Artificial intelligence aided decision-making method and system for standardized evaluation

Abstract

The invention discloses an artificial intelligence auxiliary decision-making method and system for standardized assessment, and the method comprises the steps: obtaining a decision-making scene and a decision-making target, constructing a standardized assessment model, carrying out the standardized assessment of decision variables through the standardized assessment model, and building a decision variable assessment system based on an assessment result. A neural network is adopted to construct a decision agent based on the decision variable evaluation system, a decision option set is obtained based on the decision agent, coordinated filtering is performed on the decision option set, initial decision options are obtained, the range of the initial decision options is narrowed, and real-time return and long-term return are balanced in the narrowed range to obtain an optimal decision. And optimizing the decision agent through reinforcement learning according to the adoption condition of each optimal decision. The method not only can provide comprehensive and accurate information support for the decision making process, but also can deeply analyze a complex decision making scene through an intelligent algorithm, and accurately excavates key factors and potential trends.

Description

Technical Field

[0001] The present invention relates to the field of intelligent decision-making, and in particular to an artificial intelligence-assisted decision-making method and system for standardized evaluation. Background Art

[0002] In today's society, decision-making problems in various fields are becoming more and more complex. Whether it is corporate strategic planning, project investment decisions, or government public policy formulation, many factors need to be considered. Traditional decision-making methods often rely on the experience and subjective judgment of decision makers, and it is difficult to comprehensively and accurately analyze complex decision-making problems. Therefore, it is necessary to introduce more scientific and objective decision-making methods to meet complex decision-making needs.

[0003] Traditional decision-making methods often rely on expert experience and subjective judgment, while the introduction of artificial intelligence technology provides new ideas and methods for construction project decision-making. It can realize rapid data processing, intelligent analysis and accurate prediction, and provide decision makers with more scientific and objective decision-making basis.

[0004] In AI-assisted decision-making, standardized evaluation is crucial to ensure the objectivity, accuracy, and reliability of decision-making. Therefore, it is necessary to establish a set of standardized evaluation methods to provide a basis for decision makers to choose appropriate algorithms. Summary of the invention

[0005] The purpose of the present invention is to provide an artificial intelligence-assisted decision-making method and system for standardized evaluation.

[0006] To achieve the above object, the present invention is implemented according to the following technical solutions:

[0007] The first aspect of the present invention provides an artificial intelligence-assisted decision-making method for standardized evaluation, comprising:

[0008] S100 obtains decision scenarios and decision goals;

[0009] S200 determines a standardized evaluation index according to the decision scenario, generates a multi-objective function according to the decision goal, and constructs a standardized evaluation model by combining the multi-objective function and the standardized evaluation index;

[0010] S300 performs variable analysis on the decision scenario to obtain decision variables, performs standardized evaluation on the decision variables using the standardized evaluation model, and establishes a decision variable evaluation system based on the evaluation results;

[0011] S400 uses a neural network to construct a decision agent based on the decision variable evaluation system, obtains a decision option set based on the decision agent, and performs coordinated filtering on the decision option set to obtain an initial decision option;

[0012] S500 narrows the scope of the initial decision options through unsupervised learning, and obtains the optimal decision by weighing the immediate reward and the long-term reward based on the decision maker agent within the narrowed scope;

[0013] S600 optimizes the decision-making agent through reinforcement learning according to the adoption of each optimal decision.

[0014] As a further method, the method of generating a multi-objective function according to the decision objective includes:

[0015] Obtain the expected level of each decision target and construct a multi-objective function by minimizing the deviation between each decision target and the expected level. The expression of the multi-objective function is:

[0016]

[0017] Where n is the number of decision targets, ω i is the weight coefficient of the i-th decision target, y i is the actual value of the i-th decision target, d i is the expected level of the i-th decision target, α i For z ij The degree of influence on the decision-making target deviation, m i is the number of direct cost factor sets of the ith decision objective, z ij is the jth factor in the set of direct cost factors of the i-th decision objective, β i for u ik The degree of influence on the decision-making target deviation, l i is the number of indirect cost factor sets for the i-th decision objective, u ik is the kth factor in the set of indirect cost factors for the i-th decision objective.

[0018] As a further method, the method of constructing a standardized evaluation model by combining the multi-objective function and the standardized evaluation index includes:

[0019] The Pearson coefficient between standardized evaluation indicators is calculated to measure the linear correlation between standardized evaluation indicators, and a linear regression correlation function is constructed for standardized evaluation indicators with a Pearson coefficient greater than 0.4;

[0020] The standardized evaluation indicators are substituted into the multi-objective function, and the linear regression correlation function between the standardized evaluation indicators is used as a constraint condition to form a standardized evaluation model.

[0021] As a further method, the method of performing standardized evaluation on the decision variables through the standardized evaluation model and establishing a decision variable evaluation system based on the evaluation results includes:

[0022] Conduct standardized evaluation on decision variables through standardized evaluation model to obtain standardized scores for each decision variable;

[0023] The weight value of each decision variable is calculated by the entropy weight method according to the standardized score of the decision variable and the variability of the decision variable;

[0024] Construct a standardized matrix of decision variables based on the weight value of each decision variable;

[0025] Calculate the standard gradient threshold of the decision variable, and establish a decision indicator system based on the standardized matrix and the standard gradient threshold. The calculation formula of the standard gradient threshold is:

[0026]

[0027] Among them, α is the scaling factor, i is the index of the evaluation object, m is the number of decision variables, and λ j is the weight value of the jth decision variable, y ij is the standardized score of the i-th evaluation object on the j-th decision variable, is the standardized score of the evaluation of the ith evaluation object on all decision variables, and ψ is the number of evaluation objects.

[0028] As a further method, the method of constructing a decision-making agent using a neural network based on the decision variable evaluation system includes:

[0029] The network architecture of decision-making agents based on neural networks is as follows:

[0030] The input layer is used to receive decision variables. The number of nodes in the input layer is equal to the number of decision variables.

[0031] The middle layer is used to integrate the decision indicator system and perform gradient division on the decision variables of the input layer;

[0032] The model layer creates a decision model based on the decision tree and makes decisions based on the gradient classification of the middle layer;

[0033] The output layer is used to output the decision results. If it is a decision variable value problem, the output is the decision value of the decision variable; if it is a decision variable selection problem, the output decision value is converted from 0 to 1. When the decision value is 0, it means that the decision variable is not selected, and the decision value is 1, which means that the decision variable is selected.

[0034] Collect historical decision data to train the decision agent, and use the back-propagation algorithm to calculate the gradient of the loss function to the bias of each layer of the decision agent network architecture;

[0035] The stochastic gradient descent algorithm is used to update the parameters of the decision agent according to the gradient of each layer bias to obtain the final decision agent.

[0036] As a further method, the method of obtaining a decision option set based on the decision agent, and performing coordinated filtering on the decision option set to obtain an initial decision option includes:

[0037] Obtain the internal adjustable data and external influencing data during decision-making, combine them as decision data, input the decision data into the decision agent, and obtain a decision option set that satisfies the decision model;

[0038] Calculate the similarity between decision options in the decision option set. The expression is:

[0039]

[0040] Where D(A,B) is the similarity between decision options A and B. The smaller the value, the more similar the two decision options are. n is the number of features in the decision options. ξ i is the feature weight factor of the i-th feature, a i and b i They are respectively represented as the values ​​of decision options A and B on the i-th feature, δ is the distance adjustment parameter, and are respectively represented as vectors of decision options A and B, is the Euclidean distance between decision options A and B, is the sum of the moduli of decision options A and B, ∈ is the covariance adjustment parameter, ∑ -1 It is expressed as the inverse matrix of the covariance between decision options A and B, s i is the feature scaling factor, used to adjust the scale of the i-th feature, is the importance adjustment parameter, I i is the information gain of the i-th feature in the decision option, H i is the information entropy of the i-th feature in the decision option, η is the adjustment factor to avoid the denominator of the importance function term being zero when the information entropy is zero, and is taken as 0.01;

[0041] The similarity is normalized from 0 to 1, and a partitioning criterion is set. The decision option set is clustered according to the partitioning criterion to obtain several clusters. From each cluster, the decision option combination with the largest information gain is selected as the initial decision option.

[0042] As a further method, the Gaussian mixture model is applied to group the initial decision options to obtain the probability of the group to which each decision option belongs, which is expressed as:

[0043]

[0044] Among them, p(k|x i ) is expressed as the probability that the i-th decision option belongs to the k-th group, π k is the mixing coefficient of the kth Gaussian distribution, Expressed as the probability density function of the kth Gaussian distribution, x i is the feature vector of the i-th decision option, μ k is the mean vector of the kth Gaussian distribution, ∑ k is the covariance matrix of the kth Gaussian distribution, υ is the number of groups;

[0045] The t-distributed random neighbor embedding is used to reduce the dimension and extract the group features. The group features are represented by the central variable and the covariance matrix, and the expression is:

[0046]

[0047] Among them, c k is the central variable of the kth group in the low-dimensional space, N k is the number of data points in the kth group, N is the number of decision options in the initial decision options, p(k|x i ) is expressed as the probability that the i-th decision option belongs to the k-th group, y i It is represented as the representation of the i-th decision option in the low-dimensional space after dimensionality reduction. is the covariance matrix of the kth group in the low-dimensional space, and T represents the mapping transformation operation;

[0048] The isolation forest algorithm is used to remove outliers that do not meet the group characteristics and narrow the range of decision options.

[0049] As a further method, the method for weighing the immediate reward and the long-term reward based on the decision maker agent in the narrowed scope includes:

[0050] Create a game model based on the Nash equilibrium algorithm, defining immediate returns and long-term returns as game participants;

[0051] Obtaining the immediate return function and the long-term return function, and incorporating the immediate return and the long-term return function into the game model as the profit function of the game participants;

[0052] The game strategy method is used to solve the game model and obtain the Nash equilibrium solution. Among the Nash equilibrium solutions, the decision plan with the largest sum of the immediate reward function and the long-term reward function is selected as the optimal decision.

[0053] As a further method, the method of optimizing the decision-making agent by reinforcement learning according to the adoption of each optimal decision includes:

[0054] Construct a reward and punishment mechanism based on the adoption of each optimal decision. If the decision is successfully adopted, a positive reward will be given, and if the decision is not adopted, a penalty will be given;

[0055] The decision agent is optimized to maximize the future cumulative reward and is updated by minimizing the KL divergence. The expression is:

[0056]

[0057] Among them, π new is the updated decision, ∏ is the strategy set adopted by the decision agent, and D KL Expressed as KL divergence, π(·|s t ) is the decision agent in state s t The strategy under t is the probability distribution of the decision agent at time t, i.e., the number of decisions made, is the state of the decision agent, the current strategy selects different actions, α is the reward adjustment parameter, Q(s t ,·) means in state s t The action value function for taking different actions under π (s t ) is a normalization parameter that ensures that the probability distribution of the target strategy sums to 1.

[0058] A second aspect of the present invention provides an artificial intelligence-assisted decision-making system for standardized evaluation, comprising:

[0059] Decision collection module, used to obtain decision scenarios and decision goals;

[0060] The standardized evaluation module is used to determine the standardized evaluation indicators according to the decision-making scenario, generate a multi-objective function according to the decision-making goal, and build a standardized evaluation model by combining the multi-objective function and the standardized evaluation indicators;

[0061] The decision variable evaluation module is used to perform variable analysis on the decision scenario, obtain decision variables, perform standardized evaluation on the decision variables through a standardized evaluation model, and establish a decision variable evaluation system based on the evaluation results;

[0062] The decision agent module is used to construct a decision agent using a neural network based on the decision variable evaluation system, and coordinate and filter the decision data through the decision agent to obtain the initial decision options;

[0063] The decision-making process module is used to narrow the scope of initial decision options through unsupervised learning, and obtain the optimal decision based on the trade-off between immediate rewards and long-term rewards of the decision-maker agent within the narrowed scope;

[0064] The reinforcement learning module is used to optimize the decision-making agent through reinforcement learning based on the adoption of the best decision each time.

[0065] Compared with the prior art, the embodiments of the present invention have at least the following advantages or beneficial effects:

[0066] (1) The present invention determines standardized evaluation indicators according to decision-making scenarios and generates multi-objective functions according to decision-making goals. The standardized evaluation model constructed can comprehensively consider the impact of various factors on decision-making and improve the accuracy and reliability of decision-making.

[0067] (2) The present invention helps decision-makers make more sustainable decisions by weighing the immediate returns and long-term returns, avoiding the problem of focusing only on short-term benefits and ignoring long-term impacts, and making decisions more forward-looking.

[0068] (3) The present invention continuously updates the decision-making agent, so that the decision-making agent can gradually learn to avoid being punished while pursuing high-value actions, so that the decision-making agent can continuously learn and improve. As the decision-making continues, the agent can adjust its own strategy based on feedback, improving the adaptability and flexibility of decision-making. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] Figure 1 The present invention is a flowchart of the steps of an artificial intelligence-assisted decision-making method for standardized evaluation in an embodiment of the present invention.

[0070] Figure 2 This is a logical diagram of an artificial intelligence-assisted decision-making system for standardized evaluation in an embodiment of the present invention. DETAILED DESCRIPTION

[0071] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0072] Reference Figure 1 As shown, the present invention provides an artificial intelligence-assisted decision-making method for standardized evaluation, comprising:

[0073] S100 obtains decision scenarios and decision goals;

[0074] In actual evaluation, an implementation example of decision making for electronic products:

[0075] Decision-making scenario: An electronic product manufacturer needs to decide on the production plan for the next quarter. The company faces problems such as market demand uncertainty, raw material price fluctuations, and production capacity constraints. At the same time, market competition is fierce, and customers have high requirements for product quality and delivery time;

[0076] Decision-making goal: maximize profits while meeting market demand, while ensuring product quality and on-time delivery. Specific goals include: profit maximization: maximize profits by reasonably arranging production quantities, controlling costs, and increasing sales revenue; market demand satisfaction rate is not less than 90%: ensure that the products produced can meet market demand and avoid inventory backlogs or shortages; product quality qualification rate is not less than 95%: strictly control product quality and improve customer satisfaction.

[0077] S200 determines a standardized evaluation index according to the decision scenario, generates a multi-objective function according to the decision goal, and constructs a standardized evaluation model by combining the multi-objective function and the standardized evaluation index;

[0078] In the actual evaluation, the standardized evaluation indicators are determined as follows: raw material cost proportion, production efficiency, inventory turnover rate, product defective rate and market share; let profit be P, market demand satisfaction rate be D, product quality qualification rate be Q, production cost be C, production quantity be N, and the weight coefficients be 0.3, 0.2, 0.2, 0.1, and 0.2 respectively. The expected level of the decision-making target is: if the profit increases by 10%, the expected profit is 5.5 million yuan, the expected level of market demand satisfaction rate is not less than 90%, and the expected level of product quality qualification rate is not less than 95%. Further construct a multi-objective function, calculate the Pearson coefficient between the standardized indicators, and obtain the Pearson coefficient of raw material cost proportion and production efficiency as -0.4, indicating that the two have a certain negative correlation. Create a linear regression correlation function for these two standardized indicators, substitute the standardized evaluation indicators into the multi-objective function, and use the linear regression correlation function between the standardized evaluation indicators as constraints to form a standardized evaluation model.

[0079] S300 performs variable analysis on the decision scenario to obtain decision variables, performs standardized evaluation on the decision variables using the standardized evaluation model, and establishes a decision variable evaluation system based on the evaluation results;

[0080] It is necessary to explain that by evaluating the decision variables through the standardized evaluation model, we can understand the degree of influence of each decision variable on the decision target and its performance under the standardized evaluation index. Based on the evaluation results of the decision variables, a comprehensive decision variable evaluation system is established. This system can help decision makers better understand the importance of each decision variable, their relationship and their impact on the decision results;

[0081] In the actual evaluation, the decision variables include the purchase volume of raw materials, the investment in production equipment upgrades, and the length of employee overtime. The decision variables are evaluated through a standardized evaluation model and analyzed to obtain the following results: if the purchase volume of raw materials is increased, the proportion of raw material costs can be reduced according to market price fluctuations and historical data forecasts, but inventory costs may increase. Assuming that the purchase volume of raw materials increases by 20%, it can be evaluated that the proportion of raw material costs can be reduced by 5%, but the inventory turnover rate can be reduced by 3%. Substitute it into the multi-objective function to calculate the score;

[0082] In the actual evaluation, after obtaining the standardized score of each decision variable, the weight value of each decision variable is calculated by the entropy weight method according to the standardized score of the decision variable and the degree of variability of the decision variable. By analyzing the company's historical data, it is determined that the standardized score of the raw material procurement volume is 0.7 and the degree of variability is 0.6; the standardized score of the production equipment upgrade investment is 0.65 and the degree of variability is 0.55; the standardized score of the employee overtime hours is 0.68 and the degree of variability is 0.6. The entropy weight method is used to calculate that the weight value of the raw material procurement volume is 0.32, the weight value of the production equipment upgrade investment is 0.3, and the weight value of the employee overtime hours is 0.38;

[0083] In the actual evaluation, a standardized matrix of decision variables is constructed based on the weight value of each decision variable, and the standard gradient threshold of decision variables is calculated. Based on the data of the company in the past three years, the scaling factor is determined to be 1.5. The evaluation object is the operating conditions of the company in each quarter. There are 12 evaluation objects in total and 3 decision variables. The weight value of the raw material procurement volume is 0.32. The standardized scores of the evaluation objects on this decision variable are [0.72, 0.78, 0.7, 0.75, 0.8, 0.76, 0.74, 0.79, 0.73, 0.77, 0.71, 0.76], and the standardized scores of the evaluation objects on all decision variables are [0.74, 0.8, 0.72, 0.76, 0.79, 0.77, 0.75, 0.81, 0.74, 0.78, 0.72, 0.77]. The standard gradient threshold of each decision variable is further calculated, and a decision indicator system is established based on the standardized matrix and standard gradient threshold.

[0084] S400 uses a neural network to construct a decision agent based on the decision variable evaluation system, obtains a decision option set based on the decision agent, and performs coordinated filtering on the decision option set to obtain an initial decision option;

[0085] It should be explained that after training, decision agents can generate a series of possible decision options for specific decision scenarios. These decision options are based on the agent's analysis and evaluation of decision variables, and represent different decision directions and strategies. For example, in corporate strategic decision-making, decision agents may give different decision options such as expanding market share, reducing costs, and launching new products; coordinated filtering of the decision option set is to screen and optimize the decision options generated by the decision agent according to specific standards and requirements, remove some less ideal decision options, and obtain an initial decision option set that has been preliminarily screened, providing a basis for further decision analysis and selection.

[0086] In the actual evaluation, the network architecture was constructed: the number of input layer nodes was 3, the middle layer had 5 nodes, and the historical decision data of the enterprise in the past five years was collected, with a total of 200 sets of data. The back propagation algorithm was used to calculate the gradient of the loss function to the bias of each layer of the decision agent network architecture, and the stochastic gradient descent algorithm was used to update the parameters of the decision agent according to the gradient of the bias of each layer. After many experiments, the learning rate was determined to be 0.009, and the final decision agent was obtained after 700 iterations.

[0087] In the actual evaluation, internal adjustable data (inventory level, production capacity) and external influencing data (market demand forecast, competitor strategy) are obtained when making decisions, and combined as decision data. Through internal enterprise data statistics, it is determined that the current inventory level is 1,000 electronic products and the production capacity is 500 pieces per day. Through market research, it is learned that the market demand growth forecast is 8%, and competitors have launched a new product line. The decision data is input into the decision-making agent to obtain a set of decision options that meet the decision-making model, a total of 11 decision options, including (increase the purchase of raw materials + increase investment in production equipment upgrades + increase employee overtime hours), (keep the purchase of raw materials unchanged + reduce investment in production equipment upgrades + do not increase employee overtime hours), etc.;

[0088] In the actual evaluation, the decision options have 5 features, and the feature weight factors are [0.22, 0.2, 0.18, 0.2, 0.2]. The distance adjustment parameter is 0.4, the covariance adjustment parameter is 0.3, the feature scaling factor is [1.15, 1.12, 1.18, 1.16, 1.14], the importance adjustment parameter is 0.35, the adjustment factor is 0.01, and the similarity between the decision options is calculated. The similarity is normalized from 0 to 1, and the partition standard is set to 0.45. The decision option set is clustered according to the partition standard to obtain several clusters. From each cluster, the decision option combination with the largest information gain is selected as the initial decision option.

[0089] S500 narrows the scope of the initial decision options through unsupervised learning, and obtains the optimal decision by weighing the immediate reward and the long-term reward based on the decision maker agent within the narrowed scope;

[0090] It is necessary to explain that the use of unsupervised learning to process the initial decision options can automatically identify similar decision options and cluster or group them by analyzing the characteristics and potential relationships of the decision options, thereby narrowing the scope of the decision options. For example, it may be found that some decision options have similar characteristics in some aspects, and they are grouped together, which can reduce the number of decision options that need to be further considered;

[0091] In the actual evaluation, the Gaussian mixture model is applied to group the initial decision options, and the probability of each decision option belonging to a group is obtained. The dimension is reduced by t-distributed random neighbor embedding, and the group characteristics are extracted. The outliers that do not meet the group characteristics are removed using the isolation forest algorithm to narrow the range of decision options. The expression of the immediate reward function is obtained as follows: Where ρ is the coefficient of profit on immediate return, set to 0.7, P is the profit function, is the coefficient of cost on immediate return, set to -0.3, Q is the product quality pass rate function; the expression of the long-term return function is: R2 = θ × P + ζ × Q, where θ is the coefficient of product quality pass rate on long-term return, set to 0.6, ζ is the coefficient of product quality pass rate on long-term return, set to 0.4; the immediate return and long-term return functions are incorporated into the game model as the profit function of the game participants;

[0092] In the actual evaluation, the game strategy method is used to solve the game model and obtain the Nash equilibrium solution. There are four decision options in the narrowed decision options:

[0093] Plan 1: moderately increase the purchase volume of raw materials, make a certain degree of investment in production equipment upgrades, and reasonably arrange employee overtime hours; Plan 2: keep the purchase volume of raw materials unchanged, reduce the investment in production equipment upgrades, and do not increase employee overtime hours; Plan 3: significantly increase the purchase volume of raw materials, increase the investment in production equipment upgrades, and extend employee overtime hours; Plan 4: slightly increase the purchase volume of raw materials, moderately invest in production equipment upgrades, and appropriately increase employee overtime hours; calculate the sum of their immediate return function and long-term return function respectively, and find that Plan 4 performs best under the comprehensive consideration of immediate return and long-term return. It can better meet market demand and product quality requirements while ensuring a certain profit. The optimal decision is: slightly increase the purchase volume of raw materials, and increase the purchase volume by 10% on the original basis. That is, the monthly purchasing volume will increase from 800 pieces to about 880 pieces, investment will be made in upgrading production equipment to 700,000 yuan, and the overtime hours of employees will be appropriately increased to 10 hours per month. Flexible overtime arrangements can be made, such as arranging overtime during periods of heavier production tasks, and giving employees corresponding overtime subsidies and benefits to improve their enthusiasm and work efficiency.

[0094] S600 optimizes the decision-making agent through reinforcement learning according to the adoption of each optimal decision.

[0095] In the actual evaluation, a reward and punishment mechanism is established according to the adoption of each optimal decision. If the optimal decision is successfully adopted and the company's profit growth, market demand satisfaction rate and product quality qualification rate meet expectations, a positive reward is given, and the reward value is set to 8. If the decision is not adopted or the effect after adoption is not good, a penalty is given, and the penalty value is set to -4. The decision-making agent is optimized to maximize the future cumulative rewards, and the decision-making agent is updated by minimizing the KL divergence. By continuously updating the parameters of the decision-making agent, it can better adapt to different decision-making scenarios and goals.

[0096] In this embodiment, the method for generating a multi-objective function according to the decision objective includes:

[0097] Obtain the expected level of each decision target and construct a multi-objective function by minimizing the deviation between each decision target and the expected level. The expression of the multi-objective function is:

[0098]

[0099] Where n is the number of decision targets, ω i is the weight coefficient of the i-th decision target, y i is the actual value of the i-th decision target, d i is the expected level of the i-th decision target, α i For z ijThe degree of influence on the decision-making target deviation, m i is the number of direct cost factor sets of the ith decision objective, z ij is the jth factor in the set of direct cost factors of the i-th decision objective, β i for u ik The degree of influence on the decision-making target deviation, l i is the number of indirect cost factor sets for the i-th decision objective, u ik is the kth factor in the set of indirect cost factors for the i-th decision objective.

[0100] In this embodiment, the method of constructing a standardized evaluation model by combining the multi-objective function and the standardized evaluation index includes:

[0101] The Pearson coefficient between standardized evaluation indicators is calculated to measure the linear correlation between standardized evaluation indicators, and a linear regression correlation function is constructed for standardized evaluation indicators with a Pearson coefficient greater than 0.4;

[0102] The standardized evaluation indicators are substituted into the multi-objective function, and the linear regression correlation function between the standardized evaluation indicators is used as a constraint condition to form a standardized evaluation model.

[0103] In this embodiment, the method of performing standardized evaluation on the decision variables through the standardized evaluation model and establishing a decision variable evaluation system based on the evaluation results includes:

[0104] Conduct standardized evaluation on decision variables through standardized evaluation model to obtain standardized scores for each decision variable;

[0105] The weight value of each decision variable is calculated by the entropy weight method according to the standardized score of the decision variable and the variability of the decision variable;

[0106] Construct a standardized matrix of decision variables based on the weight value of each decision variable;

[0107] Calculate the standard gradient threshold of the decision variable, and establish a decision indicator system based on the standardized matrix and the standard gradient threshold. The calculation formula of the standard gradient threshold is:

[0108]

[0109] Among them, α is the scaling factor, i is the index of the evaluation object, m is the number of decision variables, and λ j is the weight value of the jth decision variable, y ij is the standardized score of the i-th evaluation object on the j-th decision variable, is the standardized score of the evaluation of the ith evaluation object on all decision variables, and ψ is the number of evaluation objects.

[0110] In this embodiment, the method of constructing a decision agent using a neural network based on the decision variable evaluation system includes:

[0111] The network architecture of decision-making agents based on neural networks is as follows:

[0112] The input layer is used to receive decision variables. The number of nodes in the input layer is equal to the number of decision variables.

[0113] The middle layer is used to integrate the decision indicator system and perform gradient division on the decision variables of the input layer;

[0114] The model layer creates a decision model based on the decision tree and makes decisions based on the gradient classification of the middle layer;

[0115] The output layer is used to output the decision results. If it is a decision variable value problem, the output is the decision value of the decision variable; if it is a decision variable selection problem, the output decision value is converted from 0 to 1. When the decision value is 0, it means that the decision variable is not selected, and the decision value is 1, which means that the decision variable is selected.

[0116] Collect historical decision data to train the decision agent, and use the back-propagation algorithm to calculate the gradient of the loss function to the bias of each layer of the decision agent network architecture;

[0117] The stochastic gradient descent algorithm is used to update the parameters of the decision agent according to the gradient of each layer bias to obtain the final decision agent.

[0118] In this embodiment, the method of obtaining a decision option set based on the decision agent, and performing coordinated filtering on the decision option set to obtain an initial decision option includes:

[0119] Obtain the internal adjustable data and external influencing data during decision-making, combine them as decision data, input the decision data into the decision agent, and obtain a decision option set that satisfies the decision model;

[0120] Calculate the similarity between decision options in the decision option set. The expression is:

[0121]

[0122] Where D(A,B) is the similarity between decision options A and B. The smaller the value, the more similar the two decision options are. n is the number of features in the decision options. ξ i is the feature weight factor of the i-th feature, a i and b i They are respectively represented as the values ​​of decision options A and B on the i-th feature, δ is the distance adjustment parameter, and are respectively represented as vectors of decision options A and B, is the Euclidean distance between decision options A and B, is the sum of the moduli of decision options A and B, ∈ is the covariance adjustment parameter, ∑ -1 It is expressed as the inverse matrix of the covariance between decision options A and B, s i is the feature scaling factor, used to adjust the scale of the i-th feature, is the importance adjustment parameter, I i is the information gain of the i-th feature in the decision option, H i is the information entropy of the i-th feature in the decision option, η is the adjustment factor to avoid the denominator of the importance function term being zero when the information entropy is zero, and is taken as 0.01;

[0123] The similarity is normalized from 0 to 1, and a partitioning criterion is set. The decision option set is clustered according to the partitioning criterion to obtain several clusters. From each cluster, the decision option combination with the largest information gain is selected as the initial decision option.

[0124] In this embodiment, the Gaussian mixture model is applied to group the initial decision options to obtain the probability of the group to which each decision option belongs, which is expressed as:

[0125]

[0126] Among them, p(k|x i ) is expressed as the probability that the i-th decision option belongs to the k-th group, π k is the mixing coefficient of the kth Gaussian distribution, Expressed as the probability density function of the kth Gaussian distribution, x i is the feature vector of the i-th decision option, μ k is the mean vector of the kth Gaussian distribution, ∑ k is the covariance matrix of the kth Gaussian distribution, υ is the number of groups;

[0127] The t-distributed random neighbor embedding is used to reduce the dimension and extract the group features. The group features are represented by the central variable and the covariance matrix, and the expression is:

[0128]

[0129] Among them, c k is the central variable of the kth group in the low-dimensional space, N k is the number of data points in the kth group, N is the number of decision options in the initial decision options, p(k|x i ) is expressed as the probability that the i-th decision option belongs to the k-th group, y iIt is represented as the representation of the i-th decision option in the low-dimensional space after dimensionality reduction. is the covariance matrix of the kth group in the low-dimensional space, and T represents the mapping transformation operation;

[0130] The isolation forest algorithm is used to remove outliers that do not meet the group characteristics and narrow the range of decision options.

[0131] In this embodiment, the method for weighing the immediate return and the long-term return based on the decision maker agent within the narrowed scope includes:

[0132] Create a game model based on the Nash equilibrium algorithm, defining immediate returns and long-term returns as game participants;

[0133] Obtaining the immediate return function and the long-term return function, and incorporating the immediate return and the long-term return function into the game model as the profit function of the game participants;

[0134] The game strategy method is used to solve the game model and obtain the Nash equilibrium solution. Among the Nash equilibrium solutions, the decision plan with the largest sum of the immediate reward function and the long-term reward function is selected as the optimal decision.

[0135] In this embodiment, the method of optimizing the decision-making agent through reinforcement learning according to the adoption of each optimal decision includes:

[0136] Construct a reward and punishment mechanism based on the adoption of each optimal decision. If the decision is successfully adopted, a positive reward will be given, and if the decision is not adopted, a penalty will be given;

[0137] The decision agent is optimized to maximize the future cumulative reward and is updated by minimizing the KL divergence. The expression is:

[0138]

[0139] Among them, π new is the updated decision, ∏ is the strategy set adopted by the decision agent, and D KL Expressed as KL divergence, π(·|s t ) is the decision agent in state s t The strategy under t is the probability distribution of the decision agent at time t, i.e., the number of decisions made, is the state of the decision agent, the current strategy selects different actions, α is the reward adjustment parameter, Q(s t ,·) means in state s t The action value function for taking different actions under π (s t ) is a normalization parameter that ensures that the probability distribution of the target strategy sums to 1.

[0140] Reference Figure 2 As shown, the present invention provides an artificial intelligence-assisted decision-making system for standardized evaluation, comprising:

[0141] Decision collection module, used to obtain decision scenarios and decision goals;

[0142] A standardized evaluation module, used to determine a standardized evaluation index according to the decision scenario, generate a multi-objective function according to the decision goal, and construct a standardized evaluation model by combining the multi-objective function and the standardized evaluation index;

[0143] A decision variable evaluation module, used to perform variable analysis on the decision scenario, obtain decision variables, perform standardized evaluation on the decision variables through the standardized evaluation model, and establish a decision variable evaluation system based on the evaluation results;

[0144] A decision agent module, which is used to construct a decision agent using a neural network based on the decision variable evaluation system, and coordinate and filter the decision data through the decision agent to obtain an initial decision option;

[0145] A decision-making process module, used for narrowing the range of loss initial decision options through unsupervised learning, and obtaining the optimal decision based on the trade-off between immediate return and long-term return of the decision-maker agent within the narrowed range;

[0146] The reinforcement learning module is used to optimize the decision-making agent through reinforcement learning according to the adoption of each optimal decision.

[0147] The above contents are merely examples and explanations of the structure of the present invention. The technicians in this technical field may make various modifications or additions to the specific embodiments described or replace them in a similar manner. As long as they do not deviate from the structure of the invention or exceed the scope defined by the claims, they should all fall within the protection scope of the present invention.

Claims

1. An artificial intelligence-assisted decision-making method for standardized evaluation, characterized in that: The following steps are involved: Obtain decision scenarios and decision goals; Determine a standardized evaluation index according to the decision scenario, generate a multi-objective function according to the decision goal, and construct a standardized evaluation model by combining the multi-objective function and the standardized evaluation index; Performing variable analysis on the decision scenario to obtain decision variables, performing standardized evaluation on the decision variables using the standardized evaluation model, and establishing a decision variable evaluation system based on the evaluation results; Based on the decision variable evaluation system, a decision agent is constructed using a neural network, a decision option set is obtained based on the decision agent, and the decision option set is coordinated and filtered to obtain an initial decision option; Narrowing the scope of the initial decision options through unsupervised learning, and obtaining the optimal decision based on the trade-off between immediate rewards and long-term rewards of the decision-maker agent within the narrowed scope; The decision-making agent is optimized through reinforcement learning according to the adoption of each optimal decision.

2. The method for artificial intelligence-assisted decision-making based on standardized evaluation according to claim 1, characterized in that: The method for generating a multi-objective function according to the decision-making objective comprises: Obtain the expected level of each decision target and construct a multi-objective function by minimizing the deviation between each decision target and the expected level. The expression of the multi-objective function is: Where n is the number of decision targets, ω i is the weight coefficient of the i-th decision target, y i is the actual value of the i-th decision target, d i is the expected level of the i-th decision target, α i For z ij The degree of influence on the decision-making target deviation, m i is the number of direct cost factor sets of the ith decision objective, z ij is the jth factor in the set of direct cost factors of the i-th decision objective, β i for u ik The degree of influence on the decision-making target deviation, l i is the number of indirect cost factor sets for the i-th decision objective, u ik is the kth factor in the set of indirect cost factors for the i-th decision objective.

3. The artificial intelligence-assisted decision-making method for standardized evaluation according to claim 1, characterized in that: The method of constructing a standardized evaluation model by combining the multi-objective function and the standardized evaluation index comprises: The Pearson coefficient between standardized evaluation indicators is calculated to measure the linear correlation between standardized evaluation indicators, and a linear regression correlation function is constructed for standardized evaluation indicators with a Pearson coefficient greater than 0.4; The standardized evaluation indicators are substituted into the multi-objective function, and the linear regression correlation function between the standardized evaluation indicators is used as a constraint condition to form a standardized evaluation model.

4. The method for artificial intelligence-assisted decision-making based on standardized evaluation according to claim 1, characterized in that: The method of performing standardized evaluation on the decision variables through the standardized evaluation model and establishing a decision variable evaluation system based on the evaluation results includes: Conduct standardized evaluation on decision variables through standardized evaluation model to obtain standardized scores for each decision variable; The weight value of each decision variable is calculated by the entropy weight method according to the standardized score of the decision variable and the variability of the decision variable; Construct a standardized matrix of decision variables based on the weight value of each decision variable; Calculate the standard gradient threshold of the decision variable, and establish a decision indicator system based on the standardized matrix and the standard gradient threshold. The calculation formula of the standard gradient threshold is: Among them, α is the scaling factor, i is the index of the evaluation object, m is the number of decision variables, and λ j is the weight value of the jth decision variable, y ij is the standardized score of the i-th evaluation object on the j-th decision variable, is the standardized score of the evaluation of the ith evaluation object on all decision variables, and ψ is the number of evaluation objects.

5. The method for artificial intelligence-assisted decision-making based on standardized evaluation according to claim 1, characterized in that: The method of constructing a decision-making intelligent agent using a neural network based on the decision variable evaluation system includes: The network architecture of decision-making agents based on neural networks is as follows: The input layer is used to receive decision variables. The number of nodes in the input layer is equal to the number of decision variables. The middle layer is used to integrate the decision indicator system and perform gradient division on the decision variables of the input layer; The model layer creates a decision model based on the decision tree and makes decisions based on the gradient classification of the middle layer; The output layer is used to output the decision results. If it is a decision variable value problem, the output is the decision value of the decision variable; if it is a decision variable selection problem, the output decision value is converted from 0 to 1. When the decision value is 0, it means that the decision variable is not selected, and the decision value is 1, which means that the decision variable is selected. Collect historical decision data to train the decision agent, and use the back-propagation algorithm to calculate the gradient of the loss function to the bias of each layer of the decision agent network architecture; The stochastic gradient descent algorithm is used to update the parameters of the decision agent according to the gradient of each layer bias to obtain the final decision agent.

6. The method for artificial intelligence-assisted decision-making based on standardized evaluation according to claim 1, characterized in that: The method of obtaining a decision option set based on the decision agent, and performing coordinated filtering on the decision option set to obtain an initial decision option comprises: Obtain the internal adjustable data and external influencing data during decision-making, combine them as decision data, input the decision data into the decision agent, and obtain a decision option set that satisfies the decision model; Calculate the similarity between decision options in the decision option set. The expression is: Where D(A,B) is the similarity between decision options A and B. The smaller the value, the more similar the two decision options are. n is the number of features in the decision options. ξ i is the feature weight factor of the i-th feature, a i and b i They are respectively represented as the values ​​of decision options A and B on the i-th feature, δ is the distance adjustment parameter, and are respectively represented as vectors of decision options A and B, is the Euclidean distance between decision options A and B, is the sum of the moduli of decision options A and B, ∈ is the covariance adjustment parameter, ∑ -1 It is expressed as the inverse matrix of the covariance between decision options A and B, s i is the feature scaling factor, used to adjust the scale of the i-th feature, is the importance adjustment parameter, I i is the information gain of the i-th feature in the decision option, H i is the information entropy of the i-th feature in the decision option, η is the adjustment factor to avoid the denominator of the importance function term being zero when the information entropy is zero, and is taken as 0.01; The similarity is normalized from 0 to 1, and a partitioning criterion is set. The decision option set is clustered according to the partitioning criterion to obtain several clusters. From each cluster, the decision option combination with the largest information gain is selected as the initial decision option.

7. The method for artificial intelligence-assisted decision-making based on standardized evaluation according to claim 1, characterized in that: The method of narrowing down the range of the initial decision options by unsupervised learning comprises: Apply the Gaussian mixture model to group the initial decision options and obtain the probability of the group to which each decision option belongs. The expression is: Among them, p(k|x i ) is expressed as the probability that the i-th decision option belongs to the k-th group, π k is the mixing coefficient of the kth Gaussian distribution, Expressed as the probability density function of the kth Gaussian distribution, x i is the feature vector of the i-th decision option, μ k is the mean vector of the kth Gaussian distribution, ∑ k is the covariance matrix of the kth Gaussian distribution, υ is the number of groups; The t-distributed random neighbor embedding is used to reduce the dimension and extract the group features. The group features are represented by the central variable and the covariance matrix, and the expression is: Among them, c k is the central variable of the kth group in the low-dimensional space, N k is the number of data points in the kth group, N is the number of decision options in the initial decision options, p(k|x i ) is expressed as the probability that the i-th decision option belongs to the k-th group, y i It is represented as the representation of the i-th decision option in the low-dimensional space after dimensionality reduction. is the covariance matrix of the kth group in the low-dimensional space, and T represents the mapping transformation operation; The isolation forest algorithm is used to remove outliers that do not meet the group characteristics and narrow the range of decision options.

8. The method for artificial intelligence-assisted decision-making based on standardized evaluation according to claim 1, characterized in that: The method for weighing the immediate return and the long-term return based on the decision maker agent within the narrowed scope includes: Create a game model based on the Nash equilibrium algorithm, defining immediate returns and long-term returns as game participants; Obtaining the immediate return function and the long-term return function, and incorporating the immediate return and the long-term return function into the game model as the profit function of the game participants; Use game strategy method to solve the game model and obtain the Nash equilibrium solution. Among the Nash equilibrium solutions, select the decision plan with the largest sum of the immediate reward function and the long-term reward function as the optimal decision.

9. The method for artificial intelligence-assisted decision-making based on standardized evaluation according to claim 1, characterized in that: The method for optimizing the decision-making agent through reinforcement learning according to the adoption of each optimal decision includes: Construct a reward and punishment mechanism based on the adoption of each optimal decision. If the decision is successfully adopted, a positive reward will be given, and if the decision is not adopted, a penalty will be given; The decision agent is optimized to maximize the future cumulative reward and is updated by minimizing the KL divergence. The expression is: Among them, π new is the updated decision, ∏ is the strategy set adopted by the decision agent, and D KL Expressed as KL divergence, π(·|s t ) is the decision agent in state s t The strategy under t is the probability distribution of the decision agent at time t, i.e., the number of decisions made, is the state of the decision agent, the current strategy selects different actions, α is the reward adjustment parameter, Q(s t ,·) means in state s t The action value function of taking different actions under π (s t ) is a normalization parameter that ensures that the probability distribution of the target strategy sums to 1.

10. An artificial intelligence-assisted decision-making system for standardized evaluation, used to execute an artificial intelligence-assisted decision-making method for standardized evaluation according to claims 1-9, characterized in that: The system comprises: Decision collection module, used to obtain decision scenarios and decision goals; The standardized evaluation module is used to determine the standardized evaluation indicators according to the decision-making scenario, generate a multi-objective function according to the decision-making goal, and build a standardized evaluation model by combining the multi-objective function and the standardized evaluation indicators; The decision variable evaluation module is used to perform variable analysis on the decision scenario, obtain decision variables, perform standardized evaluation on the decision variables through a standardized evaluation model, and establish a decision variable evaluation system based on the evaluation results; The decision agent module is used to construct a decision agent using a neural network based on the decision variable evaluation system, and coordinate and filter the decision data through the decision agent to obtain the initial decision options; The decision-making process module is used to narrow the scope of initial decision options through unsupervised learning, and obtain the optimal decision based on the trade-off between immediate rewards and long-term rewards of the decision-maker agent within the narrowed scope; The reinforcement learning module is used to optimize the decision-making agent through reinforcement learning based on the adoption of the best decision each time.

Images