Você está na página 1de 7

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

net/publication/326655333

Different Methodologies in Treating Uncertainty

Conference Paper · July 2018

CITATIONS READS

0 20

1 author:

Areeg S. Abdalla
Cairo University
16 PUBLICATIONS   67 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Satisfiability Algorithms View project

Uncertainty weighted of support vector machine View project

All content following this page was uploaded by Areeg S. Abdalla on 29 July 2018.

The user has requested enhancement of the downloaded file.


Proceedings of The 12th International Multi-Conference on Society, Cybernetics and Informatics (IMSCI 2018)

Different Methodologies in Treating Uncertainty


Areeg Abdalla
Department of Mathematics, Faculty of Science, Cairo University
Giza, 12613, Egypt
areeg@sci.cu.edu.eg
ABSTRACT
The probability theory was the first theory to conceive
Uncertainty is unavoidable when dealing with data. The any uncertainty-based information. And for a long period
errors in measurements, limitations of measuring tools, or of time, the probability was the only method to treat un-
imprecise definition of linguistic variables may result in certainty. However the probability concerns with only the
different types of uncertainty. These ambiguities may be chances of existence of an element to a certain set [6],
due to vagueness in data which results from the imprecise
boundaries of data sets; inconsistency that reflects con-
which is not sufficient in treating several uncertainty
flict and contradiction between sets; qualitative descrip- types. In addition to the classical probability theory,
tion of data which sometimes taken by expertise; or some uncertainty-based information is now very well unders-
other type. Ignoring dealing with these types of uncer- tood in fuzzy set theory, possibility theory and others [9].
tainty affects the reliability of research and the validity of Soft computing methodologies are tolerant for impreci-
the results. sion, partial truth, inconsistency and other uncertainty
This article presents three approaches to treat uncertainty types [7]. While hard (crisp) computing is useful in appli-
using fuzzy logic, intuitionistic logic, and neutrosophic cations in which the data is accurately described by ma-
logic and their methodologies in treating these kinds of thematical models. Soft computing techniques mimic
ambiguity. Fuzzy logic and neutrosophic logic are used in human mind in forming conclusions from inaccurate or
building Rule-based Classification Systems. Different approximated data and in forming propositions using
comparisons are presented to illustrate the importance of approximate reasoning. These techniques include –not
choosing the suitable logic to tackle the uncertainty in limited to- fuzzy logic, artificial neural networks, genetic
different data sets. These approaches are applied on six algorithms and machine learning. Soft computing tech-
real world data sets; Iris, Wine, Wisconsin Diagnostic nologies have been applied successfully on data with
Breast Cancer, Seeds, Pima, and Statlog (Heart); which several types of uncertainty. Hybridization between soft
are available on UCI Machine Learning Repository web computing techniques and these logics have shown great
site. The results show that the type of uncertainty in the success in ambiguous data.
data set plays a great role in choosing the appropriate Each of these techniques may tackle only one certain type
logic. of uncertainty. Therefore, hybrid systems of them are
Keywords: Uncertainty, Intuitionistic Fuzzy Logic, used successfully in applications. The advent of very high
Neutrosophic Logic, Classification. performance processors makes it possible to applications
of soft computing to expand fast.
1. INTRODUCTION One of the main advantages of soft computing is that it
opened the door for applications utilizing non-bivalent
In treating data, in any form, one continually comes into logic. Allowing the truth values to take more than the
possession of information that he/ she sums up in general Boolean two values is a very old idea. For example, the
propositions. And it would be a hudge mistake to take Kleene's three-valued logic uses True, False and Unde-
these general propositions, as a guarantee, without any cided [1]. It allows the fact to be undecided rather than to
dispute.Whether this data is recorded by instruments, or be just true or false. There have been other versions of
collected by humans, the information obtained has some the three-valued logic which postulate that some facts
sort of deficiency. The information might be vague, may be intermediate between true and false. But the
incomplete, imprecise or contradictory. Which in turn three-valued logic did not seem to have many applica-
results in different types of uncertainity[9]. In general, tions like other logics. Once we have got familiar with the
these various information deficiencies determine the type idea of a three-valued logic it seems a natural generaliza-
of the associated uncertainty. But, In decision-making tion to have many-valued logic. That allows the facts to
projects, the problems of uncertainty and hesitancy have different degrees of truth. The only problem with
usually turn out to be unavoidable[11]. In clustering[8],it the many-valued logics is that there are many of them.
is usually the case that some data points are hard to Each one is designed to overcome a particular problem.
identify their beloningness to exactely one cluster. We don’t have a unified definition for the logic opera-
Therefore, considering degrees of belongingness to a data tions. In literature, there has been some other generaliza-
point to more than one cluster would solve the problem. tions of fuzzy logic, like interval valued fuzzy logic,
The need for considering uncertainity in application intuitionistic interval valued fuzzy logic, L-fuzzy logic
encouraged mathematicians to put theoritical foundation and intuitionistic L-fuzzy logic. Each is supported by its
for different theories to treat uncertainity. logical operations. And each has found its way to applica-
tions [1].

59
Proceedings of The 12th International Multi-Conference on Society, Cybernetics and Informatics (IMSCI 2018)

The next three sections introduce brief descriptions of fuzzy rules from the knowledge base to generate fuzzy
the Fuzzy Logic, Intuitionistic Fuzzy Logic and the outputs. The fuzzy rule base systems are in the form of
Neutrosophic Logic. A case study on Rule Based “IF-THEN” rules written using linguisitic values. The last
Classification System is discussed in section 4. Results stage is the defuzzification of the fuzzy outputs to crisp
of appling this system using Fuzzy Logic and output.
Neutrosophic Logic are presented in section 5. At the
end, section 6 concludes the results obtained and 3. INTUITIONISTIC LOGIC
discusses some ideas for future work.
The intuitionistic fuzzy Logic was introduced by K. Ata-
2. FUZZY LOGIC nassov in 1986, as one sort of generalization of FL. The
fuzzy logic was very successful in handling uncertainties
Fuzzy logic (FL) deals with imprecise or vague. It facili- arising from vagueness of a fact. Yet, it cannot model all
tates the human common sense reasoning. The proposi- sorts of uncertainties happening in different real observa-
tions in FL can serve as a basis for decision support. It tions specially problems involving imprecise information.
allows the truth values to be any number in the real In defining intuitionistic fuzzy set (IFS), besides the de-
interval [0, 1] instead of taking just two values 0 and 1 in gree of membership ߤ‫ܣ‬ሺ‫ݔ‬ሻ‫א‬ሾͲǡͳሿof each element ‫ܺאݔ‬to
the Boolean logic. This idea of generalization was not a set A, Atanassov considered a degree of non-
new when introduced by Zadeh. There already has been membership ߥ‫ܣ‬ሺ‫ݔ‬ሻ‫א‬ሾͲǡͳሿ, such that ‫ Ͳ ܺאݔ ׊‬൑ ߤ‫ ܣ‬ሺ‫ݔ‬ሻ൅
the many-valued logic, which allows the truth values to ߥ‫ܣ‬ሺ‫ݔ‬ሻ൑ͳǤ ˆሺ‫ݔ‬ሻൌͳǦߤ‫ܣ‬ሺ‫ݔ‬ሻthe IFS is reduced to a fuzzy
be any countable number in [0,1]. However, it was not set. The Intuitionistic fuzzy sets have the ability to handle
untill mid sixties that the truth values take any imprecise information resulted from incomplete or incon-
uncountable real value in [0,1] [9]. sistent data [2]. Later, Atanassov in 1989, introduced the
interval-valued intuitionistic fuzzy logic, which received
The truth value of an atomic proposition p is tv(p)‫[ ג‬0, little attention from the practical point of view.
1] for any proposition in fuzzy logic. tv(p) =1 or tv(p) =
0 mean that p is absolutely true or false, respectively, 4. NEUTROSOPHIC LOGIC
preserving the Boolean truth meanings while tv(p) = 0.65
just means that the truth of p is 0.65. Since the real world Neutrosophy is one of the new theories that deals with
propositions are often only partly true, FL is very uncertainity. It was Introduced by Smarandache in 1995.
representative. The Neutrosophy theory treats uncertainty results from
Opertaions on FL vague, imprecise, incomplete and inconsistent data at the
same time[13]. Therefore the Neutrosophic Logic is a
To form sentences in FL, more proposition may be built very reach logic that generalizes the concept of the classic
from atomoic proposition using the negation, defined as, Boolean Logic, fuzzy logic, intuitionistic fuzzy logic.
tv(~p) = 1- tv(p) and t-norm and t-conorm functions for Table 1 shows this generalization and gives the different
the “and” and “or” operations, respectively; where t- types of uncertainity in which each logic is used.
norm, t-conorm: ሾͲǡͳሿ ൈ  ሾͲǡͳሿ ՜ ሾͲǡͳሿ. The definitions
of the t-norm and t-conorm functions are application and In neutrosophic logic (NL), a propositin has a degree of
person dependent. Researchers define them in many truth (T), a degree of falsity (F) in addition to a degree of
ways, however they have to meet certain conditions. The indeterminacy (I). That is any proposition ൏A൐, is to be
t-norm has to satisty the following boundary, commuta- considered with the negation of the proposition <Anti A>
tivity, monotonicity, and associativity axioms [9]: as well as a spectrum of neutralities <Neut A>. The later
two forms the term <Non A> which keeps the believe of
1. t-norm (a, 1) = a; the proposition balanced and neutralized [3][4] [3][4] [8]
2. t-norm (a,b) = t-norm (b,a); [13]. NL is very close to human thinking and it has been
developed to represent mathematical models which can
3. if b ” c , then t-norm (a, b) ” t-norm (a, c); deal with uncertainty, vagueness, ambiguity, imprecision.
4. t-norm(a,t-norm(b, c))=t-norm (t-norm(a, b), c). That is the knowledge which comes from observations is
mostly characterized by imprecise data, as a result of the
Simillarily any t-conorm has to satisty the following four imprecision of humans or inaccurate measurments.
axioms: Therefore, Neutrosophic Logic is perfect in treating
1. t-conorm (a, 0) = a; problems that involve imprecision, partial truth in data. In
2. t-conorm (a, b) = t-conorm (b, a); addition to that, it can treat incompleteness,
3. if b ” c ,then t-conorm (a,b)” t-conorm (a, c); inconsistency, redundancy, and contradictions in data.
4.t-conorm(a,t-conorm(b,c))= t-conorm(t-conorm(a,b),c). The neutrosophic values T, I, and F, are real subsets of
the non-standard unit interval ]-0,1+[. However, for
t-norm(a,b) = min(a,b) and t-conorm(a,b) = max(a,b) are pratical reasons the non-standard unit interval is replaced
the most used functions for t-norm and t-conorm, by the unit interval [3].
respectively.
Systems that are build using FL start with a fuzzification In intuitionistic fuzzy sets, the incorporated uncertainty -
stage to transform the crisp input value into a fuzzy represented by the falsty degree- is dependent on the
linguistic value. A step that is mandatory done since all degree of belongingness. But, here, the uncertainty in
existing measurments are in crisp numerical values. Then neurosophic fuzzy sets is presented independentely. A
the inference engine takes these fuzzy inputs and calls the neutrosophic set A in X is defined by the truth ( TA ), the

60
Proceedings of The 12th International Multi-Conference on Society, Cybernetics and Informatics (IMSCI 2018)

indeterminacy ( I A ) and falsity ( FA ) membership propositions[9]. And like the classical logic, the most
common inference rule is the generalized modus ponens.
functions. These TA , I A and FA are real standard or
These systems drive their rules directly from numerical

non-standard subsets of ] 0,1+ [ with no restriction on data using soft computing techniques. One of the earilest
their sum, i.e. systems was by Kosko[10]. Then many applications have
been introduced for Fuzzy Rule Based Classification

0 ≤ supTA ( x ) + supI A ( x ) + supFA ( x ) ≤ 3+. Systems[11][5][12]. These fuzzy systems have been
generalized using different logics. Yet, the most recent
Opertaions on NL ones use Neutrosophic Logic. In [3], generalization of a
Like other non-bivalent logics, the connectives are fuzzy rule based classification system to the
defined in many ways[13]. Two functions N-norm and corresponding neutrosophic system, is done using the
N-conorm are used; where N-norm and N-conorm are truth, indeterminacy,and falsity membership funtions. It
turns out that the overall classification accuracy has been
from ሺሿ ିͲǡ ͳା ሾൈሿ ିͲǡ ͳା ሾൈሿ ିͲǡ ͳା ሾሻଶ to ]-0,1+[. improved using the neutrosophic logic, especially, in data
 sets with intereleaved and overlapped classes. An
Any N-norm or N-conorm has to satisfy the four axioms
for the boundary, commutativity, monotonicity, and asso- improvement of this system in [4], in which genetic
ciativity. An example of the logical connectives is to algorithm is used in designing and optimizing the
define them as [1]: knowledge base of the system.

̱ሺ‫ݐ‬ଵ ǡ ݅ଵ ǡ ݂ଵ ሻ ൌ ሺ݂ଵ ǡ ݅ଵ ǡ ‫ݐ‬ଵ ሻ 6. RESULTS AND DISCUSSIONS


N-norm൫ሺ‫ݐ‬ଵ ǡ ݅ଵ ǡ ݂ଵ ሻǡ ሺ‫ݐ‬ଶ ǡ ݅ଶ ǡ ݂ଶ ሻ൯ ൌ ሺ‹ሼ‫ݐ‬ଵ ǡ ‫ݐ‬ଶ ሽ ǡ ͳ െ
Fuzzy Logic and Neutrosophic Logic have been used in
ሺ‹ሼ‫ݐ‬ଵ ǡ ‫ݐ‬ଶ ሽ ൅ ƒšሼ݂ଵ ǡ ݂ଶ ሽሻǡ ƒšሼ݂ଵ ǡ ݂ଶ ሽሻ rule based classification of six different world wide data-
N-conorm൫ሺ‫ݐ‬ଵ ǡ ݅ଵ ǡ ݂ଵ ሻǡ ሺ‫ݐ‬ଶ ǡ ݅ଶ ǡ ݂ଶ ሻ൯ ൌ ሺƒšሼ‫ݐ‬ଵ ǡ ‫ݐ‬ଶ ሽ ǡ ͳ െ sets; Iris, Wine, Wdbc, Seeds, Pima, and Statlog(Heart).
ሺƒšሼ‫ݐ‬ଵ ǡ ‫ݐ‬ଶ ሽ ൅ ‹ሼ݂ଵ ǡ ݂ଶ ሽሻǡ ‹ሼ݂ଵ ǡ ݂ଶ ሽሻ They all have vague boundries between their classes.
Moreover, some of them have intersected areas with
Similar to the FL systems, any NL system strarts with a misplaced objects.
neutrosofication stage to transform the crisp input value As in any real world problem, the data here contains
into a neutrosophic value. Then the inference engine runs imbalanced data sets, where one class (or more)
on neutrosophic based “IF-THEN” rules. And the result represents large number of the examples (majority class)
has to pass through de-neutrosophication step to transfer while the other classes contain just few examples (mi-
the output to crisp output. nority classes) [12].This drives the classifier to be skewed
Table 1: A comparison between the above logic and the towards the majority class. Therefore, it is important
different types of uncertainity they measure -when dealing with real world data sets- to select an ap-
propriate measure of performance. The most common
Intuitionistic method is analysis based on the confusion matrix [12][8].
Fuzzy Neutrosophic
Fuzzy Logic Table 2 shows a confusion matrix for classification of
Logic Logic
two classes A, B, where:
Zadeh, Atanssov, Smarandache, True Positive(TP) is the percentage of correctly classified
Logic
founder 1965 1983 1999 examples in class A,
False Negative(FN) is the percentage of examples classi-
Membership Truth Truth Degree fied in A while it should be in B, False Positive(FP) is the
Truth percentage of examples classified in B while it should be
Degree Indeterminacy
Degree in A, and True Negative(FN) is the percentage of correct-
Degree
Falsity ly classified examples in class B.
Degree Falsity Degree
Table 2: Confusion matrix for Classes A,B
vagueness vagueness, vagueness,
Uncertainity
Type it Imprecision Imprecision,
Actual Class A Actual Class B
treats Inconsistency,
Incompleteness Class A True False
Prediction Positive(TP) Negative(FN)
5. CASE STDUDY Class B False True
RULE BASED CLASSIFICATION SYSTEM Prediction Positive(FP) Negative(TN)

்௉ା்ே
One of the earliest applications of fuzzy logic is Rule Accuracy = ்௉ା்ேାி௉ାிே is the most commonly used
Based Systems. In such systems, the core consists of metric for empirical evaluations but if one of these
fuzzy “IF-THEN” rules. Fuzzy sets are used to form the classes were a majority class, the minority class would
antecedent and the consequent parts of the “IF-THEN” have a little impact [12]. Therefore, three other measures
rules and a logical fuzzy implication is used, as well[11]. have been used [8]:
Reasoning based on fuzzy propositions is referred to as ்௉ ்௉
The Precision = , the Sensitivity = , and the
approximate reasoning. The fundamental components of ்௉ାி௉ ்௉ାிே
்ே
approximate reasoning are these “IF-THEN” fuzzy Specifcity =்ேାி௉

61
Proceedings of The 12th International Multi-Conference on Society, Cybernetics and Informatics (IMSCI 2018)

Fig. 1 shows the total accuracy of the two rule based the center of each class, which reflects inconsistency of
classification systems one is using fuzzy logic and the the data. Fig. 4 shows that the NL system has reached
other is using neutrosophic logic. It is clear that NL gives better results than the FL one.
a more accurate classification than FL for the six data The Seeds dataset has three classes with very vague
sets. Which explains the importance of using the boundries. i.e. the misplased objects lie only on the
indetermincy term for these data sets. The other measures edges. As a result, we can see the precision and specifity
for each data set are explained later. of the first class in the FL is better than the ones using
NL, Fig 5. However, the overall measures were better
using NL.
Pima dataset has two classes, where the two classes are
interleaved and may have overlapped centers. It is
difficult to identify them separately. Similar to the Wdbs,
the indeterminacy term, here, gives better results. The
results of the classification using NL, with respect to the
three measures, is always better than the classification
using FL, Fig. 6.
Statlog(Heart) dataset contains two almost overlapped
classes. It is difficult to identify one from the other. It is
an example of the imprecision and inconsistency that
arises in data. Therefore, there is a big difference between
the results of the NL and the FL, Fig. 7.
Figure 1: Accuracy of Classifing the data sets in Fuzzy
and Neutrosophic Logics 7. CONCLUSION AND FUTURE WORK
Studying well the data set of any project is important.
The Iris dataset contains three classes Iris-Setosa, Iris- And treating the uncertainty in the data is essential and
Versicolour, and Iris-Virginica. The Iris-Setosa is affects the reliability of the results. This article compares
completely separated from the other two classes. And between fuzzy, intuitionistic, and neutrosophic logics and
therefore, the classification using FL has the same result introduces the uncertainty types each logic can handle. A
as the classification using NL for the three measures; case study of Rule Based Classification System -applied
Precision, Sensitivity, and Specificity. However, The to six world wide data sets-, is presented to show the
second and third classes do have vague boundries and importance of choosing the appropriate logic according to
intersected areas. As a result, the Classification using NL the data set. The results showed that using neutrosophic
gives better results than the classification using FL. logic is in general better. Fuzzy logic suits systems that
The Wine dataset contains three classes. Most objects of has only vague data, i.e. the boundaries between the
the first are separated from the other two classes. But classes are unclear. However, the neutrosophic logic -in
some objects in second and the third class are misplaced most measures- gives better results for intersected data
near the center of the wrong class. Here the sets.
indeterminacy term, in NL, plays a good role. The In the future work, building hybrid systems with other
difference between the two classification systems is soft computing techniques seems promissing in extracting
shown in Fig. 3. rules for the rule based classification systems.
The Wdbc dataset has two classes class M, and class B.
There were objects belong to both classes and also nearby

62
Proceedings of The 12th International Multi-Conference on Society, Cybernetics and Informatics (IMSCI 2018)

Figure 2: The measures for the Iris dataset Figure 3: The measures for the Wine dataset

Figure 4: The measures for the Wdbc dataset


Figure 5: The measures for the Seeds dataset

63
Proceedings of The 12th International Multi-Conference on Society, Cybernetics and Informatics (IMSCI 2018)

Figure 6: The measures for the Pima dataset Figure 7: The measures for the Statlog dataset

8. REFRENCES 8. Hassanien, Aboul Ella & Basha, Sameh


H.& Abdalla, Areeg Generalization of
1. Ashbacher, Charles. Introduction to Fuzzy C-Means Based on Neutrosophic
Neutrosophic Logic, American Research Logic, Studies in Informatics and Con-
Press. 2002. trol, ISSN 1220-1766, vol. 27(1),
2. Atanassov, Krassimir T. Type-1 Fuzzy 2018.pp. 43-54.
Sets and Intuitionistic Fuzzy Sets. Algo- 9. Klir, George J. Uncertainty and Infor-
rithms, 10, 106. 2017. mation. John Wiley & Sons, Inc. Neutro-
3. Basha, S. H.& Abdalla, Areeg& Hasse- sophic Logic. Neutrosophy, Neutrosophic
nian, Aboul Ella. NRCS:Neutrophic Rule- Set, Neutrosophic Probability, American
based Classification System, Proceedings Research Press, 4th edition. 2006.
of SAI Intelligent Systems Confe- 10. B. Kosko. Neural Networks and Fuzzy
rence(IntelliSys), Lecture Notes in Net- Systems, Prentice Hall., Englewood NJ.
works and Systems, 15, 2016. 627–639. 1992.
4. Basha, Sameh & Abdalla, Areeg & Has- 11. Mendel, Jerry M. Uncertain Rule-based
sanien, Aboul Ella. GNRCS: Hybrid Clas- Fuzzy Systems. 2nd edition. Springer In-
sification System based on Neutrosophic ternational Publishing. 2017. ISBN 978-3-
Logic and Genetic Algorithm(ICENCO) 319-51369-0.
IEEE.2016. 53-58.10.1109. 12. J. A. Sanz, D. Bernardo, F. Herrera, H.
5. Basha, S.H., Baz, S.M., Hassanien, A.E., Bustince, H. Hagras "A Compact Evolu-
Sahlol, A.T. Neutrosophic Rule-Based tionary Interval-Valued Fuzzy Rule-Based
Prediction System for Assessment of Pol- Classification System for the Modeling
lution on Benthic Foraminifera in Burul- and Prediction of Real-World Financial
lus Lagoon in Egypt. 12th International Applications With Imbalanced Da-
Conference on Computer Engineering and ta", IEEE Transactions on Fuzzy Sys-
Systems (ICCES), 2017. 663-668. tems, vol. 23, no. 4, 2015. 973-990.
6. Brian R. Gaines. Fuzzy and probability 13. Smarandache, F. Neutrosophic set, a ge-
Uncertainty Logics. Information and neralization of the intuitionistic fuzzy sets,
Control, 38, 1978. 154-169. International Journal of Pure and Ap-
7. Ibrahim, Dogan. An overview of soft plied Mathematics, 24, 2005. 287–297.
computing. 12th International Conference
on Application of Fuzzy Systems and Soft
Computing, ICAFS Vienna, Austria,
2016. 29-30.

64

View publication stats