Multi-agent collaborative decision-making method and system based on generative large model

Abstract

The present application relates to the technical field of data processing, in particular to a multi-agent collaborative decision-making method and system based on a generative large model, comprising: constructing multiple agents with different roles by a generative large model, and performing multi-round debate on a decision-making task; collecting inclination probability of each round and confidence of the last round of each agent, and obtaining a static aggregation weight; obtaining inclination probability stability, deviation, net convergence and effective persuasion based on the inclination probability, and obtaining a dynamic reliability correction factor; correcting the static aggregation weight by using the dynamic reliability, and calculating the score of each candidate option by combining the inclination probability of the last round of each agent, and taking the candidate option with the highest score as the final decision output. The present application solves the problem of low accuracy, stability and credibility when processing complex decision-making tasks.

Description

Technical Field

[0001] This invention relates to the field of data processing technology. More specifically, this invention relates to a multi-agent collaborative decision-making method and system based on generative large models. Background Technology

[0002] With the rapid development of generative large model technology, it has been applied in fields such as intelligent question answering, code generation, text analysis, solution design, scientific reasoning, and complex task planning. Compared with traditional rule-based or single-model decision-making methods, generative large models can combine massive amounts of knowledge and contextual information to analyze complex problems from multiple perspectives and output highly logical reasoning results. However, when faced with complex decision-making tasks involving multiple candidate solutions, multiple constraints, and multi-objective optimization, a single generative large model is easily affected by factors such as limited reasoning paths, knowledge bias, random sampling fluctuations, and context window limitations. The same problem may yield different decision results in different reasoning processes, making it difficult to guarantee the stability and reliability of the decision.

[0003] Existing multi-agent collaborative decision-making technologies typically employ multiple agents with different roles to engage in multiple rounds of interaction around the same decision-making task. Each agent continuously adjusts its own judgment based on the views of other agents, and finally aggregates the results through majority voting, average probability fusion, fixed weighting, or based on the confidence level of the last round to determine the final decision outcome.

[0004] However, while the above-mentioned solutions can alleviate the inference bias that is prone to occur in single generative large models to some extent and improve the analytical ability of complex problems, most existing technologies focus more on the final output of the agents and pay insufficient attention to the evolution of viewpoints throughout the debate process. They usually only use the output of the final round as the basis for decision-making, or only allocate the influence of each agent in the final aggregation based on fixed weights, historical experience weights, and confidence scores in the final round. They do not make full use of the stability of judgments, the trend of viewpoint changes, and the influence of the agents on the formation of consensus in the overall discussion, which are shown by the agents throughout the debate process. As a result, a large amount of effective information contained in the interaction of the whole process is not fully explored and utilized, which limits the further improvement of the performance of multi-agent collaborative decision-making, and leads to problems of low accuracy, stability and reliability when dealing with complex decision-making tasks. Summary of the Invention

[0005] To address the issues of low accuracy, stability, and reliability in handling complex decision-making tasks as mentioned in the background art, the present invention provides solutions in the following aspects.

[0006] In a first aspect, the present invention provides a multi-agent collaborative decision-making method based on a generative large model, comprising: constructing multiple agents with different roles using a generative large model, and conducting multiple rounds of debate on a decision task containing multiple candidate options; collecting the tendency probability of each agent towards each candidate option in each round and the confidence score in the final round, and exponentially processing the confidence score to obtain the static aggregation weight of each agent; obtaining the tendency probability stability of each agent throughout all rounds of debate based on the tendency probability; and using the average tendency probability of the remaining agents (excluding the current agent) towards each candidate option in each round as the group consensus, and calculating... The deviation between the current agent's probabilities of favoring each candidate option and the group consensus in each round is calculated. The net convergence is obtained based on the difference between the mean of these deviations in the first and second halves of the multi-round debate. The effective persuasiveness of each agent is acquired, and this effective persuasiveness is positively correlated with the stability of the probabilities of favoring and the net convergence. A dynamic reliability correction factor is obtained based on the stability of the probabilities of favoring and the effective persuasiveness. The static aggregation weights are corrected using the dynamic reliability correction factor to obtain corrected aggregation weights. The scores of the corrected aggregation weights and the probabilities of favoring each agent in the final round are calculated, and the candidate option corresponding to the highest score is used as the final decision output.

[0007] The above technical solution comprehensively utilizes the changes in the tendency probability of each agent throughout the multi-round debate process, the degree of consistency with the overall judgment, and the impact on the formation of consensus on the overall opinion to conduct dynamic reliability evaluation of each agent. Based on this, the aggregation weight is adaptively adjusted so that the final decision not only reflects the judgment results of each agent in the last round, but also fully reflects the comprehensive performance of each agent throughout the debate process. This avoids the problem that reliability differences cannot be effectively distinguished when making decisions based solely on fixed weights or the output of the last round, improves the rationality, stability, and accuracy of the aggregation results, and thus enhances the reliability and credibility of multi-agent collaborative decision-making in generative large models.

[0008] Furthermore, obtaining the static aggregate weight of each agent includes: taking the natural index of the confidence score of any agent in the last round to obtain a confidence index value; summing the confidence index values ​​of all agents to obtain a total confidence index; calculating the ratio of the confidence index value to the total confidence index, and using the ratio as the static aggregate weight of the agent.

[0009] Furthermore, the stability of each agent's tendency probability throughout all rounds of debate is obtained, including: calculating the absolute difference of any agent's tendency probability for each candidate option between any two adjacent rounds; summing the absolute differences of the tendency probabilities for each candidate option and dividing by 2 to obtain the magnitude of the change in the tendency probability of any agent; averaging the magnitude of the change in the tendency probability of any agent from the second round to the last round to obtain the average magnitude of the change in the tendency probability; and subtracting the average magnitude of the change in the tendency probability from 1 to obtain the stability of the tendency probability of any agent.

[0010] The above technical solution quantifies the consistency of agents' judgments by statistically analyzing the changes in the tendency probability of each agent in adjacent rounds throughout the debate process. Agents with smaller changes in tendency probability receive higher stability evaluations, while agents with larger changes receive lower stability evaluations. This effectively distinguishes the decision stability of different agents, reduces the impact of single-round fluctuations on the evaluation results, provides a more objective data basis for subsequent reliability assessments, and further improves the stability and credibility of multi-agent collaborative decision-making results.

[0011] Furthermore, the calculation of the deviation includes: for any target agent, averaging the probabilities of inclination towards each candidate option for all agents except the target agent in any round to obtain the group consensus for any round; calculating the absolute difference between the group consensus and the probability of inclination towards each candidate option for the target agent in any round, summing the absolute differences of each candidate option and dividing by 2 to obtain the deviation of the target agent in any round, wherein the target agent is any agent among multiple agents with different roles.

[0012] The above technical solution compares the overall judgment formed by the target agent with that of other agents, and quantifies the degree of consistency between the target agent and the overall judgment. This allows the degree of deviation to objectively reflect the differences in viewpoints of agents in each round of debate, effectively distinguishing between agents that are more consistent with the overall judgment and those that deviate significantly. This provides a basis for subsequent analysis of the changing trends of agents in the debate process and reliability assessment, further improving the accuracy and credibility of multi-agent collaborative decision-making results.

[0013] Further, the net convergence quantity includes: dividing all debate rounds into a first half and a second half in sequence; calculating the average deviation of each round in the first half to obtain the average deviation of the first half; calculating the average deviation of each round in the second half to obtain the average deviation of the second half; and subtracting the average deviation of the second half from the average deviation of the first half to obtain the net convergence quantity.

[0014] The above-mentioned technical solution quantifies the convergence of viewpoints of agents throughout the debate process by comparing the overall deviation between the two stages before and after the debate. It can effectively reflect whether agents gradually form a consensus as the debate progresses, reduce the impact of single-round deviation fluctuations on the evaluation results, provide a more comprehensive basis for subsequent reliability assessment, and further improve the stability and credibility of multi-agent collaborative decision-making results.

[0015] Furthermore, the method of obtaining the effective persuasiveness of each agent includes: adding the net convergence quantity to 1 and dividing by 2 to obtain the mapping convergence quantity; and using the product of the tendency probability stability and the mapping convergence quantity as the effective persuasiveness.

[0016] The above technical solution comprehensively evaluates the actual contribution of agents in the debate process by taking into account the stability of the agents' judgments and their impact on the formation of consensus in the overall opinion. This ensures that only agents that maintain stable judgments and can promote overall consensus receive higher evaluation results, thereby improving the ability to distinguish the overall performance of agents. This provides a more accurate basis for subsequent reliability assessment and aggregation weight correction, further enhancing the reliability and accuracy of multi-agent collaborative decision-making results.

[0017] Furthermore, the dynamic reliability correction factor includes: weighting and summing the tendency probability stability and effective persuasion with a preset matching coefficient to obtain the reliability of each agent; averaging the reliability of all agents to obtain the mean reliability; calculating the ratio of the reliability of each agent to the mean reliability, and using the ratio as the dynamic reliability correction factor for each agent.

[0018] The above technical solution evaluates the reliability of each agent by comprehensively considering the probability stability and effective persuasiveness, and constructs a dynamic reliability correction factor based on the overall reliability level. This allows the evaluation results of each agent to be dynamically adjusted according to their actual performance during the debate process, effectively improving the objectivity and adaptability of the reliability evaluation, providing a more accurate basis for aggregate weight correction, and further enhancing the stability, reliability, and accuracy of the multi-agent collaborative decision-making results.

[0019] Further, the modified aggregate weight includes: calculating the product of the dynamic reliability correction factor of any agent and its static aggregate weight to obtain the modified product of any agent; summing the modified products of all agents to obtain the total modified product; calculating the ratio of the modified product of any agent to the total modified product; and using the ratio as the modified aggregate weight of any agent.

[0020] The above technical solution integrates dynamic reliability evaluation results with static aggregated weights and makes unified adjustments to the integrated results. This allows the final weights of each agent to simultaneously reflect the performance of the final round of decision-making and the overall reliability throughout the debate process. This enhances the ability of aggregated weights to represent the actual decision-making contribution of agents, reduces the impact of agents with lower reliability on the final decision results, improves the rationality and adaptability of weight allocation, provides more accurate and stable weight support for the final decision, and further improves the reliability and accuracy of multi-agent collaborative decision-making results.

[0021] Furthermore, the step of using the candidate option corresponding to the maximum score as the final decision output includes: for any candidate option, calculating the product of the probabilities of each agent in the last round towards any candidate option and its corrected aggregate weights, summing the products of all agents to obtain the score of any candidate option, and using the candidate option corresponding to the maximum score as the final decision output.

[0022] In a second aspect, the present invention provides a multi-agent cooperative decision-making system based on a generative large model, including a memory and a processor. The memory stores computer program instructions, which, when executed by the processor, implement the multi-agent cooperative decision-making method based on a generative large model as described above.

[0023] The beneficial effects of this invention are as follows: This invention utilizes the changes in the probabilities of multiple agents throughout the debate process, their consistency with the overall judgment, and their actual contributions to forming a consensus to dynamically evaluate the reliability of each agent. Based on this, the aggregation weights are adaptively adjusted, so that the final decision takes into account both the agent's final round judgment results and the overall debate performance. This effectively improves the ability to distinguish the decision contributions of different agents, reduces the impact of agents with lower reliability on the final result, and thus enhances the accuracy, stability, and credibility of the multi-agent collaborative decision-making results. Attached Figure Description

[0024] Figure 1 This is a flowchart illustrating a multi-agent collaborative decision-making method based on a generative large model according to an embodiment of the present invention; Figure 2 This is a schematic diagram illustrating the deviation of each agent in each round of debate in a multi-agent collaborative decision-making method based on a generative large model according to an embodiment of the present invention. Detailed Implementation

[0025] Implementation examples of a multi-agent collaborative decision-making method based on generative large models.

[0026] like Figure 1As shown in the flowchart of the multi-agent cooperative decision-making method based on a generative large model according to an embodiment of the present invention, the method includes the following steps: S1: Construct multiple agents with different roles using a generative large model, and conduct multiple rounds of debate on a decision-making task containing multiple candidate options; collect the probabilities of each agent's preference for each candidate option in each round and the confidence score in the final round, and obtain the static aggregate weight of each agent after exponentializing the confidence scores.

[0027] In a preferred embodiment, the generative large model is a computational model trained on massive text corpora that can generate natural language output based on input prompts.

[0028] It is understood that the decision-making task includes multiple candidate options, which refer to mutually exclusive alternative answers to the same question. For example, for a decision-making task regarding whether to approve a loan application, the candidate options could be approval or rejection. The computer organizes multiple agents to conduct multiple rounds of debate on this decision-making task. Each round of debate refers to the complete number of iterations in which all agents sequentially express their opinions and read each other's views. In each round, each agent first reads the opinion text of all agents from the previous round, and then updates its own probability of inclination towards the multiple candidate options; for example, in the first round, the inclination to approve is 0.6, and the inclination to reject is 0.4.

[0029] As an exemplary numerical scenario, assume the number of agents is 3, including the proposer (agent 1), the challenger (agent 2), and the reviewer (agent 3); the number of candidate options is 2; and there are 4 rounds of debate. The computer collects the following probabilities for each round: Agent 1: 0.6 and 0.4, 0.65 and 0.35, 0.7 and 0.3, 0.72 and 0.28; Agent 2: 0.3 and 0.7, 0.4 and 0.6, 0.5 and 0.5, 0.55 and 0.45; Agent 3: 0.5 and 0.5, 0.55 and 0.45, 0.6 and 0.4, 0.62 and 0.38. In the final round of debate, the computer asks each agent to report a confidence score, which is a self-assessed value of the agent's confidence in its own conclusion in the final round.

[0030] The process of obtaining the static aggregate weight of each agent is as follows: take the natural index of the confidence score of any agent in the last round to obtain the confidence index value; sum the confidence index values ​​of all agents to obtain the total confidence index; calculate the ratio of the confidence index value to the total confidence index, and use the ratio as the static aggregate weight of any agent.

[0031] By exponentializing the confidence scores of each agent's final output, the difference between agents with high confidence and those with low confidence is appropriately amplified. Subsequently, the total amount of the exponentialized results of all agents is used to normalize the weights of each agent, ensuring that the aggregated weights of each agent are on a uniform scale and maintain overall balance. This enhances the contribution of high-confidence agents to the final decision result while retaining the participation of all agents in the decision-making process. It also avoids imbalances in weight allocation caused by differences in confidence scales or uneven numerical distribution among different agents, thereby improving the discriminative power, stability, and comparability of the aggregated weights.

[0032] S2: Obtain the stability of the tendency probability of each agent throughout all rounds of debate, calculate the deviation of the current agent's tendency probability for each candidate option in each round from the group consensus; obtain the net convergence amount based on the difference between the mean of the deviation in the first half and the second half of the multi-round debate.

[0033] In a preferred embodiment, obtaining the stability of the tendency probability of each agent throughout all rounds of debate includes: calculating the absolute difference of the tendency probability of any agent for each candidate option between any two adjacent rounds; summing the absolute differences of the tendency probabilities of each candidate option and dividing by 2 to obtain the magnitude of the change in the tendency probability of any agent; averaging the magnitude of the change in the tendency probability of any agent from the second round to the last round to obtain the average magnitude of the change in the tendency probability; and subtracting the average magnitude of the change in the tendency probability from 1 to obtain the stability of the tendency probability of any agent.

[0034] Using the change in the tendency probability of each agent in consecutive debate rounds as the evaluation criterion, this method calculates the magnitude of the change in the tendency probability of each candidate option between adjacent rounds and further statistically analyzes the average change level throughout the entire debate process. This comprehensively reflects the decision consistency and viewpoint maintenance ability of each agent throughout the debate. The reverse result of the change level is then used as the tendency probability stability, so that agents with smaller tendency probability changes obtain higher stability, while agents that frequently adjust their tendency probabilities obtain lower stability. This effectively distinguishes agents that maintain stable judgments and those that repeatedly change their judgments, reducing evaluation bias caused by local fluctuations or single-round anomalies, improving the quantitative accuracy of agent decision stability, and providing an objective and stable data foundation for subsequent reliability assessment and aggregation weight correction. Ultimately, this enhances the stability and credibility of multi-agent collaborative decision-making results.

[0035] like Figure 2 The diagram shows the deviation of each agent in each round of debate in the multi-agent collaborative decision-making method based on a generative large model according to an embodiment of the present invention.

[0036] The calculation of the deviation includes: for any target agent, averaging the probabilities of inclination towards each candidate option for all agents except the target agent in any round to obtain the group consensus for any round; calculating the absolute difference between the group consensus and the probability of inclination towards each candidate option for the target agent in any round, summing the absolute differences of each candidate option and dividing by 2 to obtain the deviation of the target agent in any round, wherein the target agent is any agent among multiple agents with different roles.

[0037] Using the average probability of each candidate option held by all agents except the target agent in each round as a reference, the degree of consistency between the target agent and the overall judgment is quantitatively evaluated by comparing the degree of difference between the target agent and the overall judgment. The differences on each candidate option are combined to form the deviation, so that the evaluation result can reflect the overall deviation of the target agent on all candidate options, rather than judging based on a single candidate option. This avoids the influence of local bias changes on the evaluation result, improves the ability to characterize the consistency of the agent's views and the objectivity of the evaluation result, and provides a reliable basis for subsequent analysis of the changing trend of the agent and the overall judgment during the debate process.

[0038] The net convergence quantity includes: dividing all debate rounds into a first half and a second half in sequence; calculating the average deviation of each round in the first half to obtain the average deviation of the first half; calculating the average deviation of each round in the second half to obtain the average deviation of the second half; and subtracting the average deviation of the second half from the average deviation of the first half to obtain the net convergence quantity.

[0039] The debate process is divided into two halves: the first half and the second half. The average deviation rate within each half is calculated, and the overall deviation changes between the two halves are compared to quantify the trend of the agent gradually approaching or deviating from the overall judgment during continuous interaction. This allows the evaluation results to reflect not only the degree of consistency with the overall judgment at a specific moment but also the direction and magnitude of the evolution of viewpoints throughout the debate. This effectively distinguishes between situations where consensus is gradually formed through thorough communication and situations where disagreements persist, reducing the impact of single-round fluctuations on the evaluation results. It also improves the ability to characterize the evolutionary features of the agent's collaborative decision-making process, providing a more comprehensive and stable basis for subsequent reliability assessments and aggregate weight adjustments, further enhancing the accuracy, consistency, and credibility of multi-agent collaborative decision-making results.

[0040] S3: Obtain the effective persuasiveness of each agent, and obtain a dynamic reliability correction factor based on the stability of the tendency probability and the effective persuasiveness.

[0041] In a preferred embodiment, obtaining the effective persuasiveness of each agent includes: adding the net convergence quantity to 1 and dividing by 2 to obtain the mapping convergence quantity; and using the product of the tendency probability stability and the mapping convergence quantity as the effective persuasiveness.

[0042] By mapping the net convergence quantity to a uniform numerical range, it is easier to comprehensively evaluate it in conjunction with the stability of the tendency probability. Subsequently, the mapped convergence result and the stability of the tendency probability are used together as the evaluation criteria for effective persuasiveness. Only when an agent maintains relatively stable judgments and can promote the overall judgments to gradually converge during the debate can it obtain high effective persuasiveness. Agents that only have stable judgments but do not promote overall consensus, or whose judgments change frequently while promoting overall convergence, have their effective persuasiveness limited accordingly. This enables an objective evaluation of the agent's overall performance and improves the ability to distinguish the agent's actual contribution to decision-making.

[0043] The dynamic reliability correction factor includes: weighting and summing the tendency probability stability and effective persuasion with a preset matching coefficient to obtain the reliability of each agent; averaging the reliability of all agents to obtain the mean reliability; calculating the ratio of the reliability of each agent to the mean reliability, and using the ratio as the dynamic reliability correction factor for each agent.

[0044] By comprehensively evaluating the probabilistic stability and effective persuasiveness of each agent, reliability can reflect both the stability of the agent's judgment throughout the debate process and its actual contribution to the formation of consensus in the overall collaborative decision-making. On this basis, a relative comparison is made with the overall reliability level of all agents, and a dynamic reliability correction factor is constructed based on the difference between the reliability of each agent and the overall average level. This allows the correction result to be adaptively adjusted according to the agent's performance in different decision-making tasks and different debate processes, rather than relying on fixed weights for evaluation. This effectively enhances the ability to distinguish between high-performing and low-performing agents and improves the objectivity and adaptability of reliability evaluation.

[0045] S4: Use the dynamic reliability correction factor to correct the static aggregate weight to obtain the corrected aggregate weight, calculate the score of the corrected aggregate weight and the tendency probability of each agent in the last round, and take the candidate option corresponding to the maximum score as the final decision output.

[0046] In a preferred embodiment, the corrected aggregate weight includes: calculating the product of the dynamic reliability correction factor of any agent and its static aggregate weight to obtain the corrected product of any agent; summing the corrected products of all agents to obtain the total corrected product; calculating the ratio of the corrected product of any agent to the total corrected product; and using the ratio as the corrected aggregate weight of any agent.

[0047] By combining dynamic reliability correction factors with static aggregation weights, the static aggregation weights can be dynamically adjusted based on the actual performance of the agents throughout the debate process, while retaining the confidence evaluation results of the agents in the final round. Furthermore, the correction results are normalized to ensure that the aggregated weights of all agents maintain a uniform scale and overall balance after correction. This avoids imbalances in weight allocation or overall weight shifts caused by dynamic adjustments. Consequently, agents with higher reliability receive higher decision contributions in the final aggregation process, while those with relatively lower reliability contribute less. This achieves an effective fusion of static and dynamic evaluation, improves the ability of aggregation weights to represent the comprehensive performance of agents, enhances the rationality and adaptability of the final weight allocation, provides a more accurate and stable weight basis for subsequent candidate option score calculations, and further improves the accuracy, reliability, and stability of multi-agent collaborative decision-making results.

[0048] The step of taking the candidate option with the highest aggregate score as the final decision output includes: For any candidate option, calculating the product of the probability of each agent's preference for any candidate option in the last round and its corrected aggregate weight, summing the products of all agents to obtain the score of any candidate option, and taking the candidate option corresponding to the highest score as the final decision output.

[0049] The above scheme calculates the probabilities of each agent's final output and their corresponding adjusted aggregation weights for each candidate option, and summarizes the overall support of all agents for that candidate option to form the final score for each candidate option. The final decision is then determined based on the scores of each candidate option. This ensures that the final output not only comprehensively considers the support of each agent for the candidate option, but also fully reflects the overall reliability and decision contribution of each agent throughout the debate process. It avoids the problem of not being able to effectively distinguish the influence of different agents when making decisions based solely on simple majority voting or average support. This increases the contribution of highly reliable agents to the final decision and reduces the interference of less reliable agents on the decision result. As a result, the final decision result can more comprehensively reflect the comprehensive judgment formed by multi-agent collaborative debate, improving the accuracy, stability, and credibility of multi-agent collaborative decision-making.

[0050] The present invention comprehensively analyzes the stability of each agent's judgment, its consistency with the overall judgment, and its actual contribution to forming a consensus throughout the multi-round debate process. It performs dynamic reliability evaluation on each agent and adaptively adjusts the aggregation weights based on the final round decision results. This ensures that the final decision fully reflects the comprehensive performance of each agent throughout the collaborative decision-making process, reduces the impact of agents with lower reliability on the decision results, improves the rationality, stability, and accuracy of the aggregated decision, and thus enhances the reliability and credibility of multi-agent collaborative decision-making in generative large models.

[0051] Example of a multi-agent collaborative decision-making system based on generative large models: The multi-agent cooperative decision-making system based on a generative large model, according to an embodiment of the present invention, includes a processor and a memory. The memory stores computer program instructions, which, when executed by the processor, implement the multi-agent cooperative decision-making method based on a generative large model as described above.

[0052] The multi-agent collaborative decision-making system based on generative large models also includes other components well known to those skilled in the art, such as communication interfaces. Their settings and functions are known in the art and will not be described in detail here.

[0053] While this specification has shown and described numerous embodiments of the invention, it will be apparent to those skilled in the art that such embodiments are provided by way of example only. Many modifications, alterations, and alternatives will occur to those skilled in the art without departing from the spirit and essence of the invention. It should be understood that various alternatives to the embodiments of the invention described herein may be employed in the practice of this invention.

Claims

1. A multi-agent cooperative decision-making method based on generative large models, characterized in that, include: Multiple agents with different roles are constructed using a generative large model, and multiple rounds of debate are conducted on decision-making tasks that include multiple candidate options. Collect the probability of each agent's preference for each candidate option in each round and the confidence score in the last round. After exponentializing the confidence scores, obtain the static aggregate weight of each agent. Based on the aforementioned tendency probability, the stability of the tendency probability of each agent throughout all rounds of debate is obtained; the average tendency probability of the other agents (excluding the current agent) for each candidate option in each round is taken as the group consensus, and the deviation between the tendency probability of the current agent for each candidate option in each round and the group consensus is calculated. The net convergence quantity is obtained by the difference between the mean of the deviation degree in the first half and the second half of the multi-round debate; the effective persuasiveness of each agent is obtained, and the effective persuasiveness is positively correlated with the stability of the tendency probability and the net convergence quantity. A dynamic reliability correction factor is obtained based on the aforementioned tendency probability stability and effective persuasiveness. The static aggregate weights are corrected using the dynamic reliability correction factor to obtain the corrected aggregate weights. The scores of the corrected aggregate weights and the tendency probabilities of each agent in the last round are calculated, and the candidate option corresponding to the maximum score is used as the final decision output.

2. The multi-agent cooperative decision-making method based on a generative large model according to claim 1, characterized in that, The process of obtaining the static aggregate weights of each agent includes: The confidence index value is obtained by taking the natural index of the confidence score of any agent in the last round. The confidence index values ​​of all agents are summed to obtain the total confidence index. Calculate the ratio of the confidence index value to the sum of the confidence indices, and use the ratio as the static aggregation weight of any agent.

3. The multi-agent cooperative decision-making method based on a generative large model according to claim 1, characterized in that, Obtain the stability of the tendency probability for each agent throughout all rounds of debate, including: Calculate the absolute difference in the probability of any agent's preference for each candidate option between any two adjacent rounds, sum the absolute differences in the probability of each candidate option's preference, and divide by 2 to obtain the magnitude of the change in the probability of any agent's preference. The average change in the tendency probability of any agent from the second round to the last round is averaged to obtain the average change in tendency probability. Subtract the average change in tendency probability from 1 to obtain the stability of the tendency probability of any agent.

4. The multi-agent cooperative decision-making method based on a generative large model according to claim 1, characterized in that, The calculation of the deviation includes: For any target agent, in any round, the probabilities of the inclinations of the remaining agents (excluding the target agent) towards each candidate option are averaged to obtain the group consensus for any round. Calculate the absolute difference between the group consensus and the target agent's tendency probability for each candidate option in any round. Sum the absolute differences of each candidate option and divide by 2 to obtain the deviation of the target agent in any round. The target agent is any agent among multiple agents with different roles.

5. The multi-agent cooperative decision-making method based on a generative large model according to claim 1, characterized in that, The net convergence includes: All debate rounds are divided into the first half and the second half in sequence; the average deviation of each round in the first half is calculated to obtain the average deviation of the first half; the average deviation of each round in the second half is calculated to obtain the average deviation of the second half. The net convergence is obtained by subtracting the average deviation of the second half from the average deviation of the first half.

6. The multi-agent cooperative decision-making method based on a generative large model according to claim 1, characterized in that, The acquisition of the effective persuasiveness of each agent includes: Add the net convergence quantity to 1 and divide by 2 to obtain the mapping convergence quantity; the product of the tendency probability stability and the mapping convergence quantity is taken as the effective persuasive power.

7. The multi-agent cooperative decision-making method based on a generative large model according to claim 1, characterized in that, The dynamic reliability correction factor includes: The reliability of each agent is obtained by weighting and summing the tendency probability stability and effective persuasiveness using preset matching coefficients. The reliability of all agents is averaged to obtain the mean reliability; the ratio of the reliability of each agent to the mean reliability is calculated, and the ratio is used as the dynamic reliability correction factor for each agent.

8. The multi-agent cooperative decision-making method based on a generative large model according to claim 1, characterized in that, The corrected aggregate weights include: Calculate the product of the dynamic reliability correction factor of any agent and its static aggregate weight to obtain the correction product of any agent. Summate the correction products of all agents to obtain the total correction product. Calculate the ratio of the correction product of any agent to the total correction product. Use the ratio as the corrected aggregate weight of any agent.

9. The multi-agent cooperative decision-making method based on a generative large model according to claim 1, characterized in that, The step of using the candidate option corresponding to the maximum score as the final decision output includes: For any candidate option, calculate the product of the probability of each agent's preference for any candidate option in the last round and its adjusted aggregate weight. Sum the products of all agents to obtain the score of any candidate option, and take the candidate option with the maximum score as the final decision output.

10. A multi-agent collaborative decision-making system based on a generative large model, characterized in that, It includes a memory and a processor, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, the multi-agent cooperative decision-making method based on any one of generative large models as described in any one of claims 1 to 9 is implemented.

Images