1. Introduction
Decision-making can be complex, often involving the assessment of multiple factors and their interconnections. Numerous methods in academic literature address multi-criteria decision-making (MCDM) problems, one of which is the analytic hierarchy process (AHP) [
1]. AHP has gained popularity across various fields due to its straightforward nature and ability to decompose intricate decision problems into hierarchical structures [
2]. However, AHP assumes that criteria are independent, which may not hold true in many real-world situations.
To tackle the independence assumption in AHP, researchers have developed several techniques to incorporate dependencies between criteria, such as employing the analytic network process (ANP) [
3] or combining AHP with methods such as fuzzy cognitive maps (FCMs) [
4] and Bayesian networks (BNs) [
5]. While these approaches provide valuable insights, they frequently neglect the evolving nature of dependencies between criteria over time. Dynamic Bayesian networks (DBNs) extend BNs by modeling dependencies between variables across time [
6]. DBNs have been effectively applied in various domains, including finance [
7] and environmental modeling [
8]. To our knowledge, no studies have explored the integration of DBNs and AHP to model time dependencies between criteria in MCDM problems and, hence, are considered here to account for the time-dependent criteria in the AHP.
This paper introduces an innovative method that combines DBNs and AHP to model dynamic dependencies between criteria in multi-criteria decision-making problems. First, we use AHP to derive independent weights for criteria. Then, we establish the dependencies between criteria and derive the DBN parameters. Next, considering time dependency, we combine the initial weights and DBN parameters to update the criteria weights. Finally, we use the revised weights to rank alternatives and make decisions. Our contributions are as follows. The major novelty of the proposed method is to provide extra information for the AHP and adjust the weights derived from the AHP to account for this information. Compared to the data of the AHP or ANP, which are quantified from expert opinions, the time-dependent relationship between criteria can be derived either from expert opinion or historical data. In addition, the concept presented here can be naturally extended to the ANP for handling interdependent and time-varying criteria since the information considered here is distinct from PCMs used in the AHP/ANP. The fact that past papers have not addressed this issue also contributes to the novelty of the paper. We use the AHP as our base model here because it is a well-established and extensively utilized technique for MCDM. In addition, our method can easily be applied in others, besides the AHP/ANP, since the information considered here is distinct from conventional MCDM methods. Its application spans a multitude of domains, and both academics and practitioners have embraced it. This broad acceptance and popularity serve as a solid basis for our suggested method.
The paper’s organization is as follows:
Section 2 delves into the literature on AHP and DBNs;
Section 3 elaborates on our proposed method for amalgamating DBNs and AHP;
Section 4 showcases a numerical example illustrating the implementation of our approach;
Section 5 juxtaposes our method with established approaches; and lastly,
Section 6 concludes the paper and explores potential avenues for future research.
3. Integrating DBNs and AHP: Proposed Method
In this section, we outline our suggested approach for combining DBNs and AHP. The integration seeks to overcome the independence assumption constraint in AHP by including time-dependent dependencies between criteria through the use of DBNs. By fusing DBNs, which are capable of modeling criteria dependencies over time, with AHP, which concentrates on deriving weights for hierarchical decision-making, we enable the computation of time-dependent weights that consider the interdependencies among criteria.
The proposed method comprises the following steps:
Step 1. Problem structuring: define the decision problem and identify the relevant criteria and alternatives. Decision problems are considered hierarchical structures, with the goal at the top, followed by criteria and sub-criteria (if applicable), and alternatives at the bottom.
Step 2. Pairwise comparison and initial weight calculation: conduct pairwise comparisons of criteria using the AHP scale (1–9) based on expert opinion. Construct pairwise comparison matrices for each criterion at the same hierarchical level and calculate the initial weights of the criteria using the eigenvector method, i.e., Equation (1).
Step 3. Designing the DBN: identify criteria with time dependencies and define the time steps for the DBN based on the decision problem’s temporal characteristics. For example, create a two-layered DBN structure, i.e., two time steps, with the first layer representing the initial state of the criteria and the second layer representing the criteria at the next time step. Represent dependencies between criteria with directed edges within and between layers, as shown in
Figure 1.
Based on
Figure 1, we need to acquire information of each transition matrix, e.g.,
P(
|
),
P(
|
), etc.
Step 4. Learning DBN parameters: here, we assume the transition probabilities between statuses are given by experts, but we will consider historical data to derive the transition probabilities between statuses in our numerical example. Take
, for example; we need experts to give the option to fill up the below transition matrix:
| = Status1 | = Status2 |
= Status1 | P(= Status1| = Status1) | P( = Status1| = Status2) |
= Status2 | P(= Status2| = Status1) | P( = Status1| = Status2) |
Sum | 1 | 1 |
Indeed, we can use linguistic variables to help experts to give their opinion above. For example, we can give the scores for linguistic probability: high, medium, and low as 3, 2, and 1, respectively. Then, we normalize the linguistic score matrix to a probability matrix.
Step 5. Inference in DBN: perform inference in the DBN to obtain the probability distribution of each criterion, taking into account the dependencies between criteria and their evolution over time. The marginal steady-state distribution of each criterion can be derived using the Markov chain theory. However, the joint steady-state distribution for a criterion can be determined by the marginal steady-state distribution as follows. The Markov chain theory derives from the marginal distributions of two criteria with two statuses, namely,
C1 = [
p11,
p12] and
C2 = [
p21,
p22]. Then, we can calculate the joint probability of
C1 and
C2,
P(
C1,
C2), as:
P(C1,C2) | C2 = Status1 | C2 = Status2 |
C1 = Status1 | p11p21 | p11p22 |
C1 = Status2 | p12p21 | p12p22 |
Assume
C3 also has two statuses, and
C3(
t + 1) is affected by
C1(
t) and
C2(
t), then the marginal probability of
C3 for the
ith status can be calculated by the joint probability of
C1 and
C2 as:
Step 6. Synthesis of time-dependent weights: combine the initial weights obtained from the AHP (Step 2) with the time-dependent weights from the DBN (Step 5) to calculate adjusted weights for each criterion. Synthesize the modified weights to rank alternatives, considering the time dependencies between criteria.
Let the marginal distributions of criteria, e.g.,
C1–
C3 here, be presented as shown in
Table 1.
The relative advantage (RA) of different statuses indicates the relative importance of criteria under the same status. Next, we can derive the total relative advantage scores (RS) of criteria as the base to adjust the initial weights of criteria. Here, we use the weighted average method to calculate the total relative advantage scores of criteria for simplicity.
Finally, the modified weights of the criteria can be calculated as follows:
By incorporating time dependencies between criteria, the proposed method for integrating DBNs and AHP offers a comprehensive framework to address the limitations of the independence assumption in AHP. This approach enables more accurate and informed decision-making in complex, dynamic environments where criteria dependencies change over time.
4. Numerical Example
In this section, we present a numerical example to demonstrate the application of the proposed method for integrating DBNs and the AHP. The decision problem involves selecting the best location for a new warehouse among five alternatives (A1, A2, A3, A4, and A5) based on eight criteria (C1 to C8). The criteria are as follows. C1: proximity to customers; C2: proximity to suppliers; C3: infrastructure quality; C4: labor availability; C5: real-estate costs; C6: regulatory environment; C7: environmental impact; and C8: local economic conditions. We will now apply the proposed method to this decision problem, explaining each step in detail.
Step 1: Problem structuring
The decision problem is structured hierarchically using AHP, with the goal at the top (selecting the best warehouse location), followed by the eight criteria (
C1 to
C8), and the five alternatives (A1 to A5) at the bottom, as shown in
Figure 2.
The problem above indicates criteria and alternatives are independent. Here, we focus on handling the time dependence between criteria.
Step 2: Pairwise comparison and local weight calculation
Let an expert perform pairwise comparisons of the criteria using the AHP scale (1–9) using the following PCM information:
PCM | C1 | C2 | C3 | C4 | C5 | C6 | C7 | C8 | Local weights |
C1 | 1 | 2 | 3 | 3 | 2 | 5 | 3 | 3 | 0.25 |
C2 | 1/2 | 1 | 2 | 2 | 1 | 4 | 2 | 2 | 0.15 |
C3 | 1/3 | 1/2 | 1 | 1 | 1/2 | 3 | 1 | 1 | 0.10 |
C4 | 1/3 | 1/2 | 1 | 1 | 1/2 | 3 | 1 | 1 | 0.10 |
C5 | 1/2 | 1 | 2 | 2 | 1 | 4 | 2 | 2 | 0.15 |
C6 | 1/5 | 1/4 | 1/3 | 1/3 | 1/4 | 1 | 1/2 | 1/2 | 0.05 |
C7 | 1/3 | 1/2 | 1 | 1 | 1/2 | 2 | 1 | 1 | 0.10 |
C8 | 1/3 | 1/2 | 1 | 1 | 1/2 | 2 | 1 | 1 | 0.10 |
We can use the eigenvector method to calculate the local weights of the criteria from the AHP. Then, the initial weights will be adjusted and updated based on the information on the time dependence between criteria.
Step 3: Designing the DBN
Next, we need to identify the relationship between criteria with time-dependent dependencies based on expert opinion. In this example, we assume that
C1,
C2, and
C8 exhibit time-dependent dependencies for simplicity. Hence, the weights of
C3–
C7 are static and will not be changed further. We then define the time steps for the DBN (e.g., monthly) and create a two-layered DBN structure. The structure represents two consecutive time steps, with the initial state of the criteria in the first layer and their state at the next time step in the second layer. The dependencies between criteria are represented by directed edges within and between layers, as shown in
Figure 3:
In this structure, expert opinion indicates that C1 and C2 at time t influence themselves at time t + 1, and C1 and C2 at time t also influence C8 at the t + 1 time step.
Step 4: Learning DBN parameters
Assume we use expert opinion to estimate each criterion’s transition probabilities between states. The experts provide the following conditional probability tables for C1, C2, and C8. Note that in the given context below, low (L), medium (M), and high (H) refer to each criterion’s different states or levels. These levels represent the degree to which the criterion is satisfied or its intensity in the decision-making process. For instance, for C1 (proximity to customers), a low level would imply that the warehouse location is relatively far from the customers.
In this example, we used expert opinion to estimate the transition probabilities for the time-dependent criteria in the DBN. These probabilities describe the dynamics of the criteria over time and their influence on other criteria in the network. Assume we have the following historical series for C1: (L, L, L, M, L, L, L, M, M, L, H, M, M, H, H, H, H, H, M, H, H, M, M, M, M, L, M, L, L, M, L, L, L, L, L, M, M, M, H, M, H, H, H, H, M, H, H, H, M, L, L, L, L, L, L, M, H, H, H, M, H, M), we can calculate the transition probabilities for C1 as:
:
| = Low | = Medium | = High |
Low | 0.7 | 0.3 | 0.1 |
Medium | 0.2 | 0.4 | 0.3 |
High | 0.1 | 0.3 | 0.6 |
For example, . We can use the historical data to complete the transition matrix above.
Let us assume we calculate the historical data to obtain the following transition matrices:
:
| = Low | = Medium | = High |
Low | 0.6 | 0.2 | 0.1 |
Medium | 0.3 | 0.5 | 0.3 |
High | 0.1 | 0.3 | 0.6 |
:
| = Low | = Medium | = High | = Low | = Medium | = High |
Low | 0.7 | 0.4 | 0.1 | 0.6 | 0.3 | 0.1 |
Medium | 0.2 | 0.4 | 0.3 | 0.3 | 0.5 | 0.3 |
High | 0.1 | 0.2 | 0.6 | 0.1 | 0.2 | 0.6 |
Step 5: Updating the criteria weights using the DBN
Using the learned DBN parameters, we compute the updated weights for criteria
C1,
C2, and
C8, considering their dependencies. For simplicity, let us assume the state probabilities for
C1,
C2, and
C8 provided by an expert as shown in
Table 2:
Then, we can use the steady-state distributions of
C1 and
C2 to calculate the joint probability of
C1 and
C2 as:
P(C1,C2) | P(C2 = Low) | P(C2 = Medium) | P(C2 = High) |
P(C1 = Low) | 0.20 | 0.12 | 0.08 |
P(C1 = Medium) | 0.20 | 0.12 | 0.08 |
P(C1 = High) | 0.10 | 0.06 | 0.04 |
The joint probability matrix,
P(
C1,
C2), can then be used to derive the marginal probability distribution of
C8 as:
After calculating the steady-state distribution of
C8, we can derive the marginal probabilities for criteria, as shown in
Table 3.
Next, we can calculate the RA and RS of the criteria, as shown in
Table 4.
Finally, we can derive the modified weights of the criteria as follows:
Note that we adjust the second term of the modified weights formula to ensure that the sum weights of C1, C2, and C8 after DBN should be the same as the original sum weights of C1, C2, and C8.
Step 6: Ranking the alternatives
We now have the updated weights for all criteria, including the time-dependent criteria. The following performance matrix for the alternatives with respect to the criteria and the weighted sum for each alternative using the updated criteria weights can also be given as shown in
Table 5:
The proposed method’s final rank shows the scores of A1, A3, and A4. Compared with the conventional AHP, the proposed method can account for extra information on time dependencies between criteria and modify the initial weight based on the information. The proposed method also reduces to the conventional AHP by ignoring the time-dependency information between criteria.
5. Discussions
Interdependency between criteria has been an important issue in MCDM. For example, ref. [
22] presented a hybrid modified TOPSIS integrated with preemptive goal programming to address interdependencies in the supplier selection process, a multiple-criteria decision-making issue. By comparing the proposed methodology to the analytical hierarchy process (AHP), the results show improved total value of purchasing while maintaining the equal total cost of purchasing, emphasizing the importance of considering interdependencies in supplier selection. Ref. [
23] proposed a non-orthogonal coordinate system-based multi-criteria decision-making method to address interdependent criteria in sustainability assessments. The method was applied to an electricity production case study, comparing results with traditional methods and conducting sensitivity analysis. Ref. [
24] presented a non-orthogonal coordinate system-based multi-criteria decision-making method that accounts for the interactions and interdependencies among the criteria in sustainability assessments. This approach offers a more comprehensive framework for prioritizing industrial systems, as it incorporates the complex relationships between evaluation criteria and sustainability-oriented decision-makers.
In this numerical example, we have showcased the practical application of our proposed method for combining DBNs and the AHP. This comprehensive framework addresses the limitations of the independence assumption in the AHP by integrating time dependencies between criteria using DBNs. It enables more accurate and informed decision-making in complex, dynamic environments where criteria dependencies evolve over time—something that previous AHP-related papers have not been able to address.
In this section, we compare our proposed method with existing approaches in the literature, focusing on the differences and advantages our method brings to the decision-making process. Conventional AHP and Fuzzy AHP [
1], ref. [
25] are widely used techniques for multi-criteria decision-making. However, both methods assume criteria independence, which may not always hold true in real-life scenarios. Some studies, such as [
26], have combined AHP with BNs to model dependencies among criteria. In addition, ref. [
27] presented a method for quantitative risk assessment of gas explosions in underground coal mines using BNs and FAHP. The proposed approach combines subjective and objective expert information for fuzzification, calculates real-time probabilities of potential risk events and risk factors, and determines the most critical risk factors through sensitivity analysis. Although this approach can capture static dependencies, it lacks the ability to model time-varying relationships.
In conclusion, our proposed method for integrating DBNs and AHP addresses some limitations found in existing approaches, particularly in modeling time-varying dependencies among criteria. By incorporating DBNs, our approach can capture complex and evolving relationships in multi-criteria decision-making problems, offering a more accurate and robust tool for decision-makers across various domains. Note that in this paper, we only consider two-layer DBNs, and the concept can be extended to consider more layers to handle more complicated time-dependent criteria in practice. In addition, the availability of transition matrices is still the major limitation of the proposed method.