Multi-objective resource allocation method based on multi-modal data and multi-agent cooperation

Abstract

The application discloses a kind of multi-objective resource configuration methods based on multi-modal data and multi-agent cooperation, comprising: collecting the multi-modal data of multiple candidate targets and preprocessing;To each target, feature extraction is carried out to multi-modal data, and the characteristics are fused and analyzed using multiple professional analysis agents, to generate preliminary analysis result;Based on preliminary analysis result, multi-agent debate is executed, to generate the multi-agent decision result of target;Multi-agent decision result is weighted and fused, to obtain the single-target configuration comprehensive score of each candidate target and corresponding expected contribution value, risk value and strategic value value;According to single-target configuration comprehensive score, form configuration candidate pool, determine the correlation between candidate configuration targets;According to optimization target, construct multi-objective combination optimization model, and solve the optimal weight vector of each candidate configuration target under certain constraint condition.The application effectively improves the decision quality and overall efficiency of multi-objective resource configuration.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence, decision optimization and resource management, and specifically to a multi-objective resource allocation method based on multimodal data and multi-agent collaboration. It is applicable to resource allocation scenarios where configuration objectives are selected from multiple candidate objectives, and the configuration weights of each objective are determined under the constraints of total resource amount and objective weight. Background Technology

[0002] With the application of artificial intelligence technology in resource decision-making, intelligent resource allocation methods are increasingly being used to assist in the analysis and resource allocation of multiple candidate objectives. However, existing methods still have shortcomings in utilizing multi-source information, making decisions from multiple perspectives, and optimizing multi-objective combinations.

[0003] Existing resource analysis methods primarily rely on text reports and numerical indicators for judgment, making insufficient use of image data such as trend charts, Gantt charts, resource allocation diagrams, and report screenshots. This makes it difficult to comprehensively extract complementary information contained in text, image, and numerical data. Furthermore, the data quality and consistency of different modalities vary. If a fixed fusion method is used for each modality, data with lower reliability can easily adversely affect the analysis results.

[0004] Existing single-model or single-perspective decision-making methods typically directly generate evaluation results for candidate objectives, lacking the interaction of evidence between two opposing perspectives—supporting and opposing allocations—and also lacking the mutual correction between radical, conservative, and neutral resource strategies. Consequently, it is difficult to fully identify the allocation value, expected contribution, and allocation risk of candidate objectives.

[0005] In the process of multi-agent result fusion, existing methods usually use fixed weights or manually specified weights, which makes it difficult to determine the corresponding decision weights based on the confidence of the agent's output results, the reliability of the input data, and the historical evaluation accuracy of the agent, thus affecting the rationality of the single-objective evaluation results.

[0006] Furthermore, existing resource allocation methods often evaluate each candidate objective separately, which can easily lead to the direct superposition of multiple single objective results. This fails to fully consider the correlation between candidate objectives and the impact of the correlation on the overall portfolio risk. It is also difficult to coordinate the expected contribution, strategic value, and allocation risk of candidate objectives under constraints such as total resources, upper and lower limits of objective weights, and objective concentration, so as to obtain the allocation weight of each candidate objective. Summary of the Invention

[0007] In view of this, the purpose of this invention is to provide a multi-objective resource allocation method based on multimodal data and multi-agent collaboration. By multimodal feature fusion, multi-agent two-layer debate, single-objective comprehensive scoring and multi-objective combination optimization, the allocation weight of each candidate allocation objective is determined based on the correlation of candidate objectives and resource allocation constraints.

[0008] The objective of this invention is achieved through the following technical solution:

[0009] This invention provides a multi-objective resource allocation method based on multimodal data and multi-agent cooperation, comprising the following steps:

[0010] Text, image, and numerical data of multiple candidate targets are collected and preprocessed. Text features, visual features, and numerical features are extracted respectively. The reliability of each modality data is determined based on the data quality and cross-modal consistency of each modality data. Based on the reliability and the contribution of each modality data to the corresponding analysis task, the multimodal fusion weights of different professional analysis agents are determined. The multimodal fusion weights are used to form the comprehensive features corresponding to each professional analysis agent, and preliminary analysis results including configuration value evaluation, risk evaluation, expected contribution evaluation, strategic value evaluation, and confidence level are generated.

[0011] Next, for each candidate objective, a multi-agent debate is performed based on the preliminary analysis results: the supporting and opposing agents update their evaluation results based on the scores and evidence output by the other party, resulting in a debate result on the configuration value; the aggressive, conservative, and neutral resource allocation agents generate and update resource allocation suggestions based on the configuration value debate result, resulting in a multi-agent decision result; the decision weights are determined based on the confidence level of each agent's output, the reliability of the corresponding input data, and the historical evaluation accuracy of each agent, and the configuration value, expected contribution, risk, and strategic value are weighted and fused to obtain a single-objective configuration comprehensive score for the candidate objective; candidate objectives are screened based on the single-objective configuration comprehensive score to form a configuration candidate pool and the correlation between each candidate configuration objective in the pool is determined; with the optimization objective of maximizing the sum of configuration benefits of each candidate configuration objective and minimizing the overall configuration risk of the multi-objective combination, a multi-objective combination optimization model is constructed, and the configuration weights of each candidate configuration objective are solved under the constraints of total resources, upper and lower limits of objective weights, and objective concentration.

[0012] The beneficial effects of this invention are:

[0013] First, this invention comprehensively utilizes text data, image data, and numerical data of candidate targets, and determines the fusion weight based on the reliability of each modal data and its contribution to the corresponding analysis task, which can improve the rationality of multimodal information utilization and reduce the impact of low-reliability data on the analysis results.

[0014] Second, by supporting the debate on the configuration value between configuration agents and opposing configuration agents, as well as the debate on resource strategies between radical, conservative, and neutral resource allocation agents, this invention enables agents with different positions to update their own results based on the scores, evidence, and resource allocation suggestions output by other agents, thereby improving the objectivity of candidate target evaluation.

[0015] Third, this invention determines decision weights based on the confidence level of each agent's output, the reliability of the source data corresponding to the output, and the historical evaluation accuracy of the agent. It also performs weighted fusion of configuration value, expected contribution, risk, and strategic value. Compared with fixed weights or manually assigned weights, this invention can improve the rationality of the comprehensive score of a single objective.

[0016] Fourth, this invention screens candidate targets based on the comprehensive score of single-objective configuration and incorporates the correlation between candidate configuration targets into the overall risk of multi-objective combination. Under the constraints of total resources, upper and lower limits of target weights and target concentration, it optimizes the combination of configuration benefits and overall configuration risks, avoiding the simple superposition of multiple single-objective results and realizing the unified solution of configuration weights for each candidate configuration target. Attached Figure Description

[0017] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will now be described in further detail with reference to the accompanying drawings, wherein:

[0018] Figure 1 This is a diagram of the overall processing architecture of the method of the present invention;

[0019] Figure 2 This is a flowchart of a single-objective analysis.

[0020] Figure 3 It is an interactive diagram of a two-layer debate mechanism;

[0021] Figure 4 This is a flowchart of candidate target selection and multi-target combination optimization. Detailed Implementation

[0022] The preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings. It should be understood that the preferred embodiments are for illustrative purposes only and are not intended to limit the scope of protection of the present invention.

[0023] It should be noted that the "resources" in this invention include, but are not limited to, quantifiable or convertible into a unified resource unit, such as budget funds, computing power, equipment capacity, man-hours, material inventory, supply chain quotas, and project scheduling resources; the "resource allocation targets" include, but are not limited to, projects, tasks, departments, production lines, supplier solutions, cloud computing operations, or business units.

[0024] To enable those skilled in the art to clearly implement this invention, the key terms, data fields, and quantification methods involved in this specific embodiment are uniformly defined below. Unless otherwise specified, the following terms in this invention are understood according to the definitions in Table 1.

[0025]

[0026] Figure 1 This is the overall architecture diagram of a multi-objective resource allocation system. (For example...) Figure 1 As shown, the overall architecture of the multi-objective resource allocation system provided by the present invention can be divided into: an input layer, a parallel single-objective analysis layer, a single-objective decision layer, a multi-objective combined optimization layer, and an output layer.

[0027] The input layer receives multimodal data from N candidate targets. The system executes multimodal data acquisition and specialized analysis by an intelligent agent (i.e., ...) for each candidate target. Figure 1 The analysis engine (such as analysis engine 1 and analysis engine 2) integrates analysis and multi-agent debate to obtain the multi-agent decision results for each candidate objective.

[0028] Each candidate objective corresponds to a single-objective decision result (i.e. Figure 1 In the process of decision 1, decision 2, etc., each decision is determined by the confidence of the output results of each agent, the reliability of the source data corresponding to the output results, and the historical evaluation accuracy of each agent. The outputs of each agent are weighted and fused to obtain the configuration value fusion value, expected contribution fusion value, risk fusion value and strategic value fusion value of each candidate target, and to determine the comprehensive configuration score of a single target.

[0029] The multi-objective combination optimization layer filters candidate objectives based on the comprehensive score of single-objective configuration, forms a configuration candidate pool, and determines the correlation between each candidate configuration objective in the configuration candidate pool. Based on the configuration benefits and overall configuration risks of each candidate configuration objective, a multi-objective combination optimization model is constructed, and the configuration weights of each candidate configuration objective are solved under the constraints of total resources, upper and lower limits of objective weights, and objective concentration. The output layer outputs the results.

[0030] Combination Figures 1 to 4 The multi-objective resource allocation method based on multimodal data and multi-agent cooperation in this invention includes the following specific steps:

[0031] S1: Multimodal data acquisition and preprocessing

[0032] Obtain N candidate targets (i.e., "candidate resource allocation targets", also known as "resource allocation targets" or simply "targets"), where N≥2.

[0033] Multimodal data for these N candidate objectives are collected and preprocessed. The multimodal data includes text data, image data, and numerical data. Text data includes project / task reports, performance or financial reports, environmental feedback, logs, meeting minutes, and external comments; image data includes trend charts, Gantt charts, resource allocation diagrams, flowcharts, and report screenshots; numerical data includes historical performance indicators, base period costs, current costs, output data, risk indicators, and resource consumption indicators.

[0034] Preprocessing includes data cleaning, format standardization, missing value handling, normalization, and time alignment. Numerical data undergoes time alignment, outlier handling, and standardization according to the execution cycle. The processed data forms a preprocessed multimodal dataset corresponding to each candidate target. , , Indicates the first A unified feature dataset for each resource allocation objective. , and They represent the first The target consists of preprocessed text data, image data, and numerical data.

[0035] S2: Multimodal Feature Extraction and Professional Analysis: Agent Fusion Analysis

[0036] For each of the N candidate targets, multiple specialized analytical agents are created. The multimodal feature extraction module in each analysis engine extracts features from the multimodal data corresponding to each target. Then, the extracted features are fused and analyzed by the multiple specialized analytical agents in each analysis engine to generate the preliminary analysis results of each specialized analytical agent for that target.

[0037] The multiple specialized analytical agents include a fundamental analysis agent, a performance analysis agent, and an environmental analysis agent. Each objective The corresponding specialized analytical agents include fundamental analysis agents. Performance analysis intelligent agent and environmental analysis intelligent agents .

[0038] Each specialized analytical agent employs a structure of "multimodal encoder + task prompt / rule base + inference decision module + scoring head + memory module". The multimodal encoder is responsible for converting image, text, and numerical data into vectors; the task prompt / rule base defines the specialized perspective that the analytical agent focuses on; the inference decision module can be implemented using a large language model, a rule-based inference engine, or a deep neural network; the scoring head outputs standardized scores; and the memory module stores the analytical agent's historical prediction results, historical evidence, and historical errors. While the three specialized analytical agents share the same underlying structure, they differ in the task prompts, feature selection weights, scoring head parameters, and evaluation priorities they use.

[0039] Fundamental analysis agents focus on assessing the necessity of target configuration, strategic alignment, long-term capacity building, inter-departmental synergy value, and policy window. Performance analysis agents focus on assessing target achievement rate, input-output efficiency, progress trends, technological maturity, and efficiency improvement. Environmental analysis agents focus on assessing the volatility of the external environment in which the target operates, dependency conflicts, risk events, and execution uncertainty.

[0040] The analytical results of the three specialized analytical agents may differ. For example, the fundamental analysis agent might give a high strategic value score, while the environmental analysis agent might give a high risk coefficient. This disagreement is not treated as an error, but rather as input for a subsequent two-tiered debate—the agents supporting and opposing the allocation will each use these points of disagreement to engage in adversarial arguments.

[0041] For each of the N candidate targets, S2 specifically includes the following steps S21 to S25.

[0042] In S21: Multimodal feature extraction is performed on the image data, text data, and numerical data of each candidate target to obtain the visual features, text features, and numerical features corresponding to the candidate target.

[0043] The multimodal feature extraction module includes an image recognition module, a text understanding module, and a numerical analysis module. The image recognition module is used to extract visual features such as trend patterns, chart structures, and visual annotations from the image data of resource allocation targets.

[0044] The image recognition module is used to extract visual features such as trend patterns, progress status, structural relationships and visual annotations from trend charts, Gantt charts, resource allocation diagrams, flowcharts and report screenshots, and encodes the extraction results into visual feature vectors.

[0045] The text understanding module extracts semantic information, key performance indicators (KPIs), and sentiment / risk profiles from textual data related to resource allocation objectives. This textual data includes at least one or more of the following: project / task reports, performance or financial reports, environmental feedback, logs, meeting minutes, and external comments. The text understanding module may include sub-modules for text cleaning, named entity recognition, key indicator extraction, semantic embedding, sentiment / risk classification, and evidence sentence extraction. After processing by these sub-modules, the text understanding module outputs a text feature vector, which includes at least the following features: number of positive events, number of negative events, risk word intensity, strategic matching keyword score, execution obstacle keyword score, and evidence sentence credibility. Based on the number of negative events and risk word intensity, these can be weighted and fused to obtain the negative event intensity. This text feature vector is then input into a multimodal fusion module for cross-modal fusion with image feature vectors and numerical feature vectors.

[0046] The numerical analysis module is used to quantitatively analyze the historical performance series, cost series, output data, resource consumption series, and risk series of each resource allocation target, extracting the target's performance ratio and trend change indicators. The potential for improvement in achieving a stage target is determined by the ratio of the difference between the stage target value and the current achieved value to the stage target value, reflecting the available space for further improvement in the current stage. When this indicator is positive, it can be expressed as the ratio of the stage target value minus the current achieved value to the stage target value, and is limited to a preset range.

[0047] Performance ratios are used to measure the input-output efficiency, achievement status, and resource utilization efficiency of resource allocation objectives within a specific execution cycle. Performance ratios include input-output ratio, objective achievement rate, cost deviation rate, cost reduction rate, output per unit of resource, resource load rate, quality pass rate, and task on-time rate. Specifically, the cost reduction rate is determined by the ratio of the difference between the cost of the previous execution cycle and the cost of the current execution cycle to the cost of the previous execution cycle; output per unit of resource is determined by the ratio of output data to resource consumption data within the same execution cycle.

[0048] Trend change indicators are used to describe the direction and magnitude of changes in key indicators over the execution cycle, including recent trends. The linear regression slope, month-on-month growth rate, moving average difference, and volatility for each execution period.

[0049] The degree of external environmental volatility is determined by a weighted average of the risk tendencies extracted from environmental feedback and external comments by the text understanding module and the risk volatility calculated by the numerical analysis module based on the relevant risk sequences of the external environment.

[0050] In addition, equipment utilization is determined based on the ratio of actual equipment usage time to equipment availability time extracted from the resource consumption sequence by the numerical analysis module; the degree of supply chain constraints is determined based on the characteristics of execution obstacles such as supply chain disruptions, material shortages, and delivery delays extracted from project / task reports, logs, and environmental feedback by the text understanding module; and the degree of critical path blocking is determined based on the critical task delay status and task dependency blocking status extracted from Gantt charts and flowcharts by the image recognition module.

[0051] The delay cost is determined based on the task delay duration extracted from the Gantt chart by the image recognition module, and the unit time cost obtained from the cost sequence by the numerical analysis module.

[0052] Therefore, the single-objective reasoning processes of each objective are independent of each other, but they can share the results of combined-level correlation calculations.

[0053] Correlation is used to measure the statistical association between different resource allocation objectives, such as the rate of change in performance, the rate of change in resource consumption, or a series of risk events, rather than a simple correlation between indicators within a single objective. Specifically, let Indicates the first The resource allocation objective is in the first The contribution change rate or performance return rate for the execution cycle, then the first The resource allocation objective and the first Correlation coefficient of resource allocation objectives Represented as: ,

[0054] in, and These represent the standardized performance sequences of the two objectives over the historical execution period. and These represent the standard deviations of the two performance series, respectively. When the correlation coefficient between the two objectives is greater than the set value (e.g., 0.7), the two resource allocation objectives are considered to be highly correlated, indicating that they may be jointly affected by the same external factors, underlying technologies, supply chain conditions, resource bottlenecks, or changes in market demand. When combining resource allocation, it is necessary to control the combined weight of the two to avoid the simultaneous failure of both objectives causing a significant impact on the overall allocation.

[0055] For resource allocation objectives lacking sufficient historical execution data, similar historical objective data from the same category, industry, or department can be used as priors, or the correlation between the objective and other objectives can be initialized based on simulation results and expert evaluation results. After accumulating new actual execution data, the correlation is updated on a rolling basis according to the updated performance change rate, resource consumption change rate, or risk event sequence. Through the calculation of single-objective indicators such as the above-mentioned objective achievement rate, cost deviation rate, resource load rate, trend change index, and volatility, the numerical feature vectors corresponding to each objective can be obtained. These numerical feature vectors are then input into the multimodal fusion module for cross-modal fusion with image feature vectors and text feature vectors. The correlation results between objectives are saved separately as shared combined-level features for extraction in step S5 and used to construct the risk correlation matrix.

[0056] For ease of description, for the first For each resource allocation target, the image data, text data, and numerical data, after being encoded as described above, yield image feature vectors, text feature vectors, and numerical feature vectors, respectively denoted as:

[0057] , ,

[0058] in, , and They represent the first Image feature vectors, text feature vectors, and numerical feature vectors of each resource allocation target. , and These represent image encoding functions, text encoding functions, and numerical encoding functions, respectively.

[0059] In S22: Determine the reliability of each modal data based on the data quality of each modal data and the cross-modal consistency between each modality.

[0060] In S22, for the first Modalities of candidate targets First, the data quality score for this modality is determined based on factors such as image clarity, OCR confidence, text source credibility, missing value rate, and outlier ratio. The cross-modal consistency score is determined based on the degree of information consistency between this mode and other modes. The data reliability of this modality is determined based on the data quality score and cross-modal consistency score. in, and The weighting parameters are shared by the various professional analytical agents to ensure the reliability of data for each modality.

[0061] In S23: For the first The first of the resource allocation objectives A specialized analytical agent, based on the reliability of each modal data and the impact of each modal data on the first... The contribution of each specialized analytical agent to the corresponding analytical task is determined. Each professional analysis agent corresponds to the modal fusion weights.

[0062] For the The resource allocation objective, its first A professional analysis agent corresponds to the modality Fusion weights It can be represented as:

[0063] in, Representing modes For the The contribution of each specialized analytical agent to the corresponding analytical task; c is the weight adjustment parameter.

[0064] In S24: For each specialized analytical agent, the visual features, text features, and numerical features of the target are weighted and fused using the fusion weights of each modality to obtain the comprehensive feature vector of the target corresponding to that specialized analytical agent. The first of the resource allocation objectives The comprehensive feature vector (i.e., "comprehensive feature") corresponding to a specialized analytical agent can be represented as:

[0065]

[0066] in, , and They represent the first Each professional analysis agent corresponds to the image, text, and numerical feature transformation matrix. Indicates the first Each professional analyzes the bias terms corresponding to the intelligent agent. Represents a non-linear activation function. Indicates the first The first of the resource allocation objectives A comprehensive feature vector corresponding to a professional analytical agent.

[0067] S25: Each specialized analytical agent generates a preliminary analysis result for the target based on its corresponding comprehensive feature vector. In some embodiments, each specialized analytical agent can also generate a preliminary analysis result for the target based on the target's comprehensive feature vector and a specialized task template (i.e., each specialized analytical agent has a specific task template). The preliminary analysis result includes at least a configuration value score (i.e., "configuration value value"), a risk score (i.e., "risk value"), an expected contribution score (i.e., "expected contribution value"), a strategic value score (i.e., "strategic value value"), a growth potential score, an efficiency score, a cost overrun risk score, a comprehensive risk score, a resource bottleneck degree score, and a confidence level.

[0068] The growth potential score is determined based on the slope of the target's performance trend (obtained through linear regression of historical performance sequences), the year-on-year growth rate of historical performance, and the potential for improvement in achieving the stage target. The efficiency score is determined based on the input-output ratio, cost reduction rate, output per unit of resources, and on-time delivery rate (i.e., "task on-time rate"). The strategic value score is determined based on strategic alignment, long-term capability building, departmental synergy value, and policy window period. Long-term capability building, departmental synergy value, and policy window period are primarily evaluated by the fundamental analysis agent based on relevant textual features extracted by the text understanding module, combined with other modal features. The cost overrun risk score is determined based on the cost deviation rate and delay costs. The comprehensive risk score is determined based on the intensity of negative events and the degree of external environmental volatility. The resource bottleneck score is determined based on resource load factor, equipment utilization rate, supply chain constraints, and critical path congestion. All of the above scores are standardized to a preset scoring range.

[0069] After obtaining the above scores, the confidence level of the preliminary analysis results output by the professional analytical agent is further determined. This confidence level is jointly determined by the completeness of evidence, modal consistency, data reliability, and the historical evaluation accuracy of the professional analytical agent. The first of the resource allocation objectives The confidence level of the output of a professional analytical agent It can be represented as:

[0070]

[0071] in, Indicates the first The data reliability score relied upon by a professional analytical agent, for the first The first candidate target A specialized analytical agent, whose data reliability score is based on. (i.e., in the previous formula) The reliability of the modal data corresponding to the agent is weighted and determined, i.e.:

[0072]

[0073] in, This indicates the first step determined in step S22. The first candidate target Reliability of modal data This represents the modal fusion weights determined in step S23.

[0074] Indicates the first A professional analytical agent's cross-modal consistency score of the data it relies on, the first The first candidate target A professional analytical agent relies on cross-modal consistency scoring of the data. (Right now According to the modality fusion weights corresponding to the agent, the cross-modal consistency scores of each modality obtained in step S22 are weighted and determined, that is:

[0075]

[0076] Indicates the first The evidence coverage score of the data relied upon by each specialized analytical agent is obtained by weighted summing of the temporal coverage score and the modal coverage score of the evidence. The temporal coverage score is the ratio of the length of the historical period in which the specialized analytical agent actually has valid evidence to the total execution period length; the modal coverage score is the ratio of the number of modalities in which the specialized analytical agent actually has valid evidence to the total number of modalities (i.e., 3).

[0077] This represents the historical evaluation accuracy score of the professional analytical agent. The historical evaluation accuracy score is determined based on the historical evaluation results of the professional analytical agent in completed resource allocation tasks. After obtaining the actual execution results of the resource allocation objectives, the previously output allocation value evaluation, risk evaluation, expected contribution evaluation, and strategic value evaluation of the professional analytical agent are compared with the corresponding actual evaluation results, and the historical evaluation accuracy score is calculated based on the normalized error between the two. The historical evaluation accuracy score can be expressed as:

[0078]

[0079] in, This indicates the number of historical resource allocation tasks that have been completed. Indicates the first The professional analytical agent in the first Normalized evaluation error in each historical task. When no historical evaluation record exists, Set to a preset initial value and update it after obtaining the actual execution result.

[0080] It should be noted that the coefficients of each term in this formula are merely examples and are not intended to limit the invention.

[0081] S3: Multi-Agent Two-Level Debate

[0082] refer to Figure 3 For each of the N candidate targets, a two-layer debate is performed based on the multiple preliminary analysis results. The first layer is the configuration value debate, where supporting and opposing configuration agents engage in adversarial debate. In each round of adversarial debate, the supporting configuration agent determines the supporting configuration score of the candidate target based on the preliminary analysis results output by multiple professional analysis agents, generates corresponding supporting evidence, and outputs the confidence level of the supporting configuration result. The opposing configuration agent determines the opposing configuration score of the candidate target based on the preliminary analysis results output by multiple professional analysis agents, generates corresponding opposing evidence, and outputs the confidence level of the opposing configuration result. After multiple rounds of debate, the configuration value debate result is obtained, including the final supporting configuration score, supporting evidence, confidence level of the supporting configuration result, opposing configuration score, opposing evidence, and confidence level of the opposing configuration result. The second layer is the resource strategy debate, where aggressive resource allocation agents, conservative resource allocation agents, and neutral resource allocation agents generate suggested resource ratios, expected contribution values, risk values, and corresponding confidence levels of resource allocation suggestions for candidate targets based on the configuration value debate result. They update their resource allocation suggestions with each other, and after multiple rounds of debate, the final resource allocation suggestions and confidence levels for each resource allocation agent are obtained.

[0083] Specifically, in each round of the first-level debate, the support configuration agent obtains growth potential scores, efficiency scores, and strategic value scores from the preliminary analysis results output by multiple professional analysis agents for the target. Based on the confidence level of each preliminary analysis result output by the final professional analysis agents, the above scores are weighted and fused to obtain the growth potential fusion value, efficiency fusion value, and strategic value fusion value of the target, thereby determining the support configuration score for the target in that round.

[0084] For the Each resource allocation objective is based on the confidence level of the output results of the analytical agents from various professional backgrounds. The various scores output are weighted and fused together, that is:

[0085] , , ,

[0086] in, Indicates the first K professional analytical agents for the first (out of K) The confidence level of the output results of each resource allocation target; , and These represent the growth potential score, efficiency score, and strategic value score output by the analytical agent in this field, respectively. , and These represent the weighted fusion values ​​for growth potential, efficiency, and strategic value, respectively.

[0087] The support configuration score can be represented as:

[0088] ,

[0089] in, Indicates the first Support configuration score for each resource allocation target; , and These are weighting parameters, and their sum is 1. They can be determined based on business priorities or expert presets.

[0090] In each round of the first-level debate, the opposing configuration agent obtains the cost overrun risk score, comprehensive risk score, and resource bottleneck score generated by each professional analysis agent from the preliminary analysis results output by multiple professional analysis agents for the target. Based on the confidence level of the final output results of each professional analysis agent, the above scores are weighted and fused to obtain the cost overrun risk fusion value, comprehensive risk fusion value, and resource bottleneck score fusion value of the target, and thus determine the opposing configuration score of the target in that round.

[0091] The score for opposing the configuration can be expressed as:

[0092] ,

[0093] in:

[0094] , ,

[0095] Indicates the first The opposition score for each resource allocation objective. , and They represent the first Each professional analytical agent outputs a cost overrun risk score, a comprehensive risk score, and a resource bottleneck level score. This represents the fusion value of the cost overrun risk for this objective. This represents the overall risk fusion value of the target. This represents the fusion value indicating the degree of resource bottleneck for the target. , and These are weight parameters, and their sum is 1.

[0096] The system supports multiple rounds of adversarial debate between supporting and opposing agents. Each round includes four stages: viewpoint generation, evidence citation, rebuttal of the opposing viewpoint, and confidence level update. If the changes in both the supporting and opposing configuration scores are less than a threshold for two consecutive rounds... or reach the maximum number of debate rounds. Then the debate will cease.

[0097] That is, the stopping condition can be expressed as: and ,or ,

[0098] in, and They represent the first The scores for supporting and opposing configurations after the debate on the value of configuration.

[0099] The first-level debate outputs the final supporting configuration score, supporting evidence, opposing configuration score, opposing evidence, and corresponding confidence level, yielding the configuration value debate result, which serves as the input for the second-level resource strategy debate and step S4 for single-objective decision fusion.

[0100] In the second-level debate, in each round, the radical resource allocation agent, the conservative resource allocation agent, and the neutral resource allocation agent (i.e., the three-party resource allocation agents, also known as the three-party agents) output the suggested resource ratio, expected contribution value, risk value, and corresponding confidence level of the resource allocation suggestion for the target. Through multiple rounds of debate, when the maximum difference of the three suggested ratios in a certain round is less than a set threshold, or when the change of the suggested resource ratio output by each resource allocation agent in two consecutive rounds is less than a preset threshold, the strategy debate is considered to have converged, the second-level debate is stopped, and the final resource strategy suggestion for the target and its confidence level are obtained.

[0101] Among them, the aggressive resource allocation agent proposes a higher allocation ratio or allocation upper limit when there is high confidence and high growth potential; the conservative resource allocation agent proposes a low allocation ratio when there is high risk, high volatility or conflict of evidence; and the neutral resource allocation agent proposes a compromise resource range based on historical resource efficiency and risk level.

[0102] When the three intelligent agents engage in multi-round game debate, each of them outputs a suggested resource ratio, expected contribution value, risk value, and confidence level.

[0103] Suppose the suggested ratios given by the three resource allocation agents are as follows: , and These correspond to the radical resource allocation agent, the conservative resource allocation agent, and the neutral resource allocation agent, respectively.

[0104] In each round of debate, each resource allocation agent updates its own resource allocation recommendations based on the suggested resource ratios, expected contribution values, and risk values ​​output by other resource allocation agents.

[0105] After the second layer of debate, the resource allocation suggestions output by the radical resource allocation agent, the conservative resource allocation agent, and the neutral resource allocation agent are obtained respectively, and are used as input for step S4.

[0106] In steps S2 and S3, the outputs of each specialized analytical agent, the supporting configuration agent, the opposing configuration agent, and the radical, conservative, and neutral resource allocation agents are all based on their respective fixed analytical perspectives and fixed debate rules. The output formats of different agents differ, requiring unified mapping and weighted fusion.

[0107] The preliminary analysis results output by the professional analysis agent, the debate results on the configuration value output by the supporting and opposing configuration agents, and the resource allocation suggestions output by the radical, conservative, and neutral resource allocation agents all include the confidence level of the corresponding output results. The confidence level of the professional analysis agent's output is determined according to the method described in step S25. The confidence levels of the outputs of the supporting and opposing configuration agents, as well as each resource allocation agent, are jointly determined based on the confidence levels of their upstream agents (the upstream agents of the supporting and opposing configuration agents are the respective professional analysis agents, and the upstream agents of each resource allocation agent are the supporting and opposing configuration agents) and the stability of the agent's output results over two consecutive rounds. When only one round of debate is completed, the weighted average of the confidence levels of the upstream agents' outputs is used as the initial confidence level.

[0108] Additionally, it should be noted that the supporting configuration agent, opposing configuration agent, and the three-party resource allocation agents (radical, conservative, and neutral) used in step S3 all adopt a similar structure to the aforementioned professional analysis agent, except that the task prompts, feature selection weights, scoring head parameters, and evaluation focus are different, which will not be elaborated here.

[0109] S4: Multi-agent decision fusion and single-objective scoring.

[0110] Specifically, in S4, for each candidate target, the decision weight of each agent is determined based on the confidence of the output results of each agent, the reliability of the source data corresponding to the output results, and the historical evaluation accuracy of each agent. The configuration value, expected contribution, risk, and strategic value of each agent's output are weighted and fused to determine the single-target configuration comprehensive score of the candidate target.

[0111] For each resource configuration target, S4 specifically includes steps S41 to S43.

[0112] S41: Determine the decision weight of each agent based on the confidence level of the output results of each agent, the reliability of the corresponding source data, and the historical evaluation accuracy of each agent.

[0113] Before calculating the decision weights, the output confidence, corresponding source data reliability, and historical evaluation accuracy of each agent are normalized to the interval 0-1. For the first... The candidate target, the first The decision weights of each agent are represented as follows:

[0114]

[0115] in, Indicates the first The first agent on the first The confidence level of the output of each candidate target. This indicates the reliability of the source data corresponding to the agent's output. This indicates the accuracy of the agent's historical evaluations. , , To adjust the parameters, This represents the set of intelligent agents participating in the decision-making fusion.

[0116] Specifically, for a specialized analytical agent, the reliability of its corresponding source data is the data reliability score on which the specialized analytical agent relies, as determined in step S25; for agents that support configuration, oppose configuration, and are radical, conservative, or neutral in resource allocation, the reliability of their corresponding source data is determined by weighting the source data reliability of the directly upstream agent according to the output confidence level of the upstream agent.

[0117] ,

[0118] in, Indicates the first The set of direct upstream agents of each agent. and These represent the upstream agents' views on... No. The confidence level of the output results of each candidate target and the reliability of the corresponding source data. When the upstream agent is a professional analytical agent... That is , That is .

[0119] Since different agents output results in different forms, it is necessary to first map their output results to a unified decision vector. In S42: Obtain the decision vector for each agent's output to the target. The decision vector includes at least the agent's configuration value, risk value, expected contribution value, and strategic value for the target. The original outputs of different types of agents are converted into decision vectors with consistent direction and uniform dimensions through a preset index mapping function. For vector items that do not obtain effective output, weight normalization and weighted fusion are performed only based on the agents that have effective output for that vector item.

[0120] As an example, the first The agent for the first The output vector (i.e., the "decision vector") of each objective can be represented as: ,in, Indicates the first The agent for the first The configuration value of the output of each objective. Indicates the first The agent for the first The risk value of the output of each target Indicates the first The agent for the first The expected contribution of the output of each objective. Indicates the first The agent for the first The strategic value of the output of each objective. Indicates the first The agent for the first The recommended resource ratio for the output of each objective Indicates the first The agent for the first The confidence level of the output of each target.

[0121] A corresponding index mapping function is set for the raw outputs of different types of agents to map the heterogeneous outputs of each agent to a decision index space with consistent direction and unified dimensions. Specifically, the support configuration score of the agent supporting configuration is directly mapped to the positive configuration value score; the opposition configuration score of the agent opposing configuration is mapped to the configuration value score after a reverse transformation. For example, when all scores are standardized to the range of 0-100, the configuration value score corresponding to the agent opposing configuration can be expressed as:

[0122] in, Indicates the first The score for opposing configurations of each objective.

[0123] For each item in the decision vector, the corresponding item's output for the objective by each agent is calculated according to each agent's decision weight. Perform weighted fusion to obtain the weighted fusion value of this item.

[0124] For example, the merged configuration value. Risk fusion value Expected contribution fusion value Integration with strategic value They can be represented as:

[0125] , , , .

[0126] In S43: The configuration value fusion value, expected contribution fusion value and strategic value fusion value are positively weighted and reversely corrected according to the risk fusion value, thereby determining the single-objective configuration comprehensive score of the target.

[0127] For example, ,in, Indicates the first The formula represents the comprehensive score for a single-objective resource allocation goal. The weights (0.40, 0.25, 0.20, 0.15) are typical values, reflecting the relative importance of the allocation value score, expected contribution value, strategic value coefficient, and risk score within the overall single-objective allocation score. These weights are not fixed but can be adjusted based on application scenarios, business preferences, or historical data calibration.

[0128] In step S4, the system independently generates a single-objective decision report for each candidate objective. However, the above decisions are based solely on the individual objective itself, without considering the correlation and dependencies between objectives, as well as the resource competition and risk diversification needs at the portfolio level. For example, two objectives with high individual scores may belong to the same industry or share the same supply chain. If both are assigned high weights, the portfolio as a whole will face concentrated risk should the industry or supply chain experience fluctuations. Therefore, subsequent steps require summarizing the single-objective decision reports of all N objectives, elevating the single-objective decision results from an "individual perspective" to a "portfolio perspective." Thus, portfolio-level decision fusion is necessary.

[0129] S5: Candidate Target Screening and Relevance Determination

[0130] refer to Figure 4 In step S5, the N candidate targets are filtered based on their single-target configuration comprehensive scores to form a configuration candidate pool. For example, the top five candidate configuration targets can be selected from high to low based on their single-target configuration comprehensive scores and added to the configuration candidate pool.

[0131] After forming the configuration candidate pool, the correlation coefficients between each candidate configuration target are determined based on the performance change rate, resource consumption change rate, or risk event sequence of each candidate configuration target in the historical execution cycle, forming a correlation coefficient matrix corresponding to the configuration candidate pool.

[0132] If the candidate pool includes For each candidate configuration target, a risk fusion value is constructed based on the risk fusion value of each candidate configuration target and the correlation coefficient between the targets. × Risk correlation matrix Among them, the first in the risk correlation matrix Line number The elements of a column are represented as follows:

[0133]

[0134] in, Indicates the first The candidate configuration target and the first The correlation coefficient between the candidate configuration targets and These represent the risk scales obtained by normalizing the risk fusion values ​​of the two candidate configuration targets.

[0135] Step S5 obtains the correlation between the configuration candidate pool and the candidate configuration objectives, and uses it as the input for step S6 to construct the multi-objective combined optimization model.

[0136] S6: Multi-objective combinatorial optimization and allocation weight solution

[0137] Let the number of candidate targets be The weight vector can then be represented as: ,in, Indicates the first element in the final configuration target set. The weight of each configuration target, .

[0138] The expected contribution vector (obtained from the fused expected contribution values ​​of each objective) can be represented as: .

[0139] The strategic value vector (obtained by the fusion of the strategic values ​​of each objective) can be represented as: .

[0140] The risk correlation matrix (as shown above, obtained from the risk fusion values ​​of each objective and the correlation between objectives) can be represented as: .

[0141] The expected contribution of the portfolio can be expressed as: .

[0142] The strategic value of a combination can be expressed as: .

[0143] The portfolio risk variance can be expressed as: .

[0144] The portfolio risk level (standard deviation) can be expressed as: .

[0145] In this invention, the configuration benefit of each candidate configuration target can be determined based on the expected contribution fusion value and strategic value fusion value of each candidate configuration target; then, the overall configuration risk of the multi-objective combination can be determined based on the risk fusion value of each candidate configuration target and the correlation between the candidate configuration targets; and a multi-objective combination optimization model is constructed with the optimization objective of maximizing the sum of the configuration benefits of each candidate configuration target and minimizing the overall configuration risk.

[0146] Under the constraints that the sum of the configuration weights satisfies the preset total resource amount, the configuration weights of each candidate configuration target are within the corresponding upper and lower weight limits, and the target concentration does not exceed the preset threshold, the multi-objective combination optimization model is solved to obtain the configuration weights of each candidate configuration target.

[0147] As an example, its objective function can be written as: ,

[0148] in, , and These represent the weighting coefficients for expected contribution, portfolio risk, and strategic value, respectively.

[0149] During the weight optimization process, constraints are set on the total amount of resources, the upper and lower limits of the weight of a single target, and the concentration of the target.

[0150] The total resource constraint can be expressed as: ,in, This indicates the number of candidate configuration targets in the configuration candidate pool. Indicates the first The configuration weight of each candidate configuration target.

[0151] The upper and lower bound constraints of a single target weight can be expressed as: , .

[0152] in, and They represent the first The lower and upper weight limits of each candidate configuration target.

[0153] The target concentration constraint can be expressed as:

[0154] in, This indicates the degree of weight concentration among the candidate configuration targets. This indicates the preset upper limit of target concentration.

[0155] Under the above constraints, the configuration weights of each candidate configuration objective can be obtained by solving the multi-objective combined optimization model.

[0156] The process of this invention will be illustrated below through specific examples.

[0157] A diversified enterprise group has an annual project budget of 10 million yuan and wants to build a resource allocation portfolio that includes multiple strategic projects. The goal is to increase the portfolio's expected contribution and strategic value while reducing its overall risk.

[0158] In this embodiment, the optimization objective is to improve the expected contribution and strategic value of the portfolio while reducing the overall risk of the portfolio. The expected contribution vector is constructed by fusing the expected contribution values ​​of each candidate configuration objective. The strategic value integration values ​​will form a strategic value vector. And construct a risk correlation matrix based on the correlation between candidate configuration targets. Let the configuration weight vector be... Therefore, in this embodiment, the multi-objective combined optimization model described in step S6 is used to solve for the configuration weights of each candidate configuration objective.

[0159] The objective function in this embodiment is expressed as follows:

[0160]

[0161] in, , and These are adjustment coefficients for expected contribution, portfolio risk, and strategic value, respectively.

[0162] Candidate target preparation:

[0163] Select 9 projects from different departments or categories as candidate configuration targets:

[0164] Baijiu Brand Upgrade and Channel Expansion Project

[0165] Internet game business recovery and user growth project

[0166] New energy vehicle production capacity expansion project

[0167] Battery supply chain digital optimization project

[0168] Bank Digital Transformation and Risk Control System Upgrade Project

[0169] Insurance Product Innovation and Customer Service Optimization Project

[0170] Photovoltaic module R&D and capacity expansion project

[0171] Pharmaceutical CRO Service Expansion and Clinical Trial Acceleration Project

[0172] Consumer electronics supply chain integration and Apple manufacturing optimization project

[0173] Steps S1-S4: Multimodal analysis, multi-agent debate, and single-objective scoring

[0174] For the above nine candidate configuration targets, step S1 collects text data such as project reports and performance reports, image data such as trend charts and Gantt charts, and numerical data such as historical performance, cost, risk and resource consumption, and completes data cleaning, format unification, missing value handling, normalization and time alignment.

[0175] Step S2 extracts textual features, visual features, and numerical features respectively. These features are then fused and analyzed by the fundamental analysis agent, performance analysis agent, and environmental analysis agent based on the reliability of each modality's data and its contribution to the corresponding analysis task, yielding preliminary analysis results for each candidate configuration target. These preliminary analysis results include growth potential scores, efficiency scores, strategic value scores, cost overrun risk scores, comprehensive risk scores, resource bottleneck severity scores, and corresponding confidence levels.

[0176] Step S3 involves performing configuration value debate and resource strategy debate based on the multiple preliminary analysis results. Specifically, the supporting and opposing configuration agents weight and fuse their respective scores according to the confidence level of each preliminary analysis result to obtain the supporting configuration score and the opposing configuration score. Each resource allocation agent further outputs the suggested resource ratio, expected contribution value, risk value, and confidence level to obtain the multi-agent decision result.

[0177] Step S4 determines the decision weights based on the confidence level of each agent's output, the reliability of the source data corresponding to the output, and the historical evaluation accuracy of each agent. It then weights and fuses the configuration value, expected contribution, risk, and strategic value to obtain the configuration value fusion value, expected contribution fusion value, risk fusion value, strategic value fusion value, and single-objective configuration comprehensive score for each candidate configuration target. Specifically, the source data reliability of the professional analysis agent is determined by the reliability of the multimodal data it relies on; the source data reliability of the downstream agent is determined by weighting the source data reliability and output confidence of the directly upstream agent; and the historical evaluation accuracy is determined based on the error between historical output results and actual results.

[0178] Step S5: Candidate target screening and relevance determination

[0179] (1) Candidate pool screening

[0180] In this embodiment, the weighting coefficients for configuration value, expected contribution, risk, and strategic value are 0.35, 0.30, 0.20, and 0.15, respectively. The comprehensive score for each candidate configuration objective is calculated according to the following formula:

[0181]

[0182] in, , , and They represent the first The table shows the configuration value fusion value, expected contribution fusion value, risk fusion value, and strategic value fusion value of each candidate configuration target. All fusion values ​​in Table 2 are the results obtained after weighted fusion in step S4 according to the decision weights of each agent.

[0183] Table 2. Comprehensive Scores and Ranking Results of Candidate Configuration Targets

[0184]

[0185] Based on the comprehensive score of the single-objective configuration, the top five candidate configuration objectives were selected from high to low and entered into the configuration candidate pool. These objectives are: Baijiu brand upgrade and channel expansion project, battery supply chain digital optimization project, Internet game business recovery and user growth project, consumer electronics supply chain integration and Apple OEM optimization project, and bank digital transformation and risk control system upgrade project.

[0186] (2) Determination of the relevance of candidate configuration targets

[0187] Step S2 has calculated and saved the correlation coefficients between the targets based on the standardized contribution change rate sequence of each candidate configuration target over the most recent 60 execution cycles. Step S5 extracts the correlation sub-matrix corresponding to the configuration candidate pool from the saved correlation results, without recalculating the correlation between targets. The correlation coefficients are obtained according to the following formula:

[0188] ,

[0189] Where L=60, and These represent the average rate of change of standardized contribution for the two projects over 60 execution cycles.

[0190] Table 3. Correlation coefficients of some targets in the candidate pool.

[0191]

[0192] Table 3 lists the correlation coefficients of some targets in the configuration candidate pool. The correlation coefficients of the remaining candidate configuration targets are extracted from the correlation results saved in step S2 in the same way.

[0193] Step S6: Multi-objective combination optimization and allocation weight solution

[0194] The expected contribution fusion value, strategic value fusion value, and risk correlation matrix formed in step S5 of each candidate configuration objective in the configuration candidate pool are input into the multi-objective combination optimization model.

[0195] Expected contribution vector: (These correspond to the following projects respectively: Baijiu brand upgrade and channel expansion project, battery supply chain digital optimization project, internet game business recovery and user growth project, consumer electronics supply chain integration and Apple OEM optimization project, and bank digital transformation and risk control system upgrade project.)

[0196] Strategic value vector:

[0197]

[0198] The order of candidate configuration targets corresponding to each element is consistent with the expected contribution value vector.

[0199] Step S5 constructs a risk correlation matrix based on the correlation coefficients between candidate configuration targets and the risk scale obtained by normalizing the risk fusion values ​​of each candidate configuration target. , its first Line number The elements of a column are represented as follows:

[0200]

[0201] in, Indicates the first The candidate configuration target and the first The correlation coefficient between the candidate configuration targets and These represent the risk scales obtained by normalizing the risk fusion values ​​of the two candidate configuration targets. The risk correlation matrix is ​​used to calculate the overall combined risk. .

[0202] The constraints are as follows: the sum of the configuration weights of all candidate configuration targets is 100%; the upper limit of the configuration weight of a single candidate configuration target is 30%, and the lower limit is 5%; the concentration of weights among all candidate configuration targets does not exceed the preset upper limit of target concentration, i.e.:

[0203]

[0204] The aforementioned expected contribution vector, strategic value vector, and risk correlation matrix are input into a multi-objective combined optimization model. The model is then solved under constraints on total resource allocation, upper and lower limits of individual objective weights, and objective concentration to obtain the allocation weights of each candidate objective. The allocation resource amount for each candidate objective is the product of its allocation weight and an annual project budget of 10 million yuan.

[0205] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A multi-objective resource allocation method based on multimodal data and multi-agent collaboration, characterized in that, Includes the following steps: S1: Collect multimodal data of multiple candidate targets and preprocess them. The multimodal data includes text data, image data, and numerical data. S2: For each candidate target, extract the text features, visual features, and numerical features corresponding to the multimodal data, and use multiple professional analytical agents to perform fusion analysis on the features to generate multiple preliminary analysis results corresponding to the candidate target; S3: Perform multi-agent debate based on the multiple preliminary analysis results. The multi-agent debate includes a debate on the configuration value between supporting and opposing configuration agents, and a debate on resource strategies between multiple resource allocation agents, to obtain the multi-agent decision result corresponding to the candidate target. S4: The decision results of the multi-agent system are weighted and fused to obtain the single-objective configuration comprehensive score of each candidate objective and the corresponding expected contribution value, risk value and strategic value; S5: Based on the comprehensive score of the single-target configuration, the candidate targets are screened to form a configuration candidate pool, and the correlation between each candidate configuration target in the configuration candidate pool is determined; S6: With the optimization objectives of improving the expected contribution and strategic value of the configuration target and reducing the configuration risk, a multi-objective combined optimization model is constructed based on the expected contribution value, risk value and strategic value value of each candidate configuration target and the correlation between the targets. The configuration weight of each candidate configuration target is solved under the constraints of total resources and target weight.

2. The method according to claim 1, characterized in that, Step S2 specifically includes: S21: Extract features from the image data, text data, and numerical data of each candidate target to obtain visual features, text features, and numerical features; S22: Determine the reliability of each modal data based on the data quality of each modality and the cross-modal consistency between each modality; S23: For different professional analytical agents, determine the corresponding multimodal fusion weights based on the reliability of each modal data and the contribution of each modal data to the corresponding analytical task; S24: Based on the multimodal fusion weights, visual features, text features, and numerical features are weighted and fused to obtain the comprehensive features corresponding to each professional analytical agent; S25: Each professional analytical agent generates preliminary analysis results based on its corresponding comprehensive characteristics. The preliminary analysis results include at least configuration value evaluation, risk evaluation, expected contribution evaluation, strategic value evaluation, and confidence level. The specialized analytical agents include fundamental analysis agents, performance analysis agents, and environmental analysis agents.

3. The method according to claim 1, characterized in that, In the configuration value debate in step S3, The support configuration agent determines the support configuration score of candidate targets based on the preliminary analysis results output by multiple professional analysis agents, generates corresponding supporting evidence, and outputs the confidence level of the support configuration results. The opposing configuration agent determines the opposing configuration score of the candidate target based on the preliminary analysis results output by multiple professional analysis agents, generates corresponding opposing evidence, and outputs the confidence level of the opposing configuration result. The supporting and opposing agents update their evaluation results based on the scores and evidence output by the other party, until the changes in the supporting and opposing configuration scores in two consecutive rounds are both less than a preset threshold, or the maximum number of debate rounds is reached. The result is a debate on the configuration value, which includes the supporting configuration score, supporting evidence, confidence level of the supporting configuration result, opposing configuration score, opposing evidence, and confidence level of the opposing configuration result.

4. The method according to claim 3, characterized in that, In step S3, the plurality of resource allocation agents include aggressive resource allocation agents, conservative resource allocation agents, and neutral resource allocation agents, in the resource strategy debate, Based on the results of the configuration value debate, the radical resource allocation agent, the conservative resource allocation agent, and the neutral resource allocation agent respectively generate the suggested resource ratio, expected contribution value, risk value, and confidence level of the corresponding resource allocation suggestion for the candidate target. Each resource allocation agent updates its own resource allocation recommendations based on the suggested resource ratios, expected contribution values, and risk values ​​output by other resource allocation agents. When the maximum difference in the suggested resource ratios output by each resource allocation agent is less than a preset threshold, or when the change in suggested resource ratios in two consecutive rounds is less than a preset threshold, the resource strategy debate is stopped, and the final resource allocation suggestions and their confidence levels for each resource allocation agent are obtained.

5. The method according to claim 4, characterized in that, Step S4 specifically includes: S41: Determine the decision weight of each agent based on the confidence level of the output results of each agent, the reliability of the source data corresponding to the output results, and the historical evaluation accuracy of each agent. Specifically, for professional analytical agents, the reliability of the source data is determined based on the reliability of each modal data they rely on and the corresponding modal fusion weights; for agents that support configuration, agents that oppose configuration, and agents that allocate resources, the reliability of the source data is determined by a weighted average of the source data reliability and output confidence of their direct upstream agents. The accuracy of the historical evaluation is determined based on the normalized error between the output results of each agent in the completed resource allocation task and the corresponding actual execution results, and is updated after obtaining new actual execution results. S42: Based on the decision weights, the configuration value, expected contribution, risk and strategic value output by each agent are weighted and fused to obtain the configuration value fusion value, expected contribution fusion value, risk fusion value and strategic value fusion value corresponding to the candidate target; S43: The configuration value fusion value, expected contribution fusion value and strategic value fusion value are positively weighted, and the risk fusion value is reversely corrected to obtain the single-target configuration comprehensive score of the candidate target.

6. The method according to claim 5, characterized in that, Step S6 specifically includes: Based on the expected contribution fusion value and strategic value fusion value of each candidate configuration objective, the configuration benefit of each candidate configuration objective is determined; based on the risk fusion value of each candidate configuration objective and the correlation between the candidate configuration objectives, the overall configuration risk of the multi-objective combination is determined; with the optimization objective of maximizing the sum of the configuration benefits of each candidate configuration objective and minimizing the overall configuration risk, a multi-objective combination optimization model is constructed. Under the constraints that the sum of the configuration weights satisfies the preset total resource amount, the configuration weights of each candidate configuration target are within the corresponding upper and lower weight limits, and the target concentration does not exceed the preset threshold, the multi-objective combination optimization model is solved to obtain the configuration weights of each candidate configuration target.

Images