MPE Mathematical Problems in Engineering 1563-5147 1024-123X Hindawi Publishing Corporation 10.1155/2015/489207 489207 Research Article Research on Evaluation Method Based on Modified Buckley Decision Making and Bayesian Network Yang Neng-pu Han Mei Chen Shi-yong Liu Xiao-hua Kang Liu-jiang Hung Chih-Cheng School of Traffic and Transportation, Beijing Jiaotong University, Beijing 100044 China njtu.edu.cn 2015 5102015 2015 20 10 2014 21 12 2014 01 01 2015 5102015 2015 Copyright © 2015 Neng-pu Yang et al. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

This work presents a novel evaluation method, which can be applied in the field of risk assessment, project management, cause analysis, and so forth. Two core technologies are used in the method, namely, modified Buckley Decision Making and Bayesian Network. Based on the modified Buckley Decision Making, the fuzzy probabilities of element factors are calibrated. By the forward and backward calculation of Bayesian Network, the structure importance, probability importance, and criticality importance of each factor are calculated and discussed. A numerical example of risk evaluation for dangerous goods transport process is given to verify the method. The results indicate that the method can efficiently identify the weakest element factor. In addition, the method can improve the reliability and objectivity for evaluation.

1. Introduction

The applications of evaluation models almost extend to every aspect of research, for example, forecasting traffic flows , analyzing the causes of traffic accidents , evaluating the risk of transport process , and discussing the origins of driver fatigue . In these topics, two targets attract researchers’ interest: investigating the top-event occurrence rate and exploring the influence degree of element factors. Analytic hierarchy process (AHP), fuzzy analytic hierarchy process (FAHP), fishbone diagram model (FDM), fault tree analysis (FTA), and Bayesian Network (BN) have proved to be efficient approaches for studying them . In these models, the probabilities of element factors are the foundation of quantitative analysis. They are calibrated in the traditional approach by the following three steps. (1) Construct judgment matrix by introducing 1~9 contrast scaling (K). (2) Normalize processing of the matrix to calculate the eigenvector. (3) Assign each element factor a value of element in eigenvector as its probability value correspondingly. However, in the first step, it has natural deficiency by using 1~9 contrast scaling to construct the judgment matrix. To illustrate this point, an intensity of importance scale is given in Table 1.

Scale and its implication.

1~9 contrast scaling (K) Intensity of importance
1 Equal importance
3 Moderate importance
5 Obvious importance
7 Strong importance
9 Extreme importance
2, 4, 6, 8 Intermediate values between two adjacent judgments

Suppose  a,  b  , and  c  are element factors. According to Table 1, if  b  is slightly more important than  a, then  b:a=3:1. Moreover, if  c:b=9:3  (in the view of scale table, the weight ratio of element factors “c” and “b” is  9:3), then the importance of factor “c” compared with “b” belongs to the level of “strong importance.” At the same time,  9:3  is equal to  3:1  mathematically. In other words, the weight ratio of factors “c” and “b” is  3:1, so factor “c” is a little more important than “b.” Therefore, it is conflicting between the above two standpoints. Furthermore, if the relationship between factors “c” and “b” applies to nexus between “c” and “a,” then the weight ratio of factors “c” and “a” is  9:1. Although factor “c” is 9 times of “a” in weight ratio, it is not definite that factor “c” is extremely more important than “a.” Therefore, it inevitably causes the confusion when using the traditional approach. In addition, the judgment matrix constructed by the traditional approach is a deterministic matrix, which cannot well represent the fuzzy comparisons between element factors ideally.

Aiming at the defects of traditional method, the 9/9~9/1 contrast scaling will be introduced into the Buckley matrix; in addition, K-value will be replaced by trapezoidal fuzzy number [3, 10]. Then the modified Buckley Decision Making method will be applied to calibrate the probabilities of element factors. Using this technique, the calculation becomes more rigorous and makes the importance degree much more scientific despite of its complexity.

In the field of reasoning model, BN can be mapped easily from the AHP, FAHP, FDM, and FTA models. Owing to the conditional probability tables of BN, the diagram calculation process of AHP, FAHP, FDM, and FTA can be transformed into a logical table calculation process, which helps to calculate intelligently. Based on the chain rule in BN, the evaluation value of top-event as well as the structure (probability, criticality) importance of each factor will be obtained by the forward and backward calculation of BN. The influence degree of each element factor can be effectively indicated by the previous three importance indexes.

The remaining parts of this paper are organized as follows. Section 2 introduces the new evaluation method. In the method, modified Buckley Decision Making method is applied to calibrate the probability of each element factor; BN is served as the reasoning computation model; structure importance, probability importance and criticality importance are designed as the evaluation indicators; Bucket Elimination algorithm is modified to solve the BN. In Section 3, a numerical example is given to verify the method. Conclusions are finally drawn in Section 4, along with recommendations for future research.

2. The New Evaluation Method 2.1. Modified Buckley Decision Making

The probabilities of element factors are fundamental to the evaluation process, which rely on judgment matrix. Based on the Buckley Decision Making method, the trapezoidal fuzzy number is introduced to construct the judgment matrix. If element factor Ix is extremely more important than Iy, then the initialized fuzzy judgment matrix can be constructed as(1)IxIyA0=IxIy11111919181888991111.Then, the formula  K=9/10-K  is applied to modify each element in the upper right corner of the matrix . The elements in the lower left corner of the matrix are derived by construction rules of Buckley matrix . Therefore, the modified fuzzy judgment matrix is(2)IxIyA=IxIy111119192929929291911111.

Similarly, based on the modified Buckley Decision Making method, the matrix contained contrast between all the element factors which is defined as(3)B=x11x1nxn1xnn,where  xij=αij,βij,γij,δij,  x11~xnn=1,1,1,1, and  xji=1/xij=1/δij,1/γij,1/βij,1/αij.

Let(4)αi=k=j=1nαkj1/n,α=i=1nαi=k;βi=k=j=1nβkj1/n,β=i=1nβ(i=k);γi=k=j=1nγkj1/n,γ=i=1nγ(i=k);δi=k=j=1nδkj1/n,δ=i=1nδ(i=k).

Then the fuzzy weight  xk  of element factor  Ik  is calculated: (5)xk=(αi=kδ,βi=kγ,γi=kβ,δ(i=k)α).Performing the same procedure for  i  from 1 to n, we have(6)X={x1,x2,,xk,,xn}.

Then, the value of element in  X  is assigned to each element factor as its fuzzy probability value correspondingly.

2.2. Bayesian Network

Bayesian Network is a directed acyclic graph with a series of conditional probability tables (CPTs) [3, 12, 13]. It has become widely used in the field of evaluation because of the conditional independence of BN nodes and bidirectional reasoning mechanism. In addition, AHP, FAHP, FDM, and FTA models can be mapped to a BN easily, which helps to establish BN models. Roughly speaking, in most of AHP, FAHP, FDM, and FTA models, the relationship between subnodes and parent nodes only involves “OR gate” and “AND gate” (the logical relationships between events and causes in AHP, FAHP, and FDM models are represented by means of logical “AND” and “OR” gates). Therefore, it is critical to discuss how an “OR gate” and an “AND gate” convert into an equivalent BN. The conversions are shown in Figures 1 and 2.

OR gate.

AND gate.

According to Figure 1, we have (7)Ps=1i1=0,i2=0=0,Ps=1else=1.

According to Figure 2, we have(8)Ps=1i1=1,i2=1=1,Ps=1else=0.

Based on the conversions, each gate (OR or AND) in the AHP, FAHP, FDM, and FTA assigns the equivalent CPT to the corresponding node in the BN and then the whole BN is established . Taking the conditional independence of BN nodes into consideration, we get the joint probability distribution function using chain rule, which is the foundation of forward and backward calculation.

2.3. Evaluation Indicators

In the evaluation method, three indicators, that is, structure importance  (QiSt), probability importance  (QiPt), and criticality importance  (QiCt), are designed to evaluate the importance of each factor.  QiSt  analyzes the influence of each factor with respect to model structure.  QiPt  analyzes the influence of each factor with respect to probability of each factor and model structure.  QiCt  analyzes the influence of each factor with respect to sensitivity and probability, which can reflect the fact that reducing the occurrence probability of a large probability event is easier than a rare event . The calculation formulas of the above three indicators are defined by (9), where the parameters are shown in Table 2:(9)QiSt=p(T=1Ii=1,p(Ij=1)=0.5,1jin)-pT=1Ii=0,pIj=1=0.5,1jin,QiPr=p(T=1Ii=1)-p(T=1Ii=0),QiCr=p(Ii=1)*p(T=1Ii=1)-p(T=1Ii=0)pT=1.

Parameters of calculation formulas.

Parameters Standing for
T Top-event
I i Element factor i
I j Element factor j
n Total numbers of all element factors
I i = 1 Element factor i occurs
I j = 1 Element factor j occurs
p ( T = 1 ) Occurrence probability of top-event
p ( I i = 1 ) Occurrence probability of element factor i
p ( I j = 1 ) Occurrence probability of element factor j
p ( T = 1 ) Conditional probability of top-event when it happens
2.4. Evaluation Methodology

Among various techniques for solving BN, Bucket Elimination has proved to be one of the most efficient approaches . Considering the dimorphism of BN mapped from AHP, FAHP, FDM, and FTA model, the Bucket Elimination algorithm can be modified to reduce the computational difficulty and to improve the efficiency of calculation. Therefore, we define calculation rules of the modified Bucket Elimination algorithm firstly; then, based on the rules and modified Buckley Decision Making, put forward the evaluation process.

(1) Calculation Rules. In the modified Bucket Elimination algorithm, three calculation rules are defined as follows (see (10) ~ (12)), where the parameters are shown in Table 3.

Parameters of calculation rules.

Parameters Standing for Note
C i The subnode event If the node represents element factor, Ci=Ii
C j The parent node event /
P i j The connection event /
p i j The conditional probability table of Pij /
p ( C i = 1 ) The occurrence probability of Ci p ( C i = 1 ) = x i , p ( C i = 0 ) = 1 - x i
p ( C j = 1 ) The occurrence probability of Cj p ( C j = 1 ) = x i , p ( C j = 0 ) = 1 - x i
Rule 1.

To deal with “singer-factor,” if we have CiPijCj, then(10)PCj=1=pCi×pij=xipij,wherepij=0or1.

Rule 2.

To deal with “AND gate,” if we have  i=k-mi=k+nCiPijCj, then(11)pCj=1=i=k-mi=k+npCi×pij=xk-m×xk-m+1××xk××xk+n×1.

Rule 3.

To deal with “OR gate,” if we have i=k-mi=k-mCiPijCj, then(12)pCj=1=i=k-mi=k+npCi×pij=1-(1-xk-m)×(1-xk-m+1)hhhhhh××1-xk××1-xk+n×1.

(2) Evaluation Process. The calculation process of the new evaluation method is shown in Figure 3. According to the evaluation process, the evaluation indicators are derived, which indicate the importance of each element factor. Based on the results, we can efficiently identify the weakest element factor.

Calculation process of the new evaluation method.

3. Numerical Example

Figure 4 shows an FTA model of risk evaluation for dangerous goods transport process, where  Ii  represents an element factor. The corresponding importance of each factor  (Ii)  is given by experts. Based on the modified Buckley Decision Making method, the fuzzy probabilities of all element factors are calibrated (see Table 4).

Fuzzy probabilities of element factors.

Element factor Fuzzy probability
I 1 (0.021 07, 0.028 52, 0.038 31, 0.053 23)
I 2 (0.018 97, 0.026 64, 0.035 25, 0.048 03)
I 3 (0.024 65, 0.033 06, 0.043 22, 0.061 38)
I 4 (0.002 50, 0.003 34, 0.004 07, 0.005 90)
I 5 (0.023 28, 0.031 16, 0.041 59, 0.057 45)
I 6 (0.006 82, 0.009 68, 0.013 09, 0.019 67)
I 7 (0.034 91, 0.045 97, 0.056 38, 0.078 31)
I 8 (0.036 44, 0.048 75, 0.059 69, 0.079 29)
I 9 (0.003 82, 0.005 28, 0.007 30, 0.010 42)

Fault tree.

According to Section 2.2, the BN is established (see Figure 5).

Bayesian Network.

As illustrated in Figure 5, the joint probability distribution function is derived in(13)pt=p(t,S1,,S4,I1,,I9)=p(tS1,S2,I9)p(I9)p(S1I1,I2)p(I1)p(I2)·p(S2S3,S4)p(S3I3,I4,I5)p(I3)p(I4)·p(I5)p(S4I6,I7,I8)p(I6)p(I7)p(I8).

Based on Section 2.4, the evaluation value  p(t)=(0.00800,0.01271,0.01933,0.03295)  and three importance indexes are calculated as follows (see Figures 68).

Structure importance.

Probability importance. Note: x was only marked-out partial numbers because of limited margin.

Criticality importance. Note: x was only marked-out partial numbers because of limited margin.

According to Figure 6, the structure importance of each element factor is sorted to be(14)Q9St>Q3St=Q4St=Q5St=Q6St=Q7St=Q8St>Q1St=Q2St.

According to Figure 7, the probability importance of each element factor is sorted to be(15)Q2Pr>Q1Pr>Q9Pr>Q3Pr>Q8Pr>Q5Pr>Q7Pr>Q6Pr>Q4Pr.

According to Figure 8, the criticality importance of each element factor is sorted to be(16)Q2Cr>Q1Cr>Q8Cr>Q7Cr>Q3Cr>Q5Cr>Q9Cr>Q6Cr>Q4Cr.

As shown in (14), most structure importance of element factors is equal. Therefore, it is not proper to indicate the influence extent of different factors using only structure importance. By a comparison of (14), (15), and (16), we find that  Q9St  and  Q9Pr  are relatively large while  Q9Cr  is relatively small. It indicates that  I9  has great effects on  p(t), but it is difficult to reduce the influence of  I9  on  p(t)  by taking measures. In addition,  i1,2,Q1Pr,Q2Pr>QiPr and  Q1Cr,Q2Cr>QiCr. It can be concluded that I1 and I2 have larger effects on  p(t)  than the other factors, and it can ameliorate the value of  p(t)  efficiently by reducing the occurrence probabilities of  I1  and  I2. Therefore, in the actual productions, it is significantly important to monitor the procedures of  I1  and  I2  to ensure safety.

4. Conclusion

Based on the combination of modified Buckley Decision Making and Bayesian Network, we present a new evaluation method in this paper, which can be widely used in the field of risk assessment, project management, cause analysis, and so forth. By using modifier formula  K=9/10-K, 9/9~9/1 contrast scaling is successfully introduced into the Buckley matrix. In addition, the elements in Buckley matrix are trapezoidal fuzzy number, which can well represent the fuzzy comparisons between element factors ideally. According to the modified Buckley Decision Making method, we can calibrate the probability of each element factor more logically. As has been said, the probabilities of factors are fundamental to evaluation. Therefore, the application of this core technology makes our evaluation method more reliable and objective. Based on the bidirectional reasoning mechanism of BN, the top-event occurrence rate (or evaluation value) and influence indexes of element factors are gained by the forward and backward calculation. Then, according to the sequence of these influence indexes, the impact of each element factor can be represented. Therefore, we can identify the weakest element factor efficiently. In addition, the application of CPTs in BN can make the diagram calculation of AHP, FAHP, FDM, and FTA into a logical table calculation, which is beneficial to the intelligence of calculating techniques.

There are many interesting directions in which we can extend our work. Ongoing and future research that we are pursuing are to construct a discrete time dynamic Bayesian Network model, combining the modified Buckley Decision Making method.

Conflict of Interests

The authors declare that there is no conflict of interests regarding the publication of this paper.

Acknowledgment

This research was supported by China Railways Technology R&D Program (no. 2014X002-C).

Queen C. M. Albers C. J. Intervention and causality: forecasting traffic flows using a dynamic Bayesian network Journal of the American Statistical Association 2009 104 486 669 681 2-s2.0-66549102138 MR2751447 10.1198/jasa.2009.0042 Najib L. Abdullah L. A Lambda-Max of consistency test in Fuzzy Analytic Hierarchy Process (FAHP) for weights of road accidents causes Proceedings of the 20th National Symposium on Mathematical Sciences: Research in Mathematical Sciences: A Catalyst for Creativity and Innovation (SKSM '12) December 2012 Washington, DC, USA AIP Publishing 426 434 10.1063/1.4801157 2-s2.0-84876923478 Yang N.-P. Yang Y.-F. Feng W. Risk assessment of railway dangerous goods transport process based on fuzzy Bayesian Network Journal of the China Railway Society 2014 36 7 8 15 Murphy S. Leach D. Z. The extent to which heavy goods vehicle driver training is focused on reducing the casual factors of driver stress and fatigue Proceedings of the LRA Annual Conference and PHD Workshop September 2013 Birmingham , UK Tang S.-S. Li X.-P. Study on method for assessment of vulnerability of railway emergency rescue system Journal of the China Railway Society 2013 35 7 14 20 2-s2.0-84881587516 10.3969/j.issn.1001-8360.2013.07.003 Singh A. R. Mishra P. K. Jain R. Khurana M. K. Robust strategies for mitigating operational and disruption risks: a fuzzy AHP approach International Journal of Multicriteria Decision Making 2012 2 1 1 28 10.1504/ijmcdm.2012.045080 2-s2.0-84893063931 Distefano N. Leonardi S. Risk assessment procedure for civil airport International Journal for Traffic and Transport Engineering 2014 4 1 62 75 10.7708/ijtte.2014.4(1).05 Tupper L. L. Chowdhury M. Sharp J. Tort liability risk prioritization through the use of fault tree analysis Proceedings of Transportation Research Board 93rd Annual Meeting 2014 Washington, DC, USA no. 14-5153 McCall J. C. Trivedi M. M. Driver behavior and situation aware brake assistance for intelligent vehicles Proceedings of the IEEE 2007 95 2 374 387 10.1109/jproc.2006.888388 2-s2.0-48749088973 Buckley J. J. Fuzzy hierarchical analysis Fuzzy Sets and Systems 1985 17 3 233 247 10.1016/0165-0114(85)90090-9 MR819361 Ying L. Fuzzy comprehensive evaluation on safety of railway dangerous goods transport Railway Freight Transport 2011 29 10 32 37 Friedman N. Geiger D. Goldszmidt M. Bayesian network classifiers Machine Learning 1997 29 2-3 131 163 10.1023/a:1007465528199 2-s2.0-0031276011 Borsuk M. E. Stow C. A. Reckhow K. H. A Bayesian network of eutrophication models for synthesis, prediction, and uncertainty analysis Ecological Modelling 2004 173 2-3 219 239 2-s2.0-1542548177 10.1016/j.ecolmodel.2003.08.020 Bobbio A. Portinale L. Minichino M. Ciancamerla E. Improving the analysis of dependable systems by mapping fault trees into Bayesian networks Reliability Engineering and System Safety 2001 71 3 249 260 10.1016/s0951-8320(00)00077-6 2-s2.0-0035283768 Dechter R. Bucket elimination: a unifying framework for reasoning Artificial Intelligence 1999 113 1-2 41 85 2-s2.0-0033188982 MR1724112 10.1016/s0004-3702(99)00059-4