Multi-agent report generation method based on two-stage scheduling strategy
By employing a two-stage scheduling strategy and closed-loop adaptive optimization, the computational complexity and adaptive capability of multi-agent systems in knowledge-intensive document generation are addressed, achieving efficient and stable report generation.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-28
- Publication Date
- 2026-04-07
AI Technical Summary
Multi-agent systems suffer from high computational complexity, insufficient adaptability, high risk of error accumulation, and uncertainty in the generation of knowledge-intensive documents, making it difficult to meet the requirements of real-time performance and scalability.
A multi-agent report generation method based on a two-stage scheduling strategy is adopted. By combining global guidance and local cooperation, a closed-loop adaptive optimization system is constructed. Through multi-dimensional evaluation and constraint-driven approaches, dynamic optimization of report generation is achieved.
It improves the coordination and stability of report generation, ensures real-time evaluation and continuous optimization of report quality, solves the problem of insufficient adaptation of multi-agent systems in complex environments, and improves computational efficiency and interpretability.
Smart Images
Figure CN121809414A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of artificial intelligence and intelligent automation technology, and more specifically, to a multi-agent report generation method based on a two-stage scheduling strategy. Background Technology
[0002] Multi-agent systems are distributed computing systems in which multiple agents interact in the same environment in a cooperative or adversarial manner to maximize the completion of tasks and achieve specific goals. They are currently widely used in various application scenarios such as task scheduling, resource allocation, and collaborative decision support in complex environments. As the interaction between intelligent agents and the physical environment becomes increasingly close, the complexity of continuous multi-task control problems is also increasing. Especially in the knowledge-intensive document generation scenario, the multi-agent collaborative mechanism has many problems in the application of dynamic tasks and complex generation tasks: (1) The core optimization algorithm has computational bottlenecks, such as excessive execution cost and high computational complexity, which leads to low efficiency when processing large-scale and high-dimensional calculations, making it difficult to meet the real-time and scalability requirements of practical applications; (2) The decision-making mechanism relies on simplified static parameters and fixed formulas, and fails to incorporate dynamic environmental information such as task execution time and real-time resource load, resulting in the system lacking online self-correction capabilities and failing to achieve adaptive optimization in complex and ever-changing environments; (3) In order to realize complex functions, multiple heterogeneous models are connected in series, resulting in high risk of error accumulation and prominent process efficiency bottlenecks; and the output of its core generation module (such as large models) lacks deterministic guarantee, requiring a lot of post-processing correction, which affects the reliability of the final result and the stability of the system.
[0003] Therefore, there is an urgent need for a knowledge-intensive document generation solution based on a multi-agent collaborative system, which can dynamically optimize and adaptively adjust the working strategy parameters of the agents, realize real-time quality assessment and continuous optimization of the report generation process, and thus improve the workflow coordination and the quality of the generated results under multiple constraints. Summary of the Invention
[0004] To achieve the above objectives, this application provides a multi-agent report generation method based on a two-stage scheduling strategy, comprising the following steps: Obtain the report generation requirements and constraints; the report generation constraints include: total word count constraint, total page count constraint, and number of charts constraint; Load the multi-agent collaboration model, which is used to generate reports based on report generation requirements and report generation constraints. Initialize the multi-agent cooperative model and generate The initial population consists of several different working schemes, represented as... ; Perform t stage scheduling operations based on the initial population and generate intermediate reports. ; A complete report is generated based on the intermediate reports during the final stage of scheduling.
[0005] Before loading the multi-agent cooperative model, the construction of the multi-agent cooperative model includes the following steps: Define a multi-agent cooperative model, which consists of m agents with heterogeneous functions, as follows: The intelligent agents include: database query agent, web page retrieval agent, chart generation agent, report writing agent, and report evaluation agent. The joint policy space of a multi-agent cooperative model is determined to be composed of the policy vectors of the agents, and is expressed as: ; Determine the timing of the multi-agent cooperative model during the scheduling process. The joint strategy is represented by the working strategy vector as follows: ,in, and These represent the working strategy vectors of the database query agent, web page retrieval agent, chart generation agent, and report content writing agent at time t, respectively. Define constraints that match the report generation constraints; Determine the overall objective of the multi-agent cooperative model.
[0006] Furthermore, the parameters of the working strategy vector are represented as follows: , , , ,in, The query depth of the database query agent. The retrieval time range for the database query agent. The scope of data retrieved by the database query agent. The number of search keywords for the webpage retrieval agent. To ensure the source authority of web retrieval agents, Chart types for generating agents for charts. The number of charts for which agents are generated. The level of detail in the report regarding the intelligent agent should be considered. The citation density of the agent is used to write the report.
[0007] Furthermore, the constraints include: word count constraint for the generated report: Page number constraints for generated reports: Constraints on the number of charts in the generated report: .
[0008] Furthermore, the overall objective is to maximize the quality assessment function of the report under multiple constraints, expressed as: ,in, For the report The quality; in, The quantitative dimensions include completeness. and rationality , is represented as: ,in and , representing the relative importance weights of completeness and rationality, respectively.
[0009] Among them, integrity Quantification methods include: The definition report contains There are 10 chapters, each represented as 10 chapters. ; Predefined set of key concepts ,in Indicates the first The total number of key concepts in each chapter; Definition and Calculation Chapter Integrity score ; Integrity For each chapter The weighted average of the integrity scores is calculated as follows: ,in This refers to the relative importance of each chapter in the overall integrity assessment.
[0010] The phase scheduling includes: performing global guided learning on the multi-agent cooperative model to update the policies of all agents; performing local guided learning on the multi-agent cooperative model to select some agents to optimize their policies based on the report generation task; determining whether the report generation task meets the stopping requirements, and if it does not meet the stopping requirements, iteratively executing phase scheduling; otherwise, stopping phase scheduling.
[0011] Global guided learning includes: Let the current population be: ; The report quality score for each solution is calculated as follows: ; The individual with the best performance is selected, denoted as: ; The remaining individuals are determined by the leader ( The joint strategy space and population average () The update is performed based on the gap between the individual and the data. During the individual update process, for each individual... Represented as: ,in: and A random number between 0 and 1. This is the global guiding step size, typically set to 1 or 2. This indicates progressive element-wise multiplication; After the global bootstrapping phase ends, if a new solution is found... Better than the old solution If the new solution is found, then accept the new solution; otherwise, keep the old solution.
[0012] Localized guided learning includes: Two individuals are randomly selected from the population. and ,in ,if , then represents an individual Superior ,So Towards Proximity, or conversely, distance, indicates an individual. Superior ,So Towards near.
[0013] Furthermore, the stop request is expressed as: ,in, , The convergence threshold, This indicates the maximum number of iterations.
[0014] This invention addresses the multi-agent report generation task by innovatively proposing a two-stage dynamic scheduling mechanism and a closed-loop adaptive optimization system. This effectively solves the problems of inefficient collaboration and insufficient adaptability in traditional agent architectures when generating high-quality professional reports. The main improvements are in three aspects: 1) It adopts a scheduling mode combining global guidance and local collaboration, balancing global optimization and local exploration to achieve efficient collaboration among heterogeneous agents in a dynamic environment, improving the coordination and stability of report generation; 2) It constructs a "generation-evaluation-feedback-adjustment" closed-loop mechanism, combining multi-dimensional evaluation and constraint-driven fusion strategies. Constraints such as word count, length, and number of charts are quantified as penalty terms and integrated into the optimization objective, maximizing content quality while ensuring structural standardization; 3) It designs a modular and pluggable algorithm framework, supporting plug-and-play use of various optimization algorithms such as TLBO and PSO. Furthermore, it enhances the interpretability and controllability of the method through visual recording of parameter evolution trajectories and collaborative contribution distribution, improving adaptability and deployment efficiency across different application scenarios. Attached Figure Description
[0015] Figure 1This is a flowchart illustrating the steps of a multi-agent collaborative report generation method provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the scheduling process during the execution phase of a multi-agent cooperative model provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the individual workflow of a multi-agent collaborative model provided in an embodiment of the present invention. Detailed Implementation
[0016] This invention aims to generate high-quality professional reports through multi-agent collaborative processes, addressing issues such as unstable quality, poor logical consistency, and weak controllability in current automated report generation methods under complex structural constraints. The invention models the professional report generation process as a constrained multi-agent collaborative optimization problem, optimizing the overall quality of the generated report while satisfying constraints on report length, structure, and the number of charts. Supported constraints include total word count, total page count, and the number of charts.
[0017] The specific implementation of the present invention will now be described in detail with reference to the accompanying drawings.
[0018] The multi-agent report generation method based on a two-stage scheduling strategy provided by this invention is as follows: Figure 1 As shown, it includes the following steps: Step S100: Obtain report generation requirements and report generation constraints; among which, the report generation constraints include: total word count constraint, total page count constraint, and number of charts constraint; Step S110: Load the multi-agent cooperation model, which is used to generate a report according to the report generation requirements and report generation constraints; Before loading the multi-agent cooperative model, perform step S105: Construct the multi-agent cooperative model, including the following steps: 1) Define a multi-agent cooperative model: A multi-agent cooperative model consists of m agents with heterogeneous functions, and the set of agents is represented as: Each intelligent agent Each has an independent adjustable working strategy parameter vector To accomplish different tasks; among them, Represents intelligent agents Adjustable parameter dimensions; The agents used to generate reports include: a database query agent, a web page retrieval agent, a chart generation agent, and a report writing agent. The adjustable parameters include search depth, search time range, number of search keywords, and number of charts; meanwhile, the multi-agent collaborative model also includes a report evaluation agent to assess report quality.
[0019] 2) Determine the joint policy space for the multi-agent cooperative model: The joint policy space of a multi-agent cooperative model consists of the policy vectors of the agents, and is represented as: This matrix defines a complete multi-agent cooperation scheme, in which all agents jointly complete the execution configuration for report generation under given task conditions.
[0020] 3) After determining the joint policy space, determine the time step of the multi-agent cooperative model during the scheduling process. (i.e., the first) A joint strategy (in the next iteration); Generally, in a multi-agent cooperative model involving m agents, the joint policy at time t is determined by vectors. In the application scenario of generating reports in this invention, the joint strategy is represented as follows: , in, and These represent the working strategy vectors of the database query agent, web page retrieval agent, chart generation agent, and report content writing agent at time t, respectively. The parameters of the working strategy vector are represented as follows: , , , , in, The query depth of the database query agent. The retrieval time range for the database query agent. The scope of data retrieved by the database query agent. The number of search keywords for the webpage retrieval agent. To ensure the source authority of web retrieval agents, Chart types for generating agents for charts. The number of charts for which agents are generated. The level of detail in the report regarding the intelligent agent should be considered. The citation density of the agent is used to write the report.
[0021] Detailed descriptions of each decision variable are shown in Table 1: Table 1. Parameter Symbols and Decision Variable Descriptions
[0022] 4) Define constraints to meet the requirements for report length, table size, and other data, matching the report generation constraints obtained in step S100. These constraints include: Word count constraints for generated reports: , Page constraints for generated reports: , Constraints on the number of charts in the generated report: ; for and The numerical statistics are completed by the report evaluation agent.
[0023] 5) Determine the task objective of the multi-agent cooperative model, i.e., the overall objective; In this step, use The report text generated from the joint policy space has the overall objective of maximizing the report's quality evaluation function under multiple constraints, expressed as: ,in, For the report The quality.
[0024] Report quality The quantitative dimensions include completeness. and rationality , is represented as: ,in and , representing the relative importance weights of completeness and rationality, respectively.
[0025] Specifically, quantifying the completeness The methods include: The definition report contains There are 10 chapters, each represented as 10 chapters. ; Predefined set of key concepts ,in Indicates the first The total number of key concepts in each chapter; Define Chapter Integrity score for: ,in, Indicating in the joint strategy space Below are key concepts when generating reports. Coverage; calculation The generated report can be segmented into words, and then it can be statistically analyzed whether all key concept implementations are covered.
[0026] The overall report completeness is the weighted average of the completeness of each chapter. The calculation method is as follows: ,in This refers to the relative importance of each chapter in the overall integrity assessment.
[0027] To assess the reasonableness of the text report, this invention uses the Vectara-HHEM hallucination detection model to evaluate the authenticity of the output professional report in order to quantify reasonableness.
[0028] Based on the above definitions, the overall objective of the multi-agent cooperative model in generating reports is formalized as a constrained multi-agent dynamic scheduling task, and the corresponding objective function is expressed as: , , Where t represents the iteration period for report generation.
[0029] In each iteration, constraints are used to ensure that the report quality does not monotonically decrease, forming a dynamic closed loop of quality improvement. After several consecutive iterations, the following condition is met: If the report quality is considered stable, further adjustments will cease.
[0030] Because the report generation process in this invention is subject to word count constraints, page count constraints, and icon count constraints, and these constraints can be appropriately violated within a certain range but should be satisfied as much as possible: the impact of such violations is directly reflected in the objective function value, and a penalty term is introduced into the objective function, expressed as: , in This is a penalty coefficient used to balance quality and the degree of constraint violation.
[0031] Through the above steps, a multi-agent collaborative model is constructed.
[0032] In step S110, the multi-agent collaborative model constructed through the above steps is loaded, along with the relevant pre-plan templates and key concept libraries. Then, step S120 is initiated to initialize the multi-agent collaborative model and begin the task of generating a report.
[0033] Step S120: Initialize the multi-agent cooperative model and generate... The initial population consists of several different working schemes, represented as... .
[0034] Each individual in the population is represented by a policy vector, where an individual represents a report generation scheme. An individual is represented as: .
[0035] In application scenarios ranging from report generation to report evaluation, an individual's workflow is as follows: Figure 3 As shown, each individual It corresponds to a complete workflow configuration and can independently generate a complete professional report.
[0036] When initializing the population, for arrive Randomly generate m individuals, represented as: Each parameter component is generated uniformly and randomly within its domain or based on experience.
[0037] After initializing the population, the initial fitness value for each individual needs to be calculated. This is done using the symbol... Let represent the fitness value of individual k at the t-th iteration. The initial fitness value of the k-th individual is calculated as follows: .
[0038] Step S130: Perform t stage scheduling operations based on the initial population and generate an intermediate report. ; Specifically, the process of scheduling the execution phase based on the multi-agent cooperative model and generating reports is as follows: Figure 2 As shown, the process includes the following: 1) Perform global guided learning on the multi-agent cooperative model to achieve policy updates for all agents; The global guided learning of the multi-agent cooperative model refers to: based on the previous iteration, identifying the optimal working scheme, and using the strategy configuration of the working scheme to guide other working schemes to converge towards the high-quality region.
[0039] This stage automatically selects the best-performing joint strategy matrix as a reference sample based on the report quality evaluation results of each scheme in the current iteration. Through parameter difference propagation and guided update mechanisms, it guides the entire population to converge towards the high-quality reporting region. This process ensures the consistency and robustness of the global search direction.
[0040] The specific implementation process is as follows: Let the current population be: ; The report quality score for each solution is as follows: ; The individual with the best performance is selected, denoted as: ; The remaining individuals are determined by the leader ( The joint strategy space and population average () The update is performed based on the gap between the individual and the data. During the individual update process, for each individual... Represented as: ,in: and A random number between 0 and 1. This is the global guiding step size, typically set to 1 or 2. This indicates progressive element-wise multiplication; After the global bootstrapping phase ends, if a new solution is found... Better than the old solution If the new solution is found, then accept the new solution; otherwise, keep the old solution.
[0041] This process ensures that the overall search direction always moves towards the high-quality report area, meaning the global strategy evolution satisfies: .
[0042] After global guidance is complete, the local collaboration phase begins. The goal of this phase is to promote mutual learning and strategy diversification among collaborative schemes, enabling different agents to explore possible directions for improvement in their local spaces, thereby enhancing adaptability and robustness.
[0043] 2) Perform local guided learning on the multi-agent cooperative model, and select some agents to optimize the strategy based on the report-generated task.
[0044] During this process, two individuals are randomly selected from the population. and ,in ,if , then represents an individual Superior ,So Towards Proximity, or conversely, distance, indicates an individual. Superior ,So Towards near.
[0045] The computation method for performing local parameter interaction and complementary updates is as follows: ,in, Indicates the local cooperation intensity coefficient; This represents the local disturbance coefficient.
[0046] Similarly, after the local learning phase ends, if a new solution is found... Superior If the new solution is found, then accept the new solution; otherwise, keep the old solution.
[0047] In this phase, several agent schemes are randomly selected for parameter exchange, task collaboration, or knowledge transfer, enabling agents to share local experiences and policy fragments. This "lateral communication" mechanism maintains population diversity, avoids premature stagnation in a single-structure search, and thus achieves higher accuracy in reporting detailed optimization and local correction.
[0048] 3) Determine whether the report generation task meets the stopping requirements. If it does not meet the stopping requirements, iterate through the execution phase scheduling; otherwise, stop the phase scheduling.
[0049] After each round of compilation iteration, a quality assessment is initiated to generate a comprehensive quality score by quantitatively scoring the currently generated report from multiple dimensions, including indicators such as content completeness, logical rationality, and compliance with length and chart constraints. Overall quality score: In each iteration, ensure that the reporting quality function Q(x,t) monotonically increases at iteration time t until the stopping requirement is met.
[0050] The stopping requirement is determined by quality convergence, maximum number of iterations, and whether there is an improvement in global quality during the iteration process; The stop request is expressed as: ,in, , The convergence threshold, This indicates the maximum number of iterations.
[0051] If any of the stopping conditions in the stopping requirements are met, the two-phase iteration ends; otherwise, it returns to the global guidance phase and starts the next scheduling cycle.
[0052] During the phase scheduling process, the agent's strategy is no longer statically fixed, but continuously fine-tuned based on real-time feedback, gradually bringing the report content closer to optimal performance. Since each update only performs asynchronous adjustments on some parameters or some agents, response speed and computational efficiency are significantly improved, avoiding the latency accumulation caused by global synchronous updates. Simultaneously, the local update mechanism allows the model to optimize in real time without interrupting report generation, achieving parallel collaboration among "generation—evaluation—optimization".
[0053] Step S140: Generate a complete report based on the intermediate reports during the end-stage scheduling.
[0054] This invention provides specific embodiments applicable to emergency response to sudden flood and drought events, aiming to quickly generate dynamic emergency response plan reports that are structurally rigorous, information-rich, and compliant with standards.
[0055] First, define the task of generating the emergency response plan, specify the report topic as "Flood and Drought Prevention Emergency Response Plan", and set key constraints, including the total word count range (e.g., 8,000-12,000 words), page limit (no more than 20 pages), and chart requirement (at least 5 situation diagrams and flowcharts); at the same time, load the plan templates and key concept library related to this event type.
[0056] Secondly, the multi-agent collaborative model is loaded, initialized, and a population is generated. The multi-agent collaborative model includes: a PostgreSQL database query agent, a Neo4j knowledge graph query agent, a CRAG (Retrieval Augmentation Generation) agent, a report writing agent, a chart generation agent, and a report evaluation agent. Each agent initializes its policy parameters (such as query depth, retrieval scope, writing style, chart type, etc.); an initial population containing NP independent working schemes (i.e., individuals) is randomly generated, with each individual representing a complete combination of agent collaborative strategies.
[0057] Step 3: Perform a two-phase scheduling optimization iteration; Each round includes the following key phases: Global guidance phase: Select the individual with the highest report quality score in the current population as the "guide" and guide other individuals to update their own parameters based on their policy vector, so as to make the population as a whole converge towards the high-quality plan area.
[0058] Local collaboration phase: Randomly select individuals from the population for pairing. Through parameter exchange and strategy complementarity, enhance local exploration capabilities, avoid getting trapped in local optima, and improve the accuracy of the report in terms of detailed description and structural connection.
[0059] Step 4: Evaluation Feedback and Dynamic Parameter Adjustment: After each iteration, the evaluation agent performs multi-dimensional quantitative scoring on the generated contingency plan report, including content completeness (coverage of key handling steps and resource allocation plans), rationality (process connection and causal relationship), and format compliance (compliance of word count and number of charts). The scoring results are fed back to the scheduler in real time, driving the dynamic adjustment of strategy parameters in the next round, forming a closed loop of "generation-evaluation-optimization".
[0060] Step 5: Meet the stopping criteria and output a candidate report. The optimization process terminates when the report quality score improvement rate is found to be lower than the set threshold for multiple consecutive rounds, or when the maximum number of iterations is reached.
[0061] At this point, the multi-agent collaborative model outputs an emergency response plan report generated using the current optimal parameters and strategies. After review and revision, the final version of the "Flood and Drought Emergency Response Plan" is generated for actual emergency command and decision support.
[0062] This invention employs a two-stage dynamic scheduling mechanism combining a global guidance phase and a local collaboration phase to generate reports, effectively balancing the global optimization and local exploration capabilities of a multi-agent architecture. Compared to the rigidity of existing centralized control architectures and the tendency of distributed strategies to fall into local optima, the two-stage dynamic scheduling mechanism achieves efficient collaboration and adaptive scheduling of heterogeneous agents in dynamic task environments, significantly improving the coordination, stability, and final output quality of the report generation process. In terms of quality control, this invention adopts a closed-loop adaptive optimization mechanism based on quantitative feedback—"generation-evaluation-feedback-adjustment"—enabling online self-correction and continuous performance improvement. This effectively overcomes the shortcomings of existing methods in terms of insufficient adaptability in dynamic environments, achieving monotonic improvement in report generation quality and enhanced robustness during the iterative process.
[0063] In terms of architecture design, this invention adopts a modular and pluggable scheduling algorithm framework, enabling the multi-agent collaborative model to possess high versatility and scalability. This includes two core components: a "policy parameter space" and a "fitness evaluation function," which support plug-and-play use of various optimization algorithms, including the Teaching and Learning Optimization (TLBO) algorithm and the Particle Swarm Optimization (PSO) algorithm. During computation, algorithm combinations can be flexibly selected or switched based on task complexity and real-time resource status, avoiding the problems of high coupling and poor scalability in existing technologies, and significantly improving adaptability and deployment efficiency in different application scenarios.
[0064] In the collaborative process of multiple agents, a multi-dimensional evaluation system and constraint-driven fusion mechanism are adopted, enabling this invention to achieve an optimal balance between report generation quality and multiple structural constraints. Constraints such as word count, page count, and number of charts are transformed into quantifiable penalty terms and incorporated into the objective function for weighted optimization. By dynamically adjusting the penalty coefficients and evaluation weights, the invention automatically balances "constraint satisfaction" and "quality improvement" in each iteration, effectively solving the problem of existing methods struggling to balance quality and constraints in complex document generation tasks. This maximizes content integrity and logical consistency while ensuring the standardization of the report structure. Furthermore, an interpretable and controllable agent parameter update and collaboration path recording mechanism is employed, making this invention highly transparent and user-friendly for human intervention.
[0065] During the phased scheduling process, the evolution trajectory of policy parameters, quality change curves, and agent collaborative contribution distribution are output after each iteration, enabling operations and maintenance personnel to clearly understand the optimization process and make targeted adjustments. Compared with existing technologies based on "black box" methods such as deep reinforcement learning, this invention significantly improves interpretability and controllability while maintaining optimization efficiency, providing reliable support for debugging and optimizing complex report generation tasks.
[0066] The above-disclosed embodiments are merely a few specific examples of the present invention. However, the present invention is not limited thereto, and any variations that can be conceived by those skilled in the art should fall within the protection scope of the present invention.
Claims
1. A multi-agent report generation method based on a two-stage scheduling strategy, characterized in that, Includes the following steps: Obtain the report generation requirements and constraints; wherein, the report generation constraints include: total word count constraint, total page count constraint, and number of charts constraint; Load a multi-agent collaboration model, which is used to generate a report based on the report generation requirements and report generation constraints; The multi-agent cooperative model is initialized and generated. The initial population consists of several different working schemes, represented as... ; Based on the initial population, perform t stage scheduling operations and generate intermediate reports. The stage scheduling includes: performing global guided learning on the multi-agent cooperative model to update the policies of all agents; performing local guided learning on the multi-agent cooperative model to select some agents to optimize their policies based on the report generation task; determining whether the report generation task meets the stopping requirements; if it does not meet the stopping requirements, iteratively executing stage scheduling; otherwise, stopping stage scheduling. A complete report is generated based on the intermediate reports during the final stage of scheduling.
2. The multi-agent report generation method according to claim 1, characterized in that, Before loading the multi-agent cooperative model, the multi-agent cooperative model is constructed, including the following steps: Define a multi-agent cooperative model, which consists of m agents with heterogeneous functions, and represent it as follows: The intelligent agents include: a database query intelligent agent, a web page retrieval intelligent agent, a chart generation intelligent agent, a report writing intelligent agent, and a report evaluation intelligent agent; The joint policy space of a multi-agent cooperative model is determined to be composed of the policy vectors of the agents, and is expressed as: ; Determine the timing of the multi-agent cooperative model during the scheduling process. The joint strategy is represented by the working strategy vector as follows: ,in, and These represent the working strategy vectors of the database query agent, web page retrieval agent, chart generation agent, and report content writing agent at time t, respectively. Define constraints that match the report generation constraints; Determine the overall goal of the multi-agent cooperative model.
3. The multi-agent report generation method according to claim 2, characterized in that, The parameters of the working strategy vector are represented as follows: , , , ,in, The query depth of the database query agent. The retrieval time range for the database query agent. The scope of data retrieved by the database query agent. The number of search keywords for the webpage retrieval agent. To ensure the source authority of web retrieval agents, Chart types for generating agents for charts. The number of charts for which agents are generated. The level of detail in the report regarding the intelligent agent should be considered. The citation density of the agent is used to write the report.
4. The multi-agent report generation method according to claim 2, characterized in that, The constraints include: word count constraint for the generated report: Page number constraints for generated reports: Constraints on the number of charts in the generated report: .
5. The multi-agent report generation method according to claim 4, characterized in that, The overall objective is to maximize the quality assessment function of the report under multiple constraints, expressed as: ,in, For the report The quality; in, The quantitative dimensions include completeness. and rationality , is represented as: ,in and , representing the relative importance weights of completeness and rationality, respectively.
6. The multi-agent report generation method according to claim 5, characterized in that, The integrity Quantification methods include: The definition report contains There are 10 chapters, each represented as 10 chapters. ; Predefined set of key concepts ,in Indicates the first The total number of key concepts in each chapter; Definition and Calculation Chapter Integrity score ; Integrity For each chapter The weighted average of the integrity scores is calculated as follows: ,in This refers to the relative importance of each chapter in the overall integrity assessment.
7. The multi-agent report generation method according to claim 6, characterized in that, The overall objective is formalized as a constrained multi-agent dynamic scheduling task objective function, and a penalty term is introduced into the objective function, expressed as: ,in, Here, t represents the penalty coefficient, and t represents the iteration period for report generation.
8. The multi-agent report generation method according to claim 1, characterized in that, The globally guided learning includes: Let the current population be: ; The report quality score for each solution is calculated as follows: ; The individual with the best performance is selected, denoted as: ; The remaining individuals are determined by the leader ( The joint strategy space and population average () The update is performed based on the gap between the individuals. During the individual update process, for each individual... Represented as: ,in: and A random number between 0 and 1. This is the global guidance step size, typically set to 1 or 2. This indicates progressive element-wise multiplication; After the global bootstrapping phase ends, if a new solution is found... Better than the old solution If the new solution is found, then accept the new solution; otherwise, keep the old solution.
9. The multi-agent report generation method according to claim 1, characterized in that, The local guided learning includes: Two individuals are randomly selected from the population. and ,in ,if , then represents an individual Superior ,So Towards Proximity, or conversely, distance, indicates an individual. Superior ,So Towards near.
10. The multi-agent report generation method according to claim 1, characterized in that, The stop requirement is expressed as follows: ,in, , The convergence threshold, This indicates the maximum number of iterations.