An autonomous decision-making method, device, equipment and storage medium combining large model and self-game technology

Abstract

The present invention provides an autonomous decision-making method, device, equipment and storage medium that combines a large model with self-game technology. First, structured sensor data and unstructured intelligence data are acquired, and a large language model based on the MoE architecture is used to process multimodal information. The results are combined with the initial prompt words to generate a structured decision input; then, the first-layer commander agent generates multiple candidate decision strategies based on the structured input, and assigns them to the second-layer execution agent for simulation execution; the third-layer evaluation agent performs a multi-dimensional collaborative evaluation on the execution results, and returns feedback information to the commander agent to adjust the strategy when the preset threshold is not met; finally, a simulated game is performed on the candidate strategy set that meets the scoring conditions, and the execution results and scoring data are re-input into the MoE large language model, and the final decision plan is generated by fusion, and the scoring threshold is used as the termination condition for output. This solves the problem that existing autonomous decision-making cannot meet the needs of complex and dynamic battlefield environments.

Description

Technical Field

[0001] The present invention relates to the field of large models, and in particular to an autonomous decision-making method, device, equipment and storage medium that combine large models with self-game technology. Background Art

[0002] With the rapid development of artificial intelligence (AI) technology, demand for autonomous decision-making systems in complex and dynamic environments, such as military, finance, and industrial control, is growing. Traditional decision-making systems lack the ability to effectively integrate heterogeneous data from multiple sources, making it difficult to simultaneously process different types of input data, such as text, images, and time series signals. This single data source leads to information gaps or misjudgments during the decision-making process, impacting decision quality and response speed. Decision delays in existing systems are particularly prominent in dynamic environments requiring rapid response, such as battlefield situational awareness.

[0003] In view of this, this application is filed. Summary of the Invention

[0004] The present invention discloses an autonomous decision-making method, device, equipment and storage medium that combine a large model with self-game technology, aiming to solve the problem that existing autonomous decision-making cannot cope with the needs of complex and dynamic battlefield environments.

[0005] The first embodiment of the present invention provides an autonomous decision-making method combining a large model with self-game technology, including:

[0006] Acquiring multimodal data, processing the multimodal data using a large language model based on a MoE architecture, and combining the processed data with an initial prompt word to generate a structured decision input, wherein the multimodal data includes structured sensor data and unstructured data;

[0007] Calling a pre-built three-layer asynchronous hierarchical agent to sequentially process the structured decision input, wherein the three-layer asynchronous hierarchical agent includes a first-layer commander agent, a second-layer execution agent, and a third-layer evaluation agent. The first-layer commander agent generates multiple candidate decision strategies based on the structured decision input and assigns them to corresponding second-layer execution agents. The third-layer evaluation agent performs a collaborative evaluation on the decision solutions generated by the second-layer execution agents. If the evaluation fails, feedback information is returned to the first-layer commander agent to regenerate a decision strategy.

[0008] A simulated game is conducted on multiple candidate decision strategies that have passed the collaborative evaluation, and the execution results and scoring data of each strategy in the simulation environment are obtained. The execution results and scoring data are re-input into the large language model of the MoE architecture, and a fusion decision scheme is generated based on the empirical data of multiple strategies. The fusion decision scheme is simulated and evaluated, and a final decision is output when the score is greater than or equal to the highest score among the candidate decision strategies.

[0009] Preferably, the multimodal data input is obtained, the multimodal data is processed using a large language model of the MoE architecture, and the processed data is combined with the initial prompt word to generate a structured decision input, specifically:

[0010] Obtaining multimodal data input, the multimodal data including structured real-time situation data and unstructured data, wherein the structured real-time situation data includes sensor data, location information, and state parameters, and the unstructured data includes image, audio, and video data;

[0011] Combining the initial prompt words with the structured real-time situation data to form basic input data, wherein the initial prompt words include role settings, task objectives, constraints, precautions, output format, and historical decision actions;

[0012] When unstructured data exists, the image, audio, and video data are first preprocessed using a multimodal large model, converted into descriptive information in text format, and then fused with the basic input data;

[0013] The fused data is input into the large language model of the MoE architecture for cleaning and screening, key information related to the current task is extracted, and the structured decision input is generated.

[0014] Preferably, the first-layer commander agent generates multiple candidate decision strategies based on the structured decision input and assigns them to corresponding second-layer execution agents, specifically:

[0015] The first-layer commander agent receives the structured decision input, parses the available agent resource information and task requirement information, splits the complex task into multiple subtasks using a task decomposition algorithm, and generates N asynchronous and parallel candidate strategy solutions based on resource constraints, time windows, and risk assessments, where each strategy solution includes a task priority matrix, a resource allocation vector, and an execution timing diagram;

[0016] Based on the agent capability description information in the structured decision input, dynamically construct an agent capability matching matrix M(i, j), where i represents the type and difficulty coefficient of the subtask, and j represents the type and capability parameters of the second-layer execution agent. M(i, j) calculates the similarity between the capability vector of agent j and the requirement vector of task i to obtain a fitness score;

[0017] The KM algorithm is used to achieve optimal task allocation based on the agent capability matching matrix M(i, j), and the execution instructions containing specific execution parameters, constraints and coordination interfaces are distributed to the corresponding second-layer execution agents;

[0018] After receiving the execution instruction, the second-layer execution agent calls the corresponding professional decision-making module to generate a specific action plan according to the task type in the instruction.

[0019] Preferably, the calling of a corresponding professional decision module according to the task type in the instruction to generate a specific action plan is as follows:

[0020] When the mission type is flight control, the aircraft control agent generates the flight path and maneuver strategy based on the flight dynamics model and trajectory planning algorithm;

[0021] When the task type is multi-unit collaboration, the formation coordination agent uses a distributed consensus algorithm to handle the coordination problem and generate the formation and communication protocol;

[0022] When the task type is a special operation, the functional module agent generates an operation plan based on the characteristics of the weapon system and the parameters of the detection equipment;

[0023] Each second-layer execution agent shares decision progress and resource usage in real time through a preset state synchronization protocol. When a policy conflict or resource competition is detected, a local negotiation algorithm based on game theory is initiated to adjust the policy and generate a specific action plan that includes spatiotemporal coordinates, action sequences, resource requirements, and expected effects.

[0024] Preferably, the third-layer evaluation agent performs a collaborative evaluation on the decision-making scheme generated by the second-layer execution agent. If the evaluation fails, feedback information is returned to the first-layer commander agent to regenerate the decision-making strategy, specifically:

[0025] The third-layer evaluation agent receives the specific action plan generated by the second-layer execution agent, performs a synergy analysis on the specific action plan, and calculates synergy effect evaluation indicators by comparing and analyzing the performance differences between the joint execution mode and the individual execution mode, wherein the synergy effect evaluation indicators include efficiency improvement rate, risk reduction rate, resource utilization rate, and goal achievement degree;

[0026] The comprehensive synergy score of the synergy effect evaluation index is calculated by weighted fusion.

[0027] When the comprehensive collaborative score is greater than or equal to the preset collaborative threshold, a simulated game will be conducted through a combination of specific action plans for collaborative evaluation. When the comprehensive collaborative score is less than the preset collaborative threshold, a feedback report containing a collaborative defect analysis is generated and returned to the first-level commander intelligent agent to regenerate candidate strategy plans.

[0028] The second embodiment of the present invention provides an autonomous decision-making device that combines a large model with self-game technology, including:

[0029] a structured strategy generation unit, configured to acquire multimodal data, process the multimodal data using a large language model based on the MoE architecture, and combine the processed data with an initial prompt word to generate a structured decision input, wherein the multimodal data includes structured sensor data and unstructured data;

[0030] a hierarchical agent processing unit, configured to call a pre-built three-layer asynchronous hierarchical agent to sequentially process the structured decision input, wherein the three-layer asynchronous hierarchical agent includes a first-layer commander agent, a second-layer execution agent, and a third-layer evaluation agent; the first-layer commander agent generates multiple candidate decision strategies based on the structured decision input and assigns them to corresponding second-layer execution agents; the third-layer evaluation agent performs a collaborative evaluation on the decision solutions generated by the second-layer execution agents, and returns feedback information to the first-layer commander agent to regenerate a decision strategy if the evaluation fails;

[0031] A decision generation unit is used to simulate the game of multiple candidate decision strategies that have passed the collaborative evaluation, obtain the execution results and scoring data of each strategy in the simulation environment, re-input the execution results and scoring data into the large language model of the MoE architecture, generate a fusion decision plan based on the empirical data of multiple strategies, simulate and evaluate the fusion decision plan, and output a final decision when the score is greater than or equal to the highest score among the candidate decision strategies.

[0032] The third embodiment of the present invention provides an autonomous decision-making device that combines a large model with self-game technology, characterized in that it includes a memory and a processor, the memory stores a computer program, and the computer program can be executed by the processor to implement an autonomous decision-making method that combines a large model with self-game technology as described in any one of the above items.

[0033] The fourth embodiment of the present invention provides a computer-readable storage medium storing a computer program, which can be executed by a processor of a device where the computer-readable storage medium is located to implement an autonomous decision-making method combining a large model and self-game technology as described in any one of the above items.

[0034] Based on the autonomous decision-making method, device, equipment and storage medium that combine a large model with self-game technology provided by the present invention, multimodal data is first acquired, and the multimodal data is processed by a large language model of MoE architecture, and the processed data is combined with the initial prompt word to generate a structured decision input, wherein the multimodal data includes structured sensor data and unstructured data; then, a pre-built three-layer asynchronous hierarchical agent is called to process the structured decision input in sequence, wherein the three-layer asynchronous hierarchical agent includes a first-layer commander agent, a second-layer execution agent and a third-layer evaluation agent, and the first-layer commander agent generates multiple candidate agents based on the structured decision input. The decision-making strategy is generated and assigned to the corresponding second-layer execution agent. The third-layer evaluation agent performs a collaborative evaluation on the decision-making scheme generated by the second-layer execution agent. If the evaluation fails, feedback information is returned to the first-layer commander agent to regenerate the decision-making strategy. Finally, a simulated game is conducted using multiple candidate decision-making strategies that have been collaboratively evaluated to obtain the execution results and scoring data of each strategy in the simulation environment. The execution results and scoring data are re-input into the large language model of the MoE architecture. A fusion decision-making scheme is generated based on the empirical data of multiple strategies. The fusion decision-making scheme is simulated and evaluated, and the final decision is output when the score is greater than or equal to the highest score among the candidate decision-making strategies. This solves the problem that existing autonomous decision-making cannot meet the needs of complex and dynamic battlefield environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 This is a flowchart of an autonomous decision-making method combining a large model and self-game technology provided by the first embodiment of the present invention;

[0036] Figure 2 This is a module diagram of an autonomous decision-making device that combines a large model with self-game technology, provided by the second embodiment of the present invention. DETAILED DESCRIPTION

[0037] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0038] In order to better understand the technical solution of the present invention, the embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0039] It should be understood that the embodiments described are only a portion of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by persons of ordinary skill in the art without creative work are within the scope of protection of the present invention.

[0040] The terms used in the embodiments of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention. The singular forms "a", "an", "the" and "the" used in the embodiments of the present invention and the appended claims are also intended to include plural forms unless the context clearly indicates otherwise.

[0041] It should be understood that the term "and / or" as used herein is merely a description of the relationship between associated objects, indicating that three possible relationships exist. For example, "A and / or B" can represent: A exists alone, A and B exist simultaneously, or B exists alone. Furthermore, the character " / " in this document generally indicates that the associated objects are in an "or" relationship.

[0042] The word "if," as used herein, may be interpreted as "at the time of" or "when" or "in response to determining" or "in response to detecting," depending on the context. Similarly, the phrases "if it is determined" or "if (stated condition or event) is detected" may be interpreted as "when it is determined" or "in response to the determination" or "when detecting (stated condition or event)" or "in response to detecting (stated condition or event)," depending on the context.

[0043] The "first" and "second" mentioned in the embodiments are merely used to distinguish similar objects and do not represent a specific ordering of the objects. It is understood that the specific order or precedence of "first" and "second" can be interchanged where appropriate. It should be understood that the objects distinguished by "first" and "second" can be interchanged where appropriate, so that the embodiments described herein can be implemented in an order other than that illustrated or described herein.

[0044] The specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0045] The present invention discloses an autonomous decision-making method, device, equipment and storage medium that combine a large model with self-game technology, aiming to solve the problem that existing autonomous decision-making cannot cope with the needs of complex and dynamic battlefield environments.

[0046] See also Figure 1The first embodiment of the present invention provides an autonomous decision-making method that combines a large model with self-game technology, which can be executed by an autonomous decision-making device, and in particular, by one or more processors within the autonomous decision-making device, to implement at least the following steps:

[0047] S101, acquiring multimodal data, processing the multimodal data using a large language model based on a MoE architecture, and combining the processed data with an initial prompt word to generate a structured decision input, wherein the multimodal data includes structured sensor data and unstructured data;

[0048] In this embodiment, the autonomous decision-making device can be a terminal with data processing capabilities such as a server, desktop computer, or laptop computer. The autonomous decision-making device can be installed with a corresponding operating system and application software, and the functions required by this embodiment can be realized through the combination of the operating system and application software.

[0049] It's important to note that in military combat decision-making support scenarios, autonomous decision-making devices receive multi-dimensional data streams from the battlefield environment in real time. Structured, real-time situational data captures enemy targets' latitude and longitude coordinates, altitude information, velocity vectors, and identification codes through radar networks. Simultaneously, status parameters such as fuel level, weapon loadout status, and communication link quality are obtained from friendly units. Unstructured data includes battlefield imagery captured by drones, communication audio intercepted by electronic reconnaissance equipment, and video streams transmitted back by satellites.

[0050] In a military scenario, the role is set as a "joint operations commander," with a clear mission objective of "controlling the target area with minimal casualties." Constraints include "avoiding civilian areas" and "prioritizing the protection of high-value friendly targets." Historical decision actions record tactical choices and effectiveness evaluations from previous time periods, providing empirical reference for current decision-making.

[0051] When an emergency occurs on the battlefield, the multimodal large model processes reconnaissance imagery in real time, converting visual information such as "3 enemy armored vehicles, 2.5 kilometers from our position, moving southeast at 40 km / h" into precise text descriptions. Audio data processing converts enemy radio communications into intelligence text such as "Enemy call sign Alpha-7 requesting fire support detected, suspected target of our forward observation post."

[0052] The advantage of the MoE architecture lies in its ability to activate the most appropriate expert module based on mission characteristics. When processing fused data, the system automatically identifies the urgency and relevance of the information. For example, when the keyword "enemy missile launch signs" is detected, the military expert module is prioritized and quickly extracts key information such as threat level, estimated impact time, and available interception methods. Routine logistics supply information is handled by the resource management expert module, focusing on supply route security and transportation timeliness. After processing by the MoE model, the raw multi-source data is reorganized into hierarchical decision-making inputs.

[0053] S102: Calling a pre-built three-layer asynchronous hierarchical agent to sequentially process the structured decision input, wherein the three-layer asynchronous hierarchical agent includes a first-layer commander agent, a second-layer execution agent, and a third-layer evaluation agent. The first-layer commander agent generates multiple candidate decision strategies based on the structured decision input and assigns them to corresponding second-layer execution agents. The third-layer evaluation agent performs a collaborative evaluation on the decision solutions generated by the second-layer execution agents. If the evaluation fails, feedback information is returned to the first-layer commander agent to regenerate a decision strategy.

[0054] Taking a military combat decision-making support scenario as an example, the first-level commander agent receives a structured decision input containing the current battlefield situation: "Three enemy armored targets are located 15 kilometers northeast, and the enemy possesses two first-tier fighter squadrons, one ground artillery battery, and one electronic jamming unit." The system immediately initiates a task decomposition algorithm. Using the analytic hierarchy process (AHP), the system intelligently decomposes the complex task of "destroying the enemy armored cluster" into three concurrently executable subtasks: "air suppression," "fire strike," and "electronic jamming." Based on resource constraints—limited fuel for 90 minutes of operation, ammunition inventories at 70%, and weather conditions allowing a four-hour operational window—the commander agent asynchronously generates five candidate strategies: "lightning assault," "continuous suppression," and "circumvention." Each strategy includes a clear task priority matrix assigning the highest priority to air suppression, a resource allocation vector specifying 70% of combat power for the primary attack, and an execution sequence diagram detailing the start time and duration of each phase.

[0055] The system extracts the first fighter's capability parameters (maximum combat radius of 1,000 kilometers, payload of 3 tons, and maneuverability of 9.2) from the structured input to form a capability vector. It also analyzes the "air suppression" subtask's requirements (high-speed maneuverability, long-range strike, and electronic countermeasures) to form a requirement vector. Cosine similarity calculations reveal that the first fighter agent has a 0.94 match with the air suppression mission, while the ground artillery agent has a only 0.23 match. The KM algorithm, through global optimization, ensures the optimal configuration: the first fighter squadron is assigned to the air suppression mission, the artillery company is responsible for ground fire support, and the electronic countermeasures unit performs the jamming mission.

[0056] When the execution order was issued, the first fighter agent received detailed combat parameters, including "route coordinates Nx°Ey°, attack altitude 8,000 meters, weapon selection AIM-120 missile, and coordination frequency with ground forces 145.5MHz." The artillery agent received precise instructions for "firing element azimuth 045°, range 12.5 kilometers, ammunition type selected high-explosive, and fire rate density 8 rounds per minute."

[0057] Furthermore, upon receiving the specific action plans from the first fighter agent, "conduct air suppression at an altitude of 8,000 meters and a speed of Mach 1.2," the ground artillery agent, "provide fire support at an azimuth of 045° and a range of 12.5 kilometers," and the electronic jamming team agent, "interfere with the 144-146MHz frequency to disrupt enemy communications," the third-tier evaluation agent immediately initiated a collaborative analysis. The system constructed a joint combat simulation, simulating a three-service coordinated effort to complete the "destroy enemy armor cluster" mission within 60 minutes, with an 8% probability of friendly losses, 75% of the standard ammunition consumption, and a 90% target destruction rate. Compared to the individual execution model, the first fighter alone took 95 minutes, had a 22% probability of losses, consumed 120% of ammunition, and achieved a target destruction rate of only 70%.

[0058] Precise calculations revealed an efficiency improvement rate of (95-60) / 95 = 36.8%, a risk reduction rate of (22-8) / 22 = 63.6%, a resource utilization rate of (120-75) / 120 = 37.5%, and a target achievement rate of 90%. Considering the urgency and strategic value of the current operation, the evaluation agent assigned the highest weight of 0.4 to risk control, 0.3 to efficiency improvement, 0.2 to resource utilization, and 0.1 to target achievement, resulting in a comprehensive coordination score of 0.4 × 0.636 + 0.3 × 0.368 + 0.2 × 0.375 + 0.1 × 0.9 = 0.519. This score exceeded the preset coordination threshold of 0.5, confirming that the air-ground-electric integrated combat plan possessed excellent coordination and automatically passed it on to the simulation game phase.

[0059] When the coordination score falls short of the target, the evaluation agent immediately generates a flaw analysis report. For example, in one scenario, the first fighter's attack timing conflicted with artillery fire, and the electronic jamming frequency band overlapped friendly communication frequencies, resulting in a coordination score of only 0.42. The evaluation agent generated detailed feedback: "There was a 15-minute overlap between the air strike and ground fire; the electronic jamming frequency affected friendly 145.5MHz coordination frequency; it was recommended to adjust the first fighter's attack window to after the artillery fire ended; and adjust the electronic jamming frequency band to 146-148MHz to avoid friendly communications." This feedback was then fed back to the first-tier commander agent to re-optimize the combat timing and frequency allocation.

[0060] S103, simulate the game for multiple candidate decision strategies that have passed the collaborative evaluation, obtain the execution results and scoring data of each strategy in the simulation environment, re-input the execution results and scoring data into the large language model of the MoE architecture, generate a fusion decision plan based on the empirical data of multiple strategies, simulate and evaluate the fusion decision plan, and output the final decision when the score is greater than or equal to the highest score among the candidate decision strategies.

[0061] Specifically, in this embodiment, the execution experience, advantages and disadvantages features and scoring data of the candidate strategies are integrated into fusion input information and re-input into the large language model of the MoE architecture; the large language model analyzes the effective decision-making elements of each strategy based on the fusion input information, extracts the advantageous features and generates a fusion decision plan; the fusion decision plan is simulated and evaluated, and the fusion score is calculated. When the fusion score is greater than or equal to the highest score of the candidate strategies at the stage, the fusion decision plan is output as the final decision; when the fusion score is less than the highest score of the candidate strategies at the stage, the evaluation result is returned to this step as new input information to re-make the fusion decision until the output condition is met.

[0062] Taking the aforementioned integrated air-ground-electric combat scenario as an example, the system needs to verify the effectiveness of multiple candidate strategies that have passed collaborative evaluation in real combat environments through simulated game theory. This is because theoretical analysis alone cannot fully predict unexpected situations and enemy countermeasures in a complex battlefield environment. Once the three strategies of "Lightning Assault Strategy," "Continuous Suppression Strategy," and "Encirclement and Detour Strategy" have all passed the collaborative evaluation, the system constructs a high-fidelity battlefield simulation environment to simulate dynamic confrontation scenarios, including terrain fluctuations, weather changes, and enemy reactions.

[0063] During the simulated game, the "Lightning Assault" strategy exposed insufficient fuel reserves in its first fighter plane when facing a rapid, mobile enemy counterattack, resulting in a combat sustainability score of only 72 points. The "Continuous Suppression" strategy, while stable, exhibited low time efficiency, earning a score of 78. The "Encirclement and Detour" strategy demonstrated the best adaptability in complex terrain conditions, achieving an overall score of 85. The system not only recorded the final score of each strategy but, more importantly, collected detailed execution data, including key information such as the effectiveness of each strategy in responding to the enemy's air defense missile launch in the 23rd minute of the simulation, the impact of deteriorating weather conditions on the combat plan in the 45th minute, and the ammunition consumption curves and target destruction timing for different strategies.

[0064] This rich simulation data was then fed back into the MoE architecture's large language model. The system chose to feed this empirical data back to the large model because of its powerful pattern recognition and knowledge integration capabilities, enabling it to extract optimal decision-making logic from the success factors and failure lessons of multiple strategies. Analysis by the large model revealed the advantages of the "outflanking and encircling" strategy in terrain adaptability, the "sustained suppression" strategy in stability, and the "lightning assault" strategy in time efficiency. Using a deep learning algorithm, the key elements of these three strategies were intelligently integrated to create an innovative "phased hybrid strategy": Lightning assaults were used to quickly establish an advantage in the early stages (e.g., the first 30 minutes), followed by sustained suppression to consolidate gains in the middle stages, and then flexibly adjusted to outflanking and encircling strategies to achieve the ultimate objective based on terrain conditions in the later stages.

[0065] When the system simulated and evaluated this fusion decision-making solution, it achieved an overall score of 92 in the same simulation environment, significantly exceeding the previous highest score of 85 for a single strategy. Because the fusion solution's score met the pre-set output condition of "not less than the highest score among the candidate strategies," the system confirmed its combat feasibility and automatically output the "phased hybrid strategy" as the final operational decision to the commander. This demonstrates that the fusion decision-making mechanism of simulated game theory and experience-based learning ensures that the system can generate an optimal solution in a complex and dynamic environment that maintains the advantages of each individual strategy while circumventing its limitations.

[0066] In a possible implementation of the present invention, the calling of a corresponding professional decision module according to the task type in the instruction to generate a specific action plan is as follows:

[0067] When the mission type is flight control, the aircraft control agent generates the flight path and maneuver strategy based on the flight dynamics model and trajectory planning algorithm;

[0068] When the task type is multi-unit collaboration, the formation coordination agent uses a distributed consensus algorithm to handle the coordination problem and generate the formation and communication protocol;

[0069] When the task type is a special operation, the functional module agent generates an operation plan based on the characteristics of the weapon system and the parameters of the detection equipment;

[0070] Each second-layer execution agent shares decision progress and resource usage in real time through a preset state synchronization protocol. When a policy conflict or resource competition is detected, a local negotiation algorithm based on game theory is initiated to adjust the policy and generate a specific action plan that includes spatiotemporal coordinates, action sequences, resource requirements, and expected effects.

[0071] Taking the aforementioned military combat scenario as an example, when the first fighter agent receives the command to "execute air suppression mission," the system automatically identifies the mission type as flight control and immediately invokes the built-in flight dynamics model to calculate the optimal trajectory. Based on the first fighter's flight performance parameters, such as a thrust-to-weight ratio of 1.096, a maximum overload of 9G, and a turning radius of 2.1 kilometers, and considering the current threat environment of "enemy air defense missile range of 25 kilometers and radar detection range of 40 kilometers," the aircraft control agent uses the A* algorithm and particle swarm optimization to generate a three-dimensional flight path: after taking off from the base, climbing to 12,000 meters to avoid low-altitude threats, then engaging the enemy at a speed of Mach 1.3 at a direction 30 degrees north-east, and descending to 8,000 meters at a distance of 35 kilometers to enter the attack route. The agent also develops a contingency strategy for "immediately executing a barrel roll and deploying jammers in the event of a missile threat."

[0072] Meanwhile, upon receiving the "Multi-unit Coordinated Advance" command, the ground armored unit's formation coordination agent identified the need to coordinate the operations of four main battle tanks and immediately activated a distributed consensus algorithm. The system established a state matrix based on each tank's current position, fuel status, and ammunition inventory. Using a neighborhood communication mechanism, the units reached consensus on the formation parameters: a wedge formation, 500-meter spacing, and a unified advance speed of 30 km / h. A three-level communication protocol consisting of a primary frequency of 144.2 MHz, a backup frequency of 146.8 MHz, and an emergency call frequency of 148.5 MHz was established to ensure reliable information transmission. Upon receiving the "Specialized Electronic Countermeasures Operation" command, the functional module agent immediately accessed the electronic warfare equipment database. Based on the jammer's technical specifications of "5000 watts of power, a frequency range of 100-300 MHz, and a 15-kilometer interference radius," combined with the detected enemy communication band characteristics, it generated a precise operation plan: "Noise jamming targeting the 144-146 MHz band, for 45 minutes, with power gradually increasing to maximum."

[0073] Each executing agent reports key information, including current position, mission progress, and resource consumption, to each other every 30 seconds through a pre-set state synchronization protocol. When the system detects a seven-minute overlap between the first fighter's attack window and the ground forces' fire support opportunity, it immediately initiates a local negotiation algorithm based on a Nash equilibrium. After multiple rounds of negotiation, the fighter agent and the artillery agent ultimately reach a compromise: the first fighter initiates its attack three minutes earlier, while the artillery fires four minutes later. This ensures the air strike's suddenness while minimizing the risk of friendly fire. After coordinated optimization, the agents ultimately generate a complete action plan, including: the first fighter drops its bomb at coordinates a°b north latitude and c°d'e east longitude; the ground forces execute the action sequence of reconnaissance, attack, and consolidation; the estimated consumption of 1,200 liters of aviation fuel and 24 155mm artillery shells; the expected destruction of 85% of enemy targets; and a guaranteed casualty rate of less than 5%.

[0074] In one possible implementation of the present invention, after the final decision is output, the actual execution effect is continuously monitored; the deviation data between the actual execution result and the expected result is collected; the deviation data is fed back to step S101 for optimizing the initial prompt word strategy and data processing parameters; and at the same time, the deviation data is fed back to step S102 for dynamically adjusting the intelligent body architecture parameters and decision generation strategy.

[0075] See also Figure 2 The second embodiment of the present invention provides an autonomous decision-making device that combines a large model with self-game technology, including:

[0076] a structured strategy generation unit 201 for acquiring multimodal data, processing the multimodal data using a large language model based on the MoE architecture, and combining the processed data with an initial prompt word to generate a structured decision input, wherein the multimodal data includes structured sensor data and unstructured data;

[0077] The hierarchical agent processing unit 202 is configured to call a pre-built three-layer asynchronous hierarchical agent to sequentially process the structured decision input. The three-layer asynchronous hierarchical agent includes a first-layer commander agent, a second-layer execution agent, and a third-layer evaluation agent. The first-layer commander agent generates multiple candidate decision strategies based on the structured decision input and assigns them to corresponding second-layer execution agents. The third-layer evaluation agent performs a collaborative evaluation on the decision solutions generated by the second-layer execution agents. If the evaluation fails, the evaluation agent returns feedback to the first-layer commander agent to regenerate a decision strategy.

[0078] The decision generation unit 203 is used to simulate the game of multiple candidate decision strategies that have passed the collaborative evaluation, obtain the execution results and scoring data of each strategy in the simulation environment, re-input the execution results and scoring data into the large language model of the MoE architecture, generate a fusion decision plan based on the empirical data of multiple strategies, simulate and evaluate the fusion decision plan, and output the final decision when the score is greater than or equal to the highest score among the candidate decision strategies.

[0079] The third embodiment of the present invention provides an autonomous decision-making device that combines a large model with self-game technology, characterized in that it includes a memory and a processor, the memory stores a computer program, and the computer program can be executed by the processor to implement an autonomous decision-making method that combines a large model with self-game technology as described in any one of the above items.

[0080] The fourth embodiment of the present invention provides a computer-readable storage medium storing a computer program, which can be executed by a processor of a device where the computer-readable storage medium is located to implement an autonomous decision-making method combining a large model and self-game technology as described in any one of the above items.

[0081] Based on the autonomous decision-making method, device, equipment and storage medium that combine a large model with self-game technology provided by the present invention, multimodal data is first acquired, and the multimodal data is processed by a large language model of MoE architecture, and the processed data is combined with the initial prompt word to generate a structured decision input, wherein the multimodal data includes structured sensor data and unstructured data; then, a pre-built three-layer asynchronous hierarchical agent is called to process the structured decision input in sequence, wherein the three-layer asynchronous hierarchical agent includes a first-layer commander agent, a second-layer execution agent and a third-layer evaluation agent, and the first-layer commander agent generates multiple candidate agents based on the structured decision input. The decision-making strategy is generated and assigned to the corresponding second-layer execution agent. The third-layer evaluation agent performs a collaborative evaluation on the decision-making scheme generated by the second-layer execution agent. If the evaluation fails, feedback information is returned to the first-layer commander agent to regenerate the decision-making strategy. Finally, a simulated game is conducted using multiple candidate decision-making strategies that have been collaboratively evaluated to obtain the execution results and scoring data of each strategy in the simulation environment. The execution results and scoring data are re-input into the large language model of the MoE architecture. A fusion decision-making scheme is generated based on the empirical data of multiple strategies. The fusion decision-making scheme is simulated and evaluated, and the final decision is output when the score is greater than or equal to the highest score among the candidate decision-making strategies. This solves the problem that existing autonomous decision-making cannot meet the needs of complex and dynamic battlefield environments.

[0082] For example, the computer programs described in the third and fourth embodiments of the present invention can be divided into one or more modules, which are stored in the memory and executed by the processor to implement the present invention. The one or more modules can be a series of computer program instruction segments capable of performing specific functions, which describe the execution process of the computer program in the autonomous decision-making device that implements a combination of a large model and self-game technology. For example, the device described in the second embodiment of the present invention.

[0083] The processor may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor, etc. The processor is the control center of the autonomous decision-making method combining a large model with self-game technology, and utilizes various interfaces and lines to connect the various parts of the autonomous decision-making method combining a large model with self-game technology.

[0084] The memory can be used to store the computer programs and / or modules. The processor implements the various functions of an autonomous decision-making method that combines a large model with self-playing technology by running or executing the computer programs and / or modules stored in the memory and accessing the data stored in the memory. The memory may primarily include a program storage area and a data storage area. The program storage area may store an operating system and at least one application required for a function (such as a sound playback function or a text conversion function); the data storage area may store data generated based on the use of the mobile phone (such as audio data and text message data). Furthermore, the memory may include high-speed random access memory and non-volatile memory, such as a hard disk, internal memory, a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, at least one disk storage device, a flash memory device, or other volatile solid-state storage device.

[0085] If the implemented module is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the present invention can implement all or part of the process steps in the above-mentioned method embodiments by using a computer program to instruct the relevant hardware. The computer program can be stored in a computer-readable storage medium. When executed by a processor, the computer program can implement the steps of each of the above-mentioned method embodiments. The computer program includes computer program code, which can be in source code form, object code form, executable file, or some intermediate form. The computer-readable medium can include: any entity or device capable of carrying the computer program code, recording medium, USB flash drive, mobile hard drive, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electric carrier signal, telecommunication signal, and software distribution medium. It should be noted that the content of the computer-readable medium can be appropriately increased or decreased based on the requirements of legislation and patent practice in a jurisdiction. For example, in some jurisdictions, based on legislation and patent practice, computer-readable media does not include electric carrier signals and telecommunication signals.

[0086] It should be noted that the device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed across multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the present embodiment. In addition, in the drawings of the device embodiments provided by the present invention, the connection relationship between the modules indicates that there is a communication connection between them, which may be specifically implemented as one or more communication buses or signal lines. A person of ordinary skill in the art can understand and implement the present invention without inventive effort.

[0087] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in the present invention should be included in the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be based on the scope of protection of the claims.

Claims

1. An autonomous decision-making method combining a large model and self-game technology, characterized in that: include: Acquiring multimodal data, processing the multimodal data using a large language model based on a MoE architecture, and combining the processed data with an initial prompt word to generate a structured decision input, wherein the multimodal data includes structured sensor data and unstructured data; Call the pre-built three-layer asynchronous hierarchical agent to process the structured decision input in sequence, wherein the three-layer asynchronous hierarchical agent includes a first-layer commander agent, a second-layer execution agent and a third-layer evaluation agent. The first-layer commander agent generates multiple candidate decision strategies based on the structured decision input and assigns them to the corresponding second-layer execution agent. Specifically, the first-layer commander agent receives the structured decision input, parses the available agent resource information and task requirement information, splits the complex task into multiple subtasks through the task decomposition algorithm, and generates N asynchronous parallel candidate strategy solutions based on resource constraints, time windows and risk assessments, wherein each strategy solution contains a task priority matrix, a resource allocation vector and an execution timing diagram; based on the agent capability description information in the structured decision input, dynamically constructs an intelligent agent. The agent capability matching matrix M(i, j) is constructed, where i represents the type and difficulty coefficient of the subtask, and j represents the type and capability parameters of the second-layer execution agent. M(i, j) calculates the similarity between the capability vector of agent j and the requirement vector of task i to obtain a fitness score. The KM algorithm is used to achieve optimal task allocation based on the agent capability matching matrix M(i, j), and distribute execution instructions containing specific execution parameters, constraints, and coordination interfaces to the corresponding second-layer execution agents. After receiving the execution instructions, the second-layer execution agents call the corresponding professional decision modules according to the task type in the instructions to generate specific action plans. The third-layer evaluation agent performs a collaborative evaluation on the decision plans generated by the second-layer execution agents. If the evaluation fails, feedback information is returned to the first-layer commander agent to regenerate the decision strategy. A simulated game is conducted on multiple candidate decision strategies that have passed the collaborative evaluation, and the execution results and scoring data of each strategy in the simulation environment are obtained. The execution results and scoring data are re-input into the large language model of the MoE architecture, and a fusion decision scheme is generated based on the empirical data of multiple strategies. The fusion decision scheme is simulated and evaluated, and a final decision is output when the score is greater than or equal to the highest score among the candidate decision strategies.

2. The autonomous decision-making method combining a large model and self-game technology according to claim 1, characterized in that: The multimodal data is obtained, the multimodal data is processed using a large language model of the MoE architecture, and the processed data is combined with the initial prompt word to generate a structured decision input, specifically: Obtaining multimodal data input, the multimodal data including structured real-time situation data and unstructured data, wherein the structured real-time situation data includes sensor data, location information, and state parameters, and the unstructured data includes image, audio, and video data; Combining the initial prompt words with the structured real-time situation data to form basic input data, wherein the initial prompt words include role settings, task objectives, constraints, precautions, output format, and historical decision actions; When unstructured data exists, the image, audio, and video data are first preprocessed using a multimodal large model, converted into descriptive information in text format, and then fused with the basic input data; The fused data is input into the large language model of the MoE architecture for cleaning and screening, key information related to the current task is extracted, and the structured decision input is generated.

3. The autonomous decision-making method combining a large model and self-game technology according to claim 1, characterized in that: The specific action plan is generated by calling the corresponding professional decision module according to the task type in the instruction, specifically: When the mission type is flight control, the aircraft control agent generates the flight path and maneuver strategy based on the flight dynamics model and trajectory planning algorithm; When the task type is multi-unit collaboration, the formation coordination agent uses a distributed consensus algorithm to handle the coordination problem and generate the formation and communication protocol; When the task type is a special operation, the functional module agent generates an operation plan based on the characteristics of the weapon system and the parameters of the detection equipment; Each second-layer execution agent shares decision progress and resource usage in real time through a preset state synchronization protocol. When a policy conflict or resource competition is detected, a local negotiation algorithm based on game theory is initiated to adjust the policy and generate a specific action plan that includes spatiotemporal coordinates, action sequences, resource requirements, and expected effects.

4. The autonomous decision-making method combining a large model and self-game technology according to claim 1, characterized in that: The third-layer evaluation agent performs a collaborative evaluation on the decision-making scheme generated by the second-layer execution agent. If the evaluation fails, feedback information is returned to the first-layer commander agent to regenerate the decision-making strategy. Specifically: The third-layer evaluation agent receives the specific action plan generated by the second-layer execution agent, performs a synergy analysis on the specific action plan, and calculates synergy effect evaluation indicators by comparing and analyzing the performance differences between the joint execution mode and the individual execution mode, wherein the synergy effect evaluation indicators include efficiency improvement rate, risk reduction rate, resource utilization rate, and goal achievement degree; The comprehensive synergy score of the synergy effect evaluation index is calculated by weighted fusion. When the comprehensive collaborative score is greater than or equal to the preset collaborative threshold, a simulated game will be conducted through a combination of specific action plans for collaborative evaluation. When the comprehensive collaborative score is less than the preset collaborative threshold, a feedback report containing a collaborative defect analysis is generated and returned to the first-level commander intelligent agent to regenerate candidate strategy plans.

5. An autonomous decision-making device combining a large model with self-game technology, characterized in that: include: a structured strategy generation unit, configured to acquire multimodal data, process the multimodal data using a large language model based on the MoE architecture, and combine the processed data with an initial prompt word to generate a structured decision input, wherein the multimodal data includes structured sensor data and unstructured data; The hierarchical agent processing unit is used to call a pre-built three-layer asynchronous hierarchical agent to process the structured decision input in sequence, wherein the three-layer asynchronous hierarchical agent includes a first-layer commander agent, a second-layer execution agent and a third-layer evaluation agent. The first-layer commander agent generates multiple candidate decision strategies based on the structured decision input and assigns them to the corresponding second-layer execution agent. Specifically, the first-layer commander agent receives the structured decision input, parses the available agent resource information and task requirement information, splits the complex task into multiple subtasks through the task decomposition algorithm, and generates N asynchronous parallel candidate strategy solutions based on resource constraints, time windows and risk assessments, wherein each strategy solution includes a task priority matrix, a resource allocation vector and an execution timing diagram; based on the agent capability description information in the structured decision input, Dynamically construct an agent capability matching matrix M(i, j), where i represents the type and difficulty coefficient of the subtask, and j represents the type and capability parameters of the second-layer execution agent. M(i, j) calculates the similarity between the capability vector of agent j and the requirement vector of task i to obtain a fitness score. A KM algorithm is used to achieve optimal task allocation based on the agent capability matching matrix M(i, j), and execute instructions containing specific execution parameters, constraints, and coordination interfaces are distributed to corresponding second-layer execution agents. After receiving the execution instructions, the second-layer execution agents call corresponding professional decision modules based on the task type in the instructions to generate specific action plans. The third-layer evaluation agent performs a collaborative evaluation on the decision plans generated by the second-layer execution agents. If the evaluation fails, feedback information is returned to the first-layer commander agent to regenerate the decision strategy. A decision generation unit is used to simulate the game of multiple candidate decision strategies that have passed the collaborative evaluation, obtain the execution results and scoring data of each strategy in the simulation environment, re-input the execution results and scoring data into the large language model of the MoE architecture, generate a fusion decision plan based on the empirical data of multiple strategies, simulate and evaluate the fusion decision plan, and output a final decision when the score is greater than or equal to the highest score among the candidate decision strategies.

6. An autonomous decision-making device that combines a large model with self-game technology, characterized in that: It includes a memory and a processor, wherein the memory stores a computer program, and the computer program can be executed by the processor to implement an autonomous decision-making method combining a large model and self-game technology as described in any one of claims 1 to 4.

7. A computer-readable storage medium, characterized in that A computer program is stored, and the computer program can be executed by a processor of the device where the computer-readable storage medium is located to implement an autonomous decision-making method that combines a large model and self-game technology as described in any one of claims 1 to 4.

Images