Multi-agent strategic scenario generation system based on large language model

Abstract

The invention provides a multi-agent strategic scenario generation system based on a large language model, and the system comprises the following steps: carrying out the modeling of a scene and a system, and carrying out the system modeling of an attempt to record a case, deduce a situation, an event, a scenario, a strategic-battle-tactical-level red-blue party subsystem and a white party subsystem; the method comprises the following steps: initializing a red party attempt to record cases and a blue party attempt to record cases, when a user uses the system, prompting that relevant information of military deduction needs to be input, and after the information is input according to requirements, the system uses a global attempt record generation agent to receive the information input by the user and generate an attempt record in a JSON format. According to the multi-agent strategic scenario generation system based on the large language model provided by the invention, the strategy-battle-tactical multi-level agent vertical cooperation relationship-based game subsystem is designed, and the task granularity of scenarios designed by the large language model is refined through the multi-level game subsystem; and the scenario design quality is effectively improved.

Description

Technical Field

[0001] The invention belongs to the field of simulation and game deduction, and particularly relates to a multi-agent deduction scenario generation system based on a large language model. Background Art

[0002] A military scenario is a detailed description of a scenario. By creating a hypothetical situation, it simulates possible military actions and events. The content of military scenarios includes: attempt to file a case, basic scenario and supplementary scenario. The attempt to file a case is based on in-depth analysis and comprehensive assessment of multi-dimensional information such as the international strategic pattern, national security environment, and the balance of power between the enemy and us. It aims to clarify operational intentions, set campaign or combat objectives, and preliminarily plan the strategic path and key steps required to achieve the goals. The basic scenario is a product that is further refined and concretized on the basis of the attempt to file a case. It is a detailed operational plan formed through rigorous battlefield environment simulation, force deployment design and action sequence arrangement. The supplementary scenario is an important supplement and improvement to the basic scenario. It mainly predicts and provides plans for various variables that may arise in actual combat. This document is organized according to the chronological order of battlefield events, covering key elements such as combat time, combat area and specific events.

[0003] Strategy, campaign and tactics are three different levels of concepts in military theory, corresponding to different time and space scopes, command levels and action purposes. The strategic level usually refers to the overall planning and guidance of a war or conflict by a country or group over a longer period of time (such as several years to decades). It focuses on the overall, directional and long-term goal setting and resource allocation, involving the comprehensive use of multiple fields such as politics, economy, diplomacy, culture, and how to serve the realization of national strategic interests through military means. The campaign level is the organization and implementation of a series of combat activities in a shorter time span (such as several months to weeks) and in a specific area. The plan is the concretization and decomposition of the strategic goal, aiming to achieve a series of intermediate and local military goals to support the gradual realization of the strategic intention. The tactical level refers to the actual combat operations of the troops in specific battlefield spaces (such as battlefields, urban blocks, sea areas, etc.) and in a short period of time (usually a few hours to a few days) in actual combat. The tactical level focuses on specific combat skills and operating methods, including the planning and execution of details such as troop organization, firepower allocation, mobile routes, and attack methods. In summary, the three levels of strategy, campaign, and tactics constitute a complete planning and execution system for military operations from top to bottom, and jointly promote the development of war or conflict towards the established goals.

[0004] A large language model is a complex system built on artificial intelligence technology that is designed to understand, generate, and process natural language. This model is trained with large-scale data sets, so that it can learn and simulate complex patterns and structures in human language. Large language models have a wide range of applications, including but not limited to text generation, language translation, sentiment analysis, dialogue systems, question-answering systems, and text summarization. A notable feature of these models is their adaptability and flexibility, which can handle a variety of different types and styles of language tasks. With the improvement of computing power and the increase in data availability, large language models have shown significant effectiveness in processing complex language tasks.

[0005] The excellent content generation and natural language understanding capabilities of large language models make them of great application value in the field of military scenario design. However, the existing large language model technology still has the following problems when designing military scenarios:

[0006] Large language model alignment problem: When generating long texts, large language models often have logical inconsistencies. This problem is particularly fatal in military scenario design because military operations require strict logic and consistent strategies. For example, the model may describe the enemy and one’s own action plans in the same scenario, but these plans may conflict with each other in time and space.

[0007] Insufficient mission details. When generating military scenarios, large language models often lack specific mission details, including troop deployment, specific parameters of weapons and equipment, detailed steps for mission execution, etc. These details are crucial for military simulation and actual combat plans.

[0008] The text length is controlled. When existing large language models generate long texts, the text length is prone to out-of-control problems. The generated content may be too short and not meet actual needs. In military scenario design, it is necessary to ensure the amount of text generated in order to provide more sufficient battlefield information.

[0009] Therefore, it is necessary to provide a multi-agent strategic scenario generation system based on a large language model to solve the above technical problems. Summary of the invention

[0010] The present invention provides a multi-agent strategic scenario generation system based on a large language model, which solves the problems of agent decision-making limitations and imperfect feedback mechanism in traditional military simulation.

[0011] In order to solve the above technical problems, the present invention provides a multi-agent strategic scenario generation system based on a large language model, comprising the following steps:

[0012] S1. Modeling scenarios and systems requires systematic modeling of attempt filing, situation simulation, events and scenarios, strategic-campaign-tactical level red-blue subsystems and white-side subsystems;

[0013] S2. Initialize the red and blue team's attempt to file a case. When the user uses the system, he is prompted to enter the relevant information of the military deduction. After entering the information as required, the system uses the global attempt to file a case to generate an agent, receives the user input information, and generates an attempt to file a case in JSON format. The Markov subchain corresponding to this step is:

[0014] Input→IF→MX→AX

[0015] Among them, Input represents user input, IF represents attempt to file a case, MX represents basic military assumptions, and AX represents supplementary assumptions. This step mainly completes the transformation from user input to attempt to file a case in the system. After the attempt to file a case is generated, the python json toolkit is used to segment the attempt to file a case, and the attempt to file a case from different perspectives of the red and blue sides is obtained, which is used to generate the basic assumptions in the next stage;

[0016] S3. Based on the basic assumption of the attempt to file a case, the present invention uses a cyclic generation workflow of the deduction event-situation. First, each red and blue party subsystem converts its own attempt to file a case into a round situation, and then loops to generate the automatic confrontation workflow. In the loop, one party's troops convert the received deduction event into a dynamic situation; the round situation after the static situation and the dynamic situation are combined is converted into a round strategy; the round strategy is further converted into a round tactic; the round tactic is converted into a deduction event and transmitted to the other party's troops, and the cycle continues. After n cycles set by the user, the confrontation workflow ends;

[0017] S4. Supplementary scenarios are constructed based on basic scenarios. Supplementary scenarios are improvements on basic scenarios. They mainly predict different scenarios in actual combat and provide pre-selected plans. We construct supplementary scenarios through three forms: re-simulation, result tree, and man-in-the-loop. Finally, we combine the attempt to file a case, basic scenario, and supplementary scenario to output a complete military scenario.

[0018] Preferably, the S1 comprises the following steps:

[0019] S11. Modeling the attempt to file a case. The attempt to file a case is a pre-setting before the military scenario simulation, which includes the strategic layout with the country as the main body and the detailed description of the initial scene of the local war campaign. The attempt to file a case is modeled as a tuple including the strategic background and the campaign description. Its mathematical model is:

[0020] IF=(SC,CD)

[0021] Among them, IF stands for attempt to establish a case, and it includes an introduction to the strategic background SC and a focused campaign description CD;

[0022] S12. Modeling of deduction situation. Deduction situation refers to the overall situation obtained by analyzing and integrating the battle situation. It is an abstract concept extracted based on specific integration. In view of the needs of multi-faceted and complex situation analysis, this patent divides the situation into static situation, dynamic situation and round situation and models them separately;

[0023] S13. Modeling of simulation events and basic assumptions. The military assumptions designed in this patent are mainly event-driven. Its characteristic is that the simulation process is mainly promoted based on key events in the battlefield or strategic environment. The simulation events and basic assumptions are modeled in turn below;

[0024] S14. Construct a red-blue subsystem based on a vertical collaboration relationship of multi-level agents in strategy, campaign and tactics. This subsystem is responsible for playing the roles of the red and blue parties, and conducting battle deduction and game from the perspective of one party. It mainly includes: verification agent, situation summary agent, strategy generation agent, campaign generation agent and tactical generation agent;

[0025] S15. Construct the white side subsystem, which is mainly responsible for receiving global information and implementing corresponding functional modules from a global perspective. The subsystem mainly includes the attempt to file a case agent and the global event generation-verification agent.

[0026] Preferably, the S11 comprises the following steps:

[0027] S111. Modeling the strategic background. In military scenarios, the strategic background is intended to describe the macro-environmental conditions and strategic situational basis of the campaign or conflict setting, including the description of the regional security environment, the analysis of the causes of war, and the speculation of the opponent's intentions. The strategic background is modeled as follows:

[0028] SC=(RS,CC,AS)

[0029] Among them, SC is the strategic background, RS is the description of the regional security environment, CC is the analysis of the causes of war, and AS is the opponent's strategy;

[0030] S112. Modeling the campaign description. The campaign description is a summary of information at the campaign level. It requires a detailed description of the environmental information of the battle to be held in the scenario, the military armaments of both sides, and the military plan. It is modeled as follows:

[0031] CD=(MA,EI,MP)

[0032] Among them, CD represents campaign description, MA represents military armament, EI represents environmental information, and MP represents military plan. Military armament describes the basic military strength of a party in the military scenario process. The formula is:

[0033] MA=(MA branch ,MA force , M.A. equipment )

[0034] Among them, MA branch Indicates the type of soldier, MA force Indicates the armed forces in the region, MA equipment Indicates the equipment resource information of the party;

[0035] Environmental information EI describes the objective environment of the military scenario, and its formula is:

[0036] EI=(EI time ,EI zone , EI weather )

[0037] Among them, EI time Indicates the assumed time, EI zone Represents the hypothetical area, EI weather Indicates weather composition;

[0038] The military plan describes the macro-strategic measures of the entire force, and its formula is:

[0039] IS=(IS goal ,IS method ,IS principle )

[0040] Among them, IS goal Indicates combat objectives, IS method Indicates basic tactics, IS principle Indicates the principle of warfare;

[0041] After modeling the mission scenario, it is necessary to model the simulated situations and events in the military scenario.

[0042] Preferably, the S12 comprises the following steps:

[0043] S121. Model the static situation. The static situation is an overview of the information of both sides before the battle is played. It will not change with the battle. The mathematical definition of the static situation is:

[0044] MS static =(MA, EI, MP)

[0045] Among them, MS staticRepresents the static situation, which is an abstract summary extracted from the military armament MA, environmental information EI and military plan MP in the attempt to file a case IF;

[0046] S122. Model the dynamic situation. The dynamic situation is defined as a comprehensive evaluation of the current scenario based on the changing battle situation in the simulation, mainly including time, dynamic changes in military power and basic conditions. The dynamic situation is time-sensitive and variable. In this patent, the definition of the dynamic situation is:

[0047] MS dynamic =(E Time , M equipment-d , S dynamic )#(6)

[0048] Among them, MS dynamic Indicates dynamic situation, E Time Indicates the current time, M equipment-d Indicates the current round of military deployment, S dynamic Indicates the basic situation of the current battle;

[0049] S123, modeling the round situation, which is used to describe the overall battle situation in a round. In this patent, the formula is:

[0050]

[0051] Among them, MS Round Indicates the turn situation, MS static Indicates static situation, MS dynamic represents the dynamic situation, n represents the current round number, in the initialization stage (n=0), the round situation is the static situation, and as the scenario generation begins, the round situation becomes a combination of the static situation and the dynamic situation;

[0052] S124. Model the ending situation. The ending situation is an evaluation summary of the overall battle situation at the end of the simulation, the completion of the strategic goals of both sides, and the loss of confrontation forces after the military scenario is completed. The ending situation is defined as:

[0053] MS end =(SO, OC, FL)

[0054] Among them, MS end Indicates the end situation, SO indicates the overall battle situation, OC indicates the completion of strategic goals, and FL indicates the loss of opposing forces.

[0055] Preferably, the S14 comprises the following steps:

[0056] S131. Modeling of deduction events. Deduction events are the basic units of basic assumptions. They describe a certain action of a certain party and the changes in the subject and environment caused by the action. In summary, the definition of basic assumptions is:

[0057] ME=(ME entity ,ME action , M.E. task ,ME interaction )

[0058] Among them, ME is the deduction event, which is composed of entity ME entity , Action ME action , Mission ME task and interactive ME interaction composition;

[0059] S132. Model the basic scenario. The basic scenario is an imagined or pre-set war or conflict scenario. The basic scenario is generated by the set of all simulation events and the situation at the end of the simulation. Its mathematical definition is:

[0060]

[0061] MX represents the basic scenario, which includes the attempt to file a case IF, a set of events in n rounds And the end situation MS end .

[0062] Preferably, the S14 comprises the following steps:

[0063] S141. Construct a verification agent model. The present invention uses a generation-verification adversarial collaboration model, in which the generation agent is responsible for generating an initial solution, and the verification agent gives feedback on the initial solution and passes it to the generation agent for regeneration. When the number of modifications exceeds a certain number, the verification solution is output. In this patent, the verification agent function is defined as receiving a large model input and outputting corresponding modification suggestions. The mathematical model is:

[0064] V({Input}) llm →Advice

[0065] Among them, V({Input}) llm It is the model of the verification agent. The model input {Input} changes according to the usage scenario of the verification agent, and the corresponding prompt words will also change. Advice represents the modification suggestions for the verification agent output, which mainly includes: whether the model output maintains temporal consistency and spatial consistency; whether the model output conforms to the overall logic; whether the grammar, sentence structure and wording of the model output are appropriate;

[0066] S142. Construct a strategy generation agent. The main function of the strategy generation agent is to make an evaluative summary based on the intentions received by the troops and generate the army's macro strategy. The mathematical model is:

[0067] Z(IF) llm →MS static

[0068] Among them, Z(IF) llm represents the strategy generation agent, which takes the attempt to file a case IF as input and outputs a static situation;

[0069] S143. Construct a situation summary agent. The main function of the situation summary agent is to follow the continuous progress of the deduction events and dynamically summarize the battlefield. It takes the deduction events as input and outputs the dynamic situation. The mathematical model is:

[0070] T(ME) llm →MS dynamic

[0071] Among them, T(ME) llm is the situation summary agent, ME is the event deduction agent, MS dynamic It is a dynamic situation;

[0072] S144. Construct a campaign generation agent. The campaign generation agent is responsible for continuously adjusting the troops' campaign strategy according to the dynamic changes in the battlefield situation. Its implementation form is: to receive the round situation and the modification suggestions of the verification agent, generate the overall campaign strategy of the current round. The campaign agent is modeled as:

[0073] C(MS round ,Advice) llm →C round

[0074] Among them, C(MS round ,Ans) llm It is the working model of the campaign agent, which uses the turn-based MS round and verification agent advice as input, output round battle C round , turn-based battles mainly include: Military objectives: This dimension focuses on clear, quantifiable military objectives, such as destroying enemy forces, occupying key areas, or controlling important resources in strategic operations;

[0075] Resources and force allocation: involves the effective allocation of existing resources and personnel, involving decisions on force size, organizational structure, fire support, and logistics to achieve strategic objectives;

[0076] Time factors: Considering time constraints and urgency issues, some goals need to be achieved immediately, while others allow for longer-term pursuit;

[0077] Comparative analysis of enemy and friendly forces: Analyze the relative strengths, weaknesses, capabilities, and potential vulnerabilities of friendly and enemy forces to assess the feasibility of strategic objectives and their attendant risks;

[0078] S145. Construct a tactical generation agent. The function of the tactical generation agent is to receive the round battle situation and the modification suggestions of the verification agent, and generate the tactical strategy of the current round. The tactical agent is modeled as:

[0079] T(C round ,Advice) llm →T round

[0080] Among them, T(C round ,Advice) llm represents a tactical agent, which uses a turn-based strategy C round , modify the advice to input and output the round tactical strategy T round ;

[0081] Turn-based tactical strategies mainly include: combat maneuvers, which refer to specific actions and movements on the battlefield aimed at gaining tactical advantages;

[0082] Fires Employment: Effectively manage and employ firepower resources to support the conduct of tactical operations;

[0083] Terrain utilization: maximize the advantages of the battlefield's natural topographic features to serve the realization of tactical objectives;

[0084] Force coordination: Ensure seamless cooperation and coordination between various troop units and elements during tactical engagements.

[0085] Preferably, the S15 comprises the following steps:

[0086] S151. Construct an agent for generating an attempt to file a case. The main function of the agent for generating an attempt to file a case is to receive and verify user input and output an attempt to file a case. The model of the agent is:

[0087] I({Input}) llm →IF

[0088] Among them, I({Input}) llm It is the agent for generating the attempted case. It receives user input and outputs the JSON format of the attempted case. At the same time, when the user input information is incomplete, the model will feedback the missing information details, try to complete the gaps, and prompt the user to verify and modify;

[0089] S152. Construct a global event generation agent. The main function of the global event generation agent is to receive the turn strategies and tactics output by the red and blue subsystems, and consider the previous events to generate the latest event. The formula is:

[0090] E(C round , T round , {ME1, ME2...ME n-1}) llm →ME n

[0091] Among them, E(C round , T round ,{ME1,ME2...ME n-1}) llm is a global event generating agent, which receives C at the nth round. round , T round And a list of all previous events {ME1, ME2...ME n-1}, output event ME n ;

[0092] S153. Construct a global situation summary agent. The function of the global situation summary agent is to generate a summary evaluation of the battle situation after the game is over. It receives a list of all game events and the attempts of both parties to file a case, and generates an ending situation. The formula is:

[0093]

[0094] in, It is a digital model of the global situation summary agent, which attempts to file a case IF, all the deduction events Input, output end state MS end .

[0095] Preferably, S3 comprises the following steps:

[0096] S31. The Red and Blue subsystems generate a turn situation by attempting to file a case and deduce events. The Red and Blue subsystems use the strategy generation agent to generate a static situation based on the attempt to file a case. If the current round is the initial round, the turn situation is the static situation. If it is not the initial round, there is an event input from the opponent's troops. At this time, the situation summarizer is called to generate a dynamic situation based on the event, and the static situation and the dynamic situation are combined into a turn situation. The process of this step is as follows:

[0097] Z(IF) llm →MS static

[0098] T(ME) llm →MS dynamic

[0099]

[0100] Among them, the strategy generation agent Z(IF) llm Generate static situation MS based on attempt to file case IF static , situation generation agent T(ME) llm Generate dynamic situation MS based on simulated events ME dynamic , the turn situation is MS round , it is a dynamic situation MS dynamic and static situation MS static Under different conditions, different combinations

[0101] Through this step, the transition from attempting to file a case to a round situation is achieved;

[0102] S32, the Red and Blue Party subsystem generates a round campaign strategy through the round situation. The Red and Blue Party subsystem generates the round campaign strategy in the form of a campaign generation-verification agent loop workflow. First, the campaign generation agent takes the round situation as input and outputs the initial campaign strategy. The formula for this step is:

[0103] C(MS round ) llm →C1

[0104] Among them, C(MS round ) llm The strategy generating agent starts with MS round The round situation is input and the initial round strategy C1 is output;

[0105] In the next cycle, the verification agent takes the initial campaign strategy as input and outputs modification suggestions; the campaign initial agent takes the modification suggestions and the previous round of campaign strategy as input and outputs the modified campaign; the verification agent takes the modified campaign as input and outputs modification suggestions. After the number of rounds set by the user is looped, the final round of the campaign is output. The formula for this stage is:

[0106] V(C k-1 ) llm →Advice k

[0107] C(C k-1 , Advice m ) llm →C k →C round (k=m)

[0108] Among them, V(C k-1 ) llm →Advice kThe recursive formula indicates that the verification agent uses the strategy C of the previous round in the kth round of loops. k-1 Input, output current modification suggestions Advice k , C(C k-1 , Advice k ) llm It means that the verification agent takes the previous round of campaign strategy and modification suggestions as input, and outputs the modification strategy C for the Kth round of cycles k , when k = m, output the final round strategy C round ;

[0109] S33, the red and blue subsystem generates a round tactical strategy through a round campaign strategy. The red and blue subsystem generates a round tactical strategy in the form of a tactical generation-verification agent loop workflow. The workflow is basically the same as the previous step, and the formula is expressed as:

[0110] T(C round ) llm →T1

[0111] V(T k-1 ) → Advice k

[0112] T(T k-1 , Advice k ) llm →T k →T round (k=m)

[0113] Among them, T(C round ) llm It is a tactic generation agent. In the initial round, it takes the round strategy as input and outputs the initial tactic T1. In the next k cycles, it verifies the agent V(T k-1 ) Accept the k-1 The modification strategy T k-1 , output modification suggestions Advice k , tactical generation agent T(T k-1 , Advice k ) llm First k-1 Round modification strategy T k-1 , the kth modification suggestion Advice k , output T k , when k = m, output round strategy T round ;

[0114] S34, the White side subsystem generates deduction events through round-based tactical strategies. The White side subsystem generates deduction events in the form of a global event generation-verification agent loop workflow. This workflow is basically the same as the previous step, and its formula is expressed as:

[0115] E(T round ) llm →ME1

[0116] V(E k-1 ) → Advice k

[0117] E(E k-1 , Advice k ) llm →ME k →ME round (k=m)

[0118] Among them, E(T round ) llm →ME1 represents the process of the global event generation agent generating the initial event. After the next k cycles, the generation-verification flow loop generates modification suggestions and modification strategies, and outputs the deduction event ME for this round when k=m round ;

[0119] S35, after n cycles, the end situation is generated, the events generated by one unit will enter the other unit, and the cycle of steps S31-S34 will be restarted. Each cycle will output the deduction events of the round. After n cycles set manually, all deduction events will be input into the global situation summary agent, and the end situation will be output. The process formula is:

[0120]

[0121] in, represents the overall situation summary agent, IF is an attempt to file a case, It is the event library composed of all deduction events, MS end To end the situation, the attempt to file a case, the event library and the end situation are combined to form a basic scenario.

[0122] Preferably, S4 comprises the following steps:

[0123] S41. Supplementary scenarios are constructed through replay. The main principle of the replay module is to control a local variable in the attempt case and replay it to observe the impact of a certain variable on the overall battle situation. There are two forms of replay: user-defined replay and large-model-defined replay.

[0124] S42. Constructing supplementary scenarios based on the result tree. After the user chooses to use the result tree model to construct supplementary scenario plans in the present invention, the system will file a case based on the initially set strategic intent, and use advanced backtracking algorithms and probability deduction technology to re-simulate and analyze the entire combat sequence. In this process, for each key deduction event node, the system not only records in detail at least three different potential development results that may occur, but also recursively conducts in-depth reasoning and derivative expansion on each possible result in a logical order, forming a multi-level, all-round result deduction tree structure;

[0125] S43. Supplementary assumptions are constructed based on the form of man-in-the-loop. The present invention designs and implements a dynamic supplementary assumption mode constructed based on the man-in-the-loop mode. The core of the system is that users can deeply participate in the simulated battlefield environment according to their own strategic needs and preset camp identities, and receive multi-dimensional and all-round battlefield situation information accurately pushed by the system in real time. This information covers but is not limited to key data such as enemy and friendly position distribution, resource stocks, and environmental changes. After receiving this situation information, users can make flexible arrangements at the tactical level according to actual conditions, such as formulating marching routes, adjusting defense layouts, and optimizing force allocation. At the same time, they can also efficiently deploy various resources at the strategic level, including but not limited to core elements such as weapons and equipment, logistics supplies, and human resources. Every decision made by the user is real-time and interactive. Once the decision instruction is issued, it will be seamlessly connected to the system and used as an important input parameter in the next round of simulation calculation process.

[0126] Preferably, the S41 comprises the following steps:

[0127] S411. Users can initiate deep reconstruction of intention cases through customized strategies. In the stage of strengthening hypothetical scenario construction, users can flexibly specify one or more sets of variable parameters in the intention plan, and the system will perform re-calculation operations based on the updated intention plan to generate more detailed and rich supplementary hypothetical scenarios. For example, in military simulation applications, users can freely set different weather conditions and force deployment configurations to trigger the system's new calculation and analysis of the established strategic results;

[0128] S412. By using a large language model to deeply customize and recalculate basic scenarios, the model can provide diversified variable adjustment suggestions for established intention plans and present these strategy options to users. Users have decision-making power in this process and can choose to adopt any change plan recommended by the model, or make detailed manual adjustments based on these suggestions. Once the intention plan is confirmed and revised by the user, the system will re-implement the deduction process based on the updated plan, thereby generating a set of more comprehensive and supplementary scenarios that meet actual needs.

[0129] Compared with the related art, the multi-agent strategic scenario generation system based on a large language model provided by the present invention has the following beneficial effects:

[0130] The present invention provides a multi-agent strategic scenario generation system based on a large language model, and designs a game subsystem of a multi-level intelligent agent vertical collaboration relationship based on strategy-campaign-tactics. Through the multi-level game subsystem, the task granularity of the scenario designed by the large language model is refined, effectively improving the quality of the scenario design.

[0131] Using a generative-verification adversarial collaborative agent framework, compared to the traditional single large model, this agent framework can timely correct errors and inconsistencies in the generated agent in content creation, enhance logical consistency, effectively improve the output quality of the large language model, and improve the overall deduction assumption level.

[0132] It shows more stable robustness, uses a multi-agent red-blue confrontation framework to simulate the system and create military scenarios. Compared with the direct game and generation of a single intelligent model, it considers multiple variables and uncertain factors, making the simulation closer to the actual battlefield situation.

[0133] By using a better feedback mechanism, using replays, outcome trees, and people-in-the-loop, we can deeply explore the potential outcomes in battlefield simulations, the possible effectiveness of each decision, and the impact of different factors on simulation results, to achieve closed-loop feedback that effectively guides users to improve future action plans. BRIEF DESCRIPTION OF THE DRAWINGS

[0134] Figure 1 A schematic diagram of a preferred embodiment of a multi-agent strategic scenario generation system based on a large language model provided by the present invention;

[0135] Figure 2 It is a structural diagram of the system framework;

[0136] Figure 3 A schematic diagram of the workflow for event generation;

[0137] Figure 4 Generate a schematic diagram of the workflow for the basic scenario;

[0138] Figure 5 Generate schematic diagrams showing scenario details for this technique and the baseline technique.

[0139] Figure 6 Schematic diagram of replying to reviews for the model. DETAILED DESCRIPTION

[0140] The present invention will be further described below in conjunction with the accompanying drawings and implementation modes.

[0141] Please refer to Figure 1 , Figure 2 , Figure 3 , Figure 4 , Figure 5 and Figure 6 ,in, Figure 1 A schematic diagram of a preferred embodiment of a multi-agent strategic scenario generation system based on a large language model provided by the present invention; Figure 2 It is a structural diagram of the system framework; Figure 3 A schematic diagram of the workflow for event generation; Figure 4 Generate a schematic diagram of the workflow for the basic scenario; Figure 5 Generate schematic diagrams showing scenario details for this technique and the baseline technique. Figure 6 A multi-agent strategic scenario generation system based on a large language model includes the following steps:

[0142] S1. Modeling scenarios and systems requires systematic modeling of attempt filing, situation simulation, events and scenarios, strategic-campaign-tactical level red-blue subsystems and white-side subsystems;

[0143] S2. Initialize the red and blue team's attempt to file a case. When the user uses the system, he is prompted to enter the relevant information of the military deduction. After entering the information as required, the system uses the global attempt to file a case to generate an agent, receives the user input information, and generates an attempt to file a case in JSON format. The Markov subchain corresponding to this step is:

[0144] Input→IF→MX→AX

[0145] Among them, Input represents user input, IF represents attempt to file a case, MX represents basic military assumptions, and AX represents supplementary assumptions. This step mainly completes the transformation from user input to attempt to file a case in the system. After the attempt to file a case is generated, the python json toolkit is used to segment the attempt to file a case, and the attempt to file a case from different perspectives of the red and blue sides is obtained, which is used to generate the basic assumptions in the next stage;

[0146] S3. Based on the basic assumption of the attempt to file a case, the present invention uses a cyclic generation workflow of the deduction event-situation. First, each red and blue party subsystem converts its own attempt to file a case into a round situation, and then loops to generate the automatic confrontation workflow. In the loop, one party's troops convert the received deduction event into a dynamic situation; the round situation after the static situation and the dynamic situation are combined is converted into a round strategy; the round strategy is further converted into a round tactic; the round tactic is converted into a deduction event and transmitted to the other party's troops, and the cycle continues. After n cycles set by the user, the confrontation workflow ends;

[0147] S4. Supplementary scenarios are constructed based on basic scenarios. Supplementary scenarios are improvements on basic scenarios. They mainly predict different scenarios in actual combat and provide pre-selected plans. We construct supplementary scenarios through three forms: re-simulation, result tree, and man-in-the-loop. Finally, we combine the attempt to file a case, basic scenario, and supplementary scenario to output a complete military scenario.

[0148] The S1 comprises the following steps:

[0149] S11. Modeling the attempt to file a case. The attempt to file a case is a pre-setting before the military scenario simulation, which includes the strategic layout with the country as the main body and the detailed description of the initial scene of the local war campaign. The attempt to file a case is modeled as a tuple including the strategic background and the campaign description. Its mathematical model is:

[0150] IF=(SC,CD)

[0151] Among them, IF stands for attempt to establish a case, and it includes an introduction to the strategic background SC and a focused campaign description CD;

[0152] S12. Modeling of deduction situation. Deduction situation refers to the overall situation obtained by analyzing and integrating the battle situation. It is an abstract concept extracted based on specific integration. In view of the needs of multi-faceted and complex situation analysis, this patent divides the situation into static situation, dynamic situation and round situation and models them separately;

[0153] S13. Modeling of simulation events and basic assumptions. The military assumptions designed in this patent are mainly event-driven. Its characteristic is that the simulation process is mainly promoted based on key events in the battlefield or strategic environment. The simulation events and basic assumptions are modeled in turn below;

[0154] S14. Construct a red-blue subsystem based on a vertical collaboration relationship of multi-level agents in strategy, campaign and tactics. This subsystem is responsible for playing the roles of the red and blue parties, and conducting battle deduction and game from the perspective of one party. It mainly includes: verification agent, situation summary agent, strategy generation agent, campaign generation agent and tactical generation agent;

[0155] S15. Construct the white side subsystem, which is mainly responsible for receiving global information and implementing corresponding functional modules from a global perspective. The subsystem mainly includes the attempt to file a case agent and the global event generation-verification agent.

[0156] The S11 comprises the following steps:

[0157] S111. Modeling the strategic background. In military scenarios, the strategic background is intended to describe the macro-environmental conditions and strategic situational basis of the campaign or conflict setting, including the description of the regional security environment, the analysis of the causes of war, and the speculation of the opponent's intentions. The strategic background is modeled as follows:

[0158] SC=(RS,CC,AS)

[0159] Among them, SC is the strategic background, RS is the description of the regional security environment, CC is the analysis of the causes of war, and AS is the opponent's strategy;

[0160] S112. Modeling the campaign description. The campaign description is a summary of information at the campaign level. It requires a detailed description of the environmental information of the battle to be held in the scenario, the military armaments of both sides, and the military plan. It is modeled as follows:

[0161] CD=(MA,EI,MP)

[0162] Among them, CD represents campaign description, MA represents military armament, EI represents environmental information, and MP represents military plan. Military armament describes the basic military strength of a party in the military scenario process. The formula is:

[0163] MA=(MA branch ,MA force , M.A. equipment )

[0164] Among them, MA branch Indicates the type of soldier, MA force Indicates the armed forces in the region, MA equipment Indicates the equipment resource information of the party;

[0165] Environmental information EI describes the objective environment of the military scenario, and its formula is:

[0166] EI=(EI time , EI zone , EI weather )

[0167] Among them, EI time Indicates the assumed time, EI zone Represents the hypothetical area, EI weather Indicates weather composition;

[0168] The military plan describes the macro-strategic measures of the entire force, and its formula is:

[0169] IS=(IS goal ,IS method ,IS principle )

[0170] Among them, IS goal Indicates combat objectives, IS method Indicates basic tactics, IS principle Indicates the principle of warfare;

[0171] After modeling the mission scenario, it is necessary to model the simulated situations and events in the military scenario.

[0172] The S12 comprises the following steps:

[0173] S121. Model the static situation. The static situation is an overview of the information of both sides before the battle is played. It will not change with the battle. The mathematical definition of the static situation is:

[0174] MS static =(MA,EI,MP)

[0175] Among them, MS static Represents the static situation, which is an abstract summary extracted from the military armament MA, environmental information EI and military plan MP in the attempt to file a case IF;

[0176] S122. Model the dynamic situation. The dynamic situation is defined as a comprehensive evaluation of the current scenario based on the changing battle situation in the simulation, mainly including time, dynamic changes in military power and basic conditions. The dynamic situation is time-sensitive and variable. In this patent, the definition of the dynamic situation is:

[0177] MS dynamic =(E Time , M equipment-d , S dynamic )#(6)

[0178] Among them, MS dynamic Indicates dynamic situation, E Time Indicates the current time, M equipment-d Indicates the current round of military deployment, S dynamic Indicates the basic situation of the current battle;

[0179] S123, modeling the round situation, which is used to describe the overall battle situation in a round. In this patent, the formula is:

[0180]

[0181] Among them, MS Round Indicates the turn situation, MS static Indicates static situation, MS dynamic represents the dynamic situation, n represents the current round number, in the initialization stage (n=0), the round situation is the static situation, and as the scenario generation begins, the round situation becomes a combination of the static situation and the dynamic situation;

[0182] S124. Model the ending situation. The ending situation is an evaluation summary of the overall battle situation at the end of the simulation, the completion of the strategic goals of both sides, and the loss of confrontation forces after the military scenario is completed. The ending situation is defined as:

[0183] MS end =(SO, OC, FL)

[0184] Among them, MS end Indicates the end situation, SO indicates the overall battle situation, OC indicates the completion of strategic goals, and FL indicates the loss of opposing forces.

[0185] The S14 comprises the following steps:

[0186] S131. Modeling of deduction events. Deduction events are the basic units of basic assumptions. They describe a certain action of a certain party and the changes in the subject and environment caused by the action. In summary, the definition of basic assumptions is:

[0187] ME=(ME entity ,ME action , M.E. task ,ME interaction )

[0188] Among them, ME is the deduction event, which is composed of entity ME entity , Action ME action , Mission ME task and interactive ME interaction composition;

[0189] S132. Model the basic scenario. The basic scenario is an imagined or pre-set war or conflict scenario. The basic scenario is generated by the set of all simulation events and the situation at the end of the simulation. Its mathematical definition is:

[0190]

[0191] MX represents the basic scenario, which includes the attempt to file a case IF, a set of events in n rounds And the end situation MS end .

[0192] The S14 comprises the following steps:

[0193] S141. Construct a verification agent model. The present invention uses a generation-verification adversarial collaboration model, in which the generation agent is responsible for generating an initial solution, and the verification agent gives feedback on the initial solution and passes it to the generation agent for regeneration. When the number of modifications exceeds a certain number, the verification solution is output. In this patent, the verification agent function is defined as receiving a large model input and outputting corresponding modification suggestions. The mathematical model is:

[0194] V({Input}) llm →Advice

[0195] Among them, V({Input}) llm It is the model of the verification agent. The model input {Input} changes according to the usage scenario of the verification agent, and the corresponding prompt words will also change. Advice represents the modification suggestions for the verification agent output, which mainly includes: whether the model output maintains temporal consistency and spatial consistency; whether the model output conforms to the overall logic; whether the grammar, sentence structure and wording of the model output are appropriate;

[0196] S142. Construct a strategy generation agent. The main function of the strategy generation agent is to make an evaluative summary based on the intentions received by the troops and generate the army's macro strategy. The mathematical model is:

[0197] Z(IF) llm →MS static

[0198] Among them, Z(IF) llm represents the strategy generation agent, which takes the attempt to file a case IF as input and outputs a static situation;

[0199] S143. Construct a situation summary agent. The main function of the situation summary agent is to follow the continuous progress of the deduction events and dynamically summarize the battlefield. It takes the deduction events as input and outputs the dynamic situation. The mathematical model is:

[0200] T(ME) llm →MS dynamic

[0201] Among them, T(ME) llm is the situation summary agent, ME is the event deduction agent, MS dynamic It is a dynamic situation;

[0202] S144. Construct a campaign generation agent. The campaign generation agent is responsible for continuously adjusting the troops' campaign strategy according to the dynamic changes in the battlefield situation. Its implementation form is: to receive the round situation and the modification suggestions of the verification agent, generate the overall campaign strategy of the current round. The campaign agent is modeled as:

[0203] C(MS round ,Advice) llm →C round

[0204] Among them, C(MS round ,Ans) llm It is the working model of the campaign agent, which uses the turn-based MS roundand verification agent advice as input, output round battle C round , turn-based battles mainly include: Military objectives: This dimension focuses on clear, quantifiable military objectives, such as destroying enemy forces, occupying key areas, or controlling important resources in strategic operations;

[0205] Resources and force allocation: involves the effective allocation of existing resources and personnel, involving decisions on force size, organizational structure, fire support, and logistics to achieve strategic objectives;

[0206] Time factors: Considering time constraints and urgency issues, some goals need to be achieved immediately, while others allow for longer-term pursuit;

[0207] Comparative analysis of enemy and friendly forces: Analyze the relative strengths, weaknesses, capabilities, and potential vulnerabilities of friendly and enemy forces to assess the feasibility of strategic objectives and their attendant risks;

[0208] S145. Construct a tactical generation agent. The function of the tactical generation agent is to receive the round battle situation and the modification suggestions of the verification agent, and generate the tactical strategy of the current round. The tactical agent is modeled as:

[0209] T(C round ,Advice) llm →T round

[0210] Among them, T(C round ,Advice) llm represents a tactical agent, which uses a turn-based strategy C round , modify the advice to input and output the round tactical strategy T round ;

[0211] Turn-based tactical strategies mainly include: combat maneuvers, which refer to specific actions and movements on the battlefield aimed at gaining tactical advantages;

[0212] Fires Employment: Effectively manage and employ firepower resources to support the conduct of tactical operations;

[0213] Terrain utilization: maximize the advantages of the battlefield's natural topographic features to serve the realization of tactical objectives;

[0214] Force coordination: Ensure seamless cooperation and coordination between various troop units and elements during tactical engagements.

[0215] The S15 comprises the following steps:

[0216] S151, construct an attempt to file a case generating agent, the main function of the attempt to file a case generating agent is to receive and

[0217] Verify user input and output the attempt to file a case. The model of this agent is:

[0218] I({Input}) llm →IF

[0219] Among them, I({Input}) llm It is the agent for generating the attempted case. It receives user input and outputs the JSON format of the attempted case. At the same time, when the user input information is incomplete, the model will feedback the missing information details, try to complete the gaps, and prompt the user to verify and modify;

[0220] S152. Construct a global event generation agent. The main function of the global event generation agent is to receive the turn strategies and tactics output by the red and blue subsystems, and consider the previous events to generate the latest event. The formula is:

[0221] E(C round , T round , {ME1, ME2...ME n-1}) llm →ME n

[0222] Among them, E(C round , T round ,{ME1,ME2...ME n-1}) llm is a global event generating agent, which receives C at the nth round. round , T round And a list of all previous events {ME1, ME2...ME n-1}, output event ME n ;

[0223] S153. Construct a global situation summary agent. The function of the global situation summary agent is to generate a summary evaluation of the battle situation after the game is over. It receives a list of all game events and the attempts of both parties to file a case, and generates an ending situation. The formula is:

[0224]

[0225] in, It is a digital model of the global situation summary agent, which attempts to file a case IF, all the deduction events Input, output end state MS end .

[0226] The S3 comprises the following steps:

[0227] S31. The Red and Blue subsystems generate a turn situation by attempting to file a case and deduce events. The Red and Blue subsystems use the strategy generation agent to generate a static situation based on the attempt to file a case. If the current round is the initial round, the turn situation is the static situation. If it is not the initial round, there is an event input from the opponent's troops. At this time, the situation summarizer is called to generate a dynamic situation based on the event, and the static situation and the dynamic situation are combined into a turn situation. The process of this step is as follows:

[0228] Z(IF) llm →MS static

[0229] T(ME) llm →MS dynamic

[0230]

[0231] Among them, the strategy generation agent Z(IF) llm Generate static situation MS based on attempt to file case IF static , situation generation agent T(ME) llm Generate dynamic situation MS based on simulated events ME dynamic , the turn situation is MS round , it is a dynamic situation MS dynamic and static situation MS static Under different conditions, different combinations

[0232] Through this step, the transition from attempting to file a case to a round situation is achieved;

[0233] S32, the Red and Blue Party subsystem generates a round campaign strategy through the round situation. The Red and Blue Party subsystem generates the round campaign strategy in the form of a campaign generation-verification agent loop workflow. First, the campaign generation agent takes the round situation as input and outputs the initial campaign strategy. The formula for this step is:

[0234] C(MS round ) llm →C1

[0235] Among them, C(MS round ) llm The strategy generating agent starts with MS round The round situation is input and the initial round strategy C1 is output;

[0236] In the next cycle, the verification agent takes the initial campaign strategy as input and outputs modification suggestions; the campaign initial agent takes the modification suggestions and the previous round of campaign strategy as input and outputs the modified campaign; the verification agent takes the modified campaign as input and outputs modification suggestions. After the number of rounds set by the user is looped, the final round of the campaign is output. The formula for this stage is:

[0237] V(C k-1 ) llm →Advice k

[0238] C(C k-1 , Advice m ) llm →C k →C round (k=m)

[0239] Among them, V(C k-1 ) llm →Advice k The recursive formula indicates that the verification agent uses the strategy C of the previous round in the kth round of loops. k-1 Input, output current modification suggestions Advice k , C(C k-1 Advice k ) llm It means that the verification agent takes the previous round of campaign strategy and modification suggestions as input, and outputs the modification strategy C for the Kth round of cycles k , when k = m, output the final round strategy C round ;

[0240] S33, the red and blue subsystem generates a round tactical strategy through a round campaign strategy. The red and blue subsystem generates a round tactical strategy in the form of a tactical generation-verification agent loop workflow. The workflow is basically the same as the previous step, and the formula is expressed as:

[0241] T(C round ) llm →T1

[0242] V(T k-1 ) → Advice k

[0243] T(T k-1 , Advice k ) llm →T k →T round (k=m)

[0244] Among them, T(C round ) llm It is a tactic generation agent. In the initial round, it takes the round strategy as input and outputs the initial tactic T1. In the next k cycles, it verifies the agent V(T k-1 ) Accept the k-1 The modification strategy T k-1 , output modification suggestions Advice k, tactical generation agent T(T k-1 , Advice k ) llm First k-1 Round modification strategy T k-1 , the kth modification suggestion Advice k , output T k , when k = m, output round strategy T round ;

[0245] S34, the White side subsystem generates deduction events through round-based tactical strategies. The White side subsystem generates deduction events in the form of a global event generation-verification agent loop workflow. This workflow is basically the same as the previous step, and its formula is expressed as:

[0246] E(T round ) llm →ME1

[0247] V(E k-1 ) → Advice k

[0248] E(E k-1 , Advice k ) llm →ME k →ME round (k=m)

[0249] Among them, E(T round ) llm →ME1 represents the process of the global event generation agent generating the initial event. After the next k cycles, the generation-verification flow loop generates modification suggestions and modification strategies, and outputs the deduction event ME for this round when k=m round ;

[0250] S35, after n cycles, the end situation is generated, the events generated by one unit will enter the other unit, and the cycle of steps S31-S34 will be restarted. Each cycle will output the deduction events of the round. After n cycles set manually, all deduction events will be input into the global situation summary agent, and the end situation will be output. The process formula is:

[0251]

[0252] in, represents the overall situation summary agent, IF is an attempt to file a case, It is the event library composed of all deduction events, MS end To end the situation, the attempt to file a case, the event library and the end situation are combined to form a basic scenario.

[0253] The S4 comprises the following steps:

[0254] S41. Supplementary scenarios are constructed through replay. The main principle of the replay module is to control a local variable in the attempt case and replay it to observe the impact of a certain variable on the overall battle situation. There are two forms of replay: user-defined replay and large-model-defined replay.

[0255] S42. Constructing supplementary scenarios based on the result tree. After the user chooses to use the result tree model to construct supplementary scenario plans in the present invention, the system will file a case based on the initially set strategic intent, and use advanced backtracking algorithms and probability deduction technology to re-simulate and analyze the entire combat sequence. In this process, for each key deduction event node, the system not only records in detail at least three different potential development results that may occur, but also recursively conducts in-depth reasoning and derivative expansion on each possible result in a logical order, forming a multi-level, all-round result deduction tree structure;

[0256] S43. Supplementary assumptions are constructed based on the form of man-in-the-loop. The present invention designs and implements a dynamic supplementary assumption mode constructed based on the man-in-the-loop mode. The core of the system is that users can deeply participate in the simulated battlefield environment according to their own strategic needs and preset camp identities, and receive multi-dimensional and all-round battlefield situation information accurately pushed by the system in real time. This information covers but is not limited to key data such as enemy and friendly position distribution, resource stocks, and environmental changes. After receiving this situation information, users can make flexible arrangements at the tactical level according to actual conditions, such as formulating marching routes, adjusting defense layouts, and optimizing force allocation. At the same time, they can also efficiently deploy various resources at the strategic level, including but not limited to core elements such as weapons and equipment, logistics supplies, and human resources. Every decision made by the user is real-time and interactive. Once the decision instruction is issued, it will be seamlessly connected to the system and used as an important input parameter in the next round of simulation calculation process.

[0257] The S41 comprises the following steps:

[0258] S411. Users can initiate deep reconstruction of intention cases through customized strategies. In the stage of strengthening hypothetical scenario construction, users can flexibly specify one or more sets of variable parameters in the intention plan, and the system will perform re-calculation operations based on the updated intention plan to generate more detailed and rich supplementary hypothetical scenarios. For example, in military simulation applications, users can freely set different weather conditions and force deployment configurations to trigger the system's new calculation and analysis of the established strategic results;

[0259] S412. By using a large language model to deeply customize and recalculate basic scenarios, the model can provide diversified variable adjustment suggestions for established intention plans and present these strategy options to users. Users have decision-making power in this process and can choose to adopt any change plan recommended by the model, or make detailed manual adjustments based on these suggestions. Once the intention plan is confirmed and revised by the user, the system will re-implement the deduction process based on the updated plan, thereby generating a set of more comprehensive and supplementary scenarios that meet actual needs.

[0260] See also Figure 4 The events generated by the loop are entered into the event library. After the loop ends, the global situation summary agent outputs the basic assumptions.

[0261] The Markov subchain corresponding to this step is:

[0262] IF→MX

[0263] Among them, IF means attempt to file a case, and MX means basic assumption. This step mainly completes the process of the system generating basic assumptions based on the attempt to file a case.

[0264] Appendix 1: Complete scenarios for this technique and baseline model generation

[0265] This technique generates complete scenarios:

[0266] Attempt to file a case

[0267] In this military scenario, titled "The Battle of Kleita Pass," a fierce clash unfolds at dawn at a strategic mountain pass in the Kleita region. Weather conditions are extremely severe, with intermittent snowfall significantly reducing visibility and making the terrain extremely treacherous.

[0268] The Reds deployed a conventional force, equipped with heavy armor and mechanized infantry. Their arsenal included T-90 main battle tanks, BMP-3 infantry fighting vehicles, Mi-28 attack helicopters for air support, AK-12 assault rifles for soldiers, and 2S19 Msta-S self-propelled artillery. Their strategic goal was to control the mountain pass, cut off the supply lines of the Azure Alliance and split their forces. To achieve this goal, they planned to use armored columns to break through the mountain pass while using air superiority to suppress any resistance. They pursued the principle of overwhelming force to quickly break through the enemy's defenses.

[0269] The Blue Force responds with highly advanced and mobile infantry forces equipped with exoskeleton technology and supported by cyber warfare forces. Their equipment includes unmanned aerial vehicles (UAVs) for reconnaissance and precision strikes, directed energy weapons, electromagnetic pulse grenades, and stealthy vertical take-off and landing (VTOL) aircraft for rapid deployment. Their combat objectives are to defend mountain passes and maintain control of important supply lines. In executing a defensive strategy, they plan to use their familiarity with the terrain to set up ambush points, employ guerrilla tactics, including strike retreat attacks, and maximize the use of intelligence collected by drones to effectively strike enemy assets. In addition, they will use cyber warfare capabilities to disrupt the Red Force's command and control systems. The guiding principles of the Blue Force are flexibility, adaptability, and the use of asymmetric tactics to offset the opponent's advantages in numbers and firepower.

[0270] Round 1

[0271] Red side battle decision

[0272] The “military objective”: to quickly control the mountain pass within the first hours of dawn, isolate enemy forces and establish an unassailable stronghold, while ensuring that logistical routes remain open and secure to provide continued support.

[0273] “Resource and Force Allocation”: “Deploy T-90 main battle tanks as the vanguard to break through the enemy defenses in the mountain pass, followed by BMP-3 infantry fighting vehicles carrying mechanized infantry. Keep Mi-28 attack helicopters on standby to conduct reconnaissance and air fire support missions at any time. Strategically deploy 2S19 Msta-S artillery units in the rear to carry out concentrated fire strikes on the enemy’s strong defenses or incoming reinforcements. Develop detailed logistical contingency plans, including equipment maintenance and supply line management in adverse weather conditions.”

[0274] "Time Factor": "Perform operations at dawn when visibility is lowest to maximize surprise, with the goal of achieving operational objectives before the enemy can launch a significant counterattack or receive reinforcements."

[0275] Red side tactical decision

[0276] "Combat Action": Deploy T-90 main battle tanks in a spearhead formation to break through the enemy's defenses at the narrowest part of the pass, and mechanized infantry on BMP-3s to quickly occupy and expand the controlled area. Coordinate Mi-28 attack helicopters to conduct real-time reconnaissance and immediately strike enemy targets. After the breakthrough, split the troops to block the escape route while surrounding the enemy positions on both sides of the pass.

[0277] "Firepower deployment": T-90 main battle tanks are given priority to carry out suppressive fire strikes on enemy strongholds, and 2S19Msta-S artillery units provide concentrated fire support after confirmation based on aerial reconnaissance. Mi-28 attack helicopters are kept on standby to quickly respond to anti-tank threats and provide close air support to advancing armored forces.

[0278] "Terrain Utilization": Take advantage of the terrain's natural bottlenecks to draw enemy forces into the annihilation zone where our heavy armor can decisively engage. Deploy infantry units on high ground and in bottlenecks to counter guerrilla tactics and prevent flanking maneuvers, ensuring clear logistical routes are maintained through less challenging terrain.

[0279] Force Coordination: Implement strict communications protocols to synchronize the actions and firepower of all units, integrating air-ground joint operations to achieve surprise and overwhelming advantage. Enhance intelligence gathering to anticipate and eliminate the Blue Alliance's guerrilla tactics and attempts to use terrain to ambush. Maintain redundancy in communication systems and ensure close coordination between combat units and logistics teams to promptly address maintenance and supply needs.

[0280] Blue side battle decision

[0281] Military objectives: Rapidly capture and control the mountain pass within the next two hours, eliminate enemy artillery positions, disrupt enemy command and control systems through immediate cyber warfare, and maintain uninterrupted control of supply routes.

[0282] Resource and Force Allocation: Continuously deploy drones for real-time reconnaissance and precision strikes against enemy strongholds, communication nodes, and key assets. Deploy exoskeleton-enhanced infantry units and cyber warfare experts to infiltrate enemy defenses, disrupt communication networks, and establish reinforced ambush points. Use T-90 tanks and BMP-3 infantry fighting vehicles to form a vanguard formation to actively break through enemy defenses, while Mi-28 attack helicopters provide continuous air cover and suppressive fire. Deploy artillery positions to conduct immediate counter-artillery strikes and close support for ground forces to strike at enemy strongholds.

[0283] Time Factor: Execute the operation quickly, taking advantage of the limited window of early dawn and deteriorating weather, to prioritize the critical tasks of capturing mountain passes and destroying enemy command structures before the enemy can take advantage of improved visibility or receive reinforcements.

[0284] Comparison of forces: Although the Red Force’s forces are technologically advanced and possess enhanced mobility, firepower, and intelligence capabilities, they may face an enemy that is superior in numbers and well-entrenched. Through rapid, targeted strikes and asymmetric warfare, exploiting the enemy’s mobility, real-time intelligence, and vulnerability to cyber attacks, the Red Force can minimize exposure to the enemy’s heavy firepower and maximize disruption to its operational coordination.

[0285] Blue team tactical decision

[0286] Combat Maneuver: Deploy exoskeleton-enhanced infantry and cyber warfare units to infiltrate enemy lines, disrupt command networks, and pave the way for a synchronized advance led by T-90 tanks and BMP-3 infantry fighting vehicles protected by Mi-28 attack helicopters.

[0287] Fire Employment: Strategically employ directed energy weapons and electromagnetic pulse (EMP) grenades against fortified enemy positions, particularly artillery positions. Coordinate Mi-28 attack helicopters to conduct precision suppressive fire on priority targets identified by real-time drone reconnaissance, while ensuring that artillery positions are prepared to conduct immediate counter-fire actions and provide close support to advancing ground forces.

[0288] Terrain Exploitation: Use the rugged terrain to establish strong ambush points at key bottlenecks and reinforcement routes in mountain passes. Take advantage of intermittent snowfall and reduced visibility to enhance concealment and conduct raids from advantageous high ground and natural cover positions.

[0289] Force Coordination: Maintain continuous communications between all combat units, facilitate real-time intelligence sharing, and make rapid tactical adjustments based on drone feedback, with an emphasis on seamless coordination within the initial two-hour window when the effect of surprise is most important.

[0290] Round 2

[0291] Red side battle decision

[0292] Military Objective: Quickly gain control of a strategic mountain pass to cut off the Blue Alliance's supply lines and isolate its forces, while defending against enemy infiltrators and neutralizing cyber warfare threats.

[0293] Resource and Force Allocation: Consider deploying armored columns led by T-90 tanks and BMP-3 infantry fighting vehicles on the main axes of advance, with dedicated Mi-28 helicopter units for aerial surveillance, anti-infiltration measures, and close air support. Enhance infantry squads with integrated drone surveillance equipment and dedicated cybersecurity teams to detect, deter, and counter Blue's asymmetric tactics and cyber attacks. Assign some artillery units to standby and conduct precision counter-fire strikes on identified energy weapon sites and potential electromagnetic pulse (EMP) threats at any time.

[0294] Time Factor: Execute the Decisive Phase within the first few hours of dawn, taking advantage of surprise and completing control of the pass before Blue can effectively consolidate or reinforce its defensive positions.

[0295] Comparison of forces: While the Red Force has significant advantages in conventional firepower and armor, the Blue Force's exoskeleton-equipped infantry and stealth capabilities present unique challenges in current weather conditions and terrain. Their use of directed energy weapons and cyber warfare requires a comprehensive strategy that combines traditional force projection with advanced surveillance, robust electronic warfare capabilities, and rapid response mechanisms to mitigate enemy advantages.

[0296] Red side tactical decision

[0297] Combat Maneuver: Rapidly advance armored columns led by T-90 tanks, followed by infantry units in BMP-3 infantry fighting vehicles to immediately secure positions. Coordinate Mi-28 helicopters to conduct aerial reconnaissance and provide precision close air support to eliminate identified enemy positions and stealth assets. Deploy infantry squads equipped with drones to conduct active surveillance operations and quickly respond to infiltrators, seamlessly integrating with cybersecurity teams to detect, analyze and mitigate cyber threats in real time.

[0298] Firepower deployment: Use concentrated firepower from T-90 tanks and 2S19 Msta-S self-propelled howitzers to suppress and destroy Blue's fortified positions and energy weapon positions. Coordinate Mi-28 attack helicopters to conduct precision strikes on enemy stealth assets, ensuring continuous air cover and suppression during the advance into the mountain pass. Maintain a dedicated artillery reserve to conduct counter-fire strikes at any time to respond to detected directed energy weapons or EMP threats, and develop backup communication protocols in case of network interruption.

[0299] Terrain Utilization: Use the natural cover provided by mountainous terrain and intermittent snowfall to obscure operations and reduce enemy visibility. Strategically position artillery units on high ground within mountain passes to obtain optimal indirect fire support and establish hardened observation posts as part of an early warning system. Equip infantry units with cold weather equipment and give them specialized mountain warfare training to effectively respond to complex environments.

[0300] Force Coordination: Implement a resilient unified command structure that integrates all arms, including armor, infantry, artillery, aviation, and cybersecurity, into a tightly coordinated force. Enhance secure communications channels with redundancy to facilitate real-time intelligence sharing between units and specialized departments. Develop contingency plans to respond to cyber attacks, including alternate communication methods and pre-planned responses to asymmetric tactics and cyber threats. Continuously optimize situational awareness and adjust tactics based on changing battlefield conditions to maintain tactical advantage.

[0301] Blue side battle decision

[0302] Military objective: quickly capture and maintain control of the mountain pass, eliminate enemy resistance, and destroy the Blue side's command structure to gain the operational initiative.

[0303] Resources and Force Allocation: Utilize T-90 main battle tanks to quickly break through while BMP-3 infantry fighting vehicles immediately consolidate positions. Deploy cyber warfare forces to counter Blue's electronic warfare efforts and protect lines of communication. Coordinate artillery fire to suppress enemy strongholds and pave the way for advancing forces. Stealth vertical take-off and landing aircraft insert exoskeleton-enhanced infantry into strategic high points to defend against ambushes. Utilize drones for persistent reconnaissance, identify priority targets, and conduct precision strikes on enemy command centers.

[0304] Time Factor: Take advantage of dawn and adverse weather conditions to quickly capture key points in the pass before the Blues can organize an effective counterattack. Prioritize securing control of the pass while establishing reinforced positions and persistent surveillance to achieve long-term control.

[0305] Comparison of forces: Anticipate Blue's guerrilla tactics and ambushes, and use our numerical superiority in heavy armor and advanced equipment (exoskeleton infantry, directed energy weapons, stealth vertical take-off and landing aircraft) to mitigate their advantages. Cyber ​​forces will balance potential interference with our command network. By exploiting environmental factors and technological advantages, minimize casualties and maximize combat effects.

[0306] Blue team tactical decision

[0307] Combat maneuver: Rapidly coordinate advances at dawn using T-90 main battle tanks through mountain passes, taking advantage of surprise and reduced visibility. Meanwhile, stealthy vertical takeoff and landing aircraft insert exoskeleton-enhanced infantry into strategic high points, reinforce key positions, and anticipate possible ambush sites to secure supply routes and bottlenecks.

[0308] Firepower deployment: Under the guidance of real-time intelligence from drones, artillery fire is accurately allocated to suppress enemy positions and eliminate threats before the enemy organizes a counterattack. Directed energy weapons are reserved to counter the blue side's armored resistance and destroy its hardened assets, ensuring continuous destruction of its combat capability.

[0309] Terrain Utilization: Use rough terrain as a natural barrier to lead enemy forces into pre-determined kill zones where the firepower of BMP-3 infantry fighting vehicles and other assets can be concentrated. Ensure all units are proficient in setting and evading ambushes at key points within the mountain pass.

[0310] Force Coordination: Maintain an impenetrable communications network protected by active cyber warfare measures to ensure seamless coordination between all unit elements. Continuously update battlefield intelligence collected by drones to guide combat maneuvers, accurately and timely deploy firepower, and adjust strategies to maintain control of the mountain pass. To ensure long-term control, implement a troop rotation and supply system to rest and resupply between engagements, thereby maintaining the endurance and effectiveness of our forces in prolonged operations.

[0311] Round 3

[0312] Red side battle decision

[0313] Resources and Force Allocation: Deploy a dedicated reconnaissance squad with drone coverage to locate and monitor enemy stealth VTOL aircraft and infiltrators. Assign a mechanized infantry unit, closely supported by Mi-28 attack helicopters, to eliminate infiltrators and protect important communication nodes. Organize a concentrated armored assault led by T-90 tanks and BMP-3 infantry fighting vehicles to penetrate mountain passes, with 2S19 Msta-S self-propelled howitzers providing precision suppressive fire to preemptively eliminate known and suspected ambush sites. Maintain a reserve platoon to respond to emerging directed energy weapons and electromagnetic pulse (EMP) threats.

[0314] Time Factor: Ensure full security of communications networks and eliminate all infiltrators within the first hour (07:45, Day 1). The main assault on the pass should begin no later than 08:00, Day 1, taking advantage of improved visibility at dawn to overwhelm enemy defenses.

[0315] Comparison of forces: Blue relies on stealth, asymmetric tactics, and cyber warfare, while we have a decisive advantage in firepower, armor, and air superiority. By leveraging real-time intelligence and adapting to enemy tactics, we will identify and exploit Blue weaknesses. Given the vulnerability of our communications systems and reliance on technology, enhanced cybersecurity measures and backup plans are critical to maintaining command and control throughout the operation.

[0316] Red side tactical decision

[0317] Firepower deployment: Preemptive strikes on identified enemy positions, using 2S19 Msta-S self-propelled howitzers for precision suppressive fire, and adjusting firepower based on real-time updates from reconnaissance teams. Coordinated Mi-28 attack helicopters to provide continuous close air support and rapidly respond to any hostile activity detected on the ground.

[0318] Terrain Utilization: Use intermittent snowfall to provide cover for advancing forces, especially when moving through areas not under direct enemy surveillance. Strategically place artillery positions and observation posts on high ground within mountain passes to optimize battlefield surveillance and fire control. Prioritize movement over less difficult terrain to reduce delays caused by equipment failure or adverse conditions.

[0319] Force Coordination: Strengthen command structures and establish multiple redundant communications channels to ensure all units maintain situational awareness and can quickly adapt to changes. Implement real-time intelligence sharing and regular situation reporting protocols. Develop contingency plans to respond to GPS jamming or communication outages, assign reserve platoons to perform specific tasks, and immediately respond to directed energy weapons and electromagnetic pulse attacks, including manually overriding critical systems when necessary.

[0320] Blue side battle decision

[0321] Military objective: Quickly capture and firmly control the mountain pass, decisively eliminate the Blue side's exoskeleton infantry and cyber warfare forces, and ensure continuity of command and control by maintaining control of communication nodes.

[0322] Resource and Force Allocation: Command T-90 tanks and BMP-3 infantry fighting vehicles to advance synchronously, taking advantage of improved visibility and coordinating with 2S19 Msta-S systems for precision artillery support. Deploy mechanized infantry equipped with exoskeletons to reinforce occupied positions and conduct aggressive sweeps to eliminate infiltrator threats. Use stealthy vertical take-off and landing aircraft to send rapid response teams deep into enemy territory, provide protection from directed energy weapons (DEW) and electromagnetic pulse (EMP), and systematically disrupt the Blue team's command structure. Coordinate Mi-28 helicopters to provide continuous close air support and synchronize drone reconnaissance to obtain real-time intelligence and conduct precision strikes against high-priority enemy assets.

[0323] Time Factor: Accomplish initial objectives quickly (capture mountain passes, protect communications nodes) within Day 1 while weakening enemy resistance; Progressively advance from Day 1 to Day 2, pursuing comprehensive area control.

[0324] Force Comparison: Given Red's advantages in firepower, mobility, and advanced technology, including drones and DEW / EMP countermeasures, these advantages should be used to offset Blue's familiarity with the terrain and possible reliance on guerrilla tactics. By pre-seizing control points, leveraging intelligence-driven mobility, and disrupting Blue's command networks, we will mitigate its potential numerical superiority and asymmetric threats.

[0325] Blue team tactical decision

[0326] Firepower deployment: Utilize the 2S19 Msta-S self-propelled howitzer system to cooperate with the advancing armored forces to accurately strike the identified blue team strongholds and exoskeleton infantry assembly points. Directed energy weapons carried by the rapid reaction team will be actively used to paralyze key enemy assets, while electromagnetic pulse (EMP) grenades will be tactically deployed to destroy the blue team's electronic systems and communication nodes.

[0327] Terrain Exploitation: Strategically deploy troops on high ground to provide indirect fire support and direct Blue to pre-determined kill zones where Red combined forces can be most effective. Optimize drone flight paths to proactively identify Blue ambush sites and guerrilla tactics along high ground and valleys and feed real-time intelligence back to coordinating forces.

[0328] Force Coordination: Integrate cyber warfare tactics into synchronized maneuvers to actively jam or hijack Blue’s communications networks to further disrupt command and control. Continuously update artillery missions and close air support from Mi-28 helicopters based on real-time drone intelligence to ensure a rapid and decisive response to changing battlefield conditions. Facilitate seamless communications between all Red units to implement a unified and flexible tactical approach that leverages our technological advantages and offsets Blue’s familiarity with the terrain.

[0329] Event List

[0330] Round 1

[0331] "Day 1, 05:30, strategically important mountain passes in the Kleita area - intermittent snowfall and reduced visibility,

[0332] Red Team: T-90 main battle tanks begin to move through the narrowest passages under the cover of dawn and low visibility. BMP-3 infantry fighting vehicles follow closely behind, occupying positions behind the tanks' advance. Mi-28 attack helicopters are in the air, providing real-time reconnaissance and suppressive fire for enemy strongholds. Artillery units remain on standby at predetermined strategic locations, ready to provide concentrated fire support when needed. Red Team forces coordinate operations to traverse complex terrain, establish defensive positions to prevent potential guerrilla-style counterattacks, and strictly adhere to communication protocols to ensure synchronized operations and rapid response to dynamic combat scenarios. "

[0333] "Day 1, 06:45, strategically important mountain passes in the Kleita area - continued intermittent snowfall, visibility still significantly reduced;

[0334] BLUE: Exoskeleton-equipped infantry units quickly infiltrated Red lines, aiming to cut communications networks and create chaos. At the same time, cyber warfare operators launched complex attacks against Red command infrastructure. Stealthy vertical take-off and landing aircraft (VTOL) deployed troops to tactically advantageous positions, setting up ambush points on possible reinforcement routes. Real-time intelligence from unmanned aerial vehicles (UAVs) enabled directed energy weapons and electromagnetic pulse (EMP) grenades to accurately strike Red positions. Blue artillery units remained ready to execute counterattack missions and support advancing friendly forces at any time. Over the next two hours, Blue carefully coordinated its asymmetric attack, using severe environmental conditions to offset Red's numerical advantage and heavy armor advantage, while maintaining secure and efficient communications to exploit the changing battle dynamics. "

[0335] Round 2

[0336] "Day 1, 08:00, strategically important mountain passes in the Kleita region - continued intermittent snowfall, visibility remains low

[0337] Red Force: As visibility improves, T-90 main battle tanks and BMP-3 infantry fighting vehicles drive through the pass to initiate the main armored push. This push is closely coordinated with precision suppressive fire from 2S19 Msta-S artillery units, striking targeted and suspected enemy positions. After thoroughly clearing out infiltrators, mechanized infantry units secure and consolidate the area around the communications node, forming a protective barrier. Mi-28 attack helicopters remain alert in the air, ready to provide air support for any enemy activity. Reconnaissance teams work in conjunction with unmanned aerial vehicles (UAVs) to continuously feed real-time intelligence to the command center, ensuring tactics can be quickly adapted. Quick reaction platoons remain alert and ready to respond to any emerging directed energy weapon (DEW) or electromagnetic pulse (EMP) threats, thereby strengthening the resilience of the Red Force's command and control system throughout the operation."

[0338] "Day 1, 10:00, strategically important mountain pass in the Kleita area - intermittent snowfall

[0339] BLUE: In response to the Red’s armored advance and artillery assault, exoskeleton-equipped infantry conduct precision strikes against Red’s communications nodes. Meanwhile, cyber warfare units escalate electronic jamming operations. Stealth vertical takeoff and landing aircraft (VTOL) rapidly deploy small strike teams to strategic locations to ambush advancing Red mechanized infantry and T-90 tanks. Real-time intelligence from unmanned aerial vehicles (UAVs) guides the precise use of directed energy weapons and electromagnetic pulse (EMP) grenades, targeting high-value targets in an effort to degrade Red’s air support and command capabilities. Blue’s artillery batteries maintain a defensive posture, ready to conduct counterattack missions in response to Red’s offensive, all coordinated through secure and reliable communications channels.”

[0340] Round 3

[0341] "Day 1, 12:00, strategically important mountain passes in the Kleta region - intermittent snowfall continues

[0342] Red: After quickly regrouping and redeploying forces, mechanized infantry and armored columns advanced with air support from Mi-28 attack helicopters to seize strategic high ground and control key bottlenecks. Meanwhile, special forces penetrated deep behind Blue lines to locate and destroy stealth vertical take-off and landing aircraft (VTOL) deployment points. Electronic warfare (EW) teams immediately launched precision operations against Blue cyber assets within the first hour of regrouping. Meanwhile, Red artillery units such as 2S19 Msta-S artillery began to conduct suppressive fire missions against targeted directed energy weapon (DEW) assets and enemy artillery positions. In a coordinated advance, Red's infantry and armored forces advanced with precision support from close air support to systematically eliminate Blue's exoskeleton infantry. Taking advantage of adverse weather conditions and dangerous terrain, Red consolidated its defenses at key mountain pass locations and led enemy forces into pre-set kill zones. Additionally, reserve infantry reinforcements were deployed to key areas as part of an integrated defense plan and impending counterattack, while secure communications networks were maintained to ensure seamless coordination between artillery, air and ground forces.”

[0343] "Day 1, 14:30, strategically important mountain pass in the Kleita area - intermittent snowfall

[0344] Blue: In rapid action, exoskeleton-equipped infantry and special forces quickly seized and fortified positions abandoned by the Red as the battle line shifted. They effectively targeted Red’s communications infrastructure and VTOL deployment points with surgical strikes. Meanwhile, cyber warfare units continued to degrade Red’s C4ISR systems. According to a three-hour strategic timeline, Blue coordinated precision artillery and air strikes targeting targeted directed energy weapon (DEW) assets and enemy artillery positions, successfully neutralizing long-range threats. Taking advantage of the rugged terrain and low visibility from intermittent snowfall, they set up ambushes along anticipated reinforcement routes, targeting Red’s armored columns and mechanized infantry. As part of a multi-layered defense strategy, reserve infantry reinforcements reinforced critical bottlenecks while maintaining secure and jam-resistant communications networks to allow for real-time intelligence sharing and rapid response to changing combat conditions.”

[0345] Complete scenario generated by ChatGPT3.5

[0346] Scene Setting

[0347] Fighting in the Cleta region at dawn, the conflict between the Iron Vanguard and the Azure Alliance presents two completely different military strategies and tactical confrontations. The Iron Vanguard, with its heavy armored columns and powerful firepower advantage, attempts to quickly break through the mountain pass, cut off the Azure Alliance's supply lines and split its defenses. However, faced with the complex terrain of the Cleta region and the Azure Alliance's precise response, the Iron Vanguard faces severe challenges.

[0348] The Kleta region is located in a rugged mountain range with narrow valleys and dense forests. These terrains provide excellent defensive advantages, but also limit the mobility of large armored vehicles. The weather conditions are severe, and intermittent snowfall reduces visibility, adding uncertainty and danger to the battlefield.

[0349] Iron Vanguard Tactics and Challenges

[0350] The Iron Vanguard plans to quickly break through the mountain pass with the main armored column and use air support to strike the Azure Alliance's defenses. They rely on the heavy armor and mobility of the T-90 main battle tanks and BMP-3 infantry fighting vehicles, while Mi-28 attack helicopters provide air fire support. Its tactical goal is to quickly destroy the enemy's defenses with overwhelming firepower and numerical superiority, split the Azure Alliance forces and control strategic locations.

[0351] However, faced with complex terrain and the highly mobile response of the Azure Alliance, the Iron Vanguard faced multiple challenges:

[0352] Terrain Restrictions: Narrow valleys and dense forests limit the mobility and firepower of armored units, making them easy targets for the Azure Alliance's guerrilla tactics.

[0353] Weather impact: Snowfall reduces visibility and increases the risks and difficulties for armored forces in mountain warfare, especially for helicopter flight and fire support.

[0354] The Azure Alliance used its drones for real-time reconnaissance and deployed guerrilla tactics through infantry squads wearing exoskeletons. They used directed energy weapons and electromagnetic pulse grenades to destroy the Iron Vanguard's communications and electronic equipment, effectively weakening its command and combat effectiveness.

[0355] Azure Alliance Counterattack and Strategy

[0356] The tactics of the Azure Alliance focus on flexibility and technological advancement:

[0357] Defensive layout: Using terrain advantages to set up ambush points and defensive positions, relying on highly mobile infantry squads to conduct guerrilla warfare. They collect intelligence through drones and use cyber warfare capabilities to disrupt the enemy's command system and communication links.

[0358] Precision strikes: Through directed energy weapons and precision strikes, focus on destroying the Iron Vanguard Army’s key assets, such as main battle tanks and command centers, effectively weakening its combat capabilities.

[0359] Tactical Retreat: Using the mobility of high-tech infantry, the Azure Alliance launches short and effective attacks and quickly withdraws, avoiding long direct confrontations with armored forces.

[0360] in conclusion

[0361] The battle in the Cleta region reflects the combination of technology and tactics in modern warfare, as well as the profound impact of terrain and weather conditions on battlefield situations. The confrontation between the Iron Vanguard and the Azure Alliance showcases the contrast between traditional armored forces and high-tech mobile forces, with each side seeking strategic advantages by exploiting its own strengths and the opponent's weaknesses.

[0362] Complete scenario of iFlytek Spark generation

[0363] Military Scenario: Battle of the Kleita Pass

[0364] 1. Background

[0365] In the fictional region of Kleita, two opposing armies - the Iron Vanguard and the Azure Alliance - are fighting over strategic locations. The Iron Vanguard, led by General Orlov, has powerful armored forces and highly mechanized infantry; the Azure Alliance is known for its flexible and technologically advanced combat methods, including highly mobile infantry, drone reconnaissance and precision strikes, infantry squads wearing exoskeletons, etc. Both sides hope to win this battle, control the mountain pass and cut off the other side's supply lines.

[0366] II. War Objectives and Plans

[0367] Iron Vanguard: Control the mountain pass, cut off the supply lines of the Azure Alliance and split their forces. Plan to use armored columns to break through the mountain pass, while using air support to suppress any resistance from above. The principle of war is to quickly break through the enemy lines with overwhelming force.

[0368] Azure Alliance: Defend mountain passes and maintain control of supply routes. Will use knowledge of the terrain to set up ambush points and defensive positions. Rely on drones to gather intelligence and conduct precision strikes on key enemy assets such as command centers and tank battalions. Tactics include guerrilla warfare, strike-and-retreat attacks using the high-tech mobility of infantry, and use cyber warfare capabilities to disable and confuse the enemy's command system. Principles of warfare emphasize flexibility, adaptability, and asymmetric warfare to offset the Iron Vanguard's numerical advantage and superior firepower.

[0369] Battle

[0370] At dawn, both sides began to deploy their forces. The Iron Vanguard used its armored forces and air support to try to quickly break through the Azure Alliance's defenses. The Azure Alliance relied on the advantages of terrain and drones to set up ambush points and defensive positions, waiting for the Iron Vanguard to fall into the trap.

[0371] During the battle, both sides suffered losses, but the Iron Vanguard maintained their advantage. They successfully broke through part of the Azure Alliance's defenses, but found that the Azure Alliance had set up a strong encirclement.

[0372] Faced with the encirclement of the Azure Alliance, the Iron Vanguard Army was in trouble. They tried to break through, but were met with stubborn resistance from the Azure Alliance. In the end, the Iron Vanguard Army was forced to retreat, and the two sides launched a fierce pursuit battle near the mountain pass.

[0373] After a day of fighting, both sides suffered heavy casualties. Although the Iron Vanguard failed to completely occupy the mountain pass, it successfully stopped the attack of the Azure Alliance; the Azure Alliance suffered heavy losses during the retreat, but successfully saved the mountain pass and supply line.

[0374] Compared with the related art, the multi-agent strategic scenario generation system based on a large language model provided by the present invention has the following beneficial effects:

[0375] The present invention provides a multi-agent strategic scenario generation system based on a large language model, and designs a game subsystem of a multi-level intelligent agent vertical collaboration relationship based on strategy-campaign-tactics. Through the multi-level game subsystem, the task granularity of the scenario designed by the large language model is refined, effectively improving the quality of the scenario design.

[0376] Using a generative-verification adversarial collaborative agent framework, compared to the traditional single large model, this agent framework can timely correct errors and inconsistencies in the generated agent in content creation, enhance logical consistency, effectively improve the output quality of the large language model, and improve the overall deduction assumption level.

[0377] It shows more stable robustness, uses a multi-agent red-blue confrontation framework to simulate the system and create military scenarios. Compared with the direct game and generation of a single intelligent model, it considers multiple variables and uncertain factors, making the simulation closer to the actual battlefield situation.

[0378] By using a better feedback mechanism, using replays, outcome trees, and people-in-the-loop, we can deeply explore the potential outcomes in battlefield simulations, the possible effectiveness of each decision, and the impact of different factors on simulation results, to achieve closed-loop feedback that effectively guides users to improve future action plans.

[0379] The above descriptions are merely embodiments of the present invention and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A multi-agent strategic scenario generation system based on a large language model, characterized by: The following steps are involved: S1. Modeling scenarios and systems. It is necessary to model the attempt to file a case, simulate the situation, events and scenarios, and the strategic-campaign-tactical level red-blue subsystem and white-side subsystem; S2. Initialize the red and blue team's attempt to file a case. When the user uses the system, he is prompted to enter the relevant information of the military deduction. After entering the information as required, the system uses the global attempt to file a case to generate an agent, receives the user input information, and generates an attempt to file a case in JSON format. The Markov subchain corresponding to this step is: Input→IF→MX→AX Among them, Input represents user input, IF represents attempt to file a case, MX represents basic military assumptions, and AX represents supplementary assumptions. This step mainly completes the transformation from user input to attempt to file a case in the system. After the attempt to file a case is generated, the python json toolkit is used to segment the attempt to file a case, and the attempt to file a case from different perspectives of the red and blue sides is obtained, which is used to generate the basic assumptions in the next stage; S3. Based on the basic assumption of the attempt to file a case, the present invention uses a cyclic generation workflow of the deduction event-situation. First, each red and blue party subsystem converts its own attempt to file a case into a round situation, and then loops to generate the automatic confrontation workflow. In the loop, one party's troops convert the received deduction event into a dynamic situation; the round situation after the static situation and the dynamic situation are combined is converted into a round strategy; the round strategy is further converted into a round tactic; the round tactic is converted into a deduction event and transmitted to the other party's troops, and the cycle continues. After n cycles set by the user, the confrontation workflow ends; S4. Construct supplementary scenarios based on basic scenarios. Supplementary scenarios are the improvement of basic scenarios. They mainly predict different scenarios in actual combat and provide pre-selected solutions. We construct supplementary scenarios through three forms: replay, result tree, and man-in-the-loop. Finally, we combine the attempt to file a case, basic scenario, and supplementary scenario to output a complete deduction scenario.

2. The multi-agent strategic scenario generation system based on a large language model according to claim 1, characterized in that: The S1 comprises the following steps: S11. Modeling the attempt to file a case. The attempt to file a case is a pre-setting before the military scenario simulation, which includes the strategic layout with the country as the main body and the detailed description of the initial scene of the local war campaign. The attempt to file a case is modeled as a tuple including the strategic background and the campaign description. Its mathematical model is: IF=(SC,CD) Among them, IF stands for attempt to establish a case, and it includes an introduction to the strategic background SC and a focused campaign description CD; S12. Modeling the deduction situation. The deduction situation refers to the overall situation obtained by analyzing and integrating the battle situation. It is an abstract concept extracted based on specific integration. In view of the needs of multi-faceted and complex situation analysis, the situation is divided into static situation, dynamic situation and turn situation for modeling respectively. S13. Modeling of simulation events and basic assumptions. The designed military assumptions are mainly event-driven. Their characteristics are that the simulation process is mainly based on key events in the battlefield or strategic environment. The simulation events and basic assumptions are modeled in turn below. S14. Construct a red-blue subsystem based on a vertical collaboration relationship of multi-level agents in strategy, campaign and tactics. This subsystem is responsible for playing the roles of the red and blue parties, and conducting battle deduction and game from the perspective of one party. It mainly includes: verification agent, situation summary agent, strategy generation agent, campaign generation agent and tactical generation agent; S15. Construct the white side subsystem, which is mainly responsible for receiving global information and implementing corresponding functional modules from a global perspective. The subsystem mainly includes the attempt to file a case agent and the global event generation-verification agent.

3. The multi-agent strategic scenario generation system based on a large language model according to claim 2, characterized in that: The S11 comprises the following steps: S111. Modeling the strategic background. In the scenario, the strategic background is intended to describe the macro-environmental conditions and strategic situation basis of the campaign or conflict setting, including the description of the regional security environment, the analysis of the causes of the war, and the speculation of the opponent's intentions. The strategic background is modeled as follows: SC=(RS,CC,AS) Among them, SC is the strategic background, RS is the description of the regional security environment, CC is the analysis of the causes of war, and AS is the opponent's strategy; S112. Modeling the campaign description. The campaign description is a summary of campaign-level information, which requires a detailed description of the environmental information of the battle to be held in the scenario, the military armaments of both sides, and the military plan. It is modeled as follows: CD=(MA,EI,MP) Among them, CD represents campaign description, MA represents military armament, EI represents environmental information, and MP represents military plan. Military armament describes the basic military strength of a party in the military scenario process. The formula is: MA=(MA branch , MA force , MA equipment ) Among them, MA branch Indicates the type of soldier, MA force Indicates the armed forces in the region, MA equipment Indicates the equipment resource information of the party; Environmental information EI describes the objective environment of the military scenario, and its formula is: NO=(NO time ,NO zone ,NO weather ) Among them, EI time Indicates the assumed time, EI zone Represents the hypothetical area, EI weather Indicates weather composition; The military plan describes the macro-strategic measures of the entire force, and its formula is: IS=(IS goal ,IS method ,IS principle ) Among them, IS goal Indicates combat objectives, IS method Indicates basic tactics, IS principle Indicates the principle of warfare; After modeling the mission scenario, it is necessary to model the simulated situations and events in the military scenario.

4. The multi-agent strategic scenario generation system based on a large language model according to claim 2, characterized in that: The S12 comprises the following steps: S121. Model the static situation. The static situation is an overview of the information of both sides before the battle is played. It will not change with the battle. The mathematical definition of the static situation is: MS static =(MA,EI,MP) Among them, MS static Represents the static situation, which is an abstract summary extracted from the military armament MA, environmental information EI and military plan MP in the attempt to file a case IF; S122. Model the dynamic situation. The dynamic situation is defined as a comprehensive evaluation of the current scenario based on the changing battle situation in the simulation, mainly including time, dynamic changes in military power and basic conditions. The dynamic situation is time-sensitive and variable. The definition of the dynamic situation is: MS dynamic =(E Time ,M equipment-d ,S dynamic )#(6) Among them, MS dynamic Indicates dynamic situation, E Time Indicates the current time, M equipment-d Indicates the current round of military deployment, S dynamic Indicates the basic situation of the current battle; S123, modeling the round situation, which is used to describe the overall battle situation in a round. In this patent, the formula is: Among them, MS Round Indicates the turn situation, MS static Indicates static situation, MS dynamic represents the dynamic situation, n represents the current round number, in the initialization stage (n=0), the round situation is the static situation, and as the scenario generation begins, the round situation becomes a combination of the static situation and the dynamic situation; S124. Model the ending situation. The ending situation is an evaluation summary of the overall battle situation at the end of the simulation, the degree of completion of the strategic goals of both sides, and the loss of confrontation forces after the military scenario is completed. The ending situation is defined as: MS end =(SO,OC,FL) Among them, MS end Indicates the end situation, SO indicates the overall battle situation, OC indicates the completion of strategic goals, and FL indicates the loss of opposing forces.

5. The multi-agent strategic scenario generation system based on a large language model according to claim 2, characterized in that: The S14 comprises the following steps: S131. Modeling of deduction events. Deduction events are the basic units of basic assumptions. They describe a certain action of a certain party and the changes in the subject and environment caused by the action. In summary, the definition of basic assumptions is: ME=(ME entity ,ME action ,ME task ,ME interaction ) Among them, ME is the deduction event, which is composed of entity ME entity , Action ME action , Mission ME task and interactive ME interaction composition; S132. Model the basic scenario. The basic scenario is an imagined or pre-set war or conflict scenario. The basic scenario is generated by the set of all simulation events and the situation at the end of the simulation. Its mathematical definition is: MX represents the basic scenario, which includes the attempt to file a case IF, a set of events in n rounds And the end situation MS end .

6. The multi-agent strategic scenario generation system based on a large language model according to claim 2, characterized in that: The S14 comprises the following steps: S141, construct a verification agent model. The present invention uses a generation-verification adversarial collaboration model, in which the generation agent is responsible for generating the initial solution, the verification agent gives feedback on the initial solution and passes it to the generation agent for regeneration. When the number of modifications exceeds a certain number, the verification solution is output. The verification agent function is defined as receiving the input of the large model and outputting the corresponding modification opinions. The mathematical model is: V({Input}) llm →Advice Among them, V({Input}) llm It is the model of the verification agent. The model input {Input} changes according to the usage scenario of the verification agent, and the corresponding prompt words will also change. Advice represents the modification suggestions for the verification agent output, which mainly includes: whether the model output maintains temporal consistency and spatial consistency; whether the model output conforms to the overall logic; whether the grammar, sentence structure and wording of the model output are appropriate; S142. Construct a strategy generation agent. The main function of the strategy generation agent is to make an evaluative summary based on the intentions received by the troops and generate the army's macro strategy. The mathematical model is: Z(IF) llm →MS static Among them, Z(IF) llm represents the strategy generation agent, which takes the attempt to file a case IF as input and outputs a static situation; S143. Construct a situation summary agent. The main function of the situation summary agent is to follow the continuous progress of the deduction events and dynamically summarize the battlefield. It takes the deduction events as input and outputs the dynamic situation. The mathematical model is: T(ME) llm →MS dynamic Among them, T(ME) llm is the situation summary agent, ME is the event deduction agent, MS dynamic It is a dynamic situation; S144. Construct a campaign generation agent. The campaign generation agent is responsible for continuously adjusting the troops' campaign strategy according to the dynamic changes in the battlefield situation. Its implementation form is: to receive the round situation and the modification suggestions of the verification agent, generate the overall campaign strategy of the current round. The campaign agent is modeled as: C(MS round ,Advice) llm →C round Among them, C(MS round ,Ans) llm It is the working model of the campaign agent, which uses the turn-based MS round and verification agent advice as input, output round battle C round , turn-based battles mainly include: Military objectives: This dimension focuses on clear, quantifiable military objectives, such as destroying enemy forces, occupying key areas, or controlling important resources in strategic operations; Resources and force allocation: involves the effective allocation of existing resources and personnel, involving decisions on force size, organizational structure, fire support, and logistics to achieve strategic objectives; Time factors: Considering time constraints and urgency issues, some goals need to be achieved immediately, while others allow for longer-term pursuit; Comparative analysis of enemy and friendly forces: Analyze the relative strengths, weaknesses, capabilities, and potential vulnerabilities of friendly and enemy forces to assess the feasibility of strategic objectives and their attendant risks; S145. Construct a tactical generation agent. The function of the tactical generation agent is to receive the round battle situation and the modification suggestions of the verification agent, and generate the tactical strategy of the current round. The tactical agent is modeled as: T(C round ,Advice) llm →T round Among them, T(C round ,Advice) llm represents a tactical agent, which uses a turn-based strategy C round , modify the advice to input and output the round tactical strategy T round ; Turn-based tactical strategies mainly include: combat maneuvers, which refer to specific actions and movements on the battlefield aimed at gaining tactical advantages; Fires Employment: Effectively manage and employ firepower resources to support the conduct of tactical operations; Terrain utilization: maximize the advantages of the battlefield's natural topographic features to serve the realization of tactical objectives; Force coordination: Ensure seamless cooperation and coordination between various troop units and elements during tactical engagements.

7. The multi-agent strategic scenario generation system based on a large language model according to claim 1, characterized in that: The S15 comprises the following steps: S151. Construct an agent for generating an attempt to file a case. The main function of the agent for generating an attempt to file a case is to receive and verify user input and output an attempt to file a case. The model of the agent is: I({Input}) llm →IF Among them, I({Input}) llm It is the agent for generating the attempted case. It receives user input and outputs the JSON format of the attempted case. At the same time, when the user input information is incomplete, the model will feedback the missing information details, try to complete the gaps, and prompt the user to verify and modify; S152. Construct a global event generation agent. The main function of the global event generation agent is to receive the turn strategies and tactics output by the red and blue subsystems, and consider the previous events to generate the latest event. The formula is: E(C round ,T round ,{ME1,ME2…ME n-1 }) llm →ME n Among them, E(C round ,T round ,{ME1,ME2…ME n-1 }) llm is a global event generating agent, which receives C at the nth round. round ,T round And the list of all previous events {ME1,ME2…ME n-1 }, output event ME n ; S153. Construct a global situation summary agent. The function of the global situation summary agent is to generate a summary evaluation of the battle situation after the game is over. It receives a list of all game events and the attempts of both parties to file a case, and generates an ending situation. The formula is: in, It is a digital model of the global situation summary agent, which attempts to file a case IF, all the deduction events Input, output end state MS end .

8. The multi-agent strategic scenario generation system based on a large language model according to claim 1, characterized in that: The S3 comprises the following steps: S31. The Red and Blue subsystems generate a turn situation by attempting to file a case and deduce events. The Red and Blue subsystems use the strategy generation agent to generate a static situation based on the attempt to file a case. If the current round is the initial round, the turn situation is the static situation. If it is not the initial round, there is an event input from the opponent's troops. At this time, the situation summarizer is called to generate a dynamic situation based on the event, and the static situation and the dynamic situation are combined into a turn situation. The process of this step is as follows: Z(IF) llm →MS static T(ME) llm →MS dynamic Among them, the strategy generation agent Z(IF) llm Generate static situation MS based on attempt to file case IF static , situation generation agent T(ME) llm Generate dynamic situation MS based on simulated events ME dynamic , the turn situation is MS round , it is a dynamic situation MS dynamic and static situation MS static Under different conditions, different combinations Through this step, the transition from attempting to file a case to a round situation is achieved; S32, the Red and Blue Party subsystem generates a round campaign strategy through the round situation. The Red and Blue Party subsystem generates the round campaign strategy in the form of a campaign generation-verification agent loop workflow. First, the campaign generation agent takes the round situation as input and outputs the initial campaign strategy. The formula for this step is: C(MS round ) llm →C1 Among them, C(MS round ) llm The strategy generating agent starts with MS round The round situation is input and the initial round strategy C1 is output; In the next cycle, the verification agent takes the initial campaign strategy as input and outputs modification suggestions; the campaign initial agent takes the modification suggestions and the previous round of campaign strategy as input and outputs the modified campaign; the verification agent takes the modified campaign as input and outputs modification suggestions. After the number of rounds set by the user is looped, the final round of the campaign is output. The formula for this stage is: V(C k-1 ) llm →Advice k C(C k-1 ,Advice m ) llm →C k →C round (k=m) Among them, V(C k-1 ) llm →Advice k The recursive formula indicates that the verification agent uses the strategy C of the previous round in the kth round of loops. k-1 Input, output current modification suggestions Advice k , C(C k-1 ,Advice k ) llm It means that the verification agent takes the previous round of campaign strategy and modification suggestions as input, and outputs the modification strategy C for the Kth round of cycles k , when k = m, output the final round strategy C round ; S33, the red and blue subsystem generates a round tactical strategy through a round campaign strategy. The red and blue subsystem generates a round tactical strategy in the form of a tactical generation-verification agent loop workflow. The workflow is basically the same as the previous step, and the formula is expressed as: T(C round ) llm →T1 V(T k-1 )→Advice k T(T k-1 ,Advice k ) llm →T k →T round (k=m) Among them, T(C round ) llm It is a tactic generation agent. In the initial round, it takes the round strategy as input and outputs the initial tactic T1. In the next k cycles, it verifies the agent V(T k-1 ) Accept the k-1th modification strategy T k-1 , output modification suggestions Advice k , tactical generation agent T(T k-1 ,Advice k ) llm Take the modified strategy T in the k-1th round k-1 , the kth modification suggestion Advice k , output T k , when k = m, output round strategy T round ; S34, the White side subsystem generates deduction events through round-based tactical strategies. The White side subsystem generates deduction events in the form of a global event generation-verification agent loop workflow. This workflow is basically the same as the previous step, and its formula is expressed as: E(T round ) llm →ME1 V(E k-1 )→Advice k E(E k-1 ,Advice k ) llm →ME k →ME round (k=m) Among them, E(T round ) llm →ME1 represents the process of the global event generation agent generating the initial event. After the next k cycles, the generation-verification flow loop generates modification suggestions and modification strategies, and outputs the deduction event ME for this round when k=m round ; S35, after n cycles, the end situation is generated, the events generated by one unit will enter the other unit, and the cycle of steps S31-S34 will be restarted. Each cycle will output the deduction events of the round. After n cycles set manually, all deduction events will be input into the global situation summary agent, and the end situation will be output. The process formula is: in, represents the overall situation summary agent, IF is an attempt to file a case, It is the event library composed of all deduction events, MS end To end the situation, the attempt to file a case, the event library and the end situation are combined to form a basic scenario.

9. The multi-agent strategic scenario generation system based on a large language model according to claim 1, characterized in that: The S4 comprises the following steps: S41. Supplementary scenarios are constructed through replay. The main principle of the replay module is to control a local variable in the attempt case and replay it to observe the impact of a certain variable on the overall battle situation. There are two forms of replay: user-defined replay and large-model-defined replay. S42. Constructing supplementary scenarios based on the result tree. After the user chooses to use the result tree model to construct supplementary scenario plans in the present invention, the system will file a case based on the initially set strategic intent, and use advanced backtracking algorithms and probability deduction technology to re-simulate and analyze the entire combat sequence. In this process, for each key deduction event node, the system not only records in detail at least three different potential development results that may occur, but also recursively conducts in-depth reasoning and derivative expansion on each possible result in a logical order, forming a multi-level, all-round result deduction tree structure; S43. Supplementary assumptions are constructed based on the form of man-in-the-loop. The present invention designs and implements a dynamic supplementary assumption mode constructed based on the man-in-the-loop mode. The core of the system is that users can deeply participate in the simulated battlefield environment according to their own strategic needs and preset camp identities, and receive multi-dimensional and all-round battlefield situation information accurately pushed by the system in real time. This information covers but is not limited to key data such as enemy and friendly position distribution, resource stocks, and environmental changes. After receiving this situation information, users can make flexible arrangements at the tactical level according to actual conditions, such as formulating marching routes, adjusting defense layouts, and optimizing force allocation. At the same time, they can also efficiently deploy various resources at the strategic level, including but not limited to core elements such as weapons and equipment, logistics supplies, and human resources. Every decision made by the user is real-time and interactive. Once the decision instruction is issued, it will be seamlessly connected to the system and used as an important input parameter in the next round of simulation calculation process.

10. The multi-agent strategic scenario generation system based on a large language model according to claim 9, characterized in that: The S41 comprises the following steps: S411. Users can initiate deep reconstruction of intention cases through customized strategies. In the stage of strengthening hypothetical scenario construction, users can flexibly specify one or more sets of variable parameters in the intention plan, and the system will perform re-calculation operations based on the updated intention plan to generate more detailed and rich supplementary hypothetical scenarios. For example, in military simulation applications, users can freely set different weather conditions and force deployment configurations to trigger the system's new calculation and analysis of the established strategic results; S412. By using a large language model to deeply customize and recalculate basic scenarios, the model can provide diversified variable adjustment suggestions for established intention plans and present these strategy options to users. Users have decision-making power in this process and can choose to adopt any change plan recommended by the model, or make detailed manual adjustments based on these suggestions. Once the intention plan is confirmed and revised by the user, the system will re-implement the deduction process based on the updated plan, thereby generating a set of more comprehensive and supplementary scenarios that meet actual needs.

Images