Professional Documents
Culture Documents
Time
Time
a r t i c l e i n f o a b s t r a c t
Article history: As an important tool for information fusion, evidence theory has been widely applied in many areas.
Received 19 September 2017 Unlike spatial evidence combination, temporal evidence combination is sequential and dynamic, which
Revised 22 June 2018
calls for new temporal evidence combination rule. In this paper, temporal evidence is combined based
Accepted 23 June 2018
on relative reliability evaluation and evidence discounting. We first recall the method of evidence reli-
Available online 2 July 2018
ability evaluation based on intuitionistic fuzzy multiple criteria decision making (ERE-IFMCDM). Based
Keywords: on ERE-IFMCDM method, relative reliability factors of evidence sources in neighboring time nodes can
Dempster–Shafer evidence theory be obtained. Then, according to evidence discounting and Dempster’s combination rule, a method for
Intuitionistic fuzzy sets temporal evidence combination based on the relative reliability factor is developed. Numerical examples
Temporal information fusion and simulation demonstrate that the proposed method is time sensitive, which can reflect the dynamic
Reliability evaluation feature of temporal information fusion. Moreover, the proposed method can deal with conflict in tempo-
ral fusion much better, which can enhance the anti-interference performance of the fusion system. The
fusion framework proposed in this paper provides a dynamic perspective of the fusion of temporal un-
certain information. It is helpful for the structural design of intelligent fusion system based on multiple
temporal-spatial information.
© 2018 Elsevier Ltd. All rights reserved.
1. Introduction mation can be fused effectively. So evidence theory has been ap-
plied widely in uncertain reasoning and uncertain information pro-
Dempster–Shafer evidence theory, shorten as evidence theory cessing. The application of evidence theory has been applied in
or D-S theory, has been proposed for decades (Dempster, 1967; many areas such as pattern recognition (Liu, Pan, Dezert, Han, &
Shafer, 1976). In evidence theory, uncertain information is de- He 2018; Liu, Pan, Dezert, & Martin, 2018), expert systems (Deng &
scribed by series quantified belief measurements such as basic Shi, 2003; Klein & Colot, 2010; Li, Zhou, Hu, Chang, Zhou, & Zhao
probability assignment (BPA) function, belief function, plausibility 2017; Murphy, 20 0 0; Yang, Han, & Han, 2013; Zhou, Xu, Wen, &
function, commonality function, etc. All of these functions have Lv, 2012), decision making (Han, Han, Deng, & Yang, 2011; Wang &
specific physical meanings. And these functions are closely related Song, 2018), etc. In recent years, besides the physical meanings of
to each other. Compared with the strict axiomatic system in prob- concepts in evidence theory, its relationship with other uncertainty
ability theory, the axiomatic system for evidence theory is much theories has also aroused the interest of researchers (Song, Wang,
looser. Unlike Bayesian reasoning, evidential reasoning does not Lei, & Xue, 2014).
depend on the priori knowledge. Shortcomings of Bayesian reason- However, if evidence bodies to be combined are highly conflict-
ing can be partly overcome by evidence theory (Khaleghi, Khamis, ing, counter-intuitive results may be obtained by classical Demp-
Karray F, & Razavi, 2013). Moreover, evidence theory can deal with ster’s rule. This problem was firstly proposed by Zadeh (1986), and
linguistic fuzzy concept. So evidence theory outperforms many is considered as a roadblock in the development of evidence the-
other methods in uncertain reasoning. In evidence theory, belief ory. Many researchers have been dedicated to solving this problem.
can be assigned to all subsets of discernment frame. Information Lots of alternative combination rules have been proposed (Deng
provided by independent evidence sources can be combined by & Shi, 2003; Han et al., 2011; Murphy, 20 0 0; Song, Wang, Wu,
Dempster’s rule (Dempster, 1967). By this way, uncertain infor- Quan, & Huang, 2018; Zhou et al., 2012). All of these modified
combination rules can be divided into two classes, i.e., modifying
the combination rule (Denoeux, 2008; Florea, Jousselme, & Bosse,
∗
Corresponding author. 2009; Lefevre, Colot, & Vannoorenberghe, 2002; Smarandache &
E-mail address: yafei_song@163.com (Y. Song).
https://doi.org/10.1016/j.eswa.2018.06.048
0957-4174/© 2018 Elsevier Ltd. All rights reserved.
C.-l. Fan et al. / Expert Systems With Applications 113 (2018) 264–276 265
Dezert, 2009; Smets, 2000; Yager, 1987) and modifying the evi- temporal-spatial evidence combination rules for target identifica-
dence sources (Deng, Shi, Zhu, & Liu, 2004; Haenni, 2005; Klein tion. But some of them directly used evidence combination rules
& Colot, 2010; Murphy, 20 0 0; Yang et al., 2013). The difference of designed for spatial evidence combination to combine temporal ev-
these methods lies in their perspectives on the cause of counter- idence, which cannot reflect the sequential characteristic of tem-
intuitive combination results. poral evidence and the influence of time. Other temporal evidence
Taking a closer examination on existing evidence combination combination rules were proposed by analyzing the relationship be-
rules, we can find that most of them mainly focus on the combi- tween all evidence bodies, where the temporal evidence sequence
nation of spatial evidence sources, i.e., uncertain information col- is treated as spatial evidence sources collected at the same time.
lected from multiple sources. For an expert system, aggregating all The relationship between temporal evidence sources is not fully
information from multiple sources in a time sequence can enhance exploited. Moreover, the conflict in temporal evidence sequence
its reliability in intelligent reasoning. So the fusion of temporal in- cannot be well handled by these methods. Therefore, further in-
formation collected sequentially is significant for the design of in- vestigation on temporal evidence combination is necessary to meet
telligent expert systems. But the combination of temporal uncer- the timely requirement of temporal information fusion. The con-
tain information is rarely investigated. Nevertheless, in a multi- struction of a specific combination method for temporal evidence
sensor system, the performance of sensors is usually unsteady and is desirable to reflect the distinct characteristic of temporal infor-
may be influenced by interference information. So the results pro- mation fusion.
vided by sensors at a single time node are usually imprecise and This paper aims to solve challenges in temporal evidence com-
uncertain. In order to get more reliable fusion results, it is nec- bination and facilitate the design of intelligent expert systems
essary to fuse the temporal information which is sequentially ob- based on information fusion. From the research on evidence the-
tained from multi-sensors at multiple time nodes. ory, we can notice that the key of evidence combination is how to
Thus, the fusion of uncertain information is essentially a se- deal with conflicting information. Since evidence reliability evalua-
quential fusion process of temporal-spatial information based on tion is an important method for handling conflict, we will research
multi-sensor and multi-time-node. In the fusion of spatial infor- temporal evidence combination through evaluating the reliability
mation, the information obtained from all sensors can be fused at factor of evidence sequence.
the same time, which is time-independent. But the temporal infor- With the development of intuitionistic fuzzy theory (Atanassov,
mation reaches the fusion system sequentially over time. They are 1986; Bustince & Burillo, 1996; Song, Wang, Lei, & Xue, 2015a;
not collected at the same time. Moreover, considering the real-time Song, Wang, Quan, & Huang, 2017), its connection with other un-
requirement on most information fusion systems, it is impossible certainty theories is attracting more and more attention. Consid-
to commence the fusion after all timely information is obtained. ering the relationship between evidence theory and intuitionistic
So the fusion of temporal information is explicitly sequential, dy- fuzzy sets (Song et al., 2014), we will propose an evidence reliabil-
namic and real-time. The historical fusion results are inherited and ity evaluation method in the framework of intuitionistic fuzzy sets.
updated over the time flow. Based on this evaluation method, the reliability of two neighbor-
We have noted that some modified combination rules such ing BPAs in the evidence sequence will be evaluated. Then relative
as Yager’s combination rule (Yager, 1987) and Denœux’s cautious reliability factors can be obtained. Shafer’s discounting rule will be
conjunctive rule (Denoeux, 2008) can deal with the conflict in- adopted to modify the original BPAs, where the relative reliabil-
formation better. The cautious conjunctive rule is also commuta- ity factor is used as the discounting factor. Finally, the discounted
tive and associative. It seems that they can be applied in tempo- BPAs are combined by Dempster’s rule. Based on this method, we
ral evidence combination. However, these combination rules are also construct a framework for temporal-spatial information fusion.
designed to reduce the influence of conflicting evidence sources, It is proved that the influence of time on temporal evidence com-
rather than treating the temporal fusion in a dynamic and com- bination is considered by the proposed method. Moreover, our pro-
prehensive view. Dynamically, the current conflicting evidence may posed dynamic combination method can deal with the conflict in
convey much useful information when compared with subsequent temporal evidence sequence. Relative reliability is first considered
information, which is not considered in these methods. More- in temporal information fusion, which complies with the dynamic
over, the reliability of evidence sources is not taken into consid- characteristic of temporal information fusion. This is a successful
eration by the combination rules proposed by Yager (1987) and attempt to combine temporal evidence sources in the framework
Denœux (2008). So these combination rules may not be suitable of intuitionistic fuzzy set and evidence discounting.
for temporal information fusion due to their static perspective. The main contribution of this paper can be summarized as fol-
Although temporal information fusion based on evidence the- lowing:
ory has been concerned by some researchers (Hong & Lynch, 1993;
• The relationship between evidence theory and intuitionistic
Hong, Gao, & Li, 2011; Liu, Zhu, Li, & Zhuang, 2006; Wu, Cheng, Qu,
fuzzy sets is explored to facilitate evidence combination.
Pan, & Liu, 2012), a combination rule designed specifically for tem-
• The reliability of BPAs are evaluated in the framework of intu-
poral evidence combination is still an insistent demand. Hong and
itionistic fuzzy multiple criteria decision making.
Lynch (1993) studied the model of temporal-spatial uncertain in-
• A possibility-based intuitionistic fuzzy value ranking method is
formation fusion based on evidence theory. Aiming at target iden-
proposed and applied in evidence reliability evaluation.
tification fusion, three models were proposed to fuse temporal-
• When only two evidence sources are available, relative relia-
spatial information. They are recursive centralized fusion model,
bility evaluation is proposed. This is helpful for the fusion of
recursive distributed fusion model without feedback, and recursive
temporal information.
distributed fusion model with feedback. In Hong and Lynch (1993),
• The fusion of temporal uncertain information is investigated
characteristics and performance of these models were analyzed,
from a dynamic perspective. It is demonstrated that our pro-
but they did not propose any specific evidence combination rules
posed fusion framework is sensitive to the time factor in the
for temporal information fusion. Hong et al. (2011) reinvesti-
fusion process.
gated and modified these three models. They proposed a mixed
temporal-spatial information fusion model, where classical Demp- In other words, the proposed method provides an important
ster’s combination rule was applied in the fusion of temporal in- framework for the fusion of temporal information. Moreover, it can
formation. Based on the recursive centralized fusion model, other be also extended to develop more effective combination rules with
researchers (Liu et al., 2006; Wu et al., 2012) have developed the help of the breakthrough in evidence reliability evaluation.
266 C.-l. Fan et al. / Expert Systems With Applications 113 (2018) 264–276
To facilitate our exposition, the rest of this paper is organized ∈ [0, 1]. Pl(A) − Bel(A) can be used to quantify the level of igno-
as following. Some background knowledge on evidence theory and λ · m ( A ), A ⊂
rance about mλ (A ) = { . The difference be-
intuitionistic fuzzy sets is recalled in Section 2. The relationship λ · m(A ) + 1 − λ, A =
between intuitionistic fuzzy sets and BPAs is also analyzed. An evi- tween Bel(A) and provides a measurement of uncertainty about
dence reliability evaluation method is proposed based on intuition- the level of belief in a decision.
istic fuzzy multiple criteria decision making model in Section 3.
Definition 1. (Shafer, 1976). If a BPA m on the discernment frame
This provides the basis of temporal evidence reliability evaluation.
is not fully reliable, and its reliability factor is λ,λ ∈ [0, 1], the BPA
We propose a method to evaluate the relative reliability factor in
can be modified by discounting rule. The discounted BPA can be
Section 4. Temporal combination method based on relative relia-
expressed as:
bility factor is also proposed in this section. Numerical examples
and simulation are presented in Section 5 to show the perfor- λ · m ( A ), A ⊂
mλ ( A ) = (1)
mance of our proposed combination rule. We conclude this paper λ · m(A ) + 1 − λ, A =
in Section 6.
In the BPA mλ , the basic probability mass assigned to in-
creases, while other focal elements are assigned less basic prob-
2. Preliminaries
ability mass. We can note that the influence of unreliability in m
is transformed to the degree of ignorance in mλ . x indicates that
In this section, we will present some basic knowledge related to
the evidence source is completely reliable, then the BPA will re-
evidence theory and intuitionistic fuzzy sets to facilitate the expo-
main unchanged after discounting. If A, the evidence is fully unre-
sition. The following three main points are included in this section:
liable, then the BPA will become m() = 1, which means that the
(1) an interpretation of Dempster–Shafer evidence theory, (2) basic
evidence is useless for decision-making.
concepts on intuitionistic fuzzy sets, and (3) a brief review of the
In multi-source information system, multiple sources can pro-
relationship between intuitionistic fuzzy sets and BPAs.
vide BPAs m1 ,m2 ,, mp at the same time. One source can also
obtain BPA sequence mt1 , mt2 , · · · , mtr . The combination of multi-
2.1. Dempster–Shafer evidence theory ple BPAs can lead to a new BPA. This process is called evidence
updating.
In the mathematical model of evidence theory, we need to de-
termine the discernment framework firstly. Based on discernment Definition 2. (Dempster, 1967). Let m1 and m2 be two indepen-
frame, the support degree of each set can be quantified. Then all dent BPAs on the discernment frame = {θ 1 ,θ 2 ,, θ n }. The new
evidence sources are combined by Dempster’s rule. Discernment BPA generated by combining m1 and m2 is denoted as m = m1 m2 ,
frame is the basis of evidence modeling and evidence combination. shorten as m12 . Then ∀A⊆, we have:
Propositions and sets are connected by discernment frame, which
1
1−k
m1 (B )m2 (C ), A = ∅
also enhances the transformation from abstract logic concepts to m 1 2 ( A ) = B∩C=A (2)
intuitive set concept. In evidence theory, all mutually exclusive re- 0, A = ∅
sults of a decision problem constitute a finite set = {θ 1 ,θ 2 ,,
θ n }, called discernment frame. The following belief measurements where
are significant in evidence theory. k= m1 (B )m2 (C ) (3)
Let = {θ 1 ,θ 2 ,, θ n } be the frame of discernment. If a func- B∩C=∅
tion m : 2 → [0, 1] satisfies the following conditions: (1)m(∅) = 0, is the conflict degree between m1 and m2 . k = 1 indicates that m1
(2)∀A⊆,0 ≤ m(A) ≤ 1, and (3) m(A ) = 1, it is a basic probabil- and m2 are completely contradictory. In the case of k = 1, m1 and
A⊆
ity assignment (BPA) function. m2 cannot be combined by Dempster’s combination rule.
Basic probability assignment function is also called as belief Dempster’s rule can be extended to combine more than two
structure, basic belief assignment (BBA) function or mass function. evidence sources. Its commutativity and associativity provide con-
Since the BPA reflects the degree to which the evidence body sup- venience for the combination of multiple evidence sources. When
ports each set, BPA and evidence body often corresponds to each combining more than two evidence sources, we can combine them
other. ∀A⊆,m(A) is called the basic probability mass (BPM) of A. one by one, or combine them in parallel.
For a BPA m on the discernment frame = {θ 1 ,θ 2 ,, θ n }, Basic probability assignment function, belief function and plau-
∀A⊆, if m(A) > 0, A is the focal element of m. If |A| = 1, m(A) sibility function can be used to quantify belief degree from differ-
> 0, A refers to a singleton focal element. If |A| ≥ 2, m(A) > 0, it ent perspectives. They can also be applied separately or simulta-
is a composite focal element. All focal elements constitute the core neously to make a decision. We can also transform them to the
of m, denoted as C. We can say m is focusing on the core C. Partic- distribution on discernment frame for a sound decision. Some re-
ularly, if the focal elements of m are all singleton subsets of , m searchers have proposed decision rules based on quantified belief
degenerate into a Bayesian BPA, whose mathematical formulation functions. However, because of the existence of composite focal el-
is identical to the probability distribution on . ements, these rules cannot propose precise decision. Decision re-
Two important belief measures associated with a BPA m on sult with much uncertainty is with little help for the final deci-
= {θ 1 ,θ 2 ,, θ n } are belief function and plausibility function. Be- sion. So it is necessary to transform BPAs to probabilities. Based
lief function is a map Bel : 2 → [0, 1], satisfying Bel(∅) = 0, ∀A⊆, on such sense, many transformation methods have been put for-
ward (Daniel, 2005). Among them, the pignistic transformation is
A = ∅, and Bel (A ) = m(X ). Belief function Bel(A) reflects the
X⊆A widely applied. The pignistic transformation maps a BPA m to the
degree to which the evidence supports that A is true. Plausi- so called pignistic probability function. It is defined as following.
bility function is a function P l : 2 → [0, 1], satisfying P l (A ) =
Definition 3. (Smets & Kennes, 1994). For a BPA m(∅) = 0 on the
m(X ) = 1 − Bel (Ā ). The plausibility function Pl(A) can be re-
X∩A=∅ discernment frame = {θ 1 ,θ 2 ,, θ n }, ∀A⊆, its pignistic proba-
garded as the belief degree that A is not false. bility is defined as:
The interval λ represents the confidence interval of mλ (A ) = |A ∩ B| m(B )
λ · m ( A ), A ⊂ Bet Pm (A ) = (4)
{ . λ also describes the uncertainty about λ |B| 1 − m ( ∅ )
λ · m(A ) + 1 − λ, A = B ⊆
C.-l. Fan et al. / Expert Systems With Applications 113 (2018) 264–276 267
Particularly, given m(∅) = 0 and θ ∈ , we have: interval representation of intuitionistic fuzzy sets. So we can ana-
1 lyze the relationship between intuitionistic fuzzy sets and evidence
m (B )
Bet Pm ({θ } ) = (5) theory based on this similarity relation.
|B| 1 − m ( ∅ )
θ ∈B For an intuitionistic fuzzy set A ∈ IFSs(X), if π A (x) = 0, we have
vA (x) = 1 − μA (x), then the IFS A reduces to Zadeh’s fuzzy set. So
2.2. Intuitionistic fuzzy sets fuzzy sets and classical sets can be viewed as special intuitionistic
fuzzy sets. Consequently, both of crisp sets and fuzzy sets can be
In classical sets, given an element x ∈ X = {x1 ,x2 ,, xn } and a express in a unified form in the framework of intuitionistic fuzzy
set A⊆X, then the relationship between x and A is x ∈ A or x∈A. In sets.
other words, the membership degree of an element x belonging to Let crisp set A be the nonempty subset of the universe
A is 0 or 1. This kind of two-valued logic is the foundation of mod- X = {x1 ,x2 ,, xn }, i.e., A⊆X and A = ∅. Then A can be expressed
ern mathematics. By extending the range of membership function as an intuitionistic fuzzy set A˜ = {< x, μA˜ (x ), vA˜ (x ) > |x ∈ X }. For A˜ ,
to the interval [0, 1], Zadeh (1965) proposed the concept of fuzzy if x ∈ A, we have μA˜ (x ) = 1, vA˜ (x ) = 0; if x∈A, we have μA˜ (x ) =
set. 0,vA˜ (x ) = 1.
For example, in the universe of discourse X = {x1 ,x2 ,x3 ,x4 }, sub-
Definition 4. (Zadeh, 1965). Let X = {x1 ,x2 ,, xn } be the universe sets A = {x1 ,x2 } and B = {x1 ,x2 ,x3 } can be expressed by of intuition-
of discourse. A fuzzy set A on X is defined as: istic fuzzy sets A˜ and B˜, respectively:
A = { x, μA (x )|x ∈ X } (6) A˜ = { x1 , 1, 0, x2 , 1, 0, x3 , 0, 1, x4 , 0, 1},
where, μA (x): X → [0, 1] is membership function, representing the
degree to which x belongs to A. B˜ = { x1 , 1, 0, x2 , 1, 0, x3 , 1, 0, x4 , 0, 1}.
All fuzzy sets in the universe X can be expressed by FSs(X). In The unified form of crisp sets, fuzzy sets and intuitionistic fuzzy
fuzzy set A, vA (x) = 1 − μA (x) is the non-membership degree of x. sets illustrates the flexibility of intuitionistic fuzzy sets in portray-
In the framework of fuzzy sets, the relationship between element ing uncertain information. This enables us to model all kinds of
and fuzzy set is determined merely by membership degree. Higher uncertainty in the framework of intuitionistic fuzzy set.
membership degree μA (x) indicates greater degree to which x be-
longs to A. When μA (x) ∈ {0, 1}, the fuzzy set A degrades into 2.3. The relationship between IFS and BPA
a classical set, which is called as crisp set in this paper. In such
sense, fuzzy sets can be considered as the extension of classical Recently, more and more attention has been focused on the re-
sets, and classical sets are special fuzzy sets. lationship between intuitionistic fuzzy sets and evidence theory.
For a better description on uncertain information, the concept Li, Yang, and Yang (2004) tried to analyze evidence theory from
of intuitionistic fuzzy set was proposed by Atanassov (1986). the view of vague set. They considered evidence theory as a spe-
Definition 5. (Atanassov, 1986). Let X = {x1 ,x2 ,, xn } be the uni- cial vague set. Since vague sets and intuitionistic fuzzy set are
verse of discourse. An intuitionistic fuzzy set (IFS) A over X is de- equivalent, evidence theory can be also regarded as special intu-
fined as: itionistic fuzzy set. In Dymova and Sevastjanov (2010; 2012), con-
cepts in intuitionistic fuzzy sets are interpreted in the framework
A = { x, μA (x ), vA (x )|x ∈ X } (7) of evidence theory. Ranking method of intuitionistic fuzzy val-
ues and decision rule are also proposed based on evidence theory
where μA (x): X → [0, 1] and vA (x): X → [0, 1] are the membership
(Dymova & Sevastjanov 2010; 2012). Yager (2014) proposed an in-
function and non-membership function of x, respectively. μA (x)
tuitionistic view of the Dempster–Shafer belief structure. Xing and
and vA (x) satisfy the following condition:
Liu (2009) presented the method of determining membership and
μA ( x ) + v A ( x ) ≤ 1 (8) non-membership function in intuitionistic fuzzy sets by defining
generalized belief function and plausibility function. Next, we will
In intuitionistic fuzzy set, the condition of μA (x) + vA (x) ≤ 1 analyze the further relationship between BPAs and IFSs.
leads to another parameter, i.e., hesitation degree π A (x). The hesi- In evidence theory, the concepts of plausibility and belief are
tation degree π A (x) of x can be defined as: used to define the upper and lower probability bounds of the im-
precise probability of any subset A in the discernment frame . As
πA ( x ) = 1 − μ A ( x ) − v A ( x ) (9)
mentioned in Section 2, Bel (A) is the lower probability, denoted
For convenience, we designate all intuitionistic fuzzy sets in X as P− (A), and the plausibility is the upper probability, denoted as
as IFSs(X). Intuitionistic fuzzy set in the discourse universe X = {x} P+ (A). Thus here, the probability P(A) lies in an interval, i.e., P(A)
can be denoted as μA (x),vA (x) or μA ,vA in short. A = μA ,vA is ∈ [Bel(A), Pl(A)]. Apparently, [Bel(A), Pl(A)] is a sub-interval of the
also used to express an intuitionistic fuzzy value (IFV). unit interval [0, 1].
It is noteworthy that besides Definition 5, there are other pos- A similar situation in fuzzy set theory is type two fuzzy set
sible representations of IFSs proposed in literature. It has been (Mendel & Bob, 2002). In this environment, imprecise membership
proved that intuitionistic fuzzy sets are equivalent to vague sets grades associated with a fuzzy set are expressed as interval val-
(Bustince & Burillo, 1996). We can also use an interval representa- ues, which is also called interval-value fuzzy set. In this case, if
tion = {θ 1 ,θ 2 ,, θ n } of intuitionistic fuzzy set M in instead F is an interval-valued fuzzy subset of the space X, then F(x), the
of pair θ i (Hong & Kim, 1999). This approach is equivalent to the membership degree of the element x belonging to F is an interval
interpretation of interval valued fuzzy sets, where Bel(θ i ) and θ i F(x) = [FL (x), FU (x)].
∈ M represent the lower bound and upper bounds of member- Another related formulation for imprecise membership grades
ship degree, respectively. Obviously, Bel (θ̄i ) is a valid interval, since of a fuzzy set F was introduced by Atanassov (1986), where intu-
1 − Pl(θ i ) always holds for θ i . This indicates the equivalence rela- itionistic membership grades were presented. For an intuitionistic
tion between intuitionistic fuzzy set and interval valued fuzzy set. fuzzy set F defined in X, the membership of x in F can also be re-
We have used the interval [Bel(A), Pl(A)] to express the belief in- garded as the degree of support for membership, denoted as FY (x).
terval of a proposition in evidence theory. This is similar to the The non-membership can be considered as the degree of support
268 C.-l. Fan et al. / Expert Systems With Applications 113 (2018) 264–276
against membership, denoted as FN (x). In such sense, the expres- However, in most of applications as target recognition, there is
sion μF (x),vF (x) is identical to FY (x),FN (x), which can be referred little information available to provide criteria for the assessment of
as an intuitionistic statement about the membership of x in F. evidence reliability. Moreover, the working environment of sensors
The relationship between these two representations of impre- may change, which will affect the performance of sensors. So the
cise fuzzy membership grades has been investigated (Cornelis, evaluation of evidence reliability is usually uncertain. In the infor-
Atanassov, & Kerre, 20 03; Deschrijver & Kerre, 20 03). It is pointed mation fusion system based on evidence theory, we need to eval-
out that interval-valued fuzzy sets and intuitionistic fuzzy sets uate evidence theory based on the dynamic outputs of sensors. In
are equivalent to each other. Particularly, the interval-valued this paper, we call it as dynamic reliability evaluation.
membership grade [FL (x), FU (x)] is identical to the intuitionistic Many methods have been proposed to evaluate the dynamic re-
membership grade pair FY (x),FN (x), such that FY (x )= FL (x) and liability of evidence based on the principle of majority. In the ma-
FN (x) = 1 − FU (x). jority principle, if one evidence body is supported by others, it can
The analogy between the imprecise interval-valued member- be assigned high reliability (Guo et al., 2006). Most existing dy-
ship grade of a fuzzy set F(x) ∈ [FL (x),FU (x)], the intuitionistic mem- namic reliability evaluation methods are constructed by defining
bership grade pair FY (x),FN (x), and the imprecise interval-valued different measurements of conflict and evidence distance to de-
probability of the belief structure P(A) ∈ [Bel(A), Pl(A)] inspires us script the similarity between evidence bodies. Since the measure-
to consider an intuitionistic representation of the interval associ- ment of evidence conflict is still an open problem, different evi-
ated with the probability of the subset A. Following this analogy, dence conflict measurements may lead to different evaluation re-
we get an intuitionistic representation of the imprecise value asso- sults. We can also note that in the situation of two evidence bodies
ciated with the probability of A as Bel(A), 1 − Pl(A). are given, their reliability factors are all 1 according to above men-
Specifically, in terms of set theory, the discernment frame in tioned methods. This result is independent to the details of BPA. If
evidence theory corresponds to the discourse universe in intu- these two evidence bodies are highly conflicting, we cannot deter-
itionistic fuzzy sets. We can consider a BPA m in the discernment mine which one is unreliable. So the principle of majority is not
frame = {θ 1 ,θ 2 ,, θ n } as an intuitionistic fuzzy set M over the suitable for this situation, which is common in temporal evidence
discourse universe = {θ 1 ,θ 2 ,, θ n }. Then, for θ i ∈ , Bel(θ i ) combination.
represents the membership degree of θ i ∈ M, while Bel (θ̄i ), i.e., In temporal evidence combination, we need to evaluate the re-
1 − Pl(θ i ), can be considered as the non-membership degree of θ i liability factors of historical fusion result and new collected infor-
with respect to M. In such way, focal elements of BPA m can be mation based on their relationship. Unlike dynamic reliability eval-
simplified as singleton focal elements with intuitionistic fuzzy be- uated based on majority principle and static (absolute) reliability
lief degree < Bel(θ i ),1 − Pl(θ i ) > . Moreover, intuitionistic fuzzy be- assessed with the help of training data, the reliability for coupled
lief degree < Bel(θ i ),1 − Pl(θ i ) > represents the degree to which the temporal evidence is called as relative reliability. So it is important
object θ i matching the true solution of a decision problem. For ex- to seek an effective method for relative reliability evaluation.
ample, in the application of target recognition, < Bel(θ i ),1 − Pl(θ i ) To achieve a comprehensive and objective evaluation on the dy-
> can be considered as the matching degree between θ i and the namic reliability of evidence, we have proposed a new method for
real identification of target. evidence reliability evaluation based on intuitionistic fuzzy multi-
On the other hand, an IFS A = <x, μA (x),vA (x) > in dis- ple criteria decision making model (Wang, Zhu, Song, & Lei, 2016).
course universe X = {x} can be regarded as the answer of ques- The implementation process of this method (Wang et al., 2016)
tion “Does x belong to A? ” . In this problem, the discern- shows its potential in relative reliability evaluation. For the sake
ment frame is = {Yes, No}. According to definitions of mem- of discussion, we will recall this method in this section.
bership function and non-membership function, we can get:
m({Yes}) = μA ,m({No}) = vA ,m() = π A . So A = <x, μA (x),vA (x) > cor-
3.1. Ranking IFVs based on possibility
responds to a BPA as:
m({Yes} ) = μA
Definition 6. (Nakahara, 1998). Given two interval values
m({No} ) = vA (10)
m() = πA α = [aL ,aU ] and β = [bL ,bU ], satisfying aL ≤ aU , and bL ≤ bU ,
then the possibility to which α is not less than β , denoted as
Obviously, this transformation can be extended to intuitionistic P(α ≥ β ) , can be defined by:
fuzzy sets in any discourse universe. Hence, we can analyze intu-
itionistic fuzzy operations and measurements based on evidence aU − bL
P(α ≥β ) = min 1, max ,0 (11)
theory. We must note that the transformation from BPAs to intu- (a − aL ) + (bU − bL )
U
Definition 7. Two intuitionistic fuzzy values are given as 3.2. Evidence reliability evaluation based on IFMCDM
A = μA ,vA and B = μB ,vB . When π A + π B = 0, the possibility of
A ≥ B, shorten as P(A ≥ B) is defined as: Assume the set of all alternatives is A = {A1 ,A2 ,, Am }.
1 − v − μ X = {x1 ,x2 ,, xn } denotes the set of all criteria. The evaluation
A B of alternative Ai under criterion xj is an intuitionistic fuzzy value
P(A≥B ) = min 1, max ,0 (12)
πA + πB μij ,vij , i = 1, 2, , m, j = 1, 2, , n. μij ,vij can be also regarded
as the degree to which Ai matches the ideal alternative with re-
If π A + π B = 0, A and B are precise real numbers. We
spect to criterion xj . Then the intuitionistic fuzzy multiple criteria
can get the following relations: μA > μB ⇒P(A ≥ B) = 1,μA <
decision making (IFMCDM) model can be expressed by a decision
μB ⇒P(A ≥ B) = 0,μA = μB ⇒P(A ≥ B) = 0.5.
matrix based on alternative set A and criteria set X:
For two intuitionistic fuzzy values A = μA ,vA and B = μB ,vB ,
we can get the following properties of P(A ≥ B) :
F = μi j , vi j m×n
(1) 0 ≤ P(A ≥ B) ≤ 1; x1 x2 ··· xn
⎛ ⎞
(2) P(A ≥ A) = 0.5; A1 μ11 , v11 μ12 , v12 ··· μ1 n , v 1 n
(3) P ( A ≥ B ) = 1 ⇔ μA ≥ 1 − v B ; ⎜ μ21 , v21 μ22 , v22 ··· μ2n , v2n ⎟ (16)
= A2 ⎜ ⎟
(4) P(A ≥ B) + P(B ≥ A) = 1. .. .. .. ..
.
⎝ . . .
⎠
It is noteworthy that we can get P(A ≥ B) = 1⇒μA ≥ 1 − vB ac- Am μm 1 , v m 1 μm 2 , v m 2 ··· μmn , vmn
cording to property (3). Considering 1 − vA ≥ μA and 1 − vB ≥ μB ,
we can get 1 − vA ≥ μA ≥ 1 − vB ≥ μB . Then it follows that μA As introduced earlier, a BPA m in the discernment frame
≥ μB ,vA ≤ vB , i.e., P(A ≥ B) = 1⇒μA ≥ μB ,vA ≤ vB . But its inverse = {θ 1 ,θ 2 ,, θ n } can be transformed to intuitionistic fuzzy set.
proposition does not hold. The intuitionistic fuzzy value < Bel(θ i ),1 − Pl(θ i ) > can be consid-
Suppose N intuitionistic fuzzy values A1 ,A2 ,, AN are expressed ered as the support degree on θ i . Take the application of target
by Ai = μAi , vAi ,i = 1, 2, , N, then we can rank these intuitionis- identification fusion system based on evidence theory for exam-
tic fuzzy values according to Definition 7. This process can be pro- ple, each sensor provides the target’s identification in the form of
posed as following: BPA. If each sensor reflects one property of the target, in the BPA
(1) Calculate the possibility of Ai ≥ Aj : m provided by sensor Sk , < Bel(θ i ),1 − Pl(θ i ) > can be regarded as
the degree to which θ i matches the real identification of the target
1 − v A i − μA j with respect to the criterion Sk . Based above analysis, the combina-
P(Ai ≥A j ) = min 1, max ,0 , i, j = 1, 2, · · · , N tion of multiple evidence sources can be transformed to the model
πA i + πA j
of intuitionistic fuzzy multiple criteria decision making.
(13) Let = {θ 1 ,θ 2 ,, θ p } be the discernment frame. BPAs
to be combined are m1 ,m2 ,, mq . Suij = <μj (θ i ),vj (θ i ) >
For clarity, we let Pi j = P(Ai ≥A j ) .
is the supporting degree of θ i assigned by mj , where
(2) Construct the comparison matrix of all intuitionistic fuzzy μj (Ai ) = Belj (Ai ),vj (Ai ) = 1 − Plj (Ai ),i = 1, 2, , p,j = 1, 2, , q.
values: Then, the combination of m1 ,m2 ,, mq can be transformed to
⎡ ⎤ the following intuitionistic fuzzy multiple criteria decision making
P11 P12 ··· P1N
⎢ P21 P22 ··· P2N ⎥ model.
P=⎢ . .. .. ⎥ (14)
⎣ .. . .
⎦ F = μ j (θi ), v j (θi ) p×q
PN1 PN2 ··· PNN
⎛ m1 m2 · · · mq ⎞
Obviously, for matrix P, we have: ∀i, j ∈ {1, 2, , N}, 0 ≤ Pij ≤ θ1 Su11 Su12 · · · Su1q
1, Pij + Pji = 1, Pii = 0.5. = θ2 ⎜Su21
⎜ .
Su22 · · · Su2q ⎟
(3) Sum the elements in each row of matrix P: .. ⎝ .. .. .. ⎟⎠
. . .
N θp Su p1 Su p2 · · · Su pq
Pi = Pi j (15)
j=1
⎛ m1 m2 ··· mq ⎞
(4) Since Pi reflects the possibility to which Ai is not less than θ1 μ1 (θ1 ), v1 (θ1 ) μ2 (θ1 ), v2 (θ1 ) ··· μq (θ1 ), vq (θ1 )
other elements, we can get Pi ≥ Pj ⇒Ai ≥ Aj . So we can rank = θ2 ⎜ μ1 (θ2 ), v1 (θ2 ) μ2 (θ2 ), v2 (θ2 ) ··· μq (θ2 ), vq (θ2 )⎟
A1 ,A2 ,, An based on the ranking order of Pi . .. ⎜⎝ .
.
.
.
.
.
⎟
⎠
The following example will be employed to demonstrate the . . . .
implementation of this ranking method. θ p μ1 (θ p ), v1 (θ p ) μ2 (θ p ), v2 (θ p ) · · · μq (θ p ), vq (θ p )
(17)
Example 1. Three intuitionistic fuzzy values are given as: A1 = 0.6,
0.3, A2 = 0.45, 0.2, A3 = 0.5, 0.25. Therefore, the problem of evidence combination has been trans-
formed to the assessment of alternatives under multiple criteria.
The comparison matrix of A1 , A2 and A3 is: Accordingly, dynamic reliability of each evidence body is corre-
sponding to the weight of each criterion. Thus, we can use the
0.50 0.56 0.57
P= 0.44 0.50 0.50 method for determining criteria weights to evaluate the dynamic
0.43 0.50 0.50 reliability of evidence sources.
Although many methods have been developed for deriving cri-
The summation of each row is: terion weights (Wang et al., 2016), lots of them are based on or-
P1 = 0.50 + 0.56 + 0.57 = 1.63, P2 = 0.44 + 0.50 + 0.50 = 1.44, dered weighted averaging operation and nonlinear optimization
P3 = 0.43 + 0.50 + 0.50 = 1.43. model (Wang & Zhang, 2013). Some of them need prior knowledge
Since P1 > P2 > P3 , the ranking order of A1 , A2 , A3 is A1 > A2 > to provide the ideal solution. Apparently, these methods are not
A3 . suitable for the evaluation of evidence reliability, which is imple-
270 C.-l. Fan et al. / Expert Systems With Applications 113 (2018) 264–276
mented without any prior knowledge. So we will evaluate evidence is the ideal weighting vector. Since the eigenvector is not unique,
reliability based on self-assessment. w need to be normalized as:
In the model of multiple criteria decision making, each alter-
q
native is assessed under each criterion. We can suppose that each w = w/ wj (23)
alternative is an “Economic Person”, whose outstanding feature is j=1
2 (i )
Pj1 0.5 + P((Siu) ≥Su
β1(i) = 2
j=1 2)
(i ) 2 (i )
= 1
4.2. Evidence discounting based on relative reliability
j=1 Pj1 + j=1 Pj2 0.5 + P(Su1 ≥Su2 ) + 0.5 + P((Siu)2 ≥Su1 )
(i )
(30)
r2 · mk ( A ), A ⊂
(6) Considering P (i ) + P (i ) = 1, the weighting vector ob- mrk2 (A ) = (37)
(Su1 ≥Su2 ) (Su2 ≥Su1 ) r2 · mk ( A ) + 1 − r2 , A =
tained from θ i is:
T Then we can combine the discounted evidence bodies by
0.5 + P((Siu) ≥Su 0.5 + P((Siu) ≥Su
(i ) 2) 1) Dempster’s combination rule. This method is named as temporal
β = 1
, 2
(31)
2 2 evidence combination based on relative reliability factor (TEC-RRF).
Based on TEC-RRF, all temporal evidence bodies can be sequentially
(7) Construct matrix W = (β(1) ,β(2) ,, β(n) ). The Perron-Frobenius combined.
vector (the eigenvector associated to the maximal positive Motivated by the recursive centralized fusion model for
eigenvalue) of WWT can be used as the weighting vector of temporal-spatial information (Hong & Lynch, 1993), we can con-
mαk−1
and mk , denoted by β = (β 1 ,β 2 )T . Then β satisfies: struct a recursive combination model for temporal-spatial evidence
sources based on TEC-RRF. This model is demonstrated in Fig. 1,
β W W T = λmax · W W T (32) where the dashed line represents evidence discounting and the
dot-dash line represents evidence combination based on Demp-
(8) To keep each entry of β positive, β should be normalized by:
ster’s combination rule. Temporal-spatial information obtained by
T
β = β /(β1 + β2 ) = β 1 , β 2 (33) multi-sensor can be fused based on this model.
5.1. Numerical examples From Table 1, we can note that the final fusion results obtained
in two cases are distinctly different. In the first case, the final re-
sult supports {θ 1 } in the greatest degree. We can find that m3a
Example 2. In a target identification fusion system based on evi-
and m1 are similar in such case. They both prefer to supporting
dence theory, the discernment frame is = {θ 1 ,θ 2 ,θ 3 }. The accu-
{θ 1 }. The information obtained at t2 can be interpreted as inter-
mulative recognition result obtained at time node t1 = 10s is cor-
ference on the system, which leads to wrong recognition result. So
responding to BPA m1 . At current time node t2 = 12s, the latest
in the temporal fusion result at t2 , the supporting degree on {θ 1 }
recognition result is expressed by BPA m2 . m1 and m2 are ex-
decreases. At t3 , the system recovers from abnormal state. With
pressed as following.
the addition of m3a , the basic belief assigned to {θ 1 } climbs up in
m 1 ( { θ1 } ) = 0 . 6 , m 1 ( { θ2 } ) = 0 . 1 , m 1 ( { θ3 } ) = 0 . 3 the temporal combination result at t3 . So the target is identified as
{θ 1 } in the final decision.
For the second case, both of m3b and m2 support {θ 2 }. The BPA
m 2 ( { θ1 } ) = 0 , m 2 ( { θ2 } ) = 0 . 8 , m 2 ( { θ3 } ) = 0 . 2 . obtained at t1 indicates that this recognition result is unreliable.
Relative reliability factors of m1 and m2 can be obtained as: The target recognition system becomes more reliable with time
r1 = 1, r2 = 0.645. flow. So the information proposed later is more accuracy. Since the
According to the relative reliability factor, original BPA can be information obtained at the first time node is discounted, its influ-
discounted as follows: ence on the final fusion result can be reduced. We note that {θ 2 }
is assigned the most support degree by m3b and m2 , so {θ 2 } gets
mr11 ({θ1 } ) = 0.6, mr11 ({θ2 } ) = 0.1, mr11 ({θ3 } ) = 0.3, mr11 () = 0 the most basic probability mass in the final fusion result of second
case.
This example illustrates that the TEC-RRF method is sensitive to
mr22 ({θ1 } ) = 0, mc22 ({θ2 } ) = 0.516, mc22 ({θ3 } ) the change of BPA order. It can well handle the conflict informa-
= 0.129, mc22 () = 0.355. tion in the temporal fusion system, which can enhance the anti-
interference ability of identification fusion system.
r r
Then we can combine m11 and m22 by Dempster’s combination In order to make a comparison with Dempster’s combination
rule. The accumulative fusion result at t2 is: rule and TEC-RTRF method proposed in Song et al. (2015), we
present the temporal combination results based on these two
m12 ({θ1 } ) = 0.478, m12 ({θ2 } ) = 0.196, m12 ({θ3 } )
methods in Table 2.
= 0.326, m12 () = 0. In the first case, from Table 2 we can note when Dempster’s
We can notice that the final fusion result supports {θ 1 } in the rule is applied, the combination result at t2 does not support {θ 1 }.
greatest degree, which is consistent with m1 . If we directly use Since evidence reliability is not taken into consideration, the sys-
Dempster’s combination rule to combine m1 and m2 , we will get tem cannot recover from the abnormal state caused by the inter-
m12 ({θ 1 }) = 0. Then the BPM focused on {θ 1 } will be kept as 0 for- ference at t2 . It can be found that in the temporal fusion result at
ever, even if other BPAs supporting {θ 1 } are obtained later. The fu- t3 , the BPM assigned to {θ 1 } is still 0, although m3a supports {θ 1 }
sion system is stuck with the problem of “veto-by-one-vote”, which by the greatest degree. If we use TEC-RTRF method to combine m1
is counter-intuitive. and m2 , we can also get m12 ({θ 1 }) = 0. That is because only real-
If the real-time reliability factor is employed for the tem- time reliability factor is considered. The RTRF of m2 at t2 is 1, so
poral combination (Song, Wang, Lei, & Xing, 2015), we can m2 is unchanged after discounting. With the addition of m3a , m12
get m12 ({θ 1 }) = 0, m12 ({θ 2 }) = 0.71, m12 ({θ 3 }) = 0.29. In spite of is discounted at t3 , which can reduce its influence on the combi-
m12 ({θ 1 }) = 0, the supporting degree on {θ 1 } will climb up with nation result. Hence, the BPM assigned to {θ 1 } increases by 0.35 in
the addition of more BPAs. In other words, the problem of “veto- the temporal combination result at t3 . This indicates that the sys-
by-one-vote” can be avoided. However, if t2 is the last time node tem can recover from the abnormal state, but the recovery rate is
and the system is disturbed seriously at t2 , temporal combination slow.
based on real-time reliability factor will get a wrong result. This When m3b participates the fusion, it is regarded that the infor-
may lead to decision failure. Moreover, in the method proposed in mation gained at the first time mode has great deviation. We can
Song et al. (2015), the real rime reliability factors are determined see that both of Dempster’s rule and TEC-RTRF method can obtain
by the position in time flow, but they are independent of the dis- right fusion result. However, these combination processes are risky.
tribution of each BPA. Hence, much useful information is discarded They cannot enable the system to recover from abnormal state or
in the fusion process. We can find that conflicts between BPAs can only provide a slow recovery speed when the system is disturbed
be better handled when relative reliability factor of temporal evi- later.
dence body is taken into consideration. So TEC-RRF method is good Comparing with Dempster’s rule and TEC-RTRF, we can find
for making a sounder decision. that the TEC-RRF method can greatly enhance the anti-interference
Next, more temporal evidence bodies will be considered to an- ability of fusion system. TEC-RRF method has the highest relia-
alyze the anti-interference ability of TEC-RRF method. bility. Moreover, it is resistant to the interference occurred both
at the earlier time node and in the process of integration. Since
Example 3. On the basis of Example 2, we assume that m3 is the Dempster’s rule and TEC-RTEF method have limited ability of anti-
BPA corresponding to the identification information obtained at interference, they are not suitable for temporal evidence com-
t3 = 15s. To illustrate the performance of TEC-RRF, two cases m3a bination. Compared with Dempster’s rule, our proposed method
and m3b are considered. Two BPAs are given as: may bring much computation burden in the fusion process due to
the reliability evaluation in each time node. However, the trade-
(1) m3a ({θ 1 }) = 0.5, m3a ({θ 2 }) = 0.3, m3a ({θ 3 }) = 0.2;
off between additional computation burden and system robust-
(2) m3b ({θ 1 }) = 0.1, m3b ({θ 2 }) = 0.75, m3b ({θ 3 }) = 0.15.
ness is worthy and significant thanks to the surge of computation
According to TEC-RRF method, three BPAs obtained at three ability.
time nodes can be combined. Temporal combination results for
two cases are listed in Table 1, where r12 represents the relative
reliability factor of m12 .
C.-l. Fan et al. / Expert Systems With Applications 113 (2018) 264–276 273
Table 1
Combination results obtained by TEC-RRF for two cases.
m3 RRF Discounted BPA from m12 Discounted BPA from m3 a /m3 b Final fusion results
m3 a ({θ 1 }) = 0.5 r12 = 1 mr1212 ({θ 1 }) = 0.478 mr33aa ({θ 1 }) = 0.325 m123 a ({θ 1 }) = 0.551
m3 a ({θ 2 }) = 0.3 r3 a = 0.6499 mr1212 ({ θ 2})=0.196 mr33aa ({ θ 2})=0.195 m123 a ({θ 2 }) = 0.182
m3 a ({θ 3 }) = 0.2 mr1212 ({ θ 3})=0.326 mr33aa ({ θ 3})=0.130 m123 a ({θ 3 }) = 0.267
m3 a () = 0 mr1212 ()=0 mr33aa ()=0.350 m123 a () = 0
m3 b ({θ 1 }) = 0.1 mr1212 ({θ 1 }) = 0.311 m33bb ({θ 1 }) = 0.1 m123 b ({θ 1 }) = 0.130
r
r12 = 0.6499
m3 b ({θ 2 }) = 0.75 m123 b ({θ 2 }) = 0.704
r
mr1212 ({ θ 2})=0.127 m33bb ({ θ 2})=0.75
m3 b ({θ 3 }) = 0.15 m123 b ({θ 3 }) = 0.166
r
r3 b = 1 mr1212 ({ θ 3})=0.212 m33bb ({ θ 3})=0.15
m3 b () = 0 m123 b () = 0
r
mr1212 ()=0.350 m33bb ()=0
Table 2
Combination results obtained by Dempster’s rule and TEC-RTRF for two cases (λ = 0.1 for TEC-RTRF).
m3 a ({θ 1 }) = 0.5 m12 ({θ 1 }) = 0 m13 a ({θ 1 }) = 0 m12 ({θ 1 }) = 0 m13 a ({θ 1 }) = 0.35
m3 a ({θ 2 }) = 0.3 m12 ({θ 2 }) = 0.57 m13 a ({θ 2 }) = 0.67 m12 ({θ 2 }) = 0.71 m13 a ({θ 2 }) = 0.44
m3 a ({θ 3 }) = 0.2 m12 ({θ 3 }) = 0.43 m13 a ({θ 3 }) = 0.33 m12 ({θ 3 }) = 0.29 m13 a ({θ 3 }) = 0.21
m3 a () = 0
m3 b ({θ 1 }) = 0.1 m12 ({θ 1 }) = 0 m13 b ({θ 1 }) = 0 m12 ({θ 1 }) = 0 m13 b ({θ 1 }) = 0.05
m3 b ({θ 2 }) = 0.75 m12 ({θ 2 }) = 0.57 m13 b ({θ 2 }) = 0.87 m12 ({θ 2 }) = 0.71 m13 b ({θ 2 }) = 0.83
m3 b ({θ 3 }) = 0.15 m12 ({θ 3 }) = 0.43 m13 b ({θ 3 }) = 0.13 m12 ({θ 3 }) = 0.29 m13 b ({θ 3 }) = 0.12
m3 b () = 0
Table 3 Table 4
Recognition results of each sensor at each time node. Spatial evidence combination results at each time node.
Table 5
Temporal accumulative recognition results at each time node.
Table 6
Recognition results of each sensor at t2 and t3 .
Table 7
Spatial combination results at each time node after exchanging mt2 and mt3 .
Fig. 2. Temporal evidence accumulation results obtained by Dempster’s rule. Time Dempster’s rule EC–CF method
node m({θ 1 }) m({θ 2 }) m({θ 3 }) m({θ 1 }) m({θ 2 }) m({θ 3 })
Table 8
Temporal accumulation results at each time node after exchanging mt2 and mt3 .
6. Conclusion
Acknowledgments Li, J. P., Yang, Q. B., & Yang, B. (2004). Dempster-Shafer theory is a special case of
Vague sets theory. In Proceedings of the 2004 International Conference on Infor-
mation Acquisition (pp. 50–53).
The authors are highly grateful to the Editor-in-chief, Profes- Liu, Y., Zhu, Y., Li, X., & Zhuang, Z. (2006). Integrated Target discrimination model
sor Binshan Lin, and all anonymous reviewers for their insight- in missile defense system. Journal of Electronics & Information Technology, 28(4),
ful comments and valuable suggestions in improving this paper. 638–642.
Liu, Z., Pan, Q., Dezert, J., Han, J., & He, Y. (2018). Classifier fusion with contextual
This work is supported by the Natural Science Foundation of China reliability evaluation. IEEE Transactions on Cybernetics, 48(5), 1605–1618.
(Nos. 61703426, 61273275, 60975026, 61573375, and 61503407). Liu, Z., Pan, Q., Dezert, J., & Martin, A. (2018). Combination of classifiers with op-
timal weight based on evidential reasoning. IEEE Transactions on Fuzzy Systems,
References 26(3), 1217–1230.
Mendel, J., & Bob, J. (2002). Type-2 fuzzy sets made simple. IEEE Transactions on
Fuzzy Systems, 10, 117–127.
Atanassov, K. (1986). Intuitionistic furzy sets. Fuzzy Sets and Systems, 20(1), 87–96.
Murphy, C. K. (20 0 0). Combining belief functions when evidence conflicts. Decision
Bustince, H., & Burillo, P. (1996). Vague sets are intuitionistic fuzzy sets. Fuzzy Sets
support systems, 29(1), 1–9.
and Systems, 79(3), 403–405.
Nakahara, Y. (1998). User oriented ranking criteria and its application to fuzzy
Cornelis, C., Atanassov, K., & Kerre, E. (2003). Intuitionistic fuzzy setsand interval
mathematical programming problems. Fuzzy Sets and Systems, 94(3), 275–276.
valued fuzzy sets: A critical comparison. In Proceedings of Third European Con-
Shafer, G. (1976). A mathematical theory of evidence. Princeton, NJ: Princeton Univer-
ference on Fuzzy Logic Technology (EUSFLAT) Zittau (pp. 227–235).
sity Press.
Daniel, M. (2005). Probabilistic Transformations of Belief Functions. Lecture Notes in
Smarandache, F., & Dezert, J. (2009). In Applications and advances of DSmT for infor-
Computer Science, 3571(1), 539–551.
mation fusion: vol. 3 (pp. 4–32). Rehoboth: American Research Press.
Dempster, A. P. (1967). Upper and lower probabilities induced by a multivalued
Smets, P. (20 0 0). Data fusion in the transferable belief model. In Proceedings of the
mapping. Annals of Mathematical Statistics, 38(4), 325–339.
3rd International Conference on Information Fusion (pp. PS21–PS33).
Deng, Y., & Shi, W.-K. (2003). A modified combination rule of evidence theory. Jour-
Smets, P., & Kennes, R. (1994). The transferable belief model. Artificial Intelligence,
nal of Shanghai Jiaotong University, 37(8), 1275–1278.
66(4), 191–234.
Deng, Y., Shi, W.-K., Zhu, Z.-F., & Liu, Q. (2004). Combining belief functions based on
Song, Y., Wang, X., Lei, L., & Xing, Y. (2015). Credibility decay model in temporal
distance of evidence. Decision Support Systems, 38(3), 489–493.
evidence combination. Information Processing Letters, 115(2), 248–252.
Denoeux, T. (2008). Conjunctive and disjunctive combination of belief functions in-
Song, Y., Wang, X., Lei, L., & Xue, A. (2014). Combination of interval-valued belief
duced by nondistinct bodies of evidence. Artificial Intelligence, 17, 234–264.
structures based on intuitionistic fuzzy set. Knowledge-Based Systems, 67, 61–70.
Deschrijver, G., & Kerre, E. (2003). On the relationship between some extensions of
Song, Y., Wang, X., Lei, L., & Xue, A. (2015a). A novel similarity measure on intu-
fuzzy set theory. Fuzzy Sets and Systems, 133, 227–235.
itionistic fuzzy sets with its applications. Applied Intelligence, 42, 252–261.
Dymova, L., & Sevastjanov, P. (2010). An interpretation of intuitionistic fuzzy sets in
Song, Y., Wang, X., Lei, L., & Xue, A. (2015b). Evidence combination based on the
terms of evidence theory: Decision making aspect. Knowledge-Based Systems, 23,
degree of credibility and falsity. Journal on Communications, 36(5), 1041–1046.
772–782.
Song, Y., Wang, X., Wu, W., Quan, W., & Huang, W. (2018). Evidence combination
Dymova, L., & Sevastjanov, P. (2012). The operations on intuitionistic fuzzy val-
based on credibility and non-specificity. Pattern Analysis and Applications, 21(1),
ues in the framework of Dempster-Shafer theory. Knowledge-Based Systems, 35,
167–180.
132–143.
Song, Y., Wang, X., Quan, W., & Huang, W. (2017). A new approach to construct
Elouedi, Z., Mellouli, K., & Smets, P. (2004). Assessing sensor reliability for mul-
similarity measure for intuitionistic fuzzy sets. Soft Computing. doi:10.1007/
tisensor data fusion within the transferable belief model. IEEE Transactions on
s0 050 0- 017- 2912- 0.
Systems, Man, and Cybernetics-Part B: Cybernetics, 34(1), 782–787.
Wang, J., & Zhang, H. (2013). Multicriteria decision-making approach based on
Florea, M. C., Jousselme, A.-L., & Bosse, E. (2009). Robust combination rules for evi-
Atanassov’s intuitionistic fuzzy sets with incomplete certain information on
dence theory. Information Fusion, 10(2), 183–197.
weights. IEEE Transactions on Fuzzy Systems, 21(3), 510–515.
Fu, Y., Jia, Y., Yang, W., & Zhuang, Z. (2012). Sensor dynamic reliability evaluation
Wang, X., & Song, Y. (2018). Uncertainty measure in evidence theory with its appli-
and evidence discount. Systems Engineering and Electronics, 34(1), 212–216.
cations,. Applied Intelligence, 48(7), 1672–1688.
Guo, H., Shi, W., & Deng, Y. (2006). Evaluating sensor reliability in classification
Wang, X., Zhu, J., Song, Y., & Lei, L. (2016). Combination of unreliable evidence
problems based on evidence theory. IEEE Transactions on Systems, Man, and Cy-
sources in intuitionistic fuzzy MCDM framework. Knowledge-Based Systems, 97,
bernetics-Part B: Cybernetics, 36(5), 970–981.
24–39.
Haenni, R. (2005). Shedding new light on Zadeh’s criticism of Dempster’s rule of
Wu, J., Cheng, Y., Qu, S., Pan, Q., & Liu, Z.-G. (2012). An effective multi-platform mul-
combination. In Proceedings of the Eighth International Conference on Information
ti-radar target identification algorithm based on three level fusion hierarchical
Fusion (pp. 879–884).
structure. Journal of Northwestern Polytechnical University, 30(3), 367–372.
Han, D., Han, C., Deng, Y., & Yang, Y. (2011). Weighted combination of conflicting
Xing, Q., & Liu, F. (2009). Method of determining membership and nonmembership
evidence based on evidence variance. Acta Electronica Sinica, 39(3A), 153–157.
function in intuitionistic fuzzy sets. Control and Decision, 24(3), 393–397.
Hong, D. H., & Kim, C. (1999). A note on similarity measures between vague sets
Yager, R. R. (1987). On the Dempster–Shafer framework and new combination rules.
and between elements. Information Sciences, 115(1-4), 83–96.
Information Science, 41(2), 93–137.
Hong, L., & Lynch, A. (1993). Recursive temporal spatial information fusion with ap-
Yager, R. R. (2014). An intuitionistic view of the Dempster-Shafer belief structure.
plications to target identification. IEEE Transactions on Aerospace and Electronic
Soft Computing, 18(11), 2091–2099.
Systems, 29(2), 435–445.
Yang, W., Jia, Y., & Fu, Y. (2009). Research on fusion recognition algorithm for differ-
Hong, Z., Gao, X., & Li, X. (2011). Research on temporal-spatial information fusion
ent reliable sensors based on the belief function theory. Signal Processing, 25(11),
model based on DS theory. Signal Processing, 27(1), 14–19.
1766–1770.
Khaleghi, B., Khamis, A., Karray F, O., & Razavi, S. N. (2013). Multisensor data fusion:
Yang, Y., Han, D., & Han, C. (2013). Discounted combination of unreliable evidence
A review of the state-of-the-art. Information Fusion, 14(1), 28–44.
using degree of disagreement. International Journal of Approximate Reasoning, 54,
Klein, J., & Colot, O. (2010). Automatic discounting rate computation using a dis-
1197–1216.
sent criterion. In Proceedings of the Workshop on the Theory of Belief Functions
Zadeh, L. A. (1965). Fuzzy sets. Information and Control, 8(3), 338–353.
(pp. 1–6).
Zadeh, L. A. (1986). A simple view of the Dempster-Shafer theory of evidence and
Lefevre, E., Colot, O., & Vannoorenberghe, P. (2002). Belief functions combination
its implication for the rule of combination. AI Magazine, 7(2), 85–90.
and conflict management. Information Fusion, 3(2), 149–162.
Zhou, Z., Xu, X.-B., Wen, C.-L., & Lv, F. (2012). An Optimal Method for Combining
Li, G. L., Zhou, Z. J., Hu, C. H., Chang, L. L., Zhou, Z. G., & Zhao, F. J. (2017). A new
Conflicting Evidences. Acta Automatica Sinica, 38(6), 976–985.
safety assessment model for complex system based on the conditional general-
ized minimum variance and the belief rule base. Safety Sciences, 93, 108–120.