A multi-agent cooperative hunting method based on multiple large language models

Abstract

The present invention discloses a multi-agent cooperative roundup method based on multiple large language models, comprising: defining the state space and action space of the agents according to the roundup task; initializing the simulation environment, setting the initial position and state of each agent and the roundup object; selecting a large language model for each agent, and simultaneously obtaining or constructing a roundup task dataset; selecting a fine-tuning strategy adapted to the roundup task, and customizing the selected large language model; compiling a dedicated roundup prompt word template; using the fine-tuned roundup decision large model to control decision-making, thereby achieving multi-agent collaborative roundup; evaluating and optimizing the roundup decision large model; and deploying and testing the optimized roundup decision large model in a real environment. The present invention can effectively solve the problems faced in real roundup tasks, such as target uncertainty, complex and changeable environment, and time and resource constraints, and provides a more natural interaction method, stronger adaptability and flexibility, as well as higher interpretability and transparency, while reducing labor costs and achieving more efficient and intelligent automated collaboration.

Description

Technical Field

[0001] The present invention belongs to the technical field of artificial intelligence and multi-agent collaborative control systems, and in particular to a multi-agent cooperative capture method based on multiple large language models. Background Art

[0002] With the continuous advancement of science and technology, scholars' research on single agents has become increasingly mature. No longer content with optimizing the performance of a single agent, they have begun to explore how to further improve overall performance through the collaborative work of multiple agents. Therefore, the focus of research has gradually shifted from single-agent systems to multi-agent systems. One of the main application areas of multi-agent systems is multi-agent clusters. Multi-agent clusters can perform many functions, such as formation control, area traversal search, intelligent traffic management, and the capture of hostile targets. The multi-agent capture problem generally refers to a group of capturing agents moving towards an escaping agent, with the ultimate goal of encircling the escaping agent. As a comprehensive task, capture has broad application prospects in both civilian and military scenarios, such as post-disaster rescue, urban security, collaborative tracking, and maritime interception and capture.

[0003] However, the successful execution of roundup missions faces numerous challenges. First, there is the uncertainty of target movement. Targets may move in unpredictable ways, randomly changing their paths depending on the external environment and potential threats. Second, environmental complexity cannot be ignored. Real-world environments are often ever-changing, filled with obstacles, irregular terrain, and constantly changing environmental conditions. This requires the intelligent agent to possess advanced navigation capabilities and flexible decision-making mechanisms to cope with various unforeseen situations. Finally, time and resource constraints are also important factors that must be considered when executing roundup missions. In many application scenarios, roundup missions must be completed within strict time limits while also paying attention to energy and computing resource consumption. These constraints further increase the complexity of the task.

[0004] Current research approaches to the roundup problem can be broadly categorized into two categories: traditional methods and reinforcement learning. Traditional methods primarily rely on mathematical tools and bioinspired algorithms, which excel in theoretical analysis and mathematical proofs. However, these methods typically require precise physical models and PID control strategies, which can be limiting in certain practical applications, particularly when the environment is difficult to accurately model or exhibits significant variability.

[0005] In contrast, reinforcement learning methods exhibit unique advantages. They self-learn through interactions between agents and their environments, achieving end-to-end control strategies and thus possessing greater adaptability and flexibility. Methods such as Q-learning, deep Q-networks (DQNs), and multi-agent deep deterministic policy gradients (MADDPGs) have been widely studied and applied. Although reinforcement learning methods have performed well in certain specific scenarios, they still face a series of challenges that need to be addressed. These include: ensuring rapid convergence and excellent performance in complex and changing scenarios is particularly difficult; low sample utilization and high sampling costs in real-world environments make these algorithms difficult to generalize in practical applications; the exponential growth of the state-action space as the number of agents increases, making the training process more difficult; the design of the reward function plays a crucial role in real-world applications, requiring continuous parameter adjustment to avoid potential learning policy errors and prevent results from deviating significantly from expectations; and the generalization performance of the algorithm is also a major shortcoming that needs significant improvement.

[0006] In recent years, the rapid advancement of large language models (LLMs) has demonstrated tremendous potential for achieving human-like intelligence. This potential stems from their ability to leverage extensive training data and large model parameters. As research deepens, there is a growing trend toward using LLMs as a core driver to build intelligent agents capable of autonomous decision-making. Compared to reinforcement learning, agents based on LLMs offer a significant advantage: they are pre-trained on large amounts of data, contain rich knowledge, and can be fine-tuned to specific tasks using relatively small datasets. Furthermore, these LLM-based agents provide a natural language interface, enabling fluent communication with humans. This feature not only increases their flexibility but also significantly improves their interpretability.

[0007] In multi-agent hunting tasks, introducing a large language model to guide agent decision-making is an innovative approach. Its core concept is to replace traditional algorithmic models with a large language model. Specifically, through the large model's prompt word templates, the agent's observed environmental information and its own state are input into the large language model, which then makes decisions and guides the agent's actions. This approach has the following advantages: 1. Simplified algorithm design: Using a large language model as the decision-making core simplifies model design and implementation. Reinforcement learning algorithms typically require carefully designed reward functions and a large number of training samples. However, a large language model, with its pre-trained knowledge and reasoning capabilities, can achieve effective decision-making with fewer samples, eliminating the complex algorithm network design and debugging process. 2. End-to-end decision support: The large language model can handle the entire decision-making process, from perception to action, providing an end-to-end solution. This reduces the reliance on complex feature engineering and manually designed reward functions in traditional reinforcement learning. 3. Rapid adaptation: The large language model can quickly adapt to new environmental changes through prompt templates, improving the flexibility and efficiency of hunting tasks. Traditional reinforcement learning models, on the other hand, typically require extensive retraining and parameter adjustments when adapting to new environments. The adaptation process of reinforcement learning models relies on trial and error correction, which is not only time-consuming and computationally expensive, but can also lead to poor performance during the adaptation period. 4. Efficient Knowledge Transfer: The pre-trained nature of large language models enables efficient knowledge transfer between different tasks. Compared to reinforcement learning, which requires retraining for each new task, large language models can quickly adapt to new tasks through fine-tuning or directly applying existing knowledge. 5. Interpretable and Transparent Decisions: Traditional reinforcement learning models (especially deep reinforcement learning) are often viewed as black boxes, making their internal decision-making processes difficult to explain. This lack of transparency further complicates debugging and validating the behavior of reinforcement learning models. If the model exhibits abnormal behavior in certain situations, it can be difficult to trace and understand the root cause of the problem, which poses a significant challenge for applications requiring high confidence and safety. In contrast, large language models generate natural language descriptions, providing intuitive explanations that enable users to understand the agent's behavior and decisions. This enhanced interpretability and transparency not only improves the efficiency of system debugging and verification, but also strengthens user trust in the system, ensuring reliability and safety in safety-critical applications.

[0008] In summary, using large language models to assist multi-agent collaborative hunting decisions not only offers advantages in knowledge breadth and depth, but also offers significant advantages such as simplified algorithm design, end-to-end decision support, rapid adaptation to the environment, efficient knowledge transfer, and interpretability and transparency. These advantages bring greater flexibility and efficiency to agent decision-making, enhancing system reliability and security, making this an innovative and highly effective approach. Summary of the Invention

[0009] The purpose of the present invention is to address the problems existing in the above-mentioned prior art and provide a method for multi-agent cooperative roundup based on multiple large language models. The method innovatively uses large language models to replace traditional reinforcement learning algorithms for decision-making, thereby avoiding the problems of algorithm complexity, low sample efficiency, and poor adaptability in reinforcement learning. Through the powerful natural language processing and understanding capabilities of large language models, significant benefits are brought to multi-agent collaborative decision-making, including higher decision-making efficiency and stronger environmental adaptability. The effectiveness of task completion is evaluated by designing relevant indicators for multi-agent collaborative work, while considering the resource consumption of communication between agents and the decision-making process, and comprehensively evaluating the advantages and disadvantages of the method by considering the decision-making effect and resource cost. Finally, an innovative solution that is more suitable for multi-agent collaborative roundup tasks is obtained.

[0010] The technical solution to achieve the purpose of the present invention is: on the one hand, a multi-agent cooperative hunting method based on multiple large language models is provided, the method comprising:

[0011] Step 1: Define the state space and action space of the agent based on the roundup task;

[0012] Step 2: Initialize the simulation environment and set the initial positions and states of each agent and the captured object;

[0013] Step 3: Select a large language model for each agent as its corresponding large model for roundup decision-making, and simultaneously obtain or construct a roundup task dataset;

[0014] Step 4: Select a fine-tuning strategy suitable for the roundup task and perform customized fine-tuning on the roundup decision model;

[0015] Step 5: Write a dedicated roundup prompt word template for the fine-tuned roundup decision model;

[0016] Step 6: Use the fine-tuned large-scale capture decision model to control the decision-making and realize multi-agent collaborative capture;

[0017] Step 7: Evaluate and optimize the capture decision model based on the performance and results of the capture mission.

[0018] Step 8: Deploy the optimized model and algorithm to the actual environment for testing, monitor the performance of the intelligent agent in real time in the actual environment, and make timely adjustments and optimizations as needed.

[0019] Furthermore, step 1 specifically includes:

[0020] Step 1.1, determine the specific objectives of the roundup mission, using the multi-tuple shown in the following formula:

[0021] Target=(max S,min T,min R)

[0022] Where Target represents the specific goal of the roundup mission; S represents the roundup success rate, which is defined as the probability or ratio of successfully rounding up the target object. The formula is expressed as: T represents the average time required to successfully capture the target object, and the formula is: R represents the total amount of resources consumed to complete the task, including manpower, materials, and equipment. The formula is defined as: Resource consumption i , n' is the number of resources consumed by the task;

[0023] Step 1.2, determine the constraints of the roundup task, using the multi-tuple shown in the following formula:

[0024] Limit=(E space ,E obstacle ,A speed ,A acceleration ,A turning ,A perception ,C range ,C delay )

[0025] Among them, Limit represents the constraints of the roundup task; E space Indicates the activity space range of the agent and the captured object; E obstacle Indicates the position and size of static or dynamic obstacles i in the environment. The formula is: In two-dimensional space: E obstacle ={(x i ,y i ,l i ,w i )|i=1,2,…,n3}, where (x i ,y i ) is the center coordinate of obstacle i, l i ,w i is the length and width of obstacle i, n3 is the number of obstacles; in three-dimensional space: E obstacle ={(x i ,y i ,z i ,l i ,w i ,h i )|i=1,2,…,n}, where (x i ,y i ,z i ) is the center coordinate of obstacle i, l i ,w i ,h i is the length, width and height of obstacle i; Aspeed Represents the speed range of each agent, and the formula is expressed as: A speed ={v|v≤v max}, where v max is the maximum speed of the agent; A acceleration Represents the acceleration range of each agent, and the formula is expressed as: A acceleration ={a|a≤a max}, where a max is the maximum acceleration of the agent; A turning Represents the steering angle range of each agent, and the formula is expressed as: A turning ={θ|θ≤θ max}, where θ max is the maximum steering angle of the agent; A perception Represents the limited range of perception of each agent, and the formula is expressed as: A perception ={r|r≤r max}, where r refers to the perception radius with the agent itself as the center, r max is the maximum perception radius of the agent; C range represents the maximum communication distance between agents; C delay represents the communication delay between agents;

[0026] Step 1.3, determine the evaluation index of the roundup task, using the multi-tuple shown in the following formula:

[0027] Evaluate=(S,T,E,R,F)

[0028] Where Evaluate represents the evaluation index of the capture task; S represents the capture success rate, which is defined as the probability or ratio of successfully capturing the target object. The formula is expressed as follows: T represents the average time required to successfully capture the target object, and the formula is: E represents the ratio of effective actions in collaborative work of intelligent agents, and the formula is defined as: R represents the total amount of resources consumed to complete the task, including manpower, materials, and equipment. The formula is defined as: Resource consumption i ; F represents a function that comprehensively measures success rate, average roundup time, coordination efficiency, and resource consumption. The formula is defined as: F = αS - βT + γE - δR, where α, β, γ, and δ are the weights of each indicator respectively;

[0029] Step 1.4: Define the state space S of the agent according to the specific target Target and constraint Limit of the roundup task. i =(p i ,v i ,θ i,o i ,r i ); where p i Represents the current position of agent i in the environment. In two-dimensional space, p i =(x i ,y i ), in three-dimensional space, p i =(x i ,y i ,z i ), where x i ,y i ,z i Represents horizontal, vertical coordinates and height respectively; v i Represents the moving speed of agent i, the formula is expressed as: v i =(v xi ,v yi ) or v i =(v xi ,v yi ,v zi ), v xi ,v yi ,v zi Respectively represent the moving speed of agent i on the x, y, and z axes; θ i Indicates the orientation angle of agent i, expressed as a scalar; o i It represents the environmental information within the perception range of agent i, including the position, speed, direction angle, environmental information and remaining resources of other agents j, as well as the position, speed and direction angle of the capture object o. The formula is: i ={(p j ,v j ,θ j ,o j ,r j )|j=1,2,…,k}∪{(p o ,v o ,θ o )}, k represents the number of other agents; r i represents the remaining resources of agent i;

[0030] Step 1.5: Define the action space A of the agent based on the constraints of the capture task. i ={a ai ,a θi ,a vi ,a ci}, where A i represents the action space of agent i, a ai Represents the actions that each agent i can choose: in two-dimensional space, a ai∈{forward, backward, left, right}, in three-dimensional space, a ai ∈{forward, backward, left, right, up, down}; a θi Indicates changing the orientation angle of agent i: In two-dimensional space, a θi ∈{turn left Δθ, turn right Δθ}, in three-dimensional space a θi ∈{turn left Δθ, turn right Δθ, turn up Δθ, turn down Δθ}; a vi Indicates changing the speed of agent i: a vi ∈{accelerate, decelerate, stop}; a ci Indicates communication action: a ci ∈{send information, receive information, update own status}.

[0031] Furthermore, step 2 specifically includes:

[0032] Step 2.1, build a simulation environment to simulate the roundup task;

[0033] Step 2.2: In the simulation environment, set the initial state space S for each agent i. i0 =(ρ i0 ,v i0 ,θ i0 ,o i0 ,r i0 ), set the initial position p for the captured object t t0 =(x t0 ,y t0 ), initial velocity v t0 =(v x0 ,v y0 ), initial deflection direction θ t0 ; Among them, p i0 represents the initial position of agent i in the environment, v i0 represents the initial moving speed of agent i, θ i0 represents the initial orientation angle of agent i, o i0 represents the initial environment information within the perception range of agent i, r i0 represents the initial remaining resources of agent i; x t0 ,y t0 They represent the initial horizontal and vertical coordinates of the captured object t, v x0 ,v y0 Respectively represent the initial moving speed of the captured object t on the x and y axes.

[0034] Furthermore, step 3 specifically includes:

[0035] Step 3.1: Before starting the roundup task, check whether there is any prior expert knowledge or dataset related to the current roundup task, and then search and evaluate whether there is an existing dataset that can be used for this roundup task;

[0036] The capture task dataset refers to the collected dataset D={(d1, d2, ..., d n1 )}, where each data item d i =(r i ',e i ,m i ,t i ), where r i 'Indicates the execution results of the past roundup mission, e i Represents the characteristics of the task execution environment, m i Indicates the behavior pattern of the target object in different situations, t i Represents the action trajectory of the agent during execution;

[0037] Define the evaluation of each dataset D i Compatible i As follows: Compatible i =ω1*E i +ω2*Q i +ω3*B i , where E i is the data content matching score, Q i is the data quality score, B i is the data applicability score, ω1, ω2, ω3 are the corresponding weights, ω1 ≥ 0, ω2 ≥ 0, ω3 ≥ 0, and the sum of the three weights is 1;

[0038] In the case of only one existing dataset, determine whether the applicability of the dataset exceeds the preset threshold ε. If so, it indicates that the dataset can be used for this round-up task. Otherwise, it is considered that there is no relevant dataset. If there are multiple existing datasets, determine whether the applicability of each dataset exceeds the preset threshold ε. Then, sort and select the dataset in descending order according to the applicability, and select it for this round-up task.

[0039] Step 3.2: If there is prior expert knowledge or data set, perform the following process;

[0040] For the dataset that can be applied to this round-up task:

[0041] (1) Perform data cleaning operations;

[0042] (2) Label the data, from the dataset D = {(d1, d2, ..., d n1 )} extract each data item di The execution result of the roundup task in r i and the agent's trajectory t i , and then from the action trajectory t i Extract the state space S of each agent i and action space A i ;

[0043] Then, based on the specific target, constraint, and evaluation indicators of the roundup task, a corresponding large language model is selected for each agent as the roundup decision model;

[0044] Otherwise, perform the following steps;

[0045] Step 3.3: Based on the specific target, constraint, and evaluation metric of the roundup task, select the corresponding large language model for each agent as the roundup decision model.

[0046] Step 3.4: Construct a prompt word template for a large language model suitable for the roundup task;

[0047] Step 3.5, using the large capture decision model to make control decisions, perform multiple rounds of multi-agent collaborative capture to generate and collect capture data;

[0048] In step 3.6, the collected data from multiple rounds of capture are cleaned, and then the recorded data are labeled. The state space and action space of each agent at each step, as well as the position, speed, and deflection direction of the captured object at each step are sorted and extracted from the recorded data.

[0049] Furthermore, in steps 3.2 and 3.3, the corresponding adapted large language model is selected for each agent as the large model for roundup decision making, specifically including:

[0050] For each candidate large language model M i , its comprehensive score Score i Defined as: Score i =ω1*F i +ω2*P i +ω3*C i , where F i It is the model function fitness score, which evaluates whether the model's functions meet the requirements of the roundup task, including natural language understanding, generation ability, and context processing ability. The formula is defined as: where f ij It is a large language model M i The fitness score of the jth function, n2 is the number of functions; P iIs the model performance score, which evaluates the performance of the model. The formula is defined as: where R i is the response speed score, D i is the model accuracy score, α and β are the corresponding weights; C i is the model cost score, which evaluates the cost of the model and is defined as: Where cost i is the actual cost of the model, and max cost is the highest cost among all candidate large language models; w1, w2, and w3 are the corresponding score weights, and the sum of the three weights is 1;

[0051] Select comprehensive score Score i A high-quality large language model is used as a large model for roundup decision making.

[0052] Furthermore, step 4 specifically includes:

[0053] Step 4.1, Fine-tuning method selection: Based on existing large language model fine-tuning techniques, consider the specific target, constraint, evaluation metric, and the size and quality of the task dataset D, and select an appropriate large language model fine-tuning method.

[0054] Step 4.2, Configure the training environment: Configure GPU / TPU resources to ensure that the large-scale roundup decision model can handle large-scale decision data and real-time training requirements. Also, select a framework that supports efficient fine-tuning and inference to ensure that the framework can handle the complexity of the roundup task.

[0055] Step 4.3, loading the pre-trained model: loading the large model for roundup decision-making;

[0056] Step 4.4, data loading and processing: Divide the roundup task dataset into training set and test set;

[0057] Step 4.5, set fine-tuning parameters: Based on the fine-tuning method determined in step 4.1, set the fine-tuning hyperparameters, including learning rate, batch size, and number of training rounds;

[0058] Step 4.6, fine-tuning: using the training set to train the large model for round-up decision-making for multiple rounds, and evaluating the model performance on the test set, recording key indicators;

[0059] Step 4.7, model optimization and parameter update: Based on the evaluation results, adjust the hyperparameters and fine-tuning methods to further optimize the performance of the large roundup decision-making model in the roundup decision-making task until the optimization end conditions are met.

[0060] Furthermore, the prompt word templates in steps 3.4 and 5 are defined as follows:

[0061] T'=(D',I,H,R',F in ,F out )

[0062] Among them, T' is the prompt word template, D' is the task description module, I is the identity analysis module, H is the historical information description module, R' is the reasoning guidance process module, F in is a fixed input format, F out Is a fixed output format;

[0063] (1) Task description module D': D' = "capture mission background + capture mission objectives + specific capture environment + rules and regulations during the capture process";

[0064] (2) Identity Analysis Module I: I = “Identity Analysis + Responsibilities / Goals”;

[0065] (3) Historical information description module H: H = "historical information + historical decision";

[0066] (4) Reasoning guidance process module R': R' = "Now please think carefully about the current environment information and the state of the agent. Based on historical information and environmental changes, consider the following factors: the position and speed of the target object, the current position, speed and direction of the agent, and obstacles in the environment, and infer the optimal next action";

[0067] (5) Fixed input format F in :F in = "Current position p of agent i i , speed v i , direction θ i , perception information i and the remaining resources r i , the position p of the captured object t t , speed v t and the deflection direction θ t ”;

[0068] (6) Fixed output format F out :F out = "Next decision: {a ai ,a θi ,a vi ,a ci}, where a ai is the action that each agent i can choose, a θi is to change the orientation angle of agent i, a vi is the speed of changing agent i, a ci It is a communication action.

[0069] Furthermore, the multi-agent collaborative roundup described in step 3.5 and step 6 specifically includes:

[0070] (1) Initialization: Set the initial state S of agent i i0 =(p i0 ,v i0 ,θ i0 ,o i0 ,r i0 ) and the initial position p of the captured object t t0 =(x t0 ,y t0 ), initial velocity v t0 =(v x0 ,v y0 ) and the initial deflection direction θ t0 , and record;

[0071] (2) Observe the environment: The agent observes the surrounding environment and calculates the current position of the agent i p i =(x i ,y i ), speed v i =(v xi ,v yi ), direction θ i , perception information o i ={(p j ,v j ,θ j ,o j ,r j )|j=1,2,…,k}, agent’s remaining resources r i , the position of the captured object p t =(x t ,y t ), speed v t =(v x ,v y ), deflection direction θ t Passed to the big model of roundup decision;

[0072] (3) Decision generation: According to the prompt word template T', the corresponding prompt word T is input to all the large models of roundup decision making. i , and get the decision output F corresponding to each agent outi Each agent makes a decision based on the output of the corresponding capture decision model F outi , execute the corresponding action A i ={a ai ,a θi ,a vi ,a ci}, and record the decision, that is, the execution action, and save the intermediate reasoning process of the large model of the roundup decision;

[0073] (4) State update: After executing the agent action, update the state space of the agent and the position, speed and direction of the captured object, and record the current state;

[0074] (5) Multiple rounds of loops: Repeat steps (2), (3), and (4) until the task is completed.

[0075] Furthermore, in step 7, the evaluation index of the comprehensive roundup task is Evaluate Task and reinforcement learning evaluation indicators Evaluate RL Evaluate the effectiveness of multiple rounds of actual round-up and the performance of multiple large round-up decision-making models, and optimize the prompt word templates and model parameters;

[0076] Among them, the evaluation index Evaluate Task =(S,T,E,R), where S represents the capture success rate, which is defined as the probability or ratio of successfully capturing the target object. The formula is: T represents the average time required to successfully capture the target object, and the formula is: E represents the ratio of effective actions in collaborative work of intelligent agents, and the formula is defined as: R represents the total amount of resources consumed to complete the task, including manpower, materials, and equipment. The formula is defined as: Resource consumption i ;

[0077] Reinforcement learning evaluation indicators Evaluate RL =(Rb,Ce), where Rb represents robustness and Ce represents model computational efficiency;

[0078] The comprehensive evaluation formula is: F'=αS-βT+γE-δR+w1Rb+w2Ce, where α, β, γ, δ, w1, and w2 are the weights of each indicator, and α+β+γ+δ+w1+w2=1, and F' represents the evaluation value.

[0079] Compared with the prior art, the present invention has the following significant advantages:

[0080] (1) The present invention uses a large language model as the decision-making core, which can simplify the design and implementation process of the model.

[0081] (2) The present invention can achieve good performance with a smaller amount of data, reducing the difficulty and cost of data collection.

[0082] (3) When the present invention encounters a situation in which the new environment does not fully match the training data, it can make inferences and decisions based on its own language knowledge and structural understanding, making it more suitable for complex and changing environments.

[0083] (4) The present invention can help explain the decision-making process of the intelligent agent, making the decision more transparent and explainable.

[0084] The present invention is further described in detail below with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0085] Figure 1 This is a flowchart of the multi-agent cooperative hunting method based on multiple large language models of the present invention.

[0086] Figure 2 This is a flowchart of the present invention using multiple large language models to control decisions and perform multi-agent collaborative hunting.

[0087] Figure 3 Flowchart of the present invention for fine-tuning multiple large language models using the captured data collected from existing prior datasets / records. DETAILED DESCRIPTION

[0088] In order to make the purpose, technical solutions and advantages of this application more clear, the following further describes this application in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application.

[0089] It should be noted that if there are descriptions involving "first", "second", etc. in the embodiments of the present invention, the descriptions of "first", "second", etc. are only for descriptive purposes and cannot be understood as indicating or implying their relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first" and "second" may explicitly or implicitly include at least one of such features. In addition, the technical solutions between the various embodiments can be combined with each other, but they must be based on the ability of ordinary technicians in this field to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be deemed that such a combination of technical solutions does not exist and is not within the scope of protection required by the present invention.

[0090] In response to the problems of complex algorithms, high training data requirements, and poor environmental adaptability of traditional multi-agent collaborative hunting methods based on reinforcement learning, this paper proposes a collaborative hunting method based on collaborative decision-making of multiple large language model agents.

[0091] In one embodiment, combined Figure 1 , the method comprising:

[0092] Step 1: Define the capture mission and determine its specific objectives, constraints, and evaluation criteria. Based on the mission requirements, define the state space and possible action space of each agent.

[0093] Step 2: Initialize the simulation environment and set the initial positions and states of each agent and the captured object;

[0094] Step 3: Select a large language model for each agent as its corresponding large model for roundup decision-making, and simultaneously obtain or construct a roundup task dataset;

[0095] Step 4: Select a fine-tuning strategy suitable for the roundup task and perform customized fine-tuning on the roundup decision model;

[0096] Step 5: Write a dedicated roundup prompt word template for the fine-tuned roundup decision model;

[0097] Step 6: Use the fine-tuned large-scale capture decision model to control the decision-making and realize multi-agent collaborative capture;

[0098] Step 7: Evaluate and optimize the capture decision model based on the performance and results of the capture mission.

[0099] Step 8: Deploy the optimized model and algorithm to the actual environment for testing, monitor the performance of the intelligent agent in real time in the actual environment, and make timely adjustments and optimizations as needed.

[0100] Furthermore, in one embodiment, step 1 specifically includes:

[0101] Step 1.1, determine the specific objectives of the roundup mission, using the multi-tuple shown in the following formula:

[0102] Target=(max S,min T,min R)

[0103] Where Target represents the specific goal of the roundup mission; S represents the roundup success rate, which is defined as the probability or ratio of successfully rounding up the target object. The formula is expressed as: T represents the average time required to successfully capture the target object, and the formula is: R represents the total amount of resources consumed to complete the task, including manpower, materials, and equipment. The formula is defined as: Resource consumption i , n' is the number of resources consumed by the task;

[0104] Step 1.2, determine the constraints of the roundup task, using the multi-tuple shown in the following formula:

[0105] Limit=(E space ,E obstacle ,A speed ,A acceleration ,A turning ,A perception ,C range ,C delay )

[0106] Among them, Limit represents the constraints of the roundup task; E space Indicates the activity space range of the agent and the captured object; E obstacle Indicates the position and size of static or dynamic obstacles i in the environment. The formula is: In two-dimensional space: E obstacle ={(x i ,y i ,l i ,w i )|i=1,2,…,n3}, where (x i ,y i ) is the center coordinate of obstacle i, l i ,w i is the length and width of obstacle i, n3 is the number of obstacles; in three-dimensional space: E obstacle ={(x i ,y i ,z i ,l i ,w i ,h i )|i=1,2,…,n}, where (x i ,y i ,z i ) is the center coordinate of obstacle i, l i ,w i ,h i is the length, width and height of obstacle i; A speed Represents the speed range of each agent, and the formula is expressed as: A speed ={v|v≤v max}, where v max is the maximum speed of the agent; A acceleration Represents the acceleration range of each agent, and the formula is expressed as: A acceleration ={a|a≤a max}, where a max is the maximum acceleration of the agent; A turning Represents the steering angle range of each agent, and the formula is expressed as: A turning ={θ|θ≤θ max}, where θ max is the maximum steering angle of the agent; A perception Represents the limited range of perception of each agent, and the formula is expressed as: A perception ={r|r≤r max}, where r refers to the perception radius with the agent itself as the center, r max is the maximum perception radius of the agent; C range represents the maximum communication distance between agents; C delayrepresents the communication delay between agents;

[0107] Step 1.3, determine the evaluation index of the roundup task, using the multi-tuple shown in the following formula:

[0108] Evaluate=(S,T,E,R,F)

[0109] Where Evaluate represents the evaluation index of the capture task; S represents the capture success rate, which is defined as the probability or ratio of successfully capturing the target object. The formula is expressed as follows: T represents the average time required to successfully capture the target object, and the formula is: E represents the ratio of effective actions in collaborative work of intelligent agents, and the formula is defined as: R represents the total amount of resources consumed to complete the task, including manpower, materials, and equipment. The formula is defined as: Resource consumption i ; F represents a function that comprehensively measures success rate, average roundup time, coordination efficiency, and resource consumption. The formula is defined as: F = αS - βT + γE - δR, where α, β, γ, and δ are the weights of each indicator respectively;

[0110] Step 1.4: Define the state space S of the agent according to the specific target Target and constraint Limit of the roundup task. i =(p i ,v i ,θ i ,o i ,r i ); where p i Represents the current position of agent i in the environment. In two-dimensional space, p i =(x i ,y i ), in three-dimensional space, p i =(x i ,y i ,z i ), where x i ,y i ,z i Represents horizontal, vertical coordinates and height respectively; v i Represents the moving speed of agent i, the formula is expressed as: v i =(v xi ,v yi ) or v i =(v xi ,v yi ,v zi ), v xi ,v yi ,v ziRespectively represent the moving speed of agent i on the x, y, and z axes; θ i Indicates the orientation angle of agent i, expressed as a scalar; o i It represents the environmental information within the perception range of agent i, including the position, speed, direction angle, environmental information and remaining resources of other agents j, as well as the position, speed and direction angle of the capture object o. The formula is: i ={(p j ,v j ,θ j ,o j ,r j )|j=1,2,…,k}∪{(p o ,v o ,θ o )}, k represents the number of other agents; r i represents the remaining resources of agent i;

[0111] Step 1.5: Define the action space A of the agent based on the constraints of the capture task. i ={a ai ,a θi ,a vi ,a ci}, where A i represents the action space of agent i, a ai Represents the actions that each agent i can choose: in two-dimensional space, a ai ∈{forward, backward, left, right}, in three-dimensional space, a ai ∈{forward, backward, left, right, up, down}; a θi Indicates changing the orientation angle of agent i: In two-dimensional space, a θi ∈{turn left Δθ, turn right Δθ}, in three-dimensional space a θi ∈{turn left Δθ, turn right Δθ, turn up Δθ, turn down Δθ}; a vi Indicates changing the speed of agent i: a vi ∈{accelerate, decelerate, stop}; a ci Indicates communication action: a ci ∈{send information, receive information, update own status}.

[0112] Furthermore, in one embodiment, step 2 specifically includes:

[0113] Step 2.1, build a simulation environment for simulating the roundup mission through coding or simulation software;

[0114] Step 2.2: In the simulation environment, set the initial state space S for each agent i. i0 =(pi0 ,v i0 ,θ i0 ,o i0 ,r i0 ), set the initial position p for the captured object t t0 =(x t0 ,y t0 ), initial velocity v t0 =(v x0 ,v y0 ), initial deflection direction θ t0 ; Among them, p i0 represents the initial position of agent i in the environment, v i0 represents the initial moving speed of agent i, θ i0 represents the initial orientation angle of agent i, o i0 represents the initial environment information within the perception range of agent i, r i0 represents the initial remaining resources of agent i; x t0 ,y t0 They represent the initial horizontal and vertical coordinates of the captured object t, v x0 ,v y0 Respectively represent the initial moving speed of the captured object t on the x and y axes.

[0115] Furthermore, in one embodiment, step 3 specifically includes:

[0116] Step 3.1: Before starting the roundup task, check whether there is any prior expert knowledge or dataset related to the current roundup task, and then search and evaluate whether there is an existing dataset that can be used for this roundup task;

[0117] Here, prior expert knowledge refers to the knowledge accumulated over a long period of time in the field of roundup tasks by experts in the field of military strategy and tactics, law enforcement personnel of law enforcement agencies, and relevant researchers in multi-agent system research. This knowledge includes: the identity, behavior patterns, and activity range of the target object; the best time and method for rounding up the target object; possible emergencies and response strategies, such as the target object escaping or resisting.

[0118] The capture task dataset refers to the collected dataset D={(d1, d2, ..., d n1 )}, where each data item d i =(r i ',e i ,m i ,t i ), where r i 'Indicates the execution results of the past roundup mission, e i Represents the characteristics of the task execution environment, m iIndicates the behavior pattern of the target object in different situations, t i Represents the action trajectory of the agent during execution;

[0119] In order to quantify the evaluation process, a simple scoring system is used to evaluate the suitability of the dataset. The evaluation of each dataset D is defined as i Compatible i As follows: Compatible i =ω1*E i +ω2*Q i +ω3*B i , where E i is the data content matching score, Q i is the data quality score, B i is the data applicability score, ω1, ω2, ω3 are the corresponding weights, ω1 ≥ 0, ω2 ≥ 0, ω3 ≥ 0, and the sum of the three weights is 1;

[0120] In the case of only one existing dataset, determine whether the applicability of the dataset exceeds the preset threshold ε. If so, it indicates that the dataset can be used for this round-up task. Otherwise, it is considered that there is no relevant dataset. If there are multiple existing datasets, determine whether the applicability of each dataset exceeds the preset threshold ε. Then, sort and select the dataset in descending order according to the applicability, and select it for this round-up task.

[0121] Step 3.2: If there is prior expert knowledge or data set, perform the following process;

[0122] For the dataset that can be applied to this round-up task:

[0123] (1) Perform data cleaning operations;

[0124] (2) Label the data, from the dataset D = {(d1, d2, ..., d n1 )} extract each data item d i The execution result of the roundup task in r i and the agent's trajectory t i , and then from the action trajectory t i Extract the state space S of each agent i and action space A i ;

[0125] Then, based on the specific target, constraint, and evaluation indicators of the roundup task, a corresponding large language model is selected for each agent as the roundup decision model;

[0126] Otherwise, perform the following steps;

[0127] Step 3.3: Based on the specific target, constraint, and evaluation metric of the roundup task, select the corresponding large language model for each agent as the roundup decision model.

[0128] Considering the wide range of options available for large language models, both open source and private, a comprehensive and detailed evaluation of large language models is essential. To select a suitable large language model for the roundup task, a systematic evaluation formula is required to quantify the performance and applicability of each model. The evaluation criteria should comprehensively consider factors such as the model's functionality, performance, cost, and suitability for the roundup task.

[0129] For each candidate large language model M i , its comprehensive score Score i Defined as: Score i =ω1*F i +ω2*P i +ω3*C i , where F i It is the model function fitness score, which evaluates whether the model's functions meet the requirements of the roundup task, including natural language understanding, generation ability, context processing ability, etc. The formula is defined as: where f ij It is a large language model M i The fitness score of the jth function, n2 is the number of functions; P i Is the model performance score, which evaluates the performance of the model. The formula is defined as: where R i is the response speed score, D i is the model accuracy score, α and β are the corresponding weights; C i is the model cost score, which evaluates the cost of the model and is defined as: Where cost i is the actual cost of the model, and max cost is the highest cost among all candidate large language models; w1, w2, and w3 are the corresponding score weights, and the sum of the three weights is 1;

[0130] Select comprehensive score Score i A high-quality large language model is used as a large model for roundup decision making.

[0131] Step 3.4: Construct a prompt word template for a large language model suitable for the roundup task;

[0132] Step 3.5, using the large capture decision model to make control decisions, perform multiple rounds of multi-agent collaborative capture to generate and collect capture data;

[0133] In step 3.6, the collected data from multiple rounds of capture are cleaned, and then the recorded data are labeled. The state space and action space of each agent at each step, as well as the position, speed, and deflection direction of the captured object at each step are sorted and extracted from the recorded data.

[0134] Here, data cleaning operations include removing noise and outliers in the data set to ensure the integrity and accuracy of the data; removing invalid data, that is, rounding up unsuccessful data to ensure the quality of the data set.

[0135] Here, during the labeling process, the consistency and accuracy of the data are ensured to facilitate subsequent training and evaluation.

[0136] Furthermore, in one embodiment, in combination Figure 3 , step 4 specifically includes:

[0137] Step 4.1, Fine-tuning method selection: Based on existing large language model fine-tuning techniques, consider the specific target, constraint, evaluation metric, and the size and quality of the task dataset D, and select an appropriate large language model fine-tuning method.

[0138] Step 4.2, Configure the training environment: Configure sufficient GPU / TPU resources to ensure that the large-scale roundup decision model can handle large-scale decision data and real-time training requirements. Also, select a framework that supports efficient fine-tuning and inference, such as TensorFlow or PyTorch, to ensure that the framework can handle the complexity of the roundup task.

[0139] Step 4.3, loading the pre-trained model: loading the large model for roundup decision-making;

[0140] Step 4.4, data loading and processing: Divide the roundup task dataset into a training set (60%) and a test set (40%) to ensure fairness in model training and evaluation;

[0141] Step 4.5, set fine-tuning parameters: Based on the fine-tuning method determined in step 4.1, set the fine-tuning hyperparameters, including learning rate, batch size, and number of training rounds;

[0142] Step 4.6, fine-tuning: Use the training set to train the large round-up decision model for multiple rounds, and evaluate the model performance on the test set, recording key indicators (such as loss, accuracy, etc.);

[0143] Step 4.7, model optimization and parameter update: Based on the evaluation results, adjust the hyperparameters and fine-tuning methods to further optimize the performance of the large model in the roundup decision task until the optimization end conditions are met.

[0144] Furthermore, in one embodiment, to create a prompt word template for a large language model suitable for a roundup task, the task description, identity analysis, historical information, and reasoning guidance process can be logically linked together to form a complete prompt word template. This content should follow a fixed input and output format to ensure that the large language model can effectively understand and make decisions.

[0145] The prompt word templates in steps 3.4 and 5 are defined as follows:

[0146] T'=(D',I,H,R',F in ,F out )

[0147] Among them, T' is the prompt word template, D' is the task description module, I is the identity analysis module, H is the historical information description module, R' is the reasoning guidance process module, F in is a fixed input format, F out Is a fixed output format;

[0148] (1) Task description module D': provides detailed task background, objectives, environment information and rules, etc. D' = "capture task background + capture task objectives + specific capture environment + capture process rules";

[0149] (2) Identity Analysis Module I: Helps the model identify the roles and responsibilities of agents. I = "Identity Analysis (You are agent i, and your role in the team is decision-making) i ) + responsibilities / goals";

[0150] (3) Historical information description module H: Since the input of the large language model has a token length limit, historical information can be organized and passed to the large language model according to actual needs to help the large language model better understand the current strategy. H = "historical information + historical decision";

[0151] (4) Reasoning Guidance Process Module R': Based on existing research on large language model prompt word engineering, in order for large language models to better guide decision-making, all current factors should be fully considered and the relevant thought chain reasoning process should be compiled. R' = "Now please think carefully about the current environmental information and the state of the agent. Based on historical information and environmental changes, consider the following factors: the position and speed of the target object, the current position, speed and direction of the agent, and obstacles in the environment, and infer the optimal next action."

[0152] (5) Fixed input format F in : Define the format and content of the input, including the state of the agent and the captured object, the observed environment information, etc., to ensure that the model can understand and process the input information. in = "Current position p of agent ii , speed v i , direction θ i , perception information i and the remaining resources r i , the position p of the captured object t t , speed v t and the deflection direction θ t ”;

[0153] (6) Fixed output format F out :Define the format and content of the large language model output, such as the next action of the intelligent agent, to ensure that the decision output by the large language model can be directly used to control the intelligent agent. out = "Next decision: {a ai ,a θi ,a vi ,a ci}, where a ai is the action that each agent i can choose, a θi is to change the orientation angle of agent i, a vi is the speed of changing agent i, a ci It is a communication action.

[0154] Furthermore, in one embodiment, in combination Figure 2 , explain the steps for implementing multi-agent collaborative roundup:

[0155] (1) Initialization: Set the initial state S of agent i i0 =(p i0 ,v i0 ,θ i0 ,o i0 ,r i0 ) and the initial position p of the captured object t t0 =(x t0 ,y t0 ), initial velocity v t0 =(v x0 ,v y0 ) and the initial deflection direction θ t0 , and record;

[0156] (2) Observe the environment: The agent observes the surrounding environment and calculates the current position of the agent i p i =(x i ,y i ), speed v i =(v xi ,v yi ), direction θ i , perception information o i ={(p j ,v j ,θj ,o j ,r j )|j=1,2,…,k}, agent’s remaining resources r i , the position of the captured object p t =(x t ,y t ), speed v t =(v x ,v y ), deflection direction θ t Passed to the big model of roundup decision;

[0157] (3) Decision generation: According to the prompt word template T', the corresponding prompt word T is input to all the large models of roundup decision making. i , and get the decision output F corresponding to each agent outi Each agent makes a decision based on the output of the corresponding capture decision model F outi , execute the corresponding action A i ={a ai ,a θi ,a vi ,a ci}, and record the decision, that is, the execution action, and save the intermediate reasoning process of the large model of the roundup decision;

[0158] (4) Status update: After executing the agent action, update the agent's status information S' i =(p' i ,v' i ,θ' i ,o' i ,r' i ) and the position of the captured object p' t =(x' t ,y' t ), speed v' t =(v' x ,v' y ), direction θ' t , record the current status.

[0159] (5) Multiple rounds of loops: Repeat steps (2), (3), and (4) until the task is completed.

[0160] Furthermore, in one embodiment, in step 7, the evaluation index of the comprehensive roundup task is Evaluate Task and reinforcement learning evaluation indicators Evaluate RL Evaluate the effectiveness of multiple rounds of actual round-up and the performance of multiple large round-up decision-making models, and optimize the prompt word templates and model parameters;

[0161] Among them, the evaluation index EvaluateTask =(S,T,E,R), where S represents the capture success rate, which is defined as the probability or ratio of successfully capturing the target object. The formula is: T represents the average time required to successfully capture the target object, and the formula is: E represents the ratio of effective actions in collaborative work of intelligent agents, and the formula is defined as: R represents the total amount of resources consumed to complete the task, including manpower, materials, and equipment. The formula is defined as: Resource consumption i ;

[0162] Reinforcement learning evaluation indicators Evaluate RL =(Rb,Ce), where Rb represents robustness, reflecting the stability and adaptability of the model under different environments and conditions, and Ce represents the computational efficiency of the model, reflecting the time and computational resources required for the model to make decisions;

[0163] The comprehensive evaluation formula is: F'=αS-βT+γE-δR+w1Rb+w2Ce, where α, β, γ, δ, w1, and w2 are the weights of each indicator, which are adjusted according to specific task requirements, and α+β+γ+δ+w1+w2=1, and F' represents the evaluation value.

[0164] In one embodiment, a multi-agent cooperative hunting system based on multiple large language models is provided, the system comprising:

[0165] The first module is used to define the state space and action space of the intelligent agent according to the round-up task;

[0166] The second module is used to initialize the simulation environment and set the initial positions and states of each agent and the captured object;

[0167] The third module is used to select a large language model for each intelligent agent as its corresponding large model for roundup decision-making, and at the same time obtain or construct a roundup task dataset;

[0168] The fourth module is used to select a fine-tuning strategy suitable for the roundup task and perform customized fine-tuning on the roundup decision model;

[0169] The fifth module is used to write a dedicated roundup prompt word template for the fine-tuned roundup decision model;

[0170] The sixth module is used to control the decision-making by using the fine-tuned large-scale roundup decision model to achieve multi-agent collaborative roundups;

[0171] The seventh module is used to evaluate and optimize the capture decision-making model based on the performance and results of the capture mission;

[0172] The eighth module is used to deploy the optimized model and algorithm into the actual environment for testing, monitor the performance of the intelligent agent in real time in the actual environment, and make timely adjustments and optimizations as needed.

[0173] Regarding the specific definition of the multi-agent cooperative roundup system based on multiple large language models, please refer to the definition of the multi-agent cooperative roundup method based on multiple large language models above, which will not be repeated here. Each module in the above-mentioned multi-agent cooperative roundup system based on multiple large language models can be implemented in whole or in part by software, hardware and a combination thereof. The above-mentioned modules can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory of the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above modules.

[0174] In one embodiment, a computer device is provided, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the following is achieved:

[0175] Step 1: Define the state space and action space of the agent based on the roundup task;

[0176] Step 2: Initialize the simulation environment and set the initial positions and states of each agent and the captured object;

[0177] Step 3: Select a large language model for each agent as its corresponding large model for roundup decision-making, and simultaneously obtain or construct a roundup task dataset;

[0178] Step 4: Select a fine-tuning strategy suitable for the roundup task and perform customized fine-tuning on the roundup decision model;

[0179] Step 5: Write a dedicated roundup prompt word template for the fine-tuned roundup decision model;

[0180] Step 6: Use the fine-tuned large-scale capture decision model to control the decision-making and realize multi-agent collaborative capture;

[0181] Step 7: Evaluate and optimize the capture decision model based on the performance and results of the capture mission.

[0182] Step 8: Deploy the optimized model and algorithm to the actual environment for testing, monitor the performance of the intelligent agent in real time in the actual environment, and make timely adjustments and optimizations as needed.

[0183] For the specific limitations of each step, please refer to the limitations of the multi-agent cooperative hunting method based on multiple large language models mentioned above, which will not be repeated here.

[0184] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored, and when the computer program is executed by a processor, the computer program implements:

[0185] Step 1: Define the state space and action space of the agent based on the roundup task;

[0186] Step 2: Initialize the simulation environment and set the initial positions and states of each agent and the captured object;

[0187] Step 3: Select a large language model for each agent as its corresponding large model for roundup decision-making, and simultaneously obtain or construct a roundup task dataset;

[0188] Step 4: Select a fine-tuning strategy suitable for the roundup task and perform customized fine-tuning on the roundup decision model;

[0189] Step 5: Write a dedicated roundup prompt word template for the fine-tuned roundup decision model;

[0190] Step 6: Use the fine-tuned large-scale capture decision model to control the decision-making and realize multi-agent collaborative capture;

[0191] Step 7: Evaluate and optimize the capture decision model based on the performance and results of the capture mission.

[0192] Step 8: Deploy the optimized model and algorithm to the actual environment for testing, monitor the performance of the intelligent agent in real time in the actual environment, and make timely adjustments and optimizations as needed.

[0193] For the specific limitations of each step, please refer to the limitations of the multi-agent cooperative hunting method based on multiple large language models mentioned above, which will not be repeated here.

[0194] This invention can effectively solve problems such as target uncertainty, complex and changeable environment, and time and resource constraints faced in real-world roundup missions. It provides a more natural interaction method, stronger adaptability and flexibility, as well as higher explainability and transparency, while reducing labor costs and achieving more efficient and intelligent automated collaboration.

[0195] The above shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The above embodiments and descriptions are only illustrative of the principles of the present invention. Without departing from the spirit and scope of the present invention, any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention shall be included in the scope of protection of the present invention.

Claims

1. A multi-agent cooperative hunting method based on multiple large language models, characterized by: The method comprises: Step 1: Define the state space and action space of the agent based on the roundup task; Step 2: Initialize the simulation environment and set the initial positions and states of each agent and the captured object; Step 3: Select a large language model for each agent as its corresponding large model for roundup decision-making, and simultaneously obtain or construct a roundup task dataset; Step 4: Select a fine-tuning strategy suitable for the roundup task and perform customized fine-tuning on the roundup decision model; Step 5: Write a dedicated roundup prompt word template for the fine-tuned roundup decision model; Step 6: Use the fine-tuned capture decision-making model to control the decision-making and realize multi-agent collaborative capture; according to the capture prompt word template T', input the corresponding prompt word T to all capture decision-making models. i , and get the decision output F corresponding to each agent outi Each agent makes a decision based on the output of the corresponding capture decision model F outi , execute the corresponding action; Step 7: Evaluate and optimize the capture decision model based on the performance and results of the capture mission. Step 8: Deploy the optimized model and algorithm to the actual environment for testing, monitor the performance of the intelligent agent in real time in the actual environment, and make timely adjustments and optimizations as needed.

2. The multi-agent cooperative hunting method based on multiple large language models according to claim 1 is characterized in that: Step 1 specifically includes: Step 1.1, determine the specific objectives of the roundup mission, using the multi-tuple shown in the following formula: Target=(max S,min T,min R) Where Target represents the specific target of the roundup mission; S represents the roundup success rate, which is defined as the probability or ratio of successfully rounding up the target object. The formula is expressed as: T represents the average time required to successfully capture the target object, and the formula is: R represents the total amount of resources consumed to complete the task, including manpower, materials, and equipment. The formula is defined as: n' is the number of resources consumed by the task; Step 1.2, determine the constraints of the roundup task, using the multi-tuple shown in the following formula: Limit=(E space ,E obstacle ,A speed ,A acceleration ,A turning ,A perception ,C range ,C delay ) Among them, Limit represents the constraints of the roundup task; E space Indicates the activity space range of the agent and the captured object; E obstacle Indicates the position and size of static or dynamic obstacles i in the environment. The formula is: In two-dimensional space: E obstacle ={(x i ,y i ,l i ,w i )|i=1,2,…,n3}, where (x i ,y i ) is the center coordinate of obstacle i, l i ,w i is the length and width of obstacle i, n3 is the number of obstacles; in three-dimensional space: E obstacle ={(x i ,y i ,z i ,l i ,w i ,h i )|i=1,2,…,n}, where (x i ,y i ,z i ) is the center coordinate of obstacle i, l i ,w i ,h i is the length, width and height of obstacle i; A speed Represents the speed range of each agent, and the formula is expressed as: A speed ={v|v≤v max }, where v max is the maximum speed of the agent; A acceleration Represents the acceleration range of each agent, and the formula is expressed as: A acceleration ={a|a≤a max }, where a max is the maximum acceleration of the agent; A turning Represents the steering angle range of each agent, and the formula is expressed as: A turning ={θ|θ≤θ max }, where θ max is the maximum steering angle of the agent; A perception Represents the limited range of perception of each agent, and the formula is expressed as: A perception ={r|r≤r max }, where r refers to the perception radius with the agent itself as the center, r max is the maximum perception radius of the agent; C range represents the maximum communication distance between agents; C delay represents the communication delay between agents; Step 1.3, determine the evaluation index of the roundup task, using the multi-tuple shown in the following formula: Evaluate=(S,T,E,R,F) Where Evaluate represents the evaluation index of the capture task; S represents the capture success rate, which is defined as the probability or ratio of successfully capturing the target object. The formula is expressed as follows: T represents the average time required to successfully capture the target object, and the formula is: E represents the ratio of effective actions in collaborative work of intelligent agents, and the formula is defined as: R represents the total amount of resources consumed to complete the task, including manpower, materials, and equipment. The formula is defined as: F represents a function that comprehensively measures success rate, average roundup time, coordination efficiency, and resource consumption. The formula is defined as: F = αS - βT + γE - δR, where α, β, γ, and δ are the weights of each indicator, respectively; Step 1.4: Define the state space S of the agent according to the specific target Target and constraint Limit of the roundup task. i =(p i ,v i , θi ,o i ,r i ); where p i Represents the current position of agent i in the environment. In two-dimensional space, p i =(x i ,y i ), in three-dimensional space, p i =(x i ,y i ,z i ), where x i ,y i ,z i Respectively represent the horizontal, vertical coordinates and height; v i Represents the moving speed of agent i, the formula is expressed as: v i =(v xi ,v yi ) or v i =(v xi ,v yi ,v zi ), v xi ,v yi ,v zi Respectively represent the moving speed of agent i on the x, y, and z axes; θ i Indicates the orientation angle of agent i, expressed as a scalar; o i It represents the environmental information within the perception range of agent i, including the position, speed, direction angle, environmental information and remaining resources of other agents j, as well as the position, speed and direction angle of the capture object o. The formula is: i ={(p j ,v j ,θ j ,o j ,r j )|j=1,2,…,k}∪{(p o ,v o ,θ o )}, k represents the number of other agents; r i represents the remaining resources of agent i; Step 1.5: Define the action space A of the agent based on the constraints of the capture task. i ={a ai ,a θi ,a vi ,a ci }, where A i represents the action space of agent i, a ai Represents the actions that each agent i can choose: in two-dimensional space, a ai ∈{forward, backward, left, right}, in three-dimensional space, a ai ∈{forward, backward, left, right, up, down}; a θi Indicates changing the orientation angle of agent i: In two-dimensional space, a θi ∈{turn left Δθ, turn right Δθ}, in three-dimensional space a θi ∈{turn left Δθ, turn right Δθ, turn up Δθ, turn down Δθ}; a vi Indicates changing the speed of agent i: a vi ∈{accelerate, decelerate, stop}; a ci Indicates communication action: a ci ∈{send information, receive information, update own status}.

3. The multi-agent cooperative hunting method based on multiple large language models according to claim 2 is characterized in that: Step 2 specifically includes: Step 2.1, build a simulation environment to simulate the roundup task; Step 2.2: In the simulation environment, set the initial state space S for each agent i. i0 =(p i0 ,v i0 ,θ i0 ,o i0 ,r i0 ), set the initial position p for the captured object t t0 =(x t0 ,y t0 ), initial velocity v t0 =(v x0 ,v y0 ), initial deflection direction θ t0 ; Among them, p i0 represents the initial position of agent i in the environment, v i0 represents the initial moving speed of agent i, θ i0 represents the initial orientation angle of agent i, o i0 represents the initial environment information within the perception range of agent i, r i0 represents the initial remaining resources of agent i; x t0 ,y t0 They represent the initial horizontal and vertical coordinates of the captured object t, v x0 ,v y0 Respectively represent the initial moving speed of the captured object t on the x and y axes.

4. The multi-agent cooperative hunting method based on multiple large language models according to claim 2 is characterized in that: Step 3 specifically includes: Step 3.1: Before starting the roundup task, check whether there is any prior expert knowledge or dataset related to the current roundup task, and then search and evaluate whether there is an existing dataset that can be used for this roundup task; The capture task dataset refers to the collected dataset D={(d1, d2, ..., d n1 )}, where each data item d i =(r i ',e i ,m i ,t i ), where r i 'Indicates the execution results of the past roundup mission, e i Represents the characteristics of the task execution environment, m i Indicates the behavior pattern of the target object in different situations, t i Represents the action trajectory of the agent during execution; Define the evaluation of each dataset D i Compatible i As follows: Compatible i =ω1*E i +ω2*Q i +ω3*B i , where E i is the data content matching score, Q i is the data quality score, B i is the data applicability score, ω1, ω2, ω3 are the corresponding weights, ω1 ≥ 0, ω2 ≥ 0, ω3 ≥ 0, and the sum of the three weights is 1; In the case of only one existing dataset, determine whether the applicability of the dataset exceeds the preset threshold ε. If so, it indicates that the dataset can be used for this round-up task. Otherwise, it is considered that there is no relevant dataset. If there are multiple existing datasets, determine whether the applicability of each dataset exceeds the preset threshold ε. Then, sort and select the dataset in descending order according to the applicability, and select it for this round-up task. Step 3.2: If there is prior expert knowledge or data set, perform the following process; For the dataset that can be applied to this round-up task: (1) Perform data cleaning operations; (2) Label the data, from the dataset D = {(d1, d2, ..., d n1 )} extract each data item d i The execution result of the roundup task in r i ' and the agent's action trajectory t i , and then from the action trajectory t i Extract the state space S of each agent i and action space A i ; Then, based on the specific target, constraint, and evaluation indicators of the roundup task, a corresponding large language model is selected for each agent as the roundup decision model; Otherwise, perform the following steps; Step 3.3: Based on the specific target, constraint, and evaluation metric of the roundup task, select the corresponding large language model for each agent as the roundup decision model. Step 3.4: Construct a prompt word template for a large language model suitable for the roundup task; Step 3.5, using the large capture decision model to make control decisions, perform multiple rounds of multi-agent collaborative capture to generate and collect capture data; In step 3.6, the collected data from multiple rounds of capture are cleaned, and then the recorded data are labeled. The state space and action space of each agent at each step, as well as the position, speed, and deflection direction of the captured object at each step are sorted and extracted from the recorded data.

5. The multi-agent cooperative hunting method based on multiple large language models according to claim 4 is characterized in that: In steps 3.2 and 3.3, the corresponding large language model is selected for each agent as the large model for roundup decision making, specifically including: For each candidate large language model M i , its comprehensive score Score i Defined as: Score i =ω1*F i +ω2*P i +ω3*C i , where F i It is the model function fitness score, which evaluates whether the model's functions meet the requirements of the roundup task, including natural language understanding, generation ability, and context processing ability. The formula is defined as: where f ij It is a large language model M i The fitness score of the jth function, n2 is the number of functions; P i Is the model performance score, which evaluates the performance of the model. The formula is defined as: where R i is the response speed score, D i ' is the model accuracy score, α and β are the corresponding weights; C i is the model cost score, which evaluates the cost of the model and is defined as: Where cost i is the actual cost of the model, maxcost is the highest cost among all candidate large language models; ω1, ω2, ω3 are the corresponding score weights, and the sum of the three weights is 1; Select comprehensive score Score i A high-quality large language model is used as a large model for roundup decision making.

6. The multi-agent cooperative hunting method based on multiple large language models according to claim 1 is characterized in that: Step 4 specifically includes: Step 4.1, Fine-tuning method selection: Based on existing large language model fine-tuning techniques, consider the specific target, constraint, evaluation metric, and the size and quality of the task dataset D, and select an appropriate large language model fine-tuning method. Step 4.2, Configure the training environment: Configure GPU / TPU resources to ensure that the large-scale roundup decision model can handle large-scale decision data and real-time training requirements. Also, select a framework that supports efficient fine-tuning and inference to ensure that the framework can handle the complexity of the roundup task. Step 4.3, loading the pre-trained model: loading the large model for roundup decision-making; Step 4.4, data loading and processing: Divide the roundup task dataset into training set and test set; Step 4.5, set fine-tuning parameters: Based on the fine-tuning method determined in step 4.1, set the fine-tuning hyperparameters, including learning rate, batch size, and number of training rounds; Step 4.6, fine-tuning: perform multiple rounds of training on the large model for round-up decision-making using the training set, and evaluate the model performance on the test set, recording key indicators; Step 4.7, model optimization and parameter update: Based on the evaluation results, adjust the hyperparameters and fine-tuning methods to further optimize the performance of the large roundup decision-making model in the roundup decision-making task until the optimization end conditions are met.

7. The multi-agent cooperative hunting method based on multiple large language models according to claim 3 or 4, characterized in that: The prompt word templates in steps 3.4 and 5 are defined as follows: T'=(D',I,H,R',F in ,F out ) Among them, T' is the prompt word template, D' is the task description module, I is the identity analysis module, H is the historical information description module, R' is the reasoning guidance process module, F in is a fixed input format, F out Is a fixed output format; (1) Task description module D': D' = "capture mission background + capture mission objectives + specific capture environment + rules and regulations during the capture process"; (2) Identity Analysis Module I: I = "Identity Analysis + Responsibilities / Goals"; (3) Historical information description module H: H = "historical information + historical decision"; (4) Reasoning guidance process module R': R' = "Now please think carefully about the current environment information and the state of the agent. Based on historical information and environmental changes, consider the following factors: the position and speed of the target object, the current position, speed and direction of the agent, and obstacles in the environment, and infer the optimal next action"; (5) Fixed input format F in :F in = "Current position p of agent i i , speed v i , direction θ i , perception information i and the remaining resources r i , the position p of the captured object t t , speed v t and the deflection direction θ t ”; (6) Fixed output format F out :F out ="Next decision: {a ai ,a θi ,a vi ,a ci }, where a ai is the action that each agent i can choose, a θi is to change the orientation angle of agent i, a vi is the speed of changing agent i, a ci It is a communication action.

8. The multi-agent cooperative hunting method based on multiple large language models according to claim 3 or 4, characterized in that: The multi-agent collaborative roundup described in steps 3.5 and 6 specifically includes: (1) Initialization: Set the initial state S of agent i i0 =(p i0 ,v i0 ,θ i0 ,o i0 ,r i0 ) and the initial position p of the captured object t t0 =(x t0 ,y t0 ), initial velocity v t0 =(v x0 ,v y0 ) and the initial deflection direction θ t0 , and record; (2) Observe the environment: The agent observes the surrounding environment and calculates the current position of the agent i p i =(x i ,y i ), speed v i =(v xi ,v yi ), direction θ i , perception information o i ={(p j ,v j ,θ j ,o j ,r j )|j=1,2,…,k}, agent’s remaining resources r i , the position of the captured object p t =(x t ,y t ), speed v t =(v x ,v y ), deflection direction θ t Passed to the big model of roundup decision; (3) Decision generation: According to the prompt word template T', the corresponding prompt word T is input to all the large models of roundup decision making. i , and get the decision output F corresponding to each agent outi Each agent makes a decision based on the output of the corresponding capture decision model F outi , execute the corresponding action A i ={a ai ,a θi ,a vi ,a ci }, and record the decision, that is, the execution action, and save the intermediate reasoning process of the large model of the roundup decision; (4) State update: After executing the agent action, update the state space of the agent and the position, speed and direction of the captured object, and record the current state; (5) Multiple rounds of loops: Repeat steps (2), (3), and (4) until the task is completed.

9. The multi-agent cooperative hunting method based on multiple large language models according to claim 1 is characterized in that: In step 7, the evaluation index of the comprehensive roundup task is Evaluate Task and reinforcement learning evaluation indicators Evaluate RL Evaluate the effectiveness of multiple rounds of actual round-up and the performance of multiple large round-up decision-making models, and optimize the prompt word templates and model parameters; Among them, the evaluation index Evaluate Task =(S,T,E,R), where S represents the capture success rate, which is defined as the probability or ratio of successfully capturing the target object. The formula is: T represents the average time required to successfully capture the target object, and the formula is: E represents the ratio of effective actions in collaborative work of intelligent agents, and the formula is defined as: R represents the total amount of resources consumed to complete the task, including manpower, materials, and equipment. The formula is defined as: Reinforcement learning evaluation indicators Evaluate RL =(Rb,Ce), where Rb represents robustness and Ce represents model computational efficiency; The comprehensive evaluation formula is: F'=αS-βT+γE-δR+w1Rb+w2Ce, where α, β, γ, δ, w1, and w2 are the weights of each indicator, and α+β+γ+δ+w1+w2=1, and F' represents the evaluation value.

10. A multi-agent cooperative hunting system based on multiple large language models based on the method according to any one of claims 1 to 9, characterized in that: The system comprises: The first module is used to define the state space and action space of the intelligent agent according to the round-up task; The second module is used to initialize the simulation environment and set the initial positions and states of each agent and the captured object; The third module is used to select a large language model for each intelligent agent as its corresponding large model for roundup decision-making, and at the same time obtain or construct a roundup task dataset; The fourth module is used to select a fine-tuning strategy suitable for the roundup task and perform customized fine-tuning on the roundup decision model; The fifth module is used to write a dedicated roundup prompt word template for the fine-tuned roundup decision model; The sixth module is used to control the decision-making by using the fine-tuned large-scale roundup decision model to achieve multi-agent collaborative roundups; The seventh module is used to evaluate and optimize the capture decision-making model based on the performance and results of the capture mission; The eighth module is used to deploy the optimized model and algorithm into the actual environment for testing, monitor the performance of the intelligent agent in real time in the actual environment, and make timely adjustments and optimizations as needed.

Images