Simulation experiment planning decision-making agent and method
By planning and making decision-making intelligent agents through simulation experiments, the problem of insufficient intelligence in existing simulation experiment systems is solved. It realizes proactive cognition, agile decision-making and closed-loop optimization, improves the efficiency and credibility of simulation experiments, and reduces resource waste and duplicate experiments.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHINESE PEOPLES LIBERATION ARMY UNIT 92728
- Filing Date
- 2026-03-23
- Publication Date
- 2026-06-23
AI Technical Summary
Existing simulation experiment systems lack sufficient intelligence in experiment design, operation, and evolution. They suffer from problems such as passive response, low decision-making efficiency, lack of experimental evolution capabilities, and inefficient conversion of experimental data, leading to repetitive experiments and resource waste. They are unable to adapt to changes in environmental scenarios in real time, have rigid knowledge bases, and are difficult to support continuous optimization.
A simulation experiment was used to plan a decision-making intelligent agent, including a natural language interaction module, a knowledge base management module, a decision reasoning module, and a closed-loop optimization module, to achieve proactive information recognition, dynamic optimization, and closed-loop iteration. The natural language interaction module parses user commands through a large language model, the knowledge base management module stores and manages historical data, the decision reasoning module generates customized solutions, and the closed-loop optimization module processes results and updates knowledge.
It enables proactive cognition and agile decision-making in simulation experiments, reduces the workload of manual design, improves the credibility and guiding value of experimental results, opens up intelligent processing throughout the entire process, reduces labor costs, and realizes dynamic data-driven evolution of the experimental system.
Smart Images
Figure CN122263946A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of simulation and artificial intelligence interdisciplinary technology, specifically to a simulation experiment planning and decision-making intelligent agent and its method, and more particularly to a simulation experiment planning and decision-making intelligent agent based on a large language model, which is used to realize the full-process intelligent generation and dynamic closed-loop optimization of the experiment planning scheme, and to enhance the ability of proactive cognitive decision-making and agile evolution. It is applicable to simulation full-process experiment scenarios such as swarm and adversarial simulations. Background Technology
[0002] Simulation experiments are a crucial tool for supporting strategic analysis, equipment system demonstration, and training and development. While currently widely used simulation systems are continuously improving in terms of simulation realism and computational power, their intelligence levels in experiment design, operation, and evolution still have significant shortcomings, mainly manifested in the following three aspects: First, current simulation experimental systems generally suffer from problems such as passive response, low decision-making efficiency, lack of experimental evolution capabilities, and inefficient experimental data conversion. Experimental design relies on manual input and lacks the ability to actively acquire missing information. It is prone to incomplete experimental task specifications due to vague requirements, such as missing indicator systems and undefined factor level values, leading to repeated experiments and wasted resources. Second, the selection of equipment and algorithms and the formulation of experimental procedures require manual comparison with historical data, which is time-consuming and highly subjective, and cannot adapt to changes in environmental scenarios in real time. Third, experimental conclusions are only passively recorded and do not form a closed loop of "execution-feedback-evolution", resulting in a rigid knowledge base that is difficult to support continuous optimization in practical training. Summary of the Invention
[0003] To address the shortcomings of existing technologies, the purpose of this invention is to provide a simulation experiment planning and decision-making intelligent agent and its method.
[0004] A simulation experiment planning and decision-making intelligent agent provided by the present invention includes: The natural language interaction module is used to receive users' natural language commands, parse the semantics, match similar cases from the historical project library based on the parsing results and generate a draft of the experimental task book, and actively ask for key details during the generation process; The knowledge base management module is used to store and manage historical project libraries, device status libraries, model performance libraries, and environmental scenario libraries. The decision reasoning module is used to query the equipment status library, model performance library and environmental scenario library according to the requirements of the experimental task book draft, output a customized experimental equipment and algorithm selection recommendation scheme, and match similar experimental processes from the historical process library; The closed-loop optimization module is used to perform semantic understanding and format transformation on the unstructured results generated by the experiment, and update the knowledge base in the knowledge base management module through an execution-feedback-evolution closed-loop mechanism.
[0005] Preferably, the natural language interaction module performs semantic parsing and intent recognition through a large language model.
[0006] Preferably, in the knowledge base management module: The historical project library is used to store historical experimental schemes, including experimental scenarios, equipment systems of both parties, task processes, indicator systems, impact factors, factor level values, and number of simulations. The device status database is used to store the real-time availability status and load information of the experimental equipment; The model performance library is used to store performance data of different algorithm models in various environmental scenarios; The environmental scenario library is used to store typical environmental scenario data, including meteorological, electromagnetic, topographic, and situational parameters.
[0007] Preferably, the decision reasoning module uses a similarity matching algorithm to retrieve cases similar to the current experimental task from the historical process library and automatically reuses their task process framework and parameters.
[0008] Preferably, the closed-loop optimization module uses a large language model to perform semantic parsing on the experimental results and automatically converts them into standardized experimental conclusions according to a preset template.
[0009] Preferably, the closed-loop optimization module archives the formatted experimental conclusions to the corresponding knowledge base in the knowledge base management module for updating the model recommendation rules and strategy weights.
[0010] A simulation experiment planning and decision-making method for intelligent agents provided by the present invention includes the following steps: Step S1: Receive the user's natural language instructions, generate a draft of the experimental task book, and proactively ask follow-up questions to supplement key details; Step S2: Based on the confirmed experimental task sheet, query the knowledge base and output a customized experimental equipment and algorithm selection recommendation scheme; Step S3: Based on the experimental task book, match the historical process library to formulate a specific experimental process; Step S4: Perform the experiment, collect the experimental results, and use the language big model to automatically convert the unstructured experimental results into formatted experimental conclusions; Step S5: Based on the experimental conclusions, execute the "execution-feedback-evolution" closed loop to continuously optimize the knowledge base.
[0011] Preferably, in step S1, the experimental subject, evaluation target and background scenario are identified by parsing natural language instructions, and similar cases are matched from the historical project library as a reference for generating a draft task book.
[0012] Preferably, in step S4, the conversion process includes: inputting the original logs from multiple experimental simulations into the language model, and outputting structured experimental conclusions according to the template format requirements.
[0013] Preferably, in step S5, the optimization knowledge base includes: updating the algorithm performance records in the model performance base, the scene parameters in the environment scene base, and the experimental schemes in the historical project base based on the new data and experimental conclusions of this experiment.
[0014] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention, through the natural language interaction function of an intelligent agent, changes the traditional passive waiting mode of simulation experiment systems, actively identifies and inquires about key missing information, and guides users to improve their requirements. This effectively avoids deviations in experimental direction, parameter omissions, and subsequent repeated experiments caused by vague or incomplete requirements.
[0015] 2. This invention comprehensively considers the real-time status of equipment, the historical performance of algorithms, and the suitability of the environmental scenario to quickly generate customized equipment and algorithm recommendation solutions. Simultaneously, by intelligently matching similar historical cases, it can automatically reuse their validated experimental process frameworks and core parameters. This significantly reduces the workload of manual review, comparison, and design by experimental personnel in the solution design, resource allocation, and process orchestration stages.
[0016] 3. The results of each simulation experiment in this invention are automatically processed, analyzed, and used as feedback knowledge, which are then structurally stored in the system's knowledge base. This enables the model to be continuously updated and iterated, achieving a qualitative change in the experimental system from static experience-dependent to dynamic data-driven evolution.
[0017] 4. This invention streamlines the entire process of "task creation → equipment selection → process formulation → data transformation → knowledge evolution", reducing labor costs and significantly improving the credibility and practical guidance value of experimental results. Attached Figure Description
[0018] Other features, objects, and advantages of the present invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1 This is a diagram of the intelligent agent system architecture in an embodiment of the present invention.
[0019] Figure 2 This is a schematic diagram of closed-loop optimization of the experimental process in an embodiment of the present invention. Detailed Implementation
[0020] The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that those skilled in the art can make several changes and improvements without departing from the concept of the present invention. These all fall within the protection scope of the present invention.
[0021] This implementation takes "XXX defense mission" as a typical application scenario and focuses on the core task of "evaluating the contribution rate of newly developed equipment to the entire system". It elaborates on the operation mechanism and technical implementation path of the intelligent agent and its method in the whole process.
[0022] This invention relates to an intelligent agent system comprised of a natural language interaction module, a knowledge base management module, a decision-making and reasoning module, and a closed-loop optimization module. Its implementation process begins with the user issuing experimental instructions via natural language, which are then processed sequentially by each module, ultimately completing the experiment and achieving knowledge evolution. The entire process forms a dynamic "cognition-decision-execution-optimization" closed loop.
[0023] 1. Natural Language Interaction Module This module serves as the system entry point, responsible for intelligent interaction with the user. When the user inputs a natural language command such as "Please evaluate the systemic contribution rate of the newly developed equipment to the XXX defense mission," the module invokes the integrated language model for deep semantic analysis and intent recognition. At this time, the module calls the language big model for semantic vectorization processing. Assume the feature vector representation of the user instruction is... The template feature vectors in the system's predefined task template library are: The module determines user intent by calculating cosine similarity:
[0024] In response to the instruction to "assess the system contribution rate of newly developed equipment to the XXX defense mission," the system analyzed and identified key entities. and mission objectives ,in and These represent drones, carrier-based aircraft, and contribution rate, respectively. Let's consider the prior probability distribution of the "contribution rate calculation formula." A higher entropy value indicates greater information uncertainty, prompting the module to generate a follow-up questioning strategy:
[0025] in, Let represent the optimal follow-up strategy, q represent a specific follow-up question in the set of candidate follow-up questions, and Q represent the set of all candidate follow-up questions. Indicates the completeness of instruction information. Natural language commands input by the user; Through interaction, the system transforms ambiguous instructions into a structured set of task parameters:
[0026] Here, Scenario represents a scene. These represent the strength of the Red Team, the strength of the Blue Team, and the evaluation indicators, respectively.
[0027] 2. Knowledge Base Management Module This module provides structured knowledge support for other modules. It manages four sub-repositories: a historical project repository, an equipment status repository, a model performance repository, and an environmental scenario repository. It supports structured storage of experimental task templates, including experimental scenarios, red-blue team equipment systems, task plans, process logic, evaluation metrics, influencing factors and their levels, and simulation counts. It collects and updates the availability and load status of experimental equipment in real time; records algorithm performance data such as accuracy and time consumption under different environmental scenarios; integrates environmental parameters for typical environmental scenarios, such as meteorology, electromagnetic fields, terrain, and environmental conditions, supporting simulation initialization and parameter reuse; and achieves knowledge association, retrieval, and version management through a unified metadata model, providing traceable and reusable knowledge support for simulation and evaluation.
[0028] Using graphs to represent knowledge and define knowledge ,in These are entity nodes, such as devices, algorithms, and scenarios. For relation edges.
[0029] To support the reproduction and traceability of experiments, the system establishes a unified metadata model. For environmental scenarios, a multi-dimensional feature vector description is used:
[0030] in Electromagnetic environment complexity It can be quantified by the signal-to-noise ratio (SNR) spectral density function:
[0031] In the formula, For noise power spectral density, To interfere with bandwidth, This represents the signal power. This formula provides accurate electromagnetic environment parameters for simulation initialization.
[0032] 3. Decision Reasoning Module This module is responsible for generating specific experimental configuration schemes. It enables intelligent decision-making in experimental resource allocation and process planning; based on the task requirements, it calls upon the equipment status library, model performance library, and environmental scenario library in real time to comprehensively evaluate equipment availability, model applicability, and scenario matching, generating customized equipment selection, experimental process, and algorithm recommendation schemes; through similarity matching algorithms, it retrieves typical cases highly similar to the current task from the historical experimental process library, automatically filling in the task process framework, key influencing factors, and their level values, realizing intelligent reuse and rapid construction of process parameters, and improving experimental design efficiency and scheme reliability.
[0033] Define the decision objective function The aim is to maximize experimental efficiency. And minimize costs :
[0034] in, For decision variables (device ID, algorithm ID, etc.) , These are the weighting coefficients. Due to resource constraints, For logical constraints.
[0035] (1) Equipment selection: Query the equipment status database and calculate the status of each node. Overall availability :
[0036] in For load threshold, For availability, the system prioritizes... The most valuable computing resource.
[0037] (2) Algorithm recommendation: Based on the model performance library, the algorithm is calculated. Adaptability in defensive scenarios :
[0038] in For accuracy, Because it takes time, It is the reciprocal of the variance. , , These are all weighting coefficients, representing accuracy, speed, and stability, respectively.
[0039] (3) Process reuse: Utilizing a similarity matching algorithm, similar cases are retrieved from the historical process database. The current task is calculated. With historical mission similarity :
[0040] in, Representing the The weights of each feature, These represent the current task and the historical task, respectively. eigenvectors, when When the decision threshold is reached, the historical process framework and parameters are automatically reused.
[0041] 4. Implementation of the closed-loop optimization module This module ensures the system can learn and evolve from each experiment. It achieves fully automated iteration of the "execution-feedback-evolution" process. Based on preset templates, it uses a large language model to perform semantic parsing and understanding of the unstructured data output from the experiment, automatically transforming it into standardized experimental conclusions. In the execution phase, simulation runs are conducted according to the experimental plan generated by the decision reasoning module. In the feedback phase, experimental data results are collected in real time. In the evolution phase, newly acquired experimental data results and generated experimental conclusions are dynamically archived into the knowledge base, and model weights and strategy recommendation rules are updated, forming a traceable and adaptive intelligent evolution mechanism that drives the continuous optimization and capability leap of the overall decision system.
[0042] A Bayesian update mechanism is employed. Assume the model parameters... The prior distribution is In obtaining new experimental conclusions Then, its posterior distribution Updated to:
[0043] Combining the modules described above, taking the task of evaluating equipment contribution rate as an example, the complete process is as follows: User instructions are parsed and refined into a clear task statement by the natural language interaction module → The task statement triggers the decision reasoning module, which queries the various sub-databases of the knowledge base management module to generate resource configuration and experimental procedure plans → After user confirmation, the system executes the simulation → The raw results generated by the simulation are semantically parsed and formatted by the closed-loop optimization module → This module finally updates the knowledge base management module with the new knowledge from this experiment, completing the closed loop. The entire process demonstrates the synergy of the modules and the system's self-evolutionary capability.
[0044] The simulation experiment planning and decision-making method for intelligent agents in this invention includes five stages: generation of experimental task book, selection of equipment and algorithms, formulation of experimental process, result processing and closed-loop optimization, which fully presents the intelligent system of "proactive cognition - agile decision-making - closed-loop evolution".
[0045] 1. Experimental Project Task Description Creation Phase Users submit initial commands through a natural language interaction interface, such as: "Please assess the systemic contribution rate of the newly developed equipment to the XXX defense mission." Upon receiving the command, the agent's natural language interaction module invokes its built-in language model for deep semantic parsing and intent recognition. The parsing process identifies key entities, core objectives, and contextual background within the command.
[0046] This invention is defined using a performance increment model based on the task cycle:
[0047] in, To improve the overall system efficiency of "existing equipment + newly developed equipment", To improve the overall system efficiency of the original "existing equipment".
[0048] System effectiveness The ADC (Availability-Dependability-Capability) model is used for calculation:
[0049]
[0050] In the formula, For the first Key capability indicators (such as target detection probability) Task completion probability Probability of one's own survival ), Weight them accordingly. The intelligent agent proactively asks the user various questions. The values can be selected or the default values in the history database can be directly referenced to form an accurate task list.
[0051] Subsequently, the agent searches for similar cases in the historical project database of the knowledge base management module based on the parsed semantic elements. For example, it can match the historical archived "XX Existing Equipment Replacement Evaluation Experiment Plan". The agent will extract the structured task template information from this similar case, including the experimental scenario description, the equipment system composition of both the red and blue teams, typical task process stages, the preset indicator system (such as the main indicator "system contribution rate" and its calculation formula, the sub-indicator "task success rate", etc.), key influencing factors (such as the number of devices, search scope, etc.) and their level values, and the simulation number settings.
[0052] Based on the initial instructions and retrieved reference cases, the agent automatically generates a preliminary draft of the experimental task specification. Simultaneously, through semantic analysis, the agent proactively identifies any ambiguities or missing key information in the current draft. For example, regarding the calculation dimension of "system contribution rate," the agent proactively asks the user: "In the indicator system of this experiment, does the calculation formula for 'system contribution rate' need further refinement? For example, should it explicitly include specific dimensions such as 'reduction in operator casualty rate' and 'increased task success rate'?" After the user responds to the follow-up question, the agent automatically updates and improves the draft task specification based on the feedback information, forming the final version, thus completing the initialization of the experimental design and effectively avoiding the risk of information loss due to ambiguous requirements.
[0053] 2. Experimental Equipment and Algorithm Selection Stage After the experimental task description is confirmed, the agent enters the resource allocation decision-making phase. The decision reasoning module will comprehensively query multiple sub-bases in the knowledge base management module based on the experimental scenario, equipment system, and evaluation objectives defined in the task description.
[0054] First, the device status database is queried to obtain the real-time availability and load of each computing node, simulation software license, and dedicated hardware-in-the-loop device in the current simulation resource pool, ensuring the feasibility of the recommended solution. Second, the model performance database is queried to analyze and compare the accuracy, computational efficiency, and stability performance data of different task modeling and evaluation algorithms in similar "defense mission" scenarios from historical experiments, thereby recommending the optimal model-algorithm combination. Finally, the environmental scenario database is queried to match historical scenario data that best matches the current experimental requirements, including marine geographical environment, typical electromagnetic spectrum parameters, and initial red-blue phase distance, providing parameter basis for simulation initialization.
[0055] The decision reasoning module is based on a multi-objective optimization function. Make a selection.
[0056] For defense-oriented scenarios, the system focuses on evaluating the confidence level of each algorithm in the model library:
[0057] in, Represents the availability of system equipment. Indicates the algorithm's confidence level. This represents the accuracy function of algorithm m given input variables x and target variable y. Indicates input variables The probability density function is given. The policy reasoning module generates a customized optimal experimental equipment selection and algorithm recommendation scheme, and submits it to the user for review and confirmation.
[0058] 3. Experimental Procedure Development Stage Based on the confirmed experimental task specifications and equipment algorithm scheme, the intelligent agent further develops detailed operational procedures for the experiment. The decision-making and reasoning module uses a similarity matching algorithm in the historical procedure database to retrieve historical experimental procedure cases that are highly similar to the current experimental objectives, scenarios, and equipment systems.
[0059] The intelligent agent intelligently extracts the core task flow framework from matched similar cases. This is based on similar case matching. The system extracts the core task flow framework. The task process is abstracted as a finite state machine (FSM):
[0060] in For the set of states, For event collections, For state transition function, For the initial state of the experiment, This is the set of final states of the experiment.
[0061] The agent automatically fills in the state transition conditions, such as defining the trigger conditions for the standoff phase. The following conditions are triggered for the target to enter the kill zone:
[0062] in These represent the target location, ship location, and kill zone radius, respectively. Through a formulaic definition of the process, the agent automatically fills the extracted process framework and parameter set into the current experiment's process template, quickly constructing a detailed draft process for this experiment. This draft, after user review, fine-tuning, and final confirmation, becomes an executable experimental procedure.
[0063] 4. Experiment Execution Phase The intelligent agent automatically or assisted the user in initializing and starting the simulation experiment based on the established experimental task specifications, resource allocation plan, and experimental procedures. The system then initiates the simulation and performs real-time deductions. During the deduction process, timestamps of key events are recorded. With state vector Record entity interaction logs to form an unstructured data stream. This provides the original data source for subsequent processing.
[0064] 5. Experimental Results Processing Stage After the simulation, the generated raw data needs to be transformed into data usable for experimental conclusions. The closed-loop optimization module calls the language big data model to perform semantic parsing and understanding of the raw results, identifying data fragments related to the preset evaluation indicators. After the simulation, the closed-loop optimization module uses the language big data model to parse the raw logs. And extract feature values from the text. For example, calculate the interception success rate (a key sub-indicator of system contribution rate):
[0065] in For indicator functions, To analyze the incoming targets, To destroy the number of targets, This represents the state of the j-th target at the end of the simulation.
[0066] The module is created based on the experimental task description. The final system contribution rate is calculated, and experimental conclusions in a standardized format are generated.
[0067] 6. Closed-loop feedback and optimization stage After this experiment, the closed-loop optimization module initiated the "execution-feedback-evolution" mechanism. The complete data package generated by this experiment, including the final confirmed task description, the adopted process, the model algorithm performance, and the structured experimental conclusions, all served as feedback signals.
[0068] Assuming the algorithm in this experiment The actual error is Then update the records in the model performance database. Update performance metrics using the moving average method:
[0069] in Historical weighting factor This represents the actual error of algorithm m.
[0070] If this experimental plan is successful, the weight of the decision-making rules in this type of scenario will be increased. :
[0071] in For learning rate, The contribution rate gain.
[0072] Through the aforementioned mathematical processing, new knowledge is dynamically archived into the knowledge base management module, enabling the system to make more accurate equipment selections, algorithm recommendations, and process designs when faced with similar tasks in the future, thereby achieving a continuous leap in the decision-making capabilities of the intelligent agent.
[0073] Specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. Unless otherwise specified, the embodiments and features described in this application can be arbitrarily combined with each other.
Claims
1. A simulation experiment planning and decision-making intelligent agent, characterized in that, include: The natural language interaction module is used to receive users' natural language commands, parse the semantics, match similar cases from the historical project library based on the parsing results and generate a draft of the experimental task book, and actively ask for key details during the generation process; The knowledge base management module is used to store and manage historical project libraries, device status libraries, model performance libraries, and environmental scenario libraries. The decision reasoning module is used to query the equipment status library, model performance library and environmental scenario library according to the requirements of the experimental task book draft, output a customized experimental equipment and algorithm selection recommendation scheme, and match similar experimental processes from the historical process library; The closed-loop optimization module is used to perform semantic understanding and format transformation on the unstructured results generated by the experiment, and update the knowledge base in the knowledge base management module through an execution-feedback-evolution closed-loop mechanism based on the transformed experimental conclusions.
2. The simulation experiment planning and decision-making intelligent agent according to claim 1, characterized in that, The natural language interaction module performs semantic parsing and intent recognition through a large language model.
3. The simulation experiment planning and decision-making intelligent agent according to claim 1, characterized in that, In the knowledge base management module: The historical project library is used to store historical experimental schemes, including experimental scenarios, equipment systems for both red and blue teams, task processes, indicator systems, impact factors, factor level values, and the number of simulations. The device status database is used to store the real-time availability status and load information of the experimental equipment; The model performance library is used to store performance data of different algorithm models in various environmental scenarios; The environmental scenario library is used to store typical environmental scenario data, including meteorological, electromagnetic, topographic, and environmental status parameters.
4. The simulation experiment planning and decision-making intelligent agent according to claim 1, characterized in that, The decision reasoning module uses a similarity matching algorithm to retrieve cases similar to the current experimental task from the historical process library and automatically reuses their task process framework and parameters.
5. The simulation experiment planning and decision-making intelligent agent according to claim 1, characterized in that, The closed-loop optimization module uses a large language model to perform semantic parsing on the experimental data and automatically converts it into standardized experimental results according to a preset template.
6. The simulation experiment planning and decision-making intelligent agent according to claim 5, characterized in that, The closed-loop optimization module archives the formatted experimental conclusions to the corresponding knowledge base in the knowledge base management module, which is used to update the model recommendation rules and strategy weights.
7. A simulation experiment planning and decision-making method for an intelligent agent, based on the simulation experiment planning and decision-making intelligent agent according to any one of claims 1 to 6, characterized in that, Includes the following steps: Step S1: Receive the user's natural language instructions, generate a draft of the experimental task book, and proactively ask follow-up questions to supplement key details; Step S2: Based on the confirmed experimental task sheet, query the knowledge base and output a customized experimental equipment and algorithm selection recommendation scheme; Step S3: Based on the experimental task book, match the historical process library to formulate a specific experimental process; Step S4: Perform the experiment, collect experimental conclusions, and use the language big model to automatically convert unstructured experimental data into formatted experimental conclusions; Step S5: Based on the experimental conclusions, execute the "execution-feedback-evolution" closed loop to continuously optimize the knowledge base.
8. The simulation experiment planning and decision-making method for intelligent agents according to claim 7, characterized in that, In step S1, the experimental subject, evaluation target and background scenario are identified by parsing natural language instructions, and similar cases are matched from the historical project library as a reference for generating a draft task book.
9. The simulation experiment planning and decision-making method for intelligent agents according to claim 7, characterized in that, In step S4, the conversion process includes: inputting the original logs from multiple experimental simulations into the language model, and outputting structured experimental conclusions according to the format requirements of the evaluation template.
10. The simulation experiment planning and decision-making method for intelligent agents according to claim 7, characterized in that, In step S5, the optimization knowledge base includes: updating the algorithm performance records in the model performance base, the scene parameters in the environment scene base, and the experimental schemes in the historical project base based on the new data and experimental conclusions of this experiment.