Abstract
D-S evidence theory is an important mathematical tool for uncertainty reasoning. However, it may lead to counterintuitive conclusions when combining conflicting evidences. In order to overcome this disadvantage, one can modify the evidences before Dempster’s rule of combination. One representative method is to assign a weight to each evidence according to its credibility degree based on the concept of distance (or similarity) between two evidences. This method can gain more robust fusion results than many other known methods. However, it may fail to correctly converge according to the cardinality of the sets in the evidence. When evidence conflicts with other evidences, the evidence may lose impact on the combination result. Moreover, the combined mass is nonmonotonic even though evidence varies monotonically. Therefore, the method still leads to counterintuitive or confusing results. This paper brings forward an improved weighted averaging method involving a new similarity measure between evidences and a new combination rule. The numerical examples show the proposed method well solves the above problems.
Access provided by Autonomous University of Puebla. Download conference paper PDF
Similar content being viewed by others
Keywords
1 Introduction
D-S evidence theory is first proposed by Dempster [1] and later developed by Shafer [2]. It can be regarded as a general extension of Bayesian theory that can robustly deal with incomplete data. Due to the capability of uncertain reasoning, it is widely applied in many fields. When there are conflicts among the evidences, however, D-S evidence theory may draw a counterintuitive conclusion [3]. Generally, there are two types of methods for dealing with conflicting evidences. One is to modify Dempster’s rule of combination [4–8], while the other is to modify the evidences before using Dempster’s rule. Evidence-modifying methods can be further classified into two types, i.e. weighted averaging methods [9–11] and discounting techniques [12–14]. In this paper we study the fusion performance of weighted averaging methods. Murphy’s simple averaging method [9] can be viewed as a special case of weighted averaging methods where all the weights of the evidences are identical.
As studied in our previous work [15], compared with rule-modifying methods, weighted averaging methods are more attractive in that they can not only deal with conflicting evidences but converge towards dominant opinion with higher convergence speed. Among the three weighted averaging methods, Deng et al.’s method [10], which is based on Jousselme’s measure of distance between two evidences [16], outperform the other two [9, 11]. Since it takes the relationship among the evidences into account, reasonable combination results can be obtained even if some conflicting evidences are collected due to e.g. enemy’s disguise or bad weather.
Nevertheless, the convergence of Deng et al.’s method is still imperfect. This paper analyzes the problems and then presents an improved fusion method based on a similarity measure between two evidences and a new combination rule.
2 Deng et al.’s Weighted Averaging Fusion Method
In a practical multi-sensors system, the signals may be interfered with by many factors and to different degrees. Besides, some sensors may also be more stable than others. Therefore, the evidences obtained from the sensors are of different credibility degrees and should have different impacts on the fusion result. A reasonable way to handle this problem is to assign a weight to each evidence. When there is no prior knowledge, the relative importance of an evidence can be evaluated by the similarities between it and the other evidences.
Given a finite set of mutually exclusive and exhaustive propositions, i.e. a frame of discernment \( \Uptheta {=}\left\{ {A_{1} , A_{2} , \ldots , A_{m} } \right\} \), where \( A_{i} \) denotes a proposition. All possible subsets of \( \Uptheta \) form are a superset \( P(\Uptheta ) \) containing \( 2^{N} \) elements. Suppose \( m_{i} \) and \( m_{j} \) be two basic probability assignment functions under the same frame of discernment. Jousselme [16] propose a distance measure between two evidences as.
where \( D \) is a \( 2^{N} \times 2^{N} \) matrix with elements \( D(A,B) = \frac{|A \cap B|}{|A \cup B|},\,A,B \in P(\Uptheta ) \).
Then the similarity between two evidences can be defined as
The degree of support of an evidence by all the other evidences is defined by.
The normalization of support degree leads to the following credibility degree of evidence
Accordingly, the weighted average of the evidences is given as
As done in Murphy’s method [9], the new BPA is incorporated into Dempster’s rule of combination for \( n - 1 \) times in order to offer convergence toward certainty, if there are \( n \) evidences.
3 Analysis on Deng et al.’s Method
We illustrate the problems of Deng et al.’s weighted averaging method by several numerical examples as follows.
Example 1
Consider the following two groups of evidences under the frame of discernment\( \Uptheta = \{ A_{1} ,A_{2} ,A_{3} ,A_{4} \} \):
-
Group 1: \( m_{1} (A_{1} ) = 1,\,m_{2} (A_{1} ) = 1,\,m_{3} (\{ A_{1} ,A_{2} \} ) = 1,\,m_{4} (\{ A_{1} ,A_{2} \} ) = 1 \)
-
Group 2: \( m_{1} (A_{1} ) = 1,\,m_{2} (A_{1} ) = 1,\,m_{3} (\{ A_{1} ,A_{2} ,A_{3} \} ) = 1,\,m_{4} (\{ A_{1} ,A_{2} ,A_{3} \} ) = 1 \)
The combination results by Deng et al.’s method are shown in Table 88.1. When combining the former three evidences, \( m_{1} \oplus m_{2} \oplus m_{3} (A_{1} ) \) of Group 2 gains a bigger value than that of Group 1, which is unreasonable. Since the cardinality of \( \{ A_{1} ,A_{2} \} \) is less than that of \( \{ A_{1} ,A_{2} ,A_{3} \} \), the third evidence of Group 1 contains more certainty information about \( A_{1} \) than that of Group 2. Therefore, the combined mass on \( A_{1} \) of Group 1 should be bigger. It is also unreasonable the combined mass on \( A_{1} \) are equal for both groups when combining all the four evidences.
Example 2
Given the frame of discernment \( \Uptheta = \{ A_{1} ,A_{2} ,A_{3} ,A_{4} \} \) and two groups of evidences, each comprising four conflicting evidences with the following BPAs.
-
Group 1: \( m_{1} (A_{1} ) = 1,\,m_{2} (A_{1} ) = 1,\,m_{3} (\{ A_{2} ,A_{3} \} ) = 1,\,m_{4} (\{ A_{2} ,A_{3} \} ) = 1 \).
-
Group 2: \( m_{1} (A_{1} ) = 1,\,m_{2} (A_{1} ) = 1,\,m_{3} (\{ A_{2} ,A_{3} ,A_{4} \} ) = 1,\,m_{4} (\{ A_{2} ,A_{3} ,A_{4} \} ) = 1 \).
Table 88.2 shows the fusion results. Obviously, combining the former three evidences produces counterintuitive results in both the groups. Since \( A_{1} \) is not a focal element in the third evidence, \( m_{1} \oplus m_{2} \oplus m_{3} (A_{1} ) \) should be smaller than \( m_{1} \oplus m_{2} (A_{1} ) \). In fact, the combination leads to the following distance matrix for both the groups
According to formula (2), there is \( Cred = [0.5,0.5,0] \). Thus, the third evidence does not have any impact on the combination result.
Example 3
Suppose the first two evidences are the same as in Example 2 and the third one varies as follows.
As shown in Fig. 88.1, the combined mass on \( A_{1} \) varies nonmonotonically, which is confusing since \( m_{3} (A_{1} ) \) decreases monotonically.
4 A New Fusion Method
Let \( m_{i} \) and \( m_{j} \) be two BPAs under the frame of discernment \( \Uptheta \) containing \( N \) propositions. The similarity between \( m_{i} \) and \( m_{j} \) is defined as
where \( {\mathbf{D}} \) is a \( 2^{N} \times 2^{N} \) matrix and \( ||{\mathbf{m}}||_{{\mathbf{D}}} = \sqrt {{\mathbf{m}}'{\mathbf{Dm}}} \).
The similarity measure \( sim \) is a cosine measure which can be categorized into the inner product family of similarity measures [17]. Wen et al. define a cosine similarity measure with \( {\mathbf{D}} = {\mathbf{I}} \) [18], which does not satisfy any structural property [16]. For the proposed similarity measure, the \( {\mathbf{D}} \) matrix would quantify the interaction between the focal elements of the BPAs. As can be seen from formula (1), Jousselme’s distance also satisfied the strong structural property by constructing the matrix via Jaccard index. More choice of such indexes can be found in [19]. However, using any of these indexes still results in the problem described in Example 2. Therefore, a new index is needed.
Let \( s \) denote \( |A \cap B| \), \( t \) refer to \( |\Uptheta - (A \cup B)| \), and \( p \) to \( |\Uptheta | \), where \( A,B \in P(\Uptheta ) \). The index is defined as
Then the degree of support of an evidence by other evidences is defined by
where \( c \) is a constant which has important influence on the monotonicity of combined mass. In this paper, \( c \) takes the value of 2.
Afterwards, the credibility degree of evidence and the weighted average of the evidences can be defined similar to Deng et al.’s method. In order to improve the converging performance, we also integrate structural information into Dempster’s rule of combination as follows.
We apply the proposed fusion method to the examples discussed in Sect.88. 3 and the combination results are shown in Tables 88.3, 88.4 and Fig. 88.2, respectively.
For Example 1, the combined mass on \( A_{1} \) of Group 1 gains a bigger value than that of Group 2, no matter when the former three evidences or all the four evidences are combined.
For both the groups in Example 2, the former two evidences are identical and therefore there is \( m_{1} \oplus m_{2} (A_{1} ) = 1 \). The combined mass on \( A_{1} \) decreases when combining the former three evidences due to the high conflict among them. Obviously, the third evidence exerts an impact on the combination result as expected. Taking Group 2 as an illustration, when the former three evidences are considered, the credibility degree of the third evidence is 0.3. Besides, it is worthy of notice that the combination results of the former three evidences are different for the two groups. Similar to Example 1, the reason is also related to the cardinalities of focal elements. That is, though \( m_{3} (\{ A_{2} ,A_{3} \} ) \) and \( m_{3} (\{ A_{2} ,A_{3} ,A_{4} \} ) \) are equal, the former contains more certainty information than the latter and therefore the combined mass on \( \{ A_{2} ,A_{3} \} \) is bigger than that on \( \{ A_{2} ,A_{3} ,A_{4} \} \).
For Example 3, it can be observed from Fig. 88.2 that the combined mass on \( A_{1} \) decreases monotonically when \( m_{3} (A_{1} ) \) decreases. The combination result is much reasonable than as shown in Fig. 88.1.
5 Conclusion
Though Deng et al.’s fusion method can gain more robust results than many other known methods, it may still lead to counterintuitive or confusing results. This paper brings forward an improved weighted averaging method involving a new similarity measure between evidences and a new combination rule. The numerical examples show the proposed method well solves the problems.
References
Dempster, A.P.: Upper and lower probabilities induced by multivalued mapping. Ann. Math. Statist. 38(3), 325–339 (1967)
Shafer, G.: A mathematical theory of evidence. Princeton University Press, Princeton (1976)
Zadeh, L.A.: A simple view of the Dempster-Shafer theory of evidence and its implication for the rule of combination. AI mag. 7(2), p. 85 (1986)
Yager, R.R.: On the Dempster-Shafer framework and new combination rules. Inf. Sci. 41(2), 93–137 (1987)
Sun, Q., Ye, X.Q., Gu, W.K.: A new combination rule of evidence theory. Acta Electroni. Sinica 28(8), 116–119 (2000) (Chinese)
Deng, Y., Shi, W.K.: A modified combination rule of evidence theory. J. Shanghai Jiaotong Univ. 37(8), 1275–1278 (2003) (Chinese)
Du, F., Shi, W.K., Deng, Y.: Feature extraction of evidence and its application in modification of evidence theory. J. Shanghai Jiaotong Univ. z1, 164–168 (2004) (Chinese)
Xiang, Y., Shi, X.Z.: Modification on combination rules of evidence theory. J. Shanghai Jiaotong Univ. 33(3), 357–360 (1999) (Chinese)
Murphy, C.K.: Combining belief functions when evidence conflicts. Decis. Support Syst. 29(1), 1–9 (2000)
Deng, Y., Shi, W.K., Zhu, Z.F., Liu, Q.: Combining belief functions based on distance of evidence. Decis. Support Syst. 38(3), 489–493 (2004)
Chen, L.Z., Shi, W.K., Deng, Y., Zhu, Z.F.: A new fusion approach based on distance of evidences. J. Zhejiang Univ. Sci. 6A(5), 476–482 (2005)
Martin, A., Jousselme, A.L., Osswald, C.:Conflict measure for the discounting operation on belief functions. In: Proceedings of the 11th International Conference on Information Fusion. 1–8 (2008)
Browne, F., Bell, D., Liu, W., Jin, Y., Higgins, C., Rooney, N., Wang, H., Muller, J.: Application of evidence theory and discounting techniques to aerospace design. Adv. Comput. Intel. 543–553 (2012)
Zhou, Z., Xu, X.B., Wen, C.L., Lv, F.: An optimal method for combining conflicting evidences. Acta Automat. Sinica. 38(6), 976–985 (2012) (Chinese)
Li, Y., Wang, Y.G., Xu, X.M.: A numerical cases study of evidence fusion methods. Chinese Automation Congress, Beijing (2011)
Jousselme, A., Grenier, D., Bossé, É.: A new distance between two bodies of evidence. Inf. Fusion 2(1), 91–101 (2001)
Jousselme, A., Maupin, P.: Distances in evidence theory: comprehensive survey and generalizations. Int. J. Approx. Reason. 53(2), 118–145 (2012)
Wen, C., Wang, Y., Xu, X.: Fuzzy information fusion algorithm of fault diagnosis based on similarity measure of evidence. Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics). Beijing, pp. 506–515 (2008)
Diaz, J., Rifqi, M., Bouchon-Meunier, B.: A similarity measure between basic belief assignments. In: Proceedings of the 9th International Conference on Information Fusion. pp. 1–6 (2006)
Acknowledgments
This paper is supported by National Natural Science Foundation of China (61074087), Innovation Program of Shanghai Municipal Education Commission of China (12ZZ144), and Innovation Ability Construction Project for Teachers of School of Optical-Electrical and Computer Engineering of University of Shanghai Science and Technology (GDCX-Y1111).
Author information
Authors and Affiliations
Corresponding author
Editor information
Editors and Affiliations
Rights and permissions
Copyright information
© 2013 Springer Science+Business Media New York
About this paper
Cite this paper
Li, Y., Xu, L., Wang, Y., Xu, X. (2013). An Improved Weighted Averaging Method for Evidence Fusion. In: Wong, W.E., Ma, T. (eds) Emerging Technologies for Information Systems, Computing, and Management. Lecture Notes in Electrical Engineering, vol 236. Springer, New York, NY. https://doi.org/10.1007/978-1-4614-7010-6_88
Download citation
DOI: https://doi.org/10.1007/978-1-4614-7010-6_88
Publisher Name: Springer, New York, NY
Print ISBN: 978-1-4614-7009-0
Online ISBN: 978-1-4614-7010-6
eBook Packages: EngineeringEngineering (R0)