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

Abstract

The invention relates to an artificial intelligence and multi-agent system technology, in particular to a multi-agent collaborative decision-making method and system based on a large language model, and the system comprises a perception analysis agent, a collaborative decision-making agent, a resource scheduling agent, an execution monitoring agent and a learning optimization agent. The system carries out linkage evaluation through three quantitative indexes of global reward value variance, accuracy improvement rate and cooperation strategy consistency, sets a differentiation threshold value and yellow, orange and red three-level early warning, accurately recognizes training states such as strategy oscillation, local optimum and strategy instability, triggers a progressive response from enhanced monitoring to immediate intervention, and improves the early warning efficiency. The system failure risk caused by abnormal training is obviously reduced; after the early warning is triggered, the system starts a structured four-step analysis process including anomaly positioning, causal verification, root cause determination and report generation, and outputs an executable report including an evidence chain and optimization suggestions by calling a large language model and automatically executing verification codes, so that the operation and maintenance transparency and the decision-making efficiency are improved.

Description

Technical Field

[0001] This invention relates to artificial intelligence and multi-agent system technology, specifically to a multi-agent collaborative decision-making method and system based on a large language model. Background Technology

[0002] With the rapid development of artificial intelligence technology, large language models, with their excellent natural language understanding and generation capabilities, have become the core driving force for promoting intelligent applications. In order to solve the bottleneck of single models in handling complex, dynamic, and multi-dimensional tasks, multi-agent collaborative systems have emerged. Several key technical bottlenecks still exist in the current research and application of large language model multi-agent collaborative systems: The training process lacks a multi-dimensional, phased, and refined state assessment and early warning mechanism, making it difficult to dynamically identify typical convergence anomalies such as policy oscillations, local optima, and policy instability, resulting in delayed human intervention and low training efficiency. The lack of automated integration and quality assurance of multi-source heterogeneous monitoring data makes it difficult to efficiently align and clean data from TensorBoard, vector databases and log platforms, affecting the accuracy and real-time performance of diagnostics. The abnormal diagnosis process relies heavily on human experience and lacks automated, interpretable intelligent analysis capabilities, making it unable to quickly locate the root cause and generate actionable remediation suggestions. The system lacks adaptive optimization and closed-loop recovery capabilities, and the abnormal response stops at alarms, making it difficult to achieve a complete autonomous loop from detection and diagnosis to strategy adjustment and resource reallocation. The existing architecture is rigid, lacks scalability, and has low integration with engineering toolchains, making it difficult to quickly adapt to complex collaborative scenarios in multiple fields such as financial risk control and medical decision-making. These limitations severely restrict the reliability, adaptability, and deployment efficiency of multi-agent systems in dynamic environments; To address the aforementioned technical problems, this application proposes a solution. Summary of the Invention

[0003] The purpose of this invention is to solve the problems raised in the background art and to propose a multi-agent collaborative decision-making method and system based on a large language model.

[0004] The objective of this invention can be achieved through the following technical solutions: A multi-agent collaborative decision-making method and system based on a large language model includes the following steps: Q1: Construct a multi-agent system, which includes a perception and analysis agent, a collaborative decision-making agent, a resource scheduling agent, an execution monitoring agent, and a learning and optimization agent; Q2: Within each decision-making cycle, execute the following closed-loop process: Q2-1: The perception and analysis agent will interpret the original environmental state perceived by each agent. Transform into structured natural language description ; Q2-2: Each agent is based on Generate action intentions Collaboration suggestions The collaborative decision-making agent conducts multiple rounds of consultation to generate a consistent action plan; Q2-3: The resource scheduling agent dynamically allocates computing resources based on task complexity and system load; Q2-4: Each agent executes its tasks according to the unified action plan and allocated resources, and feeds back the execution results to the central coordination module; Q2-5: The execution monitoring agent collects global reward value, task accuracy and strategy parameter data in real time, and performs training status evaluation and anomaly warning based on preset indicators; Q2-6: When an early warning is triggered, the learning optimization agent initiates the root cause analysis process and automatically performs strategy optimization, model update, or resource configuration adjustment based on the analysis conclusions.

[0005] In a preferred embodiment of the present invention, the training status evaluation in steps Q2-5 includes the coordinated evaluation of the following three indicators: Global reward variance This is used to determine strategy oscillations; Accuracy improvement rate This is used to determine local optima. Consistency of Collaboration Strategy This is used to determine the stability of a strategy.

[0006] In a preferred embodiment of the present invention, the anomaly warning in steps Q2-5 includes a three-level warning mechanism: Yellow alert: If any indicator reaches the threshold, monitoring should be intensified. Orange alert: Two indicators are abnormal at the same time, and preliminary diagnosis is initiated; Red alert: If all three indicators are abnormal or the orange alert continues for an extended period, immediate intervention is required.

[0007] In a preferred embodiment of the present invention, the root cause analysis process in steps Q2-6 includes the following steps: Preliminary preparations: Bind warning levels and abnormal indicators, automatically acquire and format key data, including global reward value, accuracy, strategy parameters, environment logs and system design information; Large language model reasoning analysis: outlier location, causal relationship mining, root cause determination and conclusion generation based on structured Prompt; Validation and Output: Perform data validation, logical validation, and supplementary validation on the analysis results to generate an executable report containing root causes, evidence, impact chains, and optimization suggestions.

[0008] As a preferred embodiment of the present invention, the method for obtaining key data in the pre-processing includes: Export global reward value data through the TensorBoard monitoring platform and calculate its variance; Access the test set evaluation log directory, extract accuracy data, and calculate the improvement rate; Connect to the Milvus vector database, query the policy parameter vectors, and calculate the cosine similarity. Search the operation log management platform to obtain hardware status, software operation, and anomaly records.

[0009] In a preferred embodiment of the present invention, the preprocessing of the global reward value data, accuracy data, and strategy parameter data in steps Q2-5 includes: Acquired at the same time Calculate the mean of each data point. with standard deviation ; Set the fluctuation range as ; Data outside the specified range is marked as outliers. If the number of outliers... If so, the data is determined to be abnormal and re-detected; Otherwise, calculate the mean after removing outliers. This serves as valid data at that moment.

[0010] As a preferred embodiment of the present invention, a method for implementing the method as described in any one of claims 1-6 includes: Perception-analysis intelligent agents are used to transform raw environmental states into structured descriptions; Collaborative decision-making agents are used to organize multi-agent consultations and generate consistent action plans. A resource scheduling agent is used to dynamically allocate computing resources; The execution monitoring agent is used to monitor the training status, evaluate metrics, and trigger alerts in real time. The learning optimizes the agent for performing root cause analysis and adaptive optimization.

[0011] In a preferred embodiment of the present invention, the execution monitoring agent includes: The data acquisition module is used to collect global reward value, accuracy, and strategy parameter data. The indicator calculation module is used to calculate... , , ; The early warning triggering module is used to trigger a three-level early warning based on a threshold. The log management module is used to store training process data and interface with TensorBoard, Milvus, and the log platform.

[0012] Compared with the prior art, the beneficial effects of the present invention are: 1. The system uses three quantitative indicators—global reward value variance, accuracy improvement rate, and consistency of collaborative strategies—for joint evaluation, and sets differentiated thresholds and three levels of warnings (yellow, orange, and red). This mechanism can accurately identify training states such as strategy oscillation, local optima, and strategy instability, and trigger a progressive response from enhanced monitoring to immediate intervention, significantly reducing the risk of system failure caused by training anomalies. 2. The system seamlessly integrates TensorBoard, Milvus, and the log platform, automatically aligning and fusing heterogeneous data from multiple sources. It adopts an outlier detection and cleaning algorithm based on the 3σ principle, automatically triggering resampling when outlier data exceeds the limit, ensuring the quality of data input to the root cause analysis module from the source, and laying a solid foundation for accurate diagnosis. 3. After the warning is triggered, the system automatically starts a structured four-step analysis process: anomaly location, causal verification, root cause determination and report generation. By calling the large language model and automatically executing the verification code, it outputs an executable report containing the evidence chain and optimization suggestions, which greatly reduces the cost of manual investigation and improves the transparency of operation and maintenance and decision-making efficiency. 4. The system dynamically allocates analysis resources and depth based on the warning level, and automatically executes strategy updates, model retraining or resource adjustments based on root cause conclusions, forming a complete autonomous closed loop of "monitoring-diagnosis-repair", which significantly shortens the anomaly recovery time and achieves continuous self-optimization of system performance; 5. Through a collaborative process of perception analysis, negotiation and decision-making, resource scheduling and professional execution, the system demonstrates its efficient collaboration capabilities in complex tasks. Its modular design supports seamless integration with existing toolchains and has the potential for rapid engineering implementation in multiple fields such as smart healthcare and scientific research decision-making. Attached Figure Description

[0013] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to the accompanying drawings.

[0014] Figure 1 This is a system architecture diagram of the present invention. Detailed Implementation

[0015] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0016] Example: Please see Figure 1 As shown, a multi-agent collaborative decision-making method and system based on a large language model includes a perception and parsing agent, a collaborative decision-making agent, a resource scheduling agent, an execution monitoring agent, and a learning and optimization agent. When conducting multi-agent collaborative decision-making based on a large language model, following the process of "basic construction - collaborative mechanism design - training optimization - deployment verification", it is necessary to focus on the core risk points of each link and judge process anomalies through quantifiable data indicators. In the entire process of multi-agent collaborative decision-making based on large language models, the training and optimization phase is most prone to anomalies. The core reason is that the training and optimization phase is a deeply coupled link of "algorithm, model, and collaborative logic" - it not only needs to integrate the semantic reasoning ability of LLM with the policy optimization logic of multi-agent reinforcement learning, but also needs to solve multiple complex problems such as training convergence, reward mechanism design, and reflection to verify effectiveness. Moreover, anomalies often have "concealment" (such as training non-convergence is difficult to detect in the early stage) and "transmission" (training defects will directly lead to the failure of deployment and verification). Training convergence anomalies are a core risk in the training optimization phase. Corresponding evaluation metrics need to be matched according to the training cycle (early / mid / mid-late stages). The core objective is to identify three types of problems in a timely manner: "policy oscillation, performance stagnation, and policy instability" to ensure the effectiveness and generalization of multi-agent collaborative strategies. Training convergence anomalies are the most common problem in the training optimization phase, and their core manifestations are "policy oscillation, local optima, and poor stability." They can be quantified through reward value, accuracy, and policy consistency.

[0017] At the start of each decision cycle, each agent Perceive at any moment The original environmental state at that time (Such as corporate financing application texts, financial statement images, and market data streams) are input into a lightweight state interpretation module, which calls a large language model to interpret the original environment state. Transform into a structured natural language description (e.g., "Task: Financial Analysis; Input: 2023 Income Statement; Output Requirements: Solvency Rating; Priority: High").

[0018] Based on semantic state In accordance with their own goals, each agent invokes a large language model using an intent generation prompt template to generate its action intent for the current round. and requests or suggestions for cooperation with other intelligent agents. (For example, the financial agent might suggest that the credit agent provide the company's repayment records for the past three years); all intentions and suggestions generated by agents. All proposals are aggregated into a central coordination module, which then calls a large language model to comprehensively analyze all proposals, identify potential conflicts, and generate one or more rounds of negotiation instructions. The agreed actions reached through negotiation are then executed by the individual agents.

[0019] Before task execution, the system dynamically allocates computing resources (GPU, memory, bandwidth) based on task complexity, deadline, and agent load. The resource scheduling module uses an improved particle swarm optimization algorithm to solve for the optimal resource allocation scheme with the goal of minimizing task completion time and maximizing resource utilization, ensuring that high-priority tasks receive sufficient resources first.

[0020] Each specialized intelligent agent, based on the consensus-based decision and allocated resources, invokes the corresponding domain model or external tools to perform tasks. For example, the financial analysis intelligent agent invokes the financial ratio calculation model, the credit intelligent agent queries the enterprise credit database, and returns the execution results (such as "debt repayment ability: medium risk; confidence level: 85%) to the central coordination module.

[0021] Before each training round, the system pre-defines the "global reward value calculation logic" within the multi-agent reinforcement learning framework code. After each training round, the framework code automatically records the "global reward value of the current round" and writes it to a local log file in the format of "round + reward value" for easy batch reading later. During training, the system calls the "Add Scalar" interface provided by the TensorBoard tool to immediately send the calculated global reward value of each round to the TensorBoard monitoring panel. After training is complete, the system directly calls the TensorBoard history function to export global reward value data for K consecutive rounds, which can then be used to calculate the global reward value variance. The global reward value data obtained in both methods are mapped one-to-one by timestamp, and the global reward value data obtained at the corresponding time is preprocessed. The preprocessed data is then used to... The mean is calculated and used as the global reward value data for the corresponding timestamp. In the training code, the "task completion judgment rule" is defined in advance. After the "task execution phase" of each training round, the code automatically counts the "number of successful tasks in this round" and the "total number of tasks in this round". The accuracy of the current round is calculated by "number of successful tasks ÷ total number of tasks". The "rounds + accuracy" were recorded accordingly. To avoid overfitting of the training data and resulting in distorted accuracy, a fixed test set was used to evaluate the performance of the current strategy every 100 rounds, and the accuracy at that time point was calculated as "number of successes ÷ total number of tasks". The accuracy data obtained from the two methods were mapped one-to-one by timestamp, and the accuracy data obtained at the corresponding time point was preprocessed. The preprocessed data was then used to... The mean is calculated and used as the accuracy data for the corresponding timestamp. Before training begins, the parameter format of the policy model needs to be confirmed, and the logic for "converting parameters into vectors" should be written in the code beforehand. During training, after each round of training, the policy model parameter tensors of each agent in the current round are directly read through the code; these parameter tensors are converted into fixed-length vectors through the code, and the "insertion interface" of the Milvus vector database is called to store "round number + agent ID + parameter vector" into the database. When it is necessary to calculate the consistency of the cooperative policy, simply use Milvus's "similarity query function" and input "round number"... Wheel Intelligent Agent "parameter vector", query "the first" Wheel Intelligent Agent cosine similarity of the parameter vectors ; The acquired data is sorted according to the acquisition time, and corresponding items acquired at the same time are sorted. averaging the data and standard deviation The calculation, and the mean obtained from the calculation. and standard deviation Obtain the data fluctuation range for the corresponding items. The system is configured to compare the acquired data for a given item with its fluctuation range, mark data outside the fluctuation range as outliers, and record the number of outliers. ,like If the data is found to be abnormal, the data will be re-tested; if If outliers are removed, the mean of the remaining corresponding test data after outlier removal is calculated. The calculation, and the mean obtained from the calculation. As the corresponding data detected at the corresponding time; Global reward variance ,in For continuous training rounds, For the first Round global reward value, for The average reward per round; this metric is used to assess volatility from the early to mid-stages of training: at this point, the multi-agent system is in the policy exploration phase, and if this occurs continuously for 1000 rounds ( The variance of reward value This indicates that the fluctuation range of the global reward exceeds the reasonable range, and the multi-agent cooperative strategy has not formed a stable optimization direction. Therefore, it is judged as training non-convergence and strategy oscillation. Accuracy improvement rate ,in For the first Task completion accuracy This represents the accuracy rate 100 rounds ago; this metric is used to determine performance stagnation in the later stages of training: at this point, the multi-agent strategy has been initially formed, and if the accuracy rate of the first 100 rounds is low... The difference in accuracy between round 100 and round 100 Corresponding improvement rate This indicates that the strategy optimization has reached a bottleneck and cannot further improve the quality of task completion through iteration. Therefore, it is determined that the training has fallen into a local optimum. Consistency of Collaboration Strategy ,in For the number of agents, For the first Wheel Intelligent Agent The policy parameters; this metric is used to evaluate the policy stability over the entire training cycle: where Indicates the first Wheel and 50 before intelligent agents The policy parameter cosine similarity (values ​​from 0 to 1, the closer to 1, the more consistent the policies), if Mean policy similarity among agents This indicates that the collaboration strategies in different rounds differ too much, and the multi-agents cannot form a stable collaboration pattern. Therefore, it is judged as having unstable strategies and poor generalization ability. The above three indicators need to be evaluated in conjunction: if the policy oscillation in the early stage of training is not intervened in time, it may lead to local optima in the middle and late stages; while policy instability will further aggravate the fluctuation of reward value, ultimately causing the multi-agent collaborative strategy to fail to adapt to the needs of the actual scenario. If any single indicator reaches the threshold, it is determined that there may be a risk, monitoring should be strengthened, and a yellow warning should be issued; if two indicators are abnormal at the same time, it is determined that there is a high suspicion of convergence problem, preliminary diagnosis should be initiated, and an orange warning should be issued; if all three indicators are abnormal, or the orange warning status continues for more than the preset number of rounds, it is determined that there is serious training non-convergence, immediate intervention is required, and a red warning should be issued.

[0022] After an alert is triggered, root cause analysis preparation begins. First, the alert level and abnormal indicators are linked. The large language model identifies the abnormal indicator combination, threshold standards, and priorities corresponding to the current alert level by reading the "Multi-Agent Collaborative Decision System Alert Rule Comparison Table." It automatically records the alert level, trigger time, core analysis indicators, and analysis objectives in a newly created analysis document, achieving rapid location and documentation of abnormal indicators. Next, key data is filtered and formatted. The large language model exports global reward value data within a specified time range by calling the TensorBoard monitoring platform, uses built-in calculation tools to calculate the overall variance, and automatically describes the calculation logic and fluctuation characteristics of the reward values, achieving quantitative analysis of reward value anomalies. Finally, accuracy data is extracted. The large language model accesses the test set evaluation log directory on the server, automatically obtains evaluation reports from different rounds, and extracts task details and results. The system performs numerous training iterations, calculates the accuracy improvement rate, and records it in the analysis document, achieving precise calculation of accuracy changes. Next, it extracts policy parameter data. The large language model connects to the Milvus vector database, automatically queries and exports policy parameter vectors for a specified number of rounds, calculates the similarity of policy parameters between agents using the cosine similarity algorithm, filters out abnormal data, and supplements parameter meaning explanations, achieving a quantitative assessment of policy consistency. Then, it extracts environment and operation logs. The large language model accesses the operation log management platform, automatically retrieves key records such as hardware status, software operations, and error reports by time range, organizes and records them in the analysis document, achieving comprehensive monitoring of the training environment. Finally, it injects domain knowledge. The large language model automatically adds a system design description section to the analysis document, writing key information such as agent architecture, normal indicator ranges, and core operating rules, achieving a complete construction of the analysis background. After completing the preliminary preparations, the LLM root cause analysis process is initiated. First, anomaly identification is performed. The large language model reads the core analysis objectives, key data, and system design specifications from the analysis document, and, in accordance with the structured Prompt requirements, automatically outputs a list of anomaly indicators, anomaly time correlations, and a preliminary summary of anomaly characteristics in a fixed format, achieving automated identification and structured output of anomalies. Second, causal relationship mining is conducted. Based on the identified anomalies, the large language model automatically proposes causal relationship hypotheses and generates verification code. The generated Python code is executed to verify the correlation between high GPU load periods and agent parameter update failures, and the hypothesis is automatically judged based on the verification results, achieving hypothesis proposal and data verification of causal relationships between anomalies. Third, the root cause is determined. After verifying the causal relationships, the large language model further conducts in-depth analysis, combining the system design specifications and supplementary agent resource configuration information, automatically outputting the root cause of the anomalies, achieving precise root cause identification. Fourth, interpretable conclusions and recommendations are generated. According to the structured Prompt requirements, the large language model automatically outputs root cause conclusions, key evidence, impact chains, optimization suggestions, and easy-to-understand analogies in a fixed format, achieving a comprehensive presentation and comprehensibility of the analysis conclusions. To ensure the reliability of the root cause conclusions, the large language model performs three steps: data verification, logical verification, and supplementary verification. It automatically checks the hardware resource configuration table, parameter update logs, and recalculates key indicator data to verify the accuracy of the cited data. By simulating the audit process and combining domain knowledge and system design logic, it automatically audits the analysis chain to confirm the absence of logical loopholes. Furthermore, it automatically verifies the impact of potential influencing factors on the anomalies, supplementing the analysis conclusions to the final report, thus achieving reliable and comprehensive verification of the analysis results.

[0023] The analysis strategy is adjusted according to different warning levels. For yellow warnings, the large language model uses lightweight computing resources to read only core data and basic system information, and automatically and quickly outputs possible causes and brief suggestions using a concise prompt, achieving rapid preliminary analysis of low-priority anomalies. For orange warnings, the large language model uses medium computing resources to automatically filter key data and perform correlation analysis according to the pre-preparation process, outputting complete root cause conclusions, key evidence, and actionable suggestions, achieving complete analysis of medium-priority anomalies. For red warnings, the large language model uses high-performance computing resources to fully execute the pre-preparation, four-step reasoning analysis, and multiple verification processes, automatically outputting a complete actionable solution including root causes, evidence, impact chains, optimization suggestions, and emergency measures, and automatically compiling an emergency handling report and synchronizing it with relevant teams, achieving in-depth analysis and emergency response of high-priority anomalies. Based on the root cause analysis results, the system automatically performs optimization operations such as strategy updates, model retraining, and resource allocation adjustments, and enters the next decision-making cycle, forming a closed-loop learning process of "perception-decision-execution-monitoring-optimization".

[0024] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to any specific implementation. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A multi-agent collaborative decision-making method based on a large language model, characterized in that, Includes the following steps: Q1: Construct a multi-agent system, which includes a perception and analysis agent, a collaborative decision-making agent, a resource scheduling agent, an execution monitoring agent, and a learning and optimization agent; Q2: Within each decision-making cycle, execute the following closed-loop process: Q2-1: The perception and analysis agent will interpret the original environmental state perceived by each agent. Transform into structured natural language description ; Q2-2: Each agent is based on Generate action intentions Collaboration suggestions The collaborative decision-making agent conducts multiple rounds of consultation to generate a consistent action plan; Q2-3: The resource scheduling agent dynamically allocates computing resources based on task complexity and system load; Q2-4: Each agent executes its tasks according to the unified action plan and allocated resources, and feeds back the execution results to the central coordination module; Q2-5: The execution monitoring agent collects global reward value, task accuracy and strategy parameter data in real time, and performs training status evaluation and anomaly warning based on preset indicators; Q2-6: When an early warning is triggered, the learning optimization agent initiates the root cause analysis process and automatically performs strategy optimization, model update, or resource configuration adjustment based on the analysis conclusions.

2. The multi-agent cooperative decision-making method based on a large language model according to claim 1, characterized in that, The training status evaluation described in steps Q2-5 includes a coordinated evaluation of the following three metrics: Global reward variance This is used to determine strategy oscillations; Accuracy improvement rate This is used to determine local optima. Consistency of Collaboration Strategy This is used to determine the stability of a strategy.

3. The multi-agent collaborative decision-making method based on a large language model according to claim 1, characterized in that, The anomaly warning mentioned in step Q2-5 includes a three-level warning mechanism: Yellow alert: If any indicator reaches the threshold, monitoring should be intensified. Orange alert: Two indicators are abnormal at the same time, and preliminary diagnosis is initiated; Red alert: If all three indicators are abnormal or the orange alert continues for an extended period, immediate intervention is required.

4. The multi-agent collaborative decision-making method based on a large language model according to claim 1, characterized in that, The root cause analysis process described in steps Q2-6 includes the following steps: Preliminary preparations: Bind warning levels and abnormal indicators, automatically acquire and format key data, including global reward value, accuracy, strategy parameters, environment logs and system design information; Large language model reasoning analysis: outlier location, causal relationship mining, root cause determination and conclusion generation based on structured Prompt; Validation and Output: Perform data validation, logical validation, and supplementary validation on the analysis results to generate an executable report containing root causes, evidence, impact chains, and optimization suggestions.

5. The multi-agent collaborative decision-making method based on a large language model according to claim 4, characterized in that, The methods for obtaining key data in the preliminary preparation include: Export global reward value data through the TensorBoard monitoring platform and calculate its variance; Access the test set evaluation log directory, extract accuracy data, and calculate the improvement rate; Connect to the Milvus vector database, query the policy parameter vectors, and calculate the cosine similarity. Search the operation log management platform to obtain hardware status, software operation, and anomaly records.

6. The multi-agent collaborative decision-making method based on a large language model according to claim 1, characterized in that, The preprocessing of the global reward value data, accuracy data, and strategy parameter data mentioned in steps Q2-5 includes: Acquired at the same time Calculate the mean of each data point. with standard deviation ; Set the fluctuation range as ; Data outside the specified range is marked as outliers. If the number of outliers... If so, the data is determined to be abnormal and re-detected; Otherwise, calculate the mean after removing outliers. This serves as valid data at that moment.

7. A multi-agent collaborative decision-making system based on a large language model, characterized in that, For implementing the method as described in any one of claims 1-6, comprising: Perception-analysis intelligent agents are used to transform raw environmental states into structured descriptions; Collaborative decision-making agents are used to organize multi-agent consultations and generate consistent action plans. A resource scheduling agent is used to dynamically allocate computing resources; The execution monitoring agent is used to monitor the training status, evaluate metrics, and trigger alerts in real time. The learning optimizes the agent for performing root cause analysis and adaptive optimization.

8. The multi-agent collaborative decision-making system based on a large language model according to claim 7, characterized in that, The execution monitoring agent includes: The data acquisition module is used to collect global reward value, accuracy, and strategy parameter data. The indicator calculation module is used to calculate... , , ; The early warning triggering module is used to trigger a three-level early warning based on a threshold. The log management module is used to store training process data and interface with TensorBoard, Milvus, and the log platform.

Images