An approach to Fuzzy clustering of the iris petals by using Ac-means

An approach to Fuzzy clustering of the iris
petals by using Ac-means
Nicolás Enrique Salgado Guitiérrez1
, Sergio Andrés Valencia Ramírez2
and José Soriano Méndez3
1Engineering department, District University FJDC, Bogota D.C., 110231, Bogota, Colombia
nesalgadog@correo.udistrital.edu.co
2Engineering department, District University FJDC, Bogota D.C., 110231, Bogota, Colombia
savalenciar@correo.udistrital.edu.co
3Engineering department, District University FJDC, Bogota D.C., 110231, Bogota, Colombia, Advisor
Abstract This paper proposes a definition of a fuzzy partition element based on the homo-
morphism between type-1 fuzzy sets and the three-valued Kleene algebra. A new clustering method
based on the C-means algorithm, using the defined partition, is presented in this paper, which will
be validated with the traditional iris clustering problem by measuring its petals.
Keywords Fuzzy, Partition, K3 algebra, homomorphism, cluster, C-means.
1 Introduction
The iris clustering problem has been approached from different perspectives, which perform
such clustering under certain common characteristics represented by metrics.
These perspectives are described with many types of algorithms. One of the most commonly
used is C-means, which consists of creating fuzzy partitions according to the formal definition of
Ruspini[1]. This algorithm is an extension of the K-means algorithm that determines the member-
ship values given the mean distances between elements.
There are multiple definitions of fuzzy partition according to the application, most of these
definitions are posed in the universe of fuzzy sets, so that observing the relationship between fuzzy
sets and the algebra K3, the possibility of formulating an alternative way from the point of view
of this algebra becomes evident, in order to make use of its properties, this will be denoted as
A-partition.
An A-partition is a definition formulated from the three-valued K3 Kleene algebra and has appro-
priate characteristics for a fuzzy clustering as the C-means algorithm and it is expected that by
handling a finite set of states, a fuzzy set with better interpretability will be obtained because a
finite set can simplify the definition of linguistic labels.
Initially some theoretical concepts and examples related to partitions and finite algebras are de-
scribed,we are going to have a look through the different definitions proposed in the literature;
finite algebras particularly the K3-algebra; well-formed formulas and regularity; then the homo-
morphism between fuzzy sets and the algebra described; the crossing set is defined and based on
this, the A-partition is proposed; reduction methods are used on the proposed definition and then
the clustering algorithm is designed (using the C-means as as a reference framework) applying a ho-
momorphism and A-partition detailing its steps and a corresponding experimentation, comparison,
discussion and conclusions are made.
1
International Journal on Soft Computing (IJSC) Vol.12, No.2/3/4, November 2021
DOI: 10.5121/ijsc.2021.12401 1

2 Background information
The partition element definitions are an extension of the crisp definition where two important
concepts are required, covering (the union of a family of subsets forms the totality of the set
to which they belong) and disjoint sets (two sets have no elements in common). Over time the
paradigm of sets has developed from the crisp set theory from which the concept of crisp partition
is derived.[2][3][4].
A family {Ai : i ∈ I} of non-empty classical subsets of A is a partition if and only if
• Ai ∩ Aj = ∅ with i 6= j (Ai and Aj are disjoint)
•
[
i∈I
Ai = A (partition elements have covering over the whole set A).
With the advent of fuzzy set theory proposed by Zadeh [5] has given rise to several definitions
of partition element The first definition is attributed to Ruspini[1] and it is done based on the
membership functions of the fuzzy sets of a universe X. This definition is as follows.
A finite family {Ai : i ∈ I} of nonempty fuzzy subsets of the universal set X is a partition if
and only if for all x ∈ X is satisfied.
X
i∈I
µAi
(x) = 1
where µAi
(x) is the membership function of the subset Ai.
Bezdek and Harris [6], Dumitrescu [7] and Ovchinnikov [8] applied the fuzzy sets operations.
For this purpose they expressed the intersection as a t-norm 4 and the union as its respective dual
t-conorm 5 defined as:
x 5 y = 1 − (1 − x) 4 (1 − y) for x, y ∈ [0, 1].
A finite Family {Ai : i ∈ I} of non-empty fuzzy subsets of the universal set X is a partition if
and only if, for all x ∈ X, the following is true
• µAi (x) 4 µAj (x) = 0 with i 6= j (Ai and Aj are disjoint)
• 5
i∈I
µAi
(x) = 1 (the partition elements have coverage over the entire universal set).
An alternative way is the definition given by Butnariu[9] who modifies the meaning of disjoint
sets.
A finite family {Ai : i ∈ I} of non-empty fuzzy subsets of the universal set X is a partition if
and only if, for all x ∈ X, the following is true
• (
j
5
i=1
µAi (x)) 4 µAj+1 (x) = 0 with j = 1, 2, ...n − 1 ({Ai : i ∈ I} are disjoint)
• 5
i∈I
µAi (x) = 1 (the partition elements have coverage over the entire universal set).
Markechová [10]. A finite family {Ãi : i ∈ I} of non-empty fuzzy subsets of the universal set X
is called a fuzzy w-partition 1
if and only if, for all x ∈ X, the following is satisfied
• (
_
i∈I
µAi
(x))0
≤
_
i∈I
µAi
(x)
• (µAi
(x) ∧ µAj
(x)) ≤ (µAi
(x) ∧ µAj
(x))0
1Also called weak fuzzy partition
2
2

where ∨, ∧ and 0
are the minimum t-norm, the maximum t-norm and the standard complement.
Chakraborty and Das [11] proposed another definition of fuzzy partition based on the concept of
disjoint set.
A family {Ai : i ∈ I} of non-empty fuzzy subsets of the set A is a fuzzy partition of A if and only
if
• Ai ∩ Aj = ∅ with i 6= j (Ai and Aj are disjoint)
•
[
i∈I
Ai = A (the partition elements have coverage over the entire set A.)
where ∩ and ∪ are defined as the infimum and supremum, respectively.
Bhakat and Das [12]. A family {Ai : i ∈ I} of nonempty fuzzy subsets of the set A ∈ X is a q-fuzzy
partition of A if and only if, for all x ∈ X, it is satisfied that
• µAi (x) + µAj (x) ≤ 1, for i 6= j (Ai and Aj are disjoint sets)
•
[
i∈I
Ai = A (the partition elements have coverage over the set A)
Montes, et al. introduced the concept of δ-ε-partition by making use of the concept of α-cut (noted
as α
A) of the set A.
A family of fuzzy sets {Ai : i ∈ I} is a δ-ε-fuzzy partition of A, with 0 ≤ ε < δ ≤ 1 if and only if
• α
(Ai ∩ Aj) = ∅ with i 6= j (Ai and Aj are disjoint sets)
• α
(
[
i∈I
Ai) = α
A (the partition elements Ai have coverage over the whole set α
A)
for all α ∈ (ε, δ), where ∩ and ∪ are described by any t-norm and any t-conorm, respectively.
Let “=ε” an equality in terms of alpha cuts such that:
A1 =ε A2 ⇐⇒ α
A1 = α
A2 ∀ α ∈ (ε, ε0
)
A family of fuzzy sets {Ai : i ∈ I} is a fuzzy ε-partition of A, with 0 ≤ ε ≤ 0.5 if and only if
• Ai ∩ Aj =ε ∅ with i 6= j (Ai and Aj are disjoint sets)
•
[
i∈I
Ai =ε A (the partition elements Ai have coverage over the whole set ε
A)
where ∩ and ∪ are described by any t-norm and any t-conorm, respectively. In [13] we extend
the definition of δ-ε-partition for finite interval fuzzy sets.
These partition definitions have a restriction due to their covering, which limits the transition
between partition elements in a fuzzy way, since it prevents the existence of regions in which not
necessarily
[
i∈I
Ai = A, which implies that for all x there must exist at least one partition element
whose membership function that is equal to the set to be partitioned.
3 Preliminaries
Definition 1 (Homomorphism) [14] [15] Given two algebraic structures A = (A, ◦1, ..., ◦k) and
B = (B, ∗1, ..., ∗k) of the same type, a function T between them
T : A → B
is said homomorphism if and only if:
T(◦(a1, ..., an)) = ∗i(T(a1), ..., T(an))
For all i = 1, ..., k and a1, ...an ∈ A
3
3

One of the most important concepts in fuzzy set theory is the concept of α-cut which basically
allows to make a horizontal cut over the fuzzy sets and is defined as:
Definition 2 (α-cut and strong α-cut) [2][3] Given a fuzzy set A defined on X and any number
α ∈ [0, 1] the α-cut α
A and the strong α-cut α+
A are the classic sets:
α
A = {x | µA(x) ≥ α}
α+
A = {x | µA(x) > α}
Example 3.1 the figure 3.1 shows the Gaussian fuzzy set A from the universe X and two α-cut
over that set, α1 and α2 generating the sets α1
A and α2
A.
X
0
1
Membership
µA(x)
α1
]
[ α1
A
α2
]
[ α2
A
Figure 3.1. Gaussian fuzzy set A and its alpha cuts α1 and α2
Definition 3 (Fuzzy partition proposed by Ruspini) [1] A finite first-order fuzzy partition
of a set X is a finite family of fuzzy subsets {A1, A2, ..., An} of the universe X that satisfy the
condition
n
X
i=1
µi(x) = 1 for all x ∈ U.
Definition 4 (Lattice) [2][14][16][17]
A lattice L = hL; ≤i is a partially ordered set in which, for every pair of elements a, b ∈ L,
there exists a supremum element sup{a, b} and an infimum element inf{a, b}. In other words, L is
a lattice if sup{H} and inf{H} exist for every subset H of L.
The supreme element sup{a, b} and the smallest element inf{a, b} are usually denoted by the logical
operation disjunction a ∨ b and conjunction a ∧ b respectively and satisfy the following properties
for all a, b ∈ L:
a ∨ a = a
a ∧ a = a
(3.1)
a ∨ a = b ∨ a
a ∧ b = b ∧ a
(3.2)
(a ∨ b) ∨ c = a ∨ (b ∨ c)
(a ∧ b) ∧ c = a ∧ (b ∧ c)
(3.3)
a ∨ (a ∧ b) = a
a ∧ (a ∨ b) = a
(3.4)
Two other properties relating ∨ and ∧ with the partial order relationship ≤ are:
a ≤ b ⇔ a ∨ b = b
a ≤ b ⇔ a ∧ b = a
4
4

All reticule L satisfies the distributivity inequalities for all a, b, c ∈ A given by:
a ∨ (b ∧ c) ≥ (a ∨ b) ∧ (a ∨ c)
a ∧ (b ∨ c) ≤ (a ∧ b) ∨ (a ∧ c)
Definition 5 (Distributive lattice ) A Distributive lattice L = hL; ≤i is a lattice satisfying the
distributivity properties for binary operations ∨ and ∧ for all a, b, c ∈ L.
a ∨ (b ∧ c) = (a ∨ b) ∧ (a ∨ c)
a ∧ (b ∨ c) = (a ∧ b) ∨ (a ∧ c)
(3.5)
Definition 6 (bounded lattice) A bounded lattice L = hL; ≤i is the one where the elements zero
and unity exist in L, denoted by 0 and 1, such that for every element a ∈ L the identity properties
are satisfied for the operations ∨ and ∧.
a ∨ 0 = a
a ∧ 1 = a
(3.6)
From the equation (3.6) it follows that 1 is the maximum element of L and the element 0 is the
minimum element. Both are unique in L. From the property of identity and the property of
absorption we deduce the property of absorption by 1 and 0.
a ∨ 1 = 1
a ∧ 0 = 0
(3.7)
Definition 7 (Complemented lattice) A complemented lattice L = hL; ≤i is a bounded lattice,
in which for every element a ∈ L there is an element a0
∈ L called the complement of a such that
the complementation properties are satisfied.
a ∨ a0
= 1
a ∧ a0
= 0
(3.8)
The De Morgan, Kleene and Boolean algebras are defined below. These are lattice structures and
are therefore governed by the standard t-norm, t-conorm and complement operations. You can
consult [2][4][18][19] for a more detailed study.
Definition 8 (De Morgan Algebra) A De Morgan Algebra M is a algebra hA; ∨, ∧,
0
, 0, 1i, where hA; ∨, ∧, 0, 1i is a bounded distributive lattice. That is, it satisfies the properties (3.1)-
(3.7). Binary disjunction operations ( ∨ ) and conjuntion ( ∧ ) together with the unary involute
complement operation ( 0
) satisfy the property of involution and De Morgan’s laws for all a, b ∈ A.
(a0
)0
= a (3.9)
(a ∨ b)0
= a0
∧ b0
(a ∧ a)0
= a0
∨ b0 (3.10)
Definition 9 (Kleene’s algebra) A Kleene’s algebra K = hA; ∨, ∧, 0
, 0, 1i is a De Morgan algebra
that satisfies the following inequality for all a, b ∈ A:
a ∧ b0
≤ b ∨ b0
(3.11)
This inequality can also be written as (a ∧ a0
) ∧ (b ∨ b0
) = a ∧ a0
or as (a ∧ a0
) ∨ (b ∨ b0
) = b ∨ b0
.
These equations are also called Kleene’s laws.[20].
5
5

Definition 10 (Boolean algebra) A Boolean algebra B is a Kleene algebra that satisfies the
complementation properties for all a, b ∈ A.
a ∨ a0
= 1
a ∧ a0
= 0
(3.12)
Definition 11 (Three-valued algebra)
An algebra K3 = hK3; ∨, ∧, 0
, 0, 1i, with K3 = {0, u, 1}, is a three-element algebra where
hK3; ∨, ∧, 0, 1i is the lattice shown in the figure 3.2.
0
u
1
Figure 3.2. Lattice diagram of the algebra K3.
K3 is a set of three elements satisfying the total order relation 0 u 1. The conjunction,
disjunction and complement operations are described in the table 3.1. K3 is a finite Kleene algebra.
[2][4][21][19][22]. Therefore, it fulfills the properties (3.1)-(3.11). It does not satisfy the property
(3.12) since u ∨ u0
= u ∨ u = u 6= 1 and u ∧ u0
= u ∧ u = u 6= 0.
Table 3.1. Tables of operations in K3. (a) Disjunction, (b) conjunction and (c) complement.
∨ 0 u 1
0 0 u 1
u u u 1
1 1 1 1
∧ 0 u 1
0 0 0 0
u 0 u u
1 0 u 1
a a0
0 1
u u
1 0
A formula is a concatenation of constants, variables, symbols and operators that obey certain
rules to make sense.
Definition 12 (Well-formed set of formulas F) [4][23][21][24] Let V = {x, y, z, ...} be a finite
set of variables. F is the smallest set of concatenations of the constants 0 and 1, the elements of
V and and the operation symbols ∨, ∧ and 0
that satisfies
1. 0 ∈ F, 1 ∈ F and if x ∈ V then x ∈ F.
2. If f, g ∈ F then f ∨ g ∈ F,f ∧ g ∈ F and f0
∈ F
The elements of the set F are called well-formed formulas or simply formulas.
The first item determines the initial elements of the set F, from which it follows that V ∪{0, 1} ⊆ F.
The second item determines that known formulas can be concatenated to construct new formulas,
making use of the operation symbols ∨, ∧ and 0
. F is an infinite set [4].
In F the value of formulas is not set, that is, different strings have no connected meaning. For
instance the strings x ∨ y and y ∨ x are different. The value of the formulas is introduced by an
value algebra A. Some value algebras can be L M, K, B, K3 and B2.
The concept of regularity is an important aspect for the treatment of tables in K3 algebra when
we want to obtain formulas from them and in this paper it is a useful tool for the determination
of partitions based on K3.
Definition 13 [25] Let the set V = {0, u, 1}. Let ∝ be a partial order relation such that 0 ∝ u,
1 ∝ u and a ∝ a, a ∈ V , 0 and 1 are not comparable to each other. This relation can be extended to
V n
. For two elements A = (a1, . . . , an) and B = (b1, . . . , bn) of V n
, A ∝ B if and only if bi ∝ ai
for all values of i and is said B to be less or equally ambiguous than A.
6
6

Definition 14 (Regularity) [25] A function F is called regular if it fulfill the following conditions:
• Condition 1: if F is a ternary function, that is, it can be represented by a well-formed formula.
• Condition 2: if F(A) ∈ 0, 1,then F(B) = F(A) for all B that B ∝ A.2
• Condition 3: if B ∝ A then F(B) ∝ F(A), it is called a ternary logic function.
In [25] it is proved that the three conditions are equivalent. By considering two algebras:
hF(X), ∪, ∩,0
i the fuzzy sets and their standard operations and hK3, ∧, ∨,0
i the algebra and its
conjunction, disjunction and complement operations, there is a function f
f : F(X) → K3
which is a homomorphism for each of its operations. This function is stated in [27][4] given the
similarities in their structure and properties.
This homomorphism f is described as a mapping of the values {0, u, 1} to the membership
levels of the fuzzy sets.
The function f is determined by the following encoding:
Trivalent coding of operating regions:[27] given a ε ∈ [0, 1] defined by the designer, such
that 0 ε ζ, where ζ is the fixed point of the logical operation “no” (or complement (0
)). Each
fuzzy set A performs a partitioning of the input universe X in the following way:
a Coding with 0: for those x ∈ X tales que µA(x) ∈ [0, ε)
b Coding with u: for those x ∈ X such that µA(xi) ∈ [ε, ε0
]
c Coding with 1:for those x ∈ X such that µA(x) ∈ (ε, 1]
In this step it is possible to perform a trivalent encoding of all possible regions of operation by
means of a trivalent logic truth table. (0, u, 1), based on the division described above. Each coding
will correspond to a row in the table.
X
0
1
Membership
µA(x)
ε0
ε
0
u
1
u
0
Figure 3.3. Assigning Kleenean values to a fuzzy function
4 Development
Studying several definitions of fuzzy partition, present in the literature, a common pattern
is found in the determination of the properties of covering of sets, where the union operation is
used, and disjunct sets (or disjunctivity), which makes use of the intersection operation. We seek
to propose such properties in the 3-element Kleenean algebra K3, in which the conjunction and
disjunction operations are analogous to the standard intersection and union operations of fuzzy
sets given the homomorphism T. To achieve this goal, the following concept is proposed which
will be used to determine the value of the variables in K3 that are the product of the mapping of
the evaluated fuzzy sets into the regions where two or more membership functions present in the
interval [ε, ε0
] intersect.
2This condition is an extension of the definition present in [26]
7
7

4.1 Crossing set
Definition 15 (Crossing set) Let there be two non-empty fuzzy sets Ai and Aj (with i 6= j)
subsets of the universal set X. Let Жij, a subset of the universal set X, be defined as follows:
Жij = {x | ε ≤ µAi
(x) ≤ ε0
and
ε ≤ µAj
(x) ≤ ε0
}.
(4.1)
Жij is called a crossing set and Ai and Aj are called neighboring sets if Жij 6= ∅.
Example 4.1 Let the sets A1, A2 and A3 be subsets of the universal set X, with their respective
membership functions µA1
(x), µA2
(x) and µA3
(x) shown in the figure 4.1. The crossing sets Ж12
and Ж23 are defined as the elements of the universe X belonging to the shaded regions, as indicated
by the indicator functions µЖ12
(x) and µЖ23
(x). Since the sets Ж12 and Ж23 are non-empty, it
is said that A1 is a neighbor of A2 and A2 in turn is a neighbor of(no estoy seguro de esta frase)
A3. The set Ж13 is empty, therefore A1 and A3 are not neighbors.
0 1 2 3 4 5 6 7 8 9 10
1
Elements of the universe X (x)
Membership
µA1
(x) µA2
(x) µA3
(x)
ε
ε0
µЖ12
(x) µЖ23
(x)
µЖ13
(x)
Figure 4.1. Crossing sets for the example 4.1
The definition 4.1 allows one to identify the regions of the universe X where sets map to u
according to the algebra K3. These regions can be expressed in terms of the alpha-cuts and strong
alpha-cuts of the membership functions in addition to which the above regions can be employed to
determine whether a family of sets is a partition. The above properties are stated in the following
two propositions.
Proposition 1 The crossing set can be described as:
Жij = ε
Ai ∩ (ε0)+
Ai ∩ ε
Aj ∩ (ε0)+
Aj (4.2)
Proof. The alpha cuts of the set Ai are defined as follows:
ε
Ai = {x | ε ≤ µAi
(x)}
(1−ε)+
Ai = {x | ε0
µAi
(x)}
The definition is similar for Aj.
By rewriting equation 4.2 using the above expressions and making use of the definition of the
intersection operation between sets, the following results
ε
Ai ∩ (ε0)+
Ai ∩ ε
Aj ∩ (ε0)+
Aj = {x | ε ≤ µAi
(x) and
ε0
≥ µAi
(x) and
ε ≤ µAj
(x) and
ε0
≥ µAj
(x)}
ε
Ai ∩ (ε0)+
Ai ∩ ε
Aj ∩ (ε0)+Aj = {x | ε ≤ µAi
(x) ≤ ε0
and
ε ≤ µAj
(x) ≤ ε0
}.
8
8

This shows that the expressions 4.1 and 4.2 are equivalent.
Example 4.2 Let A1, A2 and A3 be the sets of example 4.1. Figure 4.2 shows how the crossing
sets are constructed by means of the alpha cuts, for which we make use is made of their indicator
functions. According to expression 4.2 the crossing sets are defined as follows
• Ж12 = ε
A1 ∩ (ε0)+
A1 ∩ ε
A2 ∩ (ε0)+
A2
• Ж23 = ε
A2 ∩ (ε0)+
A2 ∩ ε
A3 ∩ (ε0)+
A3
Figures 4.2b and 4.2c show that the crossing set Ж12 can be reduced to the expression Ж12 =
(ε0)+
A1 ∩ (ε0)+
A2; in turn, figures 4.2a and 4.2c allow us to visualize that the crossing set Ж23 can
be reduced to the expression Ж23 = ε
A2 ∩ ε
A3.
0 1 2 3 4 5 6 7 8 9 10
1
Membership
level
µ(εA1)(x) µ(εA2)(x) µ(εA3)(x)
(a) Alpha cuts
0 1 2 3 4 5 6 7 8 9 10
1
Membership
µ(ε0)+
A1
(x)
µ(ε0)+
A2
(x)
µ(ε0)+
A3
(x)
(b) Alpha cut complements
0 1 2 3 4 5 6 7 8 9 10
1
Membership
µЖ12
(x) µЖ23
(x)
(c) Crossing sets
Figure 4.2. Representation of the crossing sets of the example 4.2 by alpha cuts.
ЖT is the union of all possible crossing sets {Жij : i, j ∈ I} with i 6= j formed by a family of
fuzzy sets {Ai : i ∈ I}. In this way
ЖT =
[
i,j∈I
Жij with i 6= j
The homomorphism T : F(X) → K3 between fuzzy sets and the three-valued Kleene algebra
applied to a fuzzy set A is expressed as follows
T(µA(x)) = a (4.3)
Taking into account the above notions, we propose a definition that seeks to satisfy the classical
and fuzzy behaviors considering the sets Жij.
Definition 16 (A-partition)
Let be a fuzzy set A subset of the universal set X and Π = {Ai : i ∈ I} a finite family of
nonempty subsets of A; let ЖT be the union of all nonempty crossing sets generated by Π ; let
T : F(X) → K3 the homomorphism 4.3 such that T(µAi
(x)) = ai ∀ i ∈ I. Π is said to be a fuzzy
partition if and only if it is satisfied, ∀x ∈ X:
9
9

Covering (A-covering):
d =
_
i∈I
ai =

a ∀ x ∈ ЖT
u ∀ x ∈ ЖT
(4.4)
Disjoint sets (A-disjunctiveness):
dc =
n−1
_
j=1
n
_
i=j+1
(aj ∧ ai) =

0 ∀ x ∈ ЖT
u ∀ x ∈ ЖT
(4.5)
where n is the number of elements of Π .
The property 4.4 indicates that the disjunction of the variables associated to all the elements of
the family Π is equal to the variable associated to the set A (a ∈ K3) for the elements that belong
to ЖT and is equal to u for the elements belonging to ЖT . In turn, the property 4.5 indicates
that the disjunction of all possible conjunctions of two different variables associated with elements
of family Π is equal to 0 for elements belonging to ЖT and is equal to u for elements belonging
to ЖT . In the definition 16 it is evident that, both in the covering and in the disjoint sets, the
disjunction and conjunction of the variables of K3 is equal to u for the elements belonging to the
non-empty crossing sets.
However, the proposed definition contains both kleenean variables (a, {ai : i ∈ I}) and classical
sets (crossing sets Ж). We want to find an equivalence with purely algebraic terms and for this
purpose we look for a way to determine crossing sets through kleenean variables coming from the
homomorphism 4.3, using d, dc, a and the kleenan properties of these variables with which it is
desired to pose the conditions equivalent to the definition 16 in terms of tables and formulas. When
d and dc are equal to u, they guarantee the existence of non-empty crossing sets, since they imply
the existence of two or more sets mapping to that value for the same elements.
To observe the elements that meet the conditions of covering and disjoint sets, these are re-
stricted to the values that can be presented in a truth table. Since {Ai : i ∈ I} is a subset of A it
must be satisfied d ≤ a and it is evident dc ≤ d so values that do not satisfy these conditions are
discarded. Finally the inequality must be satisfied:
dc ≤ d ≤ a (4.6)
The variables d and dc describe tables, these together with the inequality (4.6), which eliminates
unnecessary combinations, can be used to determine the conditions for the 4.4 and 4.5 as shown in
the table 4.1. According to the membership to a crossing set, the covering condition and disjoint
sets are determined. The elements that satisfy both conditions are considered a A-partition.
Table 4.1. Definition’s table of the A-partition properties. This table shows the possible combinations taking
into account the restrictions of the inequality 4.6, in the columns Ж, A-covering, A-disjointness and A-partition X
indicates that the conditions are met, and 7 indicates that the conditions are not met.
a d dc Ж A-covering A-disjointness A-partition
0 0 0 7 X X X
u 0 0 7 7 X 7
u u 0 7 X X X
u u u X X X X
1 0 0 7 7 X 7
1 u 0 7 7 X 7
1 u u X X X X
1 1 0 7 X X X
1 1 u 7 7 7 7
1 1 1 7 X 7 7
The conditions of the definition 16 can be rewritten through an equality taking into account
the relation of the elements involved, since they can be interpreted as Kleenean truth tables and
in turn these can be seen as formulas in case they are regular.
10
10

Proposition 2 Let a fuzzy set A be a subset of the universal set X and a finite family Π = {Ai :
i ∈ I} of nonempty subsets of A; let T : F(X) → K3 be the homomorphism 4.3. Π is said to be a
fuzzy A-partition if and only if it is satisfied, ∀x ∈ X:
d = a ∧ dc
0
(4.7)
For all dc, d and a satisfying inequality 4.6.
Proof. Table 4.2 shows the truth values of a, d and dc where inequality 4.6. If we take the values
of the rows in which we have an A-partition (X) and with them we establish the disjunctive normal
form for a and dc, setting as output the variable d, ignoring the other combinations, we obtain
d = (a ∧ a0
∧ dc
0
) ∨ (a ∧ a0
∧ dc ∧ dc
0
) ∨ (a ∧ dc ∧ dc
0
) ∨ (a ∧ dc
0
).
Using some properties of the K3 algebra the simplification of the above formula is performed.
d = ((a ∧ a0
) ∨ (a ∧ a0
∧ dc) ∨ (a ∧ dc) ∨ a) ∧ dc
0
d = ((a ∧ a0
) ∨ a) ∧ dc
0
d = a ∧ dc
0
This gives the formula 4.7.
Table 4.2. Determination of the disjunctive normal form (DNF) for d respectively to a and dc for values satisfying
inequality 4.6
a dc d A-partition DNF
0 0 0 X
u 0 0 7
u 0 u X a ∧ a0
∧ dc
0
u u u X a ∧ a0
∧ dc ∧ dc
0
1 0 0 7
1 0 u 7
1 u u X a ∧ dc ∧ dc
0
1 0 1 X a ∧ dc
0
1 u 1 7
1 1 1 7
When the set A is the universal set, the variable a takes the value of 1; then the condition 4.7
is reduced to
d = dc
0
∧ a = dc
0
∧ 1
d = dc
0
(4.8)
The formula 4.7 is essentially a truth table so it is just necessary fulfill the four combinations
allowed by the inequality 4.6, besides this formula implies the disjunction of the elements d is
equivalent to the conjunction of the variable associated to the partitioned set a and the disjunction
of the conjunction of the all possible pairs of variables dc this is quite useful given that it simplifies
the definition using a logical rule.
Example 4.3 Let A be a fuzzy set, a subset of the universal set X whose membership function is
shown in figure 4.3. Let Π = {A1, A2, A3} be a family of subsets of A whose membership functions
are shown in figure 4.3b. Let a and {a1, a2, a3} be the variables obtained by applying homomor-
phism 4.3 on A and Π respectively. Π fulfills the conditions to be considered A-partition of A.
The total crossing set is defined as ЖT = Ж12 ∪ Ж23 since the set Ж13 = ∅. With a value of
ε defined, the variables have the following behavior in the non-empty crossing sets: for x ∈ Ж12
take the values of a = 1, a1 = u, a2 = u and a3 = 0, hence
d = a1 ∨ a2 ∨ a3 = u ∨ u ∨ 0 = u and
dc = (a1 ∧ a2) ∨ (a1 ∧ a3) ∨ (a2 ∧ a3) = u.
11
11

For x ∈ Ж23 the variables take the value from a = u, a1 = 0, a2 = u and a3 = u, then
d = a1 ∨ a2 ∨ a3 = 0 ∨ u ∨ u = u
dc = (a1 ∧ a2) ∨ (a1 ∧ a3) ∨ (a2 ∧ a3) = u.
For elements x not belonging to ЖT the expressions d = a and dc = 0 are fulfilled.
Table 4.3 shows the truth values of the variables a and {a1, a2, a3}, as well as the dc and d values.
It can be seen that this table satisfies proposition 2, since d is reproduced by the formula 4.7.
Table 4.3. Truth table for the example 4.3.
a1 a2 a3 a dc d
0 0 0 0 0 0
u 0 0 u 0 u
1 0 0 1 0 1
u u 0 1 u u
0 1 0 1 0 1
0 u 0 u 0 u
0 u u u u u
0 0 u u 0 u
(a) (b)
Figure 4.3. (a)Membership function µA(x) of the set to be partitioned A. (b)Partitioning elements
Π = {Ai : i ∈ I} of A and their crossing sets Ж.
Proposition 3 Let Π = {Ai | i ∈ I} a finite family of X fuzzy sets, then, using the standard
operations.
1. Crisp partition. if Π is a crisp partition of a subset A ⊆ X, then Π is a fuzzy partition
based on K3 for all ε ∈ (0, 0.5].
2. Fuzzy partitioning according to Ruspini. If Π is a fuzzy partition according to Ruspini
of a subset A ⊆ X, then Π is a A-partition for ε → 0+
.
3. Fuzzy partitioning according to Bezdek and Harris, Dumitrescu and Ovchinikov.
If Π is a fuzzy partition according to [6], one has that it is also a A-partition for any ε.
4. Fuzzy partitioning according to Butnariu. If Π is a fuzzy partition according to But-
nariu of the universal set X, then it is also a A-partition of the universal set X for any ε
with x ∈ ЖT .
5. Fuzzy partitioning according to Markechová If Π is a fuzzy partitioning according to
Markechová then is also a A-partition with ε → 0.5.
6. Fuzzy partitioning according to Chakraborty and Das If Π is a partition of A according
to Chakraborty and Das, then is also a A-partition of A for any ε with x ∈ ЖT .
7. Fuzzy partitioning according to Bhakat and Das If Π is a fuzzy partitioning according
to Bhakat and Das, then is also a A-partition for any ε.
12
12

8. ε
ε
ε-partition If Π is a ε-partition with α ∈ (ε, ε0
) of A, then is also a A-partition of A with
0 ε 0.5.
Proof.
1. Since the values of the membership functions (indicator) of the crisp partitions only have the
values {0, 1} it is noticeable that there are no crossing sets, so the mapping on K3 must meet
the following conditions d = a and dc = 0, which are equivalent to the conditions established
in the crisp definition, since the algebra between crisp sets is a Boolean lattice.
2. Given the restriction on Ruspini partitions, multiple cases can be obtained;
• when ε0
µAk
≤ 1 ⇒ {µA1
, ..., µAn
} − {µAk
} ∈ [0, ε)
• when ε ≤ µAK
≤ ε0
⇒ {µA1
, ..., µAn
} − {µAk
} ∈ [0, ε0
]
• when 0 ≤ µAk
ε ⇒ {µA1 , ..., µAn } − {µAk
} ∈ (ε, 1]
For k = 1, 2, ..., n.
with ε → 0+
• when µAk
→ 1−
⇒ {µA1
, ..., µAn
} − {µAk
} ∈ 0 ⇒ d = 1 and dc
0
= 1
• when 0 µAK
1 ⇒ {µA1
, ..., µAn
} − {µAk
} ∈ (0, 1) ⇒ d = u and dc
0
= u
• when µAk
→ 0+
⇒ {µA1
, ..., µAn
} − {µAk
} ∈ (0, 1] ⇒ d = 0 and dc
0
= 0 or d =
u and dc
0
= u
for k = 1, 2, ..., n. always satisfying in all cases the condition d = dc
0
.
3. replacing 4 with ∧ we have
µAi (x) ∧ µAj (x) = 0
n
_
i=1
µai (x) = 1
Then using the properties of the homomorphism T we have
ai ∧ aj = 0
_
i∈I
ai = 1
Then, given the definition 16 they are equivalent if there are no crossing sets.
4. Using the properties of homomorphisms
_
i∈I
µAi = 1 ←→
_
i∈I
ai = 1 and
n−1
_
j=1
n
_
i=j+1
(aj ∧ aj) (
j
W
i=1
µAi )
V
µAj+1 =
0 ←→ (
j
W
i=1
ai)
V
aj+1 = 0. The first expression implies x ∈ ЖT according to the condi-
tion 4.4. Using the properties of kleene algebra the second expression can be rewritten as
(
j
W
i=1
ai)
V
aj+1 =
j
W
i=1
(ai ∧ aj+1) = 0 which for each value of j takes a value of 0, that is to say
that all possible disjunctions take the value of 0 fulfilling the condition 4.5.
5. Taking into account the conditions proposed in the definition of weak partition and the
standard operations, equivalent conditions can be deduced to make it a partition according
to Markechová.
•
_
i∈I
µAi (x) ≥ 0.5
• µAi ∧ µAj ≤ 0.5 for all i 6= j
Given these conditions it is observable that for both to be fulfilled the intersections between
fuzzy sets must become 0.5 fulfilling the definition of A-partition for ε → 0.5.
13
13

6. The fuzzy sets with the operations supremum and infimum describe a kleene algebra, there-
fore, the sets can be treated as variables and the operations of infimum and supremum as
the operations of conjunction and disjunction respectively by reducing the expressions to
formulas 4.4 and 4.5.
7. Given the condition
[
i∈I
Ai = A we limit the second condition µAi
(x)+µAj
(x) ≤ 1 for the case
µAi
∈ (0, 1) requires that µAj
(x) ≤ 1 − µAi
(x) but µAj
(x) 6= 1, 0 that is, for that point there
is no cover so this definition is equivalent to the classical definition if the union is defined as
the supremum.
8. By mapping the sets Ai and A we obtain the variables in K3 a and ai that satisfy the following
implications
•
[
i∈I
Ai =ε A ⇒
_
i∈I
ai = a ∀ a = u
• Ai ∩ Aj =ε ∅ ⇒ ai ∧ aj = 0
5 AC-means algorithm
5.0.1 Proposed solution
The following is the A-partition clustering method using a similar structure of the c-means
algorithm.
Based on c-means, it is decided to implement a similar algorithm by changing the way the fuzzy
sets are calculated, so the generated sets are not c-partitions (partitions according to Ruspini) but
A-partitions. It should be clarified that the partitions according to Ruspini are A-partitions under
the appropriate ε, so the algorithm proposed by Bezdek also converges to A-partitions, but its sets
are more limited in their form and conditions.
The design of the macro scale algorithm looks like this:
Beginning
Step 1 Step 2 Step 3 Step 4 Step 5
finish
criteria
finish
yes
no
Figure 5.1. Flowchart of the c-means algorithm with A-partitions.
CLUSTERING ALGORITHM WITH A-PARTITIONS
The steps described below are shown as hexagons in figure 5.1
STEP 1: Pseudo-random fuzzy sets are generated within the minimum and maximum values
of the input data universe.
STEP 2: The centers are calculated in exactly the same way as the conventional c-means using
the equation cj =
PD
i=1(µm
ij xi)
PD
i=1(µij)m
, 1 ≤ i ≤ c(taken from the c-means algorithm [28]).
14
14

STEP 3: Kleenean values are assigned to ensure the generated sets lie in the three different
possible assignments {0, u, 1}; these values are determined by taking into account the distance
between the elements and the centers as follows:
1. A radius is established around the center of each of the sets, which will indicate the elements
that must belong to “1”; due to the condition of disjoint sets, within this zone the rest of the
membership values must map to “0”. If these n-spheres come into contact, it is established
that the elements that are inside two or more belong to crossing sets and must map to “u”.
2. All elements that are not inside the spheres will be determined as belonging to crossing sets
and should map to “u” for the two sets whose centers are closest.
For more clarity on this process see example 5.1.
Note: This method is not unique and other A-partition sets can be obtained otherwise.
STEP 4: Once the kleenean mappings of the fuzzy sets are obtained, sets that fulfill these
mappings are looked for, taking into account a given ε. It is evident that the family of sets that
fulfill these conditions is infinite due to T is not bijective, therefore the method chosen is one
of many and tries to ensure that the generated sets are continuous. To determine the degree of
membership of an element, functions that depend on the distance between the element in question
and the fuzzy set to which it belongs are used. A different function is implemented for each of the
possible kleenean values.
• For ai = 0
The membership function is established by taking into account the distance of the element x
from the radius and the largest of the distances of the centers where the kleenean assignment
for x is u.
µAi
(dis) = ε

1
1 − Am
(
dis
Hg
− 1) + 1

(5.1)
• For ai = u
The membership function is established taking into account the distance of the element x
from the center to be evaluated and the smallest of the distances of the centers where the
kleenean assignment for x is u.
Two possible cases must be taken into consideration, first that the evaluated element is outside
of an n-sphere or inside of 2 or more spheres, in the first case we use the equation
µAi (dis) = r
2ε − 1
r − Hp
(
Hp
dis
− 1) + ε (5.2)
in the second case:
µAi
(dis) =
2ε − 1
1 −
Hp
r
Hp
dis
+ ε −
2ε − 1
1 −
Hp
r
(5.3)
• For ai = 1
The distance dependent membership function is established µAi (dis) In this case, the range
of the function is known to be the distance dependent membership function is established
[ε0
, 1] and since it is known that the element in question is inside the sphere, it follows that
0 dis r then a parabolic conditioner is made between dis and µAi
such that µAi
(0) = 1
and µAi
(r) = ε0
using the equation.
µAi
(dis) =
−ε
r2
dis2
+ 1 (5.4)
15
15

STEP5 5: The objective function is calculated in exactly the same way as the conventional
c-means using the equation Jm(µ, c) =
D
X
j=1
N
X
j=1
(µij)m
||xi − cj||2
(objective function).
Example 5.1 (Assignment of kleenean variables in R2
) Be a group of centers {C1, C2, C3, C4}
and the elements {x1, x2, x3, x4, x5} and an established radius of r = 2.5 as shown in the figure 5.2.
y1
0
y
2
c1
c2
c3
c4
x1
x2
x3
x4
x5
d1
d2
d3
d4
Figure 5.2. Example of assigning kleenean values to elements of an iteration
Using the criteria established in STEP 3, the kleenean assignments of the elements xi are
determined.
Table 5.1. Kleenan assignments
data x1 x2 x3 x4 x5
axi




u
u
u
0








u
0
u
0








1
0
0
0








0
u
u
0








0
u
0
u




Where: axi =




aC1
(xi)
aC2 (xi)
aC3 (xi)
aC4
(xi)



 indicating what is the kleenean value after being operated by the
homomorphism of the fuzzy set whose center is Cn.
For the particular case of x5 it is remarkable that this element is outside all the n-spheres, which
indicates that the distance to the centers must be taken into account to determine the kleenean
values to be taken. In figure 5.2 it is evident that the two shortest distances are d2 and d4, hence
for the values of aC2(x5) and aC4(x5)a value of u is assigned and in the rest 0 as is visible in Table
5.1.
Note : The ways of determining the sets of A-partitions are diverse and this one is chosen
because of its simplicity in determining their kleenean variables centered on cj.
Example 5.2 (Determination of continuous fuzzy sets) Given three fuzzy sets with their
respective centers c1 = (0.3, 0.75), c2 = (0.5, 0.25), c3 = (0.8, 0.7), these sets are assigned a fuzzy
value taking into account a radius r = 0.2 and a kleenean value assignment as shown in step 3.
16
16

y1
0
y
2
c1
c2
c3
(a) (b)
Figure 5.3. (a)Centers and radii described in this example. (b)Membership function µc1 (y1, y2) with ε = 0.25
considering as input elements a matrix of values uniformly covering the universe X.
6 Results
Given the cluster design, it is required to vary two parameters, the ε and the radius, which
determine the shape of the arrays used. For this purpose, a matrix of experiments varying both
parameters will be made in order to observe their optimization landscape in this particular problem.
The experiment matrix used was 21×21 in which ε and r were varied in the intervals [0.01, 0.49]
and [0, 20] respectively. For each of the combinations, one hundred repetitions of the experiment
were made with the purpose of finding the best test subject taking into account the objective
function and the corresponding validations. In order to validate the performance of the cluster,
different criteria will be used, including some internal validation metrics. The internal validation
metrics are those validations that do not require data other than the data provided to the cluster
and seek to validate the cohesion and separation between the data; the cohesion seeks to measure
the distance between members of the same class, since a member of a cluster should be as close as
possible to the other members of the same cluster and the separation measures the distance between
the different clusters, since the clusters should be widely separated between them. There are several
approaches to measure this distance between clusters, either distance between the closest member,
distance between the most distant members or the distance between the centroids. Criteria taken
from [29].
• FSSW =
k
X
i=1
X
x∈Ci
µix||mi − xi||2
(Optimization criterion: maximum)
• FSSB =
k
X
j=1
njµj ∗ ||cj − x||2
(Optimization criterion: maximum)
• Hartigan: H = log(
SSB
SSW
)(Optimization criterion: minimum)
17
17

Table 6.1. Comparison of the internal validation results obtained with the traditional c-means method and Ac-
means; in gray those criteria that were higher compared to the other algorithm.
Used
cluster
FSSW FSSB Hartigan
c-means 21889.49 2492.323 -1.715
Cluster_n=2
Ac-means 1362.7 2550.9 -17.394
c-means 6050.571 2220.719 -1.099
Cluster_n=3
Ac-means 2044.1 2651.9 -11.7076
c-means 4161.423 2047.096 -0.741
Cluster_n=4
Ac-means 2725.5 2661 -0.7331
Table 6.2. The parameters with which the best results were obtained in each of the criteria established with the
different cluster_n tested
Cluster_n FSSW FSSB Hartigan
2
r = 0.1 r = 15 r = 9.04
ε = 0.01 ε = 0.01 ε = 0.01
3
r = 0.1 r = 10.53 r = 12.765
ε = 0.01 ε = 0.01 ε = 0.01
4
r = 0.1 r = 9.04 r = 10.53
ε = 0.01 ε = 0.01 ε = 0.01
(a) (b)
(c)
Figure 6.1. (a)Centers located by the algorithm with cluster_n=2. (b)Centers located by the algorithm with
cluster_n=3. (c)Centers located by the algorithm with cluster_n=4.
18
18

7 Discussion
In general Ac-means results stand out since there are a greater number of criteria in which
it is better than the traditional c-means method. But it should be noted that the Ac-means has
more input parameters that allow it to fit better the problem. This is not entirely convenient since
having an additional parameter implies a difficulty in the search.
By observing the result of the experiments it is clear that in order to choose the input radius of the
c-means with A-partitions similar dimensions to existing clusters should be chosen for best results.
It is possible to use this proposal in some application, with the objective of comparing it with
other methodologies present in the literature by making use of some previously established metrics.
Having an extra parameter allows the algorithm to be more flexible for the problem, however it
increases the difficulty in finding the optimal parameters. Since it was determined that the best ε
are those close to 0 is not necessary that this is defined by the user, so finally the only variable input
parameter of the algorithm is the value of the radius with which you want to make the clusters.
The FSSW is considered a separation metric while the FSSB is considered a cohesion metric,
as it is clear in table 6.1 in both results the Ac-means is better but the last criterion, Hartigan,
is a mixed criterion that uses a cohesion measure and a separation measure, in which it is evident
that the traditional method stands out.
8 Conclusions
A definition of fuzzy partition element was proposed which, being based on K3, allows a wide
flexibility since it contains other definitions given particular conditions.
Thanks to the definition proposed in terms of K3 it was possible to use the disjunctive normal
form, reducing this definition to a formula in terms of a, d and dc, using simplification methods
under the distribution and absorption properties of the kleenean algebra.
The use of A-partitions in the implementation performed, in general terms, allowed obtaining
better results compared to traditional methods, in most of the performance metrics evaluated,
which can be seen in table 6.1.
This method allows a categorization of the classes using the set of elements {0, u, 1}, which
makes it possible to obtain a clear and simple interpretation of the classification obtained, since
instead of evaluating a degree of membership in an interval of infinite values, it can be covered in
three possible algebraic values, facilitating the determination of linguistic labels.
REFERENCES
[1] E. H. Ruspini, “A new approach to clustering,” Information and Control, vol. 15, pp. 22–32,
1969.
[2] H. T. Nguyen and E. A. Walker, A First Course in Fuzzy Logic, C.
bibinitperiod H. /CRC, Ed. Chapman Hall /CRC, 2006.
[3] G. J. Klir and B. Yuan, Fuzzy Sets and Fuzzy Logic Theory and Applications, P. Hall/PTR,
Ed. Prentice Hall/PTR, 1995.
[4] O. Salazar and J. Soriano, Método de simplificación de fórmulas por medio de álgebras finitas.
UD editorial, 2018.
[5] L. Zadeh, “Fuzzy sets,” Information and Control, vol. 8, pp. 338–353, 1965.
[6] J. C. Bezdek and J. D. Harris, “Fuzzy partitions and relations; an axiomatic basis for clus-
tering,” Fuzzy Sets and Systems, vol. 1, pp. 111–127, 1978.
[7] D. Dumitrescu, “Fuzzy partitions with the connectives t∞, s∞,” Fuzzy Sets and Systems,
vol. 47, pp. 193–195, 1992.
19
19

[8] S. Ovchinnikov, “Similarity relations, fuzzy partitions, and fuzzy orderings,” Fuzzy Sets and
Systems, vol. 40, pp. 107–126, 1991.
[9] D. Butnariu, “Additive fuzzy measures and integrals i,” Journal of Mathematical Analysis
and Applications, vol. 93, pp. 436–452, 1983.
[10] D. Markechová, “The entropy of fuzzy dynamical systems and generators,” Fuzzy Sets and
Systems, vol. 48, pp. 351–363, 1992.
[11] M. Chakraborty and M. Das, “On fuzzy equivalence i,” Fuzzy Sets and Systems, vol. 11,
pp. 185–193, 1983.
[12] S. Bhakat and P. Das, “Q-similitudes and q-fuzzy partitions,” Fuzzy Sets and Systems, vol. 51,
pp. 195–202, 1992.
[13] P. Quirós, P. Alonso, I. Díaz, and S. Montes, “On δ-ε-partitions for finite interval-valued
hesitant fuzzy sets,” International Journal of Uncertainty, Fuzziness and Knowledge-Based
Systems, vol. 24, pp. 145–163, 2016.
[14] S. Burris and H. P. Sankappanavar, A Course in Universal Algebra. S. Burris and H.P.
Sankappanavar, 2012.
[15] J. Ignjatović, M. Ćirić, and S. Bogdanović, “Fuzzy homomorphisms of algebras,” Fuzzy Sets
and Systems, vol. 160, pp. 2345–2365, 2009.
[16] G. Grätzer, Lattice Theory: Foundation. Birkhäuser, 2011.
[17] G. Birkhoff, Lattice Theory (Colloquium Publications (Amer Mathematical Soc)). American
Mathematical Society, 1940.
[18] L. H. Kauffman, “De morgan algebras - completeness and recursion,” Proceedings of the
Eighth International Symposium on Multiple-Valued Logic. 82-86, 1978.
[19] D. E. Muller, “Treatment of transition signals in electronic switching circuits by algebraic
methods,” IRE Transactions on Electronic Computers, vol. EC-8, pp. 401–401, 1959.
[20] M. Mukaidono, “Algebraic structures of interval truth values in fuzzy logic,” in Proceedings
of 6th International Fuzzy Systems Conference. pp. 699-705., IEEE, 1997.
[21] M. Gehrke, C. Walker, and E. Walker, “A mathematical setting for fuzzy logics,” International
Journal of Uncertainty, Fuzziness and Knowledge-Based Systems., vol. 05, pp. 223–238, 1997.
[22] M. Gehrke, C. L. Walker, and E. A. Walker, “Normal forms and truth tables for fuzzy logics,”
Fuzzy Sets and Systems, vol. 138, no. 1, pp. 25–51, 2003.
[23] J. Font and V. Verdu, “Abstract characterization of a four-valued logic,” in Proceedings. The
Eighteenth International Symposium on Multiple-Valued Logic., IEEE Comput. Soc. Press,
1988, pp. 389–396.
[24] K. Maes and B. D. Baets, “Facts and figures on fuzzified normal forms,” IEEE Transactions
on Fuzzy Systems, vol. 13, pp. 394–404, 2005.
[25] Mukaidono, “Regular ternary logic functions—ternary logic functions suitable for treating
ambiguity,” IEEE Transactions on Computers, vol. C-35, no. 2, pp. 179–183, Feb. 1986. doi:
10.1109/tc.1986.1676738.
[26] S. C. Kleene, Introduction to Metamathematics. North Holland Publishing Co., Amsterdam,
1952.
[27] J. D. Chacón, J. Soriano, and O. Salazar, “Sistema de inferencia difusa basado en relaciones
booleanas y kleeneanas con combinador convexo,” Revista Ingenieria, 2017.
[28] J. C. Bezdek, R. Ehrlich, and W. Full, “FCM: The fuzzy c-means clustering algorithm,”
Computers Geosciences, vol. 10, no. 2-3, pp. 191–203, Jan. 1984. doi: 10.1016/0098-
3004(84)90020-7.
[29] Métricas para la validación de clustering, https://docplayer.es/71694389-Metricas-para-la-
validacion-de-clustering.html, consulted at 22/03/20.
20
20

An approach to Fuzzy clustering of the iris petals by using Ac-means

More Related Content

What's hot (15)

Similar to An approach to Fuzzy clustering of the iris petals by using Ac-means (20)

Recently uploaded (20)

An approach to Fuzzy clustering of the iris petals by using Ac-means