A decision-making method and system based on multi-agent deliberation

Abstract

A decision-making method and system based on multi-agent deliberation is proposed. This involves establishing three types of agents and configuring prior probability distributions. Based on evidence data, multiple rounds of deliberative interaction are conducted, including: concurrently calling conservative and radical agents to generate structured arguments; performing Bayesian updates on the prior probability distributions of each agent based on the structured arguments to obtain their respective posterior probability distributions; calling a neutral agent to analyze the structured arguments and generating conservative and radical question lists to drive the conservative and radical agents to revise their respective structured arguments and recalculate their respective posterior probability distributions; after each round of deliberative interaction, determining whether a preset convergence condition has been met; if so, performing fusion processing to generate a fusion probability distribution representing the decision result; if not, continuing to the next round of deliberative interaction; and outputting the decision result based on the fusion probability distribution or the total divergence degree.

Description

Technical Field

[0001] This invention relates to the field of financial risk control technology, and in particular to a decision-making method and system based on multi-agent deliberation. Background Technology

[0002] In the field of financial risk management, especially in high-value decision-making scenarios such as credit approval and customer rating, accurately, robustly, and interpretably assessing risk has become a core challenge. Currently, mainstream technical solutions fall into two main categories: automated decision-making systems based on a single artificial intelligence (AI) model and expert systems based on the experience of business experts.

[0003] The first type of solution typically uses a trained artificial intelligence (AI) model to score or classify a customer's comprehensive information in a single step. While it is highly efficient, the decision-making process is like a "black box," lacking transparency and resulting in insufficient credibility and auditability.

[0004] The second approach relies on risk control experts manually reviewing materials and making judgments based on their experience and rules. While this method offers interpretability, it is severely limited by the breadth of knowledge, experience, and subjective biases of individual experts, making it difficult to maintain consistency in standards and resulting in low efficiency, thus failing to meet the needs of large-scale, high-frequency decision-making.

[0005] Therefore, existing technologies have a significant drawback: the lack of a decision-making method that is transparent in the decision-making process and can systematically quantify and integrate conflicting evidence. Summary of the Invention

[0006] To address the aforementioned technical problems, this invention provides a decision-making method based on multi-agent deliberation, comprising the following steps:

[0007] S100. Establish conservative, radical, and neutral agents, and configure corresponding prior probability distributions for each agent.

[0008] S200: Based on the evidence data of the decision-making object, three types of intelligent agents are scheduled to conduct multiple rounds of deliberative interaction; each round of deliberative interaction includes:

[0009] S210. Conservative and radical agents are invoked in parallel to generate structured arguments, including the direction of influence and weights, based on the evidence data.

[0010] S220. Based on the structured arguments of the conservative and radical agents, Bayesian updates are performed on their respective prior probability distributions to obtain their respective posterior probability distributions.

[0011] S230. Invoke the neutral agent, analyze the structured arguments corresponding to the conservative agent and the radical agent, and generate a conservative question list and a radical question list.

[0012] S240. Based on the conservative question list and the radical question list, drive the conservative agent and the radical agent to revise their respective structured arguments;

[0013] S250. Based on the modified structured arguments of the conservative and radical agents, recalculate and update their respective posterior probability distributions.

[0014] S300. After each round of deliberative interaction, based on the posterior probability distribution of the conservative agent, the posterior probability distribution of the radical agent, and the posterior probability distribution of the neutral agent, determine whether the preset convergence condition has been met.

[0015] If yes, the fusion process is performed based on the posterior probability distributions of the three types of agents to generate a fusion probability distribution used to characterize the decision result; if no, the next round of deliberative interaction continues.

[0016] S400. Based on the fusion probability distribution or total divergence, output the decision result.

[0017] Optionally, S100 includes:

[0018] S110. Load predefined role background prompts for each type of intelligent agent;

[0019] S120. Based on the macro business strategy associated with the current decision-making task, retrieve the preset prior probability distribution library, and configure prior probability distributions representing the initial risk views of conservative agents and radical agents respectively, and configure prior probability distributions representing objective benchmarks for neutral agents.

[0020] S130. Bind the character background prompt words to the corresponding prior probability distribution to complete the creation and initialization of the three types of intelligent agents.

[0021] Optionally, S210 includes:

[0022] S211. Conservative and radical agents are invoked in parallel to output the corresponding classification results, severity level and adjustment factor for each piece of evidence data according to their respective roles and positions.

[0023] S212. Based on the classification results, query the preset evidence classification and grading standard library to obtain the benchmark weight range;

[0024] S213. Combining the adjustment factor and the benchmark weight range, calculate the weight values ​​of the evidence data for the conservative agent and the radical agent, respectively.

[0025] S214. Based on the weight values ​​and influence directions calculated by the conservative agent and the radical agent for all evidence data respectively, generate a structured argument including the influence direction and weight.

[0026] Optionally, S230 includes:

[0027] S231. The structured arguments generated by the conservative agent are forwarded to the radical agent to generate a structured list of conservative rebuttals; the structured arguments generated by the radical agent are forwarded to the conservative agent to generate a structured list of radical rebuttals.

[0028] S232. The radical agent obtains the list of conservative objections and responds radically to each one; the conservative agent obtains the list of radical objections and responds conservatively to each one.

[0029] S233. Invoke the neutral agent to quantify the degree of disagreement between the conservative agent and the radical agent regarding the same evidence data;

[0030] S234. Sort the opinions according to the degree of disagreement, and extract verifiable factual assertions from the highly disagreeing evidence.

[0031] S235. Verify the factual assertion to obtain the verification result and the verified facts;

[0032] S236. Based on the highly divergent evidence and its verification results, construct a conservative inquiry list and a radical inquiry list that require the conservative agent and the radical agent to respond based on the verified facts, respectively.

[0033] Optionally, S240 includes:

[0034] S241. Send the conservative question list and the radical question list to the conservative agent and the radical agent respectively;

[0035] S242. Receive conservative and radical question response information generated by the conservative agent and the radical agent; the conservative or radical question response information includes answers to the conservative or radical question list and instructions for modifying the respective structured arguments.

[0036] S243. Based on the modification instruction, drive the conservative agent and the radical agent to update their respective structured arguments to obtain the revised structured arguments;

[0037] S243. Based on the revised structured arguments, recalculate their respective posterior probability distributions.

[0038] Optional, preset convergence conditions include:

[0039] First convergence condition: The total divergence is less than a first preset threshold for the total divergence; wherein, the total divergence is the sum of the first KL divergence between the posterior probability distribution of the conservative agent and the posterior probability distribution of the neutral agent, and the second KL divergence between the posterior probability distribution of the radical agent and the posterior probability distribution of the neutral agent.

[0040] Second convergence condition: The posterior probability distribution of the conservative agent and the posterior probability distribution of the radical agent are concentrated on the same or adjacent one or two risk levels;

[0041] The third convergence condition is that the number of deliberative interaction rounds equals the preset round threshold.

[0042] When the first convergence condition, the second convergence condition, or the third convergence condition is met, the preset convergence condition is achieved.

[0043] Optionally, a fusion probability distribution is generated based on the posterior probability distributions of the three types of agents to represent the decision result, including:

[0044] The argumentation quality of the conservative and radical agents in a multi-round deliberative interaction process is evaluated, and conservative quality scores and radical quality scores are generated.

[0045] The fusion weight of the neutral agent is determined based on the ratio of the total divergence of the last deliberative interaction to the total divergence of the first deliberative interaction.

[0046] Based on the conservative and aggressive quality scores, the remaining weights, excluding the fusion weights allocated to the neutral agent, are dynamically allocated to the conservative and aggressive agents respectively.

[0047] Based on the fusion weights of the three types of agents, the posterior probability distributions of the three types of agents are weighted linearly pooled to obtain the fusion probability distribution.

[0048] Optionally, based on the total divergence, the decision result is output, including:

[0049] If the total degree of divergence is less than the second preset threshold of total divergence, then the neutral viewpoint mode is adopted and the decision result is output.

[0050] If the total divergence degree is greater than or equal to the second preset threshold of the total divergence degree, but less than the third preset threshold of the total divergence degree, then the decision result is output based on the weighted fusion mode.

[0051] If the total divergence is greater than or equal to the third preset threshold of the total divergence, then the decision result is output based on the high divergence marking mode.

[0052] Optionally, based on the fusion probability distribution, the decision result is output, including:

[0053] The risk level with the highest probability value is selected from the fusion probability distribution as the point decision result; the probability value corresponding to the point decision result in the fusion probability distribution is selected as the decision confidence level, and the decision result is output.

[0054] Alternatively, calculate the Shannon entropy of the fusion probability distribution; if the Shannon entropy is less than a first preset comparison threshold, output the decision result; if the Shannon entropy is greater than or equal to the first preset comparison threshold, but less than a second preset comparison threshold, output the decision result and add a prompt message; if the Shannon entropy is greater than or equal to the second preset comparison threshold, trigger the manual review process.

[0055] Corresponding to the aforementioned decision-making method based on multi-agent deliberation, the present invention provides a decision-making system based on multi-agent deliberation, comprising:

[0056] The agent construction module is used to build conservative agents, radical agents, and neutral agents, and to configure a corresponding prior probability distribution for each agent.

[0057] The deliberative interaction module is used to schedule three types of intelligent agents to conduct multiple rounds of deliberative interactions based on evidence data of the object to be decided; it includes:

[0058] The structured argument generation unit is used to call conservative and radical agents in parallel to generate structured arguments, including the direction of influence and weights, based on the evidence data.

[0059] The Bayesian update unit is used to perform Bayesian updates on the prior probability distributions of the conservative agent and the radical agent based on their structured arguments, so as to obtain their respective posterior probability distributions.

[0060] The question list generation unit is used to call the neutral agent, analyze the structured arguments corresponding to the conservative agent and the radical agent, and generate the conservative question list and the radical question list.

[0061] The argument correction unit is used to drive the conservative agent and the radical agent to correct their respective structured arguments based on the conservative question list and the radical question list.

[0062] The computing unit is used to recalculate and update the posterior probability distributions of the conservative and radical agents based on their modified structured arguments.

[0063] The judgment module is used to determine, after each round of deliberative interaction, whether the preset convergence condition has been met based on the posterior probability distributions of the conservative agent, the radical agent, and the neutral agent; if so, it performs fusion processing based on the posterior probability distributions of the three types of agents to generate a fusion probability distribution to represent the decision result; if not, it calls the deliberative interaction module to continue the next round of deliberative interaction.

[0064] The output module is used to output the decision result based on the fusion probability distribution or the total divergence degree.

[0065] Compared with the prior art, the present invention has the following beneficial effects:

[0066] (1) By constructing a multi-round adversarial deliberation architecture with three types of intelligent agents—conservative, radical, and neutral—the decision-making process is decomposed into a complete sequence of evidence extraction, argument generation, probability update, questioning correction, and convergence determination, making the decision-making path traceable throughout and solving the black box problem of opaque and unauditable decision-making processes of single AI models. Constructed arguments are generated in parallel by conservative and radical intelligent agents with opposing positions, and the posterior probability distribution is iteratively updated based on the Bayesian method, forcing multi-dimensional quantitative game on contradictory evidence and eliminating the inherent perspective bias of single models. Through probability distribution divergence measurement and convergence determination mechanism, it automatically decides whether to terminate the debate and merges the three parties' posterior probability distributions into the final decision probability distribution, realizing the systematic quantitative fusion of contradictory evidence and making up for the lack of a rigorous fusion mechanism in existing technologies. The entire process is executed automatically, which significantly improves decision-making efficiency while maintaining expert-level adversarial deliberation logic and meets the needs of large-scale and high-frequency decision-making scenarios.

[0067] (2) By loading predefined differentiated role background prompts for each type of agent, conservative agents, radical agents and neutral agents have a position-specific behavioral framework before entering the deliberation; dynamically retrieve and configure the corresponding Bayesian prior probability distribution according to the macro business strategy to ensure that the initial risk view of each agent is accurately aligned with the bank's current risk preference; bind the role background prompts with the prior probability distribution to complete the instantiation, so that the three types of agents can maintain the consistency of role behavior and have a mathematical initial position, laying a computable foundation for subsequent adversarial debate.

[0068] (3) By forcing the agent to perform a three-step process of classification, grading and output adjustment factors for each piece of evidence, the agent queries the preset evidence classification and grading standard library to obtain the benchmark weight range that matches the role position, avoiding the agent's subjective weighting in the open space and ensuring the comparability of cross decision weight standards; combined with the adjustment factors, the final weight is calculated and structured arguments are generated, so that the process of quantifying the degree of influence of evidence follows the rule library of solidified expert experience, while retaining the space for contextual fine-tuning.

[0069] (4) By calculating and ranking the weight differences between conservative and radical agents on the same evidence, verifiable factual assertions are extracted first for highly divergent evidence, so that the questioning resources are concentrated on the core points of contention in the debate and the verification efficiency is improved; the factual assertions are submitted to external verification and the verification results are obtained to ensure that the questioning is based on objective facts rather than model illusions; a list of adjudicative questions is constructed based on the verified facts, which forces both agents to revise their arguments on the basis of unavoidable facts, thereby enhancing the binding force and fact anchoring of the questioning process.

[0070] (5) By calculating the total degree of divergence, the mathematical quantification of the divergence of the three parties' viewpoints is realized; the divergence threshold, viewpoint proximity, and maximum rounds are set as triple convergence conditions, so that the system can terminate the debate in a timely manner when the viewpoints are fully converged, or forcibly terminate when the calculation limit is reached, so as to avoid unnecessary consumption of resources and ensure that the deliberation process has both convergence quality and execution efficiency.

[0071] (6) By scoring the argumentation quality of conservative and radical agents in multiple rounds of deliberation, the remaining weights are dynamically and asymmetrically allocated based on the difference in the quality of rebuttal, so that the party with more solid argumentation and more reasonable revision can gain more "discourse power" in the final decision. At the same time, the neutral weights are adjusted according to the ratio of the final total degree of disagreement to the initial total degree of disagreement, so as to ensure that the dominant position of the factual basis in the decision-making is not weakened by the performance of the rebuttal, and to take into account both the reward for argumentation quality and the anchoring of objective facts.

[0072] (7) By comparing the total degree of disagreement with the second and third preset thresholds, the system automatically selects the neutral viewpoint adoption mode, the weighted fusion mode, or the high disagreement marking mode, so that the decision-making strategy and the degree of viewpoint convergence are adaptively matched. When the disagreement is extremely small, the most objective neutral judgment is directly adopted; when the disagreement is moderate, the weighted fusion is performed; when the disagreement is huge, the system actively marks and triggers manual intervention, forming a hierarchical risk response strategy to improve the safety and robustness of decision-making in extremely complex or conflicting evidence scenarios.

[0073] (8) By selecting the risk level with the highest probability value from the fusion probability distribution as the point decision result and using the probability value as the decision confidence level, the system provides business personnel with an intuitive and quantifiable decision reference; calculates the Shannon entropy value of the fusion probability distribution and performs uncertainty classification based on the first and second preset comparison thresholds, thereby realizing the automatic self-reliability assessment of the decision system; automatically triggers the manual review process in response to high uncertainty, preventing erroneous automated decisions in scenarios with highly contradictory evidence or insufficient information, and constructing a dynamic linkage mechanism between decision uncertainty and manual intervention. Attached Figure Description

[0074] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this invention, illustrate exemplary embodiments of the invention and are used to explain the invention, but do not constitute an undue limitation of the invention. In the drawings:

[0075] Figure 1 This is a simplified flowchart of an embodiment of the decision-making method based on multi-agent deliberation of the present invention;

[0076] Figure 2 This is a framework diagram of an embodiment of the decision-making system based on multi-agent deliberation of the present invention. Detailed Implementation

[0077] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0078] like Figure 1 As shown, a decision-making method based on multi-agent deliberation according to the present invention includes the following steps:

[0079] S100. Establish conservative, radical, and neutral agents, and configure corresponding prior probability distributions for each agent.

[0080] S200: Based on the evidence data of the decision-making object, three types of intelligent agents are scheduled to conduct multiple rounds of deliberative interaction; each round of deliberative interaction includes:

[0081] S210. Conservative and radical agents are invoked in parallel to generate structured arguments, including the direction of influence and weights, based on evidence data.

[0082] S220. Based on the structured argument of conservative and radical intelligent agents, Bayesian updates are performed on their respective prior probability distributions to obtain their respective posterior probability distributions.

[0083] S230. Call upon the neutral agent, analyze the structured arguments corresponding to the conservative and radical agents, and generate a conservative question list and a radical question list.

[0084] S240. Based on the conservative and radical question lists, drive the conservative and radical agents to revise their respective structured arguments.

[0085] S250. Based on the modified structured arguments of conservative and radical agents, recalculate and update their respective posterior probability distributions.

[0086] S300. After each round of deliberative interaction, based on the posterior probability distribution of the conservative agent, the posterior probability distribution of the radical agent, and the posterior probability distribution of the neutral agent, determine whether the preset convergence condition has been met.

[0087] If yes, perform fusion processing based on the posterior probability distributions of the three types of agents to generate a fusion probability distribution to represent the decision result; if no, continue to the next round of deliberative interaction.

[0088] S400 outputs the decision result based on the fusion probability distribution or the total divergence degree.

[0089] This invention constructs a multi-round adversarial deliberation architecture with conservative, radical, and neutral agents. It decomposes the decision-making process into a complete sequence of evidence extraction, argument generation, probability updating, questioning and correction, and convergence determination, making the entire decision-making path traceable and solving the black-box problem of opaque and unauditable decision-making processes in single AI models. It utilizes the parallel generation of structured arguments by conservative and radical agents with opposing viewpoints, and iteratively updates the posterior probability distribution based on Bayesian methods, forcing multi-dimensional quantitative game analysis of contradictory evidence and eliminating the inherent perspective bias of single models. Through probability distribution divergence measurement and convergence determination mechanisms, it automatically decides whether to terminate the debate and merges the three parties' posterior probability distributions into the final decision probability distribution, achieving systematic quantitative fusion of contradictory evidence and overcoming the lack of rigorous fusion mechanisms in existing technologies. The entire process is automated, significantly improving decision-making efficiency while maintaining expert-level adversarial deliberation logic, meeting the needs of large-scale, high-frequency decision-making scenarios.

[0090] In this embodiment, S100 includes:

[0091] S110. Load predefined role background prompts for each type of intelligent agent;

[0092] S120. Based on the macro business strategy associated with the current decision-making task, retrieve the preset prior probability distribution library, and configure prior probability distributions representing the initial risk views of conservative agents and radical agents respectively, and configure prior probability distributions representing objective benchmarks for neutral agents.

[0093] S130. Bind the background prompts of the characters to the corresponding prior probability distributions to complete the creation and initialization of the three types of intelligent agents.

[0094] Preferably, the character background prompts in S110 are stored in a character prompt library, which stores a series of modular, finely designed character background prompt templates that define the character background and behavioral framework; the character background is as follows:

[0095] Conservative intelligent agent role: risk averse, emphasizing prevention by "monitoring negative deviations" and "building a chain of evidence for risk", with capital preservation as the core.

[0096] Radical intelligent agent role: risk driver, emphasizing the discovery of long-term value by "identifying leading indicators" and "reconstructing risks into growth-related pains", with the core being the construction of strategic cooperation between banks and enterprises.

[0097] Neutral intelligent agent role: Emphasizes penetrating appearances by verifying facts and constructing an integrated narrative, with the core being the development of actionable risk management solutions.

[0098] While loading predefined role background prompts for each type of intelligent agent, basic data such as customer overview information, industry background, bank credit status, and credit scheme are filled into the placeholders of the role background prompt template to form a complete and readily available context.

[0099] Behavioral framework: Used to enforce the workflow of intelligent agents, usually using thought chains (CoT) or more complex structured frameworks.

[0100] The preset prior probability distribution library mentioned in S120 stores Bayesian prior probability distributions used to quantify the initial stance of the agent. This is a discrete random variable representing a "risk level". The library can be initialized based on samples under different macroeconomic business strategies (such as "expansion period" and "contraction period"). The library stores the initial Bayesian prior probability distribution set corresponding to different macroeconomic business strategies. These distributions are mathematical expressions of the bank's risk preferences under different operating cycles. Specific examples are as follows:

[0101] Prior probability distribution of conservative (agent) This approach utilizes the risk distribution of customers who have "triggered risk warnings." These customers are already "marked" by the bank's risk identification system, and their risk distribution naturally leans towards high-risk areas. The conservative (agent) perspective is built upon statistical data from all customers who have historically generated risk, thus supporting its conservative stance.

[0102] Prior probability distribution of the radical (agent) The risk distribution of "new customers in the past year" is used. New customers have typically undergone the latest onboarding approval process and represent the type of customers the bank is currently most willing to accept and develop. Their risk distribution may skew towards the low-risk area. Using this as a starting point for aggressive investors aligns with their future-oriented and growth-embracing role.

[0103] Prior probability distribution of neutral (agent) The risk distribution is based on the bank's total existing customer base. This risk distribution is the most accurate and comprehensive reflection of the bank's current risk situation. Using this as the benchmark for the neutral agent ensures its "impartial" and "representative" role. The benchmark for the neutral agent is the bank's own average level.

[0104] It should be noted that the prior probability distribution refers to the probability distribution of the five risk levels; in a computer system, the prior probability distribution is represented as a vector, that is, a Vector type data structure, specifically displayed as [0.1,0.5,0.2,0.1,0.1].

[0105] The macro-level business strategy specifically includes an expansion phase; specific examples are as follows:

[0106] 1) Expansion period: Banks increase their risk appetite and encourage them to take on appropriate risks in exchange for growth.

[0107] Prior probability distribution of conservative (agent) :

[0108] Benchmark: Still using the distribution of customers who have "triggered risk warnings".

[0109] Adjustment: A second screening is performed on this group, selecting only a subset of customers who triggered Level 1 and Level 2 warning signals to calculate the distribution, ignoring customers who triggered Level 3 warning signals. Level 1, Level 2, and Level 3 warning signals refer to the bank's classification of warning signals for customers, representing the degree of importance of the warning. For example, "bankruptcy cases" should be a Level 1 warning signal, while "stock price fluctuations" are a Level 3 warning signal. Level 1 warning signals are signals requiring close attention, Level 2 warning signals are signals requiring attention, and Level 3 warning signals are warning signals; the classification of customer warning signals is not strongly correlated with the level of customer risk.

[0110] The conservative (agent) stance remains prudent, but its focus is directed toward truly significant risks, avoiding overreaction to minor issues, thus adapting to the business needs of the expansion phase.

[0111] Prior probability distribution of the radical (agent) :

[0112] Benchmark: The latest risk level distribution is determined manually from "new customers in the past year".

[0113] Adjustment: Based on this, a weighted shift is applied to the distribution. The probability weights of "blue" and "green" levels can be appropriately increased, while the weights of "yellow" and above can be decreased. "Red, orange, yellow, blue, and green" represent the risk exposure levels of enterprise customers from high to low risk, with red indicating high risk and green indicating low risk.

[0114] The optimistic stance of the aggressive (agent) type is reinforced, making it more proactive in seeking out "top students" from high-quality new customer groups, which aligns with the bank's strategic intention to encourage growth.

[0115] Prior probability distribution of neutral (agent) :

[0116] No adjustments are needed, or only a slight shift towards the low-risk zone is required. The neutral (agent) represents global stability and should not be overly optimistic during the expansion phase; it should maintain its role as an objective benchmark.

[0117] 2) Contraction period: Banks tighten their risk appetite and prioritize risk preservation over business growth.

[0118] Prior probability distribution of conservative (agent) :

[0119] Baseline: The distribution of customers with "triggered risk warnings" is used.

[0120] Adjustment: Expand the sample scope to include all customers who have triggered any level of alert (including the mildest warning class) in the calculation sample.

[0121] The conservative (agent) is on high alert, and any warning is seen as a potential risk signal, which is in line with the risk management tone during the contraction period.

[0122] Prior probability distribution of the radical (agent) :

[0123] Benchmark: The potentially overly optimistic "new customer" distribution will no longer be used. Instead, a distribution of "all existing customers of the bank" or a distribution of high-quality existing customers who have completed credit re-examination within the past three years and have not experienced any risk events will be adopted.

[0124] The radical (agent) stance has been forced back to a more realistic and cautious starting point. It still needs to seek opportunities, but these must be based on a more battle-tested customer base, and its optimism needs to be supported by stronger evidence.

[0125] Prior probability distribution of neutral (agent) :

[0126] Benchmark: The distribution of "all existing customers of the bank" is used.

[0127] Adjustment: A weighted shift is applied to the baseline distribution towards higher-risk areas. This can systematically transfer some of the probability quality from "blue" to "yellow" and some from "yellow" to "orange" to reflect the overall tightening of rating standards across the bank.

[0128] The objective standards of neutral (intelligent agents) become more stringent overall, causing the convergence point of debate to naturally shift towards a more deliberate area.

[0129] In this embodiment, S130 initialization refers to: initializing other state parameters of the agent, such as round (number of rounds of deliberation interaction) = 0, and argument_history (deliberation interaction history) being initially empty.

[0130] This invention loads predefined, differentiated role background prompts for each type of agent, enabling conservative, radical, and neutral agents to possess a position-specific behavioral framework before entering the deliberation phase. It dynamically retrieves and configures corresponding Bayesian prior probability distributions based on macro-business strategies, ensuring that each agent's initial risk viewpoint is precisely aligned with the bank's current risk appetite. By binding role background prompts with prior probability distributions for instantiation, the three types of agents maintain consistent role behavior while possessing a mathematically based initial stance, laying a computable foundation for subsequent adversarial debates.

[0131] In this embodiment, S210 includes:

[0132] S211. Conservative and radical agents are invoked in parallel to output the corresponding classification results, severity level and adjustment factor for each piece of evidence data according to their respective roles and positions.

[0133] S212. Based on the classification results, query the preset evidence classification and grading standard library to obtain the benchmark weight range;

[0134] S213. Combining the adjustment factor and the benchmark weight range, calculate the weight values ​​of this evidence data for the conservative agent and the radical agent, respectively.

[0135] S214. Based on the weight values ​​and influence directions calculated by the conservative and radical agents for all evidence data, respectively, generate a structured argument including the influence direction and weight.

[0136] This invention forces the intelligent agent to perform a three-step process of classification, grading, and outputting adjustment factors for each piece of evidence. It queries a preset evidence classification and grading standard library to obtain a benchmark weight range that matches the role's position, avoiding subjective weighting by the intelligent agent in an open space and ensuring comparability across decision weight standards. The final weight is calculated by combining the adjustment factors and generating structured arguments, so that the process of quantifying the degree of influence of evidence follows the rule library solidified by expert experience while retaining space for contextual fine-tuning.

[0137] Preferably, a risk control expert team and historical data statistics are used to distinguish between conservative and aggressive agent perspectives. For the same "financial indicator - profitability - moderate deterioration", the standard weight range for the conservative (agent) might be 7-9, while for the aggressive (agent) it is 4-6. An evidence classification and grading standard library is established through manual annotation, as shown in Table 1.

[0138] Table 1 Examples of Evidence Classification and Grading Standards

[0139]

[0140] In this embodiment, for any intelligent agent ,in C stands for conservative (agent), and A stands for radical (agent). Their roles in the first... During the round of deliberations, from the prior probability distribution Update to posterior probability distribution The specific calculation steps are as follows (i.e., S220 includes):

[0141] Position score vector calculation: For the first position in the deliberation... Evidence data The agent first generates its weights. and direction The system covers all risk levels (assuming there are). Calculate a position score vector (at each level). , its first The formula for calculating each element is:

[0142] ;in, , indicating evidence data For intelligent agents In other words, does it indicate an increase in risk (+1) or an increase in opportunity (-1)? It is a fixed viewpoint vector. This maps different risk levels to a numerical scale; This is a position steepness hyperparameter, with a value range of [0.5, 5.0]. In actual operation, it is used to adjust the strength of the influence of evidence data according to the business scenario.

[0143] Target probability distribution: The system will assign a position score vector. The evidence data is generated by applying the Softmax function as input. Corresponding target probability distribution The formula is:

[0144] This distribution This represents considering only the evidence data. In this case, the idealized risk view of agent i.

[0145] The consensus target distribution for this round is calculated as follows: The system performs a weighted average of the target probability distributions generated from all independent evidence data to summarize the comprehensive opinions from all evidence data in this round, forming a weighted average target distribution. The formula is:

[0146] ;in, This is the weighted set of all evidence data in this round; This is the target distribution set of all evidence data in this round; this distribution represents the "consensus" risk view that all evidence data in this round points to, with the view of important evidence accounting for a larger proportion.

[0147] Bayesian Update: Employing the standard discrete Bayesian inference formula, the system updates the prior probability distribution of agent i from the previous round. Considered as the prior probability distribution of this round, the newly formed weighted average target distribution Treating it as a likelihood function, calculate the final posterior probability distribution for this round. ;include:

[0148] Calculate the nonnormalized posterior probability vector For sets Each specific level in The specific formula is as follows:

[0149] ;in, This is a smoothing factor used to prevent the zero-probability trap and ensure the numerical stability of subsequent calculations. Its value is preferably a very small positive number. ; To update the intensity hyperparameter, the preferred value range is... .when When this occurs, it indicates that the impact of the evidence is amplified, accelerating the convergence of viewpoints; when At that time, it indicates a conservative attitude towards the evidence, slowing down the pace of updates.

[0150] right Normalization is performed to obtain the posterior probability distribution. To ensure that the sum of probabilities equals 1, the specific formula is as follows:

[0151] Through the above for each The point-by-point calculation ultimately yields the posterior probability distribution. J is the set of j.

[0152] In this embodiment, S230 includes:

[0153] S231. Forward the structured arguments generated by the conservative agent to the radical agent to generate a structured list of conservative rebuttals; forward the structured arguments generated by the radical agent to the conservative agent to generate a structured list of radical rebuttals; wherein, the task of the radical agent is: to analyze the structured arguments of the conservative agent, identify its "overly pessimistic interpretation", "ignoring positive signals" or "failing to see potential opportunities", and generate a structured list of radical rebuttals; the task of the conservative agent is: to analyze the structured arguments of the radical agent, identify its "overly optimistic assumptions", "logical leaps" or "ignoring potential risks", and generate a list of conservative rebuttals;

[0154] S232. The radical agent obtains a list of conservative objections and responds to each one in a radical manner; the conservative agent obtains a list of radical objections and responds to each one in a conservative manner.

[0155] S233. Invoke a neutral agent to quantify the degree of disagreement between conservative and radical agents regarding the same evidence data; the calculation formula is as follows:

[0156] ;in, For conservative intelligent agents regarding evidence data The given weight values, For radical intelligent agents on evidence data The given weight values;

[0157] S234. Sort the evidence according to the degree of disagreement, and extract verifiable factual assertions from the highly disagreed evidence; for example, "the company's gross profit margin has decreased by 20%" or "this patent has 3 lost cases"; preferably, the highly disagreed evidence is evidence data that is greater than the preset disagreement threshold, and the preset disagreement threshold is set according to the actual needs of the application scenario.

[0158] S235. Verify the factual assertions to obtain the verification results and the verified facts; preferably, the factual assertions are formatted as query requests and submitted to the RAG knowledge base, database API, or external tools for verification.

[0159] S236. Based on highly divergent evidence and its verification results, construct conservative inquiry lists and radical inquiry lists respectively, requiring conservative agents and radical agents to respond based on the verified facts.

[0160] It should be noted that the factual assertion described in S234 refers to the identification, through Natural Language Processing (NLP), of whether a structured argument contains named entities (e.g., specific amounts, case numbers of the litigation cases) or structured fields (e.g., names of financial indicators) for highly divergent evidence. If the argument contains named entities that can be mapped to fields in external databases (such as business, judicial, and credit databases), it is considered a verifiable factual assertion; if it only contains subjective adjectives (e.g., "business is doing well"), it is not a verifiable factual assertion.

[0161] This invention calculates and ranks the weight differences between conservative and radical agents on the same evidence, prioritizing the extraction of verifiable factual assertions from highly divergent evidence. This concentrates inquiry resources on the core points of contention in the debate, improving verification efficiency. The factual assertions are submitted for external verification, and the verification results are obtained, ensuring that the inquiry is based on objective facts rather than model illusions. Based on the verified facts, a list of adjudicative inquiries is constructed, forcing both agents to revise their arguments based on unavoidable facts, thus enhancing the binding force and factual anchoring of the inquiry process.

[0162] Furthermore, by calculating and ranking the weight differences between conservative and radical agents on the same evidence, verifiable factual assertions are extracted first for highly divergent evidence, concentrating inquiry resources on the core points of contention in the debate and improving verification efficiency. The factual assertions are submitted for external verification and the verification results are obtained, ensuring that the inquiry is based on objective facts rather than model illusions. Based on the verified facts, a list of adjudicative inquiries is constructed, forcing both agents to revise their arguments on the basis of unavoidable facts, thereby enhancing the binding force and factual anchoring of the inquiry process.

[0163] In this embodiment, S240 includes:

[0164] S241. Send the conservative question list and the radical question list to the conservative agent and the radical agent respectively;

[0165] S242, Receive conservative and radical question response information generated by conservative and radical agents respectively; the conservative or radical question response information includes answers to the conservative or radical question list and instructions for modifying the respective structured arguments; preferably, the modification instructions include instructions to modify at least one of the interpretation text of the evidence data, the direction of influence, the weight, and the underlying logical chain;

[0166] S243. Based on the modification instructions, drive the conservative agent and the radical agent to update their respective structured arguments to obtain the revised structured arguments.

[0167] S243. Based on the revised structured arguments, recalculate their respective posterior probability distributions.

[0168] In this embodiment, taking a conservative agent as an example, the formula for recalculating the respective posterior probability distribution in S250 is as follows:

[0169] ;in, It is the posterior probability distribution of the conservative agent in the previous round (or the prior probability distribution if the current round is the second round). This represents the updated posterior probability distribution of a conservative agent after the current round. The radical agent... It is also generated in the same way.

[0170] In this embodiment, the preset convergence conditions include:

[0171] First convergence condition: The total divergence is less than a first preset threshold for the total divergence; where the total divergence is the sum of the first KL divergence between the posterior probability distributions of the conservative agent and the neutral agent, and the second KL divergence between the posterior probability distributions of the aggressive agent and the neutral agent; the total divergence... The calculation formula is as follows:

[0172] ; The distribution can be calculated by a neutral agent based on all verified facts and using risk-neutral weighting standards in the classification and grading standard library, serving as the "center of gravity" of objective facts; preferably, the first preset threshold for total divergence is 0.1. When the sum of the KL divergences of the three distributions is less than 0.1, it mathematically indicates that the posterior probability distributions of each agent are very close, that is, a consensus has been reached.

[0173] Second convergence condition: The posterior probability distributions of conservative agents and radical agents are concentrated on the same or adjacent one or two risk levels.

[0174] The third convergence condition is that the number of deliberation-style interaction rounds equals the preset round threshold. Preferably, the preset round threshold is 3 rounds. In actual risk control business, if convergence is not achieved after 3 rounds of debate, continued interaction usually leads to semantic repetition or infinite loops, and should be forcibly terminated at this time.

[0175] When the first convergence condition, the second convergence condition, or the third convergence condition is met, the preset convergence condition is achieved.

[0176] Preferably, when the preset convergence condition is met, the results of the entire debate process will be integrated and output, including:

[0177] Final Set of Arguments by the Three Parties: The structured arguments finally adopted by the three parties after multiple rounds of revisions;

[0178] The final posterior probability distribution of the three parties, and the total divergence degree of the final round;

[0179] A complete history of arguments: including arguments, questions, responses, and revisions for each round, providing a complete and transparent record for human auditing and decision tracing.

[0180] This invention achieves mathematical quantification of the differences in opinions among the three parties by calculating the total degree of divergence; it sets a triple convergence condition of divergence threshold, opinion proximity, and maximum number of rounds, so that the system can terminate the debate in a timely manner when opinions are fully converged, or forcibly terminate it when the calculation limit is reached, avoiding unnecessary consumption of resources and ensuring that the deliberation process has both convergence quality and execution efficiency.

[0181] In this embodiment, a fusion process is performed based on the posterior probability distributions of the three types of agents to generate a fusion probability distribution used to characterize the decision result, including:

[0182] The system evaluates the argument quality of conservative and radical agents during multi-round deliberative interactions and generates conservative and radical quality scores. Preferably, an agent argument quality arbitrator responsible for the evaluation can be invoked to receive a complete list of questions and response logs, and then score the overall argument quality of the conservative and radical agents.

[0183] The fusion weights of neutral agents are determined based on the ratio of the total divergence of the last deliberative interaction to the total divergence of the first deliberative interaction. The calculation formula is:

[0184] ;in, The minimum weight for neutrality is set (e.g., 0.4). The preset base weight (e.g., 0.5). The total degree of disagreement in the final deliberative interaction. β represents the total degree of disagreement in the first deliberative interaction, and β is the influence coefficient of the degree of disagreement.

[0185] Based on conservative and aggressive quality scores, the remaining weights (1-) after the fusion weights assigned to the neutral agent are... In this process, the fusion weights for conservative and aggressive agents are dynamically assigned; the formula is as follows:

[0186] ;

[0187] ;

[0188] in, For conservative agents, the fusion weights For the fusion weights of radical agents, , This is a conservative quality rating. This is a radical quality rating;

[0189] Based on the fusion weights of the three types of agents, a weighted linear pooling method is applied to the (final) posterior probability distributions of the three types of agents to obtain the fusion probability distribution. The formula is as follows:

[0190] ;in, For the (final) posterior probability distribution of a neutral agent, For the (final) posterior probability distribution of a conservative agent, For the (final) posterior probability distribution of the radical agent.

[0191] This invention scores the argumentation quality of conservative and radical agents in multiple rounds of deliberation, and dynamically allocates the remaining weights asymmetrically based on the differences in the quality of rebuttal. This allows the party with more solid arguments and more reasonable revisions to gain greater "voice" in the final decision. At the same time, the neutral weights are adjusted according to the ratio of the final total degree of disagreement to the initial total degree of disagreement, ensuring that the dominant position of the factual basis in the decision-making process is not weakened by the performance of the rebuttal, thus balancing the reward for argumentation quality with the anchoring of objective facts.

[0192] In this embodiment, the decision result is output based on the total divergence degree, including:

[0193] If the total degree of divergence is less than the second preset threshold for total degree of divergence, the neutral viewpoint model is adopted and the decision result is output. Preferably, the second preset threshold for total degree of divergence is 0.05. When the degree of divergence is lower than this value, it can be considered that the views of all parties have reached a substantial consensus on the basis of facts.

[0194] If the total divergence is greater than or equal to the second preset threshold of total divergence, but less than the third preset threshold of total divergence, then the decision result is output based on the weighted fusion mode; preferably, the third preset threshold of total divergence is 0.3.

[0195] If the total degree of disagreement is greater than or equal to the third preset threshold of total disagreement, the decision result is output based on the high disagreement labeling pattern. When the degree of disagreement is higher than this value, it indicates that even after multiple rounds of debate, there are still huge differences among the parties on key judgments. There may be complex situations or evidence conflicts that the model does not understand, and manual review is required.

[0196] This invention compares the total degree of disagreement with a second preset threshold and a third preset threshold, and automatically selects a neutral viewpoint adoption mode, a weighted fusion mode, or a high-disagreement labeling mode, so that the decision-making strategy and the degree of viewpoint convergence are adaptively matched. When the disagreement is extremely small, the most objective neutral judgment is directly adopted; when the disagreement is moderate, weighted fusion is performed; when the disagreement is huge, it is actively labeled and human intervention is triggered, forming a hierarchical risk response strategy, which improves the safety and robustness of decision-making in extremely complex or conflicting evidence scenarios.

[0197] In this embodiment, based on the fusion probability distribution, the decision result is output, including:

[0198] The risk level with the highest probability value is selected as the point decision result from the fusion probability distribution; the probability value corresponding to the point decision result in the fusion probability distribution is selected as the decision confidence level, and the decision result is output; the formula is as follows:

[0199] ;

[0200] in, These correspond to {green, blue, yellow, orange, and red} respectively. Represents a set The first in Each risk level is considered; the confidence level of the final decision is then determined by the final risk level decision. The corresponding probability value;

[0201] Alternatively, calculate the Shannon entropy of the fusion probability distribution; the formula is:

[0202] ;

[0203] If the Shannon entropy is less than the first preset comparison threshold, the decision result is output; preferably, the first preset comparison threshold is 1. If the Shannon entropy is less than the first preset comparison threshold, it means that the probability distribution is highly concentrated on a certain decision, indicating that the system decision is clear and reliable, and the decision result can be output directly.

[0204] If the Shannon entropy is greater than or equal to the first preset comparison threshold, but less than the second preset comparison threshold, the decision result is output with an additional prompt message; this indicates that the decision has some ambiguity, but there is still a clear dominant option, and the system outputs a normal decision, but can add a prompt message; preferably, the second preset comparison threshold is 1.5;

[0205] If the Shannon entropy is greater than or equal to the second preset comparison threshold, the manual review process is triggered. When the Shannon entropy is greater than or equal to the second preset comparison threshold, it indicates that the probability distribution is very dispersed and there is no decision option with an absolute advantage. This indicates that the input evidence itself may be highly contradictory or lack information. The system triggers the manual review process to prevent erroneous automated decisions in ambiguous situations.

[0206] This invention selects the risk level with the highest probability value from the fusion probability distribution as the point decision result and uses this probability value as the decision confidence level, providing business personnel with an intuitive and quantifiable decision reference; it calculates the Shannon entropy value of the fusion probability distribution and performs uncertainty classification based on the first and second preset comparison thresholds, enabling the decision system to automatically assess its own reliability; it automatically triggers a manual review process in response to high uncertainty, preventing erroneous automated decisions in scenarios with highly contradictory evidence or insufficient information, and constructs a dynamic linkage mechanism between decision uncertainty and human intervention.

[0207] Preferably, a narrative generation agent is invoked to output the decision result; the narrative generation agent simulates the final decision-making process of a senior risk manager. It aims to integrate the quantified probabilities and arguments output by previous modules, which may still contain contradictions and disagreements, into a unified decision that aligns with business strategy and is actionable. Specifically, this includes:

[0208] Context splicing logic: When receiving information, this agent does not simply splice together all historical dialogues and data in a flat manner. It employs a structured "hierarchical context injection" method to organize information from different sources and of different natures into a hierarchical context structure with explicit instructions; its splicing logic follows these steps:

[0209] (a) Role and Principle Injection: At the top level of the context, the system injects instructions (i.e., System Prompt) that define its role, decision principles, and output format. This is the basic framework for all subsequent thinking of the agent.

[0210] (II) Structured Assembly of Core Decision-Making Elements: The key information required for decision-making is assembled in a structured JSON format, including:

[0211] 1) Customer static profile: customer's basic information, industry, development stage, etc.

[0212] 2) Objective decision-making benchmark: risk level assessment standards from banks retrieved from the knowledge base.

[0213] 3) Refined summary of the rebuttal process: A pre-processed summary that includes:

[0214] A) Core risk points of the conservative approach.

[0215] B) Aggressive core opportunities.

[0216] C) List of key facts verified by the neutral party.

[0217] D) Key disagreements that remain between the two sides after the debate.

[0218] 4) Quantitative decision reference: the final posterior probability distribution of the three parties output by the preceding module and the probability distribution after fusion.

[0219] (iii) Final task instruction injection: At the end of the context, give a clear final task instruction (i.e., User Prompt) that needs to be executed, such as: "Based on the above complete decision context, please make a final decision on the customer's risk level and provide detailed reasons and action plans."

[0220] like Figure 2 As shown, the present invention also provides a decision-making system based on multi-agent deliberation, which includes:

[0221] The agent construction module 10 is used to build conservative agents, radical agents and neutral agents, and to configure a corresponding prior probability distribution for each agent.

[0222] The deliberative interaction module 20 is used to schedule three types of intelligent agents to conduct multiple rounds of deliberative interaction based on the evidence data of the object to be decided; it includes:

[0223] The structured argument generation unit 21 is used to call conservative and radical agents in parallel to generate structured arguments, including the direction of influence and weights, based on evidence data.

[0224] Bayesian update unit 22 is used to perform Bayesian update on the prior probability distributions of conservative agents and radical agents based on their structured arguments, so as to obtain their respective posterior probability distributions.

[0225] The question list generation unit 23 is used to call a neutral agent, analyze the structured arguments corresponding to the conservative agent and the radical agent, and generate a conservative question list and a radical question list.

[0226] Argument revision unit 24 is used to drive conservative agents and radical agents to revise their respective structured arguments based on conservative question lists and radical question lists;

[0227] The computation unit 25 is used to recalculate and update the posterior probability distributions of the conservative and radical agents based on the modified structured arguments.

[0228] The judgment module 30 is used to determine, after each round of deliberative interaction, whether the preset convergence condition has been met based on the posterior probability distributions of the conservative agent, the radical agent, and the neutral agent; if so, it performs fusion processing based on the posterior probability distributions of the three types of agents to generate a fusion probability distribution to represent the decision result; if not, it calls the deliberative interaction module 20 to continue the next round of deliberative interaction.

[0229] Output module 40 is used to output decision results based on the fusion probability distribution or the total divergence degree.

[0230] It should be noted that the various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For system embodiments, since they are basically similar to method embodiments, the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments.

[0231] Furthermore, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0232] The foregoing description illustrates and describes preferred embodiments of the present invention. It should be understood that the present invention is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and environments, and can be altered within the scope of the inventive concept by means of the foregoing teachings or techniques or knowledge in related fields. Any modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims.

Claims

1. A decision-making method based on multi-agent deliberation, characterized in that, Includes the following steps: S100. Establish conservative, radical, and neutral agents, and configure corresponding prior probability distributions for each agent. S200: Based on the evidence data of the decision-making object, three types of intelligent agents are scheduled to conduct multiple rounds of deliberative interaction; each round of deliberative interaction includes: S210. The conservative and radical intelligent agents are invoked in parallel to generate structured arguments, including the direction of influence and weights, based on the evidence data. S220. Based on the structured argument of conservative and radical intelligent agents, Bayesian updates are performed on their respective prior probability distributions to obtain their respective posterior probability distributions. S230. Invoke the neutral agent, analyze the structured arguments corresponding to the conservative agent and the radical agent, and generate a conservative question list and a radical question list. S240. Based on the conservative question list and the radical question list, drive the conservative agent and the radical agent to revise their respective structured arguments; S250. Based on the modified structured arguments of the conservative and radical agents, recalculate and update their respective posterior probability distributions. S300. After each round of deliberative interaction, based on the posterior probability distribution of the conservative agent, the posterior probability distribution of the radical agent, and the posterior probability distribution of the neutral agent, determine whether the preset convergence condition has been met. If yes, the fusion process is performed based on the posterior probability distributions of the three types of agents to generate a fusion probability distribution used to characterize the decision result; if no, the next round of deliberative interaction continues. S400. Based on the fusion probability distribution or total divergence degree, output the decision result; S230 includes: S231. The structured arguments generated by the conservative agent are forwarded to the radical agent to generate a structured list of conservative rebuttals; the structured arguments generated by the radical agent are forwarded to the conservative agent to generate a structured list of radical rebuttals. S232. The radical agent obtains the list of conservative objections and responds radically to each one; the conservative agent obtains the list of radical objections and responds conservatively to each one. S233. Invoke the neutral agent to quantify the degree of disagreement between the conservative agent and the radical agent regarding the same evidence data; S234. Sort the opinions according to the degree of disagreement, and extract verifiable factual assertions from the highly disagreeing evidence. S235. Verify the factual assertion to obtain the verification result and the verified facts; S236. Based on the highly divergent evidence and its verification results, construct a conservative inquiry list and a radical inquiry list that require the conservative agent and the radical agent to respond based on the verified facts, respectively.

2. The decision-making method based on multi-agent deliberation according to claim 1, characterized in that, S100 includes: S110. Load predefined role background prompts for each type of intelligent agent; S120. Based on the macro business strategy associated with the current decision-making task, retrieve the preset prior probability distribution library, and configure prior probability distributions representing the initial risk views of conservative agents and radical agents respectively, and configure prior probability distributions representing objective benchmarks for neutral agents. S130. Bind the background prompt words of the role to the corresponding prior probability distribution to complete the creation and initialization of the three types of intelligent agents.

3. The decision-making method based on multi-agent deliberation according to claim 1, characterized in that, S210 includes: S211. Conservative and radical agents are invoked in parallel to output the corresponding classification results, severity level and adjustment factor for each piece of evidence data according to their respective roles and positions. S212. Based on the classification results, query the preset evidence classification and grading standard library to obtain the benchmark weight range; S213. Combining the adjustment factor and the benchmark weight range, calculate the weight values ​​of the evidence data for the conservative agent and the radical agent, respectively. S214. Based on the weight values ​​and influence directions calculated by the conservative agent and the radical agent for all evidence data respectively, generate a structured argument including the influence direction and weight.

4. The decision-making method based on multi-agent deliberation according to claim 1, characterized in that, S240 includes: S241. Send the conservative question list and the radical question list to the conservative agent and the radical agent respectively; S242. Receive conservative and radical question response information generated by the conservative agent and the radical agent; the conservative or radical question response information includes answers to the conservative or radical question list and instructions for modifying the respective structured arguments. S243. Based on the modification instruction, drive the conservative agent and the radical agent to update their respective structured arguments to obtain the revised structured arguments; S244. Based on the revised structured arguments, recalculate their respective posterior probability distributions.

5. The decision-making method based on multi-agent deliberation according to claim 1, characterized in that, The preset convergence conditions include: First convergence condition: The total divergence is less than a first preset threshold for the total divergence; wherein, the total divergence is the sum of the first KL divergence between the posterior probability distribution of the conservative agent and the posterior probability distribution of the neutral agent, and the second KL divergence between the posterior probability distribution of the radical agent and the posterior probability distribution of the neutral agent. Second convergence condition: The posterior probability distribution of the conservative agent and the posterior probability distribution of the radical agent are concentrated on the same or adjacent one or two risk levels; The third convergence condition is that the number of deliberative interaction rounds equals the preset round threshold. When the first convergence condition, the second convergence condition, or the third convergence condition is met, the preset convergence condition is achieved.

6. The decision-making method based on multi-agent deliberation according to claim 5, characterized in that, Based on the posterior probability distributions of the three types of agents, a fusion process is performed to generate a fusion probability distribution for representing the decision result, including: The argumentation quality of the conservative and radical agents in a multi-round deliberative interaction process is evaluated, and conservative quality scores and radical quality scores are generated. The fusion weight of the neutral agent is determined based on the ratio of the total divergence of the last deliberative interaction to the total divergence of the first deliberative interaction. Based on the conservative and aggressive quality scores, the remaining weights, excluding the fusion weights allocated to the neutral agent, are dynamically allocated to the conservative and aggressive agents respectively. Based on the fusion weights of the three types of agents, the posterior probability distributions of the three types of agents are weighted linearly pooled to obtain the fusion probability distribution.

7. The decision-making method based on multi-agent deliberation according to claim 5, characterized in that, Based on the total divergence, the decision results are output, including: If the total degree of divergence is less than the second preset threshold of total divergence, then the neutral viewpoint mode is adopted and the decision result is output. If the total divergence degree is greater than or equal to the second preset threshold of the total divergence degree, but less than the third preset threshold of the total divergence degree, then the decision result is output based on the weighted fusion mode. If the total divergence is greater than or equal to the third preset threshold of the total divergence, then the decision result is output based on the high divergence marking mode.

8. The decision-making method based on multi-agent deliberation according to claim 1, characterized in that, Based on the fusion probability distribution, the decision result is output, including: The risk level with the highest probability value is selected from the fusion probability distribution as the point decision result; the probability value corresponding to the point decision result in the fusion probability distribution is selected as the decision confidence level, and the decision result is output. Alternatively, calculate the Shannon entropy of the fusion probability distribution; if the Shannon entropy is less than a first preset comparison threshold, output the decision result; if the Shannon entropy is greater than or equal to the first preset comparison threshold, but less than a second preset comparison threshold, output the decision result and add a prompt message; if the Shannon entropy is greater than or equal to the second preset comparison threshold, trigger the manual review process.

9. A decision-making system based on multi-agent deliberation, characterized in that, include: The agent construction module is used to build conservative agents, radical agents, and neutral agents, and to configure a corresponding prior probability distribution for each agent. The deliberative interaction module is used to schedule three types of intelligent agents to conduct multiple rounds of deliberative interaction based on the evidence data of the object to be decided. It includes: The structured argument generation unit is used to call conservative and radical agents in parallel to generate structured arguments, including the direction of influence and weights, based on the evidence data. The Bayesian update unit is used to perform Bayesian updates on the prior probability distributions of the conservative agent and the radical agent based on their structured arguments, so as to obtain their respective posterior probability distributions. The question list generation unit is used to invoke the neutral agent, analyze the structured arguments corresponding to the conservative and radical agents, and generate a conservative question list and a radical question list, including: S231, forwarding the structured arguments generated by the conservative agent to the radical agent to generate a structured conservative rebuttal list; forwarding the structured arguments generated by the radical agent to the conservative agent to generate a structured radical rebuttal list; S232, the radical agent obtains the conservative rebuttal list and responds to each item in the radical rebuttal list; the conservative agent obtains the radical rebuttal list... S233. Rebuttal list, and provide conservative responses to each item; S234. Invoke the neutral agent to quantify the degree of disagreement between the conservative and radical agents regarding the same evidence data; S235. Sort the evidence based on the degree of disagreement, and extract verifiable factual assertions from highly disagreed evidence; S236. Verify the factual assertions to obtain verification results and verified facts; S237. Based on the highly disagreed evidence and its verification results, construct conservative and radical inquiry lists respectively, requiring the conservative and radical agents to respond based on the verified facts. The argument correction unit is used to drive the conservative agent and the radical agent to correct their respective structured arguments based on the conservative question list and the radical question list. The computing unit is used to recalculate and update the posterior probability distributions of the conservative and radical agents based on their modified structured arguments. The judgment module is used to determine, after each round of deliberative interaction, whether the preset convergence condition has been met based on the posterior probability distributions of the conservative agent, the radical agent, and the neutral agent; if so, it performs fusion processing based on the posterior probability distributions of the three types of agents to generate a fusion probability distribution to represent the decision result; if not, it calls the deliberative interaction module to continue the next round of deliberative interaction. The output module is used to output the decision result based on the fusion probability distribution or the total divergence degree.

Images