A method and equipment for collaborative optimization of manufacturing value chains based on dynamic multi-criteria decision analysis

Abstract

This invention discloses a method and apparatus for collaborative optimization of manufacturing value chains based on dynamic multi-criteria decision analysis. The method involves defining the value chain collaborative optimization decision problem; acquiring decision-maker preference information; constructing a preference learning model by linking the decision-maker's context-dependent preferences with the dynamic decision-making environment and using a probabilistic value-driven ranking procedure to describe the preference information; inferring the parameters of the decision-maker's dynamic context preference model based on maximum likelihood estimation to predict the recommendation results of value chain collaborative solutions; predicting the decision-maker's future preference model and the corresponding ranking of alternative value chain collaborative solutions based on continuous new background inputs; and selecting the optimal collaborative solution based on the ranking results of the alternative value chain collaborative solutions to perform collaborative optimization of the manufacturing value chain. This invention mitigates cognitive bias by simulating stable, static preference evolution, thereby making the predicted collaborative solution recommendations more accurate.

Description

Technical Field

[0001] This method relates to the fields of multi-criteria decision-making and machine learning, specifically to a collaborative optimization method and device for the manufacturing value chain based on dynamic multi-criteria decision analysis. Background Technology

[0002] Multi-criteria ranking aims to provide decision-makers with ranking recommendations for alternatives from best to worst, based on a set of criteria. This type of decision problem is prevalent in various fields, including medical diagnostics, supply chain management, ecological planning, renewable energy, and education. The practical significance of ranking has attracted widespread attention in multi-criteria decision analysis (MCDA), leading to the development of corresponding methodologies.

[0003] Despite significant progress in multi-criteria ranking, almost all research rests on the implicit assumption that decision-making takes place in a static environment. In other words, decision-makers' preferences for a set of alternatives are assumed to follow a fixed pattern, remaining constant throughout the decision-making process. This, in turn, justifies the use of inferring static preference models, which are used to develop ranking recommendations. However, this approach is insufficient in decision-making environments where decision-makers' preferences can continuously change with the evolving decision-making context. For example, in a medical diagnostic scenario, a doctor's treatment preferences for a patient with a chronic illness may shift as the patient's condition progresses. In portfolio decisions within a financial context, investment managers' preferences for risk and return may fluctuate with market conditions. Similarly, in Weapon Targeting Assignment (WTA) missions, a commander's weapon and strategic choices may change due to weather changes, enemy movements, or new intelligence. Furthermore, a company's supplier priorities may shift due to external factors such as changing demand, raw material supply, or geopolitical risks. Given the prevalence of dynamic decision-making environments, considering the shifting preferences of decision-makers during the decision-making process is crucial. This underscores the need for a more comprehensive decision support approach that can build models of decision-makers' evolving preferences, thereby improving the accuracy of decision recommendations. Summary of the Invention

[0004] To address the aforementioned technical shortcomings, this method proposes a novel multi-criteria scheme ranking approach for collaborative scheme recommendation in a dynamic multi-base value chain collaborative decision-making environment, where decision-makers' preferences change with the environment.

[0005] A method and device for recommending value chain collaborative solutions based on dynamic multi-criteria decision analysis, characterized by the following steps:

[0006] S1. Definition of Value Chain Collaborative Optimization Decision-Making Problem;

[0007] S2. Obtain information on decision-makers' preferences;

[0008] S3. By linking decision-makers’ context-dependent preferences with the dynamic decision-making environment and using a probabilistic value-driven ranking procedure to describe preference information, a preference learning model is constructed.

[0009] S4. Based on maximum likelihood estimation, infer the parameters of the decision-maker's dynamic situational preference model to predict the recommendation results of value chain collaborative solutions;

[0010] S5. Based on the continuous new background input, predict the decision-maker's future preference model and the corresponding ranking of alternative value chain collaboration schemes;

[0011] S6. Select the best collaboration solution based on the ranking results of the alternative value chain collaboration solutions, and carry out value chain collaboration optimization in the manufacturing industry.

[0012] Furthermore, in step S1, a collaborative solution recommendation problem is considered in a dynamic multi-site value chain collaborative decision-making environment. The decision-maker's preferences are influenced by the gradually evolving environment. The goal is to rank a limited set of alternative value chain collaborative solutions to select the optimal collaborative solution. Formally, this is achieved using... Let represent the set of collaborative alternative solutions, and assume that all alternative solutions are based on One criterion The assessment was conducted, and express exist The performance above, without loss of generality, assumes The larger, exist The better the performance, the better each time. The dynamic decision-making environment consists of a... An environment vector of observable features To describe this, we assume that the constantly changing preferences of decision-makers are described as context-dependent preference models. , These models summarize exist The performance, thus at each time A comprehensive score is obtained. Because decision-makers' environment-related preferences are time-varying, different rankings can be derived from the set of alternatives. , ,use Indicates sorting China provides alternative solutions The position makes ,and Indicates the assignment to the position The alternative options make .

[0013] Furthermore, in step S2, the decision-maker can use a continuous time period. Complete sorting This paper aims to develop a preference learning method that uses provided preference information to express indirect preference information. The preference model constructed in this way It can be applied to time Given an unseen environment Dynamically predicting future decision outcomes under certain conditions .

[0014] Furthermore, in step S3, the method uses a linear-Gaussian state-space model to link the decision-maker's preferences related to the collaborative decision-making context with the dynamic collaborative solution recommendation decision-making environment, wherein the hidden state With decision-makers in time Preference Model Highly relevant, and by... Compared to its previous state and will With the current decision-making environment Relatedly, based on the evolution of Markov processes over time, the transfer function of the linear Gaussian state-space model... as follows:

[0015]

[0016] in, It is a regression coefficient matrix used for connecting... and ; It is random noise, follows a normal distribution, and has a mean vector. and diagonal covariance matrix ,Right now, Because it is assumed that decision-makers' preferences evolve smoothly over time according to the situation, a hyperparameter is used in the model. This controls the degree of smoothness.

[0017] This method uses a popular probabilistic model, the Luce model, to define the ranking information. At each time Given an environment-related preference model Regarding the likelihood of alternatives, the Luce model assumes that decision-makers express their preferences in the following way: The probability of being assigned to the first position is Then, alternative solutions The probability of being assigned to the second position is Continue allocating until a complete sort is produced. Based on the above generation procedure, using a preference model For the complete sorting of conditions The probability can be defined as follows:

[0018]

[0019] To aggregate the performance of alternatives across multiple criteria based on the decision-maker's dynamic preferences, this method employs an environment-related additive value function model. To summarize each alternative plan The performance across all criteria is as follows:

[0020]

[0021] in, This indicates that the decision-maker's time... For alternative solutions The overall score of perceived attractiveness, It is an alternative plan In the guidelines Marginal value on, marginal value function It is monotonically, non-decreasing, and normalized, therefore the overall score is... It is bounded in the interval [0,1].

[0022] Therefore, likelihood can be used The restatement is as follows:

[0023]

[0024] Subsequently, in order to convert the preference vector With hidden state In connection with these, a softmax transformation can be applied as follows:

[0025]

[0026] Please note, and Having the same dimensions The above formula ensures that the Gaussian variable is... To the definition Parameter vectors on a 3D simplex The conversion.

[0027] Furthermore, in step S4, the method uses maximum likelihood estimation to infer model parameters and the generation process. This task can be formulated as the following optimization problem:

[0028]

[0029] in Indicates parameters The maximum likelihood estimate, Indicates likelihood To express the likelihood, a Monte Carlo method can be used to obtain an unbiased approximation of the likelihood, and then a stochastic gradient descent-based algorithm can be applied to obtain the maximum likelihood estimate. The inferred parameters can be used to predict the ranking results of collaborative schemes in future scenarios.

[0030] The maximum likelihood estimation problem can be formulated as follows:

[0031]

[0032] in , It is a hyperparameter used in and A trade-off is made between the two, and the regularization term is used. It is achieved by the change in the slope of the sectional linear marginal value function.

[0033] Furthermore, in step S5, the proposal model can be based on time. Unseen context vectors Predicting decision-makers' future preferences and corresponding decision outcomes allows Indicates parameters The maximum likelihood estimate,

[0034] Hidden state and preference vector Predictions can be made in the following ways:

[0035]

[0036] Then the decision-maker in time The preference model at that time can predict as

[0037]

[0038] Based on the predicted composite value Will By sorting from best to worst, we can arrive at the most likely decision outcome. This allows us to obtain a ranking of alternative value chain collaboration schemes that align with decision-makers' preferences in a dynamic environment.

[0039] Furthermore, in step S6, the decision result predicted by the collaborative solution decision model is used... The highest priority alternative value chain collaboration solution is selected as the best collaboration solution.

[0040] Furthermore, an electronic device is characterized by comprising a memory and a processor, the memory being used to store a computer program, and the processor running the computer program to cause the electronic device to perform the value chain optimization method as described in any one of claims 1 to 7.

[0041] This method has the following beneficial effects and advantages:

[0042] 1. This method establishes a collaborative solution recommendation model framework based on a probabilistic graphical model, integrating an evolving decision-making context, a dynamic preference model, and a probabilistic ranking procedure. The model's structure is simple, clear, and meaningful, ensuring the transparency and interpretability of the dynamic decision-making process.

[0043] 2. In step S3, this method introduces a preference learning approach. This approach combines a state-space model to capture the conditional dependence of decision-makers' preferences on the changing environment with a Markov process that links the decision-makers' preferences at each decision-making stage, reflecting the dynamic changes in preferences. This mitigates cognitive bias by simulating a stable, static preference evolution, thereby making the predicted collaborative solution recommendations more accurate.

[0044] 3. In step S4, this method utilizes reparameterization techniques and the Adam optimizer to develop an algorithm for estimating model parameters using maximum likelihood. These advancements from machine learning ensure computational efficiency when searching for optimal solutions to difficult-to-solve likelihood functions. Therefore, the dynamic situational preference model constructed from the estimated parameters can be used to recommend future decisions in unknown situations, thus enabling faster provision of optimal solutions in collaborative decision-making scenarios.

[0045] This method considers scenarios where the decision-making environment is constantly evolving, leading to changes in preferences for multiple criteria. To handle dynamic decision-making environments, it proposes a novel probabilistic preference learning method for multi-criteria ranking. This method utilizes indirect preference information conditioned on the dynamic decision-making environment to construct an evolving value-based preference model. To this end, a linear-Gaussian state-space model is used, which describes a probabilistic ranking procedure that generates preference information at different stages of the decision-making process by linking the underlying time-varying probabilistic preference model to the evolving decision-making environment. The method also develops an inference algorithm that estimates the model's parameters by following the principle of maximizing probability. By drawing on and appropriately adapting technological advancements in machine learning, the computational efficiency of parametric inference is ensured. Therefore, the parameterized dynamic contextual preference model can be used to recommend future decision outcomes in unknown decision-making situations. In summary, this method aims to solve the problem of predicting the optimal collaborative solution that aligns with the decision-maker's personal preferences in a dynamic multi-base value chain collaborative decision-making environment. Attached Figure Description

[0046] Figure 1 This is a summary diagram of the method.

[0047] Figure 2 This is the model probability diagram of the method. Detailed Implementation

[0048] The method will be further described in detail below with reference to specific embodiments. These descriptions are for explanation purposes only and not for limitation.

[0049] like Figure 1 As shown, a method and device for recommending multi-site value chain collaborative solutions based on dynamic multi-criteria decision analysis includes the following steps: defining the value chain collaborative optimization decision problem, obtaining decision-maker preference information, modeling a dynamic preference model, inferring model parameters, predicting decision-maker preferences, and selecting the optimal solution.

[0050] Consider the problem of recommending collaborative solutions in a dynamic, multi-site value chain collaborative decision-making environment. Decision-makers' preferences are influenced by the evolving environment. The goal is to rank a limited set of alternative value chain collaborative solutions to select the optimal solution. Formally, using… Let represent the set of alternative value chain collaboration solutions, and assume that all alternative solutions are based on One criterion An evaluation was conducted. Order express exist The performance on the above. Without loss of generality, assume The larger, exist The better the performance. Each time The dynamic decision-making environment consists of a... An environment vector of observable features To describe this. Further, it is assumed that the decision-maker's changing preferences are described as context-dependent preference models. , These models summarize exist The performance, thus at each time A comprehensive score is obtained. Because decision-makers' environment-related preferences are time-varying, different rankings can be derived from the set of alternatives. , .use Indicates sorting China provides alternative solutions The position makes ,and Indicates the assignment to the position The alternative options make .

[0051] Decision-makers can use continuous time periods Complete sorting This paper aims to develop a preference learning method that expresses indirect preference information in the form of provided preference information. The preference model constructed in this way It can be applied to time Given an unseen environment Dynamically predicting future decision outcomes under certain conditions .

[0052] 3.1 Preferences and Dynamic Decision-Making Environments Dependent on the Context

[0053] We assume that a gradually evolving collaborative decision-making environment influences decision-makers' environment-related preferences. A reasonable assumption is that the model representing these preferences evolves smoothly. To link decision-makers' situation-related preferences to the dynamic decision-making environment, this method uses a linear-Gaussian state-space model. Here, the hidden states... With decision-makers in time Preference Model Highly relevant, and by... Compared to its previous state and will With the current decision-making environment Related to the evolution of Markov processes over time. The transition function of a linear Gaussian state-space model. as follows:

[0054] (1)

[0055] in, It is a regression coefficient matrix used for connecting... and ; It is random noise, follows a normal distribution, and has a mean vector. and diagonal covariance matrix ,Right now, Because it is assumed that decision-makers' preferences evolve smoothly over time according to the situation, a hyperparameter is used in the model. This controls the degree of smoothness. Therefore, the linear-Gaussian state-space model not only captures the decision-making environment... For hidden states The influence of this also preserves the continuous state. and The time dependency between them. State-space models characterize the static and dynamic patterns of decision-makers' smoothly evolving preferences, thus mitigating cognitive biases when constructing preference models. Based on the transition function... It is quite obvious that... .

[0056] 3.2 Value-Driven Probabilistic Ranking Procedure

[0057] This method uses a popular probabilistic model, the Luce model, to define the sorting information. At each time Given an environment-related preference model The Luce model assumes that decision-makers express their preferences in the following way: alternatives The probability of being assigned to the first position is Then, alternative solutions The probability of being assigned to the second position is Continue allocating until a complete sort is produced. Based on the above generation procedure, using a preference model... For the complete sorting of conditions The probability can be defined as follows:

[0058] (2)

[0059] To aggregate the performance of alternatives across multiple criteria based on the decision-maker's dynamic preferences, this method employs an environment-related additive value function model. To summarize each alternative plan The performance across all criteria is as follows:

[0060] (3)

[0061] in, This indicates that the decision-maker's time... For alternative solutions The overall score of perceived attractiveness, It is an alternative plan In the guidelines The marginal value on the boundary. Marginal value function. It is monotonically, non-decreasing, and normalized, therefore the overall score is... It is bounded in the interval [0,1].

[0062] Real value function model of decision-makers It is unknown beforehand. Therefore, a technique is needed to approximate the true value function model. This approximation can be performed using various types of marginal value functions (including linear, piecewise linear, spline, and generally monotonic functions). For simplicity, this method only considers the case of piecewise linear marginal value functions. Approximating the true value function model using a piecewise linear marginal value function can be viewed as a nonparametric technique because we do not need to make any assumptions about the form of the true marginal value function. By using a piecewise linear marginal value function for approximation, the true value function model... It can be described as:

[0063] (4)

[0064] in, It is a parameter vector, defined in On the simplex of dimension 1, a context-dependent preference model is represented. The inherent characteristics; It is a feature vector, entirely composed of alternative solutions. Performance across multiple criteria is determined, particularly independent of the decision-maker's dynamic preferences. In other words, the decision-maker's preference model is context-dependent. It is composed of preference vectors Parameterized. Therefore, likelihood can be used... The restatement is as follows:

[0065] (5)

[0066] Subsequently, in order to convert the preference vector With hidden state In connection with these, a softmax transformation can be applied as follows:

[0067] (6)

[0068] Please note, and Having the same dimensions The above formula ensures that the Gaussian variable is... To the definition Parameter vectors on a 3D simplex The conversion.

[0069] 3.3 Probabilistic Graphical Model and Generation Process

[0070] like Figure 2 As shown, solid nodes represent model parameters. and hyperparameters Blank nodes represent all latent variables (including...) , and Shaded nodes represent observable data (including environment). and preference information The arrows indicate the probabilistic dependencies between nodes. Based on the above membership structure, the likelihood of the entire model can be expressed as follows:

[0071] (7)

[0072] Please note the hidden state. Model parameters can be used and hyperparameters Write it recursively as follows:

[0073] (8)

[0074] in Through this method, it is possible to observe Follows some unknown distribution Its parameters are and However, the distribution It cannot be solved in a closed-form solution.

[0075] This method uses maximum likelihood estimation to infer the parameters and generation process of the collaborative solution recommendation decision model. This task can be formulated as the following optimization problem:

[0076] (9)

[0077] in Indicates parameters The maximum likelihood estimate, Indicates likelihood This represents likelihood. Due to probability... There is no closed form of the expression, therefore it is included in the likelihood. The integral term in the equation cannot be calculated through analysis. However, an unbiased likelihood approximation can be obtained using the Monte Carlo method, and then a maximum likelihood estimate can be obtained by applying a stochastic gradient descent-based algorithm. Specifically, the integral term Randomly select according to formula (8) The sample (using) (represented) to approximate, that is

[0078] (10)

[0079] It should be noted that, although when the covariance matrix When the value is large, the approximation in equation (10) may have a high variance, but it provides information about the parameters. The unbiased estimate of the likelihood gradient allows the search algorithm to converge to a stationary point. Subsequent experimental results also validate the effectiveness of this approximation method.

[0080] Using the approximation in formula (10), the likelihood Can be rewritten as

[0081] (11)

[0082] By applying the chain rule, the regression coefficient matrix expressed in formula (11) can be derived. and initial state The likelihood gradient. However, for the covariance matrix... The likelihood gradient cannot be directly calculated because of the following in equation (11): Not definitively based on This is conditional, because To address this challenge, this method employs a reparameterization technique, which reparameterizes the random variable... Equivalent representation is

[0083] (12)

[0084] in It is A dimensional vector, such that , It is 3D random variable, This represents the Hadamard product. Formula (8) can be restated as follows:

[0085] (13)

[0086] In the formula The conditions for determining it are ,and The randomness originates only from independent random variables. The covariance matrix can be obtained using this method. The likelihood gradient.

[0087] Based on the above analysis, the maximum likelihood estimation problem can be formulated as follows:

[0088] (14)

[0089] in , It is a hyperparameter used in and A trade-off is made between these factors. Regularization term. This is achieved as a change in the slope of the slicing marginal value function. Note that the model's hyperparameters include the smoothing coefficient. Subinterval number and regularization coefficient All hyperparameters can be specified through cross-validation.

[0090] This method uses the Adam optimizer to solve the optimization problem in Equation (11). This is a state-of-the-art algorithm for solving highly complex non-convex optimization problems. Adam utilizes a series of innovative computational features such as stochastic gradient descent and adaptive learning rate to ensure minimal memory requirements and the fastest convergence speed when handling data-intensive tasks.

[0091] Finally, this method can be based on time. Unseen context vectors Predict future policymakers' preferences and the corresponding ranking of multi-criteria options. Let... Indicates parameters The maximum likelihood estimate.

[0092] Hidden state and preference vector Predictions can be made in the following ways:

[0093]

[0094] Then the decision-maker in time The preference model at that time can predict as

[0095]

[0096] Based on the predicted composite value Will By sorting from best to worst, we can arrive at the most likely decision outcome. This allows for the acquisition of multi-criteria ranking results that align with decision-makers' preferences in future scenarios within a dynamic environment.

[0097] 6. Choose the best option.

[0098] Decision results predicted by the collaborative solution recommendation decision model The highest priority alternative value chain collaboration solution is selected as the best collaboration solution.

[0099] To investigate the ability of this method and the baseline method to predict the ranking of schemes in a dynamic environment, the following experiment was designed and implemented.

[0100] 7.1 Dataset and Experiment Setup

[0101] The experimental analysis, based on the generation of random data, considered the following factors of the problem: the number of alternative solutions ( ), number of criteria ( ), length of time series ( ), Dimensions of the environment vector ( ) and the coefficient for adjusting the level of consensus among decision-makers ( These five factors are all directly related to the development of preference models. Three levels were set: 1.0986, 5.2933, and 12.2060, corresponding to a value function of the decision-maker, respectively. alternative solutions Better than the other alternative solutions The probabilities are 0.5, 0.99, and 0.99999. Table 1 lists the levels of each factor considered in the analysis.

[0102] Table 1. Factor levels considered in the simulation experiment

[0103]

[0104] For a given combination (243 combinations) of the five factors listed in Table 1, 20 random datasets were simulated, according to... The uniform distribution of the factors yielded the performance of the alternatives on each criterion. By considering each combination of the five factors, a ranking problem instance was constructed, and a value function model was randomly generated. This represents the decision-maker's true value system. Specifically, in order to generate an environment vector... , An environmental feature , The derivation process is as follows:

[0105]

[0106] in, These are the initial values ​​of environmental characteristics. It is the rate of environmental change. Then, we can obtain it according to formula (1). The sequence, and according to Derive the weight vector Finally, each criterion marginal value function It was simulated in the following way

[0107]

[0108] in, This parameter determines the curvature of the marginal value function and is randomly generated from a uniform distribution within the interval [-10, 10]. Such a procedure can simulate convex or concave marginal value functions, reflecting different risk attitudes of decision-makers.

[0109] Then, using the assumed true preference model Preference information is generated, and a proposal method is applied to construct an estimated preference model based on this information. Specifically, this method uses a true value function model to calculate... True composite value According to the probability-value-driven ranking procedure in formula (2), the alternatives at each position are determined from best to worst as follows. For the first position, from the entire set of alternatives... Randomly select one alternative plan, each alternative plan The probability of being selected is ,in, It is a coefficient used to adjust the level of consensus among decision-makers, i.e. The second position is from the set. A candidate solution is randomly selected from the options. The probability of each candidate solution being selected is... Continue assigning values ​​until a complete sort is generated. .

[0110] To quantify the performance of these methods, three sorting quality metrics were used in the experiments: Kendall's Ranking difference measure and ranking consistency measure. This experiment compares the developed dynamic preference learning method with the following two baseline methods:

[0111] 1) The UTA method is an ordinal regression method that uses linear programming to estimate the parameters of the utility function. It uses information from subjective ranking of a set of stimuli or actions (weak-order comparison judgment) and multi-criteria evaluations of these actions to evaluate the additive utility function that combines multiple criteria into a composite criterion; the method is implemented using the cvxpy solver.

[0112] 2) The ANN-UTA algorithm uses artificial neural networks (ANNs) to derive the shape of the weights and marginal value functions without specifying feature points; the optimization objective is the same as the proposed model.

[0113] Experimental environment and parameter settings: The simulation was run on a server (Intel Xeon Silver 4214 CPU 2.20GHz, 128G memory, 480G solid-state drive, 4×RTX 2080Ti 11G), and Python 3.9 was used for data simulation, model inference and result analysis.

[0114] For each simulated problem instance, the experimental dataset was divided into two parts. First, to examine the predictive performance of the new alternatives, 70% of the alternatives were randomly selected from each set for training the model, while the remaining 30% were used as new alternatives to test their predictive performance. Second, to examine the predictive performance under new conditions, the top 70% of the preference information in the time series was selected for training, and the last 30% were used as alternatives under the new conditions for testing.

[0115] Regarding the model hyperparameters, the range of hyperparameter values ​​is defined here, along with other baseline hyperparameters required for successful model training during these model runs. To find the optimal values, cross-validation was performed to verify the ranking quality of different values. Specifically, four hyperparameters were tested:

[0116] 1) Smoothing coefficient of the proposed model ;

[0117] 2) Number of subintervals in the proposal model and UTA method ;

[0118] 3) Regularization coefficients of the proposed model ;

[0119] 4) Number of monotonic block components of ANN-UTA This is the only parameter that needs to be provided with a value before training this method.

[0120] 7.2 Experimental Results Presentation

[0121] 7.2.1 Experimental Evaluation of Predicting the Ranking of New Alternatives

[0122] The above three methods are used for each random dataset to compare Kendall's Performance in terms of RAM and RDM. Table 2 lists each consideration ( - Kendall's average at different levels RAM and RDM values.

[0123] Table 2 Kendall's scalar model, UTA, and ANN-UTA in the new alternative experiment. The mean and standard deviation of RAM and RDM values

[0124]

[0125] Under different problem settings, the proposed method outperforms both UTA and ANN-UTA. On average, the proposed model achieves Kendall's... The values ​​for RAM and RDM were 0.655, 0.365, and 0.935, respectively, while those for UTA were 0.635, 0.345, and 0.933. The three measurements for ANN-UTA were 0.585, 0.306, and 0.929, which were the lowest values. A Wilcoxon test was performed to verify the statistical significance of the results. Table 2 shows that the three measurements... p The values ​​are all less than 0.001. This confirms that the proposal method significantly outperforms UTA ​​on artificial datasets, and UTA shows stronger performance compared to ANN-UTA, indicating that the method can better obtain the ranking results of the schemes that conform to the decision-maker's preferences in dynamic decision-making environments.

[0126] 7.2.2 Experimental Evaluation of Predicting the Ranking of Alternative Solutions under New Environments

[0127] Since baseline UTA and ANN-UTA cannot predict alternatives with context vectors in new environments, a multilayer perceptron (MLP) was developed to predict decision-makers' preferences based on inferred preference models derived from UTA and ANN-UTA. This model can incorporate current context vectors... Compared with the previous parameter vector Combined, output the current parameter vector. Predictive value function model It can be specified using formula (4). The optimization objective of the MLP is the same as that proposed in this paper.

[0128] Table 3 shows the average Kendall's scores at different levels for the five factors considered. RAM and RDM values. The performance of the three methods shows the same trend as the problem setting changes. Specifically, the best-performing method utilizes the results obtained by that method. Compared to the case of prediction using the new alternative, the prediction performance of all three methods increases with the length of the time series (RAM and RDM values). ) and environment vector dimension ( The proposed method improves with increasing contextual information. Compared to UTA and ANN-UTA, it uses a state-space model to link environment-related preferences with the dynamic decision-making environment. Conversely, the two baseline methods incorporate MLP to consider contextualized predictions. All three methods improve predictive performance and are consistent with expectations when more contextual information is included. In summary, this method demonstrates better predictive performance in predicting the ranking of alternatives that align with decision-makers' preferences in dynamic environments, and can effectively address this type of alternative ranking problem.

[0129] Table 3. Kendall's proposed models, UTA, and ANN-UTA in the new environment experiment. The mean and standard deviation of RAM and RDM values.

[0130]

Claims

1. A collaborative optimization method for the manufacturing value chain based on dynamic multi-criteria decision analysis, characterized in that... Includes the following steps: S1. Definition of Value Chain Collaborative Optimization Decision-Making Problem; S2. Obtain information on decision-makers' preferences; S3. By linking decision-makers’ context-dependent preferences with the dynamic decision-making environment and using a probabilistic value-driven ranking procedure to describe preference information, a preference learning model is constructed. A linear-Gaussian state-space model is used to link decision-makers' preferences related to collaborative decision-making scenarios with the dynamic collaborative solution recommendation decision-making environment, where the hidden states... With decision-makers in time Preference Model Highly relevant, and by... Compared to its previous state and will With the current decision-making environment Relatedly, based on the evolution of Markov processes over time, the transfer function of the linear Gaussian state-space model... as follows: in, It is a regression coefficient matrix used for connecting... and ; It is random noise, follows a normal distribution, and has a mean vector. and diagonal covariance matrix ,Right now, Because it is assumed that decision-makers' preferences evolve smoothly over time according to the situation, a hyperparameter is used in the model. To control the degree of smoothness; Using an environment-related additive value function model To summarize each alternative plan The performance across all criteria is as follows: in, This indicates that the decision-maker's time... For alternative solutions The overall score of perceived attractiveness, It is an alternative plan In the guidelines Marginal value on, marginal value function It is monotonically, non-decreasing, and normalized, therefore the overall score is... It is bounded in the interval [0,1]. S4. Based on maximum likelihood estimation, infer the parameters of the decision-maker's dynamic situational preference model to predict the recommendation results of value chain collaborative solutions; Using maximum likelihood estimation to infer model parameters and the generation process, this task is formulated as the following optimization problem: in Indicates parameters The maximum likelihood estimate, The likelihood function is equal to The Monte Carlo method is used to obtain an unbiased likelihood approximation, and then a stochastic gradient descent-based algorithm is applied to obtain the maximum likelihood estimate. The covariance matrix is ​​calculated using a reparameterization technique. The likelihood gradient is used to infer the parameters, which are then used to predict the ranking of collaborative schemes in future scenarios. The maximum likelihood estimation problem is formulated as follows: in , It is a hyperparameter used in and A trade-off is made between the two, and the regularization term is used. It is achieved as a change in the slope of the slicing linear marginal value function; S5. Based on the continuous new background input, predict the decision-maker's future preference model and the corresponding ranking of alternative value chain collaboration schemes; S6. Select the best collaboration solution based on the ranking results of the alternative value chain collaboration solutions, and carry out value chain collaboration optimization in the manufacturing industry.

2. The manufacturing value chain collaborative optimization method based on dynamic multi-criteria decision analysis as described in claim 1, characterized in that: In step S1, based on the collaborative solution recommendation problem in a dynamic multi-site value chain collaborative decision-making environment, the decision-maker's preferences are influenced by the gradually evolving environment. The goal is to rank a limited set of alternative value chain collaborative solutions to select the optimal collaborative solution. Formally, this is achieved using... Let represent the set of value chain collaboration solutions, and assume that all alternative solutions are based on One criterion The assessment was conducted, and the order was made express exist The performance above, without loss of generality, assumes The larger, exist The better the performance, the better each time. The dynamic decision-making environment consists of a... An environment vector of observable features To describe this, we assume that the constantly changing preferences of decision-makers are described as context-dependent preference models. , These models summarize exist The performance, thus at each time A comprehensive score is obtained. Because decision-makers' environment-related preferences are time-varying, different rankings are derived from the set of alternatives. , ,use Indicates sorting China provides alternative solutions The position makes ,and Indicates the assignment to the position The alternative options make .

3. The manufacturing value chain collaborative optimization method based on dynamic multi-criteria decision analysis as described in claim 1, characterized in that: In step S2, the decision-maker uses a continuous time period. Complete sorting The goal is to develop a preference learning method that expresses indirect preference information in a specific form, based on the provided preference information. The preference model constructed in this way It can be applied to time Given an unseen environment Dynamically predicting future decision outcomes under certain conditions .

4. The manufacturing value chain collaborative optimization method based on dynamic multi-criteria decision analysis as described in claim 1, characterized in that: The ranking information is defined using the popular probabilistic model, the Luce model. At each time Given an environment-related preference model The Luce model assumes that decision-makers express their preferences in the following way: alternatives The probability of being assigned to the first position is Then, alternative solutions The probability of being assigned to the second position is Continue allocating until a complete sort is produced. Based on the above generation procedure, using a preference model For the complete sorting of conditions The probability is defined as follows: The performance of alternatives across multiple criteria is summarized based on the decision-maker's dynamic preferences. Therefore, likelihood can be used The restatement is as follows: Subsequently, in order to transform the preference vector With hidden state In connection, the softmax transformation is as follows: and Having the same dimensions The above formula ensures that the Gaussian variable is... To the definition Parameter vectors on a 3D simplex The conversion.

5. The manufacturing value chain collaborative optimization method based on dynamic multi-criteria decision analysis as described in claim 1, characterized in that: In step S5, the proposal model is based on time. Unseen context vectors Predicting decision-makers' future preferences and corresponding decision outcomes allows Indicates parameters The maximum likelihood estimate, Hidden state and preference vector Predict as follows: Then the decision-maker in time The preference model predicts for time Based on the predicted composite value Will By sorting from best to worst, we can arrive at the most likely decision outcome. This allows us to obtain a ranking of value chain collaboration schemes that align with decision-makers' preferences in a dynamic environment.

6. The manufacturing value chain collaborative optimization method based on dynamic multi-criteria decision analysis as described in claim 1, characterized in that: In step S6, the decision result is predicted by the value chain collaborative solution recommendation decision model. The highest priority alternative value chain collaboration solution is selected as the best collaboration solution.

7. An electronic device, characterized in that, The device includes a memory and a processor, the memory being used to store a computer program, and the processor running the computer program to cause the electronic device to perform the manufacturing value chain collaborative optimization method as described in any one of claims 1 to 6.

Images