Decision-making method and device based on multiple agents, equipment and medium

Abstract

The invention provides a multi-agent-based decision-making method and device, equipment and a medium, and the method comprises the steps: collecting data of a plurality of data sources, and constructing a plurality of knowledge bases in different knowledge fields; aiming at different task requirements contained in the target social measure, adopting a corresponding decision model and a knowledge base of a corresponding knowledge field to construct a plurality of sub-agents; and controlling the plurality of sub-agents to formulate respective preliminary social measure implementation schemes for the target social measure, and performing debate aiming at the plurality of preliminary social measure implementation schemes to obtain a final social measure implementation scheme. According to the method, decision making is carried out by integrating the multiple sub-agents representing different perspectives, multi-field and multi-dimensional insights can be formed, then social measure implementation schemes of the multiple sub-agents are fused through a debate mechanism of the multiple sub-agents, the agents can generate more comprehensive and diversified decision making results, and the decision making efficiency is improved. And complex and changeable decision problems can be effectively solved.

Description

Technical Field

[0001] The present invention relates to the field of multi-agent modeling technology, and in particular to a multi-agent based decision-making method, device, equipment and medium. Background Art

[0002] Multi-agent deduction shows great potential in simulating human behavior. Multi-agent systems can simulate complex social phenomena, such as social network interactions, individual and group behaviors, and financial market dynamics. By simulating behaviors in different situations, these systems can reveal the underlying mechanisms of social phenomena and provide a basis for human decision-making. Large language models have powerful knowledge representation and natural language generation capabilities, and can perform a certain degree of reasoning and learning, which makes it possible to build more realistic and credible humanoid agents.

[0003] In the past few years, a lot of work has been done to build agents based on large models. For example, researchers created 25 agents in a virtual town and observed their behavior and interactions. Agents can perform daily activities such as making breakfast, going to work, and creating art, as well as complex social interactions such as planning and attending parties, remembering and reflecting on past events, and making plans for the future. Each agent has a perception system that can capture dynamic changes in the surrounding environment, such as the behavior of other agents, time, and environmental changes. This information is stored in the agent's memory, including short-term memory and long-term memory. Short-term memory stores recent events and information, and long-term memory stores long-term experience and historical data. Through the feedback loop mechanism, the agent can retrieve relevant information from the memory to make decisions in the current situation.

[0004] The researchers proposed the RecAgent framework, which aims to study user behavior patterns by simulating user behavior in the recommendation system in order to improve the performance of the recommendation system. The agent construction part of the framework is divided into a portrait module, a memory module, and a behavior module. The portrait module describes the basic characteristics of the user, such as gender, age, personality, interests and other information; the memory module designs three types of memory: perceptual memory records the user's immediate perception and short-term reaction, short-term memory saves recent events and user behavior, and long-term memory stores the user's long-term preferences and historical behavior; the action module is responsible for simulating various user behaviors in the recommendation system, such as browsing recommended content, interacting with other users, and posting messages on social media. Generating these behaviors through a large language model can make the user behavior simulation closer to reality.

[0005] However, existing methods all use a single agent to model a goal, which has the following shortcomings: (1) Perspective limitations: When simulating complex real-world scenarios, single-agent models are often limited by their inherent perspectives. Although large models have powerful capabilities, they may also be biased or fail to fully consider the multi-domain, multi-dimensional factors and diverse perspectives involved in the problem. This single-perspective modeling method is difficult to accurately reproduce the multi-angle thinking and comprehensive decision-making process that humans show when facing complex problems, thus limiting the effectiveness and universality of the model in solving complex decision-making problems.

[0006] (2) Hallucinations and errors: When a single agent makes a decision, it is easily affected by the hallucinations and false assumptions inherent in the large model, that is, the agent generates seemingly reasonable but actually wrong conclusions without sufficient evidence. Such errors not only reduce the reliability of the agent in the deduction, but may also have a misleading impact on the decision.

[0007] (3) Poor explainability: The decision-making process of complex problems is characterized by highly nonlinear and invisible internal mechanisms, lacking transparency and being difficult to understand and explain. Summary of the invention

[0008] The present invention provides a multi-agent based decision-making method, device, equipment and medium to solve the problem in related technologies that a single agent is limited by its inherent perspective when simulating complex real-life scenarios, which limits the effectiveness and universality of the agent in solving complex decision-making problems.

[0009] The present invention provides a decision-making method based on multi-agent, comprising the following steps.

[0010] Collect data from multiple data sources and build multiple knowledge bases in different knowledge fields; According to the different task requirements contained in the target social measures, corresponding decision-making models and knowledge bases in corresponding knowledge fields are used to construct multiple sub-agents; Control multiple sub-agents to formulate their own preliminary social measure implementation plans for the target social measures, and debate on multiple preliminary social measure implementation plans to obtain the final social measure implementation plan.

[0011] According to a multi-agent decision-making method provided by the present invention, data from multiple data sources are collected to construct multiple knowledge bases in different knowledge fields, including: Connect to multiple data sources of different types; Divide multiple knowledge domains based on data stored in multiple different types of data sources; For each knowledge domain, the knowledge base is constructed using the data of the corresponding knowledge domain.

[0012] According to a multi-agent decision-making method provided by the present invention, for different task requirements and different knowledge fields included in the target social measures, corresponding decision-making models and knowledge bases of corresponding knowledge fields are adopted to construct multiple sub-agents, including: Extracting shared knowledge from multiple knowledge bases based on targeted social measures; For each task requirement, determine the decision model corresponding to the task requirement; For each task requirement, extract the perspective features, domain knowledge, and thinking methods required to meet the task requirements from the knowledge base of the corresponding knowledge domain, and generate independent knowledge corresponding to the task requirements; Based on shared knowledge, sub-agents are constructed for each task requirement using the decision model corresponding to the task requirement and the corresponding independent knowledge.

[0013] According to a multi-agent decision-making method provided by the present invention, multiple sub-agents are controlled to formulate respective preliminary social measure implementation plans for target social measures, and debate on multiple preliminary social measure implementation plans to obtain a final social measure implementation plan, including: Control multiple sub-agents to formulate preliminary social measures implementation plans for target social measures that meet the requirements of their respective tasks; Control each sub-agent to demonstrate the preliminary social measures implementation plan formulated by each sub-agent and obtain demonstration results. The demonstration includes clarifying the decision-making basis, reasoning process, expected effects and potential risks faced by the preliminary social measures implementation plan; For each sub-agent, the control sub-agent refutes the preliminary social measure implementation plan formulated by other sub-agents to obtain a refutation result, wherein the refutation includes vulnerability detection and refutation score of the preliminary social measure implementation plan; Control each sub-agent to adjust the preliminary social measures implementation plan according to the demonstration results and refutation results of the preliminary social measures implementation plan formulated by each sub-agent, so as to obtain the revised implementation plan formulated by each sub-agent; Repeat the above-mentioned steps of argumentation, refutation, and revision until the argumentation result and / or refutation result of the revised implementation plan meets the preset conditions, and then determine the final social measure implementation plan based on the revised implementation plan currently obtained by each sub-agent.

[0014] According to a multi-agent decision-making method provided by the present invention, each sub-agent is controlled to demonstrate the preliminary social measure implementation plan formulated by each sub-agent, and the demonstration result is obtained, including: Controlling the sub-agent to analyze the preliminary social measure implementation plan based on the decision basis and reasoning process recorded when generating the preliminary social measure implementation plan; The control sub-agent scores the preliminary social measures implementation plan, and determines the expected effects of the preliminary social measures implementation plan and the potential risks it faces based on the score of the preliminary social measures implementation plan.

[0015] According to a multi-agent decision-making method provided by the present invention, for each sub-agent, the sub-agent is controlled to refute the preliminary social measure implementation plan formulated by other sub-agents to obtain the refutation result, including: For each sub-agent, the control sub-agent analyzes the preliminary social measures implementation plans formulated by other sub-agents, and determines the loopholes and rebuttal scores of the preliminary social measures implementation plans formulated by other sub-agents based on the analysis results.

[0016] According to a multi-agent decision-making method provided by the present invention, the method further comprises: Receive user feedback on the final social measures implementation plan; The sub-agents are adjusted according to the feedback information, and the adjustment content includes at least one of the number of sub-agents, sub-agent types, shared knowledge, independent knowledge, decision model types, argumentation methods and refutation methods.

[0017] The present invention also provides a multi-agent based decision-making device, comprising the following modules: The data collection module is used to collect data from multiple data sources and build multiple knowledge bases in different knowledge fields; The sub-agent generation module is used to construct multiple sub-agents based on the different task requirements contained in the target social measures by adopting the corresponding decision-making model and the knowledge base of the corresponding knowledge field; The debate module is used to control multiple sub-agents to formulate their own preliminary social measure implementation plans for the target social measures, and to debate on multiple preliminary social measure implementation plans to obtain the final social measure implementation plan.

[0018] According to a multi-agent decision-making device provided by the present invention, the data acquisition module is specifically used for: Connect to multiple data sources of different types; Divide multiple knowledge domains based on data stored in multiple different types of data sources; For each knowledge domain, the knowledge base is constructed using the data of the corresponding knowledge domain.

[0019] According to a multi-agent decision-making device provided by the present invention, the sub-agent generation module is specifically used for: Extracting shared knowledge from multiple knowledge bases based on targeted social measures; For each task requirement, determine the decision model corresponding to the task requirement; For each task requirement, extract the perspective features, domain knowledge, and thinking methods required to meet the task requirements from the knowledge base of the corresponding knowledge domain, and generate independent knowledge corresponding to the task requirements; Based on shared knowledge, sub-agents are constructed for each task requirement using the decision model corresponding to the task requirement and the corresponding independent knowledge.

[0020] According to a multi-agent decision-making device provided by the present invention, the debate module is specifically used for: Control multiple sub-agents to formulate preliminary social measures implementation plans for target social measures that meet the requirements of their respective tasks; Control each sub-agent to demonstrate the preliminary social measures implementation plan formulated by each sub-agent and obtain demonstration results. The demonstration includes clarifying the decision-making basis, reasoning process, expected effects and potential risks faced by the preliminary social measures implementation plan; For each sub-agent, the control sub-agent refutes the preliminary social measure implementation plan formulated by other sub-agents to obtain a refutation result, wherein the refutation includes vulnerability detection and refutation score of the preliminary social measure implementation plan; Control each sub-agent to adjust the preliminary social measures implementation plan according to the demonstration results and refutation results of the preliminary social measures implementation plan formulated by each sub-agent, so as to obtain the revised implementation plan formulated by each sub-agent; Repeat the above-mentioned steps of argumentation, refutation, and revision until the argumentation result and / or refutation result of the revised implementation plan meets the preset conditions, and then determine the final social measure implementation plan based on the revised implementation plan currently obtained by each sub-agent.

[0021] According to a multi-agent decision-making device provided by the present invention, the debate module controls each sub-agent to argue the preliminary social measure implementation plan formulated by each sub-agent and obtains the argumentation result, which is specifically used to: Controlling the sub-agent to analyze the plan based on the decision-making basis and reasoning process recorded when generating the preliminary social measure implementation plan; The control sub-agent scores the preliminary social measures implementation plan, and determines the expected effects of the preliminary social measures implementation plan and the potential risks it faces based on the score of the preliminary social measures implementation plan.

[0022] According to a multi-agent decision-making device provided by the present invention, the debate module controls each sub-agent to refute the preliminary social measure implementation plan formulated by other sub-agents, and when the refutation result is obtained, it is specifically used to: For each sub-agent, the control sub-agent analyzes the preliminary social measures implementation plans formulated by other sub-agents, and determines the loopholes and rebuttal scores of the preliminary social measures implementation plans formulated by other sub-agents based on the analysis results.

[0023] According to a multi-agent decision-making device provided by the present invention, the device also includes a feedback module, which is specifically used to: Receive user feedback on the final social measures implementation plan; The sub-agents are adjusted according to the feedback information, and the adjustment content includes at least one of the number of sub-agents, sub-agent types, shared knowledge, independent knowledge, decision model types, argumentation methods and refutation methods.

[0024] The present invention also provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, any of the above-mentioned multi-agent-based decision-making methods is implemented.

[0025] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements any of the above-mentioned multi-agent-based decision-making methods.

[0026] The present invention also provides a computer program product, including a computer program, which implements any of the above-mentioned multi-agent-based decision-making methods when executed by a processor.

[0027] The multi-agent-based decision-making method provided by the present invention first collects data from multiple data sources to construct multiple knowledge bases in different knowledge fields, then adopts corresponding decision models and knowledge bases in corresponding knowledge fields to construct multiple sub-agents for different task requirements contained in the target social measures, and finally controls multiple sub-agents to formulate their own preliminary social measures implementation plans for the target social measures, and debates on multiple preliminary social measures implementation plans to obtain the final social measures implementation plan. Using the method provided by the embodiment of the present invention, multiple sub-agents can be integrated. Since each sub-agent is constructed by decision models corresponding to knowledge bases in different knowledge fields and task requirements, each sub-agent can represent different perspectives. Integrating multiple sub-agents for decision-making can form multi-field and multi-dimensional insights, and then the social measures implementation plans of multiple sub-agents are integrated through the debate mechanism of multiple sub-agents, so that the agent can produce more comprehensive and diversified decision results, and effectively deal with complex and changeable decision-making problems. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0029] Figure 1 It is a flowchart of the multi-agent based decision-making method provided by the present invention.

[0030] Figure 2 It is a schematic diagram of the structure of the multi-agent provided by the present invention.

[0031] Figure 3 It is a schematic diagram of the structure of the multi-agent based decision-making device provided by the present invention.

[0032] Figure 4 It is a structural schematic diagram of the electronic device provided by the present invention. DETAILED DESCRIPTION

[0033] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the drawings of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0034] Multi-agent deduction shows great potential in simulating human behavior. Multi-agent systems can simulate complex social phenomena, such as social network interactions, individual and group behaviors, and financial market dynamics. By simulating behaviors in different situations, these systems can reveal the underlying mechanisms of social phenomena and provide a basis for human decision-making. Large language models have powerful knowledge representation and natural language generation capabilities, and can perform a certain degree of reasoning and learning, which makes it possible to build more realistic and credible humanoid agents.

[0035] In the past few years, a lot of work has been done to build agents based on large models. For example, researchers created 25 agents in a virtual town and observed their behavior and interactions. Agents can perform daily activities such as making breakfast, going to work, and creating art, as well as complex social interactions such as planning and attending parties, remembering and reflecting on past events, and making plans for the future. Each agent has a perception system that can capture dynamic changes in the surrounding environment, such as the behavior of other agents, time, and environmental changes. This information is stored in the agent's memory, including short-term memory and long-term memory. Short-term memory stores recent events and information, and long-term memory stores long-term experience and historical data. Through the feedback loop mechanism, the agent can retrieve relevant information from the memory to make decisions in the current situation.

[0036] Researchers and others proposed the RecAgent framework, which aims to study user behavior patterns by simulating user behavior in the recommendation system to improve the performance of the recommendation system. The intelligent agent construction part of the framework is divided into a portrait module, a memory module, and a behavior module. The portrait module describes the basic characteristics of the user, such as gender, age, personality, interests and other information; the memory module designs three types of memory: perceptual memory records the user's immediate perception and short-term reaction, short-term memory saves recent events and user behavior, and long-term memory stores the user's long-term preferences and historical behavior; the action module is responsible for simulating various behaviors of users in the recommendation system, such as browsing recommended content, interacting with other users, and posting messages on social media. Generating these behaviors through a large language model can make the user behavior simulation closer to reality.

[0037] However, existing methods all use a single agent to model a goal, which has the following shortcomings: (1) Perspective limitations: When simulating complex real-world scenarios, single-agent models are often limited by their inherent perspectives. Although large models have powerful capabilities, they may also be biased or fail to fully consider the multi-domain, multi-dimensional factors and diverse perspectives involved in the problem. This single-perspective modeling method is difficult to accurately reproduce the multi-angle thinking and comprehensive decision-making process that humans show when facing complex problems, thus limiting the effectiveness and universality of the model in solving complex decision-making problems.

[0038] (2) Hallucinations and errors: When a single agent makes a decision, it is easily affected by the hallucinations and false assumptions inherent in the large model, that is, the agent generates seemingly reasonable but actually wrong conclusions without sufficient evidence. Such errors not only reduce the reliability of the agent in the deduction, but may also have a misleading impact on the decision.

[0039] (3) Poor explainability: The decision-making process of complex problems is characterized by highly nonlinear and invisible internal mechanisms, lacking transparency and being difficult to understand and explain.

[0040] Therefore, a multi-agent based decision-making method is urgently needed to solve the above problems.

[0041] Combine the following Figure 1-Figure 4 Describe the multi-agent based decision-making method, device, equipment and medium of the present invention.

[0042] Figure 1 is one of the flow charts of the decision-making method based on multi-agent provided by the present invention, such as Figure 1 As shown, the method includes the following: Step 101: Collect data from multiple data sources and build multiple knowledge bases in different knowledge fields.

[0043] This step can be implemented by the data acquisition module, which adopts a modular design, supports flexible configuration and expansion, and can connect and access various types of data sources. These data sources include but are not limited to real-time environmental monitoring systems, historical databases, domain knowledge bases, social media platforms, etc., covering a wide range of information fields from physical environment to social culture. The data in these data sources provides the basic information and background knowledge required for decision-making, and provides the necessary data support for subsequent intelligent agent generation and decision-making.

[0044] In some embodiments, data from multiple data sources are collected to construct multiple knowledge bases in different knowledge fields, including: connecting multiple different types of data sources; dividing multiple knowledge fields according to the data stored in the multiple different types of data sources; and for each knowledge field, constructing a knowledge base using the data from the corresponding knowledge field.

[0045] First, identify and determine the data sources that need to be connected. These data sources can include structured data, semi-structured data, unstructured data, and real-time data, etc. The data types in these data sources can include pictures, videos, copywriting data, etc. Design appropriate data interfaces for each type of data source, and use data integration platforms or tools to integrate data from different sources and convert data from different data sources into a unified format for processing and analysis.

[0046] By analyzing the data content, the themes and domains of the data can be identified. For example, social media data may involve sentiment analysis, topic classification, etc., while sensor data may involve environmental monitoring, equipment status, etc. Define the knowledge domain model based on the themes and content of the data. A knowledge domain can be a specific discipline or business field, such as healthcare, financial markets, user behavior, etc. Then use clustering algorithms such as K-means clustering and hierarchical clustering to group the data, each group representing a knowledge domain. After that, topic modeling techniques can be applied to identify themes in the data, thereby helping to divide the knowledge domain.

[0047] Finally, a knowledge base is designed and constructed for each knowledge domain, relevant data is extracted, loaded into the knowledge base, and maintained.

[0048] The disclosed embodiments can effectively integrate and utilize information from different data sources, create a useful knowledge base, and support subsequent decision-making processes.

[0049] Step 102: According to the different task requirements contained in the target social measures, corresponding decision-making models and knowledge bases of corresponding knowledge fields are adopted to construct multiple sub-agents.

[0050] This step can be implemented by the sub-agent generation module, which divides, designs and constructs multiple sub-agents according to task requirements and knowledge domains. Each sub-agent is given a specific perspective, knowledge base and reasoning rules, and can independently complete the preliminary analysis and evaluation of the target problem. The reasoning rules can be implemented using a decision model, and the perspective information of the sub-agent is determined by the task requirements combined with the knowledge base.

[0051] Taking the decision-making task of driving route planning as an example, the target decision measure of this task is to plan a reasonable driving route for the user. This target decision measure can include many different task requirements, such as the shortest driving route, the least expected red light duration in the driving route, and the lowest probability of congestion in the driving route. For the above three path planning task requirements, the sub-agent generation module can select the knowledge base required to meet these three task requirements respectively. For example, the task requirement of the shortest driving route needs to include the knowledge base of the path distance, the task requirement of the least expected red light duration in the driving route needs to include the knowledge base of the traffic light time, and the task requirement of the least probability of congestion in the driving route needs to include the knowledge base of historical congested road conditions. It is not difficult to understand that the sub-agent makes path planning decisions for these task requirements respectively. These task requirements contain the perspective information of the sub-agent, which are planning for the perspective of the shortest distance, planning for the perspective of the least expected red light duration, and planning for the perspective of the least probability of congestion.

[0052] The corresponding decision model can adopt a machine learning model. For example, for the task requirement of the shortest distance, a machine learning model trained with the target output of the shortest distance planning path can be adopted; for the task requirement of the least expected time at red lights, a machine learning model trained with the target output of the planning path with the least time at red lights can be adopted; for the task requirement of the least congestion probability, a machine learning model trained with the target output of the planning path with the least congestion probability can be adopted.

[0053] Sub-agent 210 can be constructed by combining the task requirement of shortest distance, a knowledge base including path lengths, and a decision model 211 whose target output is the planned path with the shortest distance. Sub-agent 220 can be constructed by combining the task requirement of shortest estimated time spent at red lights, a knowledge base including traffic light times, and a decision model 221 whose target output is the planned path with the shortest time spent at red lights. Sub-agent 230 can be constructed by combining the task requirement of minimum probability of congestion, a knowledge base including historical congested road conditions, and a decision model 231 whose target output is the planned path with the minimum probability of congestion.

[0054] In some embodiments, for different task requirements and different knowledge fields included in the target social measures, corresponding decision models and knowledge bases of corresponding knowledge fields are used to construct multiple sub-agents, including: extracting shared knowledge from multiple knowledge bases according to the target social measures; for each task requirement, determining the decision model corresponding to the task requirement; for each task requirement, extracting the perspective characteristics, domain knowledge and thinking methods required to meet the task requirement from the knowledge base of the corresponding knowledge field, and generating independent knowledge corresponding to the task requirement; based on the shared knowledge, for each task requirement, using the decision model corresponding to the task requirement and the corresponding independent knowledge to construct a sub-agent.

[0055] Shared knowledge is the common foundation of all sub-agents, ensuring the basic consistency and coordination between them. For example, in the above-mentioned path planning task scenario, map information can be used as shared knowledge 201 of all sub-agents.

[0056] On the basis of shared knowledge, each sub-agent has independent knowledge, which includes the sub-agent's unique perspective features, domain knowledge and thinking mode. For example, the independent knowledge 212 of sub-agent 210 can be the domain knowledge in the knowledge base containing the route, the independent knowledge 222 of sub-agent 220 can be the domain knowledge in the knowledge base containing the traffic light time, and the independent knowledge 232 of sub-agent 230 can be the domain knowledge in the knowledge base containing the historical congested road conditions. In this way, the target agent 200 generated by the sub-agent generation module for the task of driving route planning is as follows: Figure 2 shown.

[0057] Each sub-agent is controlled by its own decision-making model, which enables each sub-agent to understand problems, analyze situational information and make decisions in a unique way, thereby promoting diversity and innovation within the intelligent agent system.

[0058] The disclosed embodiment generates sub-agents with different characteristics, thereby ensuring that the generated sub-agents not only meet the overall requirements of the task but also have unique personalities and capabilities.

[0059] Step 103: Control multiple sub-agents to formulate their own preliminary social measure implementation plans for the target social measures, and debate on multiple preliminary social measure implementation plans to obtain a final social measure implementation plan.

[0060] This step can be implemented by the debate module, which controls each sub-agent to formulate a preliminary social measures implementation plan based on its unique perspective and decision-making method, and then allows multiple sub-agents to simulate the human decision-making process to debate. By simulating the complex collision of thoughts, it ensures that the social measures implementation plans obtained from opinions and reasoning from different perspectives can be fully discussed and coordinated, thereby integrating multiple preliminary social measures implementation plans to obtain the final social measures implementation plan.

[0061] In some embodiments, multiple sub-agents are controlled to formulate their own preliminary social measures implementation plans for the target social measures, and debate on multiple preliminary social measures implementation plans to obtain a final social measures implementation plan, including: controlling multiple sub-agents to formulate preliminary social measures implementation plans for the target social measures that meet their respective corresponding task requirements; controlling each sub-agent to demonstrate the preliminary social measures implementation plans formulated by each sub-agent to obtain demonstration results, and the demonstration includes clarifying the decision-making basis, reasoning process, expected effects and potential risks faced by the preliminary social measures implementation plans; for each sub-agent, controlling the sub-agent to refute the preliminary social measures implementation plans formulated by other sub-agents to obtain rebuttal results, and the rebuttal includes vulnerability detection and rebuttal scoring of the preliminary social measures implementation plans; controlling each sub-agent to adjust the preliminary social measures implementation plans according to the demonstration results and rebuttal results of the preliminary social measures implementation plans formulated by each sub-agent, and obtaining a revised implementation plan formulated by each sub-agent; repeating the above-mentioned demonstration, refutation and revision steps until the demonstration results and / or refutation results of the revised implementation plan meet the preset conditions, and then determining the final social measures implementation plan according to the revised implementation plan currently obtained by each sub-agent.

[0062] The debate module aims to simulate the complex collision of thoughts in the brain during the human decision-making process. Through the four stages of proposal, argumentation, refutation and reflection and correction, it promotes in-depth communication and mutual error correction among sub-agents, thereby enhancing the scientificity, rationality and explainability of decision-making.

[0063] In the proposal phase, each sub-agent independently proposes a preliminary social measure implementation plan based on its unique perspective and decision-making method. This process is the starting point of decision-making and requires the sub-agent to give full play to its unique characteristics and form a preliminary concept that is innovative and feasible.

[0064] In the demonstration phase, the sub-agent needs to elaborate and demonstrate the implementation plan of the social measures it proposes, including clarifying the decision-making basis, reasoning process, expected effects, and potential risks of the plan. Through this process, not only the credibility and explainability of the plan are enhanced, but also a clear target is provided for subsequent rebuttal and evaluation.

[0065] The rebuttal stage is the core link of the debate mechanism. In the rebuttal stage, the sub-agent needs to question and refute the solutions of other agents, pointing out their shortcomings or potential problems. This process simulates the fierce confrontation and collision between different viewpoints in the human brain during human decision-making, which helps to discover loopholes and deficiencies in the solution and promote the continuous improvement and optimization of the solution. The comprehensive review stage is the key link to reflect on and revise each solution based on the debate situation.

[0066] In the reflection and correction stage, sub-agents adjust and improve their own plans based on feedback, integrating the insights of different agents to form a more mature and feasible implementation plan for social measures.

[0067] Repeat the process of argumentation, refutation and reflection and correction until the argumentation results and / or refutation results of each revised social measure implementation plan meet the preset conditions, such as reaching a consensus among the sub-intelligent agents or achieving a certain balance in the revised social measure implementation plan, so as to obtain the final social measure implementation plan.

[0068] Among them, each sub-intelligent agent is controlled to demonstrate the preliminary social measures implementation plan formulated by each of them to obtain the demonstration result, including: controlling the sub-intelligent agent to analyze the preliminary social measures implementation plan according to the decision-making basis and reasoning process recorded when generating the preliminary social measures implementation plan; controlling the sub-intelligent agent to score the preliminary social measures implementation plan, and determining the expected effect of the preliminary social measures implementation plan and the potential risks it faces according to the score of the preliminary social measures implementation plan.

[0069] Exemplarily, a scoring mechanism can be introduced into the decision model of the sub-agent to achieve the effect of demonstration. For example, the shortest distance implementation plan made by the sub-agent 210 will also generate the decision basis and reasoning process synchronously. The decision model 211 can analyze the shortest distance implementation plan, the decision basis and the reasoning process, and then score it according to the analysis results. For example, the plan score of the shortest distance implementation plan is 7 points. If the full score of the implementation plan is 10, it can be determined that the potential risk level of the shortest distance implementation plan is 3 points. Similarly, the decision model 221 of the sub-agent 220 can analyze the shortest red light duration implementation plan made by the sub-agent 220 to obtain the demonstration result of the plan. The decision model 231 of the sub-agent 230 can analyze the minimum congestion probability implementation plan made by the sub-agent 230 to obtain the demonstration result of the plan.

[0070] Among them, for each sub-agent, the sub-agent is controlled to refute the preliminary implementation plans formulated by other sub-agents to obtain the refutation results, including: for each sub-agent, the sub-agent is controlled to analyze the preliminary implementation plans formulated by other sub-agents, and determine the loopholes and refutation scores of the preliminary implementation plans formulated by other sub-agents based on the analysis results.

[0071] Exemplarily, the decision model 221 of the sub-agent 210 analyzes the shortest red light duration implementation plan made by the sub-agent 220 and the minimum congestion probability implementation plan made by the sub-agent 230 from the perspective of planning the shortest distance path, and obtains the loopholes and solution scores detected by the sub-agent 210 for the shortest red light duration implementation plan and the minimum congestion probability implementation plan. Similarly, the sub-agent 220 can obtain the loopholes and solution scores detected by it for the shortest distance implementation plan and the minimum congestion probability implementation plan, and the sub-agent 230 can obtain the loopholes and solution scores detected by it for the shortest distance implementation plan and the shortest red light duration implementation plan.

[0072] Sub-agent 210 then modifies the preliminary shortest distance implementation plan according to the loopholes and solution scores detected by sub-agent 220 and sub-agent 230 for the shortest distance implementation plan, and obtains a modified shortest distance implementation plan. Sub-agent 220 modifies the preliminary shortest red light duration implementation plan according to the loopholes and solution scores detected by sub-agent 210 and sub-agent 230 for the shortest red light duration implementation plan, and obtains a modified shortest red light duration implementation plan. Sub-agent 230 modifies the preliminary minimum congestion probability implementation plan according to the loopholes and solution scores detected by sub-agent 210 and sub-agent 220 for the minimum congestion probability implementation plan, and obtains a modified minimum congestion probability implementation plan.

[0073] If the revised shortest distance implementation plan, the revised shortest red light duration implementation plan and the revised minimum congestion probability implementation plan meet the preset conditions, the final implementation plan is determined based on the three revised implementation plans. For example, the average score of the shortest distance implementation plan is calculated based on the scheme scores of the revised shortest distance implementation plan by sub-agent 210, sub-agent 220 and sub-agent 230, the average score of the shortest red light duration implementation plan is calculated based on the scheme scores of the revised shortest red light duration implementation plan by sub-agent 210, sub-agent 220 and sub-agent 230, and the average score of the minimum congestion probability implementation plan is calculated based on the scheme scores of the revised minimum congestion probability implementation plan by sub-agent 210, sub-agent 220 and sub-agent 230. If the difference between the average scores of the three revised implementation plans is less than the preset value, the revised implementation plan with the highest average score is used as the final implementation plan. Or when the average scores of the three revised implementation plans no longer change, the revised implementation plan with the highest average score is used as the final implementation plan.

[0074] The disclosed embodiment sets up a self-argumentation and debate mechanism among multiple sub-agents, requiring each sub-agent to fully demonstrate its views and reasons, and accept questioning and discussion from other sub-agents. This process not only increases the depth of communication within the system, but also enables external observers to clearly see the entire process of decision-making, thereby improving the transparency and explainability of the decision. At the same time, it enables multiple sub-agents to review and complement each other, forming a cycle of self-correction and improvement. During the debate, sub-agents can discover and correct each other's biases and errors, thereby improving the overall fault tolerance. This mechanism significantly improves the accuracy and reliability of system decisions, making the decision results more robust and reliable.

[0075] By using the method provided in the embodiment of the present invention, multiple sub-agents can be integrated. Since each sub-agent is constructed by a knowledge base in a different knowledge field and a decision-making model corresponding to the task requirements, each sub-agent can represent a different perspective. Integrating multiple sub-agents for decision-making can form multi-field and multi-dimensional insights. Then, through the debate mechanism of multiple sub-agents, the social measures implementation plans of multiple sub-agents are integrated, so that the agents can produce more comprehensive and diverse decision-making results and effectively deal with complex and changeable decision-making problems.

[0076] In some embodiments, the method also includes: receiving user feedback information on the final social measure implementation plan; adjusting the sub-agents based on the feedback information, and the adjustment content includes at least one of the number of sub-agents, sub-agent types, shared knowledge, independent knowledge, decision model type, argumentation method and refutation method.

[0077] The disclosed embodiment can be implemented by a feedback module, which serves as a bridge connecting humans and the intelligent agent construction framework. The feedback module is responsible for fine-tuning the core components of the intelligent agent, namely the sub-agent generation module and the debate module, based on real-time or accumulated human feedback information.

[0078] Specifically, the feedback mechanism first captures and analyzes direct evaluations, preference indications, and indirect feedback from users or experts on task execution results. By deeply analyzing human feedback, the feedback module can identify possible deficiencies in the design of the agent structure, and then guide the sub-agent generation module and the debate module to make adaptive adjustments. This includes but is not limited to the number of sub-agents, sub-agent types, shared knowledge and independent knowledge, large model types, argument logic, and refutation strength.

[0079] This process not only improves the adaptability of decision-making strategies to specific task environments, but also promotes the generalization ability of intelligent agents in unknown or dynamically changing environments. Through continuous iterative optimization, intelligent agents can gradually learn to follow human expectations and norms and make more reasonable and efficient decisions.

[0080] In addition, intelligent agents have great application potential in decision-making on implementation plans, especially in complex environments. They can help decision makers simulate human cognition and behavior, thereby improving the scientificity and accuracy of the formulation and evaluation of implementation plans.

[0081] For example, agents can be used to evaluate implementation plans of measures. Agents can simulate the behavior of social members with different backgrounds and individual characteristics. Each agent can contain unique attributes, such as age, occupation, income level, social status, cultural background, education level, etc. Through these attributes, agents can show diverse behavioral patterns. In the simulation of the impact of implementation plans of measures, different individuals will react differently to social measures. Humanoid agents can produce diverse behavioral responses based on these differences, thereby more comprehensively reflecting the reactions of various groups of people in society to social measures. In addition, humanoid agents can also simulate the interaction between individuals after the implementation of social measures. Since the actual effect of social measures depends not only on the independent behavior of individuals, but also on the influence of mutually influential relationships in social networks, it is necessary to simulate the interaction between individuals to more accurately evaluate the impact of social measures.

[0082] For example, when evaluating a social measure that encourages environmental protection, the intelligent agent can simulate how changes in environmental attitudes in the social circles of different individuals after receiving the social measure affect the decisions of others, thereby predicting the promotion of environmental protection behavior of the social measure to the entire society.

[0083] For example, in emergency management, intelligent agents can be used to simulate the emergency response and social recovery process of disasters, helping managers of some places to formulate emergency plans and safety measures more effectively. It can also simulate the behavior patterns of users in crises and assist in designing more flexible response measures.

[0084] Based on the social simulation application of humanoid agents, the planners can have a more comprehensive understanding of complex social dynamics, predict the effects of social measures after implementation by simulating human behavior, foresee potential problems and make optimization adjustments, thereby improving the scientificity, rationality and implementation effect of social measures and reducing the risk of implementation plan errors. The following is a description of the multi-agent-based decision-making device provided by the present invention. The multi-agent-based decision-making device described below and the multi-agent-based decision-making method described above can be referenced to each other.

[0085] like Figure 3 As shown, the multi-agent based decision-making device 300 provided by the present invention includes the following modules: The data collection module 301 is used to collect data from multiple data sources and build multiple knowledge bases in different knowledge fields; A sub-agent generation module 302 is used to construct multiple sub-agents according to different task requirements contained in the target social measures by adopting corresponding decision models and knowledge bases in corresponding knowledge fields; The debate module 303 is used to control multiple sub-agents to formulate their own preliminary social measure implementation plans for the target social measure, and to debate on multiple preliminary social measure implementation plans to obtain a final social measure implementation plan.

[0086] According to a multi-agent-based decision-making device provided by the present invention, the data acquisition module 301 is specifically used for: Connect to multiple data sources of different types; Divide multiple knowledge domains based on data stored in multiple different types of data sources; For each knowledge domain, the knowledge base is constructed using the data of the corresponding knowledge domain.

[0087] According to a multi-agent decision-making device provided by the present invention, the sub-agent generation module 302 is specifically used for: Extracting shared knowledge from multiple knowledge bases based on targeted social measures; For each task requirement, determine the decision model corresponding to the task requirement; For each task requirement, extract the perspective features, domain knowledge, and thinking methods required to meet the task requirements from the knowledge base of the corresponding knowledge domain, and generate independent knowledge corresponding to the task requirements; Based on shared knowledge, sub-agents are constructed for each task requirement using the decision model corresponding to the task requirement and the corresponding independent knowledge.

[0088] According to a multi-agent-based decision-making device provided by the present invention, the debate module 303 is specifically used for: Control multiple sub-agents to formulate preliminary social measures implementation plans for target social measures that meet the requirements of their respective tasks; Control each sub-agent to demonstrate the preliminary social measures implementation plan formulated by each sub-agent and obtain demonstration results. The demonstration includes clarifying the decision-making basis, reasoning process, expected effects and potential risks faced by the preliminary social measures implementation plan; For each sub-agent, the control sub-agent refutes the preliminary social measure implementation plan formulated by other sub-agents to obtain a refutation result, wherein the refutation includes vulnerability detection and refutation score of the preliminary social measure implementation plan; Control each sub-agent to adjust the preliminary social measures implementation plan according to the demonstration results and refutation results of the preliminary social measures implementation plan formulated by each sub-agent, so as to obtain the revised implementation plan formulated by each sub-agent; Repeat the above-mentioned steps of argumentation, refutation, and revision until the argumentation result and / or refutation result of the revised implementation plan meets the preset conditions, and then determine the final social measure implementation plan based on the revised implementation plan currently obtained by each sub-agent.

[0089] According to a multi-agent decision-making device provided by the present invention, the debate module 303 controls each sub-agent to argue the preliminary social measure implementation plan formulated by each sub-agent and obtains the argumentation result, which is specifically used for: Controlling the sub-agent to analyze the preliminary social measure implementation plan based on the decision basis and reasoning process recorded when generating the preliminary social measure implementation plan; The control sub-agent scores the preliminary social measures implementation plan, and determines the expected effects of the preliminary social measures implementation plan and the potential risks it faces based on the score of the preliminary social measures implementation plan.

[0090] According to a multi-agent decision-making device provided by the present invention, the debate module 303 controls each sub-agent to refute the preliminary social measure implementation plan formulated by other sub-agents, and when the refutation result is obtained, it is specifically used to: For each sub-agent, the control sub-agent analyzes the preliminary social measures implementation plans formulated by other sub-agents, and determines the loopholes and rebuttal scores of the preliminary social measures implementation plans formulated by other sub-agents based on the analysis results.

[0091] According to a multi-agent decision-making device provided by the present invention, the device further includes a feedback module 304, which is specifically used to: Receive user feedback on the final social measures implementation plan; The sub-agents are adjusted according to the feedback information, and the adjustment content includes at least one of the number of sub-agents, sub-agent types, shared knowledge, independent knowledge, decision model types, argumentation methods and refutation methods.

[0092] Figure 4 An example of a physical structure diagram of an electronic device is shown in FIG. Figure 4 As shown, the electronic device may include: a processor 1110, a communications interface 1120, a memory 1130 and a communication bus 1140, wherein the processor 1110, the communications interface 1120 and the memory 1130 communicate with each other through the communication bus 1140. The processor 1110 may call the logic instructions in the memory 1130 to execute the decision-making method based on multi-agents, and the method includes: Collect data from multiple data sources and build multiple knowledge bases in different knowledge fields; According to the different task requirements contained in the target social measures, corresponding decision-making models and knowledge bases in corresponding knowledge fields are used to construct multiple sub-agents; Control multiple sub-agents to formulate their own preliminary social measure implementation plans for the target social measures, and debate on multiple preliminary social measure implementation plans to obtain the final social measure implementation plan.

[0093] In addition, the logic instructions in the above-mentioned memory 1130 can be implemented in the form of a software functional unit and can be stored in a computer-readable storage medium when it is sold or used as an independent product. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, etc. Various media that can store program codes.

[0094] On the other hand, the present invention further provides a computer program product, the computer program product comprising a computer program, the computer program can be stored on a non-transitory computer-readable storage medium, when the computer program is executed by a processor, the computer can execute the multi-agent based decision-making method provided by the above methods, the method comprising: Collect data from multiple data sources and build multiple knowledge bases in different knowledge fields; According to the different task requirements contained in the target social measures, corresponding decision-making models and knowledge bases in corresponding knowledge fields are used to construct multiple sub-agents; Control multiple sub-agents to formulate their own preliminary social measure implementation plans for the target social measures, and debate on multiple preliminary social measure implementation plans to obtain the final social measure implementation plan.

[0095] In another aspect, the present invention further provides a non-transitory computer-readable storage medium having a computer program stored thereon, which is implemented when the computer program is executed by a processor to execute the multi-agent-based decision-making method provided by the above methods, the method comprising: Collect data from multiple data sources and build multiple knowledge bases in different knowledge fields; According to the different task requirements contained in the target social measures, corresponding decision-making models and knowledge bases in corresponding knowledge fields are used to construct multiple sub-agents; Control multiple sub-agents to formulate their own preliminary social measure implementation plans for the target social measures, and debate on multiple preliminary social measure implementation plans to obtain the final social measure implementation plan.

[0096] The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Ordinary technicians in this field can understand and implement it without paying creative labor.

[0097] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, can also be implemented by hardware. Based on this understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0098] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A decision-making method based on multi-agent, characterized in that: The method comprises: Collect social measures data from multiple data sources and build multiple social measures knowledge bases in different knowledge fields; According to the different task requirements contained in the target social measures, multiple sub-agents are constructed by adopting corresponding decision-making models and social measures knowledge bases in corresponding knowledge fields; The multiple sub-intelligent agents are controlled to formulate respective preliminary social measure implementation plans for the target social measure, and debate on the multiple preliminary social measure implementation plans to obtain a final social measure implementation plan.

2. The multi-agent decision-making method according to claim 1, characterized in that: The social measure data collected from multiple data sources and multiple social measure knowledge bases in different knowledge fields are constructed, including: Connect to multiple data sources of different types; dividing a plurality of knowledge domains according to the social measure data stored in the plurality of different types of data sources; For each knowledge area, the social measures knowledge base is constructed using the data of the corresponding knowledge area.

3. The multi-agent based decision-making method according to claim 1, characterized in that: According to the different task requirements contained in the target social measures, a corresponding decision model and a social measures knowledge base in the corresponding knowledge field are used to construct multiple sub-agents, including: extracting shared knowledge from the plurality of knowledge bases according to the target social measure; For each task requirement, determining a decision model corresponding to the task requirement; For each task requirement, extract the perspective features, domain knowledge and thinking mode required to meet the task requirement from the knowledge base of the corresponding knowledge domain, and generate independent knowledge corresponding to the task requirement; Based on the shared knowledge, the sub-agent is constructed for each task requirement using the decision model corresponding to the task requirement and the corresponding independent knowledge.

4. The multi-agent decision-making method according to claim 1, characterized in that: The controlling of the plurality of sub-agents to formulate respective preliminary social measure implementation plans for the target social measure, and debating the plurality of preliminary social measure implementation plans to obtain a final social measure implementation plan, including: Controlling the plurality of sub-agents to formulate preliminary social measure implementation plans for the target social measures that meet the requirements of their respective corresponding tasks; Controlling each sub-agent to demonstrate the preliminary social measures implementation plan formulated by each sub-agent to obtain demonstration results, wherein the demonstration includes clarifying the decision basis, reasoning process, expected effects and potential risks faced by the preliminary social measures implementation plan; For each sub-agent, controlling the sub-agent to refute the preliminary social measure implementation plan formulated by other sub-agents to obtain a refutation result, wherein the refutation includes vulnerability detection and refutation score of the preliminary social measure implementation plan; Control each sub-agent to adjust the preliminary social measures implementation plan according to the demonstration results and refutation results of the preliminary social measures implementation plan formulated by each sub-agent, so as to obtain a revised implementation plan formulated by each sub-agent; Repeat the above-mentioned steps of argumentation, refutation and correction until the argumentation result and / or refutation result of the revised implementation plan meets the preset conditions, and then determine the final social measure implementation plan based on the revised implementation plan currently obtained by each sub-agent.

5. The multi-agent based decision-making method according to claim 4, characterized in that: The controlling each sub-agent to demonstrate the preliminary social measures implementation plan formulated by each sub-agent, and obtaining demonstration results, including: Controlling the sub-agent to analyze the preliminary social measure implementation plan according to the decision basis and reasoning process recorded when generating the preliminary social measure implementation plan; The sub-agent is controlled to score the preliminary social measures implementation plan, and the expected effects of the preliminary social measures implementation plan and the potential risks faced by the preliminary social measures implementation plan are determined according to the score of the preliminary social measures implementation plan.

6. The multi-agent based decision-making method according to claim 4, characterized in that: For each sub-intelligent agent, controlling the sub-intelligent agent to refute the preliminary social measure implementation plan formulated by other sub-intelligent agents, and obtaining the refutation result, including: For each sub-agent, control the sub-agent to analyze the preliminary social measure implementation plans formulated by other sub-agents, and determine the loopholes and rebuttal scores of the preliminary social measure implementation plans formulated by other sub-agents based on the analysis results.

7. The multi-agent based decision-making method according to claim 3, characterized in that: The method further comprises: receiving user feedback on the final social measure implementation plan; The sub-agents are adjusted according to the feedback information, and the adjustment content includes at least one of the number of sub-agents, sub-agent types, shared knowledge, independent knowledge, decision model types, argumentation methods and refutation methods.

8. A decision-making device based on multi-agents, characterized in that: include: The data collection module is used to collect data from multiple data sources and build multiple knowledge bases in different knowledge fields; The sub-agent generation module is used to construct multiple sub-agents based on the different task requirements contained in the target social measures by adopting the corresponding decision-making model and the knowledge base of the corresponding knowledge field; The debate module is used to control the multiple sub-agents to formulate their own preliminary social measure implementation plans for the target social measure, and to debate on the multiple preliminary social measure implementation plans to obtain a final social measure implementation plan.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, it implements the multi-agent based decision-making method as described in any one of claims 1 to 7.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, it implements the multi-agent based decision-making method as described in any one of claims 1 to 7.

Images