A method, system, equipment, and medium for generating military combat simulation schemes.

Abstract

This invention provides a method, system, device, and medium for generating military combat simulation schemes, belonging to the field of military combat simulation technology. The method includes: parsing combat tasks described in natural language according to a preset large language model to generate structured task data; automatically generating a problem description file conforming to the planning domain definition language specification based on the structured task data and a preset combat domain knowledge base; inputting the problem description file and a predefined domain definition file into a planning domain definition language planner for solving, outputting a feasible action sequence that satisfies all preset hard constraints; using the feasible action sequence as the initial strategy, executing and evaluating it in a simulation environment; iteratively optimizing the action sequence using a reinforcement learning algorithm based on the evaluation results, generating an optimized action sequence as the final simulation scheme. This method solves the problem of low feasibility and rule compliance in existing simulation schemes.

Description

Technical Field

[0001] This invention belongs to the field of military combat simulation technology, specifically relating to a method, system, equipment, and medium for generating military combat simulation schemes. Background Technology

[0002] In modern warfare, the quality of operational simulation schemes directly impacts the reference value of simulation exercises for real combat operations. High-quality simulation schemes not only need to meet basic physical constraints, tactical rules, and equipment capability limitations, but also require comprehensive optimization across multiple dimensions, including mission completion rate, resource consumption, and execution efficiency. Traditional simulation scheme development primarily relies on military experts manually compiling schemes based on their tactical knowledge and training experience. While this method ensures the professionalism and feasibility of the schemes, it suffers from problems such as long development cycles, difficulty in handling large-scale complex scenarios, and a lack of automated optimization capabilities.

[0003] With the development of artificial intelligence technology, various automated methods for generating combat simulation schemes have emerged. Methods based on rule engines or knowledge graphs can automatically generate simulation schemes that meet constraints by reasoning through predefined domain knowledge and logical rules, but they suffer from problems such as rigid rules, limited coverage, and difficulty adapting to dynamically changing battlefield situations. Methods based on large language models utilize natural language understanding and generation capabilities to directly generate simulation schemes based on predefined scenarios described in natural language, offering advantages in semantic expression and structured scene description, but lacking a rigorous verification mechanism for hard constraints; the generated content may violate physical limitations, tactical rules, or equipment capability constraints. Planner methods based on Planning Domain Definition Language (PDDL) ensure that the generated simulation schemes meet all hard constraints through formal modeling and logical reasoning, offering irreplaceable advantages in rule consistency and constraint verifiability, but struggle to directly obtain optimal solutions in complex, multi-objective simulation scenarios. Methods based on reinforcement learning continuously optimize through simulation feedback, enabling continuous improvement of simulation scheme performance in multi-objective scenarios, gradually approaching the comprehensive optimum, but lacking a rule constraint guarantee mechanism, they may generate action sequences that violate basic rule constraints during the exploration process.

[0004] In summary, existing technologies that loosely combine Large Language Model (LLM), planners, and reinforcement learning (RL) lack an end-to-end automated process and an inherent deep integration mechanism. This results in the system being unable to simultaneously and efficiently meet the key requirements in generating military combat simulation schemes, leading to low feasibility and rule compliance of the simulation schemes, and consequently, low reference value of the simulation schemes. Summary of the Invention

[0005] To address the issues of low feasibility and regulatory compliance in existing simulation schemes, this invention provides a method, system, equipment, and medium for generating military combat simulation schemes.

[0006] A first aspect of this invention provides a method for generating a military combat simulation scheme, comprising the following steps: Operational missions requiring the acquisition of natural language descriptions; The combat mission described in the natural language is semantically parsed according to the preset large language model to generate structured mission data containing mission objectives, entities, and constraints. Based on the structured task data and the pre-set operational domain knowledge base, a problem description file conforming to the Planning Domain Definition Language (PDDL) specification is generated; the problem description file and the pre-defined PDDL domain definition file are input into the planner for solving, and a feasible action sequence that satisfies all preset hard constraints is output. The feasible action sequence is used as the initial strategy, executed and evaluated in a simulation environment, and the feasible action sequence is iteratively optimized using a reinforcement learning algorithm based on the evaluation results to generate an optimized action sequence; the optimized action sequence is then used as the optimal combat simulation scheme.

[0007] Furthermore, the structured task data is organized in JSON format and contains at least one element among task objectives, execution entities, resources, and spatiotemporal constraints.

[0008] Furthermore, the problem description file is generated by mapping entities in the structured task data to PDDL objects and mapping task objectives and constraints to initial and target state predicates in PDDL.

[0009] Furthermore, the PDDL domain definition file is pre-written and defines the actions, preconditions, effects, and resource constraints of military operations.

[0010] Furthermore, the reinforcement learning algorithm employs a proximal strategy to optimize the PPO algorithm.

[0011] Furthermore, the step of iteratively optimizing the action sequence using a reinforcement learning algorithm based on the evaluation results specifically includes: adjusting the current action sequence to generate candidate sequences in each iteration; inputting the candidate sequences into a simulation platform for execution and obtaining rewards; updating the policy network based on the rewards to generate new action sequences; and obtaining optimized action sequences after multiple iterations until a preset convergence condition is met. The evaluation results include at least one indicator such as task completion rate, resource consumption, or execution time.

[0012] Furthermore, the preset convergence conditions include at least one of the following: the change in the scheme evaluation reward is lower than a first preset threshold in multiple consecutive iterations; the magnitude of the strategy update is lower than a second preset threshold; and the number of iterations reaches a preset upper limit.

[0013] A second aspect of the present invention provides a military combat simulation scheme generation system, comprising: The acquisition module is used to acquire combat mission descriptions in natural language. The parsing module is used to perform semantic parsing on the combat mission described in the natural language according to the preset large language model, and generate structured mission data containing mission objectives, entities and constraints. The solution module is used to generate a problem description file that conforms to the Planning Domain Definition Language (PDDL) specification based on the structured task data and the pre-defined operational domain knowledge base; input the problem description file and the pre-defined PDDL domain definition file into the planner for solution, and output a feasible action sequence that satisfies all preset hard constraints; The optimization module is used to execute and evaluate the feasible action sequence as the initial strategy in the simulation environment, and to iteratively optimize the feasible action sequence using a reinforcement learning algorithm based on the evaluation results to generate an optimized action sequence; the optimized action sequence is then used as the optimal combat simulation scheme.

[0014] A third aspect of the present invention provides an electronic device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the above-described method.

[0015] A fourth aspect of the present invention provides a readable storage medium storing a computer program that, when executed by a processor, implements the steps of the above-described method.

[0016] The military combat simulation scheme generation method provided by this invention has the following beneficial effects: This invention achieves natural language task understanding and automatic modeling of PDDL problems through a large language model. It ensures that the simulation scheme meets the rule constraints and obtains a feasible solution through a PDDL planner. Then, it optimizes the performance based on the feasible solution through reinforcement learning to obtain the optimal solution. This not only ensures the basic feasibility and rule compliance of the simulation scheme, but also realizes the intelligent optimization of the scheme performance, thereby significantly improving the automation level, simulation quality and engineering practicality of military combat simulation scheme generation.

[0017] Leveraging the deep semantic understanding capabilities of large language models, commanders and planners are allowed to describe operational intentions, constraints, and scenarios using the most natural language. This significantly improves the response speed and automation level of operational plan generation. The design automatically generates PDDL problem files based on structured data and combines them with predefined domain files, then uses a dedicated planner to solve and output feasible action sequences. By introducing the PDDL planner—a module strictly based on logical reasoning and constraint satisfaction—the intentions understood by the large language model are forcibly transformed into an action sequence that conforms to tactical rules, physical laws, and hard constraints of equipment capabilities pre-coded by military experts. This ensures that regardless of subsequent optimizations, the underlying logic of the plan remains legal, compliant, and executable, providing crucial credibility and security guarantees for AI-generated operational plans. The plan specifies that the feasible action sequences generated by the planner serve as the starting point, and iterative optimization is performed in a simulation platform using reinforcement learning. Starting with this safe and feasible sequence, the reinforcement learning algorithm, guided by simulation feedback, autonomously discovers and achieves a better solution that balances multiple objectives such as mission completion rate, resource consumption, and time consumption, while absolutely guaranteeing the tactical feasibility of the plan. Attached Figure Description

[0018] To more clearly illustrate the embodiments and design schemes of the present invention, the accompanying drawings required for this embodiment will be briefly described below. The drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a flowchart illustrating the method for generating a military combat simulation scheme according to an exemplary embodiment of the present invention. Detailed Implementation

[0020] To enable those skilled in the art to better understand and implement the technical solutions of the present invention, the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. The following embodiments are only used to more clearly illustrate the technical solutions of the present invention and should not be construed as limiting the scope of protection of the present invention.

[0021] The invention aims to address key issues in existing military combat simulation scheme generation technologies, such as insufficient rule constraint guarantees, limited performance optimization capabilities, and the difficulty in balancing rule constraints and performance optimization. Specifically, existing technical solutions mainly suffer from the following shortcomings:

[0022] (1) Although simulation scheme generation methods based on large language models can understand natural language task descriptions and generate action plans, the lack of a strict rule constraint verification mechanism may lead to simulation schemes that violate basic physical limitations, tactical rules, or equipment capability constraints, resulting in unexecutable schemes or serious safety hazards. For example, large language models may generate simulation schemes that do not conform to actual constraints, such as "unrefueled aircraft directly execute long-range attack missions" or "attack commands beyond weapon range." These schemes will inevitably fail in simulation or actual execution.

[0023] (2) Although the simulation scheme generation method based on PDDL planner can ensure that the generated simulation scheme meets all hard constraints through rigorous logical reasoning and constraint verification, and guarantee the legality and feasibility of the scheme at the tactical, physical and equipment levels, its rule-based search characteristics make it difficult to directly obtain the optimal performance solution in complex multi-objective scenarios. PDDL planner is better at quickly locating schemes that meet the constraints in the feasible solution space, but its ability to comprehensively optimize in multiple dimensions such as task completion rate, resource consumption and execution efficiency is limited, and it can often only generate "feasible but not optimal" simulation schemes.

[0024] (3) Although the simulation scheme generation method based on reinforcement learning can continuously optimize the scheme performance through simulation feedback and gradually approach the multi-dimensional comprehensive optimal solution, due to the lack of a rule constraint guarantee mechanism, the reinforcement learning algorithm may generate simulation schemes that violate basic rule constraints during the exploration and optimization process. If only reinforcement learning is used to directly generate simulation schemes, the algorithm may ignore physical limitations, tactical rules or equipment capability constraints in order to maximize rewards, resulting in simulation schemes that are infeasible at the rule level and cannot be executed in actual combat or simulation environments.

[0025] (4) Existing technologies lack an effective mechanism to organically combine rule constraints with performance optimization. They cannot achieve intelligent optimization of the simulation scheme's performance while ensuring the basic feasibility and rule compliance of the simulation scheme. This makes it difficult to meet the dual requirements of modern military operations for the quality and execution efficiency of the simulation scheme.

[0026] To address the aforementioned technical shortcomings, this invention proposes a method for generating military combat simulation schemes that integrates large language models, PDDL planners, and reinforcement learning techniques. The PDDL planner ensures that the simulation scheme meets the rule constraints to obtain a feasible solution. Then, reinforcement learning is used to optimize the performance of the feasible solution to obtain the optimal solution. This method not only guarantees the basic feasibility and rule compliance of the simulation scheme but also achieves intelligent optimization of the scheme's performance, thereby significantly improving the automation level, decision quality, and engineering practicality of military combat simulation scheme generation.

[0027] The technical solutions provided by the various embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0028] First, such as Figure 1 As shown, the method specifically includes the following steps: S1. Receive input data containing a natural language description of the combat mission.

[0029] Among them, the natural language description of combat missions includes scenario scenarios that conform to the standards of military simulation and deduction platforms and combat objectives based on the scenario scenarios.

[0030] S2. The input data is fed into a large language model for semantic parsing and structuring, and the output is structured task data. The structured task data includes the task objective, execution unit, spatiotemporal elements, resources and constraints extracted from the natural language description.

[0031] S3. Based on the structured task data, automatically generate a PDDL problem description file that conforms to the planning domain definition language.

[0032] S4. Input the PDDL problem description file and the predefined PDDL domain definition file into the PDDL planner for solution, and output the feasible action sequence that satisfies all hard constraints.

[0033] S5. Input the feasible action sequence into the simulation platform for simulation and execution, collect simulation process data for multi-dimensional quantitative evaluation, and output the evaluation result containing at least one performance indicator.

[0034] S6. Using the feasible action sequence as the initial strategy, construct a reward signal based on the evaluation result, and use a reinforcement learning algorithm to perform multiple rounds of iterative optimization in a simulation environment to generate the optimized action sequence.

[0035] S7. Output the optimized action sequence generated by the reinforcement learning as the final combat simulation scheme.

[0036] Based on the above inventive concept, the present invention proposes an embodiment, including the following steps: S1. System Input and Task Acquisition.

[0037] S1.1 Multi-source data acquisition and task input specifications: This system supports multiple input sources to drive the automatic generation of subsequent combat simulation plans. The input content covers the following three core data categories:

[0038] (1) Textualized combat objectives: Combat mission intentions, combat objectives, combat constraints and expected effects described by users or combat commanders in natural language, which are used to drive the large language model to perform semantic parsing and task extraction.

[0039] (2) Operational simulation scenario: includes structured or semi-structured scenario information such as the combat area, the force composition of the two sides, the initial situation, geographical and environmental conditions, electromagnetic situation, and the division of mission phases, which describes the basic situation background for generating the operational simulation plan.

[0040] (3) Structured simulation scripts for simulation platforms: Standardized structured script files for specific simulation platforms, used to define the entity objects, behavior models, weapon systems, state variables and execution processes required by the platform, providing directly runnable programmatic inputs for final simulation verification and scheme evaluation.

[0041] The system can automatically acquire the aforementioned data through manual input, file import, API synchronization, and database queries, and integrate it in a unified structure according to different sources. All inputs are formatted according to the standardized description specifications of combat simulation, ensuring that they can be directly used for mission semantic parsing, PDDL problem modeling, planning and solving, and subsequent simulation execution processes. This input mechanism is fully adaptable to the simulation scenario requirements of multiple tasks, multiple regions, multiple branches of service, and multiple spatiotemporal scales, providing a high-quality data foundation for the subsequent generation of executable, predictable, and optimizable combat simulation schemes.

[0042] S1.2 Unified scene template and hierarchical modeling of simulation scenario elements: The system adopts a unified and scalable structured template for combat simulation scenarios, and performs layered modeling of the key elements required for generating combat simulation schemes. The template covers all entities and constraints involved in natural language combat objectives, simulation scenario inputs, and structured scripts of the simulation platform, and can provide hierarchical descriptions of the following: combat spatiotemporal framework, tactical and operational objectives, force and resource deployment, platform and weapon characteristics, environmental conditions, electromagnetic and reconnaissance situation, rules of operation, and possible contingencies.

[0043] This template supports flexible expression of multi-objective, multi-stage, and multi-level event chains, covering different granularity requirements from tactical-level action steps to campaign-level operational processes. Through structured and hierarchical descriptions, the system can establish a stable mapping between natural language input and executable descriptions on the simulation platform, facilitating semantic parsing, automatic feature extraction, and task decomposition by large language models.

[0044] S1.3 Natural Language Input Parsing and Automatic Element Extraction: To adapt to diverse inputs in real-world application scenarios, the system supports users inputting mission intent, plan summaries, or direct combat instructions in Chinese natural language. Built-in professional-grade models for word segmentation, named entity recognition (NER), event extraction, and spatiotemporal normalization can automatically parse unstructured text into standard mission elements (such as mission objectives, action units, time limits, resources, and specific constraints). For example, the statement "The Blue Force must capture High Ground B within 12 hours while ensuring its own losses do not exceed 5%" can be easily split into a standard field structure, improving the system's engineering adaptability and automation level for natural language input.

[0045] S1.4 Unique Task Identifier and Full Lifecycle Traceability: Each task input is automatically generated with a unique identifier (ID) upon acceptance. The system binds this ID to multiple metadata items, including the original task content, input time, source channel, operation records, and preprocessing logs, enabling traceability and complete process management throughout the task's lifecycle. All input operations are automatically logged and archived in system logs and data metadata tables, supporting rapid task retrieval, accountability, compliance auditing, and historical task review.

[0046] S1.5 Input integrity verification and intelligent completion mechanism: During the task reception phase, the system automatically verifies the completeness and compliance of all key fields, including parameters such as task objectives, time, space, units, equipment, and constraints. For missing, ambiguous, or non-standard inputs, the system uses an automatic error correction mechanism to clearly identify the fields that need to be added and provide reasonable correction suggestions. Simultaneously, the system can integrate knowledge base and historical case data to intelligently infer and automatically fill in missing parameters, greatly improving the standardization and efficiency of the task input process and laying a solid data foundation for downstream automated generation and intelligent planning.

[0047] S2. Input standardization and preprocessing.

[0048] S2.1 Semantic Segmentation and Automatic Feature Extraction: The system uniformly processes all received combat scenario and mission input data into a semantic segmentation and feature extraction process. It integrates multi-level word segmentation, named entity recognition (NER), word normalization, and entity relation extraction algorithms to accurately parse key fields in the input text, including: combat time, mission location, participating units, equipment models, and resource constraints. For common military abbreviations, aliases, and synonyms, the system automatically performs normalization, entity mapping, and label localization based on a high-frequency terminology database and standardized libraries. For example, recognizing "052D" can automatically expand it to a destroyer category. Spatially ambiguous descriptions can be semantically inferred by combining mission context and historical examples, improving the consistency of scenario understanding and the accuracy of feature extraction.

[0049] S2.2 Automatic error correction and ambiguity resolution: To ensure data structure and semantic consistency, the system is equipped with a multi-level error correction and ambiguity resolution module that integrates a rule engine and a machine learning model. For singular or ambiguous time, geographical, and other information, rule-based normalization is employed (e.g., "midnight" is automatically mapped to "00:00"). Geographical descriptions are precisely reconstructed using scene and simulation databases. For missing, divergent, or uncertain fields, the system can automatically complete them by referring to task history, ontology knowledge base, and expert rules. If necessary, highlighted annotations require manual correction, improving the robustness of subsequent processing chains and preventing downstream decision-making biases caused by ambiguity or omissions.

[0050] S2.3 Data cleaning and conflict consistency detection: After parsing the structured elements, the system automatically initiates data cleaning and multi-version consistency checks. Through incremental log chains, partition mapping tables, and historical change comparisons, it tracks each round of input behavior, avoiding redundancy and conflicts arising from multi-channel collection, parallel editing, or version iteration. The system has real-time alerts and conflict correction mechanisms; all suspected anomalies are automatically prompted for verification, ensuring the uniqueness, consistency, and traceability of the dataset, laying a solid foundation for high-quality data inflow into the model and planner.

[0051] S2.4 Standardized data entry and full traceability: Data that has undergone multi-level normalization and cleaning is uniformly archived into the core data warehouse in a one-to-one mapping format between the original input and the structured results. The entire standardized data entry process is accompanied by detailed processing logs and mapping relationship tables, supporting data restoration and traceability at any stage. All standardized data will serve as the unified input for subsequent LLM understanding, PDDL modeling, and simulation processes, providing authoritative, unified, and high-quality basic data support for intelligent solution generation, end-to-end closed-loop optimization, and historical case accumulation.

[0052] Task parsing and structuring driven by S3 and LLM.

[0053] S3.1 Task semantic understanding and key information extraction: The standardized input data first enters the Large Language Model (LLM) parsing module. This module performs structured analysis and semantic induction of the task data based on multi-round deep semantic parsing and military scenario-specific fine-tuning weights. LLM automatically identifies and extracts specialized fields such as combat mission objectives, executing units, spatiotemporal elements, equipment resources, constraints, and environmental parameters through contextual understanding, knowledge transfer, and prompting engineering mechanisms. To further ensure the consistency and accuracy of the system's understanding, the model integrates a military terminology ontology, a key entity dictionary, and prior knowledge of actions / scenarios, significantly enhancing its ability to parse complex sentences, abbreviations, and domain-specific terms.

[0054] S3.2, Structured Field Population and Multi-Level Data Object Generation: After semantic parsing, the system automatically populates a structured task object (Task_JSON) using the task template. The content covers the task name, objective, primary / secondary participants, geographical region, timeline, resource allocation, key risks, and external conditions. This process is compatible with scenario modeling of multi-stage and multi-level operational objectives, supporting the segmented filling of sub-tasks, action phases, and operational links. For incomplete or ambiguous fields, the model will automatically invoke historical tasks for completion, expert knowledge inference, or mark them as "to be completed" according to priority, leaving room for subsequent model optimization and manual intervention, ensuring the traceability and high fidelity of task data.

[0055] S3.3 Multi-source task completion and intelligent disambiguation: LLM parsing supports multi-source completion strategies. Addressing common issues such as ambiguous expressions, units with the same name, unclear locations, and ambiguous actions in the input, the system, in addition to real-time contextual hints and domain limitations (e.g., "Red Force Main Force," "Dense Forest Area"), automatically infers results by integrating historical cases, domain expert knowledge, and statistical rules. Furthermore, it specifically records the reference paths, reasoning logic, and comparison examples for each completion process in a structured object. This mechanism effectively improves the model's adaptability to the variability and high complexity of real-world combat environments, giving the system strong adaptability and customizability.

[0056] S3.4. Archiving and Output of Structured Results: The fully parsed and completed Task_JSON objects, complete with field descriptions, data reference logs, and completion markers, are automatically archived into the central task data warehouse, ensuring traceability throughout the entire process from upstream input to structured output. All structured results are pushed to the PDDL automatic modeling module and subsequent planning, simulation, and optimization stages according to system interface specifications, guaranteeing no loss of technical information, full field coverage, and semantic closure. This lays the foundation for the generation of highly automated and reliable operational plans and multiple rounds of self-learning optimization.

[0057] S4. From structured tasks to PDDL problem modeling.

[0058] S4.1 Domain Knowledge Base and Domain Scenario Management: This system pre-constructs a PDDL domain knowledge base, covering various combat categories, equipment types, and rule constraints, maintained by military experts in conjunction with typical scenarios and standard planning specifications. Each Domain file in the knowledge base corresponds to a major combat category or mission scenario, containing object categories, action operators, equipment parameters, state predicates, and typical constraint formulas. The system automatically retrieves and schedules the most suitable Domain template based on scenario tags or mission characteristics, achieving precise mapping between the knowledge system and specific simulation scenarios, ensuring the legitimacy and high fidelity of subsequent planning calculations.

[0059] S4.2 Automatic conversion of structured task data to Problem files: The parsed and standardized Task_JSON data is mapped to a PDDL problem file via an automatic transformation engine guided by a large language model. The module performs structured decomposition on each task element, corresponding to appropriate PDDL entries. Utilizing an extensible set of mapping rules, the system automatically adapts and populates task fields to objects, states, resources, and goals, recursively completing multi-layered dependency elements. The transformation process incorporates built-in syntax and constraint checks to ensure that the generated problem file strictly conforms to PDDL specifications and can be directly used by the planner.

[0060] S4.3, Multi-Goal, Multi-Subtask and Complex Constraint Automatic Injection: Faced with complex combat missions involving multiple objectives and phases, the system dynamically analyzes mission hierarchy, logical dependencies, and priorities. The transformation engine automatically identifies and breaks down the main objective and sub-tasks, mapping them to PDDL and appropriately combining set constraint structures such as "and / or / sequence / within" to achieve joint scheduling of multiple objectives and complex constraints such as timing and resource limits. Key controlled fields are automatically mapped and injected into the init, goal, and constraints sections of the problem file by Task_JSON, fully restoring the realistic and complex constraint space of simulation decisions and improving the practical adaptability of military planning.

[0061] S4.4 PDDL file consistency and interoperability verification: After the Problem file is generated, the system automatically invokes the consistency and interoperability verification mechanism to check its compatibility with the data references, object mappings, actions, and predicates in the current scenario's Domain file. The system checks key points such as variable definitions, goal solvability, and state achievement paths, and automatically backtracks the data parsing process or triggers manual intervention for abnormal situations (such as unreachable goals, type mismatches, null object references, etc.). All verification steps include detailed error messages and logs to ensure that the downstream planning and simulation processes can be executed stably and without ambiguity.

[0062] S4.5 Archive Mapping and Full-Process Traceability: Each call to the Domain / Problem file generation and binding process is automatically archived and traced, recording file version, associated task ID, structured input, mapping parameters, and processing logs. All intermediate data and templates can be historically reviewed, supporting task reproduction, solution status restoration, abnormal path analysis, and rollback of expert intervention scenarios. This mechanism maximizes the traceability of the simulation / decision-making process, the ability to extrapolate solutions, and the reuse of results, forming the core foundation for compliance and engineering reliability.

[0063] S5, the PDDL planner generates feasible action sequences.

[0064] S5.1 Planner Scheduling and Intelligent Loading of PDDL Files: The system integrates a high-performance PDDL automatic planning module, supporting seamless integration with various mainstream planners. Upon receiving the problem.pddl file and Domain configuration, the scheduling center automatically assigns the most suitable planner engine based on the task type and user requirements. The loading process incorporates automatic detection of file consistency, syntax compliance, and model compatibility, and preset search options (such as heuristic functions, state space pruning, time or resource priorities) to ensure maximum compatibility between automatic configuration and the model. All processes implement log registration and anomaly alerts before loading and invocation, effectively preventing solution failures in complex scenarios.

[0065] S5.2 Efficient Feasibility Search and Complex Constraint Solving: The planning engine drives global solution based on PDDL descriptions. Under multi-objective, multi-level constraint problems, it employs a hybrid algorithm combining hierarchical heuristics, priority queues, and dynamic resource allocation to quickly filter and produce feasible solutions that meet military constraints. The system supports setting optimal solution metrics based on different tasks, such as execution time, action level depth, resource consumption, and risk probability. For specific soft / hard constraints, the planner has a built-in constraint consistency filtering process to ensure that any output solution meets operational, physical, and equipment requirements, and to report potential conflicts or unsolvable states. The entire search process can output real-time progress and key nodes, facilitating operation and maintenance monitoring and interactive optimization.

[0066] S5.3, Structured Archiving and Standard Output of Action Sequences: Each action plan output by the planner is structured and standardized, detailing information such as action name, participating parties, parameters, time nodes, and resource usage, and is converted into a data format suitable for subsequent simulation platforms. All output plans are automatically labeled with task numbers and timestamps, associated with input parameters, planner algorithms, and optimization configurations, and the complete chain is archived in the system task and decision database.

[0067] S6 Action Plan Simulation and Evaluation Feedback.

[0068] S6.1 Scheme format adaptation and automatic integration with simulation platform: After completing the PDDL planner output scheme, the system first converts the action sequence into a standard input format supported by the simulation platform through the plan adapter. For different simulation platforms, the system provides a plug-in customization adaptation mechanism to ensure that various platforms can efficiently and seamlessly receive the optimal combat plan, achieving automatic and integrated connection between the simulation scenario and system decision-making.

[0069] S6.2 Simulation Environment Initialization and Automated Execution: After the scheme format is adapted, the system pushes the unified action plan to the integrated simulation engine. Once the simulation platform starts, it automatically loads the relevant scenario environment, equipment parameter models, initial troop status, and operational constraints, ensuring strict synchronization between the virtual battlefield environment and the original input. Based on the mission action sequence, the simulation engine drives tactical actions in real-time calculations, dynamically reconstructing the battlefield evolution process. All operations are triggered according to preset rules, supporting the automated completion of complex military operation simulations.

[0070] S6.3, Full-process data acquisition and event monitoring: During simulation execution, the system monitors multiple key data points in real time, including unit location, equipment dynamics, mission progress, and changes in battlefield situation. A structured log collection mechanism combining timed sampling and event-driven methods is employed to accurately record every operation, state transition, and result feedback. All process data and raw simulation logs are standardized and stored in a database, supporting multi-dimensional traceability by task, plan, stage, and unit, laying a data foundation for refined evaluation and subsequent optimization.

[0071] S6.4 Simulation Multidimensional Evaluation and Quantitative Analysis: After the simulation is completed, the evaluation module automatically retrieves data from the entire execution process to perform multi-dimensional and quantitative analysis of the implementation effect of the plan. The system has a rich set of built-in evaluation dimensions, such as mission objective achievement rate, battle damage analysis, resource consumption, key node events, mission time consumption, and combat coordination index, and outputs a comprehensive score and detailed evaluation report. All indicators can be customized with weights according to needs, enabling targeted and multi-angle simulation effectiveness evaluation and significantly improving the ability to distinguish the merits of the plan.

[0072] S7. Optimization of reinforcement learning-based schemes.

[0073] S7.1, Definition of Reward Signals and Construction of Action Space: Based on the multi-dimensional performance indicators output by simulation evaluation, the system constructs a weighted reward mechanism adapted to the gradient optimization of PPO strategies. Under different operational objectives and application scenarios, the weights of each reward item can be flexibly customized, enabling the optimization process to achieve a dynamic balance between stability and strategy improvement. The action editing space remains highly open, encompassing multi-granularity controllable operations such as adding and deleting plan sequence-level actions, adjusting their order, and fine-tuning parameters, to meet the requirements of iterative, corrective, and locally refined optimization in military decision-making scenarios. This action space further serves as the foundation for the generation and evaluation of PPO strategy network actions, providing an operational entry point for subsequent continuous optimization.

[0074] S7.2, Enhanced Self-Learning Optimization Process: This system employs a closed-loop PPO (Policy Point of Interest) reinforcement self-learning mechanism: based on the current action sequence, the system generates multiple rounds of candidate adjustment schemes (such as merging steps, inserting redundant actions, adjusting resource allocation, and rearranging action order), and inputs these as policy sampling trajectories into the simulation engine. The key performance indicators obtained from the simulation are mapped to the PPO advantage estimate through a reward function, which drives the policy network update. Each fine-tuning action is immediately fed into the next round of simulation, ensuring a continuous balance between "constraint changes," "policy stability," and "improvement magnitude" in the PPO, guiding the system to automatically approach operational action schemes with higher returns and lower risks.

[0075] S7.3 Optimize the convergence criterion and link iterative simulation: To ensure the engineering controllability of the PPO strategy optimization process, the system automatically monitors the optimization results of each round and sets multi-dimensional criteria based on strategy convergence characteristics, including the improvement of the objective function, the stability of reward changes, the range of KL divergence constraints, and the extreme value of resource consumption. If the advantage improvement of consecutive rounds of strategy updates is lower than the threshold, or the KL deviation exceeds the PPO limit range, the system automatically determines that the optimization has reached the optimal or local optimum state. All scheme generation, simulation reproduction, and evaluation comparison are versioned and recorded to facilitate subsequent traceability analysis and scheme feasibility verification.

[0076] S7.4 Result Archiving and Knowledge Feedback Mechanism: During the optimization process, all intermediate solutions, policy network outputs, simulation evaluations, and performance data are archived simultaneously. The final optimal solution and its corresponding key parameters for the PPO policy network are encoded into the action knowledge base, combined with task tags, historical environments, and improvement trajectories, for rapid reuse, experience transfer, and templated application in future tasks. Typical PPO optimization paths are automatically transformed into domain knowledge and engineering cases, providing model feedback for the continuous operation of the system. If policy non-convergence or reward distortion occurs during iteration, the system will automatically roll back to a verified feasible solution version or trigger a global replanning to ensure stable and reliable output results.

[0077] S8. Solution closed-loop output and knowledge archiving.

[0078] S8.1 Standard Output and Archiving Management of Optimal Action Plan: After the PPO optimization strategy converges, the system automatically selects the action sequence with the best comprehensive indicators and numbers, labels, and archives it according to a unified structured format. Each optimal solution is accompanied by metadata such as generation time, task attributes, resource constraints, indicator scores, and PPO convergence status, facilitating rapid deployment and automatic invocation later. All output solutions and version history are encrypted and stored, and support one-click rollback to ensure data security and engineering consistency.

[0079] S8.2, System collection of full-process data and intermediate results: The system performs versioned data aggregation on the complete operational mission processing chain, including raw inputs, LLM parsing structures, PDDL issues and Domain files, planning results, simulation execution logs, and policy updates and benefit trajectories for each round of PPO. All node state transitions and scheme evolution histories are automatically recorded, providing a data foundation for intelligent analysis, result reproduction, performance comparison, and engineering strategy optimization.

[0080] S8.3 Construction and Intelligent Scheduling of Historical Knowledge Case Base: Each closed-loop output of a solution and its corresponding simulation feedback are incorporated into the case library as "knowledge units." Case management supports solution indicator feature abstraction, tagging system, similarity retrieval, and parameter transfer and reuse. Combining AI-driven knowledge clustering, template reasoning, and PPO strategy feature matching mechanisms, the system can quickly deduce and schedule historical high-quality solutions, significantly improving response speed and intelligence under complex tasks.

[0081] S8.4 Anomaly detection, manual evaluation, and closed-loop feedback correction: For abnormal or low-performance solutions, the system automatically triggers detection and alarm mechanisms to locate fault points, process conflicts, and evaluation deviations. It supports expert annotation and collaborates with the PPO replanning engine to reconstruct and correct parameters for failed solutions. All abnormal data is archived to enrich the knowledge base, providing new reinforcement samples for subsequent model training, achieving continuous enhancement and steady-state evolution of system capabilities.

[0082] Based on the above inventive concept, the present invention also provides a military combat simulation scheme generation system, comprising: The acquisition module is used to acquire combat mission descriptions in natural language.

[0083] The parsing module is used to perform semantic parsing of combat missions described in natural language based on a preset large language model, and generate structured mission data containing mission objectives, entities, and constraints.

[0084] The solver module generates a problem description file that conforms to the Planning Domain Definition Language (PDDL) specification based on structured mission data and a pre-defined operational domain knowledge base. The problem description file and the pre-defined PDDL domain definition file are then input into the planner for solving, and the output is a sequence of feasible actions that satisfies all preset hard constraints.

[0085] The optimization module is used to execute and evaluate feasible action sequences as initial strategies in a simulation environment. Based on the evaluation results, it uses a reinforcement learning algorithm to iteratively optimize the feasible action sequences and generate optimized action sequences. The optimized action sequences are then used as the optimal combat simulation scheme.

[0086] The present invention also provides a computer-readable storage medium storing a computer program that can be used to execute the above-described... Figure 1 The steps of the provided method for generating military combat simulation schemes.

[0087] This invention also provides a computer device. At the hardware level, the computer device includes a processor, an internal bus, a network interface, memory, and non-volatile memory, and may also include other hardware required for various operations. The processor reads the corresponding computer program from the non-volatile memory into memory and then executes it to achieve the above-mentioned functions. Figure 1 The steps of the provided method for generating military combat simulation schemes.

[0088] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0089] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, as well as combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0090] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0091] These computer program instructions can also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0092] It should be noted that the specific embodiments described above enable those skilled in the art to more fully understand the present invention, but do not limit the present invention in any way. Therefore, although the present invention has been described in detail in this specification, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the present invention; and all technical solutions and improvements that do not depart from the spirit and scope of the present invention are covered within the protection scope of the patent of the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. A method for generating a military campaign simulation scenario, characterized in that, The method includes the following steps: Operational missions involving the acquisition of natural language descriptions; The combat mission described in the natural language is semantically parsed according to the preset large language model to generate structured mission data containing mission objectives, entities, and constraints. Based on the structured task data and the pre-set operational domain knowledge base, a problem description file conforming to the Planning Domain Definition Language (PDDL) specification is generated; the problem description file and the pre-defined PDDL domain definition file are input into the planner for solving, and a feasible action sequence that satisfies all preset hard constraints is output. The feasible action sequence is used as the initial strategy, executed and evaluated in a simulation environment, and the feasible action sequence is iteratively optimized using a reinforcement learning algorithm based on the evaluation results to generate an optimized action sequence; the optimized action sequence is then used as the optimal combat simulation scheme.

2. The method of claim 1, wherein, The structured task data is organized in JSON format and includes at least one element among task objectives, execution entities, resources, and spatiotemporal constraints.

3. The method of claim 2, wherein, The problem description file is generated by mapping entities in the structured task data to PDDL objects and mapping task objectives and constraints to PDDL initial and target state predicates.

4. The method of claim 1, wherein, The PDDL domain definition file is pre-written and defines the actions, preconditions, effects, and resource constraints of military operations.

5. The method of claim 1, wherein, The reinforcement learning algorithm employs a proximal strategy to optimize the PPO algorithm.

6. The method of claim 1, wherein, The step of iteratively optimizing the action sequence using a reinforcement learning algorithm based on the evaluation results specifically includes: adjusting the current action sequence to generate candidate sequences in each iteration; inputting the candidate sequences into a simulation platform for execution and obtaining rewards; updating the policy network based on the rewards to generate new action sequences; and obtaining optimized action sequences after multiple iterations until a preset convergence condition is met. The evaluation results include at least one indicator such as task completion rate, resource consumption, or execution time.

7. The method of claim 1, wherein, The preset convergence conditions include at least one of the following: the change in the scheme evaluation reward is lower than a first preset threshold in multiple consecutive iterations; the magnitude of the strategy update is lower than a second preset threshold; and the number of iterations reaches a preset upper limit.

8. A military campaign simulation scenario generation system, characterized by, include: The acquisition module is used to acquire combat mission descriptions in natural language. The parsing module is used to perform semantic parsing on the combat mission described in the natural language according to the preset large language model, and generate structured mission data containing mission objectives, entities and constraints. The solution module is used to generate a problem description file that conforms to the Planning Domain Definition Language (PDDL) specification based on the structured task data and the pre-set operational domain knowledge base. The problem description file and the predefined PDDL domain definition file are input into the planner for solving, and the feasible action sequence that satisfies all preset hard constraints is output. The optimization module is used to execute and evaluate the feasible action sequence as the initial strategy in the simulation environment, and to iteratively optimize the feasible action sequence using a reinforcement learning algorithm based on the evaluation results to generate an optimized action sequence; the optimized action sequence is then used as the optimal combat simulation scheme.

9. A computer device, comprising: The method includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the method described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, implements the method described in any one of claims 1 to 7.

Images