Raymond Bisdor
Fa ulty of Law, E onomi s andFinan e, University of Luxembourg, Avenue dela
Faïen erie162a, L-1511 LUXEMBOURG.
Mar Pirlot
Fa ulté Polyte hnique de Mons,9, rue de Houdain,B-7000 Mons, Belgium.
Mar Roubens
Fa ulté Polyte hnique de Mons,9, rue de Houdain,B-7000 Mons, Belgium.
Abstra t
We explore the extension of the notion of kernel (independent, dominant or
ab-sorbent,non-empty subset)ofa digraphto valuedgraphs (orvaluedrelations). We
denevariousnaturalextensionsandshowtherelationshipbetweenthem.Thiswork
haspotential interestfor appli ationsin hoi e de isionproblems.
Keywords:De isionMaking,Outrankingmethods,Choi e,Kernel,Digraphs,
Valuedrelations, Fuzzy sets.
1 Introdu tion
Dominant kernels have been initially proposed as solutions of games by von
Neumann and Morgenstern (1944). Roy (1968)has suggested to use them as
a set of potentially good andidates among whi h to hoose when fa ing a
de ision problem. In the latter ontext, the preferen e relation on the set of
alternatives is usually built using some sort of a majority rule with vetoes
(Roy and Bouyssou, 1993), whi h results in a relation that is not ne essarily
transitive or omplete and may have y les.
Email addresses: raymond.bisdorffuni.lu(Raymond Bisdor),
(or fuzzy) one. The value atta hed to an ar
(a, b)
may express for instan e the degree with whi ha
is preferred tob
, a redibility index assigned to the preferen eofa
overb
,et .Intheperspe tiveofgeneralizingtovaluedrelations themethodsusedfor hoosingamonganitesetofalternativesonthebasisofrisp (
{0, 1}
-valued) relations,one an think of a generalizationof the notion of kernel ofvalued graphsor relations.Asisusualwhenstudyingvalued(fuzzy) ounterpartsofnotionsthathaverst
beenintrodu edforordinary( risp)sets,thereareseveralwaysofgeneralizing
thenotionofkernel.The maingoalofthis paperistoexplorethe relationship
between various natural denitions of kernels for valued graphs and see also
how they relate to kernels of risp graphs.Our analysis is restri ted to nite
graphs.
Inse tion2,were allthedenitionsofkernelsinthe risp ase,relatingmainly
the set-theoreti denition (interms of independent, dominant or absorbent,
non-empty subset) and the algebrai one by means of the so- alled kernel
equationsystem.Notethat we onsidertwosorts ofkernels,the dominantand
the absorbent one, that are just dual of one another in the sense that the
absorbent kernels of a graph are the dominant kernels of the graph obtained
by reversing the orientation of the ar s of the originalone.
Se tion3introdu esavaluationofthear sbydegreesbelongingtoanordered
set
L
;this set isequipped withanegation (anantitoneone-to-oneoperation). IfL
ontains anodd numberof elements, there isan elementthat is equal to itsnegation;thisplaysthe roleofanundeterminedlevel0
thatmaybeused forinstan e for odingla kofinformation,likeamissingar inthegraph.Onedistinguishesthe levels of
L
that are above0
(positive)and those below0
(negative).Wethen assume agraph with ar s assigned avalue inL
. IfL
has a`0
,we maysuppose,withoutlossofgenerality,that thegraphis omplete, assigningthe value0
to the missing ar s. This graph isalsoa valued relation and we use indierentlyboth terminologies.The rst idea for generalizing kernels, is to ut the relation above
0
, yielding a risp relation orgraph, and to make prot of the set-theoreti denition ofthe kernels of the obtained risp graph. We all su h kernels
L
-independent, dominant (or absorbent) hoi es, reservingthe term kernel forthe solutionsofthekernelequationsystem.Avarietyofdegreesofquali ationaredened
andasso iatedtothesesubsets;theymaybeinterpretedas hara terizingtheir
quality asa set of good alternatives(for dominantsubsets) oras aset of bad
alternatives (for absorbent subsets). We show that
L
-independent, dominant (resp.absorbent) hoi es anbedenedalternativelybymeansoftheirdegreethat are valued (fuzzy) sets. We identify a subset of solutions,the maximally
sharp ones, and we show in se tion 4 that they orrespond to the set of
L
-independent dominant (or absorbent) hoi es that were initiallydened. Theproofusingxpointargumentis onstru tive,enablingtondthe
L
-valued kernels starting from the independent dominant (or absorbent) hoi es. Thedegreeofsharpness oftheformerisshowntobeequaltothedegreeofgoodness
(or badness) of the latter. From a theoreti al point of view, this is rather
satisfa tory sin e the various generalizations of the kernel lead to essentially
the same notion, though in a non-trivial way, as the reader will see. One
should add that there are perspe tives, using
L
-valued kernels, to ontribute to methods for hoosing among a set of alternatives, but su h developmentsare left for another publi ation.
2 Kernels of risp binary relations
2.1 Independent dominant and absorbent hoi es
In this se tion, we onsider a nite set
X
= {a, b, c, d, . . .}
of alternatives endowed with abinaryrelationR
.Thisrelationisinterpretedasapreferen e relation; we do not assume any parti ular property, su h as transitivity, ofR
; the preferen e maythusresult fromthe aggregation of multipleattribute informationthroughsomekindofamajorityruleora ordingwithsomeformofa on ordan e-dis ordan e prin iplelikethey are implemented forinstan e
inthe Ele tremethods(Roy, 1968; Roy and Bouyssou, 1993).
We denoteby
a R b
the fa tthat the pair(a, b)
belongsto the relationR
; we phrasea R b
asa
is preferred tob
(this not ex luding thatb
may also be preferred toa
;R
is not supposed tobean asymmetri relation).The set
X
endowed with the relationR
an also be viewed as a dire ted graph (digraph, for short)G
= (X, R)
whereX
represents the set of verti es and(a, b)
is anar of the graphif and onlyif(a, b) ∈ R
.It has been suggested in the Ele tre I method (Roy, 1969) to use
inde-pendent dominant subsets of X as potential good hoi e if the relation
R
is a y li .Denition 1 (Independent dominantand absorbent hoi es).
that
∀a 6= b ∈ K, (a, b) 6∈ R
.A dominant hoi e or dominant subset in
G
is a hoi e setK
⊂ X
su h that∀a 6∈ K, ∃ b ∈ K, (b, a) ∈ R
.Anabsorbent hoi e orabsorbentsubset in
G
isa hoi esetK
⊂ X
su hthat∀a 6∈ K, ∃ b ∈ K, (a, b) ∈ R
.Example 1. The graph
G
1
= (X
1
, R
1
)
, whereX
1
= {a, b, c, d}
andR
1
=
{(a, b), (a, c), (a, d), (b, c), (b, d)}
is illustrated in gure 1. There is a unique independent dominant hoi e (good hoi e) that is the singleton{a}
; the unique independentabsorbent hoi e (bad hoi e)is{c, d}
.a
b
c
d
Figure1.Graph ofthe preferen erelation in example 1
Example 2. The graph
G
2
= (X
2
, R
2
)
, whereX
2
= {a, b, c, d}
andR
2
=
{(a, b), (b, c), (c, d), (d, a)}
is illustratedingure2.In thisexample,{a, c}
and{b, d}
are two independent hoi es that are both dominantand absorbent.a
b
c
d
Figure2.Graph ofthe preferen erelation in example 2
If a relation
R
is a y li , then it has a unique dominant (resp. absorbent) hoi e. If, in addition, the relation is transitive, it is a partial order and thesetofalternativesthathavenoprede essor(resp.su essor)inthegraphforms
an independent dominant (resp. absorbent) hoi e. The above existen e and
uniqueness property is amotivationfor hoosing among the alternativesthat
belong to the independent dominant hoi e. In ase one aims at eliminating
bad alternatives, asymmetri argumentwould lead to hoosing theminthe
independent absorbent hoi e set, again for a y li digraphs. We loose su h
We now introdu e an algebrai denition of a independent dominant (resp.
absorbent) hoi e. Ea h subset
K
⊆ X
is asso iated its hara teristi (row) ve torY
: X → {0, 1}
withY
(a) =
1
ifa
∈ K
0
otherwise.
(2.1) for alla
∈ X
.Denition 2 (Dominant and absorbent kernels).
Adominant kernel (Bisdorand Roubens,1996a)isa hara teristi ve tor
Y
that isa solution of the following system of Boolean equations:(Y ◦ R )(a) = ¬Y (a),
for alla
∈ X,
(2.2)where
Y
◦ R
is the max min produ t of the Boolean matri esY
andR
,i.e.(Y ◦ R )(a) = max
b
6=a
min (Y (b), R (b, a)),
for all
a
∈ X
and
¬
istheordinarynegation,i.e.¬Y (a) = 1−Y (a)
.Thesystemofequations (2.2) reads more expli itlyas:max
b
6=a
min (Y (b), R (b, a)) = 1 − Y (a),
for all
a
∈ X.
(2.3)We all absorbent kernel a hara teristi ve tor
Y
, solution of the system of Booleanequations: for alla
∈ X
,(R ◦ Y
t
)(a) = max
b6=a
min (R(a, b), Y (b)) = ¬Y (a) = 1 − Y (a),
(2.4)
where
Y
t
denotes the transposed ( olumn) hara teristi ve tor.
We denote
K
dom
(G)
(
K
abs
(G)
) the possibly empty set of dominant (resp.
ab-sorbent)kernels in
G
,i.e. the solutionsof thesystem of equations(2.2) (resp. (2.4)).Proposition 1.
Thesetofindependentdominant(resp.absorbent) hoi esof
X
endowed with therelationR
)
isthesetofdominant(resp.absorbent)kernelsofG
= (X, R )
.R
representing the relationR
into four partsR
K K
, R
K K
, R
K K
, R
KK
; for in-stan eR
K K
orresponds tothe asesR(a, b)
inwhi hbotha
andb
belong toK
;the three otherpartsare dened similarly.The system ofequation(2.2) is rewritten as[Y
K
Y
K
] ◦
R
K K
R
K K
R
KK
R
KK
= [Y
K
Y
K
].
(2.5)One veries easily that ve tor
Y
exa tly hara terizes a subsetK
that is independent and dominant if and only if the above system of equations issatised. A similar argument applies to independent absorbent hoi es and
absorbent kernels.
Inthesequelweshalluseindierentlythetermsindependentdominant(resp.
absorbent) hoi es and dominant (resp. absorbent)kernels.
2.3 Histori al note
Followingtheobservationthatthe independentabsorbent hoi einana y li
graph orresponds to the kernel of the asso iated Grundy fun tion, Riguet
(1948)introdu edthenamenoyau (kernel)forsu ha hoi eset.(Absorbent)
kernelswereinthesequelextensivelystudiedbyBerge(1958,1970)inthe
on-textofmodelingtheNimgame.More resultson(absorbent)kernels,
on ern-ingsolutionsof dierentgames,havebeen reportedby S hmidtandStröhlein
(1985, 1989). Re ently, Bang-Jensen and Gutin (2001) reviewed the link
be-tween kernel-solvability and perfe t graphs.
Internally stableand dominating hoi es,i.e.dominantkernelswere originally
introdu ed by J. von Neumann and O. Morgenstern under the name game
solution in the ontext of game theory (von Neumann and Morgenstern,
1944). B.Roy proposed the same on ept in the ontext of the multi riteria
Ele tre de ision aid methods (Roy, 1968, 1985; Roy and Bouyssou, 1993).
Ambiguous hoi es, i.e. both dominant and absorbent hoi es at the same
time,were proposedby Bisdor(2002a)aspotential luster andidatesinthe
ontextof multi riteria lustering.
Inthegeneral(non-oriented)theoryofgraphs,kernelsappear underthename
ofindependentdominatingsetsinHaynesetal.(1998).Ourorienteddominant
(resp. absorbent)version isthere identiedas aninside (resp.outside)
semi-dominatingset orsometimesmoresimplyasaninkernel (resp.outkernel).
Theabsorbentversionofthekernelequationsystem (2.4)wasrstintrodu ed
by S hmidtandStröhlein (1985,1989)inthe ontextof theirthorough
(1996a,b); Bisdor(1997).
3 Kernels in valued digraphs
3.1
L
-valued binary relationsLet
X
be a nite set of alternatives andR
˜
, a valued binary relation onX
. We assume that the values assigned byR
˜
to ea h pair of alternativesbelong toa nite ordered setL
: {c
0
, c
1
, . . . , c
M
}
withc
0
< c
1
< . . . < c
M
. The value˜
R(a, b) ∈ L
is often interpreted as the degree of redibility of the assertiona
is preferred tob
(ora
is at least as good asb
). We allsu h a relation aL
-valued binary relation(or shortlyL
-vbr).The set
L
isendowed with a negation operator¬
that mapsc
i
ontoc
M
−i
for alli
= 0, . . . , M
. This operator has a xpointc
m
iM
is equal to2m
, withm
a positive integer. Bisdor (2002a) has for efully argued that one should takeM
an even number in pra ti e; the median redibility levelc
m
is then interpreted asalogi allyundetermined value; itmayen ode amissingar . Inase
M
isodd thenegationhasnoxpointatallandthereisthusnologi ally undeterminedvalue. Fromnowon,fornotationalsimpli ity,wemapthesetL
monotoni ally ontothe following sets of integers, endowed with their naturalorder:
−m, −(m − 1), . . . , −1, 0, 1, . . . , m − 1, m
ifM
= 2m
−m, −(m − 1), . . . , −1, 1, . . . , m − 1, m
ifM
= 2m − 1;
(3.1)
weuseornotthevalue
0
dependingonwhether|L|
isoddoreven,respe tively; in ase|L|
is odd (i.e.M
= 2m = |L| − 1
), the median redibility levelc
m
is mapped onto0
. Sin e this mappingisan orderisomorphism, the operatorsmax
andmin
an be seen as a ting indierently onL
or its image on the integers;moreover, the negationoperator¬
onL
orresponds with takingthe oppositeofanintegerintheimageset. Wehen eforthidentifyL
withthesets ofintegersdened in(3.1),endowed withtheirnaturalorderandthenegationoperator,whi h isjust takingthe opposite.Wedene the subset ofpositive
levels
L
>0
(resp. negative levels
L
<0
) as being the set of positive integers
{1, 2, . . . , m}
(resp. negative integers{−m, . . . , −1}
); the set of nonnegative levelsL
≥0
(resp. nonpositive levels
L
≤0
)is dened as the omplement of
L
<0
(resp.
L
>0
).Of ourse,thereisnodieren ebetween
L
>0
andL
≥0
(orbetweenL
<0
andL
≤0
)unless
|L|
isodd(i.e.M
iseven);inthelatter ase,thedieren e onsists of the median redibility level0
. In the sequel, forx
∈ L
, we shall writex >
0
(resp.<
0, ≥ 0, ≤ 0
) forx
∈ L
Interpreting theelementsof
L
astruthvaluesorlogi allevels,assuggestedby Bisdor (2002a), leads to all the elements ofL
>0
, true levels and those of
L
<0
,falselevels. If
R(a, b) > 0
we ouldsay thattheproposition(a, b) ∈ R
isL
-true. If, on the ontrary,R
(a, b) < 0
, the proposition(a, b) ∈ R
isL
-false. IfR(a, b) = 0
, i.e. the median level, we say that the proposition(a, b) ∈ R
isL
-undetermined. The redibility level of a onjun tion (resp. a disjun tion) ofL
-valued propositions is obtained by using themin
(resp. themax
) operatordened onL
.Twosimplespe ial asesof
L
are thefollowing.IfL
hasonlytwoelements,we getba ktothe lassi alBooleanlatti eL
2
= {c
0
, c
1
} = {−1, 1}
;c
0
or−1
isthe symbol for false, whilec
1
or+1
is the symbol for true. The three-valued latti eL
3
(M
= 2m = 2
) orresponds to the ase in whi h−1
represents false,0
represents a logi allyundetermined level and+1
represents true.3.2 Independent dominant or absorbent hoi es in a
L
-valued graphWedenoteby
G
L
= (X, R)
a ompleteL
-valueddigraphwithvertexsetX
;an ar(a, b)
ofG
L
isassignedthe value
R(a, b)
.Thesimplestway ofgeneralizing thenotionsofindependentdominant(resp.absorbent) hoi einG
L
isthrough
using a ut. Dene the appli ation
τ
mappingthe latti eL
onto the Boolean latti eL
2
asfollows:thesetoflevelsL
>0
areallmappedonto
+1
(true)while thelevelsofL
≤0
are allmapped onto
−1
(false).This appli ationextendsto a transformationof theL
-valued graphG
L
= (X, R)
into the ordinary graphτ G
= (X, τ R)
;(a, b)
is an ar of the relationτ R
iR(a, b) > 0
; otherwise, there isno ar(a, b)
in the graphτ G
.Denition 3 (Independent dominant or absorbent hoi es in
G
L
). The set
K
⊆ X
is anindependent dominant (resp. absorbent) hoi e inG
L
= (X, R)
if it isanindependent dominant(resp. absorbent) hoi e in
τ G
= (X, τ R)
. Note that with this denition an independent dominant (resp. absorbent)hoi e in
G
L
= (X, R)
isan ordinary subsetofX
and not a fuzzy subsetofX
.Using the values in
L
, we an dene variousL
-valued quali ation degrees atta hed toany non-empty subsetK
ofX
.Denition 4 (Degrees of quali ation of hoi es). Let
K
be a non-empty subset ofX
.The degree of independen e of the hoi eK
is dened as∆
ind
(K) =
m
if|K| = 1
,min
b
6=a
b
∈K
min
a∈K
{¬ R(a, b)}
otherwise.
K
is onsidered tobeL
-independent if∆
ind
(K) > 0
.
The degree of dominan e of a hoi e
K
orresponds to∆
dom
(K) =
m
if K =X,
min
a
6∈K
max
b
∈K
{R(b, a)}
otherwise.
(3.3)
K
is onsidered tobeL
-dominant if∆
dom
(K) > 0
.
The degree of absorben e of a hoi e
K
is equal to∆
abs
(K) =
m
if K = X,
min
a
6∈K
max
b
∈K
{R(a, b)}
otherwise.
(3.4)
K
is onsidered tobeL
-absorbent if∆
abs
(K) > 0
.
The quali ation of
K
as being an independent and dominant hoi e orre-sponds toQ
i−dom
(K) = min(∆
ind
(K), ∆
dom
(K)).
(3.5)
The quali ation of
K
as being an independent and absorbent hoi e, orre-sponds toQ
i−abs
(K) = min
∆
ind
(K), ∆
abs
(K)
(3.6)Denition 5 (Potentially good and bad hoi es). The set
K
⊆ X
is a po-tentially good hoi e ifQ
i−dom
(K) ∈ L
>0
, i.e. if
K
isL
-independent andL
-dominant.The setK
⊆ X
is a potentially bad hoi e ifQ
i−abs
(K) ∈ L
>0
, i.e.
if
K
isL
-independent andL
-absorbent. We denote byC
p−good
(G
L
)
(respe tivelyC
p−bad
(G
L
)
) the (possibly empty) set of potentiallygood(respe tively potentially bad) hoi es inG
L
. The
or-responden e between potentially good hoi es (resp. potentially bad hoi es)
andindependentdominant(resp.absorbent) hoi esof
G
L
(denition3)isnot
entirely straightforward due to the possible existen e of the undetermined
medianlevel
0
.Proposition 2 (L.Kitainik, 1993).
Let
G
L
= (X, R)
beanL
-valued digraphand letτ G
= (X, τ R)
be its asso i-ated risp digraph.Wehave:C
i−dom
(G
L
) ⊆ K
dom
(τ G)
and
C
i−abs
(G
L
) ⊆ K
abs
(τ G).
Proof. Consider
K
6= ∅
su hthat∆
ind
(K) ∈ L
>0
∆
ind
(K) ∈ L
>0
⇔ min
b6=a
b
∈K
min
a∈K
(¬ R(a, b)) > 0
⇔ ∀a 6= b ∈ K : R(a, b) < 0
⇒ ∀a 6= b ∈ K : (a, b) 6∈ τ R
⇔ K
is anindependent hoi e inτ G
= (X, τ R).
Consider now
K
6= ∅
su h that∆
dom
(K) > 0
.∆
dom
(K) ∈ L
>0
⇔ min
a
6∈K
max
b
∈K
R(b, a) > 0
⇔ ∀a 6∈ K, ∃ b ∈ K : R(b, a) > 0
⇔ ∀a 6∈ K, ∃ b ∈ K : (b, a) ∈ τ R
⇔ K
is adominant set inτ G.
Usingthe same arguments,
∆
abs
(K) > 0 ⇔ K
isan absorbent set in
τ G.
The propositionis animmediate onsequen e of the previous results.
From the rst part of the previous proof (independen e), it learly appears
thatthe sets
C
p−good
(G
L
)
andK
dom
(τ G)
maybedierentonlyif
R
(a, b)
takes thelogi allyundeterminedvalue0
;theremayindeedexistsetsK
⊆ X
thatare independent dominant hoi es ofτ G
but have aquali ationQ
i-dom
(K)
equal to0
. A similar remark holds for independent absorbent hoi es. The follow-ing orollary establishes the pre ise orresponden e between potentially good(resp. potentially bad) hoi es and independent dominant (resp.absorbent)
hoi es in
τ G
(whi h are also the dominant (resp. absorbent) kernels inτ G
, asestablished in proposition 1).Corollary 1.
The set of potentially good(resp. potentially bad) hoi es in
G
L
= (X, R)
is theset ofindependentdominant(resp.absorbent) hoi esK
inτ G
= (X, τ R)
for whi hQ
i-dom
Consider the rst part of the proof of proposition 2.We have
∆
ind
(K) ∈ L
>m
⇔ K
is anindependent set inτ G
i
R(a, b) 6= 0
foralla, b
∈ K
.Therestoftheprooffollowsthatofproposition 2.Example 3. The following example (see gure 3) has been suggested by
BernardRoy(Lausanne, 1995;see Bisdor(2000)).Let
G
L
besu h that
X
=
{a, b, c}
,L
= {−5, . . . , 0, . . . , 5}
andR
: {R(a, b) = −5
,R(b, a) = R(c, a) =
R(c, b) = 1
,R(b, c) = R(a, c) = 5}
.a
b
c
1
5
1
1
5
Figure3.Graph ofthe preferen erelation in example 3
Table 1
Degreesof quali ation ofall hoi es in example 3
hoi e
∆
ind
∆
dom
∆
abs
Q
i−dom
Q
i−abs
{a}
5
−5
1
−5
1
{b}
5
1
−5
1
−5
{c}
5
1
5
1
5
{a, b}
−1
5
1
−1
−1
{a, c}
−5
1
5
−5
−5
{b, c}
−5
1
5
−5
−5
{a, b, c}
−5
5
5
−5
−5
Table 3 shows the degrees of quali ation for all hoi es in
G
L
. All
single-tons are
L
-independent. All hoi es, ex ept{a}
, areL
-dominant and all areL
-absorbent, ex ept{b}
. Therefore, the singletons{b}
and{c}
give poten-tially good single hoi es whereas{a}
and{c}
give potentially bad single hoi es. Choi es onsisting of pairs, as well asthe hoi e onsisting ofX
, are neither potentially good nor potentially bad, as they all la k the requiredOne may argue that the indi es
Q
i−dom
and
Q
i−abs
an be used to sele t a
good hoi e amongseveral possibleones;forinstan e,inexample3,onesees
from table 3 that
{b}
is a more onvin ing potentially good hoi e than{c}
sin e{c}
is alsoa potentiallybad hoi e. We referthe readerto Bisdorand Roubens (2004) forfurther developments of this idea.Proposition2 suggests a way of omputing the good hoi es (and the bad
ones) of
L
-valued graphs: ompute the independent dominant hoi es ofτ G
; then verify, by omputing their degree of independen e∆
ind
and he king
whether
∆
ind
>
0
,that they are potentiallygood hoi es.
3.3
L
-valued dominant and absorbent kernelsWe now address the generalization of the kernel equation systems (2.2) and
(2.4). Choi es and kernels of
L
-valued relations an be dened as fuzzy sets; the degree of membership of an elementa
ofX
to su h a set belongs to the setL
.A hara teristi ve tor
Y
˜
ofafuzzysubsetofX
isanappli ationfromX
ontoL
, i.e. a row ve tor[ ˜
Y
(x), x ∈ X]
. IfY
˜
is to be interpreted as a hoi e and depending on the ontext,Y
˜
(a)
, for anya
∈ X
, may be interpreted as the degree of truth or the redibility of the assertion elementa
belongs to the hoi eY
˜
. Values ofY
˜
(a)
above0
are interpreted as meaning thata
tends moretobelongtoY
˜
than not;the onverse interpretationholdswhenY
˜
(a)
is below0
;Y
˜
(a) = 0
represents anundetermined situation.It is rather straightforward to extend the kernel equation systems (2.2) and
(2.4) to the ase in whi h
R
isL
-valued and we a ept solutions that areL
-valued. Theseequations be ome: for alla
∈ X
:( ˜
Y
◦ R)(a) = max
b
6=a
h
min
Y
˜
(b), R(b, a)
i
= ¬ ˜
Y
(a) = − ˜
Y
(a),
(3.7)(R ◦ ˜
Y
t
)(a) = max
b
6=a
h
min
R(a, b), ˜
Y
(b)
i
= ¬ ˜
Y
(a) = − ˜
Y
(a).
(3.8)Applying the negation operator to the above equations yields the following
formulation that willbe used inthe sequel:
˜
Y
(a) = min
b
6=a
[max(− ˜
Y
(b), −R(b, a))],
∀a ∈ X
(3.9)
˜
Y
(a) = min
b
6=a
[max(−R(a, b), − ˜
Y
(b))],
∀a ∈ X.
We denote by
Y
dom
(G
L
)
(resp.Y
abs
(G
L
)
) the set of solutions of equation system 3.7 (resp. 3.8).Example4. Consider
G
L
= (X, R)
withX
= {a, b, c}
,L
= {−5, . . . , 0, . . . , 5}
(M
= 2m = 10
)andR
= {R(a, b) = −3}
,R
(a, c) = 4
,R
(b, a) = 1
,R
(b, c) =
5
,R(c, a) = 2
,R(c, b) = 3}
(see gure 4).a
b
c
1
−3
5
3
2
4
Figure4.Graph ofthe preferen erelation in example 4
The orresponding dominantkernel equation system is:
h
˜
Y
(a) ˜
Y
(b) ˜
Y
(c)
i
◦
-−3
4
1
-5
2
3
-
=
h
− ˜
Y
(a) − ˜
Y
(b) − ˜
Y
(c)
i
(3.11)Its solutionsare shown intable 2.Solution
Y
˜
0
is ompletelyL
-undetermined; Table 2Solutionsof EquationSystem (3.11).The starredsolutions aremaximally sharp.
solution
Y
˜
(a)
Y
˜
(b)
Y
˜
(c)
˜
Y
0
0
0
0
˜
Y
1
−1
1
−1
˜
Y
2
−1
2
−2
˜
Y
3
∗
−1
3
−3
˜
Y
4
−1
−1
1
˜
Y
5
∗
−2
−2
2
it doesn't hara terize any hoi e at all and we may ignore it. If we apply
the operator
τ
to the fuzzy hara teristi ve tors, solutionsY
˜
1
,Y
˜
2
andY
˜
∗
3
yield the same risp hoi e
{b}
, whereas solutionsY
˜
4
andY
˜
∗
5
yield the same risp hoi e{c}
. Note that both subsets of solutions are organized as hains ofmoreandmorelogi allydeterminedsolutions,i.e. omingin reasingly loseto the values
−5
and+5
that represent the maximal degree of falseness and truth, respe tively. In ea hgroup of solutions,we shall fo us on the maximalsolutions,
Y
˜
∗
3
andY
˜
∗
all an
L
-fun tion anyL
-valued fun tion dened onX
. We mainly onsiderL
-fun tionsY
˜
su h thatY
˜
(a) 6= 0
, for alla
∈ X
; the set of su hL
-fun tions is denoted byL
0
.Denition 6 (Sharpness). Let
Y , ˜
˜
Z
be two fun tions inL
0
. We say thatY
˜
is atleast as sharp asZ
˜
,whi h wedenote byZ 4 ˜
˜
Y
,i, for alla
∈ X
, either˜
Y
(a) ≤ ˜
Z(a) < 0
or0 < ˜
Z(a) ≤ ˜
Y
(a)
. The asymmetri part≺
of4
is dened as follows:a
≺ b
ia 4 b
and notb 4 a
.The sharpness relation
≺
isapartial order(asymmetri and transitive rela-tion) onthe set ofL
-fun tionsinL
0
.To ea h fun tion
Y
˜
inL
, we asso iate a risp subset of X, whi h we denote byK
Y
˜
and dene by:K
Y
˜
= τ ˜
Y
= {a ∈ X
su h thatY
˜
(a) > 0}.
(3.12)We all
K
˜
Y
the stri t median ut hoi e asso iated withK
˜
Y
.Forany
L
-fun tionY
˜
∈ L
0
, thedegree of sharpnessQ
sharp
( ˜
Y
)
of
Y
˜
isdened as:Q
sharp
( ˜
Y
) = min( min
a
∈K
Y
˜
˜
Y
(a), min
a
6∈K
Y
˜
− ˜
Y
(a)).
(3.13)This index interprets asa degree of sharpness of
Y
˜
sin e we have˜
Y 4 ˜
Z
⇒
Q
sharp
( ˜
Y
) ≤ Q
sharp
( ˜
Z).
(3.14)
Of ourse,wemayhave
Q
sharp
( ˜
Y
) ≤ Q
sharp
( ˜
Z)
andneither
Y 4 ˜
˜
Z
norZ 4 ˜
˜
Y
. Denition 7 (L
-valueddominantand absorbent kernels).We all
L
-dominant (resp.L
-absorbent)kernel aL
-valued hara teristi ve -torY
˜
that satises the followingthree onditions:• ˜
Y
is a solutionof equation system (3.7) (resp. (3.8));•
itismaximal with respe t tothe sharpness relation4
inthe setY
dom
(G
L
)
(resp.
Y
abs
(G
L
)
)ofadmissiblesolutionsof(3.7) (resp.(3.8)),i.e.thereisno othersolutionZ
˜
su h thatY
˜
≺ ˜
Z
;•
itbelongs toL
0
, i.e.Y
˜
(a) 6= 0
, for alla
∈ X
We denote by
F
dom
(G
L
)
(resp.F
abs
(G
L
)
) the (possibly empty) set ofL
-dominant(L
-absorbent) kernels ofG
L
.
Revisitingexample4,wehave
Y
˜
∗
3
andY
˜
∗
5
(seeTable 2)asL
-dominantkernels ofG
L
; defuzzyfying
Y
˜
∗
3
using the operatorτ
yieldsτ ˜
Y
∗
3
= {b}
. In a similar way, we getτ ˜
Y
∗
systems yields the solutions
[Y (a) = 2, Y (b) = −1, Y (c) = −2]
and[Y (a) =
−4, Y (b) = −4, Y (c) = 4]
as theL
-absorbent kernels inG
L
.
In example 3, the digraph
G
L
admits the set
F
i−dom
(G
L
) = { ˜
Y
1
∗
, ˜
Y
2
∗
}
ofL
-dominant kernels and the setF
bad
(G
L
) = { ˜
Y
∗
3
, ˜
Y
4
∗
}
ofL
-absorbent kernels (see table 3). Applying the operatorτ
to theseL
-valued kernels yields the hoi e sets that were already obtained for example 3in se tion3.2.Table 3
L
-valued kernelsin the digraphof Example 3kernel
˜
Y
(a)
Y
˜
(b)
Y
˜
(c)
L
-dominant˜
Y
1
∗
−1
5
−5
˜
Y
2
∗
−1
−1
1
L
-absorbent˜
Y
3
∗
1
−1
−1
˜
Y
4
∗
−5
−5
5
Fromthelatterexample,itappearsthat
L
-valuedkernelsandpotentiallygood or potentially bad hoi es of a valued graph are strongly related.In the nextse tion,we show the formal orresponden e between those notions.
4 Relating
L
-valued kernels and potentially good and bad hoi esInthisse tion,weshallonlydealwiththe dominant ase;our resultsextend
mutatismutandis totheabsorbent ase.Itiseasytoestablishalinkbetween
asubsetofsolutionsofthedominantkernelequationsystem(3.7)andasubset
of dominantkernels of
τ G
that we dene below.Denition8(Determinedsolutionsofthedominantkernelequationsystem).
A solution
Y
˜
of the dominantkernel equationsystem (3.7)is said determined if it belongs toL
0
i.e.Y
˜
(a) 6= 0
, for alla
∈ X
. We denote byY
dom
0
(G
L
)
the setY
dom
(G
L
) ∩ L
0
of determined solutionsof (3.7).Denition 9 (
R
-determined subsets). A subsetK
ofX
inthe valued graphG
L
= (X, R)
isR
-determined ifR(a, b) 6= 0
for alla, b
∈ K
.Lemma 1. If
Y
˜
belongs toL
0
and is a solution of the dominant kernel equation system (3.7), thenτ ˜
Y
= K
Y
˜
is aR
-determined dominant kernel ofProof. Using the fa t that
Y
˜
is a solution of the dominant kernel equation system, we applyτ
toboth sides of (3.7),yielding−τ ˜
Y
(a) = max
b6=a
[min(τ ˜
Y
(b), τ R(b, a))],
whi h is exa tly (2.2) with
Y
= τ ˜
Y
. It remains to be proven thatK
˜
Y
isR
-determined. Assume that there area, x
∈ K
Y
˜
withR(a, x) = 0
. Sin eY
˜
belongs toL
0
by hypothesis,Y
˜
(a), ˜
Y
(x) > 0
. Usingthe negative form (3.9) of the kernel equationsystem, wehave for that parti ulara
∈ K
˜
Y
:0 < ˜
Y
(a) ≤ max(− ˜
Y
(x), −R(x, a)) = 0,
a ontradi tion. 4.1 A useful transformation Theelementsof
Y
dom
0
(G
L
)
i.e.thedeterminedsolutionsofthedominantkernel equation system an be viewed as the xpoints of a transformationT
that operates onL
0
.Denition 10 (The transformation
T
). The transformationT
maps a fun -tionU
˜
∈ L
0
, ontoT ˜
U
dened by:T ˜
U
(a) = min
x
6=a
max(− ˜
U
(x), −R(x, a)),
∀a ∈ X.
The xpoints of
T
inL
0
are the solutions of (3.9) or, equivalently, of (3.7), sin eT ˜
U
= ˜
U
⇔
U
˜
(a) = min
x
6=a
max(− ˜
U
(x), −R(x, a)),
∀a ∈ X.
Lemma 2. If
K
is aR
-determined dominant kernel ofτ G
andU
˜
∈ L
0
is su h thatK
U
˜
= K
, thenT ˜
U
∈ L
0
andK
T ˜
U
= K
. Moreover, for alla
∈ K
U
˜
, we may omputeU
˜
(a) as follows:T ˜
U
(a) =
min
x
6=a; x∈K
U
˜
−R(x, a)
!
∧
min
x
6∈K
U
˜
max(− ˜
U
(x), −R(x, a))
!
(4.1)and for all
a
6∈ K
U
˜
, wehave:T ˜
U
(a) = min
x
∈K
U
˜
max(− ˜
U
(x), −R(x, a))
(4.2)•
forx
∈ K
,R(x, a) < 0
and thus, forx
∈ K
,max(¬ ˜
U
(x), −R(x, a)) =
−R(x, a)
(this alsoproves that(4.1) is equivalentto (3.9),fora
∈ K
);•
forallx
6∈ K
,max(− ˜
U
(x), −R(x, a)) ≥ − ˜
U
(x) > 0
.If
a
6∈ K
,T ˜
U
(a) < 0
be ause there isx
∈ K
su h thatR(x, a) > 0
; for su h anx
,max(− ˜
U
(x), −R(x, a)) < 0
and hen eU
˜
(a) < 0
.Moreover, for allx
not inK
,max(− ˜
U
(x), −R(x, a)) > 0
, hen e (4.2) isvalid.Corollary 2. Let
Y
˜
∈ L
0
. The fun tionY
˜
is a solution of equation system (3.9) (or equivalently of (3.7)),iY
˜
satises˜
Y
(a) =
min
x
6=a; x∈K
Y
˜
−R(x, a)
!
∧
min
x
6∈K
Y
˜
max(− ˜
Y
(x), −R(x, a))
!
,
(4.3) for alla
∈ K
˜
Y
, and˜
Y
(a) = min
x
∈K
Y
˜
max(− ˜
Y
(x), −R(x, a)),
(4.4) for alla
6∈ K
˜
Y
. Proof.[⇒]
Lemma1 indi ates thatτ ˜
Y
= K
˜
Y
is aR
-determineddominant kernel ofτ G
; thereforeY
˜
fulllsthe hypotheses of lemma2 yielding the result.[⇐]
WeprovethatK
˜
Y
isaR
-determineddominantkernelofτ G
.Inviewof (4.3)andthefa tthatY
˜
(a) > 0
whenevera
∈ K
˜
Y
,wededu ethatR(x, a) < 0
for alla, x
∈ K
˜
Y
. Similarly, for alla
6∈ K
˜
Y
, we dedu e fromY
˜
(a) < 0
and equation (4.4) that there exists at least onex
inK
˜
Y
su h thatR(x, a) > 0
. Hen eforth,K
˜
Y
isaR
-determineddominantkernelofτ G
.Usingthefa tthatK
Y
˜
is a kernel ofτ G
, lemma 2 yields thatT ˜
Y
an be omputed using (4.1) and (4.2); by hypothesis,Y
˜
satises (4.3) and (4.4); thereforeY
˜
is a xpoint ofT
,hen eY
˜
belongs toY
dom
0
(G
L
)
.4.2 Obtaining a
L
-dominant kernel from aR
-determined dominant kernel ofτ G
Weusetransformation
T
toobtainaL
-dominantkernelfromaR
-determined dominant kernel ofτ G
. The main result we need is the fa t thatT
respe ts the relation at least as sharp as on any subset ofL
-fun tionsY
˜
su h thatτ ˜
Y
is aR
-determined dominantkernel.Lemma 3. Let
K
beaR
-determineddominantkernel ofτ G
andU , ˜
˜
V
∈ L
0
be su hthatK
˜
Proof. Forany
a
∈ K
,T ˜
V
(a) =
min
x
6=a; x∈K
V
˜
−R(x, a)
!
∧
min
x
6∈K
V
˜
max(− ˜
V
(x), −R(x, a))
!
.
Sin e the fun tions
min
andmax
are nonde reasing and for allx
6∈ K
,− ˜
V
(x) ≥ − ˜
U
(x)
, wegetT ˜
V
(a) ≥ T ˜
U
(a)
.Forany
a
6∈ K
,T ˜
V
(a) = min
x∈K
˜
V
max(− ˜
V
(x), −R(x, a))
.Sin ethe fun tions
min
andmax
are nonde reasing and for allx
∈ K
,− ˜
V
(x) ≤ − ˜
U
(x)
, we getT ˜
V
(a) ≤ T ˜
U
(a)
.Startingfromthe hara teristi fun tion of a
R
-determineddominant kernel ofτ G
and applying iterativelyT
to this fun tion, one obtains, after a nite number of iterations, anelement ofY
dom
0
(G
L
)
that is at least as sharp asany other elementY
˜
ofY
dom
0
(G
L
)
su h thatK
˜
Y
= K
.Proposition3.Let
K
beaR
-determineddominantkernelofτ G
andU
˜
∈ L
0
be dened byU
˜
(a) = m
ifa
∈ K
and−m
otherwise. For some nite integern
,T
n
˜
U
= T
n+1
U
˜
andZ
˜
= T
n
˜
U
isanelementofY
dom
0
(G
L
)
su hthatK
Z
˜
= K
; we have furthermoreY 4 ˜
˜
Z
for all elementY
˜
ofY
dom
0
(G
L
)
withK
˜
Y
= K
.Proof. The fun tion
U
˜
dened in the proposition belongs toL
0
. Sin e, by lemma 2,K
T ˜
U
= K
U
˜
, it is obvious from the denition ofU
˜
thatT ˜
U 4 ˜
U
. Applying lemma 3, we getT
i+1
U 4 T
˜
i
˜
U
, for alli
. Due to the niteness ofL
andX
, there is a nite numbern
su h thatT
n+1
U
˜
is not stri tly less sharp
than
T
n
˜
U
, whi h means thatT
n+1
U
˜
= T
n
˜
U
. For su h an
,Z
˜
= T
n
˜
U
is an element ofY
dom
0
(G
L
)
. It is moreover a maximal element w.r.t. the relation4
in the subset of elementsY
˜
ofY
dom
0
(G
L
)
su h thatK
˜
Y
= K
. LetY
˜
be su h an element. We haveT ˜
Y
= ˜
Y
andY 4 ˜
˜
U
. Applying lemma 3, we getT
n
˜
Y 4 T
n
˜
U
, whi h means thatY 4 ˜
˜
Z
.It was not evident atrst sight that, for all
R
-determined dominant kernelsK
ofτ G
,therearesolutionsY
˜
of(3.9)withK
Y
˜
= K
.Itwasnotevidenteither that the set of solutionsY
˜
of (3.9) withK
˜
Y
= K
, when non-empty, admits a uniquemaximallysharpelement.Puttingtogetherlemma1andproposition3,weestablishtheexisten eofabije tive orresponden ebetween
R
-determined dominantkernels ofτ G
and determinedL
-dominantkernels ofG
L
.
Theorem 1.
K
is aR
-determined dominant kernel ofτ G
i there is a de-terminedL
-dominantkernelY
˜
su hthatK
˜
There is a link between the quali ation
Q
i-dom(K)
of a dominant kernelK
ofτ G
and the sharpness quali ation of all elementsY
˜
ofY
dom
0
(G
L
)
su h thatK
Y
˜
= K
. We rst prove that the degree of sharpness of an element ofY
dom
0
(G
L
)
isnotlargerthanthequali ationofthe orrespondingR
-determined dominantkernel.Lemma 4. If
K
is aR
-determined dominant kernel ofτ G
, any determined solutionY
˜
of thekernelequationsystem (3.7)issu hthatthe hoi eK
˜
Y
= K
in
τ G
veries:Q
sharp
( ˜
Y
) ≤ Q
i-dom(K)
(4.5)Proof. Assumethat
Q
sharp
( ˜
Y
) > Q
i-dom(K)
.LetQ
sharp
( ˜
Y
) = r
andQ
i-dom(K) =
k
,withr > k >
0
. Thequali ationQ
i-dom
(K) = k
isequal eitherto−R(b, a)
for somea, b
∈ K
or tomax
x∈K
R(x, y)
, for somey
6∈ K
.In the former ase, using (4.3), weget
r
≤ ˜
Y
(a) ≤ −R(b, a) = k
,a ontradi -tion.In the latter ase, there is
a
∈ K
su h thatR(a, y) = max
x
∈K
R(x, y) =
k
. Therefore, for allx
∈ K
,R
(x, y) ≤ k
. Using (4.4), we haveY
˜
(y) =
min
x
∈K
Y
˜
max(− ˜
Y
(x), −R(x, y))
.Forall
x
∈ K
,− ˜
Y
(x) ≤ −r ≤ −k ≤ −R(x, y)
††
;
therefore
Y
˜
(y) ≥ −k
. But, by denition ofQ
sharp
( ˜
Y
)
,
Y
˜
(y) ≤ −r
hen e−r ≥ ˜
Y
(y) ≥ −k
,a ontradi tion.Wenallyprovethatthe sharpnessofthe
L
-dominantkernelY
˜
orresponding toaR
-determined dominant kernelK
is equal to the quali ation ofK
. Theorem 2. For anyR
-determined dominant kernelK
, theL
-dominant kernelZ
˜
orrespondingtoK
is su h thatQ
sharp
( ˜
Z
) = Q
i-dom(K)
.Proof. In view of lemma 4, it su es to prove that
Q
sharp
( ˜
Y
) ≥ Q
i-dom(K)
. LetQ
i-dom
(K) = k
, withk >
0
.Dene theL
-fun tionW
˜
∈ L
0
as˜
W
(a) =
k
ifa
∈ K
−k
ifa
6∈ K.
Bylemma 2,
T ˜
W
belongstoL
0
andK
T ˜
W
= K
W
˜
= K
. Using(4.1) and (4.2),††
we showthat
W 4 T ˜
˜
W
. Indeed, fora
∈ K
,T ˜
W
(a) =
min
x
6=a;x∈K
W
˜
−R(x, a)
!
∧
min
x
6∈K
W
˜
max(− ˜
W
(x), −R(x, a)) ≥ k
!
,
and, fora
6∈ K
,T ˜
W
(a) = min
x
∈K
W
˜
max(− ˜
W
(x), −R(x, a))
= max(−k, min
x
∈K
W
˜
−R(x, a))
≤ −k.
Rememberthedenitionof
U
˜
∈ L
0
that wasused inproposition3:U
˜
(a) = m
ifa
∈ K
and−m
otherwise.WehaveW 4 ˜
˜
U
and,usinglemma3,T
i
˜
W 4 T
i
˜
U
, for alli
= 1, 2, . . .
. Sin e, for somen
,T
n
˜
U
= ˜
Z
, theL
-dominant kernel orresponding toK
, we have:˜
W 4 T
n
W 4 T
˜
n
U
˜
= ˜
Z
and, using (3.14),Q
sharp
( ˜
W
) = k ≤ Q
sharp
( ˜
Z
)
. 4.4 RemarksThe above results promptthe followingremarks.
(1) Theundetermined value
0
,whensu havalue appearsinthe setL
,plays aspe ialrole andhas tobe arefullydealtwith.Inlemma2forinstan e,if we either drop the restri tion that the dominant kernel
K
ofτ G
is aR
-determined one or thatU
˜
(x) 6= 0
for allx
∈ X
, we an no longer make sure thatT ˜
U
(a) > 0
for alla
∈ K
. This means that the trans-formed hara teristi fun tionT ˜
U
would not satisfyK
T ˜
U
= K
U
˜
. Sin e we are interested in determined solutions of the kernel equation system,wedidnotpay attentiontoundetermined solutions,and orrelatively,we
onsidered onlythe
R
-determined dominant hoi es ofτ G
.(2) Throughoutthis paper,wehaveassumed thatboththeset
X
andthesetof the real numbers (for instan e the interval
[−1, 1]
), the onstru tion of the xpoint asin proposition 3 stillworks but the onvergen e mighttake aninnite number of steps.
5 Con lusion
This paper aimsat exploringthe relationshipbetween independent dominant
(resp. absorbent) hoi es and solutions of the kernel equation system in the
ase of valued relations.Wehaveshown thatthere is aone-to-one
orrespon-den ebetween
L
-dominant(resp.L
-absorbent)kernels ofa valued graphand the (R
-determined) independent dominant (resp. absorbent) hoi es of the stri t median ut digraphτ G
. Furthermore, starting with an independent dominant (resp. absorbent) hoi e ofτ G
, the te hnique of proof provides a way of building the orrespondingL
-dominant (resp.L
-absorbent) kernel as the xpointofatransformation.This has animportant onsequen efromthealgorithmi point of view: nding
L
-valued kernels is not more di ult than nding the kernels of a risp graph. Indeed, starting with a (R
-determined) kernel ofτ G
and applying the transformation (denition 10) at mostm|X|
times yields the xpoint.Of ourse,this does not makethe determinationofthe
L
-valued kernels an easy task sin e we know that nding all kernels of a risp relation is a omputationally hard problem (this was proven by V.Chvátal; see Garey and Johnson(1979), p.204).
Theorem2alsohasinterestinitsown:thedegreeofsharpnessofa
L
-dominant kernelY
˜
isequal tothe degreeof potentialgoodnessQ
i−dom
of the asso iated
risp kernel
K
˜
Y
.The omparison of risp kernels (or equivalently of indepen-dentdominantorabsorbent hoi es ofτ G
)done attheend ofse tion4about examples3 and 4 an thus be transposed toL
-valuedkernels.Referen es
BangJensenJ.andGutin,G.,Digraphs:Theory,algorithmsandappli ations.
SpringerVerlag, London, 2001.
Berge, C., Théorie des graphes et ses appli ations.Dunod,Paris, 1958.
Berge, C., Graphes et hypergraphes. Dunod,Paris, 1970.
Bisdor,R.andRoubens,M.,Ondening fuzzykernelsfrom
L
-valuedsimple graphs.In:Pro eedings Information Pro essing and Management ofUn er-tainty,IPMU'96,Granada, July1996, 593599.
Bisdor,R. and Roubens, M.,On dening and omputing fuzzy kernels from
S ienti Publishers, Singapoure,1996, 113123.
Bisdor,R., On omputing kernelsfrom
L
-valuedsimple graphs.In: Pro eed-ings 5th European Congress on Intelligent Te hniques and Soft ComputingEUFIT'97, (Aa hen, September 1997),vol. 1,97103.
Bisdor,R.,Logi alfoundationoffuzzy preferentialsystemswith appli ation
tothe Ele tre de isionaid methods, Computers & Operations Resear h,
27,2000, 673687.
Bisdor, R., Ele tre-like lustering from a pairwise fuzzy proximity index.
European Journal of Operational Resear h, 138/2, 2002,320331.
Bisdor, R. and Roubens, M., Le temps des noyaux .... In: Bair, J. and
Henry, V. (ed.) Regards roisés sur les méthodes quantitatives de gestion.
LesÉditionsde l'Université de Liège, Liège (Belgium),2004, pp. 4154.
Garey,M.R. and Johnson, D.S., Computers and intra tability, A guide to the
theory of NP-Completeness. Freemanand Co., New York, 1979.
Haynes, T.W.,Hedetniemi,S.T.andSlater,P.J.,Fundamentalsofdomination
in graphs.Mar el Dekker In . New-York, 1998.
Kitainik,L.,Fuzzyde ision pro edures withbinary relations:towards aunied
theory. KluwerA ademi Publ.,Boston, 1993.
Riguet, J., Relations binaires, fermetures, orrespondan es de Galois. Bull.
So . Math. Fran e, 76, 1948,114155.
Roy,B.,Classementet hoixen présen ede pointsde vuemultiples.RIRO 8,
5775.
Roy, B., Algèbre moderne et théorie des graphes. Dunod, Paris, 1969.
Roy, B., Méthodologie Multi ritère d'aide à la dé ision. E onomi a, Paris,
1985.
Roy, B. and Bouyssou, D., Aide multi ritère à la dé ision: Méhodes et as.
E onomi a,Paris, 1993.
S hmidt,G. and Ströhlein,Th., On kernels ofgraphs and solutionsof games:
a synopsis based on relations and xpoints. SIAM, J. Algebrai Dis rete
Methods, 6, 1985, 5465.
S hmidt, G. and Ströhlein, Th., Relationen und Graphen. Springer-Verlag,
Berlin,1989.Englishversion:Relationsandgraphs,SpringerVerlag,Berlin,
1993.
vonNeumann, J.and Morgenstern,O., Theory of gamesand e onomi