An agent-based model optimization method and apparatus

Abstract

The application relates to the technical field of artificial intelligence, in particular to a model optimization method and device based on an intelligent agent. A model optimization instruction is understood by an intention understanding intelligent agent to obtain an intelligent agent scheduling task instruction, a target intelligent agent is called in sequence according to the scheduling sequence of the intelligent agent in the intelligent agent scheduling task instruction, data to be processed is sent to the target intelligent agent to complete output processing until a model parameter adjustment suggestion output by the last intelligent agent in the scheduling sequence is obtained, and parameters in a to-be-optimized model are adjusted based on the model parameter adjustment suggestion. Since the callable intelligent agents include a data query intelligent agent, a code execution intelligent agent, a feature mining intelligent agent, a model evaluation intelligent agent and a model optimization intelligent agent, when the model is optimized, automatic data query, feature analysis, model evaluation and optimization operations can be performed based on the callable intelligent agents, and the efficiency of model optimization is improved.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a model optimization method and apparatus based on intelligent agents. Background Technology

[0002] Timely model optimization and evaluation are crucial for ensuring the accuracy of business operations. For example, in preventing malicious user entrapment, model evaluation and optimization are key steps in ensuring the model can identify and block malicious behavior. Existing models for preventing malicious user entrapment typically identify potential numbers or behaviors by comprehensively analyzing multi-dimensional feature data such as call records, SMS content, and network behavior.

[0003] In related technologies, automated scripts are typically written to calculate common characteristics of the numbers involved in a case (such as the number of calls, call duration, active cities, and number of contacts) to reduce the workload of manual statistics, thereby assisting in model evaluation and rule adjustment. Although existing technologies can support the evaluation and optimization of models for preventing malicious user inducement to a certain extent, there are still significant shortcomings. Writing automated scripts requires time and high skill levels, and it cannot quickly identify scenarios or predict new types of malicious user inducement behavior. It also cannot meet the characteristics of rapidly changing business scenarios and high timeliness, exhibiting significant limitations.

[0004] Therefore, how to quickly complete model evaluation and optimization has become an urgent problem to be solved. Summary of the Invention

[0005] This application provides a model optimization method and apparatus based on intelligent agents to solve the problems of low efficiency in model evaluation and optimization in the prior art, which cannot meet the characteristics of rapid scene changes and strong timeliness.

[0006] This application provides an agent-based model optimization method, the method comprising: Upon receiving a model optimization instruction, the model optimization instruction includes information about the model to be optimized and the optimization requirements; The intent-understanding agent interprets the model optimization instructions to obtain agent scheduling task instructions. The agent scheduling task instructions describe the order in which agents need to be called to complete the model optimization instructions, and the tasks to be completed by each agent. The agents that can be called include: data query agent, code execution agent, feature mining agent, model evaluation agent, and model optimization agent. According to the scheduling order of the agents in the agent scheduling task instruction, the target agents are called in turn, and the data that the target agents need to process is sent to the target agents so that the target agents can complete the tasks they need to complete, until the model parameter adjustment suggestions output by the last agent in the scheduling order are obtained; Based on the proposed model parameter adjustment, the parameters in the model to be optimized are adjusted.

[0007] This application also provides an agent-based model optimization device, the device comprising: A receiving module is used to receive a model optimization instruction, which includes information about the model to be optimized and optimization requirements. The intent understanding module is used to understand the model optimization instructions using an intent understanding agent to obtain an agent scheduling task instruction. The agent scheduling task instruction describes the order in which agents need to be called to complete the model optimization instructions, and the tasks to be completed by each agent. The agents that can be called include: data query agent, code execution agent, feature mining agent, model evaluation agent, and model optimization agent. The optimization module is used to sequentially call the target agent according to the scheduling order of the agents in the agent scheduling task instruction, send the data that the target agent needs to process to the target agent, so that the target agent can complete the task it needs to complete, until the model parameter adjustment suggestion output by the last agent in the scheduling order is obtained; and adjust the parameters in the model to be optimized based on the model parameter adjustment suggestion.

[0008] This application also provides an electronic device, which includes a processor for executing a computer program stored in a memory to implement the steps of the agent-based model optimization method as described above.

[0009] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of any of the agent-based model optimization methods described above.

[0010] This application also provides a computer program product, which includes computer program code. When the computer program code is run on a computer, it causes the computer to perform the steps of any of the above-described agent-based model optimization methods.

[0011] In this embodiment, upon receiving a model optimization instruction, an intent-understanding agent is used to interpret the instruction, resulting in an agent scheduling task instruction describing the order of agents to be invoked to complete the instruction and the tasks to be performed by each agent. Following the scheduling order of the agents in this instruction, the target agents are invoked sequentially, and the data to be processed by each agent is sent to them so that they can perform their tasks. This continues until the model parameter adjustment suggestion output by the last agent in the scheduling order is obtained. Based on this suggestion, the parameters in the model to be optimized are adjusted. Since the invoked agents include data query agents, code execution agents, feature mining agents, model evaluation agents, and model optimization agents, automatic data querying, feature analysis, model evaluation, and optimization operations can be performed based on these invoked agents when model optimization is required, thus improving the efficiency of model optimization. Attached Figure Description

[0012] To more clearly illustrate the technical solutions of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 A flowchart of an agent-based model optimization process is provided for embodiments of this application; Figure 2 A flowchart of another agent-based model optimization process provided in this application embodiment; Figure 3 A schematic diagram of a model optimization device based on an intelligent agent provided in an embodiment of this application; Figure 4 This is a schematic diagram of an electronic device structure provided in an embodiment of this application. Detailed Implementation

[0014] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art are within the scope of protection of this application.

[0015] The following explanation uses the business area of ​​preventing malicious user inducement as an example. In practical applications, model optimization and evaluation face difficulties due to heavy reliance on manual analysis and the high difficulty of feature mining. To address these issues, related technologies typically employ automated scripts to calculate common characteristics of involved phone numbers (such as call count, call duration, active cities, and number of contact persons) to reduce the workload of manual analysis, thereby assisting in model evaluation and rule adjustment. Alternatively, knowledge graph technology can be introduced to perform structured modeling and correlation analysis of information such as suspicious phone numbers, device numbers, accounts, regions, and behavioral paths. The weights of node relationships reflect the network structure characteristics of malicious user inducement, thus aiding in model optimization.

[0016] Although the relevant technologies can support the evaluation and optimization of models to a certain extent, there are still obvious shortcomings, mainly in the following aspects: (1) The results of manual analysis are used to discover common features, but it is not possible to quickly identify malicious user inducement scenarios or predict new malicious user inducement behaviors. It cannot meet the characteristics of rapid scenario changes and strong timeliness, and has significant limitations.

[0017] (2) Feature association analysis based on knowledge graphs can reveal organizational structure and behavioral relationships, but its construction and updating rely on a large amount of manual annotation and expert experience, resulting in high maintenance costs and difficulty in achieving dynamic adaptive optimization of the model. In addition, knowledge graphs usually emphasize static relationships and have limited ability to predict the real-time evolution and future trends of maliciously induced user behavior, making it difficult to respond quickly to changes in the scenario.

[0018] To address the aforementioned issues, this application provides an agent-based model optimization method, apparatus, and device. The method receives a model optimization instruction, which includes information about the model to be optimized and optimization requirements. An intent-understanding agent interprets the model optimization instruction to obtain an agent scheduling task instruction. This instruction describes the order in which agents are invoked to complete the model optimization instruction, and the tasks each agent needs to perform. The invoked agents include: a data query agent, a code execution agent, a feature mining agent, a model evaluation agent, and a model optimization agent. Following the scheduling order of the agents in the agent scheduling task instruction, target agents are invoked sequentially. Data to be processed by each target agent is sent to it, enabling it to perform its required tasks, until the model parameter adjustment suggestion output by the last agent in the scheduling order is obtained. The parameters in the model to be optimized are then adjusted based on the model parameter adjustment suggestion.

[0019] The agent-based model optimization method provided in this application, upon receiving a model optimization instruction, utilizes an intent-understanding agent to comprehend the instruction, obtaining an agent scheduling task instruction describing the order of agents to be invoked to complete the optimization and the tasks to be performed by each agent. This eliminates the need for manual analysis of the model to be evaluated. Following the scheduling order of the agents in the task instruction, the target agents are invoked sequentially, and the data to be processed by each agent is sent to it, enabling it to complete its tasks. This process continues until the last agent in the scheduling order outputs a model parameter adjustment suggestion, which is then used to adjust the parameters in the model to be optimized. Since agents can invoke large models with powerful semantic understanding and data processing capabilities, the agents that can be invoked in this application include data query agents, code execution agents, feature mining agents, model evaluation agents, and model optimization agents. When model optimization is required, agents with different functions automatically analyze and evaluate the relevant information of the model to be optimized. Compared with the feature association analysis method based on knowledge graphs in related technologies, the evaluation and optimization process is more flexible and capable of adaptive optimization. Furthermore, large models process data from a wide range of sources and update at a fast rate. Therefore, the multi-agent model optimization method in this application can meet the characteristics of rapid scene changes and high timeliness, enabling rapid response to scene changes. The following describes the agent-based model optimization process in detail with reference to various embodiments.

[0020] Example 1: Figure 1 A flowchart of an agent-based model optimization process provided for embodiments of this application is shown below. Figure 1 As shown, the process includes the following steps: S101: Received a model optimization instruction, which includes information about the model to be optimized and optimization requirements.

[0021] The model optimization method based on intelligent agents provided in this application is applied to electronic devices, such as servers and PCs.

[0022] In this embodiment of the application, to improve the efficiency of model optimization, a model optimization instruction can be received. This instruction may include information about the model to be optimized and optimization requirements. Subsequently, the model requiring optimization can be located based on this information. The information about the model to be optimized may include the model name, model storage location, model function description, etc. The optimization requirements may include improving model precision, recall, processing efficiency, etc.

[0023] In this embodiment, the model optimization command can be sent by the user of the electronic device, by other electronic devices, or automatically triggered by the electronic device at a preset time interval. That is, the automatic optimization time for each model is pre-stored, and the model optimization command is automatically triggered when the automatic optimization time is reached.

[0024] S102: The intent-understanding agent understands the model optimization instruction to obtain the agent scheduling task instruction. The agent scheduling task instruction describes the order of agents to be called to complete the model optimization instruction and the task to be completed by each agent. The agents that can be called include: data query agent, code execution agent, feature mining agent, model evaluation agent, and model optimization agent.

[0025] To understand model optimization instructions, an intent-understanding agent can be deployed in this embodiment. This agent receives model optimization requests from business personnel, parses the input semantics, and identifies business objectives, data types, model categories, and optimization goals. It then combines this with a built-in anti-fraud knowledge base to structure the model optimization instructions and generate agent task scheduling instructions, providing a basis for subsequent agent scheduling and task allocation.

[0026] In this embodiment, after receiving a model optimization instruction, an intent-understanding agent can process the instruction. Since the optimization instruction contains information about the model to be optimized and its optimization requirements, the large model deployed within the intent-understanding agent can analyze the performance of the current model to be optimized based on this information and determine how to optimize it according to the optimization requirements. In this embodiment, the intent-understanding agent can output the analysis results in an agent scheduling task instruction.

[0027] The task scheduling instruction for the intelligent agent describes the order in which the intelligent agents to be invoked are needed to complete the model optimization instruction, and the tasks to be performed by each intelligent agent. In this embodiment, the intelligent agents that can be invoked may include data query agents, code execution agents, feature mining agents, model evaluation agents, and model optimization agents.

[0028] For example, if the model optimization instruction is: optimize the XX recognition model for a certain city to improve accuracy, then the determined agent scheduling task instruction could be: to address the optimization needs of the XX recognition model, the data query agent can be invoked to obtain call data of XX numbers for the past 3 months in a certain city, the code execution agent can be invoked to generate data processing code, the model evaluation agent can calculate the current model accuracy, and so on.

[0029] Among them, the data query agent can be used to automatically generate query commands for the required data based on the received task requirements, and automatically execute them on the data platform to obtain the target data.

[0030] The code execution agent can generate and execute compliant running code based on the results returned by the data query agent and the data requirements.

[0031] Feature mining agents can extract features from multiple dimensions based on the analysis results generated by the code-execution agent.

[0032] The model evaluation agent can evaluate the performance and key metrics of the model to be optimized based on the features extracted by the feature mining agent.

[0033] The model optimization agent can comprehensively analyze the model parameter adjustment suggestions based on the model evaluation results and the features extracted by the feature mining agent.

[0034] S103: According to the scheduling order of the agents in the agent scheduling task instruction, the target agents are called in turn, and the data to be processed by the target agents is sent to the target agents so that the target agents can complete the tasks they need to complete, until the model parameter adjustment suggestions output by the last agent in the scheduling order are obtained.

[0035] After receiving the agent scheduling task instruction, the agents can be scheduled sequentially according to the scheduling order described in the instruction. For example, based on the task requirements in the instruction, each relevant agent can send an initialization instruction, including the task ID, execution criteria, and data interface parameters. Upon receiving the instruction, each agent loads its corresponding knowledge base (e.g., a data query agent loads an SQL template library, a code execution agent loads an analysis code template library), initializes its execution memory, and clears residual data from historical tasks that exceed a preset time range, preparing for collaborative work.

[0036] In this embodiment, target agents can be invoked sequentially according to the scheduling order. The data that the target agent needs to process is then sent to the target agent, enabling it to perform its required tasks.

[0037] After obtaining the data output by the target agent, the next agent can be identified as the target agent. The process continues by sending the data required for processing to the target agent, until the model parameter adjustment suggestion is obtained from the last agent in the agent scheduling sequence.

[0038] For example, if the scheduling order is: Agent A -> Agent B -> Agent A -> Agent D, then upon the first invocation of an agent, the first "Agent A" in this scheduling order can be identified as the target agent. After "Agent A" completes its required task, "Agent B" can be identified as the target agent. After "Agent B" completes its required task, the next agent "Agent A" can be identified as the target agent. And within "Agent A", the next agent "Agent D" can be identified as the target agent.

[0039] S104: Adjust the parameters in the model to be optimized based on the model parameter adjustment suggestions.

[0040] In this embodiment of the application, after obtaining the model parameter adjustment suggestion, the parameters in the model to be optimized can be adjusted based on the model parameter adjustment suggestion to obtain the optimized model.

[0041] In this embodiment, upon receiving a model optimization instruction, an intent-understanding agent interprets the instruction to obtain an agent scheduling task instruction describing the order of agents to be invoked to complete the optimization and the tasks to be performed by each agent. This eliminates the need for manual analysis of the model to be evaluated. Following the scheduling order of the agents in the task scheduling instruction, the target agents are invoked sequentially, and the data to be processed by each agent is sent to them to enable them to complete their tasks. This continues until the last agent in the scheduling order outputs a model parameter adjustment suggestion, which is then used to adjust the parameters in the model to be optimized. Since the agents can invoke large models with powerful semantic understanding and data processing capabilities, in this embodiment, the invoked agents include data query agents, code execution agents, feature mining agents, model evaluation agents, and model optimization agents. When model optimization is required, agents with different functions automatically analyze and evaluate the relevant information of the model to be optimized. Compared with the feature association analysis method based on knowledge graphs in related technologies, the evaluation and optimization process is more flexible and capable of adaptive optimization. Moreover, the large model processes data from a wide range of sources and updates rapidly. Therefore, the model optimization method based on multi-agents in this application can meet the characteristics of rapid scene changes and strong timeliness, and can respond quickly to scene changes.

[0042] Example 2: To improve the model optimization effect, based on the above embodiments, in this embodiment, the step of using an intent-understanding agent to understand the model optimization instructions and obtain agent scheduling task instructions includes: The intent-understanding agent uses a large-scale corpus analysis model trained with fine-tuning to extract the target intent of the model's optimized instructions. The target intent and task template library are input into the corpus analysis model. The corpus analysis model determines the target task template in the task template library that matches the target intent, and determines the agent's scheduling task instruction based on the target task template and the target intent.

[0043] To automatically understand model optimization instructions, in this embodiment, the intent-understanding agent can invoke a large corpus analysis model during task processing, thereby understanding the model optimization instructions based on this large corpus analysis model. This large corpus analysis model can be a large model fine-tuned using corpus data from a target domain. The target domain is the business domain of the data processed by the model to be optimized.

[0044] When using intent-understanding agents to determine task instructions for agent scheduling, semantic parsing of the model optimization instructions can be performed first. Specifically, the target intent included in the model optimization instructions can be extracted from the large model by analyzing the corpus trained through fine-tuning.

[0045] Specifically, the QWQ32B model can be used. This model is fine-tuned using anti-fraud corpus and is specifically optimized for the professional terminology and requirement expressions in anti-fraud business scenarios. Upon receiving a model optimization instruction, the model can understand the core keywords (i.e., target intent) within that instruction. These core keywords include business objectives (model optimization), geographic scope (a city), model type (malicious access), model function (identification), and optimization metrics (accuracy). In other words, in this embodiment, preset intent items are pre-configured. When extracting the target intent included in the model optimization instruction using a large corpus analysis model, the target intent corresponding to each preset intent item can be determined. In the above example, "business objectives," "geographic scope," "model type," "model function," and "optimization metrics" are the preset intent items.

[0046] To accurately and quickly generate intelligent agent scheduling task instructions, a task template library can be pre-configured in this embodiment. This task template library includes task templates corresponding to different intentions. The task template can adopt a JSON structured format, specifying fields such as task ID, execution steps, involved intelligent agent type, data requirements, and output standards.

[0047] After obtaining the target intent, the target intent and the task template library can be input into the corpus analysis model. The model then searches the task template library for a target task template that matches the target intent, and determines the agent's scheduling task instruction based on the target task template and the target intent. Specifically, a match is determined when any task template includes all or part of the target intent. In this embodiment, task templates corresponding to different intents / intent combinations can be pre-saved. After determining the target intent, the saved task templates for that intent can be identified as the matching target task template. In this embodiment, when determining the matching target task template, the corpus analysis model can also analyze the received target intent and then search the task template library for semantically similar task targets as matching target task templates.

[0048] To further improve the model optimization effect, based on the above embodiments, in this embodiment of the application, after extracting the target intent of the model optimization instructions using the fine-tuned large-scale corpus analysis model, and before inputting the target intent and task template library into the large-scale corpus analysis model, the method further includes: The target intent and the pre-configured standard business processing logic are input into the corpus analysis model, which then determines whether the currently extracted target intent meets the requirements of the standard business processing logic. If so, the subsequent steps of inputting the target intent and task template library into the corpus analysis model continue.

[0049] To further improve the model optimization effect, after obtaining the target intent, before inputting the target intent and task template library into the large corpus analysis model, an intent verification mechanism can be activated to compare the extracted demand features (i.e., target intent) with the pre-configured standard business processing logic to see if they are reasonable and complete.

[0050] In this embodiment, the target intent and pre-configured standard business processing logic can be input into the corpus analysis model, so that the corpus analysis model can determine whether the currently extracted target intent meets the requirements of the standard business processing logic.

[0051] For example, if there are problems such as ambiguous requirements (e.g., no clear optimization metrics) or logical contradictions (e.g., requiring an increase in recall rate while reducing data call volume), it can be determined that the extracted target intent does not meet the requirements of standard business processing logic.

[0052] If it is determined that the extracted target intent meets the requirements of the standard business processing logic, then the subsequent steps of inputting the target intent and task template library into the corpus analysis model can be continued.

[0053] In one possible implementation, if it is determined that the currently extracted target intent does not meet the requirements of the standard business processing logic, the follow-up questions output by the corpus analysis model are obtained and output.

[0054] For example, when a user only enters "Optimize loan risk monitoring model", the system automatically asks "Would you like to prioritize optimizing the model's precision, recall, or number of detected numbers?" to ensure that the generated task template is accurate.

[0055] If a response is received from the user regarding the follow-up question, the response is added to the model optimization instruction, and the process of extracting the target intent of the model optimization instruction using the fine-tuned training of the large corpus analysis model continues until the large corpus analysis model determines that the currently extracted target intent meets the requirements of the standard business processing logic.

[0056] To further improve the accuracy of model optimization, based on the above embodiments, in this embodiment, after sending the data to be processed by the target agent to the target agent so that the target agent can complete the task it needs to perform, and before obtaining the model parameter adjustment suggestion output by the last agent in the scheduling order, the method further includes: If the execution result of the target agent is an execution failure, the execution result, the model optimization instruction, and the agent scheduling task instruction are input into the intent understanding agent, so that the corpus analysis big model in the intent understanding agent can analyze the reason for the execution failure and adjust the input data of the agent that failed to execute.

[0057] In this embodiment, the scheduling chain supports dynamic adjustment. When a node (i.e., an agent) fails to execute a task (e.g., the data query agent fails to obtain the target data), the scheduling chain automatically triggers a backtracking mechanism to reassign tasks, such as switching the data query data source or adjusting the query conditions.

[0058] In this embodiment, if the execution result of the target agent is an execution failure, the execution result, model optimization instructions, and agent scheduling task instructions can be input into the intent understanding agent, so that the corpus analysis big model in the intent understanding agent can adjust the input data of the agent that failed to execute.

[0059] For example, the prompt could be: "The currently determined agent scheduling task instruction is: first, call the data query agent to query call data of suspicious phone numbers in the past month, and then call the code execution agent to perform data preprocessing on the found call data. After that, call the feature mining agent to extract features from the preprocessed call data. The agent scheduling task instruction requires the data query agent to perform a data query task, but the agent failed to perform the task and did not find any call data for suspicious phone numbers. Please analyze the reason for the failure of the data query agent task based on the currently received model optimization instruction, and adjust the input data of the data query agent."

[0060] Example 3: To further improve the accuracy of model optimization, based on the above embodiments, in this embodiment, sending the data to be processed by the target agent to the target agent so that the target agent can perform the task it needs to complete includes: If the target agent is a data query agent, the query requirements in the agent scheduling task instruction are input into the data query agent; The data query agent breaks down the query requirement into multiple keywords and searches for target SQL statement templates that match the multiple keywords in a pre-configured Structure Query Language (SQL) statement template library. The query requirements and the target SQL statement template are input into the SQL statement generation model, which then rewrites the target SQL statement template to obtain a target SQL statement that meets the query requirements. Execute the target SQL statement and store the target result in the shared storage area.

[0061] In this embodiment, the data query agent serves as the starting point for data layer collaboration. Based on the query requirements in the agent's scheduling task instructions, it calls the built-in SQL template database and employs a similarity matching algorithm to semantically match the keywords of the query requirements with the SQL templates. The relevance score between the query terms and the template statements is calculated, and the template with the highest score is selected. This template is then rewritten by a large language model to generate a precise SQL query statement. For example, for the requirement of "obtaining call records of suspicious numbers in a certain city over the past three months," after matching the basic query template, the system rewrites and generates an SQL statement containing filtering conditions for region, time, and business type. This query is automatically executed on the big data platform, and the query results are synchronized to the result memory area.

[0062] Specifically, in this embodiment, if the agent that needs to perform the task next is a data query agent, that is, the target agent is a data query agent, the query requirements in the agent scheduling task instruction can be input into the data query agent so that the data query agent can perform corresponding query processing according to the received query requirements.

[0063] Since a query request is typically a single sentence, such as "Query call data of suspicious numbers in a certain city over the past three months," the data query agent, upon receiving this request, can break it down into multiple keywords. In other words, it performs keyword splitting based on the semantics of the query request. For example, "Query call data of suspicious numbers in a certain city over the past three months" can be split into multiple keywords: query, city, past three months, suspicious numbers, call data.

[0064] After obtaining the split keywords, target SQL statement templates matching these keywords can be found in a pre-configured SQL statement template library. For example, the BM25 (Best Matching 25) similarity matching algorithm can be used to determine the target SQL statement template. The BM25 algorithm is a probabilistic statistical information retrieval algorithm that calculates the relevance score between the query terms and the template statement, selects the optimal SQL template, and submits it to a large model for rewriting, achieving high-accuracy data querying and feature extraction. Those skilled in the art can configure the SQL statement templates in the SQL statement template library as needed.

[0065] After determining the target SQL statement template, the query requirements and the target SQL statement template can be input into the SQL statement generation model. The model then rewrites the target SQL statement template to obtain the target SQL statement that meets the query requirements. This SQL statement generation model can be obtained by fine-tuning a basic model based on a large number of SQL statement rewriting training samples.

[0066] After obtaining the target SQL statement, you can execute the target SQL statement and store the target result obtained from the execution of the target SQL statement in the shared storage area.

[0067] Example 4: To further improve the accuracy of model optimization, based on the above embodiments, in this embodiment, sending the data to be processed by the target agent to the target agent so that the target agent can perform the task it needs to complete includes: If the target agent is a code execution agent, the target result stored in the shared storage area and the code generation requirements in the agent scheduling task instruction are input into the code execution agent; The code execution agent inputs the target result and the code generation requirements into the large code generation model to obtain data processing code; The data processing code is run to process the target result, and the intermediate data obtained is stored in the shared storage area.

[0068] In order to process the target result output by the data query agent, in this embodiment of the application, the code execution agent can process the target result.

[0069] In this embodiment of the application, if the determined target agent is a code execution agent, the target result in the shared storage area and the code generation requirements in the agent scheduling task instruction can be input into the code execution agent.

[0070] The code execution agent can input the target result and code generation requirements into a large-scale code generation model to obtain data processing code. This large-scale code generation model can be obtained by fine-tuning a basic large-scale model based on a large number of target results and code generation requirements. The code generation requirements can store configured data structures and points to note during code generation. The data structure indicates the field names that the code generation model should use for different keywords. These points to note might include "every line of code should have a comment," "data preprocessing should be performed before data statistics during code generation," and may also include code framework templates for different types of data.

[0071] After obtaining the data processing code, the executing agent can run the code to process the result and obtain intermediate data. To facilitate subsequent processing of this intermediate data by other agents, the intermediate data can be stored in a shared storage area.

[0072] Specifically, when the code execution agent determines that the target result is call log data, it can match the corresponding Python data preprocessing code framework to obtain data processing code. This data processing code can perform operations such as data cleaning, missing value imputation, and outlier removal, and calculate basic metrics such as the number of calls and average call duration. These metrics can be understood as the target result. After obtaining the intermediate data, the intermediate data can be stored in the shared storage area.

[0073] Example 5: To further improve the accuracy of model optimization, based on the above embodiments, in this embodiment, sending the data to be processed by the target agent to the target agent so that the target agent can perform the task it needs to complete includes: If the target agent is a feature mining agent, the intermediate data in the shared storage area will be input into the feature mining agent; The feature mining agent extracts the independent features corresponding to each data point in the intermediate data, as well as the common features corresponding to the intermediate data, based on the pre-configured feature extraction requirements corresponding to different data. The feature mining agent uses a feature analysis big model to analyze each independent feature, the common features, and pre-configured feature importance suggestions, determines the feature importance ranking, generates a feature mining report, and stores the feature mining report in the shared storage area.

[0074] In order to analyze the features of the data, in this embodiment of the application, if the target agent is a feature mining agent, the intermediate data in the shared storage area can be input into the feature mining agent.

[0075] This feature mining agent is pre-configured with feature extraction requirements corresponding to different types of data. For example, for call data, the extracted features could be call duration, and for image data, the extracted features could be facial features of people in the image.

[0076] After receiving intermediate data, the feature mining agent can determine the target feature extraction requirements for the intermediate data based on its type and the pre-configured correspondence between different data feature extraction requirements. Then, based on these requirements, it extracts the independent features corresponding to each data point in the intermediate data, as well as the common features. The independent features can be those describing the corresponding data, while the common features are those indicating relationships between the data points in the intermediate data.

[0077] After identifying independent and common features, the feature mining agent can utilize a large feature analysis model to analyze each independent feature, common feature, and pre-configured feature importance suggestions. It then determines the feature importance ranking and generates a feature mining report, which is stored in a shared storage area. This large feature analysis model can be trained using a large amount of training data containing independent and common features. The feature importance suggestions can be: features with the highest frequency of occurrence are ranked higher, and features with higher accuracy are ranked higher.

[0078] Specifically, the analysis layer collaboratively implements a feature mining agent that receives intermediate data from the code execution agent and then performs multi-dimensional feature mining by referring to a built-in business knowledge base. This business knowledge base stores feature extraction requirements corresponding to different types of data.

[0079] For example, on the one hand, deep features of number behavior are extracted, such as call timing patterns and the clustering of contact targets; on the other hand, common features are mined, such as suspicious numbers being active between 10:00 and 16:00 on weekdays and contact targets being mostly numbers from other regions. After the mining is completed, a feature mining report containing independent features, common features, and feature importance ranking is generated and synchronized to the shared storage area.

[0080] Example 6: To further improve the accuracy of model optimization, based on the above embodiments, in this embodiment, sending the data to be processed by the target agent to the target agent so that the target agent can perform the task it needs to complete includes: If the target agent is a model evaluation agent, the information of the model to be optimized and the feature mining report in the shared storage area are input into the model evaluation agent. The information of the model to be optimized includes the model name, launch time, functional description and fixed parameters in the model to be optimized. The model evaluation agent uses a large model performance evaluation model to analyze the information in the feature mining report and the model to be optimized, determines the model evaluation report, and stores it in the shared storage area. The model evaluation report records the suggested optimization content of the model to be optimized.

[0081] To evaluate the performance of the model, in this embodiment, if the identified target agent is a model evaluation agent, the information of the model to be optimized and the feature mining report in the shared storage area can be input into the model evaluation agent. To ensure the model evaluation agent can accurately identify the model that needs optimization, in this embodiment, the information of the model to be optimized may include the model name, launch time, functional description, and fixed parameters in the model to be optimized.

[0082] After receiving the information of the model to be optimized and the feature mining report, the model evaluation agent can analyze the feature mining report and the information of the model to be optimized using a large model performance evaluation model, determine the model evaluation report, and store it in a shared storage area. This model evaluation report records suggested optimizations for the model to be optimized. These suggested optimizations describe the directions for optimizing the model. For example, the suggested optimizations may be information about the parameters to be optimized; that is, the suggested optimizations describe which parameters need to be adjusted. The large performance evaluation model can be obtained by fine-tuning a base model using a large amount of model evaluation training data.

[0083] For example, a large-scale model performance evaluation can be combined with a feature mining report to comprehensively evaluate the model to be optimized. This includes calculating core metrics such as the number of detected numbers, the number of suspicious numbers, and the number of shut-down numbers; assessing the suitability of each parameter in the current model to be optimized to the current application scenario; and identifying parameters with deviations. For instance, if a feature weight is set too low, causing some phone numbers to be missed, then that feature weight can be determined as a parameter to be optimized. Finally, a model evaluation report can be generated based on the evaluation results.

[0084] Example 7: To further improve the accuracy of model optimization, based on the above embodiments, in this embodiment, sending the data to be processed by the target agent to the target agent so that the target agent can perform the task it needs to complete includes: If the target agent is a model optimization agent, the information of the model to be optimized, the model evaluation report in the shared storage area, and the model performance requirements in the agent scheduling task instruction are input into the model optimization agent. The model optimization agent analyzes the model evaluation report, the information of the model to be optimized, and the model performance requirements using the large optimization model, determines the adjustment values ​​of the suggested optimization content, and stores them in the shared storage area. The number of parameters to be optimized included in the suggested optimization content can be one or more. Of course, the suggested optimization content may also include other content, such as deleting the average pooling layer in the model to be optimized, or adding convolutional blocks to the model to be optimized. Those skilled in the art can adjust the prompts input into the large optimization model as needed; this application embodiment does not impose any limitations on this.

[0085] In order to determine the adjustment values ​​of the parameters to be optimized recorded in the model evaluation report, in this embodiment of the application, a model optimization agent can be used for determination.

[0086] In this embodiment of the application, if the determined target agent is a model optimization agent, the information of the model to be optimized, the model evaluation report in the shared storage area, and the model performance requirements in the agent scheduling task instructions can be input into the model optimization agent.

[0087] The model optimization agent can analyze the model evaluation report, information about the model to be optimized, and model performance requirements using the large optimization model to determine the adjustment values ​​for the suggested optimization content and store them in a shared storage area. These parameter adjustment suggestions may include adjustments for one or more parameters.

[0088] Specifically, after the optimization layer collaborative implementation model optimization agent receives the model evaluation report, the information of the model to be optimized, and the model performance requirements, it can determine parameter adjustment suggestions by combining the feature mining results.

[0089] In this embodiment, operations such as adding, modifying, and adjusting the weights of fixed parameters in the model to be optimized can be performed to comprehensively improve the model's precision and recall. For example, for an anomaly detection model, the weight threshold for the proportion of out-of-town calls can be adjusted from 0.3 to 0.5.

[0090] To further improve the accuracy of model optimization, based on the above embodiments, in this embodiment, sending the data to be processed by the target agent to the target agent so that the target agent can perform the task it needs to complete includes: If the target agent is a model evaluation agent and parameter adjustment suggestions are stored in the shared storage area, then the parameter adjustment suggestions are input into the model evaluation agent. The model evaluation agent adjusts the parameters of the model to be optimized according to the parameter adjustment suggestions, and performs simulation tests on the parameter-optimized model based on each independent feature and the common feature; If the simulation test results meet the preset accuracy requirements, then the parameter adjustment suggestions will be output.

[0091] To further improve the accuracy of model optimization, in this embodiment, after obtaining the parameter adjustment suggestions for the parameters to be optimized output by the model optimization agent, the accuracy of the parameter adjustment suggestions can be evaluated using the model evaluation agent.

[0092] In this embodiment of the application, if the target agent is a model evaluation agent and parameter adjustment suggestions are stored in the shared storage area, it means that the model optimization agent has provided adjustment suggestions for the parameters to be optimized, and the parameter adjustment suggestions can be input into the model evaluation agent.

[0093] After receiving parameter adjustment suggestions, the model evaluation agent can adjust the parameters of the model to be optimized according to these suggestions, and then perform simulation tests on the optimized model based on each obtained independent feature and common feature. In other words, the extracted independent features and common features are used as test data, and this test data is used to simulate and test the optimized model.

[0094] If the simulation test results meet the preset accuracy requirements, it can be determined that the parameter adjustment suggestions given by the model optimization agent are reasonable, and these suggestions can be directly output. The preset accuracy requirement can be an improvement of more than 5% in model accuracy, and those skilled in the art can configure it as needed.

[0095] Specifically, the model evaluation agent can conduct simulation tests based on historical data. If the verification results meet the preset accuracy requirements (such as an improvement in accuracy of more than 5%), the parameter adjustment suggestions will be pushed to the big data platform for formal deployment.

[0096] If the model evaluation agent determines that the simulation test results do not meet the preset accuracy requirements, it means that the parameter adjustment suggestions given by the model optimization agent are unreasonable. In this case, the model optimization agent can readjust the parameters until the expected results are achieved.

[0097] In this embodiment of the application, the readjustment suggestions can be stored in a shared storage area.

[0098] Once the electronic device determines that the shared storage area contains readjustment suggestions, it can send a readjustment instruction to the model optimization agent. This allows the model evaluation agent to continue executing the steps of analyzing the feature mining report and the information of the model to be optimized using the large model performance evaluation model, until the simulation test results meet the preset accuracy requirements.

[0099] Example 8: To improve the efficiency of model optimization, based on the above embodiments, in this embodiment, after sequentially calling the target agent and before sending the data to be processed by the target agent to the target agent, the method further includes: Determine whether target data exists in the shared storage area, wherein the target data is data obtained after the target agent completes the target task in the past, the target task is consistent with the task required by the target agent to be completed in the agent scheduling task instruction, and the shared storage area stores the output data of each agent within a preset time range; If so, then the invocation of the target agent is determined to be complete, and the step of determining the first uninvoked agent as the target agent is continued according to the scheduling order of the agents in the agent scheduling task instruction.

[0100] To prevent the same agent from repeating the same operation within a short period of time, in this embodiment, after determining the target agent to be invoked, before sending the data to be processed by the target agent, it can be determined whether target data exists in the shared storage area. This target data is data obtained after the target agent has completed a target task in a previous call, and this target task is consistent with the task required of the target agent in the agent scheduling task instruction.

[0101] In this embodiment, the shared storage area stores the output data of each agent within a preset time range. This preset time range can be any time range, such as within a week or a day. In other words, the shared storage area stores the recent output data of each agent.

[0102] In this embodiment of the application, the shared storage area may also store data such as the execution progress, key results, and abnormal information of each intelligent agent.

[0103] Specifically, when the task scheduling instruction requires agent A to complete task A, then when the target agent is agent A, it can be determined whether the data output by agent A when completing task A exists in the shared storage area.

[0104] In this embodiment of the application, if it is determined that target data exists in the shared storage area, the call to the target agent can be skipped, and subsequent agents can directly process the target data stored in the shared storage area.

[0105] In this embodiment of the application, when it is determined that target data exists in the shared storage area, it can be determined that the call to the target agent has been completed, and the step of determining the first uncalled agent as the target agent according to the scheduling order of agents in the agent scheduling task instruction can continue to be executed.

[0106] Example 9: To facilitate understanding, the following will be combined with... Figure 2 A flowchart of another agent-based model optimization process provided for embodiments of this application is shown below. Figure 2As shown, firstly, the user inputs model optimization requirements into the intent-understanding agent. This agent is responsible for deeply understanding the user's needs, establishing an agent scheduling chain (i.e., agent scheduling task instructions) based on the actual business scenario, and establishing the workflow required for subsequent model optimization. Then, the agent scheduling module allocates tasks according to the scheduling chain, sharing key memories among the agents during execution to achieve efficient collaboration among multiple agents. Afterwards, the data query agent, code execution agent, feature mining agent, model evaluation agent, and model optimization agent collaborate according to their respective task requirements. Each agent independently possesses system prompts, a business knowledge base, and execution memory, and can interact with the big data platform storing business data. Driven by the scheduling module, they execute collaboratively. Finally, the model optimization agent comprehensively analyzes the data and outputs model optimization suggestions.

[0107] In this embodiment, the cross-agent collaborative optimization process relies on a scheduling chain. By sharing process data, key conclusions, error information, and other data during execution, it achieves efficient collaboration among the agents, covering the entire process from data acquisition to model deployment. The key results are explained below: The intent-understanding agent is responsible for receiving model optimization requests from business personnel, parsing the input semantics, and identifying the business objective, data type, model category, and optimization goal. Combining this with a built-in business knowledge base, it categorizes and structures the input intent, and generates agent scheduling task instructions, providing a basis for task allocation to the agent scheduling module.

[0108] The agent scheduling module: Based on the intent-understanding agents, this module generates agent scheduling task instructions and constructs a multi-agent collaborative scheduling sequence. Through task allocation and feedback mechanisms, this module enables collaborative work between different agents and supports memory sharing among multiple agents. This allows agents to share contextual information and execution status during task execution, thereby achieving cross-task knowledge transfer and strategy collaboration.

[0109] The data query agent is responsible for automatically generating query commands for the required data based on the task requirements assigned by the scheduling module, and then automatically executing them on the data platform. Addressing the complex SQL statement structures and variable query conditions inherent in real-world business scenarios, this agent incorporates a database of commonly used target business SQL templates. It employs the BM25 similarity matching algorithm to semantically match user intent keywords with SQL templates, selecting the template that best meets the data requirements. Subsequently, based on the similarity matching results of the BM25 algorithm, it uses a large language model to generate precise query SQL statements, automatically executes them on the data platform, and returns the query results.

[0110] The code execution agent receives the results returned by the data query agent, generates compliant Python code based on the required data, and executes it automatically. This agent has a built-in target business analysis code template library, which can automatically determine the data type, analysis purpose, and match the corresponding code framework, thereby achieving functions such as data preprocessing, indicator calculation, and visualization output. Through automated code generation and execution mechanisms, manual intervention can be significantly reduced, achieving highly efficient automation of the model analysis process.

[0111] Feature mining agent: Based on the analysis results generated by the code execution agent, multi-dimensional in-depth analysis is performed on the number and its related entities (such as account opening information, call behavior, SMS records, internet browsing trajectory, location information, etc.) to extract independent features and common features, and generate a feature mining analysis report.

[0112] Model Evaluation Agent: This agent evaluates the operational effectiveness and key metrics of various anti-fraud models currently in use, including the number of detected numbers, the number of suspicious numbers, the number of shut-down numbers, and the number of reactivated numbers. When the user does not specify a model to be optimized, the agent automatically calculates the optimization priority based on model performance metrics and selects the model with the highest priority for analysis. Simultaneously, this agent can combine feature mining results to diagnose the adaptability of model rules, identify deviations between model rules and fraud scenarios, and provide quantitative support for subsequent model optimization.

[0113] The model optimization agent receives model evaluation and feature mining results, comprehensively analyzes the correspondence between model rules and fraud scenarios, and automatically generates model optimization strategies. This agent supports adding, modifying, and adjusting weights for model rules. Optimized rules are automatically pushed to the data platform for deployment, enabling rapid response and strategy updates for new business scenarios.

[0114] The agents in the above multi-agent architecture are obtained by fine-tuning different subdomains of the target business domain (such as data query, feature mining, model metric optimization, etc.).

[0115] In this embodiment, the division of labor and cooperation among multiple agents is accomplished by an intent-understanding agent and an agent scheduling module. After the intent-understanding agent breaks down the user's needs, each agent works according to its assigned task requirements. The scheduling module is responsible for sharing execution progress, execution results, and exception information among the agents, helping each agent adjust its workflow in real time.

[0116] In this embodiment, the intent-understanding agent is combined with the agent scheduling task instructions to achieve accurate understanding of business needs and efficient agent collaboration. This two-layer design of intent-understanding agent and agent scheduling task instructions is the core of the efficient operation of the entire multi-agent system. This structure can accurately parse user needs, convert them into business requirements, and then dynamically schedule other agents to execute tasks, achieving a seamless transition from vague user needs to concrete task implementation.

[0117] In this embodiment, after the collaborative closed-loop and iterative optimization model optimization rules are deployed, the model evaluation agent initiates a continuous monitoring mechanism to collect real-time network operation data, calculate various indicators of the optimized model, compare and analyze them with those before optimization, and generate an effect evaluation report. Simultaneously, the evaluation results are fed back to the intent understanding agent and the agent scheduling module. Based on the effect evaluation report, the intent understanding agent determines whether further optimization is needed; if so, it generates a new task template and triggers the next round of collaborative optimization. The agent scheduling module stores the task execution data and anomaly handling records from this collaborative process into a knowledge base, providing a reference for subsequent similar tasks and realizing closed-loop iteration of cross-agent collaborative optimization.

[0118] The agent-based model optimization method provided in this application proposes a multi-agent collaborative architecture. Through task allocation and feedback mechanisms, it enables collaborative work between different agents, supports memory sharing among multiple agents, and allows each agent to share context information and execution status during task execution, thereby achieving cross-task knowledge transfer and strategy collaboration.

[0119] The agent-based model optimization method provided in this application embodiment automates the entire process from data acquisition to model update. The optimization results can be directly pushed to the live data platform for deployment, and the effect can be tracked through the model evaluation agent to achieve a closed-loop optimization process and shorten the model iteration cycle.

[0120] The agent-based model optimization method provided in this application can dynamically adjust model parameters and rules in a timely manner according to changes in application scenarios, avoiding business surges caused by sudden changes in scenarios and ensuring user property security. Through the division of labor and cooperation of multiple agents, automated model parameter iteration is achieved, improving the response speed of the management model to changes in application scenarios and enhancing the effectiveness of the model to be optimized.

[0121] Example 10: Based on the same inventive concept, embodiments of this application provide an agent-based model optimization device. Figure 3 This application provides a schematic diagram of a model optimization device based on an intelligent agent, the device comprising: The receiving module 301 is used to receive a model optimization instruction, which includes information about the model to be optimized and optimization requirements. The intent understanding module 302 is used to understand the model optimization instructions using an intent understanding agent to obtain an agent scheduling task instruction. The agent scheduling task instruction describes the order in which agents need to be called to complete the model optimization instructions, and the tasks to be completed by each agent. The agents that can be called include: a data query agent, a code execution agent, a feature mining agent, a model evaluation agent, and a model optimization agent. The optimization module 303 is used to sequentially call the target intelligent agents according to the scheduling order of the intelligent agents in the intelligent agent scheduling task instruction, send the data that the target intelligent agent needs to process to the target intelligent agent, so that the target intelligent agent can complete the task it needs to complete, until the model parameter adjustment suggestion output by the last intelligent agent in the scheduling order is obtained; and adjust the parameters in the model to be optimized based on the model parameter adjustment suggestion.

[0122] In one possible implementation, the intent understanding module 302 is specifically used for the intent understanding agent to extract the target intent of the model optimization instruction using a large corpus analysis model trained with fine-tuning; inputting the target intent and task template library into the large corpus analysis model, the large corpus analysis model determining the target task template in the task template library that matches the target intent, and determining the agent's scheduling task instruction based on the target task template and the target intent.

[0123] In one possible implementation, the optimization module 303 is further configured to input the target intent and the pre-configured standard business processing logic into the corpus analysis model, so that the corpus analysis model determines whether the currently extracted target intent meets the requirements of the standard business processing logic. If so, the subsequent steps of inputting the target intent and the task template library into the corpus analysis model are continued.

[0124] In one possible implementation, the optimization module 303 is further configured to input the target intent and pre-configured standard business processing logic into the corpus analysis model, so that the corpus analysis model determines whether the currently extracted target intent meets the requirements of the standard business processing logic; if not, it obtains and outputs the follow-up question content output by the corpus analysis model; if it receives a response information to the follow-up question content, it adds the response information to the model optimization instruction, and continues to execute the step of the intent understanding agent using the corpus analysis model trained by fine-tuning to extract the target intent of the model optimization instruction, until the corpus analysis model determines that the currently extracted target intent meets the requirements of the standard business processing logic.

[0125] In one possible implementation, the optimization module 303 is further configured to, if the execution result of the target agent is an execution failure, input the execution result, the model optimization instruction, and the agent scheduling task instruction into the intent understanding agent, so that the corpus analysis big model in the intent understanding agent analyzes the reason for the execution failure and adjusts the input data of the agent that failed to execute.

[0126] In one possible implementation, the optimization module 303 is specifically configured to, if the target agent is a data query agent, input the query requirements from the agent scheduling task instruction into the data query agent; the data query agent splits the query requirements into multiple keywords and searches for target SQL statement templates that match the multiple keywords in a pre-configured SQL statement template library; input the query requirements and the target SQL statement templates into a large SQL statement generation model, causing the large SQL statement generation model to rewrite the target SQL statement templates to obtain target SQL statements that meet the query requirements; execute the target SQL statement and store the obtained target results in a shared storage area.

[0127] In one possible implementation, the optimization module 303 is specifically configured to, if the target agent is a code execution agent, input the target result stored in the shared storage area and the code generation requirements in the agent scheduling task instruction to the code execution agent; the code execution agent inputs the target result and the code generation requirements into the code generation big model to obtain data processing code; runs the data processing code to process the target result and stores the obtained intermediate data in the shared storage area.

[0128] In one possible implementation, the optimization module 303 is specifically configured to, if the target agent is a feature mining agent, input intermediate data from the shared storage area into the feature mining agent; the feature mining agent extracts independent features corresponding to each data point in the intermediate data and common features corresponding to the intermediate data based on pre-configured feature extraction requirements for different data; the feature mining agent uses a feature analysis big model to analyze each independent feature, the common features, and pre-configured feature importance suggestions, determines the feature importance ranking, generates a feature mining report, and stores the feature mining report in the shared storage area.

[0129] In one possible implementation, the optimization module 303 is specifically used to input the information of the model to be optimized and the feature mining report in the shared storage area into the model evaluation agent if the target agent is a model evaluation agent. The information of the model to be optimized includes the model name, launch time, functional description, and fixed parameters in the model to be optimized. The model evaluation agent uses a model performance evaluation big model to analyze the feature mining report and the information of the model to be optimized, determines the model evaluation report, and stores it in the shared storage area. The model evaluation report records the suggested optimization content of the model to be optimized.

[0130] In one possible implementation, the optimization module 303 is specifically configured to, if the target agent is a model optimization agent, input the information of the model to be optimized, the model evaluation report in the shared storage area, and the model performance requirements in the agent scheduling task instruction to the model optimization agent; the model optimization agent uses the optimized large model to analyze the model evaluation report, the information of the model to be optimized, and the model performance requirements, determines the adjustment value of the suggested optimization content, and stores it in the shared storage area.

[0131] In one possible implementation, the optimization module 303 is specifically configured to: if the target agent is a model evaluation agent and parameter adjustment suggestions are stored in the shared storage area, input the parameter adjustment suggestions to the model evaluation agent; the model evaluation agent adjusts the parameters of the model to be optimized according to the parameter adjustment suggestions, and performs simulation testing on the parameter-optimized model based on each independent feature and the common feature; if the simulation test results meet the preset accuracy requirements, output the parameter adjustment suggestions.

[0132] In one possible implementation, the optimization module 303 is specifically used to: if the model evaluation agent determines that the simulation test results do not meet the preset accuracy requirements, store the readjustment suggestions in the shared storage area; and send a readjustment instruction to the model optimization agent so that the model evaluation agent continues to execute the step of analyzing the information of the feature mining report and the model to be optimized using the model performance evaluation large model until it is determined that the simulation test results meet the preset accuracy requirements.

[0133] In one possible implementation, the optimization module 303 is further configured to determine whether target data exists in the shared storage area, wherein the target data is data obtained after the target agent has been historically invoked to complete a target task, the target task is consistent with the task required to be completed by the target agent in the agent scheduling task instruction, and the shared storage area stores the output data of each agent within a preset time range; if so, it is determined that the invocation of the target agent has been completed, and the step of determining the first uninvoked agent as the target agent according to the scheduling order of the agents in the agent scheduling task instruction continues.

[0134] Example 11: Based on the same inventive concept, embodiments of this application provide an electronic device that can implement the steps of the agent-based model optimization method discussed above. Figure 4 This application provides a schematic diagram of an electronic device structure, such as... Figure 4 As shown, it includes: processor 401, communication interface 402, memory 403 and communication bus 404, wherein processor 401, communication interface 402 and memory 403 communicate with each other through communication bus 404. The memory 403 stores a computer program. When the program is executed by the processor 401, the processor 401 performs the following steps: Upon receiving a model optimization instruction, the model optimization instruction includes information about the model to be optimized and the optimization requirements; The intent-understanding agent interprets the model optimization instructions to obtain agent scheduling task instructions. The agent scheduling task instructions describe the order in which agents need to be called to complete the model optimization instructions, and the tasks to be completed by each agent. The agents that can be called include: data query agent, code execution agent, feature mining agent, model evaluation agent, and model optimization agent. According to the scheduling order of the agents in the agent scheduling task instruction, the target agents are called in turn, and the data that the target agents need to process is sent to the target agents so that the target agents can complete the tasks they need to complete, until the model parameter adjustment suggestions output by the last agent in the scheduling order are obtained; Based on the proposed model parameter adjustment, the parameters in the model to be optimized are adjusted.

[0135] In one possible implementation, the step of using an intent-understanding agent to understand the model optimization instructions and obtain agent scheduling task instructions includes: The intent-understanding agent uses a large-scale corpus analysis model trained with fine-tuning to extract the target intent of the model's optimized instructions. The target intent and task template library are input into the corpus analysis model. The corpus analysis model determines the target task template in the task template library that matches the target intent, and determines the agent's scheduling task instruction based on the target task template and the target intent.

[0136] In one possible implementation, after extracting the target intent of the model optimization instructions using the fine-tuned training-completed large-scale corpus analysis model, and before inputting the target intent and task template library into the large-scale corpus analysis model, the method further includes: The target intent and the pre-configured standard business processing logic are input into the corpus analysis model, which then determines whether the currently extracted target intent meets the requirements of the standard business processing logic. If so, the subsequent steps of inputting the target intent and task template library into the corpus analysis model continue.

[0137] In one possible implementation, after extracting the target intent of the model optimization instructions using the fine-tuned training-completed large-scale corpus analysis model, and before inputting the target intent and task template library into the large-scale corpus analysis model, the method further includes: The target intent and the pre-configured standard business processing logic are input into the corpus analysis model, so that the corpus analysis model can determine whether the currently extracted target intent meets the requirements of the standard business processing logic; If not, then obtain and output the follow-up questions output by the large corpus analysis model; If a response is received in response to the follow-up question, the response is added to the model optimization instruction, and the process of extracting the target intent of the model optimization instruction using the fine-tuned training corpus analysis model continues until the corpus analysis model determines the requirement of the standard business processing logic for the currently extracted target intent.

[0138] In one possible implementation, after sending the data to be processed by the target agent to enable the target agent to perform its required task, and before obtaining the model parameter adjustment suggestion output by the last agent in the scheduling order, the method further includes: if the execution result of the target agent is an execution failure, inputting the execution result, the model optimization instruction, and the agent scheduling task instruction into the intent understanding agent, so that the corpus analysis big model in the intent understanding agent analyzes the reason for the execution failure and adjusts the input data of the agent that failed to execute.

[0139] In one possible implementation, sending the data to be processed by the target agent to enable the target agent to perform its required task includes: If the target agent is a data query agent, the query requirements in the agent scheduling task instruction are input into the data query agent; The data query agent breaks down the query request into multiple keywords and searches for target SQL statement templates that match the multiple keywords in a pre-configured SQL statement template library. The query requirements and the target SQL statement template are input into the SQL statement generation model, which then rewrites the target SQL statement template to obtain a target SQL statement that meets the query requirements. Execute the target SQL statement and store the target result in the shared storage area.

[0140] In one possible implementation, sending the data to be processed by the target agent to enable the target agent to perform its required task includes: If the target agent is a code execution agent, the target result stored in the shared storage area and the code generation requirements in the agent scheduling task instruction are input into the code execution agent; The code execution agent inputs the target result and the code generation requirements into the large code generation model to obtain data processing code; The data processing code is run to process the target result, and the intermediate data obtained is stored in the shared storage area.

[0141] In one possible implementation, sending the data to be processed by the target agent to enable the target agent to perform its required task includes: If the target agent is a feature mining agent, the intermediate data in the shared storage area will be input into the feature mining agent; The feature mining agent extracts the independent features corresponding to each data point in the intermediate data, as well as the common features corresponding to the intermediate data, based on the pre-configured feature extraction requirements corresponding to different data. The feature mining agent uses a feature analysis big model to analyze each independent feature, the common features, and pre-configured feature importance suggestions, determines the feature importance ranking, generates a feature mining report, and stores the feature mining report in the shared storage area.

[0142] In one possible implementation, sending the data to be processed by the target agent to enable the target agent to perform its required task includes: If the target agent is a model evaluation agent, the information of the model to be optimized and the feature mining report in the shared storage area are input into the model evaluation agent. The information of the model to be optimized includes the model name, launch time, functional description and fixed parameters in the model to be optimized. The model evaluation agent uses a large model performance evaluation model to analyze the information in the feature mining report and the model to be optimized, determines the model evaluation report, and stores it in the shared storage area. The model evaluation report records the suggested optimization content of the model to be optimized.

[0143] In one possible implementation, sending the data to be processed by the target agent to enable the target agent to perform its required task includes: If the target agent is a model optimization agent, the information of the model to be optimized, the model evaluation report in the shared storage area, and the model performance requirements in the agent scheduling task instruction are input into the model optimization agent. The model optimization agent uses the optimization big model to analyze the model evaluation report, the information of the model to be optimized, and the model performance requirements, determines the adjustment values ​​of the suggested optimization content, and stores them in the shared storage area.

[0144] In one possible implementation, sending the data to be processed by the target agent to enable the target agent to perform its required task includes: If the target agent is a model evaluation agent and parameter adjustment suggestions are stored in the shared storage area, then the parameter adjustment suggestions are input into the model evaluation agent. The model evaluation agent adjusts the parameters of the model to be optimized according to the parameter adjustment suggestions, and performs simulation tests on the parameter-optimized model based on each independent feature and the common feature; If the simulation test results meet the preset accuracy requirements, then the parameter adjustment suggestions will be output.

[0145] In one possible implementation, the method further includes: If the model evaluation agent determines that the simulation test results do not meet the preset accuracy requirements, it will readjust the recommendations and store them in the shared storage area. Send a readjustment instruction to the model optimization agent so that the model evaluation agent continues to perform the step of analyzing the information of the feature mining report and the model to be optimized using the model performance evaluation large model until it is determined that the simulation test results meet the preset accuracy requirements.

[0146] In one possible implementation, after sequentially invoking the target agent and before sending the data to be processed by the target agent to the target agent, the method further includes: Determine whether target data exists in the shared storage area, wherein the target data is data obtained after the target agent completes the target task in the past, the target task is consistent with the task required by the target agent to be completed in the agent scheduling task instruction, and the shared storage area stores the output data of each agent within a preset time range; If so, then the invocation of the target agent is determined to be complete, and the step of determining the first uninvoked agent as the target agent is continued according to the scheduling order of the agents in the agent scheduling task instruction.

[0147] Since the principle of the above-mentioned electronic device in solving the problem is similar to that of the agent-based model optimization method, the implementation of the above-mentioned electronic device can be found in the embodiments of the method, and repeated details will not be repeated.

[0148] The communication bus mentioned in the above-mentioned electronic device can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This communication bus can be divided into address bus, data bus, control bus, etc. For ease of illustration, only one thick line is used in the figure, but this does not indicate that there is only one bus or one type of bus. Communication interface 402 is used for communication between the above-mentioned electronic device and other devices. The memory can include random access memory (RAM) or non-volatile memory (NVM), such as at least one disk storage device. Optionally, the memory can also be at least one storage device located remotely from the aforementioned processor.

[0149] The processors mentioned above can be general-purpose processors, including central processing units, network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits, field-programmable gate arrays or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc.

[0150] Example 12: Based on the same inventive concept, embodiments of this application also provide a computer-readable storage medium storing a computer program executable by a processor. When the program runs on the processor, it causes the processor to perform the following steps: Upon receiving a model optimization instruction, the model optimization instruction includes information about the model to be optimized and the optimization requirements; The intent-understanding agent interprets the model optimization instructions to obtain agent scheduling task instructions. The agent scheduling task instructions describe the order in which agents need to be called to complete the model optimization instructions, and the tasks to be completed by each agent. The agents that can be called include: data query agent, code execution agent, feature mining agent, model evaluation agent, and model optimization agent. According to the scheduling order of the agents in the agent scheduling task instruction, the target agents are called in turn, and the data that the target agents need to process is sent to the target agents so that the target agents can complete the tasks they need to complete, until the model parameter adjustment suggestions output by the last agent in the scheduling order are obtained; Based on the proposed model parameter adjustment, the parameters in the model to be optimized are adjusted.

[0151] In one possible implementation, the step of using an intent-understanding agent to understand the model optimization instructions and obtain agent scheduling task instructions includes: The intent-understanding agent uses a large-scale corpus analysis model trained with fine-tuning to extract the target intent of the model's optimized instructions. The target intent and task template library are input into the corpus analysis model. The corpus analysis model determines the target task template in the task template library that matches the target intent, and determines the agent's scheduling task instruction based on the target task template and the target intent.

[0152] In one possible implementation, after extracting the target intent of the model optimization instructions using the fine-tuned training-completed large-scale corpus analysis model, and before inputting the target intent and task template library into the large-scale corpus analysis model, the method further includes: The target intent and the pre-configured standard business processing logic are input into the corpus analysis model, which then determines whether the currently extracted target intent meets the requirements of the standard business processing logic. If so, the subsequent steps of inputting the target intent and task template library into the corpus analysis model continue.

[0153] In one possible implementation, after extracting the target intent of the model optimization instructions using the fine-tuned training-completed large-scale corpus analysis model, and before inputting the target intent and task template library into the large-scale corpus analysis model, the method further includes: The target intent and the pre-configured standard business processing logic are input into the corpus analysis model, so that the corpus analysis model can determine whether the currently extracted target intent meets the requirements of the standard business processing logic; If not, then obtain and output the follow-up questions output by the large corpus analysis model; If a response is received in response to the follow-up question, the response is added to the model optimization instruction, and the process of extracting the target intent of the model optimization instruction using the fine-tuned training corpus analysis model continues until the corpus analysis model determines the requirement of the standard business processing logic for the currently extracted target intent.

[0154] In one possible implementation, after sending the data to be processed by the target agent to enable the target agent to perform its required task, and before obtaining the model parameter adjustment suggestion output by the last agent in the scheduling order, the method further includes: if the execution result of the target agent is an execution failure, inputting the execution result, the model optimization instruction, and the agent scheduling task instruction into the intent understanding agent, so that the corpus analysis big model in the intent understanding agent analyzes the reason for the execution failure and adjusts the input data of the agent that failed to execute.

[0155] In one possible implementation, sending the data to be processed by the target agent to enable the target agent to perform its required task includes: If the target agent is a data query agent, the query requirements in the agent scheduling task instruction are input into the data query agent; The data query agent breaks down the query request into multiple keywords and searches for target SQL statement templates that match the multiple keywords in a pre-configured SQL statement template library. The query requirements and the target SQL statement template are input into the SQL statement generation model, which then rewrites the target SQL statement template to obtain a target SQL statement that meets the query requirements. Execute the target SQL statement and store the target result in the shared storage area.

[0156] In one possible implementation, sending the data to be processed by the target agent to enable the target agent to perform its required task includes: If the target agent is a code execution agent, the target result stored in the shared storage area and the code generation requirements in the agent scheduling task instruction are input into the code execution agent; The code execution agent inputs the target result and the code generation requirements into the large code generation model to obtain data processing code; The data processing code is run to process the target result, and the intermediate data obtained is stored in the shared storage area.

[0157] In one possible implementation, sending the data to be processed by the target agent to enable the target agent to perform its required task includes: If the target agent is a feature mining agent, the intermediate data in the shared storage area will be input into the feature mining agent; The feature mining agent extracts the independent features corresponding to each data point in the intermediate data, as well as the common features corresponding to the intermediate data, based on the pre-configured feature extraction requirements corresponding to different data. The feature mining agent uses a feature analysis big model to analyze each independent feature, the common features, and pre-configured feature importance suggestions, determines the feature importance ranking, generates a feature mining report, and stores the feature mining report in the shared storage area.

[0158] In one possible implementation, sending the data to be processed by the target agent to enable the target agent to perform its required task includes: If the target agent is a model evaluation agent, the information of the model to be optimized and the feature mining report in the shared storage area are input into the model evaluation agent. The information of the model to be optimized includes the model name, launch time, functional description and fixed parameters in the model to be optimized. The model evaluation agent uses a large model performance evaluation model to analyze the information in the feature mining report and the model to be optimized, determines the model evaluation report, and stores it in the shared storage area. The model evaluation report records the suggested optimization content of the model to be optimized.

[0159] In one possible implementation, sending the data to be processed by the target agent to enable the target agent to perform its required task includes: If the target agent is a model optimization agent, the information of the model to be optimized, the model evaluation report in the shared storage area, and the model performance requirements in the agent scheduling task instruction are input into the model optimization agent. The model optimization agent uses the optimization big model to analyze the model evaluation report, the information of the model to be optimized, and the model performance requirements, determines the adjustment values ​​of the suggested optimization content, and stores them in the shared storage area.

[0160] In one possible implementation, sending the data to be processed by the target agent to enable the target agent to perform its required task includes: If the target agent is a model evaluation agent and parameter adjustment suggestions are stored in the shared storage area, then the parameter adjustment suggestions are input into the model evaluation agent. The model evaluation agent adjusts the parameters of the model to be optimized according to the parameter adjustment suggestions, and performs simulation tests on the parameter-optimized model based on each independent feature and the common feature; If the simulation test results meet the preset accuracy requirements, then the parameter adjustment suggestions will be output.

[0161] In one possible implementation, the method further includes: If the model evaluation agent determines that the simulation test results do not meet the preset accuracy requirements, it will readjust the recommendations and store them in the shared storage area. Send a readjustment instruction to the model optimization agent so that the model evaluation agent continues to perform the step of analyzing the information of the feature mining report and the model to be optimized using the model performance evaluation large model until it is determined that the simulation test results meet the preset accuracy requirements.

[0162] In one possible implementation, after sequentially invoking the target agent and before sending the data to be processed by the target agent to the target agent, the method further includes: Determine whether target data exists in the shared storage area, wherein the target data is data obtained after the target agent completes the target task in the past, the target task is consistent with the task required by the target agent to be completed in the agent scheduling task instruction, and the shared storage area stores the output data of each agent within a preset time range; If so, then the invocation of the target agent is determined to be complete, and the step of determining the first uninvoked agent as the target agent is continued according to the scheduling order of the agents in the agent scheduling task instruction.

[0163] Since the principle of the computer-readable storage medium in solving the problem is similar to that of the agent-based model optimization method, the implementation of the computer-readable storage medium can be found in the embodiments of the method, and repeated details will not be described again.

[0164] Example 13: Based on the same inventive concept, this application also provides a computer program product, which includes computer program code. When the computer program code is run on a computer, it causes the computer to execute any of the agent-based model optimization methods discussed above. Since the principle of the above computer program product in solving the problem is similar to that of the agent-based model optimization method, the implementation of the above computer program product can refer to the implementation of the method, and repeated details will not be described again.

[0165] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0166] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to this application. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0167] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0168] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of user-operated steps to be executed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0169] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A model optimization method based on intelligent agents, characterized in that, The method includes: Upon receiving a model optimization instruction, the model optimization instruction includes information about the model to be optimized and the optimization requirements; The intent-understanding agent interprets the model optimization instructions to obtain agent scheduling task instructions. The agent scheduling task instructions describe the order in which agents need to be called to complete the model optimization instructions, and the tasks to be completed by each agent. The agents that can be called include: data query agent, code execution agent, feature mining agent, model evaluation agent, and model optimization agent. According to the scheduling order of the agents in the agent scheduling task instruction, the target agents are called in turn, and the data that the target agents need to process is sent to the target agents so that the target agents can complete the tasks they need to complete, until the model parameter adjustment suggestions output by the last agent in the scheduling order are obtained; Based on the proposed model parameter adjustment, the parameters in the model to be optimized are adjusted.

2. The method according to claim 1, characterized in that, The process of using an intent-understanding agent to understand the model optimization instructions and obtain agent scheduling task instructions includes: The intent-understanding agent uses a large-scale corpus analysis model trained with fine-tuning to extract the target intent of the model's optimized instructions. The target intent and task template library are input into the corpus analysis model. The corpus analysis model determines the target task template in the task template library that matches the target intent, and determines the agent's scheduling task instruction based on the target task template and the target intent.

3. The method according to claim 2, characterized in that, After extracting the target intent of the model optimization instructions from the large-scale corpus analysis model trained with fine-tuning, and before inputting the target intent and task template library into the large-scale corpus analysis model, the method further includes: The target intent and the pre-configured standard business processing logic are input into the corpus analysis model, so that the corpus analysis model can determine whether the currently extracted target intent meets the requirements of the standard business processing logic; If not, then obtain and output the follow-up questions output by the large corpus analysis model; If a response is received in response to the follow-up question, the response is added to the model optimization instruction, and the process of extracting the target intent of the model optimization instruction using the fine-tuned training corpus analysis model continues until the corpus analysis model determines the requirement of the standard business processing logic for the currently extracted target intent.

4. The method according to claim 1, characterized in that, The step of sending the data that the target agent needs to process to the target agent, so that the target agent can perform the task it needs to complete, includes: If the target agent is a data query agent, the query requirements in the agent scheduling task instruction are input into the data query agent; The data query agent breaks down the query request into multiple keywords and searches for target SQL statement templates that match the multiple keywords in a pre-configured SQL statement template library. The query requirements and the target SQL statement template are input into the SQL statement generation model, which then rewrites the target SQL statement template to obtain a target SQL statement that meets the query requirements. Execute the target SQL statement and store the target result in the shared storage area.

5. The method according to claim 1, characterized in that, The step of sending the data that the target agent needs to process to the target agent, so that the target agent can perform the task it needs to complete, includes: If the target agent is a code execution agent, the target result stored in the shared storage area and the code generation requirements in the agent scheduling task instruction are input into the code execution agent; The code execution agent inputs the target result and the code generation requirements into the large code generation model to obtain data processing code; The data processing code is run to process the target result, and the intermediate data obtained is stored in the shared storage area.

6. The method according to claim 1, characterized in that, The step of sending the data that the target agent needs to process to the target agent, so that the target agent can perform the task it needs to complete, includes: If the target agent is a feature mining agent, the intermediate data in the shared storage area will be input into the feature mining agent; The feature mining agent extracts the independent features corresponding to each data point in the intermediate data, as well as the common features corresponding to the intermediate data, based on the pre-configured feature extraction requirements corresponding to different data. The feature mining agent uses a feature analysis big model to analyze each independent feature, the common features, and pre-configured feature importance suggestions, determines the feature importance ranking, generates a feature mining report, and stores the feature mining report in the shared storage area.

7. The method according to claim 1, characterized in that, The step of sending the data that the target agent needs to process to the target agent, so that the target agent can perform the task it needs to complete, includes: If the target agent is a model evaluation agent, the information of the model to be optimized and the feature mining report in the shared storage area are input into the model evaluation agent. The information of the model to be optimized includes the model name, launch time, functional description and fixed parameters in the model to be optimized. The model evaluation agent uses a large model performance evaluation model to analyze the information in the feature mining report and the model to be optimized, determines the model evaluation report, and stores it in the shared storage area. The model evaluation report records the suggested optimization content of the model to be optimized.

8. The method according to claim 1, characterized in that, The step of sending the data that the target agent needs to process to the target agent, so that the target agent can perform the task it needs to complete, includes: If the target agent is a model optimization agent, the information of the model to be optimized, the model evaluation report in the shared storage area, and the model performance requirements in the agent scheduling task instruction are input into the model optimization agent. The model optimization agent uses the optimization big model to analyze the model evaluation report, the information of the model to be optimized, and the model performance requirements, determines the adjustment values ​​of the suggested optimization content, and stores them in the shared storage area.

9. The method according to claim 1, characterized in that, The step of sending the data that the target agent needs to process to the target agent, so that the target agent can perform the task it needs to complete, includes: If the target agent is a model evaluation agent and parameter adjustment suggestions are stored in the shared storage area, then the parameter adjustment suggestions are input into the model evaluation agent. The model evaluation agent adjusts the parameters of the model to be optimized according to the parameter adjustment suggestions, and performs simulation tests on the parameter-optimized model based on each independent feature and the common feature; If the simulation test results meet the preset accuracy requirements, then the parameter adjustment suggestions will be output.

10. The method according to claim 9, characterized in that, The method further includes: If the model evaluation agent determines that the simulation test results do not meet the preset accuracy requirements, it will readjust the recommendations and store them in the shared storage area. Send a readjustment instruction to the model optimization agent so that the model evaluation agent continues to perform the step of analyzing the information of the feature mining report and the model to be optimized using the model performance evaluation large model until it is determined that the simulation test results meet the preset accuracy requirements.

11. The method according to any one of claims 4-10, characterized in that, After sequentially invoking the target agent, and before sending the data to be processed by the target agent to the target agent, the method further includes: Determine whether target data exists in the shared storage area, wherein the target data is data obtained after the target agent completes the target task in the past, the target task is consistent with the task required by the target agent to be completed in the agent scheduling task instruction, and the shared storage area stores the output data of each agent within a preset time range; If so, then the invocation of the target agent is determined to be complete, and the step of determining the first uninvoked agent as the target agent is continued according to the scheduling order of the agents in the agent scheduling task instruction.

12. A model optimization device based on intelligent agents, characterized in that, The device includes: A receiving module is used to receive a model optimization instruction, which includes information about the model to be optimized and optimization requirements. The intent understanding module is used to understand the model optimization instructions using an intent understanding agent to obtain an agent scheduling task instruction. The agent scheduling task instruction describes the order in which agents need to be called to complete the model optimization instructions, and the tasks to be completed by each agent. The agents that can be called include: data query agent, code execution agent, feature mining agent, model evaluation agent, and model optimization agent. The optimization module is used to sequentially call the target agent according to the scheduling order of the agents in the agent scheduling task instruction, send the data that the target agent needs to process to the target agent, so that the target agent can complete the task it needs to complete, until the model parameter adjustment suggestion output by the last agent in the scheduling order is obtained; and adjust the parameters in the model to be optimized based on the model parameter adjustment suggestion.

Images