all signs from elementary
A natural question occurring here is to ask if, indeed,
sets f gS;PC included in elementary sets f gClass belong cer-
all three attributes are necessary to characterize precisely
tainly to the corresponding class, while all signs from ele-
the classes W, I and O. When attribute S or attribute PC is
mentary sets f gS;PC having a non-empty intersection with
eliminated from the description of the signs, the elemen-
elementary sets f gClass belong to the corresponding class
tary sets f gPC;SC or f gS;SC are defined, respectively, as fol-
only possibly. The two sets of certain and possible signs
lows:
are, respectively, the lower and upper approximation of the
corresponding class by attributes S and PC:
fagPC;SC ; fbgPC;SC ; fcgPC;SC ; fdgPC;SC ;
lower_approx:S;PC (W) D fag;
fagS;SC ; fb; cgS;SC ; fdgS;SC :
upper_approx:S;PC (W) D fag;
lower_approx:S;PC (I) D fbg; Using any one of the above elementary sets, it is possi-
ble to characterize (approximate) classes W, I and O with
upper_approx:S;PC (I) D fb; c; dg;
the same quality (equal to 1) as it is when using the ele-
lower_approx:S;PC (O) D ;; mentary sets f gS;PC;SC (i. e. those generated by the com-
upper_approx:S;PC (O) D fc; dg: plete set of three condition attributes). Thus, the answer to
the above question is that the three condition attributes are
The quality of approximation of the classification by not all necessary to characterize precisely the classes W, I
attributes S and PC is equal to the number of all the signs and O. It is, in fact, sufficient to use either PC and SC or S
in the lower approximations divided by the number of all and SC. The subsets of condition attributes fPC; SCg and
the signs in the table, i. e. 1/2. fS; SCg are called reducts of fS; PC; SCg because they have
 Rough Sets in Decision Making R 7759

this property. Note that the identification of reducts en- Indeed, the following decision rules can be induced:
ables us to reduce attributes about the signs from the table
to only those which are relevant. rul e #100 : if S D triangle; then Class D W fag
Other useful information can be generated from the rul e #200 : if PC D white; then Class D I fbg
identification of reducts by taking their intersection. This rul e #300 : if PC D blue; then Class D I or O fc; dg :
is called the core. In our example, the core contains at-
tribute SC. This tells us that it is clearly an indispensable at- Note that these rules can be induced from the
tribute i. e. it cannot be eliminated from the description of lower approximations of classes W and I, and from
the signs without decreasing the quality of the approxima- the set called the boundary of both I and O. In-
tion. Note that other attributes from the reducts (i. e. S and deed, for exact rule #100 , the supporting example is in
PC) are exchangeable. If there happened to be some other lower_approx:S;PC (W) D fag; for exact rule #200 it is in
attributes which were not included in any reduct, then they lower_approx:S;PC (I) D fbg; and the supporting examples
would be superfluous, i. e. they would not be useful at all in for approximate rule #300 are in the boundary of classes I
the characterization of the classes W, I and O. and O, defined as:
If, however, column S or PC is eliminated from Ta-
boundaryS;PC (I)
ble 2, then the resulting table is not a minimal representa-
tion of knowledge about the classification of the four traffic D upper_approx:S;PC(I)  lower_approx:S;PC (I)
signs. Note that, in order to characterize class W in Ta- D fc; dg ;
ble 2, it is sufficient to use the condition “S = triangle”.
Moreover, class I is characterized by two conditions (“S = boundaryS;PC (O)
circle” and “SC = red”) and class O is characterized by the
D upper_approx:S;PC(O)  lower_approx:S;PC (O)
condition “SC = white”. Thus, the minimal representation
of this information system requires only four conditions D fc; dg :
(rather than the eight conditions that are presented in Ta-
As a result of the approximate characterization of
ble 2 with either column S or PC eliminated). This repre-
classes W, I and O by S and PC, an approximate repre-
sentation corresponds to the following set of decision rules
sentation in terms of decision rules is obtained. Since the
which may be seen as classification model discovered in
quality of the approximation is 1/2, exact rules (#100 and
the data set contained in Table 2 (in the braces there are
#200 ) cover one half of the examples and the other half is
symbols of signs covered by the corresponding rule):
covered by the approximate rule (#300 ). Now, the quality
of approximation by S and SC, or by PC and SC, was equal
rul e #1 : if S D triangle; then Class D W fag to 1, so all examples were covered by exact rules (#1 to #3,
rul e #2 : if S D circle or #10 to #40 respectively).
and SC D red; then Class D I fb; cg One can see, from this simple example, that the rough
rul e #3 : if SC D white; then Class D O fdg : set analysis of data included in an information system pro-
vides some useful information. In particular, the following
results are obtained:
This is not the only representation, because an alterna-
tive set of rules is:  A characterization of decision classes in terms of cho-
sen condition attributes through lower and upper ap-
rul e #10 : if PC D yellow; then Class D W fag proximation.
0  A measure of the quality of approximation which in-
rul e #2 : if PC D white; then Class D I fbg
dicates how good the chosen set of attributes is for ap-
rul e #30 : if PC D blue;
proximation of the classification.
and SC D red; then Class D I fcg  The reducts of condition attributes including all rele-
0
rul e #4 : if SC D white; then Class D O fdg : vant attributes. At the same time, superfluous and ex-
changeable attributes are also identified.
It is interesting to come back to Table 1 and to ask what  The core composed of indispensable attributes.
decision rules represent this information system. As the  A set of decision rules which is induced from the lower
description of the four signs by S and PC is not sufficient and upper approximations of the decision classes. This
to characterize exactly all the classes, it is not surprising constitutes a classification model for a given informa-
that not all the rules will have a non-ambiguous decision. tion system.
7760 R Rough Sets in Decision Making

Dominance-Based Rough Set Approach The statement x 2 Cl t reads “x belongs to at least class
to Ordinal Classification Problems Cl t ”, while x 2 Cl t reads “x belongs to at most class
Dominance-Based Rough Set Approach (DRSA) Cl t ”. Let us remark that Cl1 D Cl m D U; Cl  D Cl
m m

and Cl1 D Cl1 . Furthermore, for t D 2; : : : ; m,
Dominance-based Rough Set Approach (DRSA) has been
proposed by the authors to handle background knowledge 
Cl t1 D U  Cl t and Cl t D U  Cl t1

:
about ordinal evaluations of objects from a universe, and
about monotonic relationships between these evaluations, The key idea of DRSA is representation (approxima-
e. g. “the larger the mass and the smaller the distance, the tion) of upward and downward unions of decision classes,
larger the gravity” or “the greater the debt of a firm, the by granules of knowledge generated by criteria. These gran-
greater its risk of failure”. Such a knowledge is typical for ules are dominance cones in the criteria values space.
data describing various phenomena. It is also character- x dominates y with respect to set of criteria P  I
istic for data concerning multiple criteria decision or de- (shortly, x P-dominates y), denoted by xD P y, if for ev-
cision under uncertainty, where the order of value sets of ery criterion i 2 P; f i (x)  f i (y). The relation of P-domi-
condition and decision attributes corresponds to increas- nance is reflexive and transitive, i. e. it is a partial preorder.
ing or decreasing preference. In case of multiple criteria Given a set of criteria P  I and x 2 U, the granules of
decision, the condition attributes are called criteria. knowledge used for approximation in DRSA are:
Let us consider a decision table including a finite uni-
verse of objects (solutions, alternatives, actions) U eval-  a set of objects dominating x, called P-dominating set,
uated on a finite set of criteria F D f f1 ; : : : ; f n g, and on DCP (x) D fy 2 U : yD P xg,
a single decision attribute d. The set of the indices of cri-  a set of objects dominated by x, called P-dominated set,
teria is denoted by I D f1; : : : ; ng. Without loss of gen- DP (x) D fy 2 U : xD P yg.
erality, f i : U ! < for each i D 1; : : : ; n, and, for all ob-
jects x; y 2 U; f i (x)  f i (y) means that “x is at least as Let us recall that the dominance principle requires that
good as y with respect to criterion f i ”, which is denoted an object x dominating object y on all considered criteria
by x  i y. Therefore, it is supposed that  i is a complete (i. e. x having evaluations at least as good as y on all con-
preorder, i. e. a strongly complete and transitive binary re- sidered criteria) should also dominate y on the decision
lation, defined on U on the basis of evaluations f i ( ). Fur- (i. e. x should be assigned to at least as good decision class
thermore, decision attribute d makes a partition of U into as y). Objects satisfying the dominance principle are called
a finite number of decision classes, Cl D fCl1 ; : : : ; Cl m g, consistent, and those which are violating the dominance
such that each x 2 U belongs to one and only one class principle are called inconsistent.
Cl t ; t D 1; : : : ; m. It is assumed that the classes are prefer- The P-lower approximation of Cl t , denoted by
ence ordered, i. e. for all r; s D 1; : : : ; m, such that r > s, P Cl t , and the P-upper approximation of Cl t , denoted
the objects from Clr are preferred to the objects from Cls . by P Cl t , are defined as follows (t D 2; : : : ; m):
More formally, if  is a comprehensive weak preference re- ˚ 
lation on U, i. e. if for all x; y 2 U; xy reads “x is at least P Cl t D x 2 U : DC
P (x)  Cl t ;
as good as y”, then it is supposed that ˚ 
P Cl t D x 2 U : D 
P (x) \ Cl t ¤ ; :

[x2Clr ; y2Cls ; r > s] ) xy ; Analogously, one can define the P-lower approxima-
tion and the P-upper approximation of Cl t as follows
where xy means xy and not yx.
(t D 1; : : : ; m  1):
The above assumptions are typical for consideration
of an ordinal classification (or multiple criteria sorting) ˚ 
P Cl t D x 2 U : D
P (x)  Cl t ;
problem. Indeed, the decision table characterized above ˚ 
includes examples of ordinal classification which consti- P Cl t D x 2 U : DC 
P (x) \ Cl t ¤ ; :
tute an input preference information to be analyzed using
DRSA. The P-lower and P-upper approximations so defined
The sets to be approximated are called upward union satisfy the following inclusion properties, for all P  I:
and downward union of decision classes, respectively:
P Cl t  Cl t  P Cl t ; t D 2; : : : ; m ;
[ [
Cl t D Cls ; Cl t D Cls ; t D 1; : : : ; m :
st st
P Cl t  Cl t  P Cl t ; t D 1; : : : ; m  1 :
 Rough Sets in Decision Making R 7761

The P-lower and P-upper approximations of Cl t and The dominance-based rough approximations of up-
Cl t  have an important complementarity property, ac- ward and downward unions of decision classes can serve
cording to which, to induce a generalized description of objects in terms
of “if . . . , then . . . ” decision rules. For a given upward
P Cl t D U  P Cl t1

and or downward union of classes, Cl t or Cls , the decision
P Cl t DUP 
Cl t1 ; t D 2; : : : ; m; rules induced under a hypothesis that objects belonging to
P(Cl t ) or P Cls are positive examples, and all the oth-
P Cl t DUP 
Cl tC1 and ers are negative, suggest a certain assignment to “class Cl t

P Cl t DUP Cl tC1 ; t D 1; : : : ; m  1 : or better”, or to “class Cls or worse”, respectively. On the
other hand, the decision rules induced under a hypothesis
The P-boundary of Cl t and Cl t , denoted by that objects belonging to P Cl t or P Cl s are positive
Bn P Cl t and Bn P Cl t , respectively, are defined as fol- examples, and all the others are negative, suggest a pos-
lows: sible assignment to “class Cl t or better”, or to “class Cls
or worse”, respectively. Finally, the decision rules induced
Bn P Cl t D P Cl t  P Cl t ; t D 2; : : : ; m; under a hypothesis that objects belonging to the intersec-
Bn P Cl t DP Cl t P Cl t ; t D 1; : : : ; m  1: tion P(Cls ) \ P(Cl t ) are positive examples, and all the
others are negative, suggest an approximate assignment to
Due to the above complementarity property, some classes between Cls and Cl t (s < t).
Bn P Cl t D Bn P Cl t1 
, for t D 2; : : : ; m. In the case of preference ordered description of ob-
For every P  C, the quality of approximation of the jects, set U is composed of examples of ordinal classifica-
ordinal classification Cl by a set of criteria P is defined as tion. Then, it is meaningful to consider the following five
the ratio of the number of objects P-consistent with the types of decision rules:
dominance principle and the number of all the objects
in U. Since the P-consistent objects are those which do 1) certain D -decision rules, providing lower profile de-
not belong to any P-boundary Bn P (Cl  t ); t D 2; : : : ; m, or scriptions for objects belonging to P(Cl t ): if f i 1 (x) 
Bn P Cl t ; t D 1; : : : ; m  1, the quality of approxima- r i 1 and . . . and f i p (x)  r i p , then x 2 Cl t ; fi1 ; : : : ;
tion of the ordinal classification Cl by a set of criteria P, i p g  I; t D 2; : : : ; m; r i 1 ; : : : ; r i p 2 <;
can be written as 2) possible D -decision rules, providing lower profile de-
ˇ S  ˇ
ˇ scriptions for objects belonging to P(Cl t ): if f i 1 (x) 
ˇU 
tD2;:::;m Bn P Cl t
P (Cl) D r i 1 and . . . and f i p (x)  r i p , then x possibly belongs to
jUj
ˇ S  ˇ
ˇ Cl t ; fi1 ; : : : ; i p g  I; t D 2; : : : ; m; r i 1 ; : : : ; r i p 2 <;
ˇU 
tD1;:::;m1 Bn P Cl t 3) certain D -decision rules, providing upper profile de-
D :
jUj scriptions for objects belonging to P(Cl t ): if f i 1 (x) 
r i 1 and . . . and f i p (x)  r i p , then x 2 Cl t ; fi1 ; : : : ;
P (Cl) can be seen as a degree of consistency of the i p g  I; t D 1; : : : ; m  1; r i 1 ; : : : ; r i p 2 <;
objects from U, when P is the set of criteria and Cl is the 4) possible D -decision rules, providing upper profile de-
considered ordinal classification. scriptions for objects belonging to P(Cl t ): if f i 1 (x) 
Each minimal (with respect to inclusion) subset P  C r i 1 and . . . and f i p (x)  r i p , then x possibly belongs
such that P (Cl) D C (Cl) is called a reduct of Cl, and is
to Cl t ; fi1 ; : : : ; i p g  I; t D 1; : : : ; m  1; r i 1 ; : : : ;
denoted by REDCl . Let us remark that for a given set U
r i p 2 <;
one can have more than one reduct. The intersection of all
5) approximate D -decision rules, providing simultane-
reducts is called the core, and is denoted by CORECl . Cri-
ously lower and upper profile descriptions for objects
teria in CORECl cannot be removed from consideration
belonging to Cls [ ClsC1 [ : : : [ Cl t , without possi-
without deteriorating the quality of approximation. This
bility of discerning to which class: if f i 1 (x)  r i 1 and
means that, in set C, there are three categories of criteria:
. . . and f i k (x)  r i k and f i kC1 (x)  r i kC1 and . . . and
 indispensable criteria included in the core, f i p (x)  r i p , then x 2 Cls [ ClsC1 [ : : : [ Cl t ; fi1 ; : : : ;
 exchangeable criteria included in some reducts, but not i p g  I; s; t 2 f1; : : : ; mg; s < t; r i 1 ; : : : ; r i p 2 <.
in the core,
 redundant criteria, neither indispensable nor ex- In the premise of a D -decision rule, there can be
changeable, and thus not included in any reduct. “ f i (x)  r i ” and “ f i (x)  r0i ”, where r i  r0i , for the same
7762 R Rough Sets in Decision Making

i 2 I. Moreover, if r i D r0i , the two conditions boil down sional condition space < p (p  n), and each decision pro-
to “ f i (x) D r i ”. file defines a dominance cone in one-dimensional decision
Since a decision rule is a kind of implication, a mini- space f1; : : : ; mg. In both cases, the cones are positive for
mal rule is understood as an implication such that there D -rules and negative for D -rules.
is no other implication with the premise of at least the Let also remark that dominance cones corresponding
same weakness (in other words, a rule using a subset of to condition profiles can originate in any point of <n ,
elementary conditions or/and weaker elementary condi- without the risk of their being too specific. Thus, contrary
tions) and the conclusion of at least the same strength (in to traditional granular computing, the condition space <n
other words, a D - or a D -decision rule assigning objects need not be discretized.
to the same union or sub-union of classes, or a D -de- Procedures for rule induction from dominance-based
cision rule assigning objects to the same or smaller set of rough approximations have been proposed in [17].
classes). In [10], a new methodology for the induction of mono-
The rules of type 1) and 3) represent certain knowl- tonic decision trees from dominance-based rough approx-
edge extracted from data (examples of ordinal classifica- imations of preference ordered decision classes has been
tion), while the rules of type 2) and 4) represent possible proposed.
knowledge; the rules of type 5) represent doubtful knowl- Application of DRSA to decision related problems
edge, because they are supported by inconsistent objects goes far beyond ordinal classification. In [27], DRSA has
only. been used for decision support involving multiple decision
Moreover, the rules of type 1) and 3) are exact if they makers, and in [28], DRSA has been applied to case-based
do not cover negative examples, and they are probabilis- reasoning. The following sections present applications of
tic otherwise. In the latter case, each rule is characterized DRSA to multiple criteria choice and ranking, to decision
by a confidence ratio, representing the probability that an under uncertainty and to interactive multiobjective opti-
object matching the premise of the rule also matches its mization. The surveys [24,26,29,51,52] include other ap-
conclusion. plications of DRSA.
Given a certain or possible D -decision rule r 
“if f i 1 (x)  r i 1 and . . . and f i p (x)  r i p , then x 2 Cl t ”,
Example Illustrating DRSA in the Context
an object y 2 U supports r if f i 1 (y)  r i 1 and . . . and
of Ordinal Classification
f i p (y)  r i p . Moreover, object y 2 U supporting decision
rule r is a base of r if f i 1 (y) D r i 1 and . . . and f i p (y) D r i p . This subsection presents a didactic example which illus-
Similar definitions hold for certain or possible D -deci- trates the main concepts of DRSA. Let us consider the fol-
sion rules and approximate D -decision rules. A deci- lowing ordinal classification problem. Students of a col-
sion rule having at least one base is called robust. Identi- lege must obtain an overall evaluation on the basis of their
fication of supporting objects and bases of robust rules is achievements in Mathematics, Physics and Literature. The
important for interpretation of the rules in multiple crite- three subjects are clearly criteria (condition attributes) and
ria decision analysis perspective. The ratio of the number the comprehensive evaluation is a decision attribute. For
of objects supporting a rule and the number of all consid- simplicity, the value sets of the criteria and of the deci-
ered objects is called relative support of a rule. The relative sion attribute are the same, and they are composed of three
support and the confidence ratio are basic characteristics values: bad, medium and good. The preference order of
of a rule, however, some Bayesian confirmation measures these values is obvious. Thus, there are three preference
reflect much better the attractiveness of a rule [25]. ordered decision classes, so the problem belongs to the
A set of decision rules is complete if it covers all con- category of ordinal classification. In order to build a pref-
sidered objects (examples of ordinal classification) in such erence model of the jury, DRSA is used to analyze a set
a way that consistent objects are re-assigned to their origi- of exemplary evaluations of students provided by the jury.
nal classes, and inconsistent objects are assigned to clusters These examples of ordinal classification constitute an in-
of classes referring to this inconsistency. A set of decision put preference information presented as decision table in
rules is minimal if it is complete and non-redundant, i. e. Table 3.
exclusion of any rule from this set makes it incomplete. Note that the dominance principle obviously applies to
Note that the syntax of decision rules induced from the examples of ordinal classification, since an improve-
rough approximations defined using dominance cones, ment of a student’s score on one of three criteria, with
is using consistently this type of granules. Each condi- other scores unchanged, should not worsen the student’s
p
tion profile defines a dominance cone in "V (pn) -dimen- overall evaluation, but rather improve it.
 Rough Sets in Decision Making R 7763

Rough Sets in Decision Making, Table 3 Rough Sets in Decision Making, Table 5
Exemplary evaluations of students (examples of ordinal classifi- Exemplary evaluations of students excluding Mathematics
cation)
Student Physics Literature Overall evaluation
Student Mathematics Physics Literature Overall S1 medium bad bad
evaluation S2 medium bad medium
S1 good medium bad bad S3 medium medium medium
S2 medium medium bad medium S4 good medium good
S3 medium medium medium medium S5 medium good good
S4 good good medium good S6 good good good
S5 good medium good good S7 bad bad bad
S6 good good good good S8 bad medium bad
S7 bad bad bad bad
S8 bad bad medium bad Rough Sets in Decision Making, Table 6
Exemplary evaluations of students excluding Physics

Rough Sets in Decision Making, Table 4 Student Mathematics Literature Overall evaluation
Exemplary evaluations of students excluding Literature S1 good bad bad
S2 medium bad medium
Student Mathematics Physics Overall evaluation S3 medium medium medium
S1 good medium bad S4 good medium good
S2 medium medium medium S5 good good good
S3 medium medium medium S6 good good good
S4 good good good S7 bad bad bad
S5 good medium good S8 bad medium bad
S6 good good good
S7 bad bad bad
S8 bad bad bad on Physics, one obtains Table 6, where no new inconsis-
tencies occur, comparing to Table 3.
The fact that no new inconsistency occurs when Math-
ematics or Physics is removed, means that the subsets
Observe that student S1 has not worse evaluations than of criteria {Physics, Literature} or {Mathematics, Litera-
student S2 on all the considered criteria, however, the ture} contain sufficient information to represent the over-
overall evaluation of S1 is worse than the overall evalua- all evaluation of students with the same quality of ap-
tion of S2. This contradicts the dominance principle, so proximation as using the complete set of three criteria.
the two examples of ordinal classification, and only those, This is not the case, however, for the subset {Mathematics,
are inconsistent. The quality of approximation of the or- Physics}. Observe, moreover, that subsets {Physics, Litera-
dinal classification represented by examples in Table 3 is ture} and {Mathematics, Literature} are minimal, because
equal to 0.75. no other criterion can be removed without new inconsis-
One can observe that in result of reducing the set tencies occur. Thus, {Physics, Literature} and {Mathemat-
of considered criteria, i. e. the set of considered subjects, ics, Literature} are the reducts of the complete set of cri-
some new inconsistencies can occur. For example, remov- teria {Mathematics, Physics, Literature}. Since Literature
ing from Table 3 the evaluation on Literature, one obtains is the only criterion which cannot be removed from any
Table 4, where S1 is inconsistent not only with S2, but also reduct without introducing new inconsistencies, it consti-
with S3 and S5. In fact, student S1 has not worse evalu- tutes the core, i. e. the set of indispensable criteria. The
ations than students S2, S3 and S5 on all the considered core is, of course, the intersection of all reducts, i. e. in our
criteria (Mathematics and Physics), however, the overall example:
evaluation of S1 is worse than the overall evaluation of S2,
S3 and S5. fLiteratureg
Observe, moreover, that removing from Table 3 the D fPhysics, Literatureg\fMathematics, Literatureg:
evaluations on Mathematics, one obtains Table 5, where
no new inconsistencies occur, comparing to Table 3. In order to illustrate in a simple way the con-
Similarly, after removing from Table 3 the evaluations cept of rough approximation, let us confine our anal-
7764 R Rough Sets in Decision Making

ysis to the reduct {Mathematics, Literature}. Let us information about evaluation of students on Mathematics
consider student S4. His positive dominance cone and Literature, the fact that student y dominates S4 or S5
C
DfMathematics; Literatureg
(S4) is composed of all the students or S6 is a possible condition to conclude that y is a “good”
having evaluations not worse than him on Mathematics student.
and Literature, i. e. of all the students dominating him with Let us observe that for the set of criteria P = {Mathe-
respect to Mathematics and Literature. Thus, matics, Literature}, the lower and upper approximations
of the set of “good” students are the same. This means
C
DfMathematics, Literatureg
(S4) D fS4; S5; S6g : that examples of ordinal classification concerning this de-
cision class are all consistent. This is not the case, how-
On the other hand, the negative dominance cone of
 ever, for the examples concerning the union of decision
student S4; DfMathematics, Literatureg
(S4), is composed of all
classes “at least medium”. For this upward union the
the students having evaluations not better than him on 
rough approximations are P(Clmedium ) D fS3; S4; S5; S6g
Mathematics and Literature, i. e. of all the students domi- 
and P(Clmedium ) D fS1; S2; S3; S4; S5; S6g. The difference
nated by him with respect to Mathematics and Literature.  
Thus, between P(Clmedium ) and P(Clmedium ), i. e. the boundary

Bn P (Clmedium ) D fS1; S2g, is composed of students with

DfMathematics, Literatureg (S4) D fS1; S2; S3; S4; S7; S8g : inconsistent overall evaluations, which has already been
noticed above. From the viewpoint of decision making,
Similar dominance cones can be obtained for all the this means that, taking into account the available infor-
students from Table 6. For example, for S2, the dominance mation about evaluation of students on Mathematics and
cones are Literature, the fact that student y is dominated by S1 and
C dominates S2 is a condition to conclude that y can obtain
DfMathematics, Literatureg
(S2) D fS1; S2; S3; S4; S5; S6g
an overall evaluation “at least medium” with some doubts.
and Until now, rough approximations of only upward
unions of decision classes have been considered. It is in-

DfMathematics, Literatureg (S2) D fS2; S7g : teresting, however, to calculate also rough approximations
of downward unions of decision classes. Let us consider
The rough approximations can be calculated using first the lower approximation of the set of students hav-
dominance cones. Let us consider, for example, the lower ing “at most medium” overall evaluation P(Clmedium 
). Ob-
approximation of the set of students having a “good” over- 
 serve that P(Clmedium ) D fS1; S2; S3; S7; S8g, because the
all evaluation P(Clgood ), with P = {Mathematics, Litera-
negative dominance cones of students S1, S2, S3, S7, and

ture}. Notice that P(Clgood ) D fS4; S5; S6g, because pos- S8 are all included in the set of students with overall eval-
itive dominance cones of students S4, S5 and S6 are all uation “at most medium”. In other words, this means that
included in the set of students with an overall evaluation there is no student dominated by S1 or S2 or S3 or S7 or S8
“good”. In other words, this means that there is no student while having an overall evaluation better than “medium”.
dominating S4 or S5 or S6 while having an overall evalu- From the viewpoint of decision making, this means that,
ation worse than “good”. From the viewpoint of decision taking into account the available information about evalu-
making, this means that, taking into account the available ation of students on Mathematics and Literature, the fact
information about evaluation of students on Mathematics that student y is dominated by S1 or S2 or S3 or S7 or S8
and Literature, the fact that student y dominates S4 or S5 is a sufficient condition to conclude that y is an “at most
or S6 is a sufficient condition to conclude that y is a “good” medium” student.
student. The upper approximation of the set of students
The upper approximation of the set of students with with an “at most medium” overall evaluation is
 
a “good” overall evaluation is P(Clgood ) D fS4; S5; S6g, P(Clmedium ) D fS1; S2; S3; S7; S8g, because the positive
because negative dominance cones of students S4, S5 and dominance cones of students S1, S2, S3, S7, and S8 have
S6 have a nonempty intersection with the set of students a nonempty intersection with the set of students having an
having a “good” overall evaluation. In other words, this “at most medium” overall evaluation. In other words, this
means that for each one of the students S4, S5 and S6, there means that for each one of the students S1, S2, S3, S7, and
is at least one student dominated by him with an over- S8, there is at least one student dominating him with an
all evaluation “good”. From the point of view of decision overall evaluation “at most medium”. From the viewpoint
making, this means that, taking into account the available of decision making, this means that, taking into account
 Rough Sets in Decision Making R 7765

the available information about evaluation of students on ered from the available information. In the current con-
Mathematics and Literature, the fact that student y is dom- text, the knowledge is interpreted as a preference model
inated by S1 or S2 or S3 or S7 or S8 is a possible condition of the jury. A characteristic feature of the syntax of de-
to conclude that y is an “at most medium” student. cision rules representing preferences is the use of expres-
Finally, the lower and upper approximations of the sions “at least” or “at most” a value; in case of extreme val-
set of students having a “bad” overall evaluation are ues (“good” and “bad”), these expressions are put in paren-
 
P(Clbad ) D fS7; S8g and P(Clbad ) D fS1; S2; S7; S8g. The theses because there is no value above “good” and below
  “bad”.
difference between P(Clbad ) and P(Clbad ), i. e. the bound-
 Even if one can represent all the knowledge using
ary Bn P (Clbad ) D fS1; S2g is composed of students with
inconsistent overall evaluations, which has already been only one reduct of the set of criteria (as it is the case
noticed above. From the point of view of decision making, using P = {Mathematics, Literature}), when considering
this means that, taking into account the available infor- a larger set of criteria than a reduct, one can obtain a more
mation about evaluation of students on Mathematics and synthetic representation of knowledge, i. e. the number of
Literature, the fact that student y is dominated by S1 and decision rules or the number of elementary conditions, or
dominates S2 is a condition to conclude that y can obtain both of them, can get smaller. For example, considering
an overall evaluation “bad” with some doubts. Observe, the set of all three criteria, {Mathematics, Physics, Litera-
  ture}, one can induce a set of decision rules composed of
moreover, that Bn P (Clmedium ) D Bn P (Clbad ) D fS1; S2g.
Given the above rough approximations with respect to the above rules 1), 2), 3) and 6), plus the following:
the set of criteria P = {Mathematics, Literature}, one can rule 7) if the evaluation on Physics is at most medium,
induce a set of decision rules representing the preferences and the evaluation on Literature is at most medium,
of the jury. The idea is that evaluation profiles of students then the overall evaluation is at most medium,
belonging to the lower approximations can serve as a base fS1; S2; S3; S7; S8g.
for some certain rules, while evaluation profiles of stu- Thus, the complete set of decision rules induced from
dents belonging to the boundaries can serve as a base for Table 3 is composed of 5 instead of 6 rules.
some approximate rules. The following decision rules have Once accepted by the DM, these rules represent his/her
been induced (between parentheses there are id’s of stu- preference model. Assuming that rules 1)–7) in our ex-
dents supporting the corresponding rule; the student be- ample represent the preference model of the jury, it can
ing a rule base is underlined): be used to evaluate new students. For example, student S9
who is “medium” in Mathematics and Physics and “good”
rule 1) if the evaluation on Mathematics is (at least)
in Literature, would be evaluated as “medium” because his
good, and the evaluation on Literature is at least
profile matches the premise of rule 2), having as conse-
medium, then the overall evaluation is (at least) good,
quence an overall evaluation at least “medium”. The over-
fS4; S5; S6g,
all evaluation of S9 cannot be “good”, because his profile
rule 2) if the evaluation on Mathematics is at least
does not match any rule having as consequence an over-
medium, and the evaluation on Literature is at
all evaluation “good” (in the considered example, the only
least medium, then the overall evaluation is at least
rule of this type is rule 1) whose premise is not matched by
medium, fS3; S4; S5; S6g,
the profile of S9.
rule 3) if the evaluation on Mathematics is at least
medium, and the evaluation on Literature is (at most)
bad, then the overall evaluation is bad or medium, DRSA on a Pairwise Comparison Table
fS1; S2g, for Multiple Criteria Choice and Ranking Problems
rule 4) if the evaluation on Mathematics is at least Multiple Criteria Choice and Ranking Problems
medium, then the overall evaluation is at least
Ordinal classification decisions are based on absolute eval-
medium, fS2; S3; S4; S5; S6g,
uation of objects on multiple criteria, however, multiple
rule 5) if the evaluation on Literature is (at most) bad,
criteria choice and ranking decisions are based on pref-
then the overall evaluation is at most medium,
erence relations between objects. Decision table including
fS1; S2; S7g,
examples of ordinal classification does not contain infor-
rule 6) if the evaluation on Mathematics is (at most) bad,
mation about preference relations between objects, thus,
then the overall evaluation is (at most) bad, fS7; S8g.
in order to apply DRSA to multiple criteria choice and
Notice that rules 1)–2), 4)–7) are certain, while rule 3) is an ranking problems, a different representation of the input
approximate one. These rules represent knowledge discov- preference information is needed.
7766 R Rough Sets in Decision Making

To handle binary relations within the rough set ap- and g[(x; y); d] 2 Td ; 8(x; y) 2 B̂. It follows that for
proach, it has been proposed in [14,15] to operate on, so- any pair of reference objects (x; y) 2 B̂ there is verified
called, pairwise comparison table (PCT), i. e. a decision ta- one and only one binary relation t 2 Td . Thus, T d induces
ble including pairs of objects for which multiple criteria a partition of B̂. In fact, information table SPCT can be seen
evaluations and a comprehensive preference relation are as decision table, since the set of considered criteria F and
known. PCT represents preference information provided decision d are distinguished.
by the DM in the form of decision examples (pairwise It is assumed that the exemplary pairwise compar-
comparisons of objects). isons provided by the DM can be represented in terms of
Similarly to ordinal classification, decision examples graded preference relations (for example “very weak pref-
concerning multiple criteria choice and ranking may be erence”, “weak preference”, “strict preference”, “strong
inconsistent with respect to the dominance principle, preference”, “very strong preference”) Pih : 8i 2 I and
however, interpretation of the inconsistency is different: 8(x; y) 2 U  U,
It occurs when preferences of a pair of objects, say (a; b), n o
on all considered criteria are not weaker than preferences Ti D Pih ; h 2 H i ;
of another pair of objects, say (c; d), on these criteria, how-
ever, the comprehensive preference of object a over ob- where H i is a particular subset of the relative integers and
ject b is weaker than the comprehensive preference of ob-
ject c over object d.  xPih y; h > 0, means that object x is preferred to ob-
ject y by degree h with respect to criterion i,
 xPih y; h < 0, means that object x is not preferred to ob-
The Pairwise Comparison Table
ject y by degree h with respect to criterion i,
Similarly to the ordinal classification, let us consider a fi-  xPi0 y means that object x is similar (asymmetrically in-
nite set of criteria F D f f1 ; : : : ; f n g, the set of their in- different) to object y with respect to criterion i.
dices I D f1; : : : ; ng and a finite universe of objects (ac-
tions, solutions, alternatives) U. For any criterion i 2 I, Of course, 8i 2 I and 8(x; y) 2 U  U, it holds:
let T i be a finite set of binary relations defined on U on h i h i
the basis of the evaluations of objects from U with respect xPih y; h > 0 ) yPik x; k  0 :
to the considered criterion i, such that 8(x; y) 2 U  U
exactly one binary relation t 2 Ti is verified; t 2 Ti has The set of binary relations T d may be defined in a sim-
the meaning of a preference relation for a pair of objects ilar way, but xPdh y means that object x is comprehensively
on a particular criterion i. More precisely, given value preferred to object y by degree h.
set V i of i 2 I, if v 0i ; v 00i 2 Vi are the respective evalua- Technically, the modeling of the binary relation Pih ,
tions of x; y 2 U on criterion i, and (x; y) 2 t, with t 2 Ti , i. e. the assessment of h, can be organized as follows:
then for each w; z 2 U having the same evaluations v 0i ; v 00i
on i; (w; z) 2 t. For interesting applications it should be  first, it is observed that criterion i is a function
jTq j  2; 8i 2 I. Furthermore, let T d be a set of binary re- f i : U ! < increasing with respect to the preferences
lations defined on U, such that at most one binary relation on i, for each i D 1; : : : ; n,
t 2 Td is verified 8(x; y) 2 U  U; t 2 Td has the mean-  then, for each i D 1; : : : ; n, it is possible to define
ing of a comprehensive preference relation for a pair of ob- a function k i : <2 ! < which measures the strength of
jects (comprehensive pairwise comparison). the preference (positive or negative) of x over y (e. g.
The preference information provided by the DM, k i [ f i (x); f i (y)] D f i (x)  f i (y)); it should satisfy the
has the form of pairwise comparisons of some refer- following properties 8x; y; z 2 U:
ence objects from B  U. These decision examples are i) f i (x) > f i (y) , k i [ f i (x); f i (z)] > k i [ f i (y);
presented in the pairwise comparison table (PCT), de- f i (z)],
fined as information table SPCT D hB̂; F [ fdg; TF [ ii) f i (x) > f i (y) , k i [ f i (z); f i (x)] < k i [ f i (z);
Td ; gi, where B̂  B  B is a non-empty set of exemplary f i (y)],
S
pairwise comparisons of reference objects, TF D i2I Ti ; d iii) f i (x) D f i (y) , k i [ f i (x); f i (y)] D 0,
is a decision corresponding to the comprehensive pair-  next, the domain of ki can be divided into inter-
wise comparison (comprehensive preference relation), vals, using a suitable set of thresholds  i ; 8i 2 I;
and g : B̂  (F [ fdg) ! TF [ Td is a total function these intervals are numbered in such a way that
such that g[(x; y); i] 2 Ti ; 8(x; y) 2 U  U and 8i 2 I, k i [ f i (x); f i (y)] D 0 belongs to interval no. 0,
 Rough Sets in Decision Making R 7767

 the value of h in the relation xPih y is then equal to the P(S) may be interpreted as the dominance relation
h
number of the interval including k i [ f i (x); f i (y)], for DCP having the largest intersection with B̂ included in the
any (x; y) 2 U  U. outranking relation S, and P(S) as the dominance relation
h
DCP including S and having the smallest intersection with
Actually, property iii) can be relaxed in order to obtain
B̂.
a more general preference model which, for instance, does
Analogously, it is possible to approximate the relation
not satisfy preferential independence. h
Sc by means of the dominance relation DP . Observe that,
To simplify the presentation, let us consider a PCT
in general, the definitions of the approximations of S and
where the set T d is composed of two binary relations de-
Sc do not satisfy the condition of complementarity, i. e. it
fined on U:
is not true, in general, that P(S) is equal to B̂  P(S c ) and
 x outranks y (denotation xSy or (x; y) 2 S), where that P(S c ) is equal to B̂  P(S). This is because S and Sc
h
(x; y) 2 B̂, are approximated using two different relations, DCP and
h
 x does not outrank y (denotation xS c y or (x; y) 2 S c ), DP , respectively. Nevertheless, the approximations thus
where (x; y) 2 B̂, obtained constitute a good basis for the generation of sim-
ple decision rules.
and S [ S c D B̂, where “x outranks y” means “x is at least
as good as y”; observe that the binary relation S is reflex-
ive, but neither necessarily transitive nor complete. In [8], Decision Rules
a more general PCT was considered, where the set T d is It is possible to represent preferences of the DM revealed
composed of multi-graded binary relations defined on U. in terms of exemplary pairwise comparisons contained in
a given PCT, using decision rules. Since approximations
Approximation by Means of S and Sc were made using graded dominance relations,
of Graded Dominance Relations it is possible to induce decision rules being propositions of
T the following type:
Let H P D i2P H i ; 8P  I. Given P  I and h 2 H P ;
8(x; y) 2 U  U it is said that x positively dominates y  DCC -decision rule: if xDCPh y, then xSy,

by degree h with respect to criteria from P iff xPic i y with  h

DC -decision rule: if not xDCP y, then xS c y,
c i  h; 8i 2 P. Analogously, 8(x; y) 2 U  U; x neg-  h y, then xSy,
DC -decision rule: if not xDP
atively dominates y by degree h with respect to criteria  h
D -decision rule: if xDP y, then xS c y,
from P iff xPic i y with c i  h; 8i 2 P. Therefore, each
P  I and h 2 H P generate two binary relations (possi- where P is a non-empty subset of I. Therefore, for example,
bly empty) on U, called positive P-dominance by degree h a DCC – decision rule is a proposition of the type: “if x
h positively dominates y by degree h with respect to criteria
(denotation DCP ) and negative P-dominance by degree h
h
(denotation DP ), respectively. They satisfy the following from P, then x outranks y”.
conditions: A constructive definition of these rules may be given,
being a kind of implication supported by the existence of
h , then (x; y) 2 D h
(P1) if (x; y) 2 DCP CR for each R  P at least one pair of objects from B̂ satisfying one of the four
and k  h; propositions listed above, and by the absence of pairs from
h h
(P2) if (x; y) 2 DP , then (x; y) 2 DR for each R  P B̂ contradicting it. Thus, for example, if
and k  h.
 there exists at least one pair (w; z) 2 B̂ such that
In [15], it has been proposed to approximate the out- wDCPh z and wSz and
ranking relation S by means of the dominance relation  there does not exist any pair (v; u) 2 B̂ such that
h
DCP . Therefore, S is considered a rough binary relation. h u and vS c u,
vDCP
The P-lower approximation of S (denotation P(S)) and h
 then “if xDCP y, then xSy” is accepted as a DCC – deci-
the P-upper approximation of S (denotation P(S)) are de- sion rule.
fined, respectively, as:
A DCC -decision rule “if xDCP h y, then xSy” is said
[ n o
h to be minimal if there does not exist any other rule “if
P(S) D DCP \ B̂  S ;
k
h2H P xDCR y, then xSy” such that R  P and k  h. Analogous
\ n o
P(S) D h
DCP \ B̂  S : definitions hold for the other cases. In other words, a mini-
h2H P
mal decision rule is a kind of implication for which there is
7768 R Rough Sets in Decision Making

no other implication whose premise is of at least the same  not uSv and not uS c v, that is unknown outranking (de-
weakness and whose consequence is of at least the same notation uS U v).
strength.
The four above situations, which together constitute
The following results show connections of the decision
the so-called four-valued outranking (see [56]), have been
rules with the P-lower and P-upper approximations of S
introduced to underline the presence and the absence of
and Sc [15]:
positive and negative reasons for the outranking. More-
 DCC -minimal decision rule “if xDCP h
y, then xSy” is over, they make it possible to distinguish contradictory sit-
supported by pairs of objects belonging to P(S) D uations from unknown ones.
DCPh
\ B̂, The following theorem underlines the operational im-
 D -minimal decision rule “if xDP h
y, then xS c y” is portance of the minimal decision rules [15]: The applica-
supported by pairs of objects belonging to P(S c ) D tion of all the decision rules obtained for a given SPCT to
DPh
\ B̂, a pair of objects (u; v) 2 M  M results in the same out-
 DC -minimal decision rule “if not xDCP h
y, then xS c y” ranking relations S and Sc as those obtained from the ap-
plication of the minimal decision rules only. Therefore, the
is supported by pairs of objects belonging to P(S) D
h \ B̂, set of the minimal decision rules totally characterizes the
DCP
h preferences of the DM contained in SPCT .
 DC -minimal decision rule “if not xDP y, then xSy”
A final recommendation can be obtained upon a suit-
is supported by pairs of objects belonging to P(S c ) D
h able exploitation of the presence and the absence of out-
DP \ B̂.
ranking S and Sc on M. A possible exploitation proce-
dure consists in calculating a specific score, called Net Flow
Score, for each object x 2 M:
Application of the Decision Rules
and Final Recommendation Snf (x) D S CC (x)  S C (x) C S C (x)  S  (x) ;
In order to obtain a recommendation in the multiple cri-
teria choice or ranking problems with respect to a set of where
objects M  U, the decision rules induced from the ap-
proximations of S and Sc (defined with respect to reference S CC (x) D jfy 2 M : there is at least one decision
rule which affirms xSygj ;
objects from B) should be applied on set M  M. The ap-
plication of the rules to any pair of objects (u; v) 2 M  M S C (x) D jfy 2 M : there is at least one decision
establishes the presence (uSv) or the absence (uS c v) of rule which affirms ySxgj ;
outranking with respect to (u; v). More precisely, S C (x) D jfy 2 M : there is at least one decision
h
rule which affirms yS c xgj ;
 from DCC -decision rule “if xDCP y then xSy” and
h S  (x) D jfy 2 M : there is at least one decision
from uDCP v, one concludes uSv,
h
rule which affirms xS c ygj :
 from DC -decision rule “if not xDCP y then xS c y” and
h v, one concludes uS c y,
from not uDCP The recommendation in multiple criteria ranking
h problems consists of the total preorder determined by
 from DC -decision rule “if not xDP y, then xSy” and
h
from not uDP v, one concludes uSv, Snf (x) on M; in multiple criteria choice problems it
 from D -decision rule “if xDP h
y, then xS c y” and consists of the object(s) x  2 M such that Snf (x  ) D
h c maxx2M Snf (x).
from uDP v, one concludes uS v.
The procedure described above has been characterized
After the application of the decision rules to each pair with reference to a number of desirable properties in [13].
of objects (u; v) 2 M  M, one of the following four situ-
ations may occur: Approximation by Means
 uSv and not uS c v,
that is true outranking (denotation of Multi-graded Dominance Relations
uS T v), The graded dominance relation introduced above assumes
 uS c v and not uSv, that is false outranking (denotation a common grade of preference for all the considered cri-
uS F v), teria. While this permits a simple calculation of the ap-
 uSv and uS c v, that is contradictory outranking (deno- proximations and of the resulting decision rules, it is lack-
tation uS K v), ing in precision. A dominance relation allowing a different
 Rough Sets in Decision Making R 7769

degree of preference for each considered criterion (multi- P(S c )  S c  P(S c ) :

graded dominance) gives a far more accurate picture of the
preference information contained in the pairwise compar- Furthermore, the following complementarity proper-
ison table SPCT [14,16,18]. ties hold:
More formally, given P  I (P ¤ ;); (x; y); (w; z) 2
P(S) D B̂  P(S c ); P(S) D B̂  P(S c ) ;
U  U; (x; y) is said to dominate (w; z) with respect to cri-
teria from P (denotation (x; y)D P (w; z)) if x is preferred
P(S c ) D B̂  P(S); P(S c ) D B̂  P(S) :
to y at least as strongly as w is preferred to z with respect
to each i 2 P. Precisely, “at least as strongly as” means
The P-boundaries (P-doubtful regions) of S and Sc are
“by at least the same degree”, i. e. hi  ki, where hi; ki 2
defined as
H i ; xPihi y and wPiki z; 8i 2 P. Let Dfig be the domi-
nance relation confined to the single criterion i 2 P. The
Bn P (S) D P(S)  P(S); Bn P (S c ) D P(S c )  P(S c ) :
binary relation Dfig is reflexive ((x; y)Dfig (x; y); 8(x; y) 2
U  U), transitive ((x; y)Dfig (w; z) and (w; z)Dfig (u; v) It is easy to prove that Bn P (S) D Bn P (S c ).
imply (x; y)Dfig (u; v); 8(x; y); (w; z); (u; v) 2 U U), and The concepts of quality of approximation, reducts and
complete ((x; y)Dfig (w; z) or (w; z)Dfig (x; y); 8(x; y); core can be extended also to the approximation of the out-
(w; z) 2 U  U). Therefore, Dfig is a complete preorder ranking relation by multi-graded dominance relations. In
on U  U. Since the intersection of complete preorders is
T particular,
a partial preorder and D P D i2P Dfig ; P  I, then the
dominance relation DP is a partial preorder on U  U. jP(S) [ P(S c )j
Let R  P  I and (x; y); (u; v) 2 U  U; then the fol- P D ˇ ˇ
ˇ B̂ˇ
lowing implication holds:
defines the quality of approximation of S and Sc by P  I.
(x; y)D P (u; v) ) (x; y)D R (u; v) :
It expresses the ratio of all pairs of objects (x; y) 2 B̂ cor-
Given P  I and (x; y) 2 U  U, let us introduce the rectly assigned to S and Sc by the set P of criteria, to all
positive dominance set (denotation DC the pairs of objects contained in B̂. Each minimal subset
P (x; y)) and the
negative dominance set (denotation D (x; y)): P0  P such that P 0 D P is called a reduct of P (denota-
P
tion RED PC T (P)). Let us remark that SPCT can have more
DC
P (x; y)) D f(w; z) 2 U  U : (w; z)D P (x; y)g ;
than one reduct. The intersection of all reducts is called the
core (denotation CORE PC T (P)).
D
P (x; y)) D f(w; z) 2 U  U : (x; y)D P (w; z)g :
Using the approximations defined above, it is then
possible to induce a generalized description of the pref-
Using the dominance relation DP , it is possible to de- erence information contained in a given SPCT in terms of
fine P-lower and P-upper approximations of the outrank- suitable decision rules. The syntax of these rules is based
ing relation S with respect to P  I, respectively, as: on the concept of upward cumulated preferences (denota-
˚  tion Pih ) and downward cumulated preferences (denota-
P(S) D (x; y) 2 B̂ : DC P (x; y)  S ; tion Pih ), having the following interpretation:
[
P(S) D DC
P (x; y) :  xPih y means “x is preferred to y with respect to i by at
(x;y)2S
least degree h”,
Analogously, it is possible to define the approxima-  xPih y means “x is preferred to y with respect to i by at
tions of Sc : most degree h”.
˚  Exact definition of the cumulated preferences, for each
P(S c ) D (x; y) 2 B̂ : D
P (x; y)  S
c
;
(x; y) 2 U  U; i 2 I and h 2 H i , is the following:
[
P(S c ) D D
P (x; y) :
 xPih y if xPik y, where k 2 H i and k  h,
(x;y)2S c
 xPih y if xPik y, where k 2 H i and k  h.
It may be proved that
Using the above concepts, three types of decision rules
P(S)  S  P(S) can be obtained:
7770 R Rough Sets in Decision Making

1. D – decision rules, being statements of the type: a strength of preference, while, for the latter, this concept
h(i p)
is meaningless. From this point of view, preferences ex-
h(i1) h(i2)
if xPi1 y and xPi2 y and : : : xPi p y; then xSy; pressed on an ordinal scale refer to ordinal utility, while
preferences expressed on a quantitative scale or a numeri-
where P D fi1; : : : ; i pg  I and (h(i1); : : : ; h(i p)) 2
cal non-quantitative scale deal with cardinal utility.
H i1   H i p ; these rules are supported by pairs of
The strength of preference ki and, therefore, the graded
objects from the P-lower approximation of S only;
preference considered in Subsect. “Approximation by
2. D – decision rules, being statements of the type:
Means of Multi-graded Dominance Relations”, is mean-
h(i1) h(i2) h(i p) ingful when the scale is quantitative or numerical non-
if xPi1 y and xPi2 y and : : : xPi p y; then xS c y;
quantitative. If the information about ki is non-available,
where P D fi1; : : : ; i pg  I and (h(i1); : : : ; h(i p)) 2 then it is possible to define a rough approximation of S
H i1   H i p ; these rules are supported by pairs of and Sc using a specific dominance relation between pairs
objects from the P-lower approximation of Sc only; of objects from U  U, defined on an ordinal scale repre-
3. D – decision rules, being statements of the type: sented by evaluations f i (x) on criterion i, for x 2 U [14].
h(i1) h(i2)
Let us explain this latter case in more details.
if xPi1 y and xPi2 y and : : : xPih(i
k
k)
y and Let I O be the set of criteria expressing preferences on
xPih(i kC1) h(i p)
y and : : : xPi p y; then xSy or xS c y ; an ordinal scale, and I N , the set of criteria expressing pref-
kC1
erences on a quantitative scale or a numerical non-quan-
where O 0 D fi1; : : : ; ikg  I; O 00 D fik C1; : : : ; i pg  titative scale, such that I O [ I N D I and I O \ I N D ;.
I; P D O 0 [ O 00 ; O 0 and O 00 not necessarily disjoint, Moreover, for each P  I; PO denotes the subset of P
(h(i1); : : : ; h(i p)) 2 H i1  H i2  H i p ; these rules are composed of criteria expressing preferences on an ordinal
supported by objects from the P-boundary of S and Sc scale, i. e. PO D P \ I O , and PN the subset of P composed
only. of criteria expressing preferences on a quantitative scale or
a numerical non-quantitative scale, i. e. PN D P \ I N . Of
course, for each P  I; P D PN [ PO and PO \ PN D ;.
Dominance Without Degrees of Preference If P D PN and PO D ;, then the definition of
The degree of graded preference considered in Sub- dominance is the same as in the case of multi-
sect. “The Pairwise Comparison Table” is defined on graded dominance (Subsect. “Approximation by Means
a quantitative scale of the strength of preference k i ; i 2 I. of Multi-graded Dominance Relations”). If P D PO and
However, in many real world problems, the existence of PN D ;, then, given (x; y); (w; z) 2 U  U, the pair
such a quantitative scale is rather questionable. Roy [45] (x; y) is said to dominate the pair (w; z) with re-
distinguishes the following cases: spect to P if, for each i 2 P; f i (x)  f i (w) and
f i (z)  f i (y). Let Dfig be the dominance relation con-
 Preferences expressed on an ordinal scale: this is the
fined to the single criterion i 2 PO . The binary rela-
case where the difference between two evaluations has
tion Dfig is reflexive ((x; y)Dfig (x; y); 8(x; y) 2 U 
no clear meaning;
U), transitive ((x; y)Dfig (w; z) and (w; z)Dfig (u; v) imply
 Preferences expressed on a quantitative scale: this is the
(x; y)Dfig (u; v); 8(x; y); (w; z); (u; v) 2 U  U), but non-
case where the scale is defined with reference to a unit
complete (it is possible that not (x; y)Dfig (w; z) and not
clearly identified, such that it is meaningful to consider
(w; z)Dfig (x; y) for some (x; y); (w; z) 2 U  U). There-
an origin (zero) of the scale and ratios between evalua-
fore, Dfig is a partial preorder. Since the intersection of
tions (ratio scale);
partial preorders is also a partial preorder and D P D
 Preferences expressed on a numerical non-quantitative T O
scale: this is an intermediate case between the previous i2P Dfig ; P D P , then the dominance relation DP is
also a partial preorder. If some criteria from P  I express
two; there are two well-known particular cases:
preferences on a quantitative or a numerical non-quantita-
 Interval scale, where it is meaningful to compare ra-
tive scale and others on an ordinal scale, i. e. if PN ¤ ; and
tios between differences of pairs of evaluations,
PO ¤ ;, then, given (x; y); (w; z) 2 U  U, the pair (x; y)
 Scale for which a complete preorder can be defined
is said to dominate the pair (w; z) with respect to criteria
on all possible pairs of evaluations.
from P, if (x; y) dominates (w; z) with respect to both PN
The preference scale has also been considered within and PO . Since the dominance relation with respect to PN is
economic theory (e. g. [47]), where cardinal utility is dis- a partial preorder on U  U (because it is a multi-graded
tinguished from ordinal utility: the former deals with dominance) and the dominance with respect to PO is also
 Rough Sets in Decision Making R 7771

Rough Sets in Decision Making, Table 7

a partial preorder on U  U (as explained above), then Information table of the illustrative example
also the dominance DP , being the intersection of these two
dominance relations, is a partial preorder. In consequence, Warehouse A1 A2 A3 A4
all the concepts introduced in the previous subsection can 1 high good no profit
be restored using this specific definition of dominance 2 medium medium no loss
relation. 3 medium medium no profit
Using the approximations of S and Sc based on the 4 low medium no loss
dominance relation defined above, it is possible to in- 5 medium good yes loss
duce a generalized description of the available preference 6 high medium yes profit
information in terms of decision rules. These decision
rules are of the same type as the rules already intro-
duced in the previous subsection, however, the condi- H i1   H i f and (r i f C1 ; : : : ; r i p ); (s i f C1 ; : : : ; s i p ) 2
tions on criteria from I O are expressed directly in terms F i f C1   F i p ; these rules are supported by pairs of
of evaluations belonging to value sets of these criteria. objects from the P-boundary of S and Sc only.
Let F i D f f i (x); x 2 Ug; i 2 I O . The decision rules have in
this case the following syntax:
1. D -decision rule, being a statement of the type: Example Illustrating DRSA in the Context
of Multiple Criteria Choice and Ranking
h(i1)
if xPi1 y and : : : xPih(i
e
e)
y and f i eC1(x)  r i eC1
The following example illustrates DRSA in the context of
and f i eC1(y)  s i eC1 and : : : f i p (x)  r i p multiple criteria choice and ranking. Six warehouses have
and f i p (y)  s q p ; then xSy ; been described by means of four criteria:
where P D fi1; : : : ; i pg  I; PN D fi1; : : : ; ieg; PO D  f 1 , capacity of the sales staff,
fie C 1; : : : ; i pg; (h(i1); : : : ; h(ie)) 2 H i1   Hi e  f 2 , perceived quality of goods,
and (r i eC1; : : : ; r i p ); (s i eC1 ; : : : ; s i p ) 2 F i eC1    f 3 , high traffic location,
F i p ; these rules are supported by pairs of objects from  f 4 , warehouse profit or loss.
the P-lower approximation of S only;
2. D -decision rule, being a statement of the type: The components of the information table S are: U D
h(i1) h(i p) f1; 2; 3; 4; 5; 6g; F D f f1 ; f2 ; f3 ; f4 g; I D f1; 2; 3; 4g; F1 D
if xPi1 y and : : : xPi p y and f i eC1(x)  r i eC1 fhigh, medium, lowg; F2 D fgood, mediumg; F3 D
and f i eC1(y)  s i eC1 and : : : f i p (x)  r i p fno, yesg; F4 D fprofit, lossg, the criterion f i (x), taking
and f i p (y)  s i p ; then xS c y ; values f1 (1) D high, f2 (1) D good, and so on.
It is assumed that the DM accepts to express pref-
where P D fi1; : : : ; i pg  I; PN D fi1; : : : ; ieg; PO D erences with respect to criteria f1 ; f2 ; f3 on a numer-
fie C 1; : : : ; i pg; (h(i1); : : : ; h(ie)) 2 H i1   Hi e ical non-quantitative scale, for which a complete pre-
and (r i eC1 ; : : : ; r i p ), (s i eC1; : : : ; s i p ) 2 F i eC1   order can be defined on all possible pairs of evalua-
F i p ; these rules are supported by pairs of objects from tions. According to this assumption, in order to build the
the P-lower approximation of Sc only; PCT, as described in Subsect. “Approximation by Means
3. D -decision rule, being a statement of the type: of Multi-graded Dominance Relation”, the DM speci-
h(i1) fies sets of possible degrees of preference; for example,
if xPi1 y and : : : xPih(i
e
e)
y and xPih(i
eC1
eC1)
y:::
H1 D f2; 1; 0; 1; 2g; H2 D f1; 0; 1g; H3 D f1; 0; 1g.
h(i f )
xPi f y and f i f C1(x)  r i f C1 Therefore, with respect to f 1 , there are the following pref-
erence relations P1h :
and f i f C1(y)  s i f C1 and : : : f i g (x)  r i g
and f i g (y)  s i g and f i gC1 (x)  r i gC1  xP12 y (and yP12 x), meaning that x is preferred to y
and f i gC1 (y)  s i gC1 and : : : f i p (x)  r i p with respect to f 1 , if f1 (x) D high and f1 (y) D low,
 xP11 y (and yP11 x), meaning that x is weakly preferred
and f i p (y)  s i p ; then xSy or xS c y ;
to y with respect to f 1 , if f1 (x) D high and f1 (y) D
where O 0 D fi1; : : : ; ieg  I; O 00 D fie C 1; : : : ; i f g  medium, or f1 (x) D medium and f1 (y) D low,
I; PN D O 0 [ O 00 ; O 0 and O 00 not necessarily dis-  xP10 y (and yP10 x), meaning that x is indifferent to y with
joint, PO D fi f C 1; : : : ; i pg; (h(i1); : : : ; h(i f )) 2 respect to f 1 , if f1 (x) D f1 (y).
7772 R Rough Sets in Decision Making

Rough Sets in Decision Making, Table 8

Analogously, with respect to f 2 and f 3 , there are the Pairwise comparison table
following preference relations P2h and P3h :
Pairs P1h P2h P3h Outranking
 xP21 y (and yP21 x), meaning that x is weakly pre-
(1,1) 0 0 0 S
ferred to y with respect to f 2 , if f2 (x) D good and (1,2) 1 1 0 S
f2 (y) D medium, (1,4) 2 1 0 S
 xP20 y (and yP20 x), meaning that x is indifferent to y with (1,5) 1 0 1 S
respect to f 2 , if f2 (x) D f2 (y), (2,1) 1 1 0 Sc
 xP31 y (and yP31 x), meaning that x is weakly preferred (2,2) 0 0 0 S
to y with respect to f 3 , if f3 (x) D yes and f3 (y) D no, (2,3) 0 0 0 Sc
 xP30 y (and yP30 x), meaning that x is indifferent to y with (2,6) 1 0 1 Sc
respect to f 3 , if f3 (x) D f3 (y). (3,2) 0 0 0 S
(3,3) 0 0 0 S
As to the comprehensive preference relation, the DM (3,4) 1 0 0 S
considers that, given two different warehouses x; y 2 U D (3,5) 0 1 1 S
f1; 2; 3; 4; 5; 6g, if x makes profit and y makes loss, then xSy (4,1) 2 1 0 Sc
and yS c x. Moreover, the DM accepts xSx for each ware- (4,3) 1 0 0 Sc
house x. As to warehouses x and y, which both make profit (4,4) 0 0 0 S
or both make loss, the DM abstains from judging whether (4,6) 2 0 1 Sc
xSy or xS c y. Therefore, the set of exemplary pairwise com- (5,1) 1 0 1 Sc
parisons supplied by the DM is B̂ D f(1, 1), (1, 2), (1, 4), (5,3) 0 1 1 Sc
(1, 5), (2, 1), (2, 2), (2, 3), (2, 6), (3, 2), (3, 3), (3, 4), (3, 5), (5,5) 0 0 0 S
(4, 1), (4, 3), (4, 4), (4, 6), (5, 1), (5, 3), (5, 5), (5, 6), (6, 2), (5,6) 1 1 0 Sc
(6, 4), (6, 5), (6, 6)}. (6,2) 1 0 1 S
At this stage, the PCT can be build as shown in Table 8. (6,4) 2 0 1 S
The I-lower approximations, the I-upper approxima- (6,5) 1 1 0 S
tions and the I-boundaries of S and Sc obtained by means (6,6) 0 0 0 S
of multi-graded dominance relations are as follows:

 I(S) D f(1, 2), (1, 4), (1, 5), (3, 4), (6, 2), (6, 4), (6, 5)}, not outrank y), ((2, 1), (2, 6), (4, 1), (4, 3), (4, 6), (5, 1),
 I(S) D f(1, 1), (1, 2), (1, 4), (1, 5), (2, 2), (2, 3), (3, 2), (5, 6)),
(3, 3), (3, 4), (3, 5), (4, 4), (5, 3), (5, 5), (6, 2), (6, 4), (6, 5),  if xP10 y, and xP10 y, (i. e. if xP10 y), then xSy or xS c y
(6, 6)}, (or, in words, if x and y are indifferent with respect
 I(S c ) D f(2, 1), (2, 6), (4, 1), (4, 3), (4, 6), (5, 1), (5, 6)}, to f 1 , then x outranks y or x does not outrank y),
 I(S c ) D f(1, 1), (2, 1), (2, 2), (2, 3), (2, 6), (3, 2), (3, 3), ((1, 1), (2, 2), (2, 3), (3, 2), (3, 3), (3, 5), (4, 4), (5, 3),
(3, 5), (4, 1), (4, 3), (4, 4), (4, 6), (5, 1), (5, 3), (5, 5), (5, 6), (5, 5), (6, 6)).
(6, 6)},
 Bn I (S) D Bn I (S c ) D f(1, 1), (2, 2), (2, 3), (3, 2), (3, 3), Let us assume now that the DM accepts to express
(3, 5), (4, 4), (5, 3), (5, 5), (6, 6)}. preferences with respect to criteria f 1 , f2 ; f3 on an ordi-
nal scale of preference, for which there is only informa-
Therefore, the quality of approximation is equal to tion about a partial preorder on all possible pairs of eval-
0.58. Moreover, there is only one reduct which is also the uations. In this case, S and Sc can be approximated in the
core, i. e. RED S (I) D CORE S (I) D f1g. way described in Subsect. “Dominance Without Degrees
Finally, the following decision rules can be induced of Preference”, i. e. without considering degrees of prefer-
(within parentheses there are the pairs of objects support- ence. The I-lower approximations, the I-upper approxi-
ing the rule): mations and the I-boundaries of S and Sc are as follows:

 if xP11 y, then xSy (or, in words, if x is at least weakly  I(S) D f(1, 1), (1, 2), (1, 4), (1, 5), (3, 4), (4, 4), (6, 2),
preferred to y with respect to f 1 , then x outranks y), (6, 4), (6, 5), (6, 6)},
((1, 2), (1, 4), (1, 5), (3, 4), (6, 2), (6, 4), (6, 5)),  I(S) D f(1, 1), (1, 2), (1, 4), (1, 5), (2, 2), (2, 3), (3, 2),
 if xP11 y, then xS c y (or, in words, if y is at least (3, 3), (3, 4), (3, 5), (4, 4), (5, 3), (5, 5), (6, 2), (6, 4), (6, 5),
weakly preferred to x with respect to f 1 , then x does (6, 6)},
 Rough Sets in Decision Making R 7773

 I(S c ) D f(2, 1), (2, 6), (4, 1), (4, 3), (4, 6), (5, 1), (5, 6)}, r; q 2 f1; : : : ; pg with r ¤ q; the classes of Cl are pref-
 I(S c ) D f(2, 1), (2, 2), (2, 3), (2, 6), (3, 2), (3, 3), (3, 5), erence-ordered according to the increasing order of
(4, 1), (4, 3), (4, 6), (5, 1), (5, 3), (5, 5), (5, 6)}, their indices,
 Bn I (S) D Bn I (S c ) D f(2, 2), (2, 3), (3, 2), (3, 3), (3, 5),  a function e : A ! Cl assigning each act a i 2 A to
(5, 3), (5, 5)}. a class Cl j 2 Cl.

Let us observe that the pairs (1, 1), (4, 4) and (6, 6) In this context, two different types of dominance can
belong now to the I-lower approximation of S and are be considered:
not contained in the I-boundaries. Therefore, the qual- 1) Classical dominance: given a p ; a q 2 A; a p dominates
ity of approximation is equal to 0.71. Moreover, there aq iff, for each possible state of the world, act ap gives
is still only one reduct which is also the core, i. e. again an outcome at least as good as act aq ; more formally,
RED S (I) D CORE S (I) D f1g. g(a p ; s j )  g(a q ; s j ), for each s j 2 S,
The following decision rules are induced from the 2) Stochastic dominance: given a p ; a q 2 A; a p dominates
above approximations and boundaries (within parenthe- aq iff, for each outcome x 2 X, act ap gives an outcome
ses there are the pairs of objects supporting the rule): at least as good as x with a probability at least as great as
the probability that act aq gives the same outcome, i. e.
 if f1 (x) is at least high and f1 (y) is at most high,
for all x 2 X,
then xSy; ((1, 1), (1, 2), (1, 4), (1, 5), (6, 2), (6, 4), (6, 5),
(6, 6)), P[S(a p ; x)]  P[S(a q ; x)]
 if f1 (x) is at least low and f1 (y) is at most low, then
xSy; ((1, 4), (3, 4), (4, 4), (6, 4)), where, for each (a i ; x) 2 A  X; S(a i ; x) D fs j 2 S :
 if f1 (x) is at most medium and f1 (y) is at least high, g(a i ; s j )  xg.
then xS c y; ((2, 1), (2, 6), (4, 1), (4, 6)(5, 1), (5, 6)),
 if f1 (x) is at most low and f1 (y) is at least medium, then In [19], it has been shown how to apply stochastic
xS c y; ((4, 1), (4, 3), (4, 6)) dominance in this context. On the basis of an a priori
 if f1 (x) is at least medium and f1 (y) is at most medium probability distribution P, one can assign to each sub-
and f1 (x) is at most medium and f1 (y) is at least set of states of the world W  S (W ¤ ;) the proba-
medium, (i. e. if f1 (x) is equal to medium and f1 (y) bility P(W) that one of the states in W is verified, i. e.
P
is equal to medium), then xSy or xS c y; ((2, 2), (2, 3), P(W) D i:s i 2W p i , and then to build up the set ˘ of all
(3, 2), (3, 3), (3, 5), (5, 3), (5, 5)). possible values P(W), i. e.

˘ D f 2 [0; 1] :  D P(W); W  Sg :

DRSA for Decision Under Uncertainty Let us define the following functions z : A  S ! ˘
Basic Concepts and z0 : A  S ! ˘ assigning to each act-state pair
(a i ; s j ) 2 A  S a probability  2 ˘ , as follows:
To apply the rough set approach to decision under uncer-
tainty, the following basic elements must be considered: X
z(a i ; s j ) D pr ;
 a set S D fs1 ; s2 ; : : : ; s u g of states of the world, or sim- r:g(a i ;s r )g(a i ;s j )

ply states, which are supposed to be mutually exclusive

and
and collectively exhaustive,
X
 an a priori probability distribution P over the states z0 (a i ; s j ) D pr :
of the world: more precisely, the probabilities of states r:g(a i ;s r )g(a i ;s j )
s1 ; s2 ; : : : ; s u are p1 ; p2 ; : : : ; p u , respectively (p1 C p2 C
C p u D 1; p i  0; i D 1; : : : ; u), Therefore, z(a i ; s j ) represents the probability of ob-
 a set A D fa1 ; a2 ; : : : ; a m g of acts, taining an outcome whose value is at least g(a i ; s j ) by
 a set X D fx1 ; x2 ; : : : ; x r g of outcomes or consequences act ai . Analogously, z0 (a i ; s j ) represents the probability
expressed in monetary terms (X  <), of obtaining an outcome whose value is at most g(a i ; s j )
 a function g : A ! X assigning to each pair act-state by act ai . On the basis of function z(a i ; s j ), function
(a i ; s j ) 2 A  S  S an outcome x 2 X, : A  ˘ ! X can be defined as follows:
 a set of classes Cl D fCl1 ; Cl2 ; : : : ; Cl p g, such that
Cl1 [ Cl2 [ : : : [ Cl p D A; Clr \ Cl q D ; for each (a i ; ) D max j : z(a g(a i ;s j ) :
i ;s j )
7774 R Rough Sets in Decision Making

Thus, (a i ; ) D x means that the outcome got by act the outcome is greater than or equal to 3 with a probability
ai is greater than or equal to x with a probability at least p at least 5/6. In fact, this is false because this probability is
(i. e. a probability p or greater). On the basis of function 4/6 (related to results 3, 4, 5 and 6). Therefore, in the con-
z0 (a i ; s j ), the function 0 : A  ˘ ! X can be defined as text of stochastic acts, an outcome expressed in positive
follows: terms refers to (a; ) giving a lower bound of an outcome
0 (“for act a there is a probability  to gain at least (a; )”),
(a i ; ) D min j : z(a g(a i ;s j ) :
i ;s j ) while an outcome expressed in negative terms refers to
0 (a; ) giving an upper bound of an outcome (“for act a
0 (a
i ; ) D x means that the outcome got by act ai is
there is a probability  to gain at most 0 (a; )”).
smaller than or equal to x with a probability at least p.
Given a p ; a q 2 A; a p stochastically dominates aq if and
Let us observe that information given by (a i ; ) and
0 (a ; ) is related. In fact, if the elements of ˘; 0 D only if (a p ; )  (a q ; ) for each  2 ˘ . This is equiv-
i
alent to the statement: given a p ; a q 2 A; a p stochastically
(0) ; (1) ; (2) ; : : : ; (d) D 1 (d D j˘ j), are reordered in
dominates aq if and only if 0 (a p ; )  0 (a q ; ) for each
such a way that 0 D (0)  (1)  (2)  : : :  (d) D 1,
 2 ˘.
then
For example, consider the game a with rolling a dice,
0 in which if the result is 1, then the gain is $7, if the result is
a i ; ( j) D a i ; 1  ( j1) :
2 then the gain is $6, and so on until the case in which
Therefore, (a i ; ( j) )  x is equivalent to 0 (a i ; 1  the result is 6 and the gain is $2. In this case game a
( j1) )  x; a i 2 A; ( j1) ; ( j) 2 ˘; x 2 X. This implies stochastically dominates game a because (a ; 1/6) D $7
that the analysis of the possible decisions can be equiva- is not smaller than (a; 1/6) D $6; (a ; 2/6) D $6
lently conducted on values of either (a i ; ) or 0 (a i ; ). is not smaller than (a; 2/6) D $5, and so on. Equiv-
However, from the point of view of representation of re- alently, game a stochastically dominates game a be-
sults, it is interesting to consider both values (a i ; ) cause 0 (a ; 1/6) D $2 is not smaller than 0 (a; 1/6) D
and 0 (a i ; ). The reason is that, contrary to intuition, $1; 0 (a ; 2/6) D $3 is not smaller than 0 (a; 2/6) D $2,
(a i ; )  x is not equivalent to the statement that by act and so on.
ai the outcome is smaller than or equal to x with a proba- DRSA can be applied in the context of decision un-
bility at least . The following example clarifies this point. der uncertainty considering as set of objects U the set
Let us consider a game a with rolling a dice, in which if the of acts A, as set of criteria (condition attributes) I the
result is 1, then the gain is 1, if the result is 2 then the gain set ˘ , as decision attribute fdg the classification Cl, as
is 2, and so on. Suppose, moreover, that the dice is equi- value set of all criteria the set X, as information func-
librated and thus each result is equiprobable with proba- tion f a function f such that f (a i ; ) D (a i ; ) and
bility 1/6. The values of (a i ; ) for all possible values of f (a i ; cl) D e(a i ). Let us observe that due to equivalence
probability are: (a i ; ( j) ) D 0 (a i ; 1  ( j1) ), one can also consider in-
(a; 1/6) D $6; (a; 2/6) D $5; (a; 3/6) D $4 ; formation function f 0 (a i ; ) D 0 (a i ; ).
The aim of the rough set approach to preferences un-
(a; 4/6) D $3; (a; 5/6) D $2; (a; 6/6) D $1 : der uncertainty is to explain the preferences of the DM
represented by the assignments of the acts from A to the
Let us remark that (a; 5/6)  $3 (indeed, (a; 5/6) D classes from Cl in terms of stochastic dominance, ex-
$2; and thus (a; 5/6)  $3 is true), however, this is not pressed by means of function . The syntax of decision
equivalent to the statement that by act a the outcome is rules obtained from this rough set approach is as follows:
smaller than or equal to $3 with a probability at least 5/6.
In fact, this is false because this probability is 3/6 (related 1) D -decision rules with the following syntax: “if
to results 1, 2 and 3). Analogously, the values of 0 (a; ) (a; p h1 )  x h1 and . . . , and (a; p h z )  x h z , then
for all possible values of probability are: a 2 Clr ” (i. e. “if by act a the outcome is at least x h1
0 0 0 with probability at least p h1 , and . . . , and the out-
(a; 1/6) D $1; (a; 2/6) D $2; (a; 3/6) D $3 ;
come is at least x h z with probability at least p h z , then
0 0 0 a 2 Clr ”) where p h1 ; : : : ; p h z 2 ˘; x h1 ; : : : ; x h z 2 X
(a; 4/6) D $4; (a; 5/6) D $5; (a; 6/6) D $6 :
and r 2 f2; : : : ; pg;
Let us remark that 0 (a; 5/6)  $5 (indeed, 2) D -decision rules with the following syntax: “if
0 (a; 5/6) 0
D $5; and thus (a; 5/6)  $3 is true), how- 0 (a; p )  x 0 (a; p )  x , then
h1 h 1 and . . . , and hz hz
ever, this is not equivalent to the statement that by act a 
a 2 Clr ” (i. e. “if by act a the outcome is at most
 Rough Sets in Decision Making R 7775

x h1 with probability at least p h1 , and . . . , and the out-  r2  “if by act a the outcome is at least 100 with prob-
come is at most x h z with probability at least p h z , then ability at least 0.50, then a is at least good”.
a 2 Clr ”) where p h1 ; : : : ; p h z 2 ˘; x h1 ; : : : ; x h z 2 X
and r 2 f1; : : : ; p  1g; r1 and r2 can be induced from the analysis of the
3) D -decision rules with the following syntax: “if same information table, because they involve different cri-
(a; p h1 )  x h1 and . . . , and (a; p h z )  x h w and teria (condition attributes). In fact, r1 involves attribute
0 (a; p
h wC1 )  x h wC1 and . . . , and
0 (a; p )  x ,
hz hz
(a; 0:25) (it can be expressed as “if (a; 0:25)  100,
then a 2 Cls [ ClsC1 [ : : : Cl t ” (i. e. “if by act a the then a is at least good”), r2 involves attribute (a; 0:50)
outcome is at least x h1 with probability at least (it can be expressed as “if (a; 0:50)  100, then a is at
p h1 , and . . . , and the outcome is at least x h w least good”). Considering the structure of the stochastic
with probability at least p h w and the outcome is dominance, the condition part of rule r1 is the weakest. In
at most x h wC1 with probability at least p h wC1 , and fact, rule r1 requires a cumulated outcome to be at least
. . . , and the outcome is at most x h z with proba- 100 with probability of 0.25, while rule r2 requires the
bility at least p h z , then a 2 Cls [ ClsC1 [ : : : Cl t ”) same outcome but with a greater probability, 0.5 against
where p h1 ; : : : ; p h w ; p h wC1 ; p h z 2 ˘; x h1 ; : : : ; x h z 2 X 0.25. Since the decision part of these two rules is the same,
and s; t 2 f1; : : : ; pg, such that s < t. r1 is minimal among these two rules. From a practical
point of view, this observation says that, if one induces
decision rules using the algorithms designed for DRSA,
According to the meaning of (a i ; p) and 0 (a i ; p) dis-
it is necessary to further filter the obtained results in or-
cussed above, D -decision rules are expressed in terms
der to remove rules which are not minimal in the specific
of (a i ; p); D -decision rules are expressed in terms of
0 (a ; p), and D
context of the DRSA analysis based on stochastic domi-
i  -decision rules are expressed in terms
nance.
of both (a i ; p) and 0 (a i ; p). Let us observe that due
to equivalence (a i ; ( j) ) D 0 (a i ; 1  ( j1) ), all above
decision rules can be expressed equivalently in terms of Example Illustrating DRSA in the Context
values of (a i ; p) or 0 (a i ; p). For example, a D -de- of Decision Under Uncertainty
cision rule r ( ) D“if (a; p h1 )  x h1 and . . . , and The following example illustrates the approach. Let us
(a; p h z )  x h z , then a 2 Clr ” can be expressed in terms consider
of 0 (a i ; p) as r ( 0 ) D“if 0 (a; ph1 )  x h1 and . . . , and
0 (a; p )  x , then a 2 Cl  ”, where, if p  a set S D fs1 ; s2 ; s3 g of states of the world,
hz hz r h r D ( j r ),
then ph r D 1  ( jr  1), with r D 1; : : : ; z, and 0 D  an a priori probability distribution P over the states of
(0); (1); (2); : : : ; (jPj) D 1 reordered in such a way the world defined as follows: p1 D 0:25; p2 D 0:35;
that 0 D (0)  (1)  (2)   (jPj) D 1. Anal- p3 D 0:40,
ogously, a D -decision rule r ( 0 ) D“if 0 (a; p h1 )  x h1  a set A D fa1 ; a2 ; a3 ; a4 ; a5 ; a6 g of acts,
and . . . , and 0 (a; p h z )  x h z , then a 2 Clr ” can be ex-  a set X D f0; 10; 15; 20; 30g of consequences
pressed in terms of (a i ; p) as r ( ) D“if (a; ph1 )  x h1  a set of classes Cl D fCl1 ; Cl2 ; Cl3 g, where Cl1 is the set
and . . . , and (a; ph z )  x h z , then a 2 Clr , where, if of bad acts, Cl2 is the set of medium acts, Cl3 is the set
p h r D ( jr ), then ph r D 1  ( jr  1), with r D 1; : : : ; z, of good acts,
and 0 D (0); (1); (2); : : : ; (jPj) D 1 reordered in  a function g : A ! X assigning to each act-state pair
such a way that 0 D (0)  (1)  (2)   (a i ; s j ) 2 A  S  S a consequence x h 2 X and a func-
(jPj) D 1. tion e : A ! Cl assigning each act a i 2 A to a class
Let us observe, however, that r ( ) is an expression Cl j 2 Cl presented in the following Table 9.
much more natural and meaningful than r ( 0 ), as well as
r ( 0 ) is an expression much more natural and meaning-
Rough Sets in Decision Making, Table 9
ful than r ( ). Another useful remark concerns minimal- Acts, consequences and assignment to classes from Cl
ity of rules, related to the specific intrinsic structure of the
stochastic dominance. Let us consider the following two pj a1 a2 a3 a4 a5 a6
decision rules: s1 0.25 30 0 15 0 20 10
s2 0.35 10 20 0 15 10 20
 r1  “if by act a the outcome is at least 100 with prob- s3 0.40 10 20 20 20 20 20
ability at least 0.25, then a is at least good”, Cl good medium medium bad medium good
7776 R Rough Sets in Decision Making

Rough Sets in Decision Making, Table 10

Acts, values of function (ai ; p) and assignment to classes from act a4 (bad). Therefore, act a3 cannot be assigned with-
Cl out doubts to the set of the class of the at least medium
acts as well as act a4 cannot be assigned without doubts
a1 a2 a3 a4 a5 a6 to the set of the classes of the (at most) bad acts. In con-
0.25 30 20 20 20 20 20 sequence, lower approximation and upper approximation
0.35 10 20 20 20 20 20 of Cl2 ; Cl3 and Cl1 ; Cl2 are equal, respectively, to
0.40 10 20 20 20 20 20
0.60 10 20 15 15 20 20  I Cl2 D fa1 ; a2 ; a5 ; a6 g D Cl2  fa3 g,
0.65 10 20 15 15 20 20  I Cl2 D fa1 ; a2 ; a3 ; a4 ; a5 ; a6 g D Cl2 [ fa4 g,
0.75 10 20 0 15 10 20  I Cl3 D fa1 ; a6 g D Cl3 ,
1 10 0 0 0 10 10  I Cl3 D fa1 ; a6 g D Cl3 ,
Cl good medium medium bad medium good  I Cl1 D ; D Cl1  fa4 g,
 I Cl1 D fa3 ; a4 g D Cl1 [ fa3 g,
 I Cl2 D fa2 ; a3 ; a4 ; a5 g D Cl2 ,
Table 10 shows the values of function (a i ; p).  I Cl2 D fa2 ; a3 ; a4 ; a5 g D Cl2 .
Table 10 is the decision table on which the DRSA is
applied. Let us give some examples of the interpretation of Since there are two inconsistent acts on a total of six
the values in Table 10. The column of act a3 can be read as acts (a3 ; a4 ), then the quality of approximation of the or-
follow: dinal classification is equal to 4/6. The second discovery
was one reduct of criteria (condition attributes) ensur-
 the value 20 in the row corresponding to 0.25 means ing the same quality of approximation as the whole set ˘
that the outcome is at least 20 with a probability of at
of probabilities: REDCl D f0:25; 0:75; 1g. This means that
least 0.25, the preferences of the DM can be explained using only the
 the value 15 in the row corresponding to 0.65 means
probabilities from REDCl . REDCl is also the core because
that the outcome is at least 15 with a probability of at
no probability value in REDCl can be removed without de-
least 0.65, teriorating the quality of approximation. The third discov-
 the value 0 in the row corresponding to 0.75 means that
ery was a set of minimal decision rules describing the DM’s
the outcome is at least 0 with a probability of at least preferences (within parentheses there is a verbal interpre-
0.75. tation of the corresponding decision rule, and the support-
Analogously, the row corresponding to 0.65, can be ing acts):
read as follows:
1) if (a i ; 0:25)  30, then a i 2 Cl3 (if the probability of
 the value 10 relative to a1 , means that by act a1 the out- gaining at least 30 is at least 0.25, then act ai is at least
come is at least 10 with a probability of at least 0.65, good) (a1 ),
 the value 20 relative to a2 , means that by act a2 the out- 2) if (a i ; 0:75)  20 and (a i ; 1)  10, then a i 2 Cl3 (if
come is at least 20 with a probability of at least 0.65, the probability of gaining at least 20 is at least 0.75 and
 and so on. the probability of gaining at least 10 is (at least) 1 (i. e.
for sure the gaining is at least 10), then act ai is at least
Applying DRSA, the following upward union and good) (a6 ),
downward union of classes are approximated: 3) if (a i ; 1)  10, then a i 2 Cl2 (if the probability of
 Cl2 D Cl2 [ Cl3 , i. e. the set of the acts at least gaining at least 10 is (at least) 1 (i. e. for sure the gaining
medium, is at least 10), then act ai is at least medium) (a1 ; a5 ; a6 ),
 Cl3 D Cl3 , i. e. the set of the acts (at least) good, 4) if (a i ; 0:75)  20, then a i 2 Cl2 (if the probability of
 Cl1 D Cl1 , i. e. the set of the acts (at most) bad, gaining at least 20 is at least 0.75, then act ai is at least
 Cl2 D Cl1 [ Cl2 , i. e. the set of the acts at most medium) (a2 ; a6 ),
medium. 5) if (a i ; 0:25)  20 (i. e. 0 (a i ; 1)  20) and (a i ;
0:75)  15 (i. e. 0 (a i ; 0:35)  15), then a i 2 Cl2 (if
The first result of the DRSA approach was a discovery the probability of gaining at most 20 is (at least) 1 (i. e.
that the decision table (Table 10) is not consistent. Indeed, for sure you gain at most 20) and the probability to gain
Table 10 shows that act a4 stochastically dominates act a3 , at most 15 is at least 0.35, then act ai is at most medium)
however act a3 is assigned to a better class (medium) than (a3 ; a4 ; a5 ),
 Rough Sets in Decision Making R 7777

6) if (a i ; 1)  0 (i. e. 0 (a i ; 0:25)  0), then a i 2 Cl2 (if An association rule r  “if f i 1 (x)  r i 1 and . . .
the probability of gaining at most 0 is at least 0.25, then and f i p (x)  r i p , then f i pC1 (x)  r i pC1 and . . . and
act ai is at most medium) (a2 ; a3 ; a4 ), f i q (x)  r i q ” holds in universe U if:
7) if (a i ; 1)  0 and (a i ; 1)  0 (i. e. (a i ; 1) D 0) and
(a i ; 0:75)  15 ( 0 (a i ; 0:35)  15), then a i 2 Cl1 [ 1) there is at least one y 2 U supporting r,
Cl2 (if the probability of gaining at least 0 is 1, then act 2) r is not contradicted in U, i. e. there is no z 2 U such
ai is at most medium) (a2 ; a3 ; a4 ). that f i 1 (z)  r i 1 and . . . and f i p (z)  r i p , while not
f i pC1 (z)  r i pC1 or . . . or f i q (z)  r i q .
Minimal sets of minimal decision rules represent the
most concise and non-redundant knowledge contained in Given the two association rules:
Table 9 (and, consequently, in Table 10). The above min-
 r1 “if f i 1 (x)  r1i 1 and . . . and f i p (x)  r1i p , then
imal set of 7 decision rules uses 3 attributes (probability
0:25; 0:75 and 1) and 11 elementary conditions, i. e. 26% f i pC1 (x)  r1i pC1 and . . . and f i q (x)  r1i q ”,
of descriptors from the original data table (Table 10). Of  r2 “if f j 1 (x)  r2j 1 and . . . and f j s (x)  r2j s , then
course, this is only a didactic example. Representation in f j sC1 (x)  r2j sC1 and . . . and f j t (x)  r2j t ”,
terms of decision rules of larger sets of exemplary acts
from real applications are more synthetic in the sense of rule r1 is not weaker than rule r2 , denoted by r1 Fr2 , if:
the percentage of used descriptors from the original deci-
˛) fi1 ; : : : ; i p g  f j1 ; : : : ; js g,
sion table.
ˇ) r1i 1  r2i 1 ; : : : ; r1i p  r2i p ,
) fi pC1 ; : : : ; i q g  f jsC1 ; : : : ; j t g,
Multiple Criteria Decision Analysis ı) r1j sC1  r2j sC1 ; : : : ; r1j t  r2j t .
Using Association Rules
In multiple criteria decision analysis, the DM is often in- Conditions ˇ and ı are formulated for criteria f i to be
terested in relationships between attainable values of cri- minimized. If criterion f i should be maximized, the cor-
teria. This information is particularly useful in multiob- responding inequalities should be reversed, i. e. r1i  r2i in
jective optimization (see [32] and Sect. “Interactive Multi- condition ˇ as well as in condition ı. Notice that F is a bi-
objective Optimization Using DRSA (IMO-DRSA)”). For nary relation on the set of association rules, which is a par-
instance, in a car selection problem, one can observe that tial preorder, i. e. it is reflexive (each rule is not weaker
in the set of considered cars, if the maximum speed is at than itself) and transitive. The asymmetric part of the re-
least 200 km/h and the time to reach 100 km/h is at most lation F is denoted by F, and r1 F r2 reads “r1 is stronger
7 s, then the price is not less than 40,000 $ and the fuel than r2 ”.
consumption is not less than 9 liters per 100 km. These re- For example, consider the following association rules:
lationships are association rules whose general syntax, in
 r1  “if the maximum speed is at least 200 km/h and
case of minimization of criteria f i ; i 2 I, is:
the time to reach 100 km/h is at most 7 s, then the price
“if f i 1 (x)  r i 1 and . . . and f i p (x)  r i p , then is not less than 40,000 $ and the fuel consumption is not
f i pC1 (x)  r i pC1 and . . . and f i q (x)  r i q ”, where less than 9 liters per 100 km”,
fi1 ; : : : ; i q g  I; r i 1 , . . . , r i q 2 <.  r2  “if the maximum speed is at least 200 km/h and
the time to reach 100 km/h is at most 7 s and the horse
If criterion f i ; i 2 I, should be maximized, the corre- power is at least 175 kW, then the price is not less than
sponding condition in the association rule should be 40,000 $ and the fuel consumption is not less than 9
reversed, i. e. in the premise, the condition becomes liters per 100 km”,
f i (x)  r i , and in the conclusion it becomes f i (x)  r i .  r3  “if the maximum speed is at least 220 km/h and
Given an association rule r  “if f i 1 (x)  r i 1 and the time to reach 100 km/h is at most 7 s, then the price
. . . and f i p (x)  r i p , then f i pC1 (x)  r i pC1 and . . . and is not less than 40,000 $ and the fuel consumption is not
f i q (x)  r i q ”, an object y 2 U supports r if f i 1 (y)  r i 1 less than 9 liters per 100 km”,
and . . . and f i p (y)  r i p and f i pC1 (y)  r i pC1 and . . . and  r4  “if the maximum speed is at least 200 km/h and
f i q (y)  r i q . Moreover, object y 2 U supporting decision the time to reach 100 km/h is at most 7 s, then the price
rule r is a base of r if f i 1 (y) D r i 1 and . . . and f i p (y) D r i p is not less than 40,000 $”,
and f i pC1 (y) D r i pC1 and . . . and f i q (y) D r i q . An associ-  r5  “if the maximum speed is at least 200 km/h and
ation rule having at least one base is called robust. the time to reach 100 km/h is at most 7 s, then the price
7778 R Rough Sets in Decision Making

is not less than 35,000 $ and the fuel consumption is not Step 3. If the DM is satisfied with one solution from the
less than 9 liters per 100 km”, sample, then this is the most preferred solution and the
 r6  “if the maximum speed is at least 220 km/h and procedure stops. Otherwise continue.
the time to reach 100 km/h is at most 7 s and the horse Step 4. Ask the DM to indicate a subset of relatively
power is at least 175 kW, then the price is not less than “good” solutions in the sample.
35,000 $”. Step 5. Apply DRSA to the current sample of solutions
classified into “good” and “others” solutions, in order
Let us observe that rule r1 is stronger than each of the to induce a set of decision rules with the following syn-
other five rules for the following reasons: tax “if f j 1 (x)  ˛ j˚1 and . . . and
 f j p (x)  ˛ j p , then so-
lution x is good”, j1 ; : : : ; j p  f1; : : : ; ng.
 r1 F r2 for condition ˛) because, all things equal else- Step 6. Present the obtained set of rules to the DM.
where, in the premise of r2 there is an additional con- Step 7. Ask the DM to select the decision rules most ade-
dition: “the horse power is at least 175 kW”, quate to his/her preferences.
 r1 F r3 for condition ˇ) because, all things equal else- Step 8. Adjoin the constraints f j 1 (x)  ˛ j 1 ; : : : ; f j p (x) 
where, in the premise of r3 there is a condition with ˛ j p coming from the rules selected in Step 7 to the set
a worse threshold value: “the maximum speed is at least of constraints imposed on the Pareto optimal set, in
220 km/h” instead of “the maximum speed is at least order to focus on a part interesting from the point of
200 km/h”, view of DM’s preferences.
 r1 F r4 for condition ) because, all thing equal else- Step 9. Go back to Step 1.
where, in the conclusion of r4 one condition is miss-
ing: “the fuel consumption is not less than 9 liters per In a sequence of iterations, the method is exploring the
100 km”, Pareto optimal set of a multiobjective optimization prob-
 r1 F r5 for condition ı) because, all thing equal else- lem or an approximation of this set. In the calculation
where, in the conclusion of r5 there is a condition with stage (Step 1), any multiobjective optimization method,
a worse threshold value: “the price is not less than which finds the Pareto optimal set or its approximation,
35,000 $” instead of “the price is not less than 40,000 $”, such as Evolutionary Multiobjective Optimization meth-
 r1 F r6 for conditions ˛), ˇ), ) and ı) because all weak ods, can be used. In the dialogue stage of the method (Step
points for which rules r2 ; r3 ; r4 and r5 are weaker than 2 to 7), the DM is asked to select a decision rule induced
rule r1 are present in r6 . from his/her preference information, which is equivalent
to fixing some upper bounds for the minimized objective
An association rule r is minimal if there is no other rule functions f j .
stronger than r with respect to F. An algorithm for induc- In Step 1, the representative sample of solutions from
tion of association rules from preference ordered data has the currently considered part of the Pareto optimal set
been presented in [20]. can be generated using one of existing procedures, such
as [33,57,58]. It is recommended to use a fine grained sam-
Interactive Multiobjective Optimization ple of representative solutions to induce association rules,
Using DRSA (IMO-DRSA) however, the sample of solutions presented to the DM in
Step 2 should be much smaller (about a dozen) in order to
This section presents a recently proposed method for In- avoid an excessive cognitive effort of the DM. Otherwise,
teractive Multiobjective Optimization using Dominance- the DM would risk to give non reliable information.
based Rough Set Approach (IMO-DRSA) [32]. Assuming The association rules presented in Step 2 help the DM
that objective functions f i; jD1;:::;n , are to minimized, the in understanding what (s)he can expect from the optimiza-
method is composed of the following steps. tion problem. More precisely, any association rule
“if f i 1 (x)  r i 1 and . . . and f i p (x)  r i p , then
Step 1. Generate a representative sample of solutions
f i pC1 (x)  r i pC1 and . . . and f i q (x)  r i q ”, where
from the currently considered part of the Pareto op-
fi1 ; : : : ; i q g  I; r i 1 ; : : : ; r i q 2 <
timal set.
Step 2. Present the sample to the DM, possibly together says to the DM that, if (s)he wants attain the values of
with association rules showing relationships between objective functions f i 1 (x)  r i 1 and . . . and f i p (x)  r i p ,
attainable values of objective functions in the Pareto then (s)he cannot reasonably expect to obtain values of ob-
optimal set. jective functions f i pC1 (x) < r i pC1 and . . . and f i q (x) < r i q .
 Rough Sets in Decision Making R 7779

With respect to the ordinal classification of solutions about his/her preferences and about the “shape” of the
into the two classes of “good” and “others”, observe that Pareto optimal set.
“good” means in fact “relatively good”, i. e. better than the
rest. In case, the DM would refuse to classify as “good”
any solution, one can ask the DM to specify some min- Example Illustrating IMO-DRSA in the Context
imal requirements of the type f j 1 (x)  ˛ j 1 and . . . and of Multiobjective Optimization
f j p (x)  ˛ j p for “good” solutions. These minimal require- To illustrate the interactive multiobjective optimization
ments give some constraints that can be used in Step 8, in procedure based on DRSA, a product mix problem is con-
the same way as the analogous constraints coming from sidered. There are three products: A, B, C which are pro-
selected decisions rules. duced in quantities denoted by x A ; x B ; and x C , respec-
The rules considered in Step 5 have a syntax corre- tively. The unit prices of the three products are p A D 20;
sponding to minimization of objective functions. In case p B D 30;p C D 25. The production process involves two
of maximization of an objective function f j , the condition machines. The production times of A, B, C on the first
concerning this objective in the decision rule should have machine are equal to t1A D 5; t1B D 8; t1C D 10, and on
the form f j (x)  ˛ j . the second machine they are equal to t2A D 8;t2B D 6;
Remark, moreover, that the Pareto optimal set reduced t2C D 2. Two raw materials are used in the production
in Step 8 by constraints f j 1 (x)  ˛ j 1 ; : : : ; f j p (x)  ˛ j p is process. The first raw material has a unit cost of 6 and
certainly not empty if these constraints are coming from the quantity required for production of one unit of A, B
one decision rule only. Since robust rules (see the glos- and C is r1A D 1; r1B D 2 and r1C D 0:75, respectively.
sary Dominance-Based Rough Set Approach (DRSA)) are The second raw material has a unit cost of 8 and the quan-
considered, the threshold values ˛ j 1 ; : : : ; ˛ j p are values tity required for production of one unit of A, B and C is
of objective functions of some solutions from the Pareto r2A D 0:5; r2B D 1 and r2C D 0:5, respectively. Moreover,
optimal set. If f j1 ; : : : ; j p g D f1; : : : ; ng, i. e. f j1 ; : : : ; j p g the market cannot absorb a production greater than 10, 20
is the set of all objective functions, then the new re- and 10 units for A, B and C, respectively. To decide how
duced part of the Pareto optimal set contains only much of A, B and C should be produced, the following ob-
one solution x such that f1 (x) D ˛1 ; : : : ; f n (x) D ˛n . If jectives have to be taken into account:
f j1 ; : : : ; j p g f1; : : : ; ng, i. e. f j1 ; : : : ; j p g is a proper sub-
set of the set of all objective functions, then the new re-  Profit (to be maximized),
duced part of the Pareto optimal set contains solutions sat-  Time (total production time on two machines – to be
isfying conditions f j 1 (x)  ˛ j 1 and . . . and f j p (x)  ˛ j p . minimized),
Since the considered rules are robust, then there is at  Production of A (to be maximized),
least one solution x satisfying these constraints. When  Production of B (to be maximized),
the Pareto optimal set is reduced in Step 8 by constraints  Production of C (to be maximized),
f j 1 (x)  ˛ j 1 ; : : : ; f j p (x)  ˛ j p coming from more than  Sales (to be maximized).
one rule, then it is possible that the resulting reduced part
of the Pareto optimal set is empty. Thus, before passing to The above product mix problem can be formulated as
Step 9, it is necessary to verify if the reduced Pareto opti- the following multiobjective optimization problem:
mal set is not empty. If the reduced Pareto optimal set is
empty, then the DM is required to revise his/her selection Maximize
of rules. The DM can be supported in this task, by informa- 20x A C 30x B C 25x C  (1x A C 2x B C 0:75x C )6
tion about minimal sets of constraints f j (x)  ˛ j coming
from the considered decision rules to be removed in order  (0:5x A C 1x B C 0:5x C )8 [Profit] ;
to get a non-empty part of the Pareto optimal set. Minimize 5x A C8x B C10x C C8x A C6x B C2x C [Time];
The constraints introduced in Step 8 are maintained
in the following iterations of the procedure, however, they Maximize xA [Production of A] ;
cannot be considered as irreversible. Indeed, the DM can
come back to the Pareto optimal set considered in one of Maximize xB [Production of B] ;
previous iterations and continue from this point. This is
in the spirit of a learning oriented conception of interac- Maximize xC [Production of C] ;
tive multiobjective optimization, i. e. it agrees with the idea
that the interactive procedure permits the DM to learn Maximize 20x A C 30x B C 25x C [Sales] ;
7780 R Rough Sets in Decision Making

Rough Sets in Decision Making, Table 11

A sample of Pareto optimal solutions proposed in the first iteration

Solution Profit Time Prod. A Prod. B Prod. C Sales Evaluation

s1 165 120 0 0 10 250 
s2 172.69 130 0.769 0 10 265.38 
s3 180.38 140 1.538 0 10 280.77 good
s4 141.13 140 3 3 4.92 272.92 good
s5 148.38 150 5 2 4.75 278.75 good
s6 139.13 150 5 3 3.58 279.58 
s7 188.08 150 2.308 0 10 296.15 
s8 159 150 6 0 6 270 
s9 140.5 150 6 2 3.67 271.67 good
s10 209.25 200 6 2 7.83 375.83 
s11 189.38 200 5 5 5.42 385.42 
s12 127.38 130 3 3 4.08 252.08 
s13 113.63 120 3 3 3.25 231.25 

subject to: 5) if Profit  148:38 and Time  150, then x B  2 (s1;

s2; s3; s5; s7; s8),
x A  10 ; x B  20 ; x C  10 6) if x A  5, then Time  150 (s5; s6; s8; s9; s10; s11),
[Market absorption limits] ; 7) if Profit  127:38 and x A  3, then Time  130 (s4;
s5; s6; s8; s9; s10; s11; s12),
xA  0 ; xB  0 ; xC  0 8) if Time  150 and x B  2, then Profit  148:38 (s4;
s5; s6; s9; s12; s13),
[Non-negativity constraints] :
9) if x A  3 and x C  4:08, then Time  130 (s4; s5;
A sample of representative Pareto optimal solutions has s8; s10; ; s11; s12),
been calculated and proposed to the DM. Observe that 10) if Sales  265:38, then Time  130 (s2; s3; s4; s5; s6;
the considered problem is a Multiple Objective Linear s7; s8; s9; s10; s11).
Programming (MOLP) problem, and thus representative
Pareto optimal solutions can be calculated using classical
linear programming looking for the solutions optimizing Then, the DM has been asked if (s)he was satisfied
each one of the considered objectives or fixing all the con- with one of the proposed Pareto optimal solutions. Since
sidered objective functions but one at a satisfying value, his/her answer was negative, (s)he was requested to indi-
and looking for the solution optimizing the remaining ob- cate a subset of relatively “good” solutions which are indi-
jective function. The set of representative Pareto optimal cated in the “Evaluation” column of Table 11.
solutions is shown in Table 11. Moreover, a set of poten- Taking into account the ordinal classification of Pareto
tially interesting association rules have been induced from optimal solutions into “good” and “others”, made by the
the sample and presented to the DM. These rules repre- DM, twelve decision rules have been induced from the
sent strongly supported relationships between attainable lower approximation of “good” solutions. The frequency
values of objective functions. The association rules are the of the presence of objectives in the premises of the rules
following (between parentheses there are id’s of solutions gives a first idea of the importance of the considered ob-
supporting the rule): jectives. These frequencies are the following:

1) if Time  140, then Profit  180.38 and Sales

4
 280:77 (s1; s2; s3; s4; s12; s13),  Profit: 12 ,
2) if Time  150, then Profit  188:08 and Sales 12
 Time: 12 ,
 296:15 (s1; s2; s3; s4; s5; s6; s7; s8; s9; s12; s13),  Production of A : 7
12 ,
3) if x B  2, then Profit  209:25 and x A  6 and 4
 Production of B : 12 ,
x C  7:83 (s4; s5; s6; s9; s10; s11; s12; s13),
5
4) if Time  150, then x B  3 (s1; s2; s3; s4; s5; s6; s7;  Production of C : 12 ,
5
s8; s9; s12; s13),  Sales: 12 .
 Rough Sets in Decision Making R 7781

The following potentially interesting decision rules 120 ) if Sales  272, then Time  130 (s10 , s20 , s30 , s40 , s50 ,
were presented to the DM: s60 , s70 , s80 ).

1) if Profit  140.5 and Time  150 and x B  2, then The DM has been asked again if (s)he was satisfied
product mix is good (s4; s5; s9), with one of the proposed Pareto optimal solutions. Since
2) if Time  140 and x A  1:538 and x C  10, then prod- his/her answer was negative, (s)he was requested again to
uct mix is good (s3), indicate a subset of relatively “good” solutions, which are
3) if Time  150 and x B  2 and x C  4:75, then product indicated in the “Evaluation” column of Table 12.
mix is good (s4, s5), Taking into account the ordinal classification of Pareto
4) if Time  140 and Sales  272:9167, then product mix optimal solutions into “good” and “others”, made by the
is good (s3, s4), DM, eight decision rules have been induced from the
5) if Time  150 and x B  2 and x C  3:67 and Sales lower approximation of “good” solutions. The frequencies
 271:67, then product mix is good (s4, s5, s9). of the presence of objectives in the premises of the rules
are the following:
Among these decision rules, the DM has selected
rule 1) as the most adequate to his/her preferences. This  Profit: 28 ,
rule permits to define the following constraints reducing  Time: 18 ,
the feasible region of the production mix problem:  Production of A: 58 ,
 20x A C30x B C25x C (x A C2x B C0:75x C )6(0:5x A C  Production of B : 38 ,
x B C 0:5x C )8  140:5 [Profit  140:5],  Production of C : 38 ,
 5x A C 8x B C 10x C C 8x A C 6x B C 2x C  150 [Time
 Sales: 28 .
 150],
 x B  2 [Production of B  2]. The following potentially interesting decision rules
were presented to the DM:
These constraints have been considered together with
the original constraints for the production mix problem, 1) if Time  125 and x A  1, then product mix is good
and a new sample of representative Pareto optimal solu- (s110 ; s120 ),
tions shown in Table 12 have been calculated and pre- 2) if x A  1 and x C  7, then product mix is good
sented to the DM, together with the following potentially (s30 ; s60 ; s110 ),
interesting association rules: 3) if x A  1:5 and x C  6:46, then product mix is good
(s30 ; s40 ; s60 ; s120 ),
10 ) if Time  140, then Profit  174 and x C  9:33 and 4) if Profit  158.25 and x A  2, then product mix is good
Sales  293:33 (s50 , s60 , s70 , s80 , s90 , s100 , s110 , s120 ), (s30 ; s40 ; s60 ),
2 ) if x A  2, then x B  3 and Sales  300:83 (s20 , s30 ,
0
5) if x A  2 and Sales  300, then product mix is good
s40 , s60 , s70 , s90 ), (s30 ; s40 ).
3 ) if x A  2, then Profit  172 and x C  8 (s20 , s30 , s40 ,
0

s60 , s70 , s90 ), Among these decision rules, the DM has selected rule
4 ) if Time  140, then x A  2 and x B  3 (s50 , s60 , s70 ,
0 4) as the most adequate to his/her preferences. This rule
s80 , s90 , s100 , s110 , s120 ), permits to define the following constraints reducing the
5 ) if Profit  158.25, then x A  2 (s10 , s30 , s40 , s50 , s60 ,
0 Pareto optimal set of the production mix problem:
s80 ),
 20x A C30x B C25x C (x A C2x B C0:75x C )6(0:5x A C
6 ) if x A  2, then Time  130 (s20 , s30 , s40 , s60 , s70 , s90 ),
0
x B C 0:5x C )8  158:25 [Profit  158:25],
70 ) if x C  7:17, then x A  2 and x B  2 (s10 , s30 , s50 , s60 ,
 x A  2 [Production of A  2].
s80 , s100 ),
8 ) if x C  6, then x A  2 and x B  3 (s10 , s30 , s40 , s50 ,
0 Let us observe that the first constraint is just strength-
s60 , s70 , s80 , s90 , s100 , s110 , s120 ), ening an analogous constraint introduced in the first iter-
9 ) if x C  7, then Time  125 and x B  2 (s10 , s30 , s50 ,
0 ation (Profit  140:5).
s60 , s80 , s100 , s110 ), Considering the new set of constraints, a new sample
10 ) if Sales  280, then Time  140 and x B  3 (s10 , s20 ,
0 of representative Pareto optimal solutions shown in Ta-
s30 , s40 , s50 , s70 ), ble 13 has been calculated and presented to the DM, to-
11 ) if Sales  279:17, then Time  140 (s10 , s20 , s30 , s40 ,
0 gether with the following potentially interesting associa-
s50 , s60 , s70 ), tion rules:
7782 R Rough Sets in Decision Making

Rough Sets in Decision Making, Table 12

A sample of Pareto optimal solutions proposed in the second iteration

Solution Profit Time Prod. A Prod. B Prod. C Sales Evaluation

s10 186.53 150 0.154 2 10 313.08 
s20 154.88 150 3 3 5.75 293.75 
s30 172 150 2 2 8 300 good
s40 162.75 150 2 3 6.83 300.83 good
s50 174 140 0 2 9.33 293.33 
s60 158.25 140 2 2 7.17 279.17 good
s70 149 140 2 3 6 280 
s80 160.25 130 0 2 8.5 272 good
s90 144.5 130 2 2 6.33 258.33 
s100 153.38 125 0 2 8.08 262.08 
s110 145.5 125 1 2 7 255 good
s120 141.56 125 1.5 2 6.46 251.46 good

100 ) if Time  145, then x A  2 and x B  2:74 and Sales information in terms of marginal rates of substitution, the
 290:2 (s200 ; s300 ; s400 ), reference point method [58] requiring a reference point
00
2 ) if x C  6:92, then x A  3 and x B  2 and Sales and weights to formulate an achievement scalarizing func-
 292:92 (s300 ; s400 ; s500 ), tion, the Light Beam Search method [33] requiring infor-
00
3 ) if Time  145, then Profit  165:13 and x A  2 and mation in terms of weights and indifference, preference
x C  7:58 (s200 ; s300 ; s400 ), and veto thresholds, being typical parameters of ELEC-
4 ) if x C  6:72, then x B  2:74 (s200 ; s300 ; s400 ; s500 ),
00 TRE methods). Eliciting such information requires a sig-
500 ) if Sales  289:58, then Profit  165:13 and Time nificant cognitive effort on the part of the DM. It is gener-
 145 and x C  7:58 (s100 ; s200 ; s300 ; s500 ). ally acknowledged that people often prefer to make exem-
plary decisions and cannot always explain them in terms
The DM has been asked again if (s)he was satisfied with
of specific parameters. For this reason, the idea of inferring
one of the presented Pareto optimal solutions shown in
preference models from exemplary decisions provided by
Table 13 and this time (s)he declared that solution s300 is
the DM is very attractive. The output result of the anal-
satisfactory for him/her. This ends the interactive proce-
ysis is the model of preferences in terms of “if . . . , then
dure.
. . . ” decision rules which is used to reduce the Pareto op-
timal set iteratively, until the DM selects a satisfactory so-
Characteristics of the IMO-DRSA lution. The decision rule preference model is very conve-
The interactive procedure presented in Sect.“Interactive nient for decision support, because it gives argumentation
Multiobjective Optimization Using DRSA (IMO-DRSA)” for preferences in a logical form, which is intelligible for
can be analyzed from the point of view of input and out- the DM, and identifies the Pareto optimal solutions sup-
put information. As to the input, the DM gives prefer- porting each particular decision rule. This is very useful
ence information by answering easy questions related to for a critical revision of the original ordinal classification of
ordinal classification of some representative solutions into representative solutions into the two classes of “good” and
two classes (“good” and “others”). Very often, in multi- “others”. Indeed, decision rule preference model speaks
ple criteria decision analysis, in general, and in interac- the same language of the DM without any recourse to tech-
tive multiobjective optimization, in particular, the prefer- nical terms, like utility, tradeoffs, scalarizing functions and
ence information has to be given in terms of preference so on.
model parameters, such as importance weights, substitu- All this implies that IMO-DRSA has a transparent
tion rates and various thresholds (see [6] for the Multi- feedback organized in a learning oriented perspective,
ple Attribute Utility Theory and [1,5,34,46] for outrank- which permits to consider this procedure as a “glass box”,
ing methods; for some well-known interactive multiob- contrary to the “black box” characteristic of many pro-
jective optimization methods requiring preference model cedures giving final result without any clear explanation.
parameters, see the Geoffrion–Dyer–Feinberg method [9], Note that with the proposed procedure, the DM learns
the method of Zionts and Wallenius [60,61] and the Inter- about the shape of the Pareto optimal set using the asso-
active Surrogate Worth Tradeoff method [2,3] requiring ciation rules. They represent relationships between attain-
 Rough Sets in Decision Making R 7783

Rough Sets in Decision Making, Table 13

A sample of Pareto optimal solutions proposed in the third iteration

Solution Profit Time Prod. A Prod. B Prod. C Sales Evaluation

s100 158.25 150 2 3.49 6.27 301.24 
s200 158.25 145 2 2.74 6.72 290.20 
s300 165.13 145 2 2 7.58 289.58 selected
s400 158.25 140 2 2 7.17 279.17 
s500 164.13 150 3 2 6.92 292.93 
s600 158.25 145.3 3 2 6.56 284.02 

able values of objective functions on the Pareto optimal tion of scales, and this ensures the meaningfulness of re-
set in logical and very natural statements. The information sults from the point of view of measurement theory (see
given by association rules is as intelligible as the decision e. g. [42]).
rule preference model, since they speak the language of the With respect to computational aspects of the method,
DM and permit him/her to identify the Pareto optimal so- notice that the decision rules can be calculated effi-
lutions supporting each particular association rule. ciently in few seconds only using the algorithms presented
Thus, decision rules and association rules give an ex- in [17,20]. When the number of objective functions is not
planation and a justification of the final decision, that too large to be effectively controlled by the DM (let us say
does not result from a mechanical application of a cer- seven plus or minus two, as suggested by Miller [37]), then
tain technical method, but rather from a mature conclu- the decision rules can be calculated in a fraction of one
sion of a decision process based on active intervention of second. In any case, the computational effort grows expo-
the DM. nentially with the number of objective functions, but not
Observe, finally, that the decision rules representing with respect to the number of considered Pareto optimal
preferences and the association rules describing the Pareto solutions, which can increase with no particularly negative
optimal set are based on ordinal properties of objective consequence on calculation time.
functions only. Differently from methods involving some
scalarization (almost all existing interactive methods), in
any step the proposed procedure does not aggregate the Conclusions
objectives into a single value, avoiding operations (such Rough set theory is a mathematical tool for dealing with
as averaging, weighted sum, different types of distance, granularity of information and possible inconsistencies in
achievement scalarization) which are always arbitrary to the description of objects. Considering this description as
some extent. Observe that one could use a method based an input data about a decision problem, the rough set ap-
on a scalarization to generate the representative set of proach permits to structure this description into lower and
Pareto optimal solutions, nevertheless, the decision rule upper approximations, corresponding to certain and pos-
approach would continue to be based on ordinal proper- sible knowledge about the problem. Induction algorithms
ties of objective functions only, because the dialogue stage run on these approximations discover, in turn, certain and
of the method operates on ordinal comparisons only. In possible decision rules that facilitate an understanding of
the proposed method, the DM gets clear arguments for the DM’s preferences, and that enable a recommendation
his/her decision in terms of “if . . . , then . . . ” decision rules concordant with these preferences.
and the verification if a proposed solution satisfies these The original version of the rough set approach, based
decision rules is particularly easy. This is not the case on indiscernibility or similarity relation, and typical in-
of interactive multiobjective optimization methods based duction algorithms considered within machine learning,
on scalarization. For example, in the methods using an data mining and knowledge discovery, deal with data
achievement scalarization function, it is not evident what describing problems of taxonomy-type classification, i. e.
does it mean for a solution to be “close” to the reference problems where neither the attributes describing the ob-
point. How to justify the choice of the weights used in the jects, nor the classes to which the objects are assigned,
achievement function? What is their interpretation? Ob- are ordered. On the other hand, multiple criteria decision
serve, instead, that the method proposed in this chapter making deals with problems where descriptions (evalua-
operates on data using ordinal comparisons which would tions) of objects by means of attributes (criteria), as well as
not be affected by any increasing monotonic transforma- decisions in classification, choice and ranking problems,
7784 R Rough Sets in Decision Making

are ordered. Moreover, in data describing multiple criteria  the decision rule preference model resulting from
decision making, there exist a monotonic relationship be- DRSA is more general than all existing models of con-
tween conditions and decisions, like “the bigger the house, joint measurement, due to its capacity of handling in-
the more expensive it is”. The generalization of the rough consistent preferences (a new model of conjoint mea-
set approach and of the induction algorithms about prob- surement is formally equivalent to the decision rule
lems in which order properties and monotonic relation- preference model handling inconsistencies),
ships are important gave birth to the Dominance-based  the decision rule preference model fulfills the postulate
Rough Set Approach (DRSA) which made a breakthrough of transparency and interpretability of preference mod-
in scientific decision support. els in decision support; each decision rule can be clearly
The main features of DRSA are the following: identified with those parts of the preference informa-
tion (decision examples) which support the rule; the
rules inform the DM in a quasi-natural language about
 preference information necessary to deal with any mul- the relationships between conditions and decisions; in
tiple criteria decision problem, or with decision under this way, the rules permit traceability of the decision
uncertainty, is asked to the DM just in terms of exem- support process and give understandable justifications
plary decisions, for the decision to be made,
 the rough set analysis of preference information sup-  the proposed methodology is based on elementary con-
plies some useful elements of knowledge about the de- cepts and mathematical tools (sets and set operations,
cision situation; these are: the relevance of attributes binary relations), without recourse to any algebraic or
or criteria, the minimal subsets of attributes or criteria analytical structures.
(reducts) conveying the relevant knowledge contained
The decision rules entering the preference model have
in the exemplary decisions, the set of indispensable at-
a special syntax which involves partial evaluation profiles
tributes or criteria (core),
and dominance relations on these profiles. The traditional
 DRSA can deal with preference information concern-
preference models, which are the utility function and the
ing taxonomy-type classification, ordinal classification,
outranking relation, can be represented in terms of equiva-
choice, ranking, multiobjective optimization and deci-
lent decision rules. The clarity of the rule representation of
sion under uncertainty,
preferences enables one to see the limits of these aggrega-
 the preference model induced from preference infor-
tion functions. Several studies [22,23,26,50] presented an
mation structured by DRSA is expressed in a natural
axiomatic characterization of all three kinds of preference
and comprehensible language of “if . . . , then . . . ” deci-
models in terms of conjoint measurement theory and in
sion rules,
terms of a set of decision rules. The given axioms of “can-
 suitable procedures have been proposed to exploit the
cellation property” type are the weakest possible. In com-
results of application of the decision rule preference
parison to other studies on the characterization of aggre-
model on a set of objects or pairs of objects in order
gation functions, these axioms do not require any prelimi-
to workout a recommendation,
nary assumptions about the scales of criteria. A side-result
 no prior discretization of quantitative condition at-
of these investigations is that the decision rule preference
tributes or criteria is necessary,
model is the most general among the known aggregation
 heterogeneous information (qualitative and quantita-
functions.
tive, ordered and non-ordered, nominal and ordinal,
quantitative and numerical non-quantitative scales of
preferences) can be processed within DRSA, Future Directions
 the proposed methodology fulfills some desirable prop- The article shows that the Dominance-based Rough Set
erties for both rough set approach (the approximated Approach is a very powerful tool for decision analysis and
sets include lower approximation and are included in support. Its potential goes, however, beyond the theoreti-
upper approximation, and the complementarity prop- cal frame considered in this article. There are many possi-
erty is satisfied), and for multiple criteria decision anal- bilities of applying DRSA to real life problems. The non-
ysis (the decision rule preference model is formally exhaustive list of potential applications includes:
equivalent to the non-additive, non-transitive and non- Decision support in medicine: In this area there are al-
complete conjoint measurement model, and to a more ready many interesting applications (see, e. g., [35,36,
general model for preferences defined on all kinds of 40,59]), however, they exploit the classical rough set
scales), approach; applications requiring DRSA, which handle
 Rough Sets in Decision Making R 7785

ordered value sets of medical signs, as well as mono- 9. Geoffrion A, Dyer J, Feinberg A (1972) An interactive approach
tonic relationship between the value of signs and the for multi-criterion optimization, with an application to the op-
degree of gravity of a disease, are in progress; eration of an academic department. Manag Sci 19(4):357–368
10. Giove S, Greco S, Matarazzo B, Słowiński R (2002) Variable con-
Customer satisfaction survey: Theoretical foundations sistency monotonic decision trees. In: Alpigini JJ, Peters JF,
for application of DRSA in this field are available Skowron A, Zhong N (eds) Rough sets and current trends in
in [30], however, a fully documented application is computing. LNAI, vol 2475. Springer, Berlin, pp 247–254
still missing; 11. Gorsevski PV, Jankowski P (2008) Discerning landslide suscep-
Bankruptcy risk evaluation: This is a field of many po- tibility using rough sets. Comput Environ Urban Syst 32:5365
12. Greco S, Matarazzo B, Słowiński R (1998) A new rough set ap-
tential applications, as can be seen from promising
proach to evaluation of bankruptcy risk. In: Zopounidis C (ed)
results reported, e. g. in [12,53,54], however, a wider Operational tools in the management of financial risks. Kluwer,
comparative study involving real data sets is needed; Dordrecht, pp 121–136
Operational research problems, such as location, rout- 14. Greco S, Matarazzo B, Słowiński R (1999) The use of rough sets
ing, scheduling or inventory management: These are and fuzzy sets in MCDM. In: Gal T, Stewart T, Hanne T (eds)
problems formulated either in terms of classification Advances in multiple criteria decision making. Kluwer, Boston,
pp 14.1–14.59
of feasible solutions (see, e. g., [11]), or in terms of
15. Greco S, Matarazzo B, Słowiński R (1999) Rough approximation
interactive multiobjective optimization, for which the of a preference relation by dominance relations. Eur J Oper Res
IMO-DRSA [32] procedure is suitable; 117:63–83
Finance: This is a domain where DRSA for decision un- 16. Greco S, Matarazzo B, Słowiński R (2000) Extension of the
der uncertainty has to be combined with interactive rough set approach to multicriteria decision support. INFOR
multiobjective optimization using IMO-DRSA; some 38:161–196
13. Greco S, Matarazzo B, Słowiński R, Tsoukias A (1998) Exploita-
promising results going in this direction have been tion of a rough approximation of the outranking relation in
presented in [31]; multicriteria choice and ranking. In: Stewart TJ, van den Hon-
Ecology: Assessment of the impact of human activity on ert RC (eds) Trends in multicriteria decision making. LNEMS,
the ecosystem is a challenging problem for which the vol 465. Springer, Berlin, pp 45–60
presented methodology is suitable; the up to date ap- 17. Greco S, Matarazzo B, Słowiński R, Stefanowski J (2001) An al-
gorithm for induction of decision rules consistent with dom-
plications are based on the classical rough set concept
inance principle. In: Ziarko W, Yao Y (eds) Rough sets and
(see, e. g., [7,43]), however, it seems that DRSA han- current trends in computing. LNAI, vol 2005. Springer, Berlin,
dling ordinal data has a greater potential in this field. pp 304–313
18. Greco S, Matarazzo B, Słowiński R (2001) Rough sets theory for
multicriteria decision analysis. Eur J Oper Res 129:1–47
19. Greco S, Matarazzo B, Słowiński R (2001) Rough set approach to
Bibliography decisions under risk. In: Ziarko W, Yao Y (eds) Rough sets and
1. Brans JP, Mareschal B (2005) PROMETHEE Methods. In: current trends in computing. LNAI, vol 2005. Springer, Berlin,
Figueira J, Greco S, Ehrgott M (eds) Multiple criteria decision pp 160–169
analysis: State of the art surveys. Springer, Berlin, pp 163–195 20. Greco S, Matarazzo B, Słowiński R, Stefanowski J (2002) Min-
2. Chankong V, Haimes YY (1978) The interactive surrogate worth ing association rules in preference-ordered data. In: Hacid M-
trade-off (ISWT) method for multiobjective decision-making. S, Ras ZW, Zighed DA, Kodratoff Y (eds) Foundations of intelli-
In: Zionts S (ed) Multiple Criteria Problem Solving. Springer, gent systems. LNAI, vol 2366. Springer, Berlin, pp 442–450
Berlin, pp 42–67 21. Greco S, Matarazzo B, Slowinski R (2002) Multicriteria classifica-
3. Chankong V, Haimes YY (1983) Multiobjective decision making tion. In: Kloesgen W, Zytkow J (eds) Handbook of data mining
theory and methodology. Elsiever Science, New York and knowledge discovery. Oxford University Press, New York,
4. Figueira J, Greco S, Ehrgott M (eds) (2005) Multiple criteria de- pp 318–328
cision analysis: State of the art surveys. Springer, Berlin 22. Greco S, Matarazzo B, Słowiński R (2002) Preference repre-
5. Figueira J, Mousseau V, Roy B (2005) ELECTRE methods. In: sentation by means of conjoint measurement & decision rule
Figueira J, Greco S, Ehrgott M (eds) Multiple criteria decision model. In: Bouyssou D, Jacquet-Lagrèze E, Perny P, Słowiński R,
analysis: State of the art surveys. Springer, Berlin, pp 133–162 Vanderpooten D, Vincke P (eds) Aiding decisions with multiple
6. Fishburn PC (1967) Methods of estimating additive utilities. criteria – essays in honor of Bernard Roy. Kluwer, Dordrecht,
Manag Sci 13(7):435–453 pp 263–313
7. Flinkman M, Michalowski W, Nilsson S, Słowiński R, Susmaga 23. Greco S, Matarazzo B, Słowiński R (2004) Axiomatic character-
R, Wilk S (2000) Use of rough sets analysis to classify Siberian ization of a general utility function and its particular cases in
forest ecosystem according to net primary production of phy- terms of conjoint measurement and rough-set decision rules.
tomass. INFOR 38:145–161 Eur J Oper Res 158:271–292
8. Fortemps P, Greco S, Słowiński R (2008) Multicriteria decision 24. Greco S, Matarazzo B, Słowiński R (2004) Dominance-based
support using rules that represent rough-graded preference Rough Set Approach to Knowledge Discovery, (I) – General
relations. Eur J Oper Res 188:206–223 Perspective, (II) – Extensions and Applications. In: Zhong N,
7786 R Rough Sets in Decision Making

Liu J (eds) Intelligent technologies for information analysis. 43. Rossi L, Słowiński R, Susmaga R (1999) Rough set approach
Springer, Berlin, pp 513–612 to evaluation of stormwater pollution. Int J Environ Pollut
25. Greco S, Pawlak Z, Słowiński R (2004) Can Bayesian confirma- 12:232–250
tion measures be useful for rough set decision rules? Eng Appl 44. Roy B (1996) Multicriteria Methodology for Decision Aiding.
Artif Intell 17:345–361 Kluwer, Dordrecht
26. Greco S, Matarazzo B, Słowiński R (2005) Decision rule ap- 45. Roy B (1999) Decision-aiding today: What should we expect.
proach. In: Figueira J, Greco S, Ehrgott M (eds) Multiple crite- In: Gal T, Stewart T, Hanne T (eds) Advances in Multiple Criteria
ria decision analysis: State of the art surveys. Springer, Berlin, Decision Making. Kluwer, Boston, pp 1.1–1.35
pp 507–563 46. Roy B, Bouyssou D (1993) Aide Multicritère à la Décision: Méth-
27. Greco S, Matarazzo B, Słowiński R (2006) Dominance-based odes et Cas. Economica, Paris
rough set approach to decision involving multiple decision 47. Shoemaker PJH (1982) The expected utility model: its variants,
makers. In: Greco S, Hata Y, Hirano S, Inuiguchi M, Miyamoto S, purposes, evidence and limitations. J Econ Lit 20:529–562
Nguyen HS, Slowinski R (eds) Rough Sets and Current Trends 48. Słowiński R (1993) Rough set learning of preferential attitude
in Computing, RSCTC. LNAI, vol 4259. Springer, Berlin, pp 306– in multi-criteria decision making. In: Komorowski J, Ras ZW
317 (eds) Methodologies for intelligent systems. LNAI, vol 689.
28. Greco S, Matarazzo B, Słowiński R (2006) Dominance-based Springer, Berlin, pp 642–651
rough set approach to case-based reasoning. In: Torra V, 49. Słowiński R, Greco S, Matarazzo B (2002) Rough set analysis of
Narukawa Y, Valls A, Domingo-Ferrer J (eds) Modelling deci- preference-ordered data. In: Alpigini JJ, Peters JF, Skowron A,
sions for artificial intelligence. LNAI, vol 3885. Springer, Berlin, Zhong N (eds) Rough sets and current trends in computing.
pp 7–18 LNAI, vol 2475. Springer, Berlin, pp 44–59
29. Greco S, Matarazzo B, Słowiński R (2007) Dominance-based 50. Słowiński R, Greco S, Matarazzo B (2002) Axiomatization of util-
rough set approach as a proper way of handling graduality ity, outranking and decision-rule preference models for multi-
in rough set theory. In: Transactions on Rough Sets VII. LNCS, ple-criteria classification problems under partial inconsistency
vol 4400. Springer, Berlin, pp 36–52 with the dominance principle. Control Cybern 31:1005–1035
30. Greco S, Matarazzo B, Słowiński R (2007) Customer satisfaction 51. Słowiński R, Greco S, Matarazzo B (2005) Rough set based de-
analysis based on rough set approach. Z Betr 16(3):325–339 cision support, chapter 16. In: Burke EK, Kendall G (eds) Search
31. Greco S, Matarazzo B, Słowiński R (2007) Financial portfolio de- methodologies: Introductory tutorials in optimization and de-
cision analysis using dominance-based rough set approach. In- cision support techniques. Springer, New York, pp 475–527
vited paper at the 22nd European Conference on Operational 52. Słowiński R, Greco S, Matarazzo B (2007) Dominance-based
Research (EURO XXII), Prague, 08–11 July rough set approach to reasoning about ordinal data, keynote
32. Greco S, Matarazzo B, Słowiński R (2008) Dominance-based lecture. In: Kryszkiewicz M, Peters JF, Rybiński H, Skowron
rough set approach to interactive multiobjective optimization. A (eds) Rough sets and intelligent systems paradigms. LNAI,
In: Branke J, Deb K, Miettinen K, Slowinski R (eds) Multiob- vol 4585. Springer, Berlin, pp 5–11
jective optimization: Interactive and evolutionary approaches. 53. Słowiński R, Zopounidis C (1995) Application of the rough set
Springer, Berlin approach to evaluation of bankruptcy risk. Int J Intell Syst Acc
33. Jaszkiewicz A, Słowiński R (1999) The “Light Beam Search” ap- Finance Manag 4:27–41
proach – an overview of methodology and applications. Eur J 54. Słowiński R, Zopounidis C, Dimitras AI (1997) Prediction of
Oper Res 113:300–314 company acquisition in Greece by means of the rough set ap-
34. Martel JM, Matarazzo B (2005) Other outranking approaches. proach. Eur J Oper Res 100:1–15
In: Figueira J, Greco S, Ehrgott M (eds) Multiple criteria decision 55. Stefanowski J (1998) On rough set based approaches to induc-
analysis: State of the art surveys. Springer, Berlin, pp 197–262 tion of decision rules. In: Polkowski L, Skowron A (eds) Rough
35. Michalowski W, Rubin S, Słowiński R, Wilk S (2003) Mobile clin- sets in data mining and knowledge discovery, vol 1. Physica,
ical support system for pediatric emergencies. J Decis Support Heidelberg, pp 500–529
Syst 36:161–176 56. Tsoukias A, Vincke P (1995) A new axiomatic foundation of the
36. Michalowski W, Wilk S, Farion K, Pike J, Rubin S, Słowiński R partial comparability theory. Theory Decis 39:79–114
(2005) Development of a decision algorithm to support emer- 57. Steuer RE, Choo E-U (1983) An interactive weighted Tcheby-
gency triage of scrotal pain and its implementation in the MET cheff procedure for multiple objective programming. Math
system. INFOR 43:287–301 Program 26:326–344
37. Miller GA (1956) The magical number seven, plus or minus two: 58. Wierzbicki AP (1980) The use of reference objectives in multi-
some limits in our capacity for processing information. Psychol objective optimization. In: Fandel G, Gal T (eds) Multiple crite-
Rev 63:81–97 ria decision making, theory and applications. Springer, Berlin,
38. Pawlak Z (1982) Rough Sets. Int J Comput Inf Sci 11:341–356 pp 468–486
39. Pawlak Z (1991) Rough Sets. Kluwer, Dordrecht 59. Wilk S, Słowiński R, Michalowski W, Greco S (2005) Supporting
40. Pawlak Z, Słowiński K, Słowiński R (1986) Rough classification triage of children with abdominal pain in the emergency room.
of patients after highly selective vagotomy for duodenal ulcer. Eur J Oper Res 160:696–709
Int J Man-Machine Stud 24:413–433 60. Zionts S, Wallenius J (1976) An interactive programming
41. Pawlak Z, Słowiński R (1994) Rough set approach to multi-at- method for solving the multiple criteria problem. Manag Sci
tribute decision analysis. Eur J Oper Res 72:443–459 22:652–663
42. Roberts F (1979) Measurement theory, with applications to de- 61. Zionts S, Wallenius J (1983) An interactive multiple objective
cision making, utility and the social sciences. Addison-Wesley, linear programming method for a class of underlying nonlin-
Boston ear utility functions. Manag Sci 29:519–523

View publication stats