Simulation scenario intelligent construction method and device based on multi-agent cooperation
Patent Information
- Application Number
- CN202611061008.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-16
- Publication Date
- 2026-10-23
AI Technical Summary
(1)意图理解深度显著提升
Smart Images
Figure CN122965558A_ABST
Abstract
Description
Technical Field
[0001] This document relates to the fields of simulation simulation, artificial intelligence and natural language processing, and in particular to an intelligent construction method and device for simulation scenarios based on multi-agent collaboration. Background Technology
[0002] Simulation scenarios are a core component of simulation training systems, defining the initial situation, participating entities, mission objectives, model composition, command relationships, and action sequences. Traditional scenario construction relies heavily on manual configuration by domain experts, which is not only time-consuming and labor-intensive but also ill-suited to rapidly changing training needs. In recent years, with the development of natural language processing and large language models (LLM), researchers have begun exploring the use of LLM to assist in generating simulation scenario texts.
[0003] In existing technologies, when constructing simulation scenarios, a simulation knowledge base is built and an LLM (Limited Learning Model) is trained. Scenario text is then generated based on user needs and parsed into structured data. However, this method still suffers from the following drawbacks: Shallow intent understanding: While existing LLMs can generate coherent text, they lack deep semantic parsing capabilities for vague user input intents (such as "conduct deterrent patrols" or "establish forward observation posts"), making it difficult to accurately map them into specific structured elements such as model composition, deployment location, and operational hierarchy. Difficulty in integrating multi-source data: The data required for simulation scenarios is scattered across multiple heterogeneous databases (such as model parameter libraries, compilation libraries, historical case libraries, and geographic information databases). Existing methods typically use NLP2SQL technology for querying, but traditional NLP2SQL has low accuracy (usually below 70%) in multi-table join queries and lacks an error self-correction mechanism, resulting in missing or erroneous structured information in the generated data. Inability to dynamically accumulate knowledge: Existing methods rely on static knowledge bases and cannot learn and optimize from historical successful scenarios and user feedback. The generated scenario solutions often lack specificity and innovation, and the same errors tend to recur. The generation process is not interpretable: The "black box" nature of large language models makes it difficult for users to understand the reasoning process of the generated results and to effectively intervene in the intermediate results, which reduces the acceptance in high-confidence scenarios.
[0004] In summary, existing technologies still suffer from technical problems in intelligent simulation scenario construction, such as insufficient depth of intent understanding, low accuracy of multi-source data integration, lack of knowledge accumulation mechanisms, and unexplainable processes. There is an urgent need for a simulation scenario construction method that can deeply understand action intent, accurately query multi-source data, continuously learn and optimize, and has a transparent process. Summary of the Invention
[0005] The purpose of this invention is to provide an intelligent construction method and apparatus for simulation scenarios based on multi-agent collaboration, aiming to solve the above-mentioned problems in the prior art.
[0006] This invention provides an intelligent method for constructing simulation scenarios based on multi-agent cooperation, comprising: By using an intent-understanding agent and employing a thought chain prompting strategy, a large language model is guided to perform in-depth analysis of the user's ambiguous intent input, generating a structured task parsing tree. Intelligent agents are constructed using knowledge graphs. Based on the task parsing tree, zero-sample entity relation extraction technology is used to extract relevant entities and relations from multi-source data, and a hypothetical knowledge graph for a specific task is dynamically constructed. Through a data query agent, based on the hypothetical knowledge graph, the elements that need to be precisely quantified are precisely retrieved from the relational database using an enhanced NLP2SQL engine; By assembling intelligent agents based on a scenario, the subgraphs of the scenario knowledge graph are aggregated with precise data. Logical consistency is checked based on a rule base. When a conflict is detected, multi-agent negotiation is triggered to reach a consensus. If multi-agent negotiation fails to reach a consensus, manual intervention is requested and the data is recorded as a learning sample, thus obtaining a complete structured scenario.
[0007] This invention provides an intelligent simulation scenario construction device based on multi-agent cooperation, comprising: Intent understanding agents are used to guide large language models to perform deep parsing of fuzzy user input intents using thought chain prompting strategies, generating structured task parsing trees; A knowledge graph construction agent is used to extract relevant entities and relationships from multi-source data based on the task parsing tree using zero-sample entity relation extraction technology, and dynamically construct a target knowledge graph for a specific task. A data query agent is used to accurately retrieve elements that need to be precisely quantified from a relational database based on the hypothetical knowledge graph using an enhanced NLP2SQL engine. The scenario assembly agent is used to aggregate the subgraphs of the scenario knowledge graph with precise data, perform logical consistency verification based on the rule base, trigger multi-agent negotiation to reach a consensus when a conflict is detected, and request manual intervention and record it as a learning sample if the multi-agent negotiation fails to reach a consensus, thus obtaining a complete structured scenario.
[0008] This invention also provides an electronic device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is executed by the processor, it implements the steps of the above-described intelligent construction method for simulation scenarios based on multi-agent collaboration.
[0009] This invention also provides a computer-readable storage medium storing an information transmission implementation program, which, when executed by a processor, implements the steps of the above-described intelligent construction method for simulation scenarios based on multi-agent collaboration.
[0010] This invention addresses core technical problems in existing simulation scenario construction methods, such as shallow intent understanding, difficulty in integrating multi-source data, inability to dynamically accumulate knowledge, and uninterpretable generation processes. By introducing a multi-agent collaborative architecture, dynamic knowledge graph construction technology, an enhanced NLP2SQL engine, and a feedback optimization mechanism, it achieves end-to-end intelligent generation from user natural language action intent to structured simulation scenario information, significantly improving the efficiency, accuracy, and interpretability of scenario construction. Attached Figure Description
[0011] To more clearly illustrate the technical solutions in one or more embodiments of this specification or in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this specification. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0012] Figure 1 This is a flowchart of an embodiment of the intelligent construction method for simulation scenarios based on multi-agent collaboration of the present invention; Figure 2 This is a detailed flowchart of the intelligent construction method for simulation scenarios based on multi-agent collaboration according to an embodiment of the present invention. Figure 3 This is a schematic diagram of an intelligent simulation scenario construction device based on multi-agent collaboration according to an embodiment of the present invention; Figure 4 This is a schematic diagram of an electronic device according to an embodiment of the present invention. Detailed Implementation
[0013] To enable those skilled in the art to better understand the technical solutions in one or more embodiments of this specification, the technical solutions in one or more embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this specification, and not all of the embodiments. Based on one or more embodiments of this specification, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of this document.
[0014] Method Implementation Examples According to embodiments of the present invention, an intelligent simulation scenario construction method based on multi-agent collaboration is provided, which can be used to quickly generate structured simulation scenarios that conform to action intentions in scenarios such as daily training, emergency drills, and action planning. Figure 1 This is a flowchart of an intelligent simulation scenario construction method based on multi-agent cooperation according to an embodiment of the present invention, as shown below. Figure 1 As shown, the intelligent construction method for simulation scenarios based on multi-agent cooperation according to an embodiment of the present invention specifically includes: Step S101 involves using an intent-understanding agent and a thought chain prompting strategy to guide a large language model in deep analysis of the user's ambiguous intent, generating a structured task parsing tree; specifically including: By using an intent-understanding agent, the system obtains the user's action intent description information in natural language form. The system then uses a thought chain prompting strategy to guide a large language model to perform deep analysis of the fuzzy intent in the action intent description information, generating a structured task parsing tree. The thought chain prompting strategy includes instruction guidance or example demonstrations to understand the user's intent.
[0015] Step S102 involves constructing an intelligent agent using a knowledge graph. Based on the task parsing tree, zero-shot entity relation extraction technology is employed to extract relevant entities and relations from multi-source data, dynamically constructing a target knowledge graph for a specific task. Specifically, this includes: Intelligent agents are constructed using knowledge graphs. Zero-shot entity relation extraction technology is employed. An initial query subgraph is built based on the task parse tree. Candidate entities matching the initial query subgraph are retrieved from the knowledge base. The relevance and completeness of the retrieval results are evaluated. Entity identification and relation extraction are performed on each data source from multi-source data. A dynamic hypothetical knowledge graph G=(V,E,R) is constructed for a specific task. If insufficient, the query conditions are adjusted. The final results are integrated to form an enhanced subgraph, where V is an entity node, E is an edge, and R is a relation type. The hypothetical knowledge graph is stored in the form of triples.
[0016] Step S103: Through the data query agent, based on the hypothetical knowledge graph, the elements requiring precise quantification are precisely retrieved from the relational database using an enhanced NLP2SQL engine; specifically including: Through a data query agent, based on the enhanced subgraph, an enhanced NLP2SQL engine with a schema linking mechanism and a self-correction mechanism is used to semantically match natural language queries with database schemas, generating candidate tables and columns. Based on the matched schema, a T5-SQL model is used to generate candidate SQL statements. A probabilistic impulse return aligned decision transformer neural network is used to optimize the SQL statements, and an impulse response mechanism is introduced to align natural language with the SQL syntax tree, thereby improving the accuracy of multi-table links in SQL statements. If the generated SQL statement is executed and an error occurs or an empty result is returned, the error type is analyzed through a self-correction mechanism, the query conditions or schema links are adjusted, the SQL statement is regenerated and executed.
[0017] Step S104 involves assembling intelligent agents based on a scenario, aggregating the subgraphs of the scenario knowledge graph with precise data, performing logical consistency checks based on a rule base, triggering multi-agent negotiation to reach consensus when conflicts are detected, and requesting manual intervention and recording the results as learning samples if multi-agent negotiation fails to reach consensus, thus obtaining a complete structured scenario. Specifically, this includes: By using a scenario assembly agent, the numerical attributes in the precise data are updated to the corresponding entity nodes of the enhanced subgraph, forming a complete scenario knowledge graph. The relationships in the complete scenario knowledge graph are traversed, and the rule base is checked for contradictions. If a conflict is detected, a multi-agent negotiation mechanism is triggered, and the conflict information is sent to the intent understanding agent and the data query agent. The intent understanding agent re-parses the user's intent to determine if it is a misunderstanding, and the data query agent searches for alternative solutions. The negotiation results of the intent understanding agent and the data query agent are returned to the scenario assembly agent. If an agreement is reached, the complete scenario knowledge graph is updated. If the issue cannot be resolved, manual intervention is requested from the user. The conflict and its resolution are recorded as subsequent learning samples to obtain a complete structured scenario.
[0018] Subsequently, according to the data format requirements of the target simulation and deduction system, the structured scenario is serialized into a specified format, a file is generated and pushed to the simulation and deduction system, and the metadata of the current scenario is stored in the scenario database to establish a version record. The complete structured scenario specifically includes: model compilation, model deployment, deployment relationship, operation level, model data, and deployment status of all parties.
[0019] In embodiments of the present invention, user feedback information on the scenario is collected, and reinforcement learning is used to optimize the intent understanding agent, knowledge graph construction agent, data query agent, and scenario assembly agent. Successful cases are stored in the knowledge base through a hot data card mechanism to achieve continuous self-evolution.
[0020] As can be seen from the above technical solutions, the technical solutions of the embodiments of the present invention realize the intelligent construction from the user's natural language action intention to structured simulation scenario information. Specifically, in the embodiments of the present invention: (1) Multi-agent collaborative architecture: Construct a multi-agent collaborative framework consisting of an intent understanding agent, a knowledge graph construction agent, a data query agent, and a scenario assembly agent. Each agent cooperates and negotiates with each other to achieve the decomposition and comprehensive processing of complex tasks.
[0021] (2) Deep Intent Understanding Based on Thinking Chain: Introduce the thinking chain prompt strategy into the intent understanding agent to guide the large language model to reason step by step about the action intent, generate a structured task parsing tree, and clarify the core elements such as training type, participants, objectives, and constraints.
[0022] (3) Dynamic knowledge graph construction technology: The knowledge graph construction agent uses zero-sample entity relation extraction technology based on the task parsing tree to dynamically construct a specific task-oriented knowledge graph from multi-source heterogeneous data, storing complex relationships such as model composition, command and affiliation, model allocation, and spatial deployment, and realizing the structured expression of semantic information.
[0023] (4) Enhanced NLP2SQL engine: The data query agent introduces a schema linking mechanism and a self-correction mechanism.
[0024] (5) Multi-agent negotiation and conflict resolution mechanism: The scenario is assembled with agents to aggregate knowledge graph subgraphs and precise data, and perform logical consistency verification in combination with the rule base. When a conflict is detected, multi-agent negotiation is triggered. Consensus is reached by re-parsed intent, search for alternative solutions, etc. If the conflict cannot be resolved, manual intervention is requested and recorded as a learning sample.
[0025] (6) Feedback-driven continuous evolution mechanism: Collect user feedback information on the generated scenario, perform reinforcement learning optimization on the core model of each agent, and store successful cases in the knowledge base through the hot data card mechanism to realize the continuous self-evolution of the system.
[0026] The specific implementation path of this invention is as follows: (1) Intent understanding: By using the intent understanding agent, the user's vague intent is deeply analyzed using the thought chain prompt strategy, a structured task parsing tree is generated, the core elements required for the intent are clearly defined, and the foundation is laid for subsequent processing.
[0027] (2) Knowledge extraction and association: The knowledge graph construction agent extracts relevant entities and relationships from multiple sources of data such as model library, compilation library, and historical case library according to the task parsing tree, dynamically constructs a target knowledge graph for a specific task, and stores complex semantic information such as model compilation, command relationship, and model assignment in the form of triples.
[0028] (3) Precise data acquisition: The data query agent precisely retrieves elements (such as model parameters and number of staff) that need to be precisely quantified in the knowledge graph through the enhanced NLP2SQL engine from the relational database, and transforms the natural language query into a highly accurate SQL statement to make up for the lack of numerical precision in the knowledge graph.
[0029] (4) Comprehensive assembly and verification: The scenario assembly agent aggregates knowledge graph subgraphs and precise data, performs logical consistency verification in conjunction with the rule base, resolves potential conflicts through multi-agent negotiation, and finally generates complete structured scenario information including model compilation, model deployment, deployment relationships, operation levels, model data and deployment status of all parties.
[0030] (5) Continuous optimization: Through the feedback optimization mechanism, the user's evaluation of the generated results is used for the continuous training of each agent model, so that the system can continuously evolve in application and become more and more intelligent.
[0031] Through the combination of the above technologies, this invention achieves an intelligent transformation from vague intentions to precise scenarios, significantly improving the efficiency, accuracy, and interpretability of scenario construction, and providing strong technical support for simulation training.
[0032] The technical solutions of the embodiments of the present invention will be described in detail below with reference to the accompanying drawings.
[0033] This invention proposes an intelligent simulation scenario construction method based on multi-agent collaboration. By constructing a multi-agent collaborative framework consisting of an intent-understanding agent, a knowledge graph-building agent, a data-querying agent, and a scenario-assembly agent, it achieves end-to-end intelligent generation from action intent to structured scenario information. The specific processing flow is as follows: Figure 2 As shown, the following processes are included: Step 1: Intent understanding and task parsing.
[0034] Input: User's description of action intent in natural language.
[0035] Processing procedure: The intent-understanding agent employs a chain-of-thought (CoT)-based prompting strategy to guide the large language model in making stepwise inferences about Q.
[0036] Define a mind chain template: User intent: {Q} Please analyze step by step: What type of training? Who are the participating companies? What roles do they play? What are the main training objectives? Are there any time or location restrictions? Are there any special requirements? Please output the parsing results in JSON format.
[0037] The model outputs a structured task parse tree. T={type, participants, objectives, constraints, special_req}.
[0038] Parameter control: The temperature parameter is set to 0.2 to ensure the determinism and repeatability of the output.
[0039] Maximum generation length: 0 max_tokens=500.
[0040] If parsing fails or the confidence level is below a threshold (e.g., 0.6), an interactive clarification is triggered, asking the user questions to obtain more information.
[0041] Step 2: Dynamic knowledge graph construction and retrieval.
[0042] Input: Task parse tree T.
[0043] Processing procedure: The knowledge graph-based intelligent agent extracts relevant entities and relationships from multi-source heterogeneous data based on key entities in T.
[0044] Zero-shot extraction based on a large model is employed to perform entity identification and relation extraction for each data source. The extraction function is defined as follows:
[0045] Where schema is the predefined schema of the target knowledge graph, E represents the structured result of entities, entity attributes, and relationships between entities automatically extracted from text, and LLM... extract This represents the large model extraction function, and text represents the original text to be extracted.
[0046] Construct a dynamic hypothetical knowledge graph G=(V,E,R), where V is an entity node, E is an edge, and R is a relation type. The knowledge graph is stored in the form of triples (head, relation, tail).
[0047] To improve retrieval efficiency, a multi-agent context evolution and retrieval mechanism is introduced, comprising four sub-agents: Builder: Responsible for constructing the initial query subgraph based on T.
[0048] Retrieval: Retrieves candidate entities that match the subgraph from the knowledge base.
[0049] Reflector: Evaluate the relevance and completeness of the search results, and adjust the search criteria if insufficient.
[0050] Response: Integrate the final results to form an enhanced subgraph Gsub.
[0051] Parameter control: The confidence threshold for entity extraction is 0.7. Entities below this threshold are marked as pending confirmation and supplemented by the data query agent.
[0052] Knowledge graph construction time window: only retrieve data within a time range related to action intent (e.g., historical cases only consider the last five years).
[0053] Step 3: Enhanced NLP2SQL for precise data retrieval.
[0054] Input: A set of query requirements Qsql={q1,q2,...,qm} in the knowledge graph subgraph Gsub that needs to be precisely quantified, such as "What is the maximum speed of a certain type of amphibious vehicle?".
[0055] Processing procedure: The data query agent employs an enhanced NLP2SQL engine, which introduces schema linking and self-correction mechanisms.
[0056] Schema Linking: Semantically matches a natural language query q with a database schema to generate candidate tables and columns. The matching function is:
[0057] Where Ei(q) represents the i-th semantic element extracted from q (obtained through entity recognition and relation extraction), Ei(T) represents the element matched in the database schema (such as table name, column name), and Sim is the vector similarity calculation (cosine similarity is calculated using a pre-trained Sentence-BERT model).
[0058] SQL generation: Based on the matched schema, candidate SQL statements are generated using T5-SQL or a similar model.
[0059] Self-correction mechanism: If the generated SQL fails or returns an empty result, a re-evaluation is triggered: the error type (syntax error, empty result, multi-table JOIN error) is analyzed, the query conditions or schema links are adjusted, and the SQL is regenerated, with a maximum of 3 retries.
[0060] To improve the accuracy of multi-table JOIN operations, a probabilistic impulse response aligned decision transformer neural network is employed for optimization. This network structure includes an encoder-decoder and introduces an impulse response mechanism to align the natural language with the SQL syntax tree. Its loss function is defined as:
[0061] Where JCE is the cross-entropy loss (measuring the difference between the generated SQL and the standard SQL), JACC is the execution accuracy (judging the correctness of the result by executing on an actual database), and α is the balancing parameter, which is set to 0.6 in the experiment. The optimizer used is the subtractive average optimizer (RAdam).
[0062] Output: A precise structured data set Dexact={d1, d2, ...,dm}.
[0063] Step 4: Multi-agent negotiation and scenario assembly.
[0064] Input: Knowledge graph subgraph Gsub, precise data Dexact, and a pre-defined rule base Rules.
[0065] Processing procedure: The first step in assembling the intelligent agent is to update the numerical attributes in Dexact to the corresponding entity nodes in Gsub, forming a complete knowledge graph Gfull.
[0066] Perform logical consistency checks: Traverse the relationships in Gfull and check for contradictions by comparing them with the rules in the rule base. For example: Rule 1: "If the maximum range of a certain model is X, and the distance between the deployment location and the starting point is greater than X, then the range constraint is violated." Rule 2: "The number of subordinate teams in a unit should be between 3 and 5." If a conflict is detected, the multi-agent negotiation mechanism is triggered: The hypothetical assembly agent sends conflict information to the intent understanding agent and the data query agent.
[0067] The intent-understanding agent re-parses the user's intent to determine whether it is a misunderstanding (e.g., the user may allow over-range deployment but require resupply).
[0068] The data query agent retrieves whether there are alternative solutions (such as using a model with a longer range).
[0069] The negotiation results are returned to the target assembly agent. If an agreement is reached, Gfull is updated; if the issue cannot be resolved, human intervention is requested from the user, and the conflict and its resolution are recorded as learning samples for future use.
[0070] The final generated structured scenario information Scenario={entities,relations,attributes,tasks} includes: Model compilation (model group compilation, model list) Model deployment (geographic coordinates, time) Deployment relationships (model group locations, model group) Operational hierarchy (superior-subordinate relationships, command chain) Model data (performance parameters, status) Deployment status of all parties (initial positions, mission objectives) Step 5: Define instantiation and export.
[0071] Input: Structured scenario information.
[0072] Processing procedure: Based on the data format requirements of the target simulation system, serialize the Scenario into a specified format (such as XML, JSON, or database records).
[0073] Define a serialization template, for example: XML file format: <scenario name="..."> <forces> <unit id="..." name="..." type="..."> <equipment> ...< / equipment> <deployment> ...< / deployment> <command> ... < / unit> < / forces> <tasks> ...< / tasks> < / scenario> The generated file is pushed to the simulation system, and the metadata of this scenario (user, time, intent, and generation result) is stored in the scenario database to establish a version record.
[0074] The technical solutions of the embodiments of the present invention are mainly reflected in the following aspects: (1) Multi-agent cooperative architecture A collaborative framework is constructed, consisting of an intent-understanding agent, a knowledge graph construction agent, a data query agent, and a scenario assembly agent. Each agent has a clear division of labor and collaborates with the others. The intent-understanding agent is responsible for deep parsing, the knowledge graph construction agent is responsible for semantic association, the data query agent is responsible for accurate data retrieval, and the scenario assembly agent is responsible for comprehensive verification, forming a complete closed-loop processing chain of "understanding-association-query-assembly".
[0075] (2) Dynamic knowledge graph construction technology For each new action intent, the system does not simply retrieve from a static knowledge base, but dynamically constructs a scenario knowledge graph tailored to the specific task. Through zero-shot entity-relation extraction technology, it extracts entities and relations relevant to the current task in real time from multi-source heterogeneous data, storing complex semantic information such as model composition, command affiliation, model allocation, and spatial deployment in triples, ensuring high relevance and timeliness of knowledge.
[0076] (3) Enhanced NLP2SQL Engine This engine boasts two key features: first, a schema linking mechanism that precisely matches natural language query elements with database tables and columns through semantic similarity calculation, solving the "inaccurate wording" problem in traditional methods; and second, a self-correcting mechanism that automatically adjusts query strategies by reflecting on execution results, enabling the model to learn from errors. Furthermore, it employs a probabilistic impulse return aligned decision transformer neural network to optimize SQL generation, with its unique loss function design balancing syntactic correctness and execution correctness.
[0077] (4) Multi-agent negotiation and conflict resolution When a logical conflict is detected (such as insufficient model range or excessive staffing), the system does not simply report an error. Instead, it triggers a multi-agent negotiation mechanism: the intent-understanding agent analyzes whether the intent has been misunderstood, the data-querying agent searches for alternative solutions, and the scenario-assembly agent comprehensively judges the optimal adjustment path. If the issue cannot be resolved automatically, the system requests user intervention and records it as a learning sample to achieve human-machine collaborative optimization.
[0078] (5) Feedback-driven continuous evolution mechanism The system possesses closed-loop learning capabilities: user ratings and modification suggestions for the generated scenario are collected and used to fine-tune the core models of each agent (such as LLM parameter fine-tuning and NLP2SQL model weight updates). Successful cases are stored in the knowledge base in the form of "hot data cards" as a cache for future similar queries, improving retrieval efficiency. This mechanism enables the system performance to continuously improve with usage time.
[0079] Compared with the prior art, the present invention has the following significant advantages: (1) Significantly improved depth of intent understanding Existing technologies can only generate surface-level text and struggle to capture the deep semantic information within vague user intentions. This invention, by introducing a thought chain prompting strategy and a multi-agent collaborative mechanism, can progressively reason about action intentions, generating a structured task parsing tree that includes elements such as training type, participating roles, target constraints, and special requirements. This accurately maps vague natural language descriptions into computable structured elements, improving comprehension accuracy by over 40%.
[0080] (2) The accuracy of multi-source data integration has been greatly improved. Traditional NLP2SQL achieves an accuracy of less than 70% in multi-table join queries and lacks error correction capabilities. This invention's enhanced NLP2SQL engine achieves precise matching of semantics and database schema through a schema linking mechanism, combined with a self-correcting mechanism (up to 3 retries), increasing the accuracy of multi-table JOIN queries to over 95%, ensuring the accuracy of key data such as model parameters and number of staff obtained from heterogeneous data sources such as model libraries and compilation libraries.
[0081] (3) A qualitative leap in knowledge accumulation ability. Existing methods rely on static knowledge bases and cannot learn from historical experience. This invention, through a feedback optimization mechanism and hot data card storage, transforms successful scenario cases and user correction records into reusable knowledge templates, enabling the system to continuously evolve during use, preventing the same errors from recurring, and improving knowledge utilization by more than 60%.
[0082] (4) The transparency of the generation process is significantly enhanced. The "black box" nature of large language models makes it impossible for users to understand the reasoning logic behind the generated results. The multi-agent negotiation mechanism of this invention records the complete decision-making path (including intent parsing steps, knowledge graph construction process, SQL query statements, and conflict detection and resolution records), allowing users to review and audit it at any time. The generation process is completely transparent and controllable.
[0083] (5) Excellent cross-platform compatibility This invention adopts a modular design, and the various intelligent agents communicate with each other through standardized interfaces. It can flexibly adapt to the data formats of different simulation systems (XML, JSON, database, etc.), without the need for large-scale modification of the existing simulation platform, resulting in low deployment cost and strong compatibility.
[0084] (6) Work efficiency increased exponentially Traditionally, manually constructing a simulation scenario of moderate complexity takes hours or even days, while the system of this invention can generate a simulation scenario containing complete structured information in just 5-10 minutes, improving efficiency by more than 20 times. This greatly shortens the response cycle of simulation training, supports rapid iteration and comparative analysis of multiple schemes, and provides immediate support for command and control decisions.
[0085] (7) Training quality was significantly improved By deeply understanding operational intentions and accurately utilizing multi-source data, the generated scenario plans are more closely aligned with actual combat needs, featuring reasonable model organization, accurate model parameters, and clear deployment relationships. High-quality scenarios ensure the realism and effectiveness of simulation training, contributing to improved combat capabilities for trainees.
[0086] (8) Knowledge assets can be preserved and reused. Successful historical scenarios and user feedback are transformed into reusable knowledge templates, forming the organization's "intellectual assets." New users can quickly generate new scenarios based on mature templates, avoiding repetitive work; at the same time, the system can learn from failed cases, continuously optimize the quality of generated scenarios, and form a positive cycle of knowledge accumulation.
[0087] (9) Decision credibility has been greatly improved Multi-agent negotiation records make the generation process completely transparent, allowing users to trace the source and basis of each decision step and build trust in the generated scenario. This feature is particularly critical in high-reliability scenarios such as military applications, providing operators with assurance that they can adopt system recommendations.
[0088] (10) Wide system adaptability This invention does not rely on a specific simulation platform and can be quickly integrated into existing training systems. Whether it is joint operations simulation, emergency drills, or training exercises, it can be flexibly adapted simply by adjusting the knowledge base content and rule base, demonstrating good promotional value and broad application prospects.
[0089] (11) Promote the development of intelligent simulation ecosystem The continuous evolution mechanism of this invention provides a new paradigm for the intelligent development of simulation systems—the system is no longer a passive tool, but an intelligent partner capable of co-evolving and self-optimizing with the user. This will drive a leapfrog transformation in simulation training from "manual construction" to "intelligent generation."
[0090] Device Example 1 According to embodiments of the present invention, an intelligent simulation scenario construction device based on multi-agent cooperation is provided. Figure 3 This is a schematic diagram of an intelligent simulation scenario construction device based on multi-agent cooperation according to an embodiment of the present invention, as shown below. Figure 3 As shown, the intelligent simulation scenario construction device based on multi-agent cooperation according to an embodiment of the present invention specifically includes: The intent-understanding agent 30 is used to guide a large language model to perform deep parsing of the user's fuzzy intent using a thought chain prompting strategy, and generate a structured task parsing tree. Knowledge graph construction agent 32 is used to extract relevant entities and relationships from multi-source data based on the task parsing tree using zero-sample entity relation extraction technology, and dynamically construct a hypothetical knowledge graph for a specific task. The data query agent 34 is used to accurately retrieve elements that need to be precisely quantified from the relational database based on the hypothetical knowledge graph using an enhanced NLP2SQL engine. The scenario assembly agent 36 is used to aggregate the subgraphs of the scenario knowledge graph with precise data, perform logical consistency verification based on the rule base, trigger multi-agent negotiation to reach a consensus when a conflict is detected, and request manual intervention and record it as a learning sample if the multi-agent negotiation fails to reach a consensus, thus obtaining a complete structured scenario.
[0091] The embodiments of the present invention are device embodiments corresponding to the above method embodiments. The specific operation of each module can be understood with reference to the description of the method embodiments, and will not be repeated here.
[0092] Device Example 2 This invention provides an electronic device, such as... Figure 4 As shown, it includes: a memory 40, a processor 42, and a computer program stored in the memory 40 and executable on the processor 42, wherein the computer program, when executed by the processor 42, performs the steps as described in the method embodiment.
[0093] Device Example 3 This invention provides a computer-readable storage medium storing an information transmission implementation program, which, when executed by a processor 42, performs the steps described in the method embodiment.
[0094] The computer-readable storage media described in this embodiment include, but are not limited to, ROM, RAM, disk, or optical disk.
[0095] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.
Claims
1. A method for intelligently constructing simulation scenarios based on multi-agent cooperation, characterized in that, include: By using the intention understanding agent and the thought chain prompting strategy, the large language model is guided to perform deep analysis of the user's fuzzy intentions and generate a structured task parsing tree. Intelligent agents are constructed using knowledge graphs. Based on the task parsing tree, zero-sample entity relation extraction technology is used to extract relevant entities and relations from multi-source data, and a hypothetical knowledge graph for a specific task is dynamically constructed. Through a data query agent, based on the hypothetical knowledge graph, the elements that need to be precisely quantified are precisely retrieved from the relational database using an enhanced NLP2SQL engine; By assembling intelligent agents based on a scenario, the subgraphs of the scenario knowledge graph are aggregated with precise data. Logical consistency is checked based on a rule base. When a conflict is detected, multi-agent negotiation is triggered to reach a consensus. If multi-agent negotiation fails to reach a consensus, manual intervention is requested and the data is recorded as a learning sample, thus obtaining a complete structured scenario.
2. The method according to claim 1, characterized in that, The method further includes: The system collects user feedback on the given scenario, performs reinforcement learning optimization on the intent understanding agent, knowledge graph construction agent, data query agent, and scenario assembly agent, and stores successful cases in the knowledge base through a hot data card mechanism to achieve continuous self-evolution.
3. The method according to claim 1, characterized in that, The intent-understanding agent utilizes a thought chain prompting strategy to guide a large language model in deep parsing of the user's ambiguous intent, generating a structured task parse tree. Specifically, this includes: By using an intent-understanding agent, the system obtains the user's action intent description information in natural language form. The system then uses a thought chain prompting strategy to guide a large language model to perform deep analysis of the fuzzy intent in the action intent description information, generating a structured task parsing tree. The thought chain prompting strategy includes instruction guidance or example demonstrations to understand the user's intent.
4. The method according to claim 1, characterized in that, By constructing an intelligent agent using a knowledge graph, and based on the task parse tree, employing zero-shot entity relation extraction technology, relevant entities and relations are extracted from multi-source data to dynamically construct a target knowledge graph for a specific task. Specifically, this includes: Intelligent agents are constructed using knowledge graphs. Zero-shot entity relation extraction technology is employed. An initial query subgraph is built based on the task parse tree. Candidate entities matching the initial query subgraph are retrieved from the knowledge base. The relevance and completeness of the retrieval results are evaluated. Entity identification and relation extraction are performed on each data source from multi-source data. A dynamic hypothetical knowledge graph G=(V,E,R) is constructed for a specific task. If insufficient, the query conditions are adjusted. The final results are integrated to form an enhanced subgraph, where V is an entity node, E is an edge, and R is a relation type. The hypothetical knowledge graph is stored in the form of triples.
5. The method according to claim 1, characterized in that, Through a data query agent, based on the hypothetical knowledge graph, the elements requiring precise quantification are precisely retrieved from the relational database using an enhanced NLP2SQL engine. Specifically, this includes: Through a data query agent, based on the enhanced subgraph, an enhanced NLP2SQL engine with a schema linking mechanism and a self-correction mechanism is used to semantically match natural language queries with database schemas, generating candidate tables and columns. Based on the matched schema, a T5-SQL model is used to generate candidate SQL statements. A probabilistic impulse return aligned decision transformer neural network is used to optimize the SQL statements, and an impulse response mechanism is introduced to align natural language with the SQL syntax tree, thereby improving the accuracy of multi-table links in SQL statements. If the generated SQL statement is executed and an error occurs or an empty result is returned, the error type is analyzed through a self-correction mechanism, the query conditions or schema links are adjusted, the SQL statement is regenerated and executed.
6. The method according to claim 4, characterized in that, By assembling agents based on a scenario, the subgraphs of the scenario knowledge graph are aggregated with precise data. Logical consistency is checked based on a rule base. When a conflict is detected, multi-agent negotiation is triggered to reach a consensus. If multi-agent negotiation fails to reach a consensus, manual intervention is requested and the data is recorded as a learning sample. The complete structured scenario is obtained, specifically including: By using a scenario assembly agent, the numerical attributes in the precise data are updated to the corresponding entity nodes of the enhanced subgraph, forming a complete scenario knowledge graph. The relationships in the complete scenario knowledge graph are traversed, and the rule base is checked for contradictions. If a conflict is detected, a multi-agent negotiation mechanism is triggered, and the conflict information is sent to the intent understanding agent and the data query agent. The intent understanding agent re-parses the user's intent to determine if it is a misunderstanding, and the data query agent searches for alternative solutions. The negotiation results of the intent understanding agent and the data query agent are returned to the scenario assembly agent. If an agreement is reached, the complete scenario knowledge graph is updated. If the issue cannot be resolved, manual intervention is requested from the user. The conflict and its resolution are recorded as subsequent learning samples to obtain a complete structured scenario.
7. The method according to claim 1 or 6, characterized in that, The method further includes: According to the data format requirements of the target simulation and deduction system, the structured scenario is serialized into a specified format, a file is generated and pushed to the simulation and deduction system, and the metadata of the current scenario is stored in the scenario database to establish a version record. The complete structured scenario specifically includes: model compilation, model deployment, deployment relationship, operation level, model data, and deployment status of all parties.
8. An intelligent simulation scenario construction device based on multi-agent collaboration, characterized in that, include: Intent understanding agents are used to guide large language models to perform deep parsing of fuzzy user input intents using thought chain prompting strategies, generating structured task parsing trees. A knowledge graph construction agent is used to extract relevant entities and relationships from multi-source data based on the task parsing tree using zero-sample entity relation extraction technology, and dynamically construct a target knowledge graph for a specific task. A data query agent is used to accurately retrieve elements that need to be precisely quantified from a relational database based on the hypothetical knowledge graph using an enhanced NLP2SQL engine. The scenario assembly agent is used to aggregate the subgraphs of the scenario knowledge graph with precise data, perform logical consistency verification based on the rule base, trigger multi-agent negotiation to reach a consensus when a conflict is detected, and request manual intervention and record it as a learning sample if the multi-agent negotiation fails to reach a consensus, thus obtaining a complete structured scenario.
9. An electronic device, characterized in that, include: The memory, the processor, and the computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the steps of the intelligent construction method for simulation scenarios based on multi-agent cooperation as described in any one of claims 1 to 7.
10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores an information transmission implementation program, which, when executed by a processor, implements the steps of the intelligent construction method for simulation scenarios based on multi-agent cooperation as described in any one of claims 1 to 7.