1. Introduction
Fuzzy logic is very successful in dealing with uncertainties related to fuzzy rules; however, this success does not extend to when referring to input–outputs, since they are assumed to be exact, i.e., values in
. In cases in which the fuzzification of inputs derives from choices made by experts and they are unsure about the value to use in to system, fuzzy logic is not always suitable. Another case lies in the situation in which the input data have impreciseness. In order to overcome this structure, interval-valued fuzzy logic has been suggested. This interval logic is a particular case of interval type-2 fuzzy logic [
1,
2,
3,
4,
5,
6].
Interval-valued fuzzy logic has been applied in a wide variety of domains; for example, Melin and Castillo [
7] used it in the context of plant control; Figueroa et al. [
8] used it for non-autonomous robots in the context of a robot football game; Lynch et al. [
9] built an interval control system for large marine diesel engines; Chourasia et al. [
10] developed a new method for assessing fetal health status based on interval type-2 fuzzy logic through fetal phonocardiography; Nguyen et al. [
11] used the wavelet feature in an interval type-2 fuzzy logic system (IT2FLS) to reduce the computation burden and time of IT2FLS; Leow et al. [
12] developed a hybrid of generalized adaptive resonance theory (GART) and an interval type-2 fuzzy logic system algorithm; among many others [
5,
6].
On the other hand, another field of great importance for this work, interval analysis, was developed in order to limit rounding errors and deal with inaccuracies. This field was first presented by Warmus [
13], Sunaga [
14], and Moore [
15]. However, it can be said that interval mathematics and analysis began with the appearance of R. E. Moore’s book,
Interval Analysis, in 1966 [
16]. Moore’s arithmetic is accepted as the standard approach and is called
standard interval arithmetic (SIA). Interval analysis has been applied in several areas, like electrical power systems [
17], mechanical engineering [
18], chemical engineering [
19], artificial intelligence [
20], multi-agent systems [
21], and geophysics [
22].
In interval analysis, there are two most important criteria, namely
correctness (accuracy) and
optimality [
23,
24]. The first criterion establishes that the result of an interval computation must always contain the value of the respective real function (see [
24] (Theorem 3.1)). Although correctness is a desirable property, not every interval method is correct. Santiago et al. [
25,
26] and Bedregal and Santiago [
27] investigate the notion of correctness for interval functions and its impact on some interval topological viewpoints. They call correctness
interval representations since interval entities (algorithms and intervals) are seen as entities that represent exact entities (functions and numbers). An interval function,
F, is said to represent a real function,
f, whenever it satisfies the following property:
and
. The second criterion establishes that the resulting interval of a computation should not be greater than necessary, which is captured by the notion of
canonical interval representation.
In order to solve the algebraic incompatibility between the real arithmetic and Moore arithmetic, Lodwick [
28,
29] defined a new interpretation for intervals called
constraint intervals. With this new approach, Lodwick proposed an alternative to Moore arithmetic on intervals in order to have
,
if
and the distributive property. This new arithmetic, called
constrained interval arithmetic (CIA), is an extension of Moore’s interval arithmetic, in the sense that they coincide in the case where there is no variable dependence and are distinct when there are dependencies. In this case, the CIA presents a smaller width interval, thus improving the overestimation of Moore’s arithmetic [
30]. Actually, CIA has been used to solve several mathematical problems, such as in [
30,
31,
32].
In order to solve, or mitigate, the problems related to overestimation and loss of information when working with intervals, in this article, we present a new approach to fuzzy logic with interval values, in which interval operators preserve some of the main algebraic properties, the overestimation problem is mitigated, and there is no loss of information. These operators were defined by using the constrained intervals introduced by Lodwick [
28]. In what follows, it is described how this study was performed and also the contributions that were studied or compared with this new approach.
The new operators, called
constrained interval operators, in the proposed new approach to interval-valued fuzzy logic are very close to Moore’s correctness. In fact, they satisfy a new correction, which was called here
constraint interval correctness, which is as efficient as Moore’s correction, in the sense of not losing information; however, they are less demanding and, in turn, work with less irrelevant information. The main properties of this operator have been verified about two orders, namely Kulisch–Miranker and Moore order. It has been found that this new approach guarantees the extension of many algebraic properties and that its fuzzy operators are extended to their respective constrained interval operators when considering the order of Kulisch–Miranker as in [
27,
33].
This paper is organized as follows:
Section 2 recalls some definitions and concepts used throughout the text in order to provide a self-contained document.
Section 3 proposes a new way of making interval-valued fuzzy logic in which the operators are called constrained interval operators. The last chapters includes the conclusions and bibliography.
3. Constraint Interval Fuzzy Operators
In this section, following the same approach as Lodwick [
28], we introduce the
constrained interval operators, in which the two lambdas, including
and
, are considered when there is no dependence. The
constrained interval operators were investigated in terms of correction, algebraic properties, and orders. In addition, a comparative study was carried out between these operators and an operator well known in the literature.
Definition 11. Let in any interval. Then,
- (i)
A function such that will be called a constraint function associated to A.
- (ii)
Associated to the interval A, we define the increasing convex constraint function : by means - (iii)
Associated to the interval A, we define the decreasing convex constraint function : by means
Observe that the same interval can be determined by different constraint functions. However, each interval is associated with a single increasing convex constraint function and two different intervals determine two different increasing convex constraint functions (i.e., there is a bijection between the space of intervals and the space of increasing (decreasing) convex constrained functions).
Remark 2. Given , for each , there exists such that ; that is, each is the image of by the function .
An analogous development can be carried out using the decreasing convex constraint functions. However, in this article, we will use increasing convex constraint functions.
Definition 12. The operator : , called the constraint interval operator, associated with the function g: is defined by, for all intervals ,
- 1.
If for any , then is - 2.
If for and for some with if , then is
We denote and as the endpointsandrespectively. It can be easily noticed that when there is no dependence on the input intervals, we have a different parameter for each interval; that is, different intervals are traversed at different levels (, for ). In the case of dependency, we have a single parameter for equal intervals, so they are traversed at the same level (). Therefore, the resulting interval has less irrelevant information, and pieces of information are not lost; see Remark 6.
Example 3. - 1.
Let N: be the standard fuzzy negation; that is, , so, , - 2.
Let T: be the minimum t-norm; that is, , so, , since and are increasing.
- 3.
Consider the aggregation function g: given by . For all , since for all , we have for all and . So,
The following proposition ensures that a constrained interval operator coincides with its best representation whenever g is increasing.
Proposition 1. Let be an increasing function. Then, , Proof. Given any
, since the function
is increasing for all
, we have
. So,
for all
with
. Therefore,
Thus, . □
Remark 3. An analogous result can be obtained if g is decreasing. In this case, we would have for all .
Remark 4. We can conclude from the previous results that, if g is monotone, is correct (Moore’s correctness) and coincides with the best interval representation of g; that is, (see Equation (1)). Therefore, given an increasing function g, we have that . However, some fuzzy logical operators like implications are neither increasing nor decreasing, in which case we have situations like:
Example 4. Consider the Łukasiewicz implication and as non-degenerate. Then, since = 1,
and , in which case we have that is not correct, since . However, the fuzzy implications satisfy the property presented in the following proposition: Proposition 2. Let I be a fuzzy implication. Then, for all .
Proof. By Definition 12, for all
,
So, since
we have, by Equation (
1),
. Now, given any
with
, we have that
Therefore, . □
Remark 5. Considering any fuzzy implication I, given with , we have that . So, by [40] (Theorem 6.2), . Example 5. Consider the Fodor implication For and , we have that and . Then, since for all ,for all . So, Therefore,since and . Now, for , we have , sosince . Therefore, . Note that is not correct since . However, in this case, elements of type , with , are not relevant (they are irrelevant); see Remark 6. We can conclude, then, that this new way of defining interval operators does not always coincide with the best representation. Also, it does not necessarily satisfy Moore’s correctness (see Example 4). However, this does not cause harm; see the following remark:
Example 6. (Moore and constraint interval correctness.) It is possible to observe that the loss of the usual correctness (Moore’s correctness) occurs only when we have variable dependence. However, this loss will not cause problems since every element of an interval operates with itself and this operation is contained in the resulting interval, i.e., the correctness with respect to a fixed element is maintained. So, everything “thrown away” is irrelevant. In other words, if the interval A represents a certain number a, we have thatthen, something like with , and , is not relevant, since we are only interested in operations with the element in which A represents; that is, operations with the element a. This kind of correctness is called here constraint interval correctness. The following proposition shows that important properties such as symmetry, (IP), and (NP) are preserved by the constrained interval operators.
Proposition 3. Let be a function. Then,
- (i)
If g is symmetric, then is also symmetric;
- (ii)
If g satisfies identity property (IP), then satisfies (I-IP);
- (iii)
If g satisfies left neutrality property (NP), then satisfies (I-NP).
Proof. Indeed,
- (i)
By the symmetry of
g, we have
for any
.
- (ii)
For any
,
since
g satisfies (IP).
- (iii)
For any
,
since g satisfies (NP). □
Proposition 4. There is function in which the exchange principle (EP) is satisfied but does not satisfy (I-EP).
Proof. Consider the Łukasiewicz implication
and the intervals
and
. From Table 1.4. in [
36],
I satisfies (EP). Now, since
, we have
So,
and, for all
,
And, since
for all
,
Therefore, . □
However, Proposition 5 guarantees that (EP) is preserved by the operator when g is continuous.
Lemma 1. Let g: be a function. Then,. In addition, if g is continuous, then equality holds. Proof. Given
, there are
such that
. As
, then, by Remark 2, there exists
such that
. We conclude that
In addition, if
g is continuous, then, given
we have
for some
. Since
, then, by the continuity of
g, there are
such that
. We conclude that
Therefore, the equality holds when g is continuous. □
Remark 7. Analogous to the previous lemma, we have that And, if g is continuous, then the equality holds.
Proposition 5. Let be a continuous function. Then:
- (i)
If g is associative, then is also associative;
- (ii)
If g satisfies the exchange principle (EP), then satisfies (I-EP).
Proof. Indeed,
- (i)
For all
,
Now, since
g is associative, we have that
Therefore, is associative.
- (ii)
For all
,
Therefore, since
g satisfies (EP),
Therefore, satisfies (I-EP). □
The binary operator g does not need to be continuous in order for its extension to preserve associativity; it suffices that g is increasing (or decreasing). See the following proposition:
Proposition 6. Let be an increasing function. Then, g is associative if and only if is also associative.
Proof. As
g is increasing, using Proposition 1, for all
,
So, for all
,
Analogously, using Proposition 1, for all
,
So, since
g is associative, we conclude that, for all
,
Conversely, for each
,
for all
; then,
and, therefore,
. Following a similar reasoning, we have that
and
So, , since is associative. □
Section 3.1 and
Section 3.2 show that the logical connectives studied are extended to their respective intervals logical connectives in relation to the order
, but the same does not occur with Moore’s order ⊴.
3.1. Properties of Some Interval Fuzzy Connectives with Respect to Order
This section will show that, considering the order of Kulisch–Miranker (see definition below), fuzzy connectives, negation, t-norms, and fuzzy implications are extended to their respective intervals logical connectives, and their main properties are maintained.
Definition 13. The Kulisch–Miranker order is defined, for all , by: The following proposition shows the connection that exists between the Kulisch–Miranker order and the functional counterpart of the intervals.
Proposition 7. Given , the following are equivalent:
- (i)
;
- (ii)
for all ;
- (iii)
for all .
Proof. Indeed, let for some ; then, and . So, since , we have and for all . Thus, for all , i.e., for all . Conversely, let for all ; then, in particular, for and , we obtain and , respectively. So, .
Straightforward, once . □
Proposition 8. There is binary operator g that satisfies (OP), whereas does not satisfy (I-OP) with respect to .
Proof. Take
such that
with
. So, since
g satisfies (OP),
So, since . □
The following result shows that the restricted operators preserve the fuzzy implications with respect to the order of Kulisch–Miranker.
Proposition 9. Let I be a fuzzy implication. Then, is an interval fuzzy implication with respect to .
Proof. Indeed, (II1) a for
, trivially
. On the other hand, if
with
, then
and
, with
or
; so, given any
,
and
, since
I satisfies (I1). So,
and
, respectively, since, by Remark 5,
. Therefore,
. The proof of (II2) is analogous. (II3) Since
I satisfies (I3),
, so
(II4) Since
I satisfies (I4), then
. So,
(II5) Since I satisfies (I5), . So, . Therefore, is an interval fuzzy implication. □
Considering the order Kulisch–Miranker, the following result shows that the restricted operators preserve the fuzzy negations.
Proposition 10. Let be a fuzzy negation. Then, is an interval fuzzy negation with respect to .
Proof. Now, given such that , we have that and . So, and , since N is antitonic. By Remark 3, and . Therefore, . We conclude that is an interval fuzzy negation with respect to . □
Proposition 11. Let be a fuzzy negation. N is strong if and only if, is a strong interval fuzzy negation.
Proof. Indeed, by Remark 3,
. So, since
N is strong, for all
,
Conversely, for each
,
for all
, so
and then
. Following a similar reasoning, we have
and, since
is strong, we have that
. □
Considering any fuzzy continuous negation, the following proposition guarantees that contrapositive properties are preserved for this negation.
Proposition 12. Let N be a continuous fuzzy negation and I be a fuzzy implication. Then,
- (i)
I satisfies (CP) with respect to N iff satisfies (I-CP) with respect to ;
- (ii)
I satisfies (LCP) with respect to N iff satisfies (I-LCP) with respect to ;
- (iii)
I satisfies (RCP) with respect to N iff satisfies (I-RCP) with respect to .
Proof. (i) For all
,
In fact,
given
, there are
such that
. By Remark 3,
for
, so, since
N is continuous and
, there exists
such that
for
. Thus,
, since
I satisfies (CP) with respect to
N. So,
. Now,
given
, there exists
such that
. Since
I satisfies (CP) with respect to
N,
. So, since
, there exists
such that
for
. Therefore,
. We conclude by Equation (
3) that
. Conversely, for each
, we have that
for all
and
. Then, for all
, we have
and
so, since
,
so
.
(ii) For all
,
In fact, for
, given
, there are
such that
. By Remark 3,
; so, since
N is continuous and
, there exists
such that
. Thus,
, since
I satisfies (LCP) with respect to
N. Now, since
, there exists
such that
; so,
. The opposite direction,
, follows analogously. Conversely, since, for each
,
for all
and
, then, for all
, we have
and
so, since
,
so
.
(iii) It follows similarly to the previous item. □
The t-norm is also preserved by the constrained interval operators; see the proposition below:
Proposition 13. Let T be a t-norm. Then, is an interval t-norm with respect to .
Proof. The symmetry and associativity are straightforward from Propositions 3 and 6, respectively. Now, to show that is increasing, consider such that ; then, and . So, for all , by the monotonicity of the t-norm, and . So, by Proposition 1, we have that and , respectively. Therefore, . From symmetry, whenever . Finally, by the boundary condition of T, for all , so . □
The following section will provide a study with respect to Moore order similar to the one performed with respect to the Kulisch–Miranker order, analyzing their similarities and differences.
3.2. Properties of Constraint Interval Fuzzy Connectives with Respect to Order ⊴
This section addresses the main properties of the constrained interval fuzzy operators with respect to the order of Moore [
16], defined below.
Definition 14. Moore order is defined by Proposition 14. Given , the following are equivalent:
- (i)
;
- (ii)
or for all ;
- (iii)
or for all .
Proof. (i) ⇔ (ii) Indeed, given such that , then or . So, it is enough to check the case where , since the other follows straight. Therefore, since and for all , then for all . Conversely, if , then, by definition, . Now, if for all , then, in particular, for all , so ; thus, .
(ii) ⇔ (iii) Straightforward, once . □
The following proposition ensures that the order property is preserved by its constrained interval operator.
Proposition 15. Let g be a binary function on . If g satisfies (OP), then satisfies (I-OP) with respect to ⊴.
Proof. Given
such that
, we have
or
for all
. If
, then
for all
. So, by Definition 12,
since
g satisfies (OP). Now, if
for all
, then
for all
since
g satisfies (OP). Therefore,
. Conversely, let
such that
; thus,
So, for all ; therefore, for all since g satisfies (OP). By Proposition 14, we conclude that . □
Unlike Proposition 9, considering Moore’s order, the fuzzy implications are not preserved by constrained interval operators; see the following proposition:
Proposition 16. There are fuzzy implications I such that is not an interval fuzzy implication with respect to ⊴.
Proof. Consider the Łukasiewicz implication
and the intervals
,
and
in
. In this case,
since
; however,
. Indeed, by Remark 5,
whenever
with
, so
,
,
, and
. Therefore,
and
So, , i.e., does not satisfy (II1). □
The following proposition shows that, analogous to Proposition 10 with respect to the order of Kulisch–Miranker, the fuzzy negations are also extended to their respective constrained interval operator with respect to Moore’s order.
Proposition 17. Let be a fuzzy negation. Then, is an interval fuzzy negation with respect to ⊴.
Proof. Indeed, analogous to Proposition 10,
and
. Now, consider
for some
. Then,
or
for all
. The first case is trivial. Now, if
for all
, then
for all
since
N is decreasing. So, in particular,
, i.e.,
Therefore, by Definition 14, . □
Proposition 18. Let be a fuzzy negation. N is strong if and only if is a strong interval fuzzy negation.
Proof. Analogous to Proposition 11. □
The contrapositive properties are again preserved for any fuzzy negation.
Proposition 19. Let N be a fuzzy negation and I be a fuzzy implication. Then,
- (i)
I satisfies (CP) with respect to N iff satisfies (I-CP) with respect to ;
- (ii)
I satisfies (LCP) with respect to N iff satisfies (I-LCP) with respect to ;
- (iii)
I satisfies (RCP) with respect to N iff satisfies (I-RCP) with respect to .
Proof. The proof follows analogous to Proposition 12. □
Unlike Proposition 13, t-norms are not preserved by the constrained interval operator considering the order ⊴.
Proposition 20. There are t-norms T such that is not an interval t-norm with respect to ⊴.
Proof. Consider a t-norm
T such that
whenever
and
. Take
and
in
as not degenerate such that
; in this case,
. Thus, for all
that are not degenerate, we have that
since
. So,
since, by Proposition 1,
. And, also, since
is not degenerate, then
, so
. Thus,
We conclude, by Equations (
5) and (
6), that
. Therefore,
is not increasing. □
It is easy to find in the literature examples of t-norms such that whenever and ; for example, the product t-norm, . Conversely, the t-norm (Łukasiewicz’s t-norm) does not satisfy this condition, since if and , then , so .
4. Final Remarks
A well-known problem in interval analysis is that, when there are variable dependencies, the resulting width of the interval is overestimated. This overestimation occurs because each of the two occurrences of one variable is treated as an independent variable.
In order to solve the problem of overestimation, using a reasoning similar to Lodwick’s [
28], we have defined the
constrained interval fuzzy operators by using the constrained intervals. By Definition 12, we can see that, when there are two or more occurrences of the same interval, it is traversed at the same level. With this, it is possible to state that there is no loss of information; see Remark 6. In view of what was presented, a new notion of correction, called
constrained interval correctness, was suggested, in which it is as effective as Moore’s correction and less demanding.
The main objective of this paper is to introduce this new approach to interval-valued fuzzy logic, in which it is possible to soften the problem of overestimation when there are dependencies of variables, and in such a way so that the information of interval connectives is maintained.
In addition, the operators were investigated in terms of correction, algebraic properties, and orders. Furthermore, the authors also proposed a comparative study between the constrained interval operators and the best interval representation , in which it was shown that . Thus, it is possible to conclude that the operator improves the overestimation problem and, of the properties studied, it only did not satisfy the order property in relation to Kulisch–Miranker order.