A discussion of the applications of fuzzy sets to game theory

by
Shane Michael Murphy
A Creative Component submitted to my graduate committee
in partial fulllment of the requirements for the degree of
MASTER OF SCIENCE
Major: Mathematics
Program of Study Committee:
Roger Maddux, Major Professor
Cliord Bergman
Leigh Tesfatsion
Iowa State University
Ames, Iowa
2006
Copyright c _Shane Michael Murphy, 2006. All rights reserved.
ii
Graduate College
Iowa State University
This is to certify that the masters Creative Component of
Shane Michael Murphy
has met the Creative Component requirements of Iowa State University
Major Professor
For the Major Program
iii
TABLE OF CONTENTS
ABSTRACT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vi
Preface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1 Vagueness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
2 Fuzzy Sets and Fuzzy Logic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.1 Universal Algebra . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.2 Introduction to Fuzzy Set Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.3 fuzzy logic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.3.1 Types of fuzzy sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
3 Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
3.1 Thinking with fuzzy sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
3.2 Dierences and similarities to probability . . . . . . . . . . . . . . . . . . . . . . . . 16
3.2.1 Fuzzy Measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
3.2.2 possibility theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
3.3 Fuzzy Relation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
3.3.1 fuzzy preference relation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
4 Game Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
4.1 crisp games . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
4.1.1 Bach or Stravinsky (Battle of the Sexes) . . . . . . . . . . . . . . . . . . . . . 25
4.1.2 Matching Pennies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
4.1.3 Prisoners Dilemma . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
4.2 fuzzy games . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
4.2.1 Butnarius Game . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
iv
4.2.2 Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
ACKNOWLEDGEMENTS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
v
LIST OF TABLES
2.1 operations , , , and . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.2 NOT operation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.3 OR operation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.4 AND operation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.5 implication . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.6 NOT operation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.7 OR operation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.8 AND operation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.9 implication . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
4.1 Bach or Stravinsky . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
4.2 Matching Pennies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
4.3 Matching Pennies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
vi
ABSTRACT
Fuzzy Game Theory is a growing eld in mathematics, economics, and computer science. In this
paper, we follow the thread of fuzzy logic to the doorstep of social sciences and game theory. We then
examine how fuzzy sets have been applied equilibrium theory of games and discuss some cases where
Nash equilibria can be show to exist, and ultimately a situation where a Nash equilibrium does not arise
unless the game is one of perfect information.
1
Preface
The pure mathematician knows that pure mathematics has an
end in itself which is more allied with philosophy.
Philip Jourdain, in the introduction to Georg Cantors
Contributions to the Founding of the Theory of Transnite Numbers
I am interested in the mathematics which can describe interactions between humans and groups of
humans in the political, sociological, and economics spheres (or social sciences in general). In con-
temporary political science, game theory (called rational choice theory) is king. The assumptions of
rational choice models are, briey, that human interactions can be modeled by players, whose choices
can be modeled by the game, who have a preference ordering concerning the outcome of those choices,
and that players will act in a way to maximize the utility of the outcome given their behavior. All utility
functions (aka payo function) from the set of outcomes to the utilities of those outcomes which satisfy
the players preference functions are allowed (Myerson (1997)). A major strength of game theoretical
analysis of social science situations is the exibility that this model is aorded by allowing agents to
order their preferences, and by giving preferences an order, but not necessarily a particular value. Al-
though like much of mathematics, the foundations of game theory can be stretched ever further into the
past. But the major source of the eld was John von Neumann, and Morgenstern and Von Neumann
(1944) marked the arrival of the subject. It has since been a major contribution in areas from economics
(John Nash, Thomas Schelling) to ethical theories (John Rawls, Amartya Sen).
In 1965, Loft Zadeh introduced (Zadeh (1965)) the world to the term fuzzy, as a formalization of
vagueness. The eld has been applied with a good deal of success to engineering, with fuzzy control
systems able to do things like cook rice or shift gears in a car with great eciency. Since fuzzy set
2
operators extend the idea of the logical and and or to non-dichotomous systems, behavioral scientists
have asked if there is a way to use fuzzy mathematical models to analyze social systems. The book by
Smithson addresses exactly that question. Suggestions about how this could be done can be found in
the computer science eld of natural language processing (Dubois and Prade (1980)).
In modern social science, game theory is the method of making mathematical models. Fuzzy Logic
gives a new tool which on the face seems to apply to this type of modeling as well. Mares (2001) mixes
the concepts in his book discussing coalition games with fuzzy pay-os (fuzzy expected utility func-
tions), developing the ideas of fuzzy core, fuzzy balancedness, and fuzzy shapely values. Others such as
Smithson and Verkuilen (2006) and Ragin (2000) examine using fuzzy sets in linear statistical models.
Mansur (1995) looks at the application of fuzzy sets to some introductory concepts of microeconomics.
3
1 Vagueness
This, of course, is the answer to the old puzzle about the man who went bald. It is supposed that at rst he was not bald, that
he lost his hairs one by one, and that in the end he was bald; therefore, it is argued, there must have been one hair the loss of
which converted him into a bald man. This, of course, is absurd. Baldness is a vague conception; some men are certainly
bald, some are certainly not bald, while between them there are men of whom it is not true to say they must be either be bald
or not bald. The law of excluded middle is true when precise symbols are employed, but it is not true when symbols are
vague, as, in fact, all symbols are.
-B. Russell Vagueness, 1923
The ship wherein Theseus and the youth of Athens returned [from Crete] had thirty oars, and was preserved by the
Athenians down even to the time of Demetrius Phalereus, for they took away the old planks as they decayed, putting in new
and stronger timber in their place, insomuch that this ship became a standing example among the philosophers, for the
logical question of things that grow; one side holding that the ship remained the same, and the other contending that it was
not the same.
-Plutarch
What is more, there cannot be anything between two contradictories, but of any one subject, one thing must either be
asserted or denied. This is clear if we rst dene what is truth and what is falsehood. A falsity is a statement of that which is
that it is not, or of that which is not that it is; and a truth is a statement of that which is that it is, or of that which is not that
it is not. Hence, he who states of anything that it is, or that it is not, will either speak truly or speak falsely. But of what is
neither being nor nonbeing it is not said that it is or that it is not.
-Aristotle Metaphysics IV
There is a sense in which human languages are vague. Logic, however, is usually presented as being
precise. In the fourth century B.C. Socratic philosopher Eubulides of Miletus discussed the paradox of
the bald man. And Plutarch reported the paradox of the ship of Theseus as coming from Greek legend.
But classic logic, even as formed by Aristotle held that a statement must be either true or false. A man
must be bald or not, a ship must be that sailed by Thesues or a dierent ship altogether. This law of
excluded middle left us with a logic which was unable to give expression to the vaguarities of human
4
language and the vaguarities of human categorization. Passing over the middle age stoic philosophers
who largely followed in Aristotles footsteps, it is important to make mention of Italian Renaissance
philosopher of language, Lorenzo Valla (1407-1457), who felt that each hair made some dierence.
John Locke (1632-1704) and Gottfried Leibniz (1646-1716) discussed the question of what sorts of
things formed a natural kind, that is which have a boundary dened by natural law. Although the pair
disagreed on the extent to which boundaries between dierent things are formed by the human mind,
both agree that borderline cases are a matter of opinion. Alexander Bain (1818-1903) addressed the
question in his Logic of Relatives (1870), concluding that a certain margin must be allowed as indeter-
mined, and open to dierence of opinion. Williamson (1996)
In the late nineteenth and early twentieth century, the concept of vagueness was reexamined by
various scientists and philosophers. A full account exists in Williamson (1996) of the historical back-
drop of fuzzy theory. Early on, the concept of vagueness was mentioned in order to exclude it. Gottlob
Freges (1848-1925) concern that without sharp borders logic rules would be broken were included
in his Grundgesetze der Arithmetik (1893-1903) inuenced Bertrand Russell (1872-1970) as well as
Polish philosophers Tadeusz Kotarbinski (1886-1981) and Kazimierz Ajdukiewicz (1890-1963), all of
whom tried to give precise denitions of vagueness. Charles Sanders Peirce (1839-1914) was work-
ing independently of Frege gave a denition of vague in a philosophical dictionary he wrote in 1901.
British philosopher Max Black (1909-1988) based his discussions of vagueness largely on examples of
borderline cases. Kotarbinski began using gradedness of truths in writings as early as 1923. Physi-
cian Ludwik Fleck (1896-1961) applied the Poles ideas to medical diagnosis.
Karl Menger (1902-1985), formerly of the Vienna circle, generalized metric spaces toward prob-
abilistic concepts, introducing the concept of a triangular norm (T-Norm) in 1942. Menger suggested
in 1951 that an item need not be absolutely an element of a set, but rather there could be a mapping
from the element to the probability that the element is in the set. He was also inuenced in a 1951
paper by Henri Poincare (1854-1912) claim that equality is not transitive in the physical reality, as
when we say two objects in the real world are equal, we mean that they are indistinguishable, but that
if one follows a chain of indistinguishable items, one need not say that the rst and the last item remain
5
indistinguishable. Menger acknowledged the fuzzy sets of Lot Zadeh (1921-), but did not explore the
dierences between probability sets and fuzzy sets in his work. In the early 1960s, Richard E. Bell-
man (1920-1984), Robert Kalaba (1926-2004), and Zadeh began exploring the concept of fuzzy sets at
RAND Corporation, leading to a joint memorandum in 1964 and a paper in the journal Systems Theory
entitled Fuzzy Sets and Systems in 1965 (Zadeh (1965)).
Now we will depart the historical development of fuzzy sets and fuzzy logic in order that we can
build our theories from the bottom. In this thesis, we will leave open the question of whether vagueness
comes only from linguistic (or human) causes.
6
2 Fuzzy Sets and Fuzzy Logic
To understand the mathematical foundations of fuzzy sets and fuzzy logic, we will start our study
with a review of the concepts of universal algebra. This will give us a foundation from which we can
understand the generalizability of fuzzyness, before we look at specic applications. We will then look
at how universal algebra serves as a basis for traditional logic, and then show how those denitions
generalize to fuzzy logic through the concept of fuzzy sets. We will then look at some of the variations
of fuzzy sets, of which we will tend to use the most simple so that its application can be most easily
understood.
2.1 Universal Algebra
An n-ary relation on the sets A
1
, . . . , A
n
is specied by giving an ordered n-tuple (a
1
, . . . , a
n
) of
elements with each element a
i
A
i
such that the n-tuple is in the relation . In this way, the relation
is a subset of the product set A
1
A
n
. The number n is called the arity of the relation. We often say
unary instead of 1-ary, binary instead of 2-ary and tertiary instead of 3-ary. For binary relations, we may
often write a
1
a
2
while for relations of any arity may be written (a
1
, . . . , a
n
) or (a
1
, . . . , a
n
).
A binary relation, , on a set X is complete if aB or ba for every a, b X, A relation is transitive
if ab and bc implies ac for every a, b, c X. A relation is reexive if aa for every a X. a
preference relation is a complete, transitive, reexive binary relation. A preference relation, , on X is
continuous if for all k given sequences (a
k
)
i
and (b
k
)
i
in X that converge to a X and b X respectively
such that a
k
b
k
, then a b. A preference relation , on R is quasiconcave if for every b R the
set a R : a b is convex. Recall, a set is convex if given two points in the set, a line segment
connecting those points is also in the set.
A function f : A
1
A
2
is a binary relation such that for each a
1
A
1
there is exactly one
a
2
A
2
such that (a
1
, a
2
) f ), and is often written f (a
1
) = a
2
. A function f (a
1
, . . . , a
n
) of n
7
variables where a
i
A
i
for all i 1, . . . , n is a function f : A
1
A
n
B. For each a
i
A
i
for i 1, . . . , n, (a
1
a
n
) A
1
A
n
and f ((a
1
a
n
)) B. We may omit one set of
brackets from this nal expression to simplify our notation.
Denition 2.1.1. An n-ary operation on the set A is a function f : A
n
A.
Denition 2.1.2. An algebra A is a pair (A, F) with A a nonempty set and F = ( f
i
: i I) a sequence
of operations on A, where I is some index set. We call A the universe or the underlying set of A, and the
f
i
are the fundamental or basic operations of the algebra. The similarity type is a function : I
such that each i I is assigned to the arity of f
i
from the fundamental operations of the algebra.
Often, an algebra (A, F) is written (A, f
1
, . . . , f
n
), and the operations are written in descending order
of their arity. So if we denote the order of the domain of a function f as f , then f
1
f
n
. With
this notation, we can see that ( f ) = f . The similarity type is often written simply as a sequence of
arities, i.e. ( f
1
, . . . , f
n
).
Denition 2.1.3. A subalgebra of an algebra A = (A, F) is a subset B = (B, G) of A if B forms a
algebra with the operations of A restricted to the set B, i.e., if for all i I and for all g
i
G, we have
g
i
= f
i

B
(i) .
Equality between two algebras A = (A, F) and B = (B, F) requires A = B and f
A
= f
B
for all
f F. If two algebras have at least the same similarity type, they are called similar. Any intersection
of subalgebras is clearly a subalgebra. So, given any subset B of A, there is a unique smallest subalgebra
containing B, namely, the subalgebra S : S subalgebra of A, S B. This is the subset generated by
B and is denoted by (B)
T
or occasionally (B).
Denition 2.1.4. A homomorphism between algebras A = (A, F) and B = (B, G) is a function :
A B such that for all i I the index set for the operations of A, and all a
1
, . . . , a
n
A, we have:
( f
i
(a
1
, . . . , a
n
)) = g
i
((a
1
), . . . , (a
n
)) for f
i
F and g
i
G. That is to sat preserves all the
operations of A. Clearly the composition of two homomorphisms is a homomorphism. If : A B
is an invertible homomorphism, then
1
: A B is also a homomorphism. In this case we call an
isomorphism, and we say that A and B are isomorphic.
8
Denition 2.1.5. Let S be any set, let A be a algebra of type , and let : S A be a function. we
say that (A, ) (or informally just A) is a free algebra (of type ) on the set S of free generators
if, for every algebra B of type and function : S B, there exists a unique homomorphism
: B A such that = .
S A
B
-

p
p
p
p
p
p
p

Note that for any set S and any similarity type , there exists a free algebra of type generated by
S (Sometimes said to be a free algebra on S ), and that this free algebra is unique up to isomorphism.
For the following denition, let A be some algebra, while F is the algebra on the set X
n
=
x
1
, . . . , x
n
. Now for any elements a
1
, . . . , a
n
A there exists a unique homomorphism : F A
with (x
i
) = a
i
for i = 1, . . . , n. If w F, then (W) A, and w is uniquely determined by a
1
, . . . , a
n
.
Thus we can dene a function w
A
(a
1
, . . . , a
n
) := (w), (note: we may omit the subscript A). In particu-
lar, if we take A = F and a
i
= x
i
for i 1, . . . , n, then is the identity and w(x
1
, . . . , x
n
) = w.
Denition 2.1.6. A T-word in variables x
1
, . . . , x
n
is an element of the free algebra of type on the
set X
n
= x
1
, . . . , x
n
of free generators. A word in the elements a
1
, . . . , a
n
of an algebra A of type is
an element w(a
1
, . . . , a
n
) A, where w is a T-word in the variables x
1
, . . . , x
n
. A algebra variable is an
element of the free generating set of a free algebra.
Denition 2.1.7. Any algebra with a similarity type = (2, 0) is a propositional algebra. We may
generally write the algebra as (X, , F). The propositional algebra P(X) of the propositional calculus
on the set X of propositional variables is the free propositional algebra of type on X.
Here we may bridge between what we have been doing and what is easily recognizable to be logic.
In this algebra, X are all free variables. the nullary operation F is often called FALSE and may also be
represented by the integer 0. The binary operation is sometimes called implication. The operations
(NOT), (AND), (OR), and (IFF) can be dened in terms of these two operations as follows:
This determines the form of our algebra of propositions. In ordinary usage, we are interested in
the truth or falsity of a statement, and rst the truth and falsity of elements of the propositional algebra
9
p p F
p q (p) q
p q (p q)
p q (p q) (q p).
Table 2.1 operations , , , and
P(X). In traditional two-valued logic, we may consider functions (called valuations) which assign to
each p P(X) one of two values, 1, or 0 (TRUE or FALSE).
Denition 2.1.8. A valuation of P(X) is a proposition algebra homomorphism v : P(X) 0, 1.
Thus we may say that p P(X) is true with respect to v if v(p) = 1 and that p P(X) is false with
respect to v if v(p) = 0.
2.2 Introduction to Fuzzy Set Theory
Conceptually we can think of a set as being fuzzy when its elements belong only partly to it. Thus a
fuzzy set A is given by rst specifying a universe of elements X which are to be discussed and a scale of
truth degrees L and a rule which associates with each element x of X a value l from L which represents
the degree to which x belongs to the fuzzy set A.
So which of these is the fuzzy set? None, in fact, but lets begin again. First we need to dene a
residuated lattice:
Denition 2.2.1. A residuated lattice is an algebra L = (L, , , , , 0, 1, ) of type (2, 2, 2, 2, 0, 0, )
where
i. 0 is the least element and 1 is the greatest element,
ii. (L, , 1, ) is a commutative monoid, is associative, commutative, and the identity x 1 = x
holds.
iii. the adjointness property holds, i.e. x y z x y z holds x, y, z L with the lattice
ordering in L.
We will later call a t-norm and implication.
10
A residuated lattice is called complete if L = (L, , , 0, 1, ) is a complete lattice, that is if every
subset S of L has both a greatest lower bound and a least upper bound. These are denoted by:

S
(meet) and

S (join).
Taking a big step ahead, a fuzzy set is a mapping A : X L for X a set and L a residuated
lattice. A complete residuated lattice with L = 0, 1 is the same propositional algebra as before. In
using a generalized residuated lattice, operations like meet (AND), join (OR), and implication can be
generalized. This generalization allows us to talk about fuzzy logic, which will be introduced in the
following section. We will later mention types of fuzzy sets to discuss an extension of the concept
where the underlying set of the residuated lattice is itself fuzzy. Another way of denoting a fuzzy set is
as a pair, one from the universal set, and one from the residuated lattice. In some cases, we may also
talk about a fuzzy set based upon some crisp set, X, such that for each subset of X, we want to dene a
degree of belonging. We will denote this by, [(X) (/(X), L), where /(X) is the powerset of X.
2.3 fuzzy logic
The subject of fuzzy logic generalizes traditional logic. In depth reviews of the algebraic basis of
fuzzy logic include Gerla (2001), Cignoli et al. (1999), Nguyen (1999), and Hjek (2001) In all forms
of logic, our main objective is often to discover whether or not a statement (or formula) is true.
In classic logic, we take the evaluation of a formula of 0 to mean that the formula is false, while
an evaluation of 1 means that the formula is true. We then evaluate our connectives , , , and
according to the following tables:

0 1
1 0
Table 2.2 NOT operation
0 1
0 0 1
1 1 1
Table 2.3 OR operation
0 1
0 0 0
1 0 1
Table 2.4 AND operation
0 1
0 1 0
1 1 1
Table 2.5 implication
To talk about a logic, it is important to talk about the connectives between the elements. We can
compare fuzzy logic to classical logic if we consider t-norms, t-conorms, and fuzzy implication to be
11
analogous to AND, OR and classic implication. For simplicity, we will use the set [0, 1] with the usual
lattice ordering as the image of our fuzzy mappings rather than a more general residuated lattice.
Denition 2.3.1. A T-norm is a binary operation T : [0, 1] [0, 1] [0, 1] with the following proper-
ties:
i. Commutativity: T(a, b) = T(b, a)
ii. Monotonicity: T(a, b) T(c, d) if a c and b d
iii. Associativity: T(a, T(b, c)) = T(T(a, b), c)
iv. 0 is the null element: T(a, 0) = 0
v. 1 is the identity element: T(a, 1) = a
Denition 2.3.2. A t-norm is called Archimedean if 0 and 1 are its only idempotent elements.
Denition 2.3.3. An Archimedean t-norm is called strict if 0 is its only nilpotent element.
Notice that a T-norm generalizes a triangle norm in a metric space in that T(a, b) T(a, 1) +T(b, 1)
for all a, b [0, 1].
It can be seen how the T-norm conforms to our idea of an AND operator, since monotonicity
requires a conjointer with a less true proposition to be less true. Associativity and commutativity
are included, and the last two properties state that truth values 0 and 1 correspond to false and true,
respectively.
Just as OR in traditional logic is in some sense dual to AND, T-conorms are dual to T-norms.
Denition 2.3.4. A T-conorm can be dened by (a, b) = T(a, b)
This, of course, generalizes De Morgans laws.
It follows that a T-conorm satises dual properties to T-norms, namely:
i. Commutativity: (a, b) = (b, a)
ii. Monotonicity: (a, b) ; (c, d) if a ; c and b ; d;
iii. Associativity: (a, (b, c)) = ((a, b), c);
12
iv. Null element: (a, 1) = 1
v. Identity element: (a, 0) = a
The T-conorm is used to represent intersection in fuzzy set theory.
There are dierent fuzzy logics that have been developed using dierent T-norms and T-conorms.
The following T-norms and T-conorms are often used:

min
(a, b) = mina, b
max
(a, b) = maxa, b

Luka
(a, b) = max0, a + b 1
Luka
(a, b) = mina + b, 1

prod
(a, b) = a b
sum
(a, b) = a + b a b

1
(a, b) =

a, if b = 1
b, if a = 1
0, else

1
(a, b) =

a, if b = 0
b, if a = 0
1, else
The rst T-norm and T-conorm are used most often, as they are simple and have some special
properties (see below). The third T-norm and the corresponding T-conorm derive from probability
theory.
Furthermore, the following relationships hold for any T-norm:

1
(a, b) (a, b)
min
(a, b)

max
(a, b) (a, b)
1
(a, b).
In other words, every T-norm lies between the drastic T-
norm (Tsub-1/sub) and the minimum T-norm (Tsubmin/sub). Conversely, every T-conorm lies
between maximum T-conorm and the drastic T-conorm.
We have discussed T-norms corresponding to AND, and T-conorms corresponding to OR. Fuzzy
implication and fuzzy negation can be discussed in terms of T-norms, and will complete the connectives
necessary to make statements in fuzzy logic.
Theorem 2.3.5. For any continuous t-norm, there is a unique operation x y such that for all x, y, z
[0, 1], we have T(x, z) y z (x y). This operation is called the residuum, and is dened
by(x y) = maxzT(x, z) y.
The proof may be found in Hjek (2001)
13
Denition 2.3.7. A fuzzy implication is a map, : [0, 1] [0, 1] [0, 1] such that
0 1
0 1 0
1 1 1
This residuum operator is the same as the implication operator in a residuated lattice.
Denition 2.3.6. A negation is a nonincresing function, , such that (0) = 1 and (1) = 0. A negation
can be generated by a t-norm quite simply. Given a t-norm, T, y : T(x, y) = 0 is a negation.
As usual, implication can be generated from an OR argument. Here, for a T-conorm, a negation,
x y ((x), y).
So as in classic logic, fuzzy logic uses a free propositional algebra, but uses these connectives rather
than the classic connectives.
Denition 2.3.8. Afuzzy valuation of P(X) is a proposition algebra homomorphismv : P(X) [0, 1].
Thus we may say that p P(X) is true with respect to v if v(p) = 1 and that p P(X) is false with
respect to v if v(p) = 0.
Another interesting non-classic logic is three-valued Lukasiewicz logic. For three-valued Lukasiewicz
logic, the set F of formulas is the same as in classical two-valued logic, however the truth evaluations
are dierent, mapping into 0, u, 1 instead of 0, 1

0 1
u u
1 0
Table 2.6 NOT operation
0 u 1
0 0 u 1
u u u 1
1 1 1 1
Table 2.7 OR operation
0 u 1
0 0 0 0
u 0 u u
1 0 u 1
Table 2.8 AND operation
0 u 1
0 1 1 1
u 1 1 u
1 0 u 1
Table 2.9 implication
Fuzzy logic uses a residuated lattice for its truth set, but usually the closed interval [0, 1] is used in
particular.
14
In both of these the law of excluded middle fails, and thus fuzzy logic is not a Boolean Algebra.
However fuzzy logic is a generalization of three-valued Lukasiewicz logic. In fact, we may state
the following theorem about the similarity between the two.
Theorem 2.3.9. The propositional calculus for three-valued Lukasiewicz logic and the propositional
calculus for fuzzy logic are the same.
Proof. For simplicity we will use [0, 1] for the truth degree set of fuzzy logic.
Truth evaluations are mappings A from F the set of formulas into the set of truth values satisfying:
A(u v) = A(u) A(v), A(u v) = A(u) A(v), A(v

) = A(v)

, u, v F.
Two formulas are equivalent if and only if they have the same values for all truth valuations. So we
need that two formulas have the same value for all truth valuations into [0, 1] if and only if they have
the same values for all truth valuations into 0, u, 1.
Let :=
x(0,1)
0, u, 1 the Cartesian product with , ,

dened componentwise. If two truth

valuations from F into dier on an element, then these functions followed by the projection of into
one of the copies of 0, u, 1 dier on an element. Likewise if two valuations from F into 0, u, 1 dier
on an element, then these two functions followed by any lattice embedding of 0, u, 1 into [0, 1] dier
on that element. There is a lattice embedding [0, 1] given by y y
x

x
where
y
x
=

0 if y < x
u if y = x
1 if y > x
If two truth valuations from F into [0, 1] dier on an element, then these two functions followed by
this embedding of [0, 1] into will dier on that element. Thus the truth values dened by the lat-
tices 0, u, 1, [0, 1], and all induce the same equivalence relation on F, and hence yield the same
propositional calculus.
The result of this is that if you want to check whether two expressions connecting n fuzzy sets with
, ,

are equivalent in fuzzy set theory, you only need to check equality between the expressions in
Lukasiewicz logic, a rather pedestrian 3
n
calculations at worst.
15
2.3.1 Types of fuzzy sets
Starting from our original denition of a fuzzy set as a mapping A : X L with L a residuated
lattice, L = (L, , , 0, 1, ), we may let L be a fuzzy set. That is, we can dene the membership of each
item from X in our fuzzy set A as a degree to which it takes a certain value. This is what is called a type
2 fuzzy set. A type 3 or type 4 fuzzy set may be dened by following this procedure further. A type 2
fuzzy set can be useful in mixing, for instance, probability theory with an arbitrary fuzzy set, where for
some element x X, we might say that the probability to which x is an element of A is .5 with is very
likely, say to a degree of .8. That is to say that if you think that there are an equal number of red and
blue balls in a bag, but you are not absolutely sure, you can say dene two functions to describe the
situation, one associating to each ball the probability that it will be red, and another the likelihood that
that probability will be correct. Up to this point when we have discussed fuzzy sets, we have meant
fuzzy sets of the rst type. We will continue with this convention in the rest of this thesis.
16
3 Applications
3.1 Thinking with fuzzy sets
Fuzzy sets and fuzzy logic is appealing when one steps back from the bivalent tenancies of mathe-
matical ideas to the world of blurred boundaries that exists in many of the sciences and social sciences.
Any denition of an object that depends on a description using linguistic hedges, their is an opening
for the type of vagueness discussed in chapter 1.
In the social science, it is often the case that theories are made with vague terms. For instance,
consider the statement that, A genocide or politicide is a sustained policy by states or their agents, or,
in civil wars, by either of the contending authorities that results in the deaths of a substantial portion
of a communal or political group. Goldstone et al. (2000) The linguistic hedges of sustained and
substantial require the reader to understand the sentence to mean that the duration of the policy and
portion of a group in the situation are only vaguely presented, but the denition may still be applicable.
Here, dening the importance of the situation in terms of a fuzzy set is appropriate. However, there
have been other ways of dealing with this situation. Let us briey discuss probability as one of these.
3.2 Dierences and similarities to probability
3.2.1 Fuzzy Measures
Before we discuss the applications of fuzzy sets and fuzzy logic to thinking in social sciences, let us
rst briey discuss fuzzy measure theory in order to tease apart one conception of fuzzy thinking from
probabilistic thinking. Fuzzy measures were rst introduced by Sugeno in 1974 (Sugeno (1974), and
have since grown into a formidable aspect of fuzzy set theory. For an in depth introduction into fuzzy
measure theory, see Wang and Klir (1993). First, let us remind ourselves about probability theory.
17
Denition 3.2.1. In probability, we consider random experiments to be experiments whose outcomes
cannot be predicted with certainty. The sample space of a random experiment is a set that includes
all possible outcomes of the experiment; the sample space plays the role of the universal set when
modeling the experiment. A probability measure (or distribution) P : [0, 1] for a random
experiment with a sample space is a real-valued function dened on the collection of events that
satises:
i. P(X) 0 for any event X with P() = 0, and P() = 1 (boundary requirements)
ii. P

iI
A
i

=
iI
P(A
i
) for A
i
: i I a set of countable, pairwise disjoint events (countable
additivity)
Property 1 denes the boundaries of the measure of the probability of an event to be 0 and 1, as
opposed to 0 and 100, simply by convention. In general, that P() = 0 and the function be countably
additive are all that is required for a function to be a measure. Now, an interpretation of a probability
measure should be fairly familiar. Given an outcome of a random experiment denoted as event X has
probability P(X), one could say that, for example, if the random experiment is repeated n times, one
could expect event X to be the outcome nP(X) times.
In some situations, however, the randomness of probability measures is not appropriate for dealing
with uncertainty. Fuzzy measures are a more generalized type of measure, which may deal with uncer-
tainty in a dierent way.
Denition 3.2.2. Given a set X and a nonempty family C of subsets of X, and a residuated lattice
L = (L, , , 0, 1, ), a fuzzy measure on (X, C) is a function g : C L such that:
i. g() = 0 and g(X) = 1 (boundary condition);
ii. for all A, B C, if A B then g(A) g(B) (monotonicity);
iii. for any increasing sequence A
1
A
2
. . . in C, if

i=1
A
i
C, then lim
i
g(A
i
) = g(

i=1
A
i
)
(continuity from below);
iv. for any decreasing sequence A
1
A
2
. . . in C, if

i=1
A
i
C, then lim
i
g(A
i
) = g(

i=1
A
i
)
(continuity from above);
18
Here, as usual, out set L is often taken to be [0, 1].
3.2.2 possibility theory
Possibility theory traces back to Zadeh (1975) and is covered in depth by Dubois and Prade (1988),
and is a dierent usage of possibility than that of modal logic. To understand possibility theory, we
will start with a evidence theory, which is based on two dual nonadditive measures: belief measures
and plausibility measures.
Denition 3.2.3. Given a universal (in our case nite) set X, a belief measure is a function Bel :
/(X) L where /(X) is the power set of X and L is the usual residuated lattice, L = (L, , , 0, 1, ),
such that Bel() = 0 and Bel(X) = 1, and Bel(

n
i=1
A
i
)

j
Bel(A
j
)

j<k
Bel(A
j

A
k
)+

j<k<l
Bel(A
j

A
k

A
l
)
. . . + (1)
n+1
Bel(

n
i=1
A
i
) for all possible families, (A
i
) of subsets of X. Due to this inequality, belief
measures are called superadditive. If X is innity, the function Bel is required to be continuous from
above.
Denition 3.2.4. Associated with each belief measure is a plausibility measure dened by the equa-
tion PlA = 1 Bel(

A) where

A is the complement of A, for all A X. Thus an denition of
plausibility measures independent of belief measures would be, a plausibility measure is a function
Pl : /(X) L such that Pl() = 0 and Pl(X) = 1, and Pl(

n
i=1
A
i
)

j
Pl(A
j
)

j<k
Pl(A
j

A
k
) +

j<k<l
Pl(A
j

A
k

A
l
) . . . + (1)
n+1
Pl(

n
i=1
A
i
) for all possible families, (A
i
) of subsets of X. Due
to this inequality, belief measures are called subadditive. If X is innity, the function Bel is required to
be continuous from below.
Immediate consequences of super- and subadditivity are that if n = 2, A
1
= A, and A
2
=

A,
Bel(A) + Bel(

A) 1 and Pl(A) + Pl(

A) 1.
Now before we discuss an interpretation of belief and plausibility measures, we must introduce a
function which while not itself being a fuzzy measure, can characterize Bel and Pl.
Denition 3.2.5. A basic probability assignment is a function m : /(X) L such that m() = 0 and

A/(X)
m(A) = 1.
An interpretation of m(A) for A /(X) is that m(A) is the proportion to which all available and rel-
evant evidence supports the claim that an element of X belongs to the set A in particular. Here nothing
19
is implied for subsets B A, whose basic probability assignment must be expressed by another value,
m(B). Notice that this is very dierent from probability, namely, it is not necessary that m(X) = 1, it is
not necessary that A B imply that m(A) m(B), and there is no necessary relationship between m(A)
and m(

A). In fact, m(X) is the proportion to which the evidence supports the claim that an element is
equally likely to be in any of the subsets of X.
In order to understand belief and plausibility measures, it is important to note that a belief mea-
sure and a plausibility measure are uniquely determined by m, where for all A /(X), Bel(A) =

BBA
m(B), and Pl(A) =

BAB
m(B). The reverse direction is also possible, m may be determined
from Bel (and dually from Pl) by m(A) =

BBA
(1)
AB
Bel(B). Thus one of the three, Bel, Pl, or m,
is sucient to determine the other two.
Now based on the relationship between m and Bel, we may say that Bel(A) represents the total evi-
dence or belief that the element belongs to A as well as subsets of A, while the Pl(A) represents the total
evidence or belief that an element belongs to a set which intersects in some way with A. Thus a belief
measure captures the idea of belief in that our degree of belief in some statement or of some answer
A to a question is equal to the sum of degrees to which we evidence supports any statement or answer
which is more specic than A, and, of course, the degree to which the evidence supports exactly A. On
the other hand a plausibility measure captures the idea of plausibility in that the degree of plausibility of
some statement or answer A to a question is equal to the sum of the degrees to which evidence supports
any set in which we are unable to determine how much more or less valid a statement independent of
A is to some subset of A as well as the degrees to which evidence supports any statement or answer
which is more specic than A, and, again, the degree to which the evidence supports exactly A.
With this interpretation, it is obvious that Pl(A) Bel(A) for all A /(X).
Denition 3.2.6. Every element A /(X) for which m(A) > 0 is sometimes called a focal element of
m because the evidence supports A. Thus we may dene a body of evidence as a pair (F, m) where F is
a set of focal elements and m is the associated basic probability assignment. Total ignorance is the case
20
where m(X) = 1, and hence m(A) = 0 for all A X. This also implies that Bel(X) = 1, and Bel(A) = 0
for all A X, while Pl() = 0 and P(A) = 1 for all A X.
3.3 Fuzzy Relation
We can generalize relations to fuzzy relations the same way we generalized sets to fuzzy sets.
Recall, a binary relation on a set X is a subset of X X. Thus we could identify a relation, , with
its characteristic function, (, ) : x X 0, 1 where a pair, (x, y) X X is in if and only if
(x, y) = 1.
Denition 3.3.1. A fuzzy relation, on a set X is a function : X X L with L a residuated
lattice.
With this denition, a fuzzy relation, : X X [0, 1], may be said to be continuous just as any
function into a Euclidean space is continuous.
It may be interesting to dene some properties for fuzzy relations. Let be a fuzzy relation on a
set X, and let x, y, z X. Then is:
i. totally reexive if rho(x, x) = 1;
ii. totally non-reexive if rho(x, x) = 0;
iii. symmetric if (x, y) = (y, x);
iv. f -transitive if (x, y) (y, x) and (y, z) (z, y) then (x, z) (z, x);
v. -transitive if (x, y) (x, z) (x, z) with the lattice ordering in L;
vi. complete if (x, y) + (y, x) = 1.
Billot (1992) maintains that reexivity under this denition is not necessary as in this denition as
if an arbitrary does not place any importance in the comparison, they may allow [0, 1]. In many
cases Nguyen (1999), what we here call totally reexive is called merely reexive. Nguyen
(1999), Zadeh (1975). and Dubois and Prade (1980) all utilize some version of -transitivity, while
f -transitivity is used by Billot (1992). The reason for diering denitions of transitivity is that f -
transitivity maintains the irrelevance of independent alternatives.
21
We will concentrate on which are -transitive. Note, this is the case if the degree to which two
elements are related is greater than the join of the degree to which each element is related to some other
element. If xy = min x, y then this says that if a relation is transitive, then an item is at least as similar
to another item as the minimum of the similarity between either of the items and any other item.
Denition 3.3.2. A fuzzy relation is a fuzzy indierence relation if it is symmetric, -transitive, and
not totally non-reexive.
3.3.1 fuzzy preference relation
One of the themes of this paper is to discuss each denition by the story the idea behind the de-
nition needs to tell. If we wish to state a preference between two objects, we may not wish for
that preference to be crisp. It may be important to consider the degree to which one object is pre-
ferred over another. For instance, consider the preferences, , between values for some country of
having an exclusive trade agreement with Lichtenstein,
L
, with Andorra,
A
, and with the United
Kingdom,
UK
. Perhaps an agreement with the UK is locally preferred to an agreement with Andorra,
(
UK
,
A
) (
A
,
UK
). Perhaps (
L
,
A
) = (
A
,
L
) = . If (0, 1], then we say that a country
would be locally but actively indierent of the relative values of agreements with Andorra and Licht-
enstein. That is to say that the country can see that the two agreements have merit and that the merit of
each is dierent from the other, but that the country does not particularly prefer an agreement with one
to the other. On the other hand, if = 0, the agent cannot particularly compare the two, perhaps their
is not enough information, and the decision makers of a country do not know enough about the two to
know the values of such an agreement. Or perhaps the country is completely uninterested in trading
with either country. Thus we will not require preference relations to be reexive, but we will require
that they not be totally non-reexive.
Denition 3.3.3. A fuzzy preference relation is a fuzzy relation that is not totally non-reexive, -
transitive
The concept of belief functions and plausibility functions can easily be adapted to be the image of
fuzzy sets, be they fuzzy preference relations, fuzzy equivalence relations, or another fuzzy mapping.
Later in this thesis, I will suggest using belief functions to dene fuzzy preference relations. For more
on how fuzzy sets might be encoded based in any of these theories, see Klir and Yuan (1995).
22
4 Game Theory
In games of strategy or chance, such as chess or poker, we might talk about a strategy set as a sort
of how-to book, which describes how to bet or what move to make, given the entire history of bets,
moves, and cards the player has seen in the game. The goal of this book would be to guarantee each
player the highest expected payo, be it a checkmate or large winnings. If the book were complete
enough, it would not be necessary to play the game, as at each stage of the game, there would be a page
which dictated the subsequent move. Each player would be able to submit their strategy and instantly
receive their winnings or forced to pay out their losses. In games of chance, the expected earnings
and losses could also be calculated, again based on the strategy outlined in each players book. Game
theory is the tool by which we might understand and model games, foreseeing outcomes based on the
rules themselves, rather than based on actually playing the game.
With this in mind, social scientists have attempted to draw parallels between games and real human
situations, and have claimed that certain political or social events have happened because essentially
the actors are playing some game and this outcome is parallel to some fortuitous payo in the game.
These fortuitous payos generally correspond to equilibria in the game, a central concept to game
theory I will discuss shortly. Thus a social scientist can dene some game which models the choices
and preferences of the parties involved in the situation being studied, and guarantee that their theorems
are logically consistent with their set of assumptions.
4.1 crisp games
Considering situations with N players, we will use strategy sets, =
1
, . . . ,
N
, to describe
all dierent strategies possible for the players. Each strategy will result in some consequence, K =
K
1
, . . . , K
N
. For each player, there is a preference relation, =
1
, . . . ,
N
, dened componentwise
on the set K. Given those consequences, payos are doled out from a payo set, =
1
, . . . ,
N
. A
23
vector of consequence functions, g = g
1
, . . . , g
N
, then associates with each strategy a consequence,
g
i
: K
i
.
Often our strategy set will be restricted to being a compact and convex subset of R
n
. Recall compact
means closed and bounded. Doing this forces our strategies to be similar to a nite set in that we can
make statements about its boundaries and we know that it has an upper and lower bound. For strategies
of games, boundedness is sensible because although a strategy may include a dense set within a range
of values, one should not expect a player to have an innitely great amount of some resource to apply.
Occasionally forcing strategy sets to be closed reduces their applicability to some game, although taking
the closure of an open strategy set will often still give the researcher a clear view of the game, with the
possibility of removing the points of closure after the analysis is complete.
The preference relation is dened to be a complete, transitive, and reexive binary relation. Often,
a vector of utility functions, u = u
1
, . . . , u
N
: then associates a payo with each strategy a
payo by u
i
(
i
) =
i
with
i

i
and
i

i
such that for x, y
i
, x
i
y if an only if u
i
(x) u
i
(y).
We will discuss games as sets, N, , K, g, , letting us specify the number of players, the choices
of strategies for each player, the consequences for each player given the strategies of all players, and an
ordering of the preferences of each player on those consequences. Notice that preferences are dened
independent of the consequences of the strategies for another player.
i
can be any possible strategy.
This means that a strategy must be dened for every possible eventuality of the game, even if that
eventuality would never occur given the rest of the strategy. For instance, it will tell a chess player
playing black how to play various variations of the Sicilian Defense, even if it rst tells the player to
respond to an opening of e4 with e5, preferring the Ruy Lopez opening.
Often situations are modeled by games to show that some set of strategies by the players is more
likely than others. One reason that a strategy would be more likely is that given some a strategy for all
other players, no one player would secure for themself a greater payo by changing their strategy.
Denition 4.1.1. for some N-player game, N, , K, g, , with the set of strategy proles, set of
payo proles, u : K the utility function, a strategy prole,

1
, . . . ,

N
is a Nash equilib-
riumif for all i 1, . . . , N, there exists no strategy
i

i
such that g
i
(

1
, . . . ,

i1
,
i
,

i+1
, . . . ,

N
)
i
g
i
(

1
, . . . ,

N
)
This is the basic equilibrium condition, and has been the most explored equilibrium in Game The-
24
ory. Another way of formulating a Nash equilibrium is by dening a best-response function. A best-
response function is a function that takes as its input the strategy of all of the opponents of some player
and outputs the one or more strategies which will give the player the highest payo.
Denition 4.1.2. In a N-player game, given a strategy set =
1
, . . . ,
N
, and a consequence function
for player i, g
i
a best-response function is a function B
i
:
1
, . . . ,
i1
,
i+1
, . . . ,
N

i
such that
B
i
=
i

i
: g
i
(
1
, . . . ,
i1
,
i

i+1
, . . . ,
N
) g
i
(
1
, . . . ,
i1
,

i
,
i+1
, . . . ,
N
) for all

i

i

We can dene Nash equilibria in terms of best-response functions.

Lemma 4.1.3. for some N-player game, N, , K, g, , with the set of strategy proles, K set of
consequences, g : K the vector of consequence functions, and = (
1
, . . . ,
N
) the set of
preference relations for each player, a strategy prole,

1
, . . . ,

N
is a Nash equilibrium if and
only if for all best-response functions, B 1, . . . , B
N
,
i
B
i
((
1
, . . . ,
i1
,
i+1
, . . . ,
N
) for all
i 1, . . . , N
To explore these concepts further, we will need some concepts from topology.
Denition 4.1.4. A preference relation,
i
, over a set, X, with X = n, is quasiconcave on X
i
if for
every x

= x

1
, . . . , x

n
X, the set, x
i
inX
i
: (x

1
, . . . , x

i1
, x
i
, x

i+1
, . . . , x

n
) (x

1
, . . . , x

n
) is convex.
One of the most famous results in Game theory is John Nashs 1951 proof of the existence of Nash
equilibria in N-player games given certain conditions. His proof used the Kakutani xed point theorem,
which gives conditions on a function such as the best-response function so that there exists some value

1
, . . . ,

N
such that

i
B
i
((

1
, . . . ,

i1
,

i+1
, . . . ,

N
) for all i 1, . . . , N.
Theorem 4.1.5. Kakutanis xed point theorem Let X be a compact convex subset of R
n
and let f :
x X be a set-valued function such that for all x X, the set f (x) is nonempty and convex and
the graph of f is closed (for all sequences x
n
and y
n
such that y
n
f (x
n
) for all n, x
n
x, and
y
n
y, we have y f (x), then there exists a xed point x

X such that x

f (x

).
Using xed point theorems gives a situation that is invariant over repeated applications of a map-
ping. In this case, we will nd that we may continue to apply the best-response function, but we will
reach a particular state which no player has a protable deviation as dened by the response function.
Now we can state and prove a theorem about the existence of Nash equilibria.
25
Theorem 4.1.6. For some N-player game, N, , K, g, , a Nash equilibrium exists if for all i
1, . . . , N the set
i
of strategies for player i is a nonempty compact, convex subset of R
n
, and each
preference relation,
i
is continuous and quasiconcave.
Proof. Dene B : by B() =
i
B
i
(
1
, . . . ,
i1
,
i+1
, . . . ,
N
) for all i 1, . . . , N where B
i
is the best response function for player i. Notice B
i
(
1
, . . . ,
i1
,
i+1
, . . . ,
N
) is nonempty since
i
is continuous and
i
is compact. Also, B
i
(
1
, . . . ,
i1
,
i+1
, . . . ,
N
) is convex by the denition of the
quasiconcavity of
i
on
i
. Also, B has a closed graph by the closedness of g. Notice, this is equivalent
to saying that for sequences of strategies, P
1
, P
2
, . . . and Q
1
, Q
2
, . . ., where P
n
P and Q
n
Q
that if for all n, Q
n
B(P
n
), then Q P. Thus, Kakutanis xed point theorem applies and B has a
xed point. So by our lemma, this xed point is a Nash equilibrium of the game.
In the case of two player games, a certain number of examples are illustrative of the properties
required for a game to have a Nash equilibrium. The games in the next sections are given in the form
of a matrix, with the choices for player 1 at the top and for player 2 along the side, and the payos for a
given pair of choices given as the corresponding ordered pair such that (x,y) implies that player 1 gets a
payo of x, and player 2 gets a payo of y. Each players preferences are here to get a higher numbered
payo.
4.1.1 Bach or Stravinsky (Battle of the Sexes)
The game classically known as the Battle of the Sexes or more recently dubbed Bach or Stravinsky
involves two players who would like to coordinate their actions so that both end up at the same venue for
a rendez-vous. It is a case where a game has two Nash-Equilibria, since given the choice of one player,
the other will want to chose the same composer, that is the best response function will recommend the
same composer as the other player.
Bach Stravinsky
Bach (5,6) (0,0)
Stravinskyy (0,0) (6,5)
Table 4.1 Bach or Stravinsky
26
4.1.2 Matching Pennies
Matching pennies is the classic zero-sum game, which means that the two players have antisym-
metric payos. In this game, both players have a penny, which they conceal from their opponent. The
they simultaneously show their opponent their penny, having placed it heads up or tails up. Player 1
gets a point if the pennies match, while player 2 receives a point if the pennies do not match. This
example shows a case where their is no Nash equilibrium. In this case, this is because the preference
relation is not quasiconcave.
Heads Tails
Heads (1,-1) (-1,1)
Tails (-1,1) (1,-1)
Table 4.2 Matching Pennies
4.1.3 Prisoners Dilemma
The most famous game is perhaps the Prisoners Dilemma. In contemporary literature, it rst arose
in work at RAND Corporation such as that of Flood (1952). Its application was popularized largely
by Schelling (1960). It has also been very popular in a repeated form (Axelrod (1985)) where one
can study concepts like learning and reputation. Borges et al. (1997) used fuzzy rules to study this
repeated or iterated prisoners dilemma. Many stories have been told to describe situations where this
arises. Perhaps the most interesting (if not compelling) is in the situation of an arms race. In this
telling, one would hope that the two players (countries) cooperate and halt production or dismantle
their arms. However if one country does so, the best response function tells the other to get themselves
an advantage by defecting (and building a large enough arsenal to destroy their opponent or at least to
force the opponent into a weak bargaining position). Thus if neither player trusts the other, they will
both defect, the unique Nash equilibrium, even though that position is not ecient (Pareto optimal).
Cooperate Defect
Cooperate (5,5) (8,1)
Defect (1,8) (3,3)
Table 4.3 Matching Pennies
27
4.2 fuzzy games
We sawbefore that sets can be turned into fuzzy sets and relations can be turned into fuzzy relations.
Thus, it seems to create a fuzzy game, we have many choices. Each choice stems from a dierent story
about how the game is to be played. Song and Kandel (1999) examines a variation on the prisoners
dilemma where the degree to which the player wishes to help or harm his partner is fuzzy. Garagic
and Cruz (2003) uses a fuzzy xed-point theorem to show that certain fuzzy matrix games have a Nash
equilibrium. Another way of making a game fuzzy is explored by Ar (2006) where outcomes of a
variation of the prisoners dilemma are broken down into nite, discrete values based on nite, discrete
levels of cooperation and defection. Another work, Nishizaki and Sakawa (2001) explores cooperative
fuzzy games as a method of conict resolution and concentrates on numerical solutions to such game.
In our case, we will start with looking at a simple fuzzy game by utilizing a fuzzy preference relation.
In this case, we might say that the degree to which a player prefers a given object over another is
fuzzy. One way to do this, is to relate each outcome by the degree of belief that it is better than another
outcome. In this case, we need to dene a new type of best-response function.
Denition 4.2.1. In a N-player game with fuzzy preference relation, , given a strategy set =

1
, . . . ,
N
, and a consequence function for player i, g
i
a fuzzy preference relation best-response
function is a function B
i
:
1
, . . . ,
N

i
such that B
i
(
1
, . . . ,
N
) =

i

i
: (
1
, . . . ,
N
)
(
1
, . . . ,
i1
,
i
,
i+1
, . . . ,
N
)
The fuzzy preference relation best-response function is obviously non-empty. In this case, our
previous proof of the existence of Nash equilibria can still hold so long as certain conditions hold:
Denition 4.2.2. A fuzzy preference relation,
i
, over a set, X, with X = n, is fuzzy quasiconcave on
X
i
if for every x

= x

1
, . . . , x

n
X, the set, x
i
inX
i
: ((x

1
, . . . , x

i1
, x
i
, x

i+1
, . . . , x

n
), (x

1
, . . . , x

n
))
((x

1
, . . . , x

n
), (x

1
, . . . , x

i1
, x
i
, x

i+1
, . . . , x

n
)) is convex.
Theorem 4.2.3. For a fuzzy preference relation, , on an N-player game, N, , K, g, , a Nash equi-
librium exists if for all i 1, . . . , N the set
i
of strategies for player i is a nonempty compact, convex
subset of R
n
, and for each preference relation,
i
,
i. is continuous;
28
ii. is fuzzy quasiconcave.
The proof follows exactly as before.
Clearly, this conception of fuzzy games amounts to reducing fuzzy preferences to crisp preferences.
Another very similar construction called an equilibrium solution with respect to the degree of attain-
ment of the aggregated fuzzy goal is discussed in Nishizaki and Sakawa (2001). There work concen-
trates on computational methods for nding such solutions. We turn to a more sophisticated conception
of a fuzzy game, and dene a Nash equilibrium concept based upon it. A generalized version of this is
due to Butnariu (1979). Exploring this work will give us another example of the importance of xed
points in analyzing equilibria in game theory. As Butnarius formulation is more general than that of
some other games found in more recent research, including that of Nishizaki and Sakawa, examples
and ideas discussed in that work could be adapted to Butnarius concept. However, we will mention an
example from Butnarius work which shows that fuzzy beliefs about a players opponents can allow us
to explore interesting games that allowing for only fuzzy preference relations among outcomes misses.
4.2.1 Butnarius Game
To start, we will explore the concept of fuzzy topology and of a fuzzy xed point theorem.
Denition 4.2.4. A fuzzy correspondence on a set X is a mapping R : X [(X) where [(X) is the
fuzzy set based on the powerset of X. We may examine the membership function, a(), of the fuzzy
subset A [(X). For a fuzzy correspondence, R(), on X, we use R
X
to denote the fuzzy subset of
X X which has a membership function r(x, y) for any (x, y) X X. We note M(X) the set of fuzzy
correspondences on X.
Denition 4.2.5. Now, given a fuzzy correspondence R() on X, a xed point, x

X, is an element
such that for all x X, r(x

, x

) r(x

, x).
Denition 4.2.6. We say that a fuzzy subset, A [(X), is convex if and only if its membership function
is concave, ie for all q [0, 1] and for all (x, y) X X, we have a(qx + (1 q)y) min a(x), a(y).
Remember that a topological vector space, X, is a vector space with continuous vector addition and
scalar multiplication. Such a topological space is Hausdor-separated if for any x X, the intersection
29
of all the closed neighborhoods of X is the single element x. Such a space is locally convex if for all
x X, every neighborhood of x is convex.
Denition 4.2.7. A fuzzy correspondence R() is convex if for the topological vector space X, locally
Hausdor-separated and C X such that C is nonempty, compact and convex, the fuzzy subset R
x

[(X) is a convex fuzzy subset of X. R() is closed if and only if the membership function, r(x, ) is
upper semicontinuous.
Now, following the lead of Butnariu (1979) and Heilpern (1981), we present a xed point theorem
in a fuzzy universe:
Theorem4.2.8. Butnarius generalized fuzzy xed point theoremFor X a real locally convex, Housdor-
separated, topological vector space, X, and any nonempty, compact, convex subset, C X, if R() is a
convex fuzzy correspondence closed in C, then R() has a xed point in C.
The proof can be found in Billot (1992), and is an adaption of the one given in Butnariu (1979) and
Butnariu (1978). It is dierent from the theorem and proof given by Heilpern (1981), which requires
that the fuzzy correspondence be, in eect, a fuzzy contraction mapping on an appropriately dened
fuzzy metric. Notice that Butnarius theorem is valid for any topological structure, including metric
spaces which are used in the fuzzy xed point theorem of Song and Kandel (1999). If we restrict X to a
subset of a Euclidean space and the fuzzy correspondence is a nonempty point to set function, we have
Kakutanis xed point theorem.
We also need to impose some particular characteristics on our fuzzy discourse. We will denote U
as the universe of objects being discussed. as before, for x U, a(x) is the membership level of x. The
product of two fuzzy subsets, A and B with membership level functions, a() and b() respectively, is
dened by [a b](x) = a(y) b(z) for x = (y, z) U and [a b](x) = 0 else. We will dene a particular
product of membership levels of fuzzy subsets where for fuzzy relations, R : A B and S : B C,
and (x, y) A, we have S R(x, z) = sup R(x, y) S (y, z) : y U for (x, z) U, S R(x, z) = 0 else.
Now, again, we return to the stories that are told to explain why grouping of mathematical structures
models some sort of game. Again we will consider N-person games. In this game, all players rst gather
and exchange information about what they will do, based on whatever logic they wish to employ. Then
all the players are cut o from communication with each other to some degree (possibly completely).
30
At this point, each player knows the choices of strategies that other players have. They can then dene
for themselves a vector of fuzzy sets which express the degree to which they believe the other players
might choose some strategy. The players will also be unsure about whether any given outcome is
preferable to another.
i. =
1
, . . . ,
N
will be the set of strategies. We will require each player, i, to have only a nite
number of strategies from which to choose from, indexed by the function, n(i). so
i
is a vector
of strategies with n(i) elements.
ii. For each player, i, we will dene the strategic arrangement as a set Y
i
[0, 1]
n(i)
such that Y
i
=

n(i)
j=1
Y
j
i
,
n(i)
j=1
Y
j
i
= 1, where Y
j
i
is the percent of time player i chooses action j. Let Y =
N
i=1
Y
i
.
This allows the players to choose mixed strategies, and allows the set of strategies to be convex.
iii. For each player, i, we dene e
i
[(
iN
Y
i
) and for all w = (w
1
, . . . , w
N
)
iN
Y
i
, as a belief
level of the strategic choice w evaluated by player i. That is to say that each player evaluates the
degree to which they believe that the other players will play each of their possible strategies. e
i
is a belief function as dened earlier.
iv. We will be considering s
i
= (A
i
, w
i
) [((
jNi
Y
j
) Y
i
) as a strategic conception for player i.
Here, A
i
is the belief of player i in the strategic arrangements of the other players, and is a fuzzy
subset of the set of all of the possible strategies of the other players. Here A
i
represents player is
belief in the strategies that would be chosen by the other players given that i plays w
i
.
v. We also want it to be the case that for A
i
[(
jNi
Y
j
) and A
i
then there exists (A
i
, w
i
)
[(
jNi
Y
j
) Y
i
) such that e
i
[A
i
](w
i
) 0. Thus no matter what the others might do, no player
will be unable to formulate a strategy.
Denition 4.2.9. A play in this game is a vector s = (s
1
, . . . , s
N
) such that s
i
[(
jNi
Y
j
) Y
i
.
For s
i
= (A
i
, w
i
) [(
jNi
Y
j
) Y
i
) and s

i
= (A

i
, w

i
) [(
jNi
Y
j
) Y
i
), s
i
is a better strategic
conception than s

i
if and only if e
i
[A
i
](w
i
) > e
i
[A

i
](w

i
). s is socially preferred to s

if s
i
is a better
strategic conception than s

i
for all i N.
Denition 4.2.10. A possible solution of a game of this type is a play, s

= (s

1
, . . . , s

N
), where s

i
=
(A

i
, w

i
) such that for any other play, s = (s
1
, . . . , s
N
), with s
i
= (A

i
, w
i
), e
i
[A

i
](w

i
) e
i
[A

i
](w
i
).
31
Thus a possible solution corresponds intuitively with equilibrium points of non-fuzzy games in that
it is marked by its advantage in terms of a preference relation. In games of this type, an equilibrium
point is a renement of this concept.
Denition 4.2.11. An Nash equilibrium of a game of this type is a possible solution, s

= (s

1
, . . . , s

N
),
where s

i
= (A

i
, w

i
) such that a

i
(w
1
, . . . , w
i1
, w
i+1
, . . . , w
N
) = 1 if w
i
= w

i
for any i N i and
a

i
(w
1
, . . . , w
i1
, w
i+1
, . . . , w
N
) = 0 elsewhere.
The additional condition, often called non-constraining collaboration implies that player i expect a
strategy w
j
from player j only when player j intends on playing that strategy. Thus a Nash equilibrium
arises after uncertainty in the belief of the moves of the other players is removed.
We will now compute two new fuzzy preference relations:
Denition 4.2.12. The individual fuzzy preference of a player i is the fuzzy relation E
i
(, ) on [(WW)
where W =
jN
Y
j
Y such that E
i
(s, s

) = [e
i
[A
i
](w

i
) e
i
[A
i
](w
i
)]
jNi
e
j
[A
j
](w

j
).
This individual fuzzy preference now takes into account what the other players would believe about
the degree of possibility of player i going against their expectations.
Denition 4.2.13. The social preference relation is a fuzzy relation E(, ) on [(W W) such that
E
s
(s, s

) =
iN
E
i
(s, s

) for all (s, s

) W W.
Now we may state and prove the main theorems presented originally by (?, Butnariu79)
Theorem 4.2.14. Let s

be a play of a game of this type, with s

i
= (A

i
, w

i
). If A

i
, for all i, then
s

is a possible solution if and only if s

is a xed point in E
s
(, ).
Lemma 4.2.15. There is no i such that if s

is a xed point in E
s
(, ), then for s

i
= (A

i
, w

i
), we have
e
i
[A

i
](w

i
) = 0.
It leaves to be understood whether any useful equilibrium concepts can be dened and shown to
exist if the strategic choices are fuzzy. We remember that a Nash equilibrium arises when for all i,
j N, A

i
= A

j
. We would call this game a game with perfect information. However, the preference
relations remain possibly fuzzy, and we have reduced out fuzzy game to the fuzzy preference relation
game discussed earlier. This suggests that the players may approach a Nash equilibrium in a fuzzy
game of this type as they exchange more information about their full range of strategic conceptions.
32
4.2.2 Example
Let us look at an example of an application of his formulation and equilibrium concept in political
science based on that provided by Butnariu (1979). For simplication, we can look at two players
named 1 and 2. In this game, the two players are interested in dening their military investments vis-
a-vis each other. Here we are looking at only the international level game in Putnams two level game
concept, ignoring completely the inuence of national bodies such as special interests and legislatures.
For the purposes of this game, we will say that military investments perceived by opponents serve as
a deterrence for escalating of military conict between the states. The two states then have as goals
both the reduction of intensity of conict and the minimization of economic energy put into military
investments toward that ends.
We can easily arrange this situation so that it ts in the conditions dening Butnarius game. First,
let
k
=
k
1
, . . . ,
k
n(k)
be dierent strategies of military investment for player k. Perhaps
k
1
is missile de-
fense,
k
2
is increased naval size,
k
3
is greater nuclear capabilities, and so on. Note that this set is crisp,
and the players are allowed to chose as strategic arrangements mixed strategies which amount to invest-
ing a certain amount of a budget in any of various strategies until the budget is used up. In the course
of relations between the two players, each players can exchange information and (possibly) make guar-
antees about their future actions. From these exchanges, each player can decide what they believe to
be the value of their actions basing these values on their beliefs about how their opponent values the
actions available to them. Normalizing each players budget, we can denote a strategic arrangement by
Y
k
[0, 1]
n(k)
. We notice that the set of degrees of belief of player k in the plausibility of each of their
opponents strategies
k
denes a fuzzy set. We may denote these belief by e
i
[(
iN
Y
i
). Fuzzy
pairs, (A
k
, w
k
) (Y
k
Y
k
) are strategic conceptions, where given player k chooses w
k
, A
k
is a fuzzy
set representing player ks level of belief that his opponent will play any given strategy given k chooses
w
k
. We will also let it be the case that for any set of information exchanged, each player can form a
strategic conception as a reply to his beliefs about the other players actions, ie that the fth condition
is satised.
This example, then, clearly denes a fuzzy game of the type Butnariu described. Thus we can say
33
that two strategic conceptions of the game can be made, and one can be preferable to the other if its
belief level is more believable. Each player will examine to what degree they believe their opponent
will do something advantageous given they act in a way that most closely approaches what they believe
is advantageous to them. We will again be able to dene a possible solution as before, a condition
where both players nd their play to be advantageous compared with any other conception. It is not
clear exactly what these possible solutions will look like in terms of the military investments of the
players without making explicit the belief functions and strategies available, but our theory shows
that the possible solutions do arise as xed points of fuzzy correspondences. So if given a chance to
play multiple games, we may expect the players to eventually discover a stable solution by iteration,
a stable play, which represents a possible solution. This is by no means a necessary outcome, as the
eistence of a xed point does not imply that it will arise in a nite number of iterations of our fuzzy
corespondence. Butnarius denition of an equilibrium point is also meaningful here, as it is a possible
solution where the uncertainty of each players beliefs in the actions of the other players is removed.
In this case, it means that it is a possible solution where the military apparatus of the opponent is
openly observed, perhaps by cameras or UN appointed monitors. Butnarius work, then, suggests that
it is these solutions that are the most stable. It is not, of course, surprising that reducing uncertainty
provides stability. Unfortunately, most of the work in fuzzy game theory has not brought us closer to
examining games with fuzzy beliefs, only fuzzy goals. A return to Butnarius work, then, may allow
us to reexamine what it is about fuzzyness that makes it so attractive to game theory.
34
5 Conclusion
In this paper we have dened fuzzy sets an looked at a couple ways they have been used, eventu-
ally discussing the topic of fuzzy games. We concentrated especially on Butnarius game because it
encompasses the more detailed work of other authors, and also because the example Butnarius game
provides is particularly seductive to those interested in social science. In particular, Butnariu examines
fuzzyness in social beings (people, societies, etc.) beliefs about the actions of social beings, including
themselves. It is true that Butnarius game suggests that to a certain degree, more information brings
more stability to games, a fact that is well known both in the study of theoretical games of a crisp type
(although not universally), as well as intuitively and in social practice (again exceptions do exist). An
exploration of social events with an eye to their formulation in the terms of Butnarius game has not
been seen, although parallel formulations do exist, and suggest that Butnarius theorems would hold
true in the real world. In this paper, I did not discuss the nature of the fuzzy correspondences which
would give rise to possible solutions of Butnarius game, and in future work, both a discussion of that
nature, and an exploration of what would make that correspondence easily denable in terms of real
world processes would be important. This would allow us to compare the mechanisms which make
Butnarius theorems work mathematically with the mechanisms which govern social processes in the
real world. This type of comparison, along with an honest assessment of how all the theorems of game
theory speak to actual causal mechanisms in the real world applications social science is so keen on,
is perhaps the next important movement in the philosophy of social science. I personally hope to see
more work on the meanings of the applications of game theory to social science. Why does game
theory model human behaivior? What mathematical tools can we use to help us better make those
models? Butnarius game shows that game theory is still relevant if we release our formulations from
the constraints of crisp sets, and where fuzzy sets better model a situation than crisp sets, the existance
of a maturing eld of fuzzy game theory will give social scientists the tools to more often move in that
35
direction.
36
Bibliography
Ar, B. (2006). Linguistic fuzzy-logic game theory. Journal of Conict Resolution, 50(1):2857.
Axelrod, R. (1985). The Evolution of Cooperation. Basic Books.
Billot, A. (1992). Economic Theory of Fuzzy Equilibria: An Axiomatic Analysis (Economic Theory of
Fuzzy Equilibria). Springer-Verlag Telos.
Borges, P. S., Pacheco, R. C., Barcia, R. M., and Khator, S. K. (1997). A fuzzy approach to the
prisoners dilemma. Biosystems, 41(2):127137.
Butnariu, D. (1978). Fuzzy games; a description of the concept. Fuzzy Sets and Systems, 1:181192.
Butnariu, D. (1979). Solution concepts for n-person fuzzy games. Advances in Fuzzy Set Theory and
Applications, pages 339359.
Cignoli, R. L., Dottaviano, I. M., and Mundici, D. (1999). Algebraic Foundations of Many-Valued
Reasoning (Trends in Logic). Springer.
Dubois, D. and Prade, H. (1988). Possibility Theory. Springer.
Dubois, D. J. and Prade, H. (1980). Fuzzy Sets and Systems : Theory and Applications (Mathematics
in Science and Engineering). Academic Press.
Flood, M. M. (1952). Some experimental games. research memorandum rm-789. Technical report,
RAND Corporation, Santa Monica, California.
Garagic, D. and Cruz, J. B. (2003). An approach to fuzzy noncooperative nash games. Journal of
Optimization Theory and Applications, 118(3):475491.
37
Gerla, G. (2001). Fuzzy Logic : Mathematical Tools for Approximate Reasoning (Trends in Logic).
Springer.
Goldstone, J. A., Gurr, T. R., Har, B., Levy, M. A., Marshall, M. G., Bates, R. H., Epstein, D. L.,
Kahl, C. H., Surko, P. T., Ulfelder, J. C., and Unger, A. N. (2000). State failure task force report:
Phase iii ndings.
Heilpern, S. (1981). Fuzzy mappings and xed point theorem. Journal of Mathematical Analysis and
Applications, 83:566569.
Hjek, P. (2001). Metamathematics of Fuzzy Logic (Trends in Logic). Springer.
Klir, G. J. and Yuan, B. (1995). Fuzzy Sets and Fuzzy Logic: Theory and Applications. Prentice Hall
PTR.
Mansur, Y. M. (1995). Fuzzy Sets and Economics: Applications of Fuzzy Mathematics to Non-
Cooperative Oligopoly. Edward Elgar Publishing.
Mares, M. (2001). Fuzzy Cooperative Games : Cooperation with Vague Expectations (Studies in
Fuzziness and Soft Computing). Physica-Verlag Heidelberg.
Morgenstern, O. and Von Neumann, J. (1944). Theory of Games and Economic Behavior. Princeton
University Press.
Myerson, R. B. (1997). Game Theory: Analysis of Conict. Harvard University Press.
Nguyen, H. T. (1999). A First Course in Fuzzy Logic, Second Edition. CRC.
Nishizaki, I. and Sakawa, M. (2001). Fuzzy and Multiobjective Games for Conict Resolution (Studies
in Fuzziness and Soft Computing). Physica-Verlag Heidelberg.
Ragin, C. C. (2000). Fuzzy-Set Social Science. University Of Chicago Press.
Schelling, T. C. (1960). The Strategy of Conict. Harvard University Press, Cambridge, MA.
Smithson, M. J. and Verkuilen, J. (2006). Fuzzy Set Theory : Applications in the Social Sciences
(Quantitative Applications in the Social Sciences). SAGE Publications.
38
Song, Q. and Kandel, A. (1999). A fuzzy approach to strategic games. Fuzzy Systems, IEEE Transac-
tions on, 7(6):634642.
Sugeno, M. (1974). Theory of Fuzzy Integrals and its Applications. PhD dissertation, Tokyo, Japan.
Wang, Z. and Klir, G. J. (1993). Fuzzy Measure Theory. Springer.
Williamson, T. (1996). Vagueness (Problems of Philosophy (Routledge (Firm)).). Routledge.
Zadeh, L. (1965). Fuzzy sets. Information and Control, 8:338353.
Zadeh, L. (1975). Fuzzy logic and approximate reasoning. Synthese, 30(3-4):407428.
39
ACKNOWLEDGEMENTS
I would like to thank Dr. Roger Maddux for his help and advice in writing and thinking about
fuzzy logic. I would like to thank Dr. Bergman for his patience as a teacher. I would like to thank Dr.
Tesfatsion for encouraging me to look at fuzzy sets as a tool in economics. I would also like to thank
Dr. Jonathan Smith for giving me the inspiration to explore fuzzy sets and game theory, Dr. Jeremy
Alm for his help in formatting and guidance through the bureaucracies of graduate school, and Mr.
Key-One Chung for constantly providing me with encouraging and supportive words.