Multi-agent simulation system of large language model

Abstract

The invention provides a multi-agent simulation system based on a large language model, and the system comprises a man-machine interaction setting module which is used for providing an interaction interface between a user and the system, enabling the user to carry out the scene setting and parameter setting used in a simulation operation process through the man-machine interaction setting module, and providing the data information needed by the operation of the system, in combination with real scene information, high-simulation simulation scenes and parameters are set for specific simulation; the multi-agent collaboration module is used for specifically performing multi-agent simulation operation based on a large language model, processing main body and environment information and realizing a development process of a real event along with time lapse; the data bus communication module is used for communicating the intelligent agent with the database and connecting the intelligent agent with the database in a bus mode so as to improve the information exchange efficiency; and the log causal analysis module is used for analyzing the simulation process and result. The simulation system provided by the invention can fully reflect the running condition of the real world.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of simulation systems, in particular to a large language model multi-agent simulation system. BACKGROUND

[0002] In the prior art platform operation analysis method, mainly through research experiments or traditional simulation. Research experiments can fully reflect the diversity of real-world user groups, but there are problems such as long cycle, many personnel, high cost, etc.; traditional simulation is mainly based on queuing theory, inventory theory and other theories of simulation, which abstractly fits specific problems, simulates the operation of real-world scenarios, which greatly improves the efficiency of simulation, but it is difficult to quantitatively fit the decision-making simulation of complex characters, and is not suitable for complex human decision-making scenarios in reality. SUMMARY

[0003] In order to solve the technical problems existing in the prior art, the present application provides a large language model multi-agent simulation system, which comprises: a man-machine interaction setting module, a multi-agent cooperation module, a data bus communication module, and a log causal analysis module.

[0004] The man-machine interaction setting module is used to provide an interactive interface between the user and the system, and the user sets the scene and parameters used in the simulation running process through the man-machine interaction setting module, and provides data information required for system running, and combines real scene information to set high-fidelity simulation scenarios and parameters for specific simulation.

[0005] The multi-agent cooperation module is used for specific multi-agent simulation running based on a large language model, processes subject and environment information, and simulates the development process of real events over time.

[0006] The data bus communication module is used for communication between the agent and the database, and connects the agent and the database in a bus manner to improve the information exchange efficiency.

[0007] The log causal analysis module is used for analyzing the simulation process and results.

[0008] Optionally, the man-machine interaction setting module is specifically used for:

[0009] 1) Feature label selection: abstractly extracting key feature labels from the real world, selecting corresponding feature labels as research settings of the simulation system, which includes three aspects: subject feature setting, environment condition setting, and emergency event setting.

[0010] 2) Feature distribution extraction and setting: Distribution setting is performed for selected feature labels to better match the real world or meet research needs, and is divided into two setting methods: fitting real data and using approximate distribution of real data;

[0011] 3) Random combination of various features: By randomly matching different indicators, the label setting of complex individuals, environments and emergencies is perfected, and the random combination process ensures that a single feature meets the distribution set in the previous step;

[0012] 4) Specific setting generation: The corresponding element configuration is generated according to the feature set generated by random combination of features. This process is generated by a large language model, which generates corresponding settings according to a pre-set configuration template by combining feature sets, including: agent configuration, environment configuration, event configuration.

[0013] Optionally, the multi-agent collaboration module includes five types of agents configured based on a large language model: management agent, support agent, guide agent, execution agent, and evaluation agent.

[0014] The management agent is configured to obtain task objectives and requirements and perform task planning, distribute and broadcast tasks to the execution agent, and serve as an information transmission role.

[0015] The support agent is configured to provide environment information to the execution agent, dynamically update the environment according to the decision of the execution agent, and provide real-time information feedback, which is transmitted to the management agent.

[0016] The guide agent is configured to establish logical guide questions for the execution agent, calculate the utility of the decision set, and provide decision constraints to enable the execution agent to make more logical and human-like decisions.

[0017] The execution agent is the main role of the simulation, representing each unique individual in the simulation scenario. It receives information from the management agent and support agent, and makes decisions under the influence of the guide agent. The specific decision objects include inter-agent decisions and agent-environment decisions.

[0018] The evaluation agent is configured to evaluate the effectiveness of task execution, generate event briefings, perform real-time data statistics, update the sub-utility weight of the execution agent, and transmit information to the management agent.

[0019] Optionally, the interaction logic of each type of agent in the multi-agent collaboration module is as follows:

[0020] 1) State update:

[0021] The execution agent delivers respective decision information, updates each agent state according to all agent decision conditions of the previous timestamp, including inter-agent influence and agent-environment influence;

[0022] 2) Task generation and distribution:

[0023] The management agent generates specific task information for each execution agent for the next stage according to the current task and state, and delivers the specific task information to the corresponding execution agent;

[0024] 3) Environment state identification:

[0025] When the execution agent receives the task information, the support agent queries the environment information, identifies the current environment state information, and delivers this information to the execution agent as part of the decision support information;

[0026] 4) Agent utility quantification:

[0027] According to the BBC happiness scale BBC-SWB, a three-dimensional utility model is built, including three sub-utility indicators: psychological happiness Ps, physical health Ph, and interpersonal relationship Re. By giving corresponding definitions, standards, and case background information, the execution agent determines the score of each sub-utility indicator for each decision, simulates the decision-making process of normal humans considering multiple aspects before making decisions, and the score is affected by the agent role image and current environment background information, and changes constantly in the iteration process to better simulate the complexity of real individuals. At the same time, different weights are set for different sub-utilities, and adjusted in the iteration process. The weighted average of the sub-utilities is obtained as the total utility U ij (t) of decision j for agent i at t simulation round.

[0028] For the candidate decision S i (t) of execution agent i at simulation round t, the total utility U i (t) of each decision j∈S ij (t) for execution agent i includes three sub-utilities P ij (t)=(Ps ij (t),Ph ij (t),Re ij (t)) T ∈R 3 :

[0029] U ij (t)=w i (t) T P ij (t)

[0030] =αi (t)*Ps ij (t)+β i (t)*Ph ij (t)+γ i (t)*Re ij (t)

[0031] where w i (t) = (a i (t), b i (t), g i (t)) T is the weight vector, and satisfies the condition w i (t) e D3, D3 = {w ≥ 0, 1 T w = 1};

[0032] 5) Decision generation based on dynamic game:

[0033] When making decisions, the execution agent not only considers its own situation to make the decision that is most beneficial to itself, but also considers the potential decisions of other agents as background information that may affect its own decision, and makes decisions based on dynamic games, including:

[0034] First, the execution agent i makes the optimal decision D i (t) according to its own information and decision set:

[0035]

[0036] where S i (t) is the decision set of the execution agent i at the t simulation round;

[0037] Then, each execution agent adds the decision information of the associated agent to its own background information, and then calculates the utility of each decision under the influence of other associated agents, and updates its optimal decision D

[0038]

[0039] where D 1 , D 2 ,..., D n represent the decision information of the n agents associated with the execution agent i at the t simulation round;

[0040] After multiple game updates, until the equilibrium state is reached or the maximum game round N is reached, the final decision of each execution agent at the current simulation round is obtained.

[0041] Optionally, the multi-agent cooperation module uses a multi-agent scheduling system to schedule, so that the simulation system realizes interaction of more agents on the basis of running a small number of large language models, so that the agents make decisions more efficiently, and the memory overhead of the large language model is saved;

[0042] The multi-agent scheduling system separates the large language model from the agent, and uses the large language model only as a carrier for thinking and calculation of the agent. In the simulation process, the large language model is flexibly scheduled, and the decision of different agents is not affected. Even if multiple agents share one large language model to process tasks, due to the differences in agent portraits, perception environment data, problem sets, and decidable sets, the specific decisions made at the end will not have high similarity, and the diversity and complexity of the multi-agent system can still be maintained.

[0043] Optionally, the scheduling process of the multi-agent scheduling system is as follows:

[0044] 1) Before the simulation starts, the hardware of the running equipment and the number of simulation multi-agents are considered to set the number of large language models as the basis of the simulation;

[0045] 2) The complexity of the task to be processed is judged, and is allocated to different large language models for processing. The task requirement complexity of the support agent and the evaluation agent is not high, and is processed by a low-parameter model. The task processed by the guide agent and the management agent is relatively complex, and is processed by a high-parameter model. The processing of the execution agent is the main part of the simulation task. The task of the execution agent is secondarily distributed according to the number of optional decision sets of each round of decision making. When the length of the decision set len(S i )≤3, the low-parameter model is used for processing; otherwise, the high-parameter model is used for processing;

[0046] 3) According to the set number of large language models, the two models are processed in parallel. The task processing is temporarily respectively entered into the complex / simple waiting queue according to the FIFO principle. If the high-parameter model is idle and has no task, the simple task can also be processed by the high-parameter model to ensure full use of resources. Since the processing effect of complex tasks needs to be ensured, the opposite is not true.

[0047] Optionally, the database includes:

[0048] The background knowledge database has read-only permission, and the data content includes background information, case information and feedback evaluation related to simulation, which is used to provide a background knowledge base for simulation, and provides corresponding background basis for decision making of the execution agent. This part of data is provided by the user through the man-machine interaction module;

[0049] The feature setting database is read-only, and the data content includes role features, task summaries, and action guidelines, which are used to save the role feature settings, and this part of data is automatically generated by the human-computer interaction module in combination with the user data system;

[0050] The task information database is read-write, and the data content includes initial commands, derivative tasks, guide questions, and weight coefficients of sub-utilities, which are used to interact with the agent, update the task information and decision set in real time, and let the agent establish a thinking chain and make decisions through guide questions;

[0051] The external environment database is read-write, and the data content includes geographic information and dynamic conditions, including space-time conditions, unexpected events, and a variable feature set of the agent, which is used to provide current state information for the agent's decision and update the relevant conditions to simulate the development of events after the agent's decision;

[0052] The running log database is read-write, and the data content includes timestamps and event briefs, which are used to record the development process of the simulation and provide data support for subsequent process analysis;

[0053] The simulation result database is read-write, and the data content includes simulation rounds, simulation result summaries, and data statistics, which are used to record the simulation result state after a certain period of time and provide data support for the final result analysis of the simulation.

[0054] Optionally, the log causal analysis module is specifically used for:

[0055] 1) Data extraction;

[0056] Extracting process data and result data of the simulation from the running log database and the simulation result database;

[0057] 2) Data processing;

[0058] The simulation result is cross-sectional data after a certain period of time, which is itself a structured result, and is convenient for secondary analysis of the result data, but the log information is process information in the running of the simulation system, which is complex and semi-structured data, and needs to be processed through a corresponding data processing process to analyze the data. The processing includes two aspects: aligning and unifying various data in the log file according to the time dimension, and analyzing various events in the system evolution process to extract key events affecting the running of the simulation and corresponding observation-plan-action information under the key events. Through the data processing stage, the log information and the result information are both cleaned as structured data, which is convenient for subsequent analysis and application;

[0059] 3) Data analysis;

[0060] Through the above data extraction and data processing, platform operators can analyze the decision-making of micro individuals, build logical causal relationships about individual decision-making according to individual observation, behavior utility calculation, utility driving factor analysis and subject interaction influence, analyze why the individual chooses the current decision instead of other optional decisions when facing specific events, and obtain behavior analysis results in the aspects of environmental factors, role attributes, subject interaction and behavior utility through the disassembled analysis of individual data.

[0061] Optionally, the system adopts a ternary architecture of "task target-organization structure-behavior routing": a group task with multi-dimensional constraints and dynamic priorities is set at the top layer, and is decomposed into several sub-tasks by a management agent and is issued; a full connection organization structure is adopted as a whole to ensure that there is a direct communication link between any two agents; and a hybrid strategy is adopted for behavior routing: sub-organizations across levels act in turn, and sub-organizations at the same level are executed in parallel and concurrently, taking into account both task timing constraints and computational efficiency.

[0062] The technical solution provided by the application has at least the following beneficial effects:

[0063] The application combines the human-like characteristics of the large language model, configures five types of agents based on the large language model, fully excavates the application scenarios of the large language model in human decision-making simulation, combines the idea of subject modeling method, builds a multi-agent simulation system with behavior consistency, system efficiency and explainability, can simulate people's decision-making behavior through the large model when studying real problems, can replace platform research and other processes to a certain extent to analyze related problems, save platform operation cost, and at the same time can study some platform operation problems that are not easy to study through research, experiment and other methods. Compared with the current simulation system, this system combines the human-like characteristics of the large model, can more fully reflect the operation of the real world, consider more influencing factors, use more real data, simulate more realistic scenarios, and the simulation results have more reference significance for platform problem analysis. At the same time, with the help of the architecture design of the simulation system, the simulation efficiency is also higher, which can save time cost. BRIEF DESCRIPTION OF DRAWINGS

[0064] In order to more clearly illustrate the technical solutions in the embodiments of the application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the application, and other drawings can be obtained by those skilled in the art without creative labor.

[0065] Figure 1 It is a large language model multi-agent simulation system structure diagram provided by the embodiment of the application;

[0066] Figure 2 is a human-computer interaction setting module provided by an embodiment of the present application;

[0067] Figure 3 is an interaction logic diagram of various agents of a multi-agent collaboration module provided by an embodiment of the present application;

[0068] Figure 4 is a scheduling process diagram of a multi-agent scheduling system provided by an embodiment of the present application;

[0069] Figure 5 is a log causal analysis module diagram provided by an embodiment of the present application. DETAILED DESCRIPTION

[0070] In order to make the technical problems, technical solutions and advantages of the present application more clear, specific embodiments will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0071] As shown in Figure 1 , an embodiment of the present application provides a multi-agent simulation system of a large language model, the system comprising: a human-computer interaction setting module, a multi-agent collaboration module, a data bus communication module, a log causal analysis module;

[0072] The human-computer interaction setting module is used to provide an interactive interface between a user and the system, and the user sets a scene, parameters used in a simulation running process, and provides data information required for system running, combines real scene information, and sets a high-fidelity simulation scenario and parameters for specific simulation;

[0073] The multi-agent collaboration module is used for specific multi-agent simulation running based on a large language model, processes subject and environment information, and simulates a development process of real events over time;

[0074] The data bus communication module is used for communication between an agent and a database, and connects the agent and the database in a bus mode to improve information exchange efficiency;

[0075] The log causal analysis module is used for analyzing a simulation process and results.

[0076] Optionally, as shown in Figure 2 , the human-computer interaction setting module is specifically used for:

[0077] 1) Feature label selection: abstractly extracting key feature labels from the real world, selecting corresponding feature labels as research settings of the simulation system, and dividing into three aspects: subject feature setting, environment condition setting, and emergency event setting;

[0078] The main feature setting of this invention embodiment: For example, when studying the behavior of platform users, it is necessary to define and set relevant characteristics of individuals. The main features should include key feature variables involved in the research problem, such as key influencing features such as user gender, age, personality, marital status, etc.

[0079] The environmental conditions set in this embodiment of the invention mainly include roads, weather, time, location, etc. Similar to the main features, only environmental conditions related to the simulation problem need to be selected. If it is uncertain whether they will have an impact, it is recommended to add all environmental condition settings to ensure the realism and reliability of the simulation results.

[0080] System operation is not only subject to conventional conditions, but also to various unforeseen events in the real world that affect people's decisions and behaviors. This invention can set up unforeseen events related to the simulation problem, such as platform public opinion events, to simulate random disturbances in the real world.

[0081] 2) Feature distribution extraction and setting: The distribution of the selected feature labels is set to better reflect the real world or meet the needs of the research. There are two setting methods: fitting real data and using an approximate distribution of real data.

[0082] The fitting of real-world data in this embodiment of the invention: Fitting is performed using real-world data from platform users, such as obtaining basic information of registered users on the platform and fitting this information to determine the distribution of simulated agent features.

[0083] The present invention adopts an approximate distribution of real-world data: for some real-world data that is not available, the approximate distribution of the indicators can be set. For example, when it is necessary to use height, IQ, etc., which have no specific real-world data as a reference, it can be set as an approximate normal distribution and the corresponding parameters can be specified.

[0084] 3) Random combination of various features: By randomly matching different indicators, the labeling of complex individuals, environments and emergencies is improved. The random combination process will ensure that a single feature conforms to the distribution set in the previous step.

[0085] 4) Specific configuration generation: Based on the feature set generated by random combination of features, the corresponding element configuration is generated. This process is generated by the large language model. The large language model will combine the feature set and generate the corresponding configuration according to the preset configuration template, including: agent configuration, environment configuration, and event configuration.

[0086] Optionally, the multi-agent collaboration module includes five types of agents configured based on a large language model: management agent, support agent, guidance agent, execution agent, and evaluation agent;

[0087] The management agent is used to acquire task objectives and requirements and perform task planning, distribute and broadcast tasks to the execution agent, and play the role of information transmission.

[0088] The supporting agent is used to provide environmental information to the executing agent, and to perform dynamic environmental updates and real-time information feedback based on the decisions of the executing agent, and the feedback information is transmitted to the management agent.

[0089] The guiding agent is used to establish logical guiding questions for the executing agent, calculate the utility of the set of decisions to be made, and provide decision constraints, so that the executing agent can make decisions in a more logical and human-like manner.

[0090] The executing agent is the main role in the simulation, representing each unique individual in the simulation scenario. It receives information from the managing agent and the supporting agent, and makes decisions under the guidance of the guiding agent. The specific decision objects include inter-agent decision-making and agent-environment decision-making.

[0091] The evaluation agent is used to evaluate the effectiveness of task execution, generate event reports and perform real-time data statistics, update the sub-utility weights of the execution agent, and transmit the information to the management agent.

[0092] Optionally, such as Figure 3 As shown, the interaction logic of various agents in the multi-agent collaborative module is as follows:

[0093] 1) Status Update:

[0094] The executing agents transmit their respective decision information and update the state of each agent based on the decision status of all agents at the previous timestamp, including the influence between agents and the influence between agents and the environment;

[0095] This process in this embodiment of the invention uses a large language model to perform information processing such as summarization, formatting, and tagging, and then updates the data file system.

[0096] 2) Task generation and distribution:

[0097] The management agent generates specific task information for the next stage for each execution agent based on the current task and state, and then transmits and distributes the specific task information to the corresponding execution agent.

[0098] In this embodiment of the invention, the information is filtered, processed, and summarized by a large language model, and then combined with historical experience information, rule information, etc. through the RAG system to generate the decision background, action set, focus elements, and decision constraints of the intelligent agent, and the information is transmitted and distributed to the corresponding execution intelligent agent.

[0099] 3) Environmental status identification:

[0100] When the executing agent receives task information, the supporting agent queries the environmental information, identifies the current environmental state information, and transmits this information to the executing agent as part of the decision support information.

[0101] 4) Quantification of agent utility:

[0102] Based on the BBC Wellbeing Scale (BBC-SWB), a three-dimensional utility model is constructed, including three sub-utility indicators: psychological well-being (Ps), physical health (Ph), and interpersonal relationships (Re). By providing corresponding definitions, standards, and case background information, the executive agent determines the scores of each sub-utility indicator for each decision, simulating the decision-making process of normal humans considering multiple aspects before making a decision.

[0103] Most similar applications currently rely on the degree of matching between problems and decisions to select the optimal decision by quantifying the importance and correlation of problems and decisions. However, this agent decision-making method not only fails to truly quantify the real utility of agents, but also fails to explore the correlation between decisions made by agents. Therefore, this embodiment of the invention builds a three-dimensional utility model based on the BBC-SWB happiness scale.

[0104] For example, in the process of simulating platform user research, when faced with platform policies, different customer groups with different profiles (gender, income, hobbies, etc.) treat the various sub-utilities of the same platform policy differently, thus their sub-utility scores will also differ, which in turn determines the differences in decision-making. For instance, when a food delivery platform introduces policies related to riders (such as not allowing late deliveries / not allowing running red lights, etc.), the rider's compliance or non-compliance is a decision, and the utility behind this decision is different. This can be simulated using the multi-agent simulation system based on a large language model in this invention. Each executing agent simulates the rider's assessment of these policies in terms of psychological well-being, physical health, and interpersonal relationships. If they comply, fewer orders may be completed, resulting in decreased economic benefits (which are ultimately reflected in psychological well-being), making them unhappy and lowering their psychological well-being sub-utility score. However, the score for physical health sub-utility will be higher, while interpersonal relationships are less relevant, with the interpersonal relationships sub-utility score potentially being 0.

[0105] The score is influenced by the agent's role profile and current environmental background information, and changes continuously during the iteration process to better simulate the complexity of real-world individuals. Different weighting coefficients are assigned to different sub-utilities and adjusted during iteration. The total utility U for decision j on agent i at simulation round t is obtained by weighted averaging of the sub-utilities. ij(t), the entire process can be mathematically abstracted as:

[0106] For the candidate decision S of agent i in simulation round t i (t), for each decision j∈S i (t) The total utility U for agent i ij (t) includes three sub-utilities P ij (t)=(Ps ij (t),Ph ij (t),Re ij (t)) T ∈R 3 :

[0107] U ij (t)=w i (t) T P ij (t)

[0108] =α i (t)*Ps ij (t)+β i (t)*Ph ij (t)+γ i (t)*Re ij (t)

[0109] Where w i (t)=(α i (t),β i (t),γ i (t)) T Let w be the weight vector, and satisfy the condition w i (t)∈Δ3, Δ3={w≥0,1} T w=1},Ps ij Ph ij Re ij Let α represent the psychological well-being sub-utility score, physical health sub-utility score, and interpersonal relationship sub-utility score of decision j for the executing agent i, respectively. i ,β i ,γ i These are the weighting coefficients for the three sub-utilities. These weighting coefficients are calculated by the evaluation agent. After making a corresponding decision in each iteration, the evaluation agent adjusts the weighting coefficients of the sub-utilities based on the state, background information, and environmental information of each executing agent (e.g., introverts are more concerned with psychological well-being, while extroverts are more concerned with interpersonal relationships, which is reflected in the changes in the magnitudes of α and γ in utility calculation; and during the simulation process, the state of the agent also changes, for example, when tired, they are more concerned with physical health, and β will be higher). These are then stored together in the task information database as the basic information for the agent's decision-making.

[0110] 5) Decision generation based on dynamic game theory:

[0111] When making decisions, the executing agent not only considers its own situation to make the decision that maximizes its own utility, but also considers the potential decisions of other agents as background information that may affect its own decision, and performs decision generation based on dynamic game theory, including:

[0112] First, the executing agent i makes the optimal decision D based on its own information and the decision set. i (t):

[0113]

[0114] Among them, S i (t) represents the set of decisions made by agent i in simulation round t;

[0115] Then, each agent incorporates the decision information from its associated agents into its own background information, calculates the utility of each decision under the influence of other associated agents, and updates its own optimal decision.

[0116]

[0117] Among them, D 1 D 2 ,...,D n This represents the decision information of n agents associated with agent i in simulation round t;

[0118] After multiple game updates, until an equilibrium state is reached or the preset maximum number of game rounds N is reached, the final decision of each agent in the current simulation round is obtained.

[0119] In this embodiment of the invention, the equilibrium state means that during the current iteration, the optimal decision of each executing agent is unchanged from the optimal decision of the previous iteration. This indicates that the whole has reached an equilibrium state, which means that for each executing agent, if other agents do not change their decisions, it will not change its current decision either.

[0120] If there is no equilibrium state or it takes many rounds to reach the equilibrium state, it will waste a lot of time. In this case, the maximum number of rounds N will be used as a limit to prevent the simulation from stalling due to a single decision.

[0121] Configuring a large language model for each agent would far exceed hardware limits, and configuring a large language model under hardware limits would restrict the number of agents, making the problem oversimplified. To solve this problem and improve simulation efficiency, this embodiment of the invention uses a multi-agent scheduling system for scheduling.

[0122] Optionally, the multi-agent collaboration module uses a multi-agent scheduling system for scheduling, enabling the simulation system to achieve interaction of more agents while running a small number of large language models, allowing agents to make decisions more efficiently, while saving memory overhead of large language models;

[0123] The multi-agent scheduling system separates the large language model from the agents, using the large language model merely as a carrier for the agents' thinking and computation. During simulation, the large language model can be flexibly scheduled without affecting the decisions of different agents. Even if multiple agents share a large language model to process tasks, due to differences in agent profiles, perceived environmental data, problem sets, and decision sets, the specific decisions made will not have high similarity, thus maintaining the diversity and complexity of the multi-agent system.

[0124] Optionally, such as Figure 4 As shown, the scheduling process of the multi-agent scheduling system is as follows:

[0125] 1) Before the simulation begins, take into account the hardware of the running equipment and the number of simulated multi-agent settings to determine the number of large language models as the basis for the simulation.

[0126] 2) The complexity of the tasks to be processed is assessed, and the tasks are assigned to different large language models for processing. The tasks of the supporting agent and the evaluating agent have low complexity and are processed using low-parameter models; while the tasks of the guiding agent and the managing agent are relatively complex and are processed using high-parameter models. The processing of the executing agent is the main part of the simulation task. The tasks of the executing agent are further subdivided based on the number of optional decision sets in each round of decision-making. When the length of the decision set len(S) is... i When )≤3, it is processed using a low-parameter model; otherwise, it is processed using a high-parameter model.

[0127] 3) Based on the number of large language models set, the two models are processed in parallel. Task processing follows the FIFO principle, and the tasks to be processed will temporarily enter the complex / simple waiting queues respectively. If the high-parameter model is idle and has no tasks, the simple tasks can also be processed by the high-parameter model to ensure full utilization of resources. However, the reverse is not possible because the processing effect of complex tasks must be guaranteed.

[0128] Optionally, the database includes:

[0129] The background knowledge database is read-only. Its content includes background information, case information and feedback evaluation related to simulation. It is used to provide a background knowledge base for simulation and to provide corresponding background basis for the decision-making of the intelligent agent. This part of the data is provided by the user through the human-computer interaction module.

[0130] The feature settings database is read-only and contains data including: role characteristics, task summary, and action guidelines. It is used to store the character's feature settings and this part of the data is automatically generated by the human-computer interaction module in conjunction with the user data system.

[0131] The task information database has read and write permissions. The data content includes: initial commands, derived tasks, guiding questions, and weight coefficients of sub-utilities. It is used to interact with the intelligent agent, update task information and decision sets in real time, and guide the intelligent agent to establish thought chains and make decisions through guiding questions.

[0132] The external environment database has read and write permissions. The data content includes geographic information and dynamic conditions. The dynamic conditions include spatiotemporal conditions, sudden events, and a set of variable features of the agent. It is used to provide the agent with current state information for decision-making and to update the relevant conditions and simulate the development of events in a timely manner after the agent makes a decision.

[0133] The runtime log database has read and write permissions. The data content includes timestamps and event summaries, which are used to record the development process of the simulation and provide data support for subsequent process analysis.

[0134] The simulation results database has read and write permissions. The data content includes: simulation rounds, simulation results summary, and data statistics. It is used to record the simulation results status after a certain period and provide data support for the final analysis of the simulation results.

[0135] Optionally, such as Figure 5 As shown, the log causal analysis module is specifically used for:

[0136] 1) Data extraction;

[0137] Extract simulation process data and result data from the runtime log database and simulation result database;

[0138] 2) Data processing;

[0139] The simulation results are cross-sectional data after several time periods. They are structured results, which facilitates secondary analysis of the results data. However, the log information is process information during the operation of the simulation system. It contains complex and semi-structured data, which requires corresponding data processing to analyze. The processing includes two aspects: aligning and unifying various data in the log file according to the time dimension, and analyzing various events in the system evolution process to extract key events that affect the simulation operation and the corresponding observation-plan-action information under key events (key decision point extraction). Through the data processing stage, both the log information and the result information are cleaned into structured data, which is convenient for subsequent analysis and application.

[0140] 3) Data analysis;

[0141] Through the data extraction and processing described above, platform operators can analyze the decisions of micro-individuals. Based on individual observation, behavioral utility calculation, utility driver factor analysis, and subject interaction influence, a logical causal relationship about individual decisions can be constructed. This allows for analysis of why individuals choose the current decision rather than other alternative decisions when faced with specific events. By decomposing and analyzing individual data, behavioral analysis results can be obtained in four aspects: environmental factors, role attributes, subject interaction, and behavioral utility.

[0142] Optionally, the system adopts a three-element architecture of "task objective - organizational structure - behavior routing": multi-dimensional constraints and dynamic priorities are set at the top level for group tasks, which are decomposed into several sub-tasks and issued by the management agent; the whole adopts a fully connected organizational structure to ensure that there is a direct communication link between any two agents; the behavior routing adopts a hybrid strategy: cross-level sub-organizations act sequentially, and sub-organizations at the same level execute in parallel and concurrently, taking into account both task timing constraints and computational efficiency.

[0143] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A multi-agent simulation system based on a large language model, characterized in that, The system comprises a human-computer interaction setting module, a multi-agent collaborative module, a data bus communication module, and a log causal analysis module. The human-computer interaction setting module is configured to provide an interactive interface between a user and the system, and the user sets the scene and parameters used in the simulation running process through the human-computer interaction setting module, and provides data information required for system running, thereby combining real scene information to set a high-fidelity simulation scenario and parameters for specific simulation. The multi-agent collaborative module is configured to perform specific multi-agent simulation running based on a large language model, process subject and environment information, and simulate the development process of real events over time. The data bus communication module is configured to communicate between the agent and the database, and connect the agent and the database in a bus mode to improve information exchange efficiency. The log causal analysis module is configured to analyze the simulation process and results.

2. The system of claim 1, wherein, The human-computer interaction setting module is specifically configured to: 1) feature label selection: abstractly extracting key feature labels from the real world, selecting corresponding feature labels as research settings for the simulation system, and dividing into three aspects: subject feature setting, environment condition setting, and emergency event setting; 2) feature distribution extraction and setting: setting the distribution of the selected feature labels to better match the real world or meet research needs, and dividing into two setting modes: fitting real data and using an approximate distribution of real data; 3) random combination of various features: improving the label setting of complex individuals, environments, and emergency events by randomly matching different indicators, and the random combination process ensures that a single feature meets the distribution set in the previous step; 4) specific setting generation: generating corresponding element configurations according to the feature set generated by random combination of features, and this process is generated by a large language model, which generates corresponding settings according to a pre-set configuration template, including agent configuration, environment configuration, and event configuration.

3. The system of claim 1, wherein, The multi-agent collaborative module comprises five types of agents configured based on a large language model: management agent, support agent, guidance agent, execution agent, and evaluation agent. The management agent is configured to obtain task targets and requirements, plan tasks, distribute and broadcast tasks to the execution agent, and serve as an information transmission role. The support agent is configured to provide environment information to the execution agent, dynamically update the environment according to the decision of the execution agent, and provide real-time information feedback, which is transmitted to the management agent. The guidance agent is configured to establish logical guidance problems for the execution agent, calculate the utility of the decision set, and provide decision constraints to enable the execution agent to make more logical and human-like decisions. The execution agent is the main role of the simulation, representing each unique individual in the simulation scenario, receiving information from the management agent and support agent, and making decisions under the guidance of the guidance agent, with specific decision objects including inter-agent decisions and agent-environment decisions. The evaluation agent is used for effect evaluation of task execution, generates event briefs and performs real-time data statistics, updates the sub-utility weight of the execution agent, and transmits information to the management agent.

4. The system of claim 3, wherein, The interaction logic of various agents of the multi-agent cooperation module is as follows: 1) State update: The execution agent transmits respective decision information, and updates the state of each agent according to the decision of all agents at the last timestamp, including the influence between agents and the influence of the agent-environment; 2) Task generation and distribution: The management agent generates specific task information for each execution agent for the next stage according to the current task and state, and transmits and distributes the specific task information to the corresponding execution agent; 3) Environment state identification: When the execution agent receives the task information, the support agent queries the environment information, identifies the current environment state information, and transmits the information to the execution agent as part of the decision support information; 4) Agent utility quantification: According to the BBC happiness scale BBC-SWB, a three-dimensional utility model is built, including three sub-utility indexes of psychological happiness Ps, physical health Ph and interpersonal relationship Re. By giving corresponding definitions, standards and case background information, the executing intelligent agent determines the score of each sub-utility index of each decision, simulates the decision-making process of normal human beings before making decisions, and the score is affected by the intelligent agent role image and the current environmental background information, which changes in the iteration process to better simulate the complexity of real individuals. At the same time, different weight coefficients are set for different sub-utilities, and adjusted in the iteration process. The total utility U of decision j to intelligent agent i at t simulation round is obtained by weighted average of sub-utilities ij (t). The whole process is mathematically abstracted as: For each decision j e S i (t), each decision j e S i (t) for the performing agent i ij (t) includes three sub-utilities P ij (t) = (Ps ij (t), Ph ij (t), Re ij (t)) T ∈ R 3 : U ij (t) = w i (t) T P ij (t) = a i (t) * Ps ij (t) + β i (t) * Ph ij (t) + γ i (t) * Re ij (t) where w i (t) = (a i (t), β i (t), γ i (t)) T is a weight vector and satisfies the condition w i (t) e Δ3, Δ3 = {w ≥ 0, 1 T w = 1}; 5) Decision generation based on dynamic game: When the execution agent makes a decision, it not only considers its own situation to make the best decision for itself, but also considers the potential decisions of other agents as background information that may affect its decision, and generates a decision based on dynamic game, including: First, the agent i makes the optimal decision D according to its own information and decision set i (t): where S i (t) is the set of decisions of the agent i at simulation round t; Then, each execution agent adds the decision information of the associated agent to its own background information, and further calculates the utility of each decision under the influence of other associated agents, and updates its own optimal decision where D 1 ,D 2 ,...,D n denotes the decision information of the n agents associated with agent i at simulation round t. After multiple game updates, until the equilibrium state is reached or the maximum number of game rounds N is reached, the final decision of each execution agent at the current simulation round is obtained.

5. The system of claim 1, wherein, The multi-agent cooperation module uses a multi-agent scheduling system for scheduling, allowing the simulation system to interact with more agents based on a small number of large language models, allowing agents to make decisions more efficiently while saving memory costs for large language models; The multi-agent scheduling system separates the large language model from the agent, and uses the large language model only as a carrier for thinking and calculation of the agent. In the simulation process, the large language model is flexibly scheduled, and the decision of different agents is not affected. Even if multiple agents share a large language model to process tasks, due to the differences in agent portraits, perception environment data, problem sets, and decidable sets, the specific decisions made at the end will not have high similarity, and the diversity and complexity of the multi-agent system can still be maintained.

6. The system of claim 5, wherein, The scheduling process of the multi-agent scheduling system is as follows: 1) Before the simulation starts, the hardware of the running device and the number of simulation multi-agents are considered as the number of large language models as the basis for simulation; 2) judge the complexity of the task that needs to be processed, and assign it to different large language models for processing, wherein the task requirements of the support agent and the evaluation agent are not high, and are processed by a low-parameter model; while the tasks processed by the guide agent and the management agent are relatively complex, and are processed by a high-parameter model; the processing of the execution agent is the main part of the simulation task, and the task of the execution agent is secondarily shunted according to the number of the optional decision set of each round of decision, when the length of the decision set len(S i )≤3, it is processed by a low-parameter model; otherwise, it is processed by a high-parameter model; 3) According to the set number of large language models, the two models are processed in parallel. The task processing is based on the FIFO principle, and the pending tasks are temporarily placed in the complex / simple waiting queue. If the high-parameter model is idle and has no task, simple tasks can also be processed by the high-parameter model to ensure full utilization of resources. Since the processing effect of complex tasks must be guaranteed, the opposite is not true.

7. The system of claim 1, wherein, The database includes: The background knowledge database has a read-only permission, and the data content includes background information, case information and feedback evaluation related to simulation, and is used to provide a background knowledge base for simulation, and to provide a corresponding background basis for the execution of the intelligent agent decision, and the data is provided by the user through the man-machine interaction module; The feature setting database has a read-only permission, and the data content includes role features, task summaries and action guidelines, and is used to save the feature settings of the role, and the data is automatically generated by the man-machine interaction module in combination with the user data system; The task information database has a read-write permission, and the data content includes initial commands, derivative tasks, guide questions and weight coefficients of sub-utilities, and is used to interact with the intelligent agent, update the task information and decision set in real time, and make the intelligent agent establish a thinking chain and make a decision through the guide question; The external environment database has a read-write permission, and the data content includes geographical information and dynamic conditions, and the dynamic conditions include space-time conditions, unexpected events and a variable feature set of the intelligent agent, and is used to provide current state information for the intelligent agent decision and update the related condition simulation event development in time after the intelligent agent decision; The running log database has a read-write permission, and the data content includes time stamps and event briefs, and is used to record the development process of the simulation, and provide data support for subsequent process analysis; The simulation result database has a read-write permission, and the data content includes simulation rounds, simulation result summaries and data statistics, and is used to record the simulation result state after a certain period, and provide data support for the final result analysis of the simulation.

8. The system of claim 1, wherein, The log causal analysis module is specifically used for: 1) data extraction; extracting process data and result data of the simulation from the running log database and the simulation result database; 2) data processing; The simulation result is cross-sectional data after a number of time periods, which is a structured result and is convenient for secondary analysis of the result data, but the log information is process information in the running of the simulation system, which is complex and semi-structured data, and needs to be analyzed through a corresponding data processing process, which includes two aspects: aligning and unifying various data in the log file according to the time dimension, and analyzing various events in the system evolution process to extract key events affecting the running of the simulation and corresponding observation-plan-action information under the key events, and through the data processing stage, the log information and the result information are cleaned as structured data, which is convenient for subsequent analysis and application; 3) data analysis; Through the above data extraction and data processing, the platform operator can analyze the decision of the micro individual, construct the logical causal relationship of the individual decision according to the individual observation, behavior utility calculation, utility driving factor analysis and subject interaction influence, analyze why the individual chooses the current decision instead of other optional decisions when facing specific events, and obtain the behavior analysis results of the simulation running in the environment factors, role attributes, subject interaction and behavior utility through the analysis of the individual data.

9. The system of any of claims 1-8, wherein, The system adopts a ternary architecture of "mission target-organization structure-behavior routing": at the top level, a multi-dimensional constrained and dynamically prioritized group mission is set, which is decomposed into several sub-tasks by the management agent and issued; The whole adopts a full connection organization structure to ensure that there is a direct communication link between any two agents; The behavior routing adopts a hybrid strategy: cross-level sub-organizations act in sequence, and sub-organizations at the same level are executed in parallel, taking into account both task timing constraints and computational efficiency.

Images