Multi-level dynamic intelligent automatic data processing and information evaluation method and system

Abstract

The invention belongs to the technical field of information, and relates to a multi-level dynamic intelligent automatic data processing and information evaluation method and system. The method comprises the steps that a database environment operation data set is constructed, targeted training and fine adjustment are conducted on a large language model in combination with reinforcement learning, and full-automatic data analysis, query and information extraction of multiple tables in a database are achieved; a multi-level agent interaction and cooperation framework is established, a scheduling agent is responsible for global decision making and strategy planning, an operation agent is responsible for execution of specific tasks, and the scheduling agent and the operation agent share information and coordinate the tasks to form a highly cooperative whole; and establishing a cyclic supervision and error correction mechanism, scheduling the intelligent agent to learn a normative workflow example of the target working environment and track the state of the operation intelligent agent in real time, identifying the behavior of the operation intelligent agent deviating from a predetermined path, and taking corrective measures. According to the method, high efficiency, intelligence and automation of data processing are realized, and the method has remarkable practical value and innovative significance.

Description

Technical Field

[0001] The present invention belongs to the field of information technology and comprehensively applies a variety of cutting-edge technologies such as big data processing and analysis, artificial intelligence agent design, multi-agent system interaction and collaboration, natural language processing (NLP), and reinforcement learning. Its core is to propose a dynamic intelligent automated data processing and information evaluation method and system based on a multi-level agent architecture. Through this architecture, the system realizes functions such as high-performance situation awareness, precise multi-module collaboration, adaptive supervision and error correction mechanisms, and high-quality stable output, and has the ability to quickly adapt to complex environments and precisely execute tasks.

[0002] The present invention provides innovative solutions for the difficulties in the scientometric evaluation of large-scale complex natural language content, such as strong data environment dynamics, diverse task requirements, and unstable result quality. By integrating the collaborative working ability of multi-agents and the optimization mechanism driven by reinforcement learning, the efficiency of big data processing, the accuracy of complex information analysis, and the reliability of output results are significantly improved, providing new ideas and practical paths for the intelligent and automated development in the field of information technology. Background Art

[0003] With the rapid development of big data technology and artificial intelligence, natural language processing (NLP) technology has gradually become an important tool in the field of scientometric evaluation, especially in the analysis and evaluation of large-scale complex natural language content. However, although the existing natural language processing technology based on large language models can recognize and process human language texts, due to the limitations of model scale, computing performance, and context processing length, it has not played a significant role in the field of scientometric evaluation.

[0004] Currently, the scientometric tasks of large-scale complex natural language content mainly rely on human logical reasoning for analysis and processing, resulting in low task execution efficiency, poor result consistency, and difficulty in meeting the rapid and precise evaluation requirements. Especially in the fusion of multi-source heterogeneous data and the parsing of complex format content, traditional methods are particularly insufficient and difficult to cope with the rapidly growing evaluation requirements and increasingly complex application scenarios. Summary of the Invention

[0005] Based on a number of technologies such as large language models, agent design, reinforcement learning, and big data analysis, the present invention proposes an innovative multi-level agent architecture, aiming to solve the problem of insufficient agent performance in complex multi-task environments, and applies it to the fields of large-scale data analysis and complex document processing, forming a system with high productivity. Through technology integration, the present invention realizes the high-efficiency, intelligent, and automated data processing, and has significant practical value and innovative significance.

[0006] The technical solution adopted by the present invention is as follows:

[0007] A multi-level dynamic intelligent automated data processing and information evaluation method, comprising the following steps:

[0008] Construct a database environment job dataset, and combine reinforcement learning to conduct targeted training and fine-tuning on the large language model to achieve fully automated data analysis, query, and information extraction of multiple tables in the database;

[0009] Establish a multi-level intelligent agent interaction and collaboration architecture, which includes a scheduling intelligent agent at the upper layer and a job intelligent agent at the lower layer. The scheduling intelligent agent is responsible for global decision-making and policy planning, and the job intelligent agent is responsible for the execution of specific tasks. The scheduling intelligent agent and the job intelligent agent share information and coordinate tasks to form a highly collaborative whole;

[0010] Establish a cyclic supervision and error correction mechanism. The scheduling intelligent agent learns the normative workflow examples of the target working environment and real-time tracks the status of the job intelligent agent, identifies the behaviors of the job intelligent agent deviating from the predetermined path, and takes corrective measures.

[0011] Further, the construction of the database environment job dataset includes constructing an inference-execution chain dataset based on the ReAct technology, and constructing a policy dataset of the intelligent agents in the multi-level intelligent agent interaction and collaboration architecture.

[0012] Further, the construction of the inference-execution chain dataset based on the ReAct technology includes:

[0013] Historical data analysis: Collect and analyze a large number of historical database query logs, and extract common database operation patterns and user behavior characteristics from them;

[0014] Workflow data integration: Obtain the log-form record data generated during the operation of the intelligent agent in the business process, and conduct in-depth analysis on it to extract dynamic trend change information;

[0015] Synthetic data generation: Use a machine learning method that combines a generative adversarial network and a variational autoencoder to generate synthetic data.

[0016] Further, the construction of the policy dataset of the intelligent agents in the multi-level intelligent agent interaction and collaboration architecture includes:

[0017] Adopt a unified quantization mechanism, and use a state quantization module to uniformly quantify communication and state data;

[0018] Utilize the reasoning-execution chain dataset based on the ReAct technology and combine the requirements in the actual application scenario to establish a policy dataset suitable for the Proximal Policy Optimization (PPO) algorithm; the policy dataset includes the state of the agent in a specific environment, the actions taken, the immediate rewards obtained, and the predicted next possible state; the data of the policy dataset comes from the policy data output and recorded by the agent, the communication data of multi-agent interaction, and the data of the selected high-quality open-source dataset.

[0019] Furthermore, the scheduling agent undertakes the overall planning responsibility from a global perspective, and multiple job agents share the specific implementation work according to their respective expertise, forming a "centralized decision-making + distributed execution" mode; the scheduling agent and the job agents maintain close contact through a standardized information transfer protocol; whenever a new task arrives or the existing plan needs to be adjusted, the scheduling agent notifies the relevant job agents; the scheduling agent automatically detects and records all online job agents and their attribute information; according to the nature of the task and the degree of matching of the capabilities of each job agent, the scheduling agent selects the most suitable job agent, and the selected job agent starts to handle the assigned task, during which it shares resources or exchanges intermediate results with other job agents; after completion, the job agent submits the final output back to the scheduling agent, and the scheduling agent integrates it or directly presents it to the user as needed.

[0020] Furthermore, the job agent includes a database interaction job agent; the database interaction job agent has the ability to comprehensively perceive the database architecture and the data structure of each table, the ability to accurately locate and analyze specific data, and the high execution ability for complex database operations.

[0021] Furthermore, the standardized information transfer protocol includes a unified quantization mechanism and a state quantization module. The unified quantization mechanism converts various types of state information into a comparable and analyzable form, and the state quantization module is responsible for collecting and processing the quantized state information in real time to support the accuracy requirements in the subsequent decision-making process.

[0022] Furthermore, the scheduling agent establishes a detailed analysis and breakdown of the overall task based on the requirements in the initialization prompt engineering provided by human engineers and a small number of high-quality normative workflow examples, and generates a corresponding response workflow framework. Then, through a standardized information transfer protocol, it sends specific task requirements and information to each job agent in a standard structured format. The job agent queries for applicable tools from the toolkit system according to the specific task requirements and information assigned by the scheduling agent. If there is no tool in the existing toolkit system that adapts to this work requirement, the code generation module in the toolkit system is activated to automatically generate a new tool. After the code generation in the toolkit system is completed, the job agent returns the specific results of the new tool execution to the scheduling agent, and the scheduling agent evaluates the results returned by the job agent and implements the loop supervision and error correction mechanism.

[0023] Furthermore, the loop supervision and error correction mechanism includes:

[0024] Based on the standardized information transfer protocol, the scheduling agent continuously monitors and evaluates the working status and progress of the job agent.

[0025] The scheduling agent uses the policy data set for reinforcement learning training to acquire the ability to identify whether the job agent deviates from the preset goal or has an operation deviation.

[0026] When the scheduling agent detects any abnormal situation, it quickly takes intervention measures to correct the problem. The intervention measures include adjusting the resource allocation strategy, re-planning the task process, optimizing the algorithm configuration, and activating the backup work path to ensure that each job agent can smoothly carry out its work process according to the predetermined goal and plan.

[0027] The scheduling agent has an adaptive ability and automatically adjusts the coping strategy according to the current environmental conditions and available resources when encountering unforeseen situations.

[0028] The scheduling agent uses its accumulated knowledge base or accesses external professional information sources to generate specific solutions for specific problems. When the job agent faces challenges, the scheduling agent puts forward optimization suggestions according to the specific situation and then conveys the optimization suggestions to the corresponding job agent through the standard information transfer protocol. In case of a serious deviation from the expected goal, the scheduling agent has the right to directly interrupt the relevant job and instruct the job agent to restart the execution process.

[0029] A multi-level dynamic intelligent automation data processing and information evaluation system, which includes:

[0030] The database automatic interaction module based on reinforcement learning is used to construct a database environment job dataset, and combined with reinforcement learning, it conducts targeted training and fine-tuning on the large language model to achieve fully automatic data analysis, query, and information extraction of multiple tables in the database;

[0031] The multi-level agent interaction and collaboration module is used to establish a multi-level agent interaction and collaboration architecture, which includes a scheduling agent at the upper layer and a job agent at the lower layer. The scheduling agent is responsible for global decision-making and policy planning, and the job agent is responsible for the execution of specific tasks. The scheduling agent and the job agent share information and coordinate tasks to form a highly collaborative whole;

[0032] The loop supervision and error correction mechanism module is used to establish a loop supervision and error correction mechanism. The scheduling agent learns the normative workflow examples of the target working environment and real-time tracks the status of the job agent, identifies the behaviors of the job agent deviating from the predetermined path, and takes corrective measures.

[0033] The beneficial effects of the present invention are as follows:

[0034] The present invention proposes a database automatic interaction technology based on reinforcement learning. By constructing a high-quality database environment job dataset and combining reinforcement learning for targeted training and fine-tuning of the large language model, it realizes fully automatic data analysis, query, and information extraction of multiple tables in the database. The system can complete the whole process from data acquisition to result presentation without manual intervention in a complex database structure, effectively improving the efficiency and accuracy of data processing. By combining the reinforcement learning algorithm with the large language model, the system can continuously optimize the behavior strategy in a simulated real interaction scenario, adjust the execution path according to the feedback of the objective function, and achieve the optimal processing effect. At the same time, by learning a large number of actual operation cases, the system gradually masters the processing methods of different data types, significantly enhancing its ability to handle unknown situations. This technology can not only meet the needs of daily complex data analysis but also quickly respond to sudden data changes or anomalies, ensuring the stability and continuity of the data processing process, and providing users with an efficient and reliable data management solution.

[0035] In terms of multi-agent interaction and cooperation, the present invention proposes a multi-level agent interaction architecture. By designing a hierarchical agent network, each agent has clear functions and responsibilities in the architecture. The upper-level agents are responsible for global decision-making and strategy planning, while the lower-level agents focus on the execution of specific tasks. Through an efficient communication mechanism, agents can share information and coordinate tasks, thus forming a highly collaborative whole. This architecture can flexibly cope with the challenges of various complex environments, demonstrating excellent performance in areas from big data analysis to real-time information processing. Compared with traditional single-agent systems, this architecture solves bottleneck problems such as response latency and resource waste that may be encountered when facing large-scale concurrent tasks, providing a more scalable, robust, and easy-to-maintain task processing platform, which is suitable for a wide range of application scenarios such as enterprise-level data analysis and intelligent decision-making support.

[0036] The present invention proposes a few-shot learning loop supervision and error correction mechanism. By combining a normative workflow example library and real-time monitoring functions, it significantly improves the learning efficiency and adaptability of the system. Domain experts or experienced operators design and create a series of customized workflow examples for the target working environment, providing high-quality learning materials for the agents, thereby reducing deviations and misoperations that may occur in multi-agent collaboration. At the same time, the system introduces quantified communication data. By real-time tracking the status of lower-level job agents, the upper-level scheduling agents can quickly identify behaviors deviating from the predetermined path and immediately take corrective measures. By adjusting task parameters, allocating resources, or sending correction instructions, etc., the system can ensure the accuracy and consistency of task execution and maintain a high degree of stability and reliability in complex environments.

[0037] Combining multiple technological innovations, the present invention effectively solves the problems of low efficiency, insufficient accuracy, and weak adaptability of current large language models in the field of data processing and complex task environments. It not only improves the intelligent level of data processing but also significantly enhances the system's autonomous decision-making ability and rapid response ability to complex environmental changes. Through the database interaction technology of reinforcement learning, the system can automatically learn the optimal data access strategy, reduce incorrect queries and unnecessary operations, and greatly improve data processing efficiency; through the multi-level agent collaboration architecture, efficient cross-module information sharing and seamless collaboration are achieved among agents, ensuring the consistency and coherence of the system; and through the loop supervision and error correction mechanism, the system can quickly adapt to new workflows and task patterns based on a small number of high-quality examples, effectively reducing the error rate and improving the quality of output results.

[0038] The present invention demonstrates significant advantages in large-scale data analysis and complex information evaluation, comprehensively improving the quality and speed of data processing, enhancing the robustness and flexibility of the system, reducing the dependence on manual intervention, and promoting the automation development of the entire process from data collection to decision support. Through highly intelligent natural language processing technology, the system provides users with a friendly and intuitive interaction experience, supporting data analysis and information evaluation requirements in various working environments and task modes. This technical solution not only improves the existing solutions but also provides cutting-edge technical support and new development directions for complex information processing. Brief Description of the Drawings

[0039] Figure 1 It is the main step flow chart of the method of the present invention.

[0040] Figure 2 It is the work flow chart of the multi-level agent interaction and collaboration architecture. Detailed Embodiment

[0041] In order to make the objectives, features, and advantages of the present invention more obvious and understandable, the present invention will be further described in detail below through specific examples and drawings.

[0042] 1. Objective of the Invention

[0043] Aiming at the scientometric tasks of large-scale complex natural language content, the present invention aims to overcome the problems of low efficiency, insufficient adaptability, and unstable result quality in the prior art. Currently, scientometric tasks still rely on static rules and a large amount of human resources to complete. Especially in the parsing and evaluation of large-scale complex natural language content, traditional methods require a large amount of human effort to adapt the environment and workflow, and it is difficult to meet the rapidly changing data requirements. This highly human-dependent method is not only costly but also has low adaptation efficiency, seriously restricting the scale and automation of task processing.

[0044] Due to the lack of an efficient system architecture and an effective supervision and error-correction mechanism in traditional automation systems, it is difficult to achieve smooth interaction and collaboration among their components, resulting in problems during operation being difficult to detect and correct in a timely manner. At the same time, with the rapid growth of data volume and the increasing complexity of content in the scientific metrology scenario, the existing technologies face multiple challenges. First, the management and analysis of large-scale complex natural language content still require a large amount of human support, which is inefficient and the consistency of results is difficult to guarantee. Second, traditional methods have insufficient parsing capabilities for unstructured or semi-structured data, making it difficult to conduct in-depth information mining and comprehensive evaluation, thus limiting the ability to effectively utilize complex information. In addition, the adaptability and robustness of existing systems are poor, and their performance is likely to drop significantly or even fail when facing unforeseen scenarios or changes in data distribution. At the same time, there are also large fluctuations in the quality and stability of the output results, which cannot meet the requirements of scientific metrology tasks for highly reliable information.

[0045] To address these problems, the present invention proposes an efficient and intelligent solution by integrating reinforcement learning-driven data interaction management, advanced natural language processing technologies, and a multi-level multi-agent collaborative architecture. The main content is as Figure 1 follows: It proposes an automatic database interaction technology based on reinforcement learning. By constructing a high-quality database environment job dataset and combining reinforcement learning to conduct targeted training and fine-tuning of large language models, it realizes full-automatic data analysis, query, and information extraction of multiple tables in the database. It proposes a multi-level agent interaction architecture. By designing a hierarchical agent network, each agent has clear functions and responsibilities in the architecture. The upper-level agents are responsible for global decision-making and strategy planning, while the lower-level agents focus on the execution of specific tasks. It proposes a few-shot learning loop supervision and error-correction mechanism. By introducing the few-shot learning loop supervision and error-correction mechanism, the present invention significantly enhances the adaptability, robustness, and the quality and stability of the output results of the system in scientific metrology tasks. This design significantly improves the parsing ability for complex natural language content and provides a method for efficiently and automatically conducting scientific metrology and evaluation of large-scale complex natural language content in the field of scientific metrology.

[0046] 2. Invention Solution

[0047] The large-scale data and complex information processing system applied in the present invention's solution constructs multiple high-quality, highly practical, and innovative datasets, which specifically include the following content:

[0048] First, it constructs a reasoning-execution chain dataset based on ReAct technology.

[0049] ReAct (Reasoning and Acting) is a methodology that combines reasoning and acting, aiming to enhance the understanding and decision-making abilities of machine learning models by constructing a complete chain from thinking to execution. This technology allows agents to not only generate reasonable responses when facing complex tasks but also self-adjust according to environmental feedback, thus achieving more efficient task processing. Compared with ordinary prompt words: The simple prompt words used in traditional methods often can only guide the model to produce answers in a single direction and lack flexibility. The ReAct technology, by introducing a multi-step reasoning process, enables the model to make decisions after comprehensively considering various factors based on understanding the problem background, greatly improving the quality and diversity of answers. Compared with CoT (Chain of Thought) prompt engineering: Although the CoT method also emphasizes the importance of logical reasoning, it is usually limited to the field of text generation and mainly relies on predefined reasoning paths. In contrast, ReAct not only supports advanced reasoning in natural language processing scenarios but also applies to cross-modal tasks such as information evaluation work in complex environments. In addition, under the ReAct framework, immediate feedback can be obtained after each action, which helps to form a closed-loop learning mechanism and further improves the learning efficiency and adaptability of the model.

[0050] To create a dataset suitable for training the reasoning-execution chain based on ReAct technology, the present invention takes the following steps:

[0051] 1) Historical data analysis: First, a large number of historical database query logs are collected and analyzed, and common database operation patterns and user behavior characteristics are extracted from them. This information provides a basis for subsequent simulation of real-world interactions.

[0052] a) Extract database operation patterns:

[0053] In the process of constructing a multi-level dynamic intelligent automation data processing and information evaluation system, in order to accurately extract common operation patterns, the present invention first focuses on the collection and preprocessing of historical data. By connecting to the log systems of multiple databases, detailed records covering various types of database operations are obtained, including but not limited to the usage of basic functions such as addition, deletion, modification, and query. Next, using natural language processing techniques and machine learning algorithms, these raw logs are cleaned, parsed, and classified to identify frequently occurring operation sequences and their context. On this basis, based on the ReAct thought chain framework, each identified operation sequence is transformed into a series of task instructions that are easy for the LLM to understand and execute. For example, when processing data queries, the whole process is broken down into specific steps such as constructing query conditions, optimizing query statements, executing queries, and processing results, and clear guidelines are provided for each step. In addition, for operations involving data modification or deletion, special attention is paid to aspects such as permission verification, transaction management, and data consistency checking to ensure that the generated operation paradigms meet both security standards and practicality. This series of steps enables the model to learn stable and reliable database operation patterns from a large amount of historical data, thus laying a solid foundation for subsequent intelligent processing.

[0054] b) Extract user behavior characteristics:

[0055] Subsequently, in order to improve the pertinence and adaptability of the model, the present invention further deeply explores user behavior characteristics. The key here is not only to rely on the daily operation records of general developers, but also to actively introduce the standardized operation processes specifically designed by developers for specific tasks. Through a customized log collection mechanism, all interaction details of each user during task execution are accurately captured, including the commands used, parameter settings, and possible problems and solutions encountered. These data are used to train a sub-model specifically for user behavior characteristic analysis, which can automatically extract operation patterns reflecting personal work habits and technical preferences on the basis of understanding user intentions. Specifically, it will dynamically adjust and optimize the generated operation guides according to information such as the user's operation frequency, preferred tools, and methods for solving specific problems. At the same time, with the help of the ReAct thought chain, these personalized operation guides are seamlessly integrated into the overall system to ensure that they can complement the general operation patterns. Finally, by combining the data analysis results from these two sources (database operation patterns and user behavior characteristics), the present invention realizes the efficient simulation and optimization of database operations, greatly improving the service quality of the model and its adaptability to business scenarios.

[0056] 2) Workflow Data Integration: Workflow data integration is a crucial part of the present invention. It not only covers static historical records but also pays particular attention to the ongoing workflow within existing systems. By monitoring these activities in real time, the system can capture the latest trend changes to ensure that the dataset always reflects the current business requirements. During the second step of the workflow data integration process for constructing the ReAct dataset, a dedicated logging mechanism is designed to track every key step and decision point when the Agent executes tasks. Specifically, whenever the Agent starts a new workflow or responds to an external event, the system automatically triggers the log generation module, which is responsible for recording detailed information, including but not limited to timestamps, operation types, involved objects, and result statuses. To maintain the richness and organization of the log content, a structured log format is introduced. By defining fixed fields to organize each log record, it is not only convenient for subsequent data processing but also improves query efficiency. After accumulating a certain amount of log data, the system uses advanced data analysis algorithms to deeply analyze these logs and extract valuable dynamic trend change information. This process starts from the preprocessing stage, where the raw logs are cleaned and standardized to ensure data consistency and integrity; then it enters the feature extraction phase, where representative behavior patterns and change rules are mined based on predefined pattern recognition rules or machine learning models. For example, analyzing the change trajectory of resource usage can evaluate the effect of system performance optimization, tracking the task execution efficiency of each Agent over time to identify the reasons for efficiency improvement or delay, providing a basis for optimizing the workflow; at the same time, monitoring the transformation of user behavior patterns, such as the relationship between the intensity of functional requirements and time periods and the frequency of preference adjustments, helps predict demand fluctuations and guide service resource allocation; in addition, the log data supports cross-version performance comparison and interaction path analysis to ensure that updates and iterations do not affect existing services and capture user operation habits to optimize the user experience design. Finally, all the extracted trend information will be integrated into the ReAct dataset, providing solid data support for the continuous improvement of the system, while also enhancing the ability to predict future changes, making the automated data processing and information evaluation system more intelligently adapt to the evolving business environment and technical requirements.

[0057] 3) Synthetic data generation: Given the possible lack of sufficient real-world cases in certain specific situations, the present invention uses several techniques, including generative adversarial networks (GANs), to produce additional synthetic samples. In the synthetic data generation process, considering the data scarcity problem that may be encountered in practical applications, the present invention not only adopts the generative adversarial network (GANs) technology to create additional synthetic samples, but also combines the machine learning method of variational autoencoders (VAEs). These techniques work together to enrich the content of the dataset and maintain its diversity, ensuring that the generated data matches the real data in distribution, thus facilitating the improvement of the generalization ability of the model. In addition, by introducing rule-based synthetic algorithms and semi-supervised learning strategies, the system can precisely control the characteristics of the synthetic data under the guidance of specific domain knowledge, enabling the creation of high-quality and highly relevant synthetic datasets even in the absence of a large amount of labeled data, further enhancing the effectiveness and applicability of model training. This method not only solves the problem of insufficient data volume but also effectively avoids overfitting, providing a solid guarantee for the accuracy and reliability of the multi-level dynamic intelligent automated data processing and information evaluation system.

[0058] In the above steps, the "historical data" in the first step refers to the database data generated by humans, including the data indirectly generated by the system audience and the behavioral data directly generated by developers, etc. The "workflow data" in the second step refers to the recorded data in the form of logs generated by the intelligent agent system during the operation of business processes. In some cases, the data generated in the first and second steps is too small, or the amount of data required for training the AI is too large and cannot be satisfied by the first two steps. In such cases, it is necessary for humans to screen a small amount of high-quality data from the first two steps, and then after specific processing, give it to the data synthesis algorithm for synthetic data generation, that is, the "synthetic data" in the third step. The above three types of data belong to three independent data from different sources.

[0059] In addition, quality control and annotation can also be carried out, that is, experienced data analysts conduct regular sampling inspections to maintain a high-quality level of data. It should be noted that this step is not a necessary step and can be selectively used based on the above steps to achieve better results.

[0060] Second, a strategy dataset of agents within a multi-level intelligent agent interaction and collaboration architecture that has undergone unified quantization processing is constructed.

[0061] In the present invention, a strategy dataset designed specifically for the multi-level intelligent agent interaction and collaboration architecture is constructed. The purpose of this dataset is to improve the overall system performance by optimizing the communication efficiency and decision-making quality among agents. The construction process and its application of this dataset will be introduced in detail below.

[0062] 1) Unified Quantization of Communication and Status Data: To ensure efficient communication between intelligent agents at different levels, the present invention first introduces a unified quantization mechanism. This mechanism allows intelligent agents to generate step-by-step streaming outputs through the inference architecture, which are then converted into concise and easily parsable status data by a specially designed status quantization module. Such status data not only simplifies the information transmission process but also improves data transmission efficiency, enabling the upper-level scheduling intelligent agent to obtain the work progress of the lower-level execution intelligent agent more quickly and promoting data sharing and collaboration among peer job intelligent agents.

[0063] 2) Based on the above-mentioned inference-execution chain dataset based on the ReAct technology and combined with the requirements in the actual application scenario, the present invention further develops a set of policy datasets applicable to the Proximal Policy Optimization (PPO) algorithm framework. This dataset contains key information such as the state, action, immediate reward obtained, and predicted next possible state of the intelligent agent in a specific environment. Half of the data in this policy dataset comes from the policy data output and recorded by the intelligent agent, and the other half comes from the communication data within the multi-intelligent agent interaction framework. Among them, the policy data output and recorded by the intelligent agent are used to optimize the performance of specific job intelligent agents, while the communication data of multi-intelligent agent interaction is used to optimize the cooperation strategy between intelligent agents and the performance of supervisory intelligent agents. In addition, this dataset also combines data from some selected high-quality open-source datasets to improve its generalization ability.

[0064] The above-mentioned policy dataset is established through the following steps:

[0065] a) Recording the Inference Process of Intelligent Agents:

[0066] Capture the streaming intermediate result outputs generated by intelligent agents when performing tasks or solving problems. These outputs can be any form of data such as decisions, evaluations, predictions, etc., reflecting the current state and action logic of the intelligent agent.

[0067] Collect the communication data between intelligent agents and ensure the consistency and interpretability of the interaction between different intelligent agents according to the standardized information transmission protocol (such as API calls, message queues, event-driven architectures, etc.).

[0068] b) Design and Application of the Unified Quantization Mechanism:

[0069] Develop a specially designed unified quantization mechanism for processing and converting the above-collected data. This mechanism aims to convert unstructured or semi-structured data into structured data with a clear format and meaning.

[0070] The structured data should include, but not be limited to, the following elements:

[0071] State: A set of information describing the environmental conditions of the agent.

[0072] Action: The behavior or decision taken by the agent based on the current state.

[0073] Immediate Reward: The immediate feedback or benefit obtained after performing a specific action in a given state.

[0074] Next Possible State: The possible future state predicted based on the current state and action.

[0075] c) Dataset Expansion and Quality Improvement:

[0076] Integrate publicly available high-quality datasets in related fields to enhance the breadth and depth of internally generated data. This step not only increases the data volume, but more importantly, improves the adaptability and generalization performance of the training model for unseen scenarios.

[0077] Select external data sources closely related to the target application scenario to ensure the relevance and applicability of their content, thereby further optimizing the performance of the final model.

[0078] d) Data Preprocessing and Cleaning

[0079] Before incorporating all data into the final policy dataset, implement necessary preprocessing steps, such as removing noise, filling in missing values, and resolving data inconsistencies.

[0080] Perform data cleaning to ensure the quality of the dataset, including checking the integrity and accuracy of the data, and handling outliers to make the data more reliable and suitable for machine learning algorithm training.

[0081] e) Continuous Iteration and Improvement

[0082] As more experimental data accumulates and new knowledge is discovered, continuously update and improve the policy dataset. By regularly evaluating the effectiveness and usability of the existing data, timely adjust the data collection methods and technical means to ensure that it can always support the latest research and development needs.

[0083] 3) PPO algorithm calculation: In the proximal policy optimization PPO algorithm, the calculation of the advantage function is achieved by the time difference estimation (TD) or generalized advantage estimation (GAE) method. The present invention accurately calculates the advantage value of the agent after taking action in each state. This process is crucial for accurately evaluating the effectiveness of the current strategy and provides a key basis for subsequent strategy adjustments. In order to guide the learning process of the PPO algorithm, the present invention designs an objective function that aims to balance the need to explore new strategies and make full use of existing experience. This balance ensures that the learning process is neither too conservative nor blindly adventurous. Based on this objective function and its gradient information, the gradient descent method or other optimization techniques are used to gradually adjust the policy parameters of the agent. The iterative update process prompts the system to evolve towards a more efficient and stable solution. Finally, the present invention integrates the above steps into a closed-loop feedback system, and restarts the training cycle with the latest version of the policy parameters after each round of training. This process is repeated until the predetermined performance index is reached. This method ensures that the PPO algorithm can continuously optimize the strategy and ultimately obtain a satisfactory control effect.

[0084] Among them, the method to accurately calculate the advantage value after the agent takes action in each state is:

[0085] a) Environmental interaction and data collection

[0086] The agent performs action a in the environment according to the current policy π and observes the next state s' and the immediate reward r.

[0087] The collected data form a series of trajectories, where T is the end point or maximum length of the sequence.

[0088] τ=(s0,a0,r0,s1,a1,r1,…,s T )

[0089] Among them, τ represents a state-action-reward sequence, also known as a trajectory or episode. τ is an ordered tuple that records the historical sequence of states, actions, and rewards generated by the agent in the process of interacting with the environment, starting from the initial state, performing a series of actions, and obtaining corresponding rewards until reaching the terminal state or reaching the preset maximum number of steps.

[0090] b) Calculate rewards and evaluate value functions

[0091] For each time step t, the cumulative reward from that moment on is calculated using a discount factor γ:

[0092]

[0093] Among them, T max represents the maximum time step of the trajectory, also known as the maximum number of steps or the time horizon. T max is a pre-set positive integer used to limit the length of a single trajectory of the interaction between the agent and the environment. When the time step t of the interaction between the agent and the environment reaches T max , even if the environment itself has not reached the termination state, the current trajectory will be forced to end. k represents the summation index, also known as the counting variable or the iteration variable, which is used to indicate the starting value and the increment step of the summation terms. It starts from the initial value of k (here it is 0), increments by 1 each time (by default), until it reaches the end value specified by the superscript of the summation symbol. In each iteration, the summation terms are calculated according to the current value of k, and the results are accumulated into the sum. r t+k represents the immediate reward obtained by the agent at time step t + k. r t+k is a scalar value representing the reward signal feedback from the environment to the agent at time step t + k after the agent executes a certain action. It reflects the quality of the action taken by the agent at that time step. Among them, t represents the current time step, and k represents the future number of steps relative to the current time step.

[0094] The value function V(s) is used to estimate the value of state s, that is, the expected future cumulative reward. For a given state s t , V(s t ) ≈ E[R t . Among them, E[R t represents the expected value of the cumulative return R t at time step t, also known as the expected return. E[R t is a mathematical expectation, which represents the weighted average of all possible future state-action-reward sequences (i.e., trajectories τ) under the given policy π and the starting state s t , and is the average value of the cumulative return R t .

[0095] c) Calculate the advantage function through temporal difference estimation

[0096] Temporal difference error (TD error):

[0097] δ t = R t + γ·V(s t+1 ) - V(s t )

[0098] δt It represents the immediate reward plus the estimated value of the next state minus the estimated value of the current state.

[0099] Advantage function A(s,a):

[0100] A(s,a) = Q(s,a) - V(s)

[0101] where Q(s,a) is the value of the state-action pair and can be estimated by methods such as TD(0), n-step TD, etc.

[0102] d) Generalized Advantage Estimation

[0103] GAE (Generalized Advantage Estimation) combines the unbiasedness of Monte Carlo estimation and the low variance advantage of TD estimation, and is defined as:

[0104]

[0105] where, represents the Generalized Advantage Estimation value calculated using the parameter λ (lambda) at time step t. is a specific estimation method for the advantage function A(s t ,a t ), which calculates by taking the weighted average of the TD errors δ t+l over multiple time steps. Among them, the weights are determined by the l-th power of the product of the discount factor γ and the parameter λ (γλ) l . λ is a hyperparameter between 0 and 1, used to balance the bias and variance of the estimation. l represents the summation index, also called the counting variable or iteration variable. L max represents the upper limit of summation, indicating the maximum value of the summation index l. L max is a pre-set value, used to limit the number of steps for accumulating TD errors in the GAE calculation. It determines how many future time steps of TD errors to consider when calculating . λ is a parameter between 0 and 1, used to balance bias and variance; when λ is close to 1, it is more biased towards Monte Carlo estimation, while when λ is close to 0, it is more inclined to TD estimation.

[0106] e) Policy update in PPO

[0107] The goal of PPO is to maximize the objective function J(θ) while keeping the Kullback-Leibler Divergence between the old and new policies small.

[0108] f) Implement policy update

[0109] In practice, the above process is batched, i.e., the policy is updated once based on multiple environment samples. In the update loop, sampling, calculating advantages, and adjusting the policy are repeated until a predetermined number of iterations or convergence conditions are reached. The PPO algorithm can effectively evaluate the advantage level of an agent taking a specific action by introducing methods of temporal difference estimation and generalized advantage estimation, and ensure the stability of the training process by controlling the update amplitude.

[0110] In the Proximal Policy Optimization (PPO) algorithm, the maximized objective function is to ensure that the update between the old and new policies is stable and does not change drastically due to a single update. The mathematical formula of the objective function is as follows:

[0111] J(θ) = E t [min(r t (θ)A t , clip(r t (θ), 1 - ∈, 1 + ∈)A t )]

[0112] Details of the specific parameters in the objective function formula are as follows:

[0113] θ: Represents the parameter vector of the current policy π. In the context of deep learning, these parameters usually refer to the neural network weights.

[0114] E t : This is an unbiased estimate of the expected value of the advantage estimate A t at time step t. In practice, this expectation is usually approximated by a finite number of time steps sampled from the environment.

[0115] r t (θ): This is the probability ratio, i.e., at time step t, the probability of taking action at according to the new parameter θ divided by the probability of taking the same action according to the old parameter θ old . It reflects the degree of difference in the selection of a specific action between the new policy and the old policy given the state s t .

[0116]

[0117] A t: This is the estimated value of the advantage function at time step t, which measures the value of taking a specific action relative to the average action in a given state. It is calculated by Generalized Advantage Estimation (GAE) or other methods.

[0118] clip(r t (θ), 1 - ∈, 1 + ∈): This is a clipping operation used to limit the range of r t so that it cannot deviate too far from 1. Here, ∈ is a small positive number that defines the clipping range. This operation helps prevent performance crashes or divergence caused by overly large updates.

[0119] min(r t (θ)A t , clip(r t (θ), 1 - ∈, 1 + ∈)A t ): This part selects the smaller of two values as part of the objective function. This is done to ensure that even when r t exceeds the clipping bounds, the new policy is not overly rewarded or punished, thus maintaining the conservativeness and smoothness of the update.

[0120] The objective function of PPO aims to achieve a stable learning process by controlling the update magnitude while maximizing the policy's performance as much as possible. By minimizing the selection between two candidate objectives - one based on the original probability ratio multiplied by the advantage estimate and the other based on the clipped probability ratio multiplied by the advantage estimate - large changes during the policy update process can be effectively prevented, thereby ensuring stability during the training process.

[0121] In addition to high-quality and highly practical datasets, the present invention also designs and applies the following innovative technologies and systems, which specifically include the following:

[0122] First, an interface for writing a targeted and standardized workflow example library for target working conditions is designed. In the present invention, a flexible and efficient interface is designed and implemented, which allows users to create and manage a standardized workflow example library for specific application scenarios (i.e., target working conditions). This functional module is of great significance for quickly adapting to new environments, improving system response speed, and enhancing the flexibility of the agent system. When an agent based on a large language model executes a task, it will quickly adapt to the standardized workflows in the example library by learning from a small number of samples in the prompt words. The agent will quickly learn the paradigms in a specific field, specific working conditions, and specific task patterns, and use this as a reference to define a tone for subsequent workflows. This adaptive learning mechanism greatly reduces the time cost and technical threshold required to design workflows from scratch for each new application scenario. In addition, the interface supports direct intervention and adjustment of the content of the example library by human administrators. When encountering unexpected new situations, the administrator can upload updated documentation according to the actual situation or directly edit existing records, enabling the entire system to immediately reflect the latest business rules or operation guides. This approach not only improves the flexibility of the system but also enhances its ability to handle complex and changing environments. Moreover, it also provides an effective means to introduce the knowledge and experience of domain experts, further enhancing the professionalism and accuracy of the decision-making process. The state quantization module in the unified quantization mechanism of the standardized communication transfer protocol is designed to not only input various types of record data mentioned above and then quantize it into the data structure for agent training, but it also has another entry. By inputting an explanatory document described in rigorous and standardized natural language, after analysis by the agent, the behavioral characteristics of human experts are summarized and stored in the form of a standard quantization data structure.

[0123] Meanwhile, this standardized workflow example library will be introduced as a reference information source into the ReAct reasoning-execution chain of the agent. As the best practice cases for the agent to refer to when dealing with specific problems, this standardized workflow example library will be designed as a reference information source for the agent's thinking. When facing an unfamiliar working condition or task pattern, during the Reasoning stage, the agent can access the standardized workflow example library to find solutions in similar situations as a reference basis, thereby optimizing its ReAct reasoning-execution chain. This will greatly improve the adaptation speed and ability of the agent system in new working conditions and new task patterns.

[0124] The premise of creating a normalized workflow example library is to do so when there is not enough valid data in the current environment or the existing data volume is insufficient to support the agent to achieve the expected performance. This means that in the absence of a sufficient number of high-quality training samples, by carefully designing and inputting a small number of representative working condition instances manually, initial learning materials can be provided for the agent, enabling it to quickly become familiar with and adapt to the target working conditions. These cases written by domain experts not only cover the operation processes of typical scenarios but also include strategic suggestions for dealing with emergencies, greatly improving the success rate of the initial deployment of the agent. The method of creating a workflow example library in the present invention is as follows: At the creation stage, a specific administrator, through the comprehensive task pattern recognition and definition work carried out, involves a detailed analysis of all potential task patterns and their variants in the target field in one step, and at the same time clarifies the specific operations to be performed under each task pattern. The administrator extracts key behavioral characteristics from domain research, historical data analysis, and literature reviews and transforms them into a data structure that conforms to a unified quantification mechanism. The design of these data structures fully considers the learning efficiency of the agent, ensuring fast and accurate task pattern adaptation even under limited sample conditions.

[0125] Second, a dynamic and diverse toolkit system provided for the agent to analyze and execute various instructions and tasks. In the present invention, a highly integrated and flexible toolkit system is designed and implemented, which is designed to provide the agent with the necessary tools and resources to support its efficient execution of various data processing and information evaluation tasks. The toolkit system not only covers diverse preset tools based on the LangChain framework but also includes SQL module tools designed specifically for database operations. In addition, for several core agents, the present invention particularly introduces a code generation and execution module, greatly enhancing the dynamic adaptability and robustness of the system.

[0126] First, by adopting LangChain as the infrastructure, the present invention enables the agent to be equipped with a series of diverse preset tools, which cover multiple aspects from text analysis, knowledge retrieval to complex logical reasoning. Each basic and general task tool based on the LangChain framework has been adaptively modified to meet the requirements of the standardized information transfer protocol framework and the data structure of the unified quantization mechanism within this system, ensuring that they are both independent and collaborative with each other, jointly constructing a powerful function set. For example, in the field of text analysis, tools such as sentiment analysis and entity recognition are provided; while in terms of knowledge retrieval, there is the ability to extract information based on large-scale corpora. Such a configuration enables the agent to select an appropriate combination of tools to work according to specific task requirements. At the same time, in order to better support operations involving structured data, the present invention adds an SQL module dedicated to database query and management to the tool kit, that is, the reinforcement learning database automatic interaction technology of the present invention. This set of tools allows the agent to directly interact with the backend database, execute complex query commands, and thus quickly obtain the required data set. At the same time, it also supports functions such as data update, insertion, and deletion, ensuring that the agent can dynamically adjust the data state during the processing.

[0127] Secondly, for the core agents undertaking critical tasks, relying solely on pre-set tools may not be sufficient to handle all possible situations. Therefore, a code generation and execution module developed based on advanced AI models is specifically equipped for them. This module allows agents to autonomously create new program codes and add them to the tool kit for use when facing new problems or situations that cannot be solved by existing tools. This mechanism endows agents with unprecedented flexibility, enabling them to continuously expand the boundaries of their capabilities to adapt to a wider range of application scenarios. The underlying technology supporting this feature is a large language model specifically trained for code generation. Such models have mastered the basic syntax structures of multiple programming languages and common software engineering practices through learning a large number of open-source projects. When an agent needs to solve a specific problem, it can obtain corresponding code snippet suggestions by inputting relevant descriptions to the model. Subsequently, after further optimization and verification, these automatically generated codes can be safely deployed to the running environment and become new tools for problem-solving. In this way, even in the absence of ready-made solutions, agents can quickly respond and take actions, greatly reducing the task stagnation time caused by the lack of appropriate tools. More importantly, with the accumulation of more and more customized tools, the overall performance of the entire system will be continuously improved, especially showing stronger adaptability and stability when dealing with complex and changing tasks. By introducing diverse pre-set tools and innovative code generation and execution mechanisms, the tool kit system proposed in the present invention not only meets the requirements of data processing and information evaluation but also has a powerful self-evolution ability, laying a solid foundation for achieving higher-level intelligent applications.

[0128] Thirdly, a database automated interaction system based on reinforcement learning. In the present invention, a database automated interaction system based on reinforcement learning is designed and implemented. This system not only enables agents based on large language models to seamlessly connect with relational databases, perceive information, and extract data but also, by introducing reinforcement learning algorithms, enables agents to continuously optimize their performance when executing data processing tasks and autonomously obtain the required information according to specific needs.

[0129] The agent uses the pre-set database connection tools within the LangChain framework to establish a communication channel with the target database. This process includes configuring necessary connection parameters (such as host name, port number, user name, and password, etc.) and sending requests through standard API interfaces to verify the connection status. Once the connection is successfully established, the agent can initiate further data query or operation instructions with the help of these tools. To ensure data security, an encrypted transmission protocol is adopted throughout the connection process, and strict permission management measures are taken for sensitive information. In addition, all operation records on external databases will be detailedly recorded for subsequent auditing and tracking.

[0130] In this design, the job agent responsible for executing database interaction tasks has the following advantages: it has a high situational awareness of the overall database data structure, a high positioning ability for specific database data, and a high implementation ability for relatively complex database operation requirements assigned by the scheduling agent. The following is a detailed introduction to the characteristics of this agent:

[0131] 1) Ability to perceive the overall database architecture and the data structures of each table

[0132] To enable the database interaction job agent to have high situational awareness, the following process is adopted in the design:

[0133] Metadata parsing: The agent first obtains the structure information of the database by executing specific SQL commands, including table names, field types, and their constraints. These metadata provide the basis for understanding the overall database architecture. These specific SQL commands are derived from the inference-execution chain datasets based on the ReAct technology collected by the system, which are the database operation behavior records generated by system users and developers during use and development. After constructing the analysis mode of the policy dataset, several sets of comprehensive database overall architecture analysis scripts are summarized to match the metadata parsing requirements of this agent.

[0134] Semantic understanding: On the basis of metadata parsing, to further enhance the high situational awareness of the database interaction job agent, the key process of semantic understanding is introduced in the design. Immediately after obtaining the structure information including table names, field types, and their constraints, the agent will use its built-in large language model to deeply analyze and interpret these raw metadata. Through this process, the agent can not only identify the association relationships between tables - such as primary-foreign key relationships, inheritance patterns, or many-to-many relationships - but also deeply understand the logical connections between fields, such as value ranges, default value settings, and possible calculation rules or trigger mechanisms. This high-level understanding is crucial for constructing a complete database knowledge graph, which enables the agent to go beyond simple data reading and writing operations and perform data analysis and processing at a more complex level. Through the profound insight into these complex relationships, the agent can more accurately locate the required information, optimize the query path, and even predict potential data anomalies or inconsistencies. In addition, based on this knowledge graph, the agent can also dynamically adjust strategies when performing data processing tasks and make optimal decisions according to changes in the context environment, thus significantly improving data operation efficiency and accuracy.

[0135] Dynamic Update: As the content of the database continues to evolve, whether it is the addition of new data records, structural changes, or schema adjustments, the agent will regularly trigger the re-scanning process automatically. During this process, it will not only synchronize the latest data changes but also perform incremental updates to the internal knowledge base to ensure that the information it holds is always up-to-date and accurate. This mechanism enables the agent to respond promptly to any changes in the database, maintaining the consistency and reliability of the services it provides externally. At the same time, by optimizing the update frequency and strategy, the agent can minimize resource consumption and improve operational efficiency while ensuring the timeliness of information.

[0136] 2) Ability to precisely locate and analyze specific data

[0137] Task Parsing and Intent Understanding: The agent first uses natural language processing technology to deeply parse the received task instructions. During this process, information such as the user's intent and the key entities involved in the task are parsed out. By combining context semantic analysis, the agent can accurately understand the user's needs and convert unstructured natural language into specific query parameters or operation commands. This process ensures that subsequent data search and analysis tasks can be executed efficiently and precisely.

[0138] Matching Strategy and Data Location: With the help of a pre-constructed comprehensive and detailed database knowledge graph, the agent can quickly determine the location of the target data after receiving the parsed task instructions. This knowledge graph not only records the table structure and its field information in the database but also includes complex relationship models between various tables, such as primary-foreign key associations, inheritance patterns, and many-to-many relationships. Using this rich metadata, the agent can quickly lock in the relevant data tables and fields based on the keywords in the task instructions and further filter out the most suitable dataset, thus achieving high-precision location of specific data.

[0139] 3) High execution ability for complex database operations

[0140] Query Optimization: Based on the understanding of the task, the agent will automatically select the optimal query path and algorithm. It will comprehensively consider factors such as the size of the data volume, access frequency, and index availability to determine the most efficient execution strategy. In addition, the agent also supports dynamic adjustment of the query plan and continuously optimizes performance based on real-time feedback.

[0141] Transaction Control: For complex operations involving multiple steps or requiring atomicity, the agent adopts a strict transaction management mechanism. All change operations are rehearsed in a temporary environment, and only when all steps are successfully completed and verified will they be officially committed to the database. This process effectively prevents data inconsistency problems caused by partial updates.

[0142] Rollback and Recovery: If an error or abnormal situation occurs during the operation, the agent can immediately initiate the rollback procedure, revoke the unconfirmed changes, and attempt to restore to the state before the operation. This mechanism greatly enhances the fault tolerance and stability of the system.

[0143] Permission Check: Before each complex operation starts, the agent will strictly check whether the operator has sufficient permissions. Only authorized operations are allowed to be executed, thus ensuring data security and privacy protection.

[0144] Audit Trail: All important database operations will be detailedly recorded to form a complete operation log. These logs can not only be used for subsequent audit work but also provide a valuable information source for troubleshooting.

[0145] To achieve the above three types of high-level capabilities, the agent needs to have a deep and refined perception and understanding of the target database. Therefore, for the data in the reasoning-execution chain based on ReAct regarding what database operations the agent reasons to execute and the relevant data of the database operations it finally implements, the present invention specially designs a knowledge graph module to record and store the relevant information. It provides the agent with a more efficient data utilization method based on the two types of record data sets of the present invention.

[0146] Fourth, multi-level agent interaction and collaboration architecture. In the present invention, a multi-level agent interaction and collaboration architecture is proposed, aiming to improve the overall performance and response speed of the system through an efficient task assignment, execution, and monitoring mechanism. The architecture consists of a Manager Agent (scheduling agent) and a Worker Agent (job agent), and realizes the effective management of tasks and the optimal allocation of resources through a clear role division. The work process of the multi-level agent interaction and collaboration architecture is as Figure 2 shown.

[0147] Scheduling Agent (Manager Agent): As the core control unit of the entire system, the Scheduling Agent is mainly responsible for the preliminary analysis, priority ranking of all incoming tasks, and reasonably allocating them to appropriate Worker Agents for execution. When receiving a new data processing or information evaluation request, the Manager Agent first uses natural language processing technology to understand the task description and determine its specific requirements. In addition, it continuously monitors the status of each Worker Agent, including the current load situation, available resource volume, etc., to ensure that tasks can be assigned to the most suitable execution location. Based on the current resource distribution in the system and historical experience data, the Manager Agent can also dynamically adjust the task allocation strategy to strive for the optimal processing efficiency. Once any potential problems or failures are detected, the Manager Agent will immediately take measures to intervene, such as reallocating tasks or notifying relevant personnel to intervene and solve the problem.

[0148] Job Agent (Worker Agent): The Job Agent is the operation unit that actually executes the specified tasks and directly participates in data collection, processing, analysis, etc. Each Worker Agent has specific domain expertise and technical capabilities. For example, some are designed to specifically handle large-scale data and analysis tasks, while others are designed with expertise in information evaluation in complex task environments. They can autonomously complete the assigned tasks after receiving instructions without additional manual intervention. After completing the tasks, the Worker Agents will promptly report the results to the Manager Agent and provide necessary status update information for subsequent work arrangements.

[0149] In the entire system, the scheduling agent undertakes the overall planning responsibility from a global perspective, while multiple job agents share the specific implementation work according to their respective specialties, forming a "centralized decision-making + distributed execution" model. The two are closely connected through a set of standardized message passing protocols to ensure smooth information exchange. The core of the standardized information transfer protocol of the present invention lies in two points: the first point is standardization, which is achieved by a unified quantization mechanism; the second point is what information to transfer. The implementation process for the second point is as follows: Whenever a new task arrives or an existing plan needs to be adjusted, the scheduling agent will promptly notify the relevant job agents. At the initial stage of system startup, the scheduling agent will automatically detect and record all online job agents and their attribute information; based on the nature of the task and the degree of ability matching of each job agent, the scheduling agent will select the most suitable executor and send a detailed description; the selected job agent will start to handle the assigned task, during which it may share resources or exchange intermediate results with other job agents; after completion, the job agent needs to submit the final output back to the scheduling agent, and the latter will further integrate it as needed or directly present it to the user. As more tasks arrive, the above steps will continue to repeat, forming a closed-loop and efficient workflow. Through such a structured organizational method, not only can the advantages of each component be fully utilized, but also complex and changing application scenarios can be effectively coped with, thus providing a more flexible and reliable service experience for users.

[0150] Fifth, the loop supervision and error correction mechanism. As a key component to ensure the stability and efficiency of the entire multi-agent system, this supervision and error correction mechanism works together with the closed-loop workflow of information transfer among multi-agents mentioned in the fourth part to enhance the self-adaptability and reliability of the system. The following specifically introduces how this mechanism realizes the effective monitoring and error correction of the operation process through the standardized information transfer protocol.

[0151] First of all, in order to ensure the smooth exchange of information between agents at different levels or of the same level, the present invention adopts a set of strictly defined standardized information transfer protocols. As mentioned above, one of the cores of this protocol lies in the design of the unified quantization mechanism and the state quantization module. The unified quantization mechanism aims to provide a general methodological framework, enabling various types of state information to be converted into comparable and analyzable forms; while the state quantization module is responsible for collecting and processing these quantized state information in real time to support the accuracy requirements in subsequent decision-making processes. The unified quantization mechanism and its subordinate state quantization module belong to the standardized steps of the standardized information transfer protocol, and their main work lies in maintaining the efficient cooperation among multi-agents and recording the two types of data sets mentioned above. The loop supervision and error correction mechanism is an advanced function designed and developed on this basis.

[0152] Among them, the specific implementation method of the unified quantization mechanism is:

[0153] 1) Data collection: First, collect various forms of data generated by each agent during its operation. These data may be unstructured or semi-structured, such as text, images, video streams, etc.

[0154] 2) Data preprocessing: Perform preliminary processing on the collected data, such as cleaning and normalization, to reduce noise and make the data more suitable for subsequent analysis.

[0155] 3) Feature extraction: Extract useful features from the preprocessed data. These features can represent important information such as the agent’s state, action, immediate reward, and next possible state.

[0156] 4) Structured transformation: Use a specific algorithm to combine the above features into a structured data set, where each entry contains state, action, immediate reward, and next possible state. This step is crucial to ensure the consistency and parseability of information among all agents.

[0157] 5) Output generation: Finally, the data that has undergone structured transformation is organized into a form that is easy to transmit and further optimized by the state quantization module so that it can be understood and utilized by other intelligent agents.

[0158] Among them, the specific working process of the state quantization module is:

[0159] 1) Input data reception and standardization

[0160] Diversified data processing: Supports database structured data, natural language text, various types of document files and other data

[0161] Time series data support: For time series data, use the sliding window method to segment and extract time-dependent features through LSTM.

[0162] Unified data interface: defines the data format under the framework of standardized information transmission protocol to ensure that different types of data can be accessed through a unified interface.

[0163] 2) Feature extraction and quantification

[0164] Mapping raw data into quantitative features lays the foundation for subsequent state modeling.

[0165] a) Feature extraction:

[0166] Numerical data: Normalize directly to the range of [0,1], and use Min-Max Scaling to ensure that numerical features have a uniform scale.

[0167] Categorical data: Using one-hot encoding, discrete categories are mapped into sparse vector form. For high-dimensional categorical data, target encoding is adopted to reduce the feature dimension.

[0168] Text data: By extracting semantic embeddings through BERT, 768-dimensional context-related feature vectors are generated.

[0169] Image data: Use the output of the penultimate layer intercepted by ResNet50 as the feature vector, and combine max pooling to further optimize the feature representation.

[0170] Time series data: Extract fixed-length vectors based on the hidden states of the LSTM model, and combine time step position embeddings to preserve sequential information.

[0171] b) Quantization process

[0172] Normalization: All numerical features are uniformly normalized using Z-score, ensuring that the feature distribution has a mean of 0 and a variance of 1.

[0173] Discretization: For continuous features, discretize them into N intervals (N is determined according to the feature distribution) based on K-Means.

[0174] Dimensionality reduction and feature selection: Use principal component analysis (PCA) for dimensionality reduction, retaining 99% of the cumulative variance contribution rate, and at the same time screening key features through the importance index of random forest.

[0175] Feature enhancement: Combine feature crosses to generate quadratic or cubic interaction features, such as "user age × active duration" to depict complex behavior patterns.

[0176] 3) State modeling and structured transformation

[0177] Integrate the quantized features to construct state variables, ensuring their consistency and interpretability among agents.

[0178] The process of constructing state variables includes:

[0179] a) State Vector:

[0180] The state vector represents the current internal state and environmental features of the agent.

[0181] Construction method: Concatenate all quantized features into a fixed-length vector. For example, concatenate text embeddings (768 dimensions), numerical features (10 dimensions), and image features (512 dimensions) into a unified 1290-dimensional vector.

[0182] Optimization: Further optimize the storage and calculation efficiency of high-dimensional vectors through feature dimensionality reduction algorithms (such as t-SNE).

[0183] b) Action Set:

[0184] The action set defines all optional behaviors of the agent in the current state.

[0185] Representation:

[0186] Discrete actions: Number the action set as discrete values (such as {0, 1, 2,...}) and generate action vectors of a fixed dimension through Action Embedding.

[0187] Continuous actions: Use a policy network to generate continuous action vectors that meet the task objectives.

[0188] Action constraints: Combine business logic or a rule engine to filter illegal actions to improve decision-making efficiency.

[0189] c) Immediate Reward:

[0190] The immediate reward quantifies the benefit or feedback of the current action and is usually calculated through a multi-dimensional scoring function:

[0191] Scoring function = α × task completion + β × time consumption + γ × resource utilization

[0192] Where α, β, and γ are adjustable parameters, and their weights are determined according to the task objectives.

[0193] Data source: The immediate reward can be directly collected from the environmental feedback or predicted by the proxy model (such as estimating the reward value based on a regression model).

[0194] d) Next State Prediction:

[0195] Describe the possible states that the agent may reach after executing the current action and model it using the following methods:

[0196] Matrix method: Construct a state transition matrix based on the Markov decision process and calculate the probability distribution of the next state.

[0197] Neural network method: Use a prediction model based on Transformer, take the current state and action as inputs, and output the feature vector of the next state.

[0198] Data correction: The predicted next state is post - processed through a regression model (such as Gradient Boosting) to ensure its rationality within a specific task domain.

[0199] 4) Data optimization and transmission

[0200] Optimization and compression: The state data is binary serialized, and the Snappy compression algorithm is used to reduce the transmission cost and ensure lossless restoration of the decompressed data.

[0201] Transmission method: A standardized information transfer protocol is used to achieve high - throughput and low - latency message distribution, while supporting data subscription and asynchronous updates between agents.

[0202] System monitoring: Monitoring tools are integrated into the multi - level agent interaction and collaboration architecture to monitor data traffic, latency, and error rate in real - time and provide an automatic alarm function.

[0203] Exception handling: When anomalies occur in the state data (such as missing features or illegal values), the module records detailed information through logs and triggers an automatic retry mechanism to ensure system reliability.

[0204] The upper - layer scheduling agent continuously monitors and evaluates the working status and progress of the lower - layer job agents based on the standardized information transfer protocol. To enhance the intelligence of this monitoring function, the scheduling agent undergoes reinforcement learning training using a strategy dataset designed specifically for its characteristics. Through this process, the scheduling agent can acquire the ability to identify whether a job agent deviates from the preset goal or has an operation deviation. Specifically, when the upper - layer scheduling agent detects any abnormal situation, it can quickly take a series of intervention measures to correct the problem. These measures are not limited to adjusting the resource allocation strategy, re - planning the task process, but also include, but are not limited to, optimizing the algorithm configuration, activating alternative working paths, etc. All these intervention means are aimed at ensuring that each job agent can carry out its work process smoothly according to the predetermined goals and plans, thus maintaining the stability and efficiency of the entire system. In addition, the scheduling agent also has a certain degree of adaptability. Through continuous autonomous learning using the two types of datasets mentioned above, it achieves an improvement in performance level and adaptability, and can automatically adjust the response strategy according to the current environmental conditions and available resources when encountering unforeseen situations. This flexibility enables the system to maintain a high degree of reliability and response speed even in the face of complex and changing operation scenarios, and thus provides users with more accurate and timely information processing services.

[0205] In addition, the upper-layer scheduling agent can also generate specific solutions for specific problems by leveraging its accumulated knowledge base or accessing external professional information sources (such as a standardized workflow example library). When the job agent faces challenges, the scheduling agent will propose optimization suggestions according to the specific situation, which may involve aspects such as algorithm improvement and process optimization. Subsequently, through the standard information transfer protocol, these modification opinions will be conveyed to the corresponding job agent. In the case of a serious deviation from the expected goal, the scheduling agent has the right to directly interrupt the relevant job and instruct the job agent to restart the execution process. In this way, the system can continuously self-correct and improve during operation, thereby enhancing the overall efficiency and service quality.

[0206] To facilitate readers' better understanding of the association and cooperation method between the various technologies and data sets of the present invention, an embodiment of the present invention is provided here for explanation of the system construction and work process.

[0207] Taking the informatization assessment system of the Chinese Academy of Sciences as an example:

[0208] The informatization assessment project of the Chinese Academy of Sciences is an assessment and information data collection project for the informatization level of each branch, research institution, and support unit within the Chinese Academy of Sciences system, hosted by the Computer Network Information Center of the Chinese Academy of Sciences. Every year, relevant domain experts specify the informatization assessment information and the corresponding index system for this year, which are distributed to each institution within the Chinese Academy of Sciences system. The relevant informatization responsible persons fill in the information, supplemented by data from other channels such as devices, sensors, and human experts. Finally, after integration and collection, analysis and processing are carried out to help relevant experts understand the informatization level of each institution within the Academy system this year.

[0209] Using the multi-level dynamic intelligent automated data processing and information assessment system of the present invention, first, through the database automated interaction job agent based on reinforcement learning, and under the condition of ensuring encrypted transmission and security protocol authentication, access the business database of the project through the preset connection tool. After access, through operations such as metadata parsing and data semantic parsing, establish a situation awareness of the complex structure of dozens of tables, hundreds of constraints, and foreign key relationships in the business database. And maintain the dynamic update of the perception of database content changes through regular scanning processes.

[0210] In this example, the administrators of each unit and data from various other sources will be filled in through the front-end web page of the informatization assessment of the Chinese Academy of Sciences and imported into the corresponding business database. There are four large databases involved, each containing about 30 tables, hundreds of dependencies, and thousands of fields of information. The pre-set intelligent agent for database automation interaction operations based on reinforcement learning will establish a perception of the entire database architecture through a series of parsing operations described above, laying a solid foundation for subsequent operations on the database data.

[0211] After establishing the situation awareness of the working database environment, the developers, according to the working requirements under this working condition, Scheduling agent carry out initialization configuration and debugging including prompt engineering and writing the example library of the normative workflow. The scheduling intelligent agent will establish a detailed analysis and breakdown of the overall task through the requirements in the initialization prompt engineering provided by human engineers and a small amount of high-quality example data of the normative workflow, and generate the corresponding response workflow framework. Based on the generated response workflow framework, the scheduling intelligent agent will send information such as the broken-down sub-tasks and specific IO requirements to each job intelligent agent in a standard structured format through a standardized information transfer protocol. Except for some pre-set fixed job intelligent agents, such as the intelligent agent for database automation interaction operations based on reinforcement learning, the job intelligent agents for other specific new tasks are all established through the job intelligent agent framework built in the multi-level intelligent agent interaction and collaboration framework: query applicable tools from the toolkit system according to the specific task requirements and information assigned by the scheduling intelligent agent; if there are no tools in the existing toolkit system that well adapt to this working requirement, the code generation module in the toolkit system will be activated to automatically generate new tools.

[0212] In this example, for the data obtained from the filling information of each institutional administrator, the system needs to perform three types of operations: For relatively simple multiple-choice questions, direct approval is carried out according to the preset. Questions involving data calculation and interaction rules across questionnaires, questions, and respondents will be sent to a specific job agent for specific processing. Questions involving uploading large attachments and requiring the understanding and evaluation of a large amount of complex natural language content in the attachments will be sent to another specific job agent for processing. The developer inputs this information to the scheduling agent in the form of prompt engineering. The latter will then start summarizing what the source data required for the work of these two types of specific agents is, and then send requests to the database automated interaction job agent based on reinforcement learning through a standardized information transfer protocol. After receiving the information, the database automated interaction job agent accurately locates the data that the scheduling agent hopes to obtain through the intention analysis and rationality reasoning of the request information, and also sends the return result through the standardized information transfer protocol. Next, the scheduling agent will send the returned database data, disassembled task requirements, IO requirements that conform to the standardized information transfer protocol, and other information to the two job agents and record them at the same time. These records will be displayed on the developer page, and the developer can review all the information in this record to ensure that the multi-agent architecture is in the correct construction process. If the developer finds that the system cannot fully meet the expected requirements when reviewing this record, the developer can, through the reserved interaction interface, edit the quantitative information with the same data structure as the system record format and upload it to the normative workflow example library. After that, the scheduling agent will preferentially refer to the normative examples provided by human developers to re-execute steps such as task parsing, data querying, and task dispatching. After ensuring that the above steps are correctly completed, since there are no preset tools in the system preset and the LangChain toolkit that are suitable for such complex tasks, the code generation module in the toolkit system is activated, and code generation is carried out according to the quantitative information sent by the scheduling agent, and the IO output is strictly restricted to ensure that the information transfer meets the standardized protocol.

[0213] The above-mentioned working steps already cover the technologies of multi-level agent interaction and collaboration architecture, including collaboration between different agents, standardized information transfer protocols, unified quantization mechanisms, etc. After the code generation of the toolkit system is completed, the job agent will return the specific results of the execution of this tool to the scheduling agent through standard information. The scheduling agent will reasonably infer from the prompts input by humans and, with reference to the normative examples, evaluate the results returned by the job agent to see if they fall within an acceptable distribution range. This is an example of the loop supervision and error correction mechanism: when the scheduling agent believes that the results returned by the job agent are not within the acceptable range, the former will generate an analysis based on the returned results and return it to the latter, and the latter will send the code generation module of the toolkit system for code correction after receiving the error correction information from the scheduling agent. This process will continue until the scheduling agent evaluates that the results returned by the job agent reach the acceptable range, so it is called loop supervision and error correction.

[0214] When the scheduling agent determines that the returned results of each job agent have met the expected requirements, it is called that the system has completed initialization at this time. From source data acquisition, establishment of the multi-agent collaboration architecture, creation of specific job agents, error-free inspection of data compliance to the final output, the system has completed the initialization of a complete process. After the initialization is completed, the results generated by the system can already be considered correct results. However, if you want the results generated by the system to be more accurate and closer to the level of human experts, you need to rely on the two types of data sets in the present invention for continuous and autonomous training and optimization.

[0215] In this example, operations such as the scheduling agent's understanding of human intentions, the rationality judgment of the data fields returned by the database interaction agent, the decomposition and assignment of human-specified tasks, and the rationality judgment of the results returned by the job agent are all based on the ReAct technology. Similarly, the execution process of the job agent for specific work requirements is also realized through the ReAct technology. The core point of this technology is to enable the scheduling agent to combine reasoning and execution, and at the same time, this intermediate process can be post-processed into a standard structure through streaming output and stored. In terms of the effect, each process of human administrators' review and correction of the scheduling agent and each process of the scheduling agent's review and correction of the lower-level job agent are used as very high-quality training data, which will be recorded in the dataset in the form of standardized ReAct streaming output. In terms of the source, for the database interaction agent, since it regularly scans the business database, the database operations indirectly or directly generated by the system's users and developers will be analyzed, and this analysis record will be recorded in the dataset in the form of standardized ReAct streaming output; during the communication and feedback process between the job agent and the scheduling agent, a large amount of data generated by ReAct, such as error correction feedback, will also be included in the communication data.

[0216] At the same time, by reusing the state quantization module in the unified quantization mechanism of the standardized information transfer protocol, the policy dataset is based on the communication data and the data in the ReAct dataset for standard quantization. Next, the proximal policy optimization algorithm is used to parse the quantized data, thereby continuously optimizing the implementation policy. This continuously optimized policy will be updated in real time to each agent to achieve the feedback mechanism and performance improvement. This process is equivalent to extracting the optimal solution idea from the reasoning-execution thinking process of ReAct through the PPO algorithm and feeding it back to the entity in the form of an implementation policy. Through the recording, quantization, processing, and parsing of the two types of datasets, self-optimization and self-feedback can be achieved without external additional data sources, mainly from the behavior data records of humans and agents during system operation.

[0217] Another embodiment of the present invention provides a multi-level dynamic intelligent automation data processing and information evaluation system, which includes:

[0218] A database automatic interaction module based on reinforcement learning, which is used to construct a database environment job dataset, combine reinforcement learning to conduct targeted training and fine-tuning on the large language model, and realize full-automatic data analysis, query, and information extraction of multiple tables in the database;

[0219] The multi-level agent interaction and collaboration module is used to establish a multi-level agent interaction and collaboration architecture, which includes a scheduling agent at the upper layer and a job agent at the lower layer. The scheduling agent is responsible for global decision-making and policy planning, and the job agent is responsible for the execution of specific tasks. The scheduling agent and the job agent share information and coordinate tasks to form a highly collaborative whole;

[0220] The cyclic supervision and error correction mechanism module is used to establish a cyclic supervision and error correction mechanism. The scheduling agent learns the normative workflow examples of the target working environment and real-time tracks the status of the job agent, identifies the behaviors of the job agent deviating from the predetermined path and takes corrective measures.

[0221] The division of the above-mentioned modules is only for illustrative purposes. In actual applications, the above functions can be assigned to different functional modules according to needs to complete all or part of the functions described in the foregoing method. The specific working processes of the above-mentioned modules can refer to the corresponding processes in the foregoing method embodiments and will not be elaborated here. Each module in the present invention can be implemented in the form of a software functional unit and can be stored in a computer-readable storage medium, including several instructions for causing a computer device to execute part or all of the steps of the method described in the present invention.

[0222] Based on the above content, the present invention proposes a multi-level agent architecture with a large language model as the core and innovatively applies it to the scientific metrology evaluation task. This architecture realizes the automated analysis and measurement of complex natural language content and significantly improves the reliability and efficiency of evaluation results. Specifically, the present invention solves the core problems of the prior art in the field of scientific metrology evaluation through the comprehensive application of the following technologies and methods:

[0223] 1) Intelligent application of natural language processing technology. By optimizing the task adaptability of the large language model, the present invention can achieve more efficient and accurate results in the evaluation of complex document information and the parsing of natural language content. This application breaks through the limitations of the model in terms of context processing length and complexity, enabling the large language model to play a key role in scientific metrology tasks.

[0224] 2) Efficient collaboration of multi-level agents. The present invention designs a multi-level agent architecture, decomposes tasks into multiple refined links, and independent execution by agents with different functions and completes the full-process tasks through efficient collaboration. This architecture not only improves the scalability of the system and task processing efficiency, but also ensures the coherence and reliability of task execution.

[0225] 3) Optimize system performance using reinforcement learning. By leveraging reinforcement learning techniques, the agent can learn the optimal task execution strategy through trial and error in a dynamic environment. This invention applies reinforcement learning to the task assignment and feedback optimization mechanism of the agent, significantly enhancing the system's adaptability to task scenario changes and the credibility of the overall evaluation results.

[0226] 4) Large-scale data processing and complex format parsing. For the large-scale complex natural language content involved in scientometric tasks, this invention realizes efficient data storage, query, and analysis capabilities by integrating intelligent database operation techniques. The system can automatically parse the logical structure of multi-source data, extract key information, and process it efficiently, significantly reducing the need for manual intervention.

[0227] 5) Enhance the reliability of scientometric evaluation. Through the combination of various innovative technologies, this invention introduces supervision and error correction mechanisms in each link of task execution to ensure the accuracy and consistency of evaluation results. At the same time, the agent can dynamically adjust task priorities and resource allocation in multi-task collaboration, further improving the system's stability in complex environments.

[0228] In summary, this invention provides an efficient and intelligent scientometric evaluation solution, which not only realizes the possibility of automated processing of large-scale complex natural language content but also significantly enhances the reliability of intelligent evaluation and measurement results. This technology fills the key gap in complex task processing in the current scientometric evaluation field and provides strong support for the rapid analysis of large-scale data and scientific decision-making.

[0229] The specific embodiments of the present invention disclosed above are intended to help understand the content of the present invention and implement it accordingly. Those of ordinary skill in the art can understand that various substitutions, changes, and modifications are possible without departing from the spirit and scope of the present invention. The present invention should not be limited to the content disclosed in the embodiments of this specification, and the protection scope of the present invention is defined by the scope of the claims.

Claims

1. A multi-level dynamic intelligent automated data processing and information evaluation method, characterized in that: The following steps are involved: Build a database environment operation data set, combine reinforcement learning to conduct targeted training and fine-tuning of large language models, and realize fully automatic data analysis, query and information extraction of multiple tables in the database; Establish a multi-level intelligent agent interaction and collaboration architecture, which includes the upper-level scheduling intelligent agent and the lower-level operation intelligent agent. The scheduling intelligent agent is responsible for global decision-making and strategy planning, and the operation intelligent agent is responsible for the execution of specific tasks. The scheduling intelligent agent and the operation intelligent agent share information and coordinate tasks to form a highly collaborative whole. Establish a cyclic supervision and error correction mechanism to schedule the intelligent agent to learn the normative workflow examples of the target working environment and track the status of the working intelligent agent in real time, identify the behavior of the working intelligent agent that deviates from the predetermined path and take corrective measures.

2. The method according to claim 1, characterized in that The construction of the database environment operation data set includes constructing a reasoning-execution chain data set based on the ReAct technology, and constructing a strategy data set of the agents in the multi-level agent interaction and collaboration architecture.

3. The method according to claim 2, characterized in that The construction of the reasoning-execution chain dataset based on ReAct technology includes: Historical data analysis: Collect and analyze a large number of historical database query logs to extract common database operation patterns and user behavior characteristics; Workflow data integration: Obtain the log data generated by the agent during the operation of the business link, and conduct in-depth analysis to extract dynamic trend change information; Synthetic data generation: Generate synthetic data using a generative adversarial network combined with a variational autoencoder machine learning method.

4. The method according to claim 2, characterized in that: The strategy data set for building the agents in the multi-level agent interaction and collaboration architecture includes: Adopt a unified quantization mechanism and use the state quantization module to uniformly quantify communication and state data; By using the reasoning-execution chain dataset based on ReAct technology and combining it with the needs in actual application scenarios, a policy dataset suitable for the proximal policy optimization (PPO) algorithm is established; the policy dataset contains the state of the agent in a specific environment, the actions taken, the immediate reward obtained, and the predicted next possible state nextstate; the data of the policy dataset comes from the policy data output and recorded by the agent, the communication data of multi-agent interactions, and the data of screened high-quality open source datasets.

5. The method according to claim 1, characterized in that The scheduling agent is responsible for overall planning from a global perspective, and multiple operating agents share specific implementation work according to their respective expertise, forming a "centralized decision-making + distributed execution" model; the scheduling agent and the operating agent maintain close contact through standardized information transmission protocols; whenever a new task arrives or an existing plan needs to be adjusted, the scheduling agent notifies the relevant operating agents; the scheduling agent automatically detects and records all online operating agents and their attribute information; based on the nature of the task and the degree of matching of the capabilities of each operating agent, the scheduling agent selects the most suitable operating agent, and the selected operating agent begins to process the assigned task, during which it shares resources or exchanges intermediate results with other operating agents; after completion, the operating agent submits the final output back to the scheduling agent, which integrates it as needed or directly presents it to the user.

6. The method according to claim 5, characterized in that The job agent includes a database interactive job agent; the database interactive job agent has the ability to comprehensively perceive the database architecture and the data structure of each table, the ability to accurately locate and analyze specific data, and the high execution capability of complex database operations.

7. The method according to claim 5, characterized in that The standardized information transmission protocol includes a unified quantization mechanism and a state quantization module. The unified quantization mechanism converts various types of state information into comparable and analyzable forms, and the state quantization module is responsible for collecting and processing quantified state information in real time to support the accuracy requirements in subsequent decision-making processes.

8. The method according to claim 5, characterized in that The scheduling agent establishes a detailed analysis and decomposition of the overall task through the requirements in the initialization prompt word engineering provided by human engineers and a small number of high-quality normative workflow examples, and generates a corresponding response workflow framework, and then sends the specific task requirements and information to each job agent in a standard structured format through a standardized information transmission protocol; The job agent queries the toolkit system for applicable tools based on the specific task requirements and information assigned by the scheduling agent. If there is no tool that meets the work requirements in the existing toolkit system, the code generation module in the toolkit system is activated to automatically generate new tools. After the code generation of the toolkit system is completed, the job agent returns the specific results of the execution of the new tool to the scheduling agent, which evaluates the results returned by the job agent and performs the above-mentioned cyclic supervision and error correction mechanism.

9. The method according to claim 8, characterized in that The cyclic supervision and error correction mechanism includes: The scheduling agent continuously monitors and evaluates the working status and progress of the operating agents based on a standardized information transmission protocol; The scheduling agent uses the policy data set for reinforcement learning training to learn the ability to identify whether the operating agent deviates from the preset target or has operational deviations; When the scheduling agent detects any abnormal situation, it will quickly take intervention measures to correct the problem, including adjusting resource allocation strategies, replanning task processes, optimizing algorithm configurations, and activating backup work paths to ensure that each operating agent can smoothly carry out its work process according to the predetermined goals and plans; The scheduling agent has the ability to adapt itself and automatically adjust its response strategy according to the current environmental conditions and available resources when encountering unforeseen situations; The scheduling agent uses its own accumulated knowledge base or accesses external professional information sources to generate specific solutions to specific problems. When the job agent faces challenges, the scheduling agent makes optimization suggestions based on the specific situation, and then communicates the optimization suggestions to the corresponding job agent through a standard information transmission protocol. For situations that seriously deviate from the expected goals, the scheduling agent has the right to directly interrupt the relevant operations and instruct the job agent to restart the execution process.

10. A multi-level dynamic intelligent automated data processing and information evaluation system, characterized in that: include: The database automatic interaction module based on reinforcement learning is used to build a database environment operation data set, and combines reinforcement learning to conduct targeted training and fine-tuning of large language models to achieve fully automatic data analysis, query and information extraction of multiple tables in the database; The multi-level agent interaction and collaboration module is used to establish a multi-level agent interaction and collaboration architecture, which includes the upper-level scheduling agent and the lower-level operation agent. The scheduling agent is responsible for global decision-making and strategy planning, and the operation agent is responsible for the execution of specific tasks. The scheduling agent and the operation agent share information and coordinate tasks to form a highly collaborative whole. The loop supervision and error correction mechanism module is used to establish a loop supervision and error correction mechanism, schedule the intelligent agent to learn the normative workflow examples of the target working environment and track the status of the working intelligent agent in real time, identify the behavior of the working intelligent agent that deviates from the predetermined path and take corrective measures.

Images