A Multi-Agent Collaborative Game Decision-Making Method Based on a Large Language Model

Abstract

This invention discloses a multi-agent collaborative game decision-making method based on a large language model, comprising: taking the visualized environmental observations of the embodied agents as input; analyzing the environmental features of the observable agents; simulating the agents' understanding and memory of the environment based on the observable environmental features using a large language model; the large language model performing comprehensive analysis based on the understanding and memory of the environment to make high-level action decisions; designing and managing communication between agents, including when to communicate, how to communicate, and the content of communication, to achieve information sharing among agents; heuristic low-level action planning, mapping high-level action decisions to low-level action instruction codes; cross-platform instruction transmission, uploading the low-level action instruction codes to the collaborative game environment, where agents perform actions to complete the current-moment environmental deduction; through this invention, leveraging the rich knowledge and rigorous and superior logical thinking of the large language model, real-time and accurate intelligent decision-making in multi-agent collaborative game scenarios can be achieved.

Description

Technical Field

[0001] This invention designs an intelligent decision-making method for multi-agent collaborative game, specifically a multi-agent collaborative game decision-making method based on a large language model. Background Technology

[0002] With the rapid development of high technologies such as artificial intelligence, research on intelligent decision-making methods for multi-agent collaborative game theory has gradually become one of the hot research areas in collaborative game scenarios. Collaborative decision-making methods for intelligent agents can be used to assist human decision-making, promote better cooperation and collaborative work, and drive social progress. In the field of military strategic game theory, intelligent decision-making methods for multi-agent collaborative game theory also play a crucial role in national defense and security. These methods can provide more intelligent, flexible, and efficient decision support in military strategic planning, helping national defense forces maintain competitiveness and advantages in complex and ever-changing security environments.

[0003] Mainstream intelligent algorithms for multi-agent collaborative game scenarios primarily rely on deep reinforcement learning (DRL). DRL-based collaborative game algorithms offer significant advantages in automation, complex environments, real-time decision support, and multi-domain applications. During training, DRL-based agents learn from interactions with the environment, adapting to different game scenarios without requiring the manual formulation of complex rules. Compared to traditional methods based on expert knowledge and logical symbolic rules, DRL-based agent decision-making algorithms greatly improve the performance and efficiency of multi-agent collaborative decision-making. However, complex collaborative game scenarios present challenges due to the high degrees of freedom in the action and state spaces of multi-agent decisions, the complexity of reward design, the difficulty in achieving satisfactory guidance for agents, unstable model convergence, and time-consuming and resource-intensive training. These problems and limitations pose challenges to traditional mainstream deep reinforcement learning-based multi-agent game learning methods. Summary of the Invention

[0004] The purpose of this invention is to provide a multi-agent collaborative game decision-making method based on a large language model, targeting intelligent decision-making in multi-agent collaborative game scenarios. This method abandons the previous mainstream reinforcement learning-based multi-agent training paradigm for collaborative games, overcoming the bottlenecks and problems of various mainstream reinforcement learning agent training methods, such as variable state and action spaces, sparse rewards, difficulty in parameter convergence, and lack of utilization of prior human knowledge. The multi-agent collaborative game decision-making method based on a large language model in this invention offers more accurate and intelligent decision planning, possesses long-term planning capabilities, and even rivals the average human decision-making level. This invention is also the first attempt to implement the capabilities of a large language model in multi-agent collaborative game scenarios.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] The present invention discloses a multi-agent cooperative game decision-making method based on a large language model for multi-agent cooperative game scenarios, the main steps of which are as follows:

[0007] (1) Given a real-time collaborative game environment and multiple embodied agents played by a large language model, the environmental state observations of the multiple embodied agents are visualized, including RGB images and streaming data. Such multi-source information is combined as the input of this method.

[0008] (2) Based on the above observation of the visualized environment state of multiple agents, analyze and extract the observable environmental state features of the agents, including the agent's own attributes (position, orientation, speed, etc.), the state of other agents on both sides (attributes, position, etc.), and environmental elements such as terrain within the visible range.

[0009] (3) Based on the observable environmental state features extracted above, a unique memory storage and update mechanism is constructed through a large language model to simulate the real-time understanding and memory of the cooperative game environment by an agent based on a large language model.

[0010] (4) The intelligent agent played by the large language model makes the next high-level action decision based on its understanding and memory of the real-time collaborative game environment and uses natural language as the carrier. There are two situations: if the intelligent agent makes a decision action, the underlying action planning of the intelligent agent is carried out; if the intelligent agent initiates a communication request, information communication between intelligent agents is carried out.

[0011] (5) For the above-mentioned information communication between intelligent agents, it is necessary to design and manage the communication between intelligent agents, including controlling when and how intelligent agents communicate and designing the specific content of the communication. After the communication is completed, the large language model continues to simulate the intelligent agents to make high-level action decisions.

[0012] (6) The agent makes a decision-making action in response to the above-mentioned intelligent agent. The intelligent agent carries out heuristic low-level action planning and outputs the corresponding low-level action instruction code in the game based on the high-level action decision made in natural language.

[0013] (7) Design a cross-platform instruction transmission scheme to upload the above-mentioned underlying action instruction encoding to the multi-agent collaborative game environment and enter the next round of cyclic decision-making corresponding to the next moment of environment deduction;

[0014] Furthermore, in the multi-agent collaborative game decision-making method based on a large language model, the real-time collaborative game environment refers to a simulated environment in which a series of agents engage in collaborative game reasoning and decision-making. This environment involves various roles controlled by multiple embodied agents completing various goals and tasks, such as in StarCraft 2 and Red Alert 3 games. The large language model represents a series of deep learning-based artificial intelligence models (such as GPT-4, LLaMA, and PanGu-Alpha), possessing rich world knowledge and powerful reasoning abilities comparable to humans to a certain extent. It features: a parameter scale of trillions, pre-training fine-tuning, context awareness, transfer learning and adversarial training, and multimodal support.

[0015] Further, the observation of the embodied agent's visualized environment state in step (1) is as follows: If the game screen is used as the carrier, the input is the RGB image P of each agent (such as the i-th agent) as the first perspective in each frame. i Image collection If streaming data is used as the carrier, such as real-time state observation data of agents represented by text formats like JSON and XML, then the input is a set of information representations of the game environment at each moment, which includes serial data of all agents and records information such as attributes, position, orientation, and health.

[0016] Based on this, for different collaborative game scenarios, the aforementioned real-time environmental state observations are obtained as inputs to this multi-embodied intelligent agent collaborative game decision-making method based on a large language model.

[0017] Furthermore, the environmental state feature parsing and extraction process described in step (2), specifically for the extraction of observable environmental state features of an agent using streaming data, uses adaptive template generation to assist the agent in aligning information of the original observable environment. This process includes the following steps:

[0018] Sub-step 1, Information Location and Capture: For game environments based on streaming data, use adaptive template generation to capture key information in the environment;

[0019] Sub-step 2, Information Alignment: By comparing the adaptively generated template with the actual streaming data, the information in the streaming data is aligned with the game environment to support the agent's understanding of the state and characteristics of the game environment, thereby obtaining the agent's observable environmental state characteristics.

[0020] Furthermore, the environmental state feature parsing and extraction process described in step (2), for the feature extraction process of the environment using the game-based RGB image as the carrier, requires an additional step of information localization and capture to obtain effective high-level information, such as the location of surrounding non-intelligent objects, the positional relationship with friendly intelligent agents, and the location and state of enemy intelligent agents within the visible range. The specific steps are as follows:

[0021] Sub-step 1: Mapping RGB image input to 3D point cloud: Using the 3D point cloud, each pixel of the RGB image observed by each agent is converted into a set of points with three-dimensional coordinates;

[0022] Sub-step 2, Object Instance Mask Segmentation: Import the shape and size features of each environment object (objects include intelligent agents and non-intelligent agents, i.e., objects), use the instance segmentation model to segment and identify objects in the 3D point cloud, and assign a unique identifier to each instance;

[0023] Sub-step 3, Object Relationship Extraction and Recognition: Based on the extracted 3D coordinate point set and instance mask segmentation results, calculate the coordinate position, size, and relative position and distance between objects to support the subsequent understanding of the intelligent agent's understanding of the object layout and relationships between objects in the environment.

[0024] Furthermore, the large language model memory storage and update mechanism described in step (3) for simulating the agent's understanding and memory of the game environment is dynamic, adaptive, and constantly updated. It mainly simulates the agent's understanding and memory of the real-time cooperative game environment based on the large language model, and its specific responsibilities are as follows:

[0025] 1. Record the overall progress of the task: In multi-task game scenarios, the agent should record the completion status of each sub-task at all times to avoid the situation of repeated execution of a single task;

[0026] 2. Self-state: It is extremely important for the agent to know itself. This module records the changes in its own state and, when necessary, communicates and shares real-time information and state with other agents through the inter-agent communication module in step (5).

[0027] 3. Observable states of other friendly and enemy agents: Track the observable states of other friendly and enemy agents, synchronize and share state information among cooperative agents, and support the exchange of enemy intelligence among cooperative agents;

[0028] 4. Scene Reproduction Memory: Responsible for storing information about all objects in the environment that the agent acquires through its own observations and communication with other agents. In agent game scenarios, the actions of agents are typical time-varying Hidden Markov Chain decision processes. The key characteristic is that the agent's decisions not only depend on the current state, but also on a series of state observations before the current time step, in order to help the agent formulate long-term plans and decisions.

[0029] Furthermore, the large language model described in step (4) acts as an intelligent agent to make high-level action decisions. Specifically, the large language model acts as an intelligent agent, and through analyzing the environment and communication results, formulates the next decision action in a high-level form using natural language as the carrier. Specific implementation details include:

[0030] Develop text prompts for agent reasoning and decision-making: Design prompts for large language models to simulate agent decision-making and reasoning, including prompt instruction header description, task progress description, state description (including: self-state description, observable environment description, enemy and friendly agent state description, self-stored scene memory description, action history, dialogue history), and instruction tail description (used to guide the large language model to make action decisions or communication requests).

[0031] Decision-making reasoning based on large language model dialogue: The constructed text prompts are input into the large language model to obtain the model's output. The model then makes subsequent action decisions or initiates task invitations to other agents. If the model makes a decision on its own next action, it skips the design and management of inter-agent communication in step (5) and proceeds to the agent-level action planning in step (6). If the model sends a request to communicate with other agents, it proceeds to the design and management of inter-agent communication in step (5) and continues the inter-agent communication process.

[0032] Furthermore, the design and management of inter-agent communication described in step (5) is used to control when, how, and the specific design of the communication content of the agents. The process includes the following steps:

[0033] Design of intelligent agent communication text prompts: Prompt design for intelligent agent communication simulation using a large language model, including prompt instruction header description, task progress description, and state description (including: self-state description, observable environment description, enemy and friendly agent state description, self-stored scene memory description, action history, dialogue history, and subtask requests to other intelligent agents).

[0034] Inter-agent communication and feedback: The constructed complete communication text prompt is input into the large language model, which is required to act as the communicating agent and provide a text description of the subsequent action arrangement in natural language. If the agent makes its own action decision, it enters the heuristic underlying action planning in step (6) and implements the action in the game scene. If the agent makes an inter-agent communication request, it enters the design and management of inter-agent communication described in step (5) and continues to carry out inter-agent information communication.

[0035] Furthermore, the heuristic agent-based low-level action planning described in step (6) maps the high-level planning in the given natural language form to the low-level action instruction encoding in the game. The specific process is as follows:

[0036] Sub-step 1, Receiving and parsing high-level action decisions: Receive high-level action decisions in natural language from step (4) or step (5), parse and extract key information and instructions from the text;

[0037] Sub-step 2, Action Decomposition and Mapping: Decompose the key instruction information of high-level action decision into a series of specific low-level actions and map them to the one-hot encoding of the action space of the intelligent agent in the game scene;

[0038] Sub-step 3, Action Sequence Generation: Based on the One-hot action codes obtained above, identify and create an ordered sequence of the above action set, and output it as the underlying action code sequence of the multi-agent game environment.

[0039] Preferably, the aforementioned heuristic underlying action planning can be continuously optimized through supervised machine learning or reinforcement learning to resolve actions and generate sequences. For example, for a reinforcement learning-based heuristic underlying action planning algorithm, the underlying action planning strategy can be adjusted in real time during the game process through environmental feedback rewards.

[0040] Furthermore, the cross-platform instruction transmission scheme described in step (7) is mainly responsible for uploading the underlying action encoding sequence to the multi-agent collaborative game environment. The specific operation process includes:

[0041] For specific scenarios, cross-platform command transmission solutions will also be developed and deployed accordingly:

[0042] Preferably, for cooperative role-playing games like StarCraft or tower defense games like Warcraft developed using the Galaxy++ language, mainstream large language models based on Python and GPT-4 or LLaMA are preferred as the language model dialogue access interface, which includes the following steps:

[0043] Sub-step 1, Action instruction encoding acquisition: First, obtain the underlying action instruction encoding sequence from the heuristic underlying action planning, which corresponds to the specific decision action sequence that the agent will act on in the game environment;

[0044] Sub-step two, Action Command Encoding Formatting: Serialize the action command encoding and adapt it to a cross-platform transmission format, such as using Protocol Buffers to efficiently sequence structured data; then, use TFRecord to encapsulate the serialized action command encoding into a cross-platform transmission format; then, use the Python-based StarCraft development kit "PYSC2" to convert the action command encoding into a format compatible with the game platform, i.e., the representational paradigm of the Galaxy++ language; finally, upload the action command encoding to the target platform "Blizzard Entertainment" or the game environment "SMAC" via network transmission or other data transmission methods.

[0045] Sub-step 3, Action Command Decoding and Implementation: The action commands are encoded, parsed, and converted into a sequence of intelligent agent action commands that the game environment can recognize. These action command sequences are executed based on the current time step, and the environment and object states are updated to enter the cooperative game scenario simulation for the next time step.

[0046] Furthermore, one cycle of the seven main steps described above can correspond to a specific time step in a cooperative game for a multi-agent decision-making scenario with discrete actions; for a multi-agent decision-making scenario based on continuous actions, each cycle of the above method will be implemented in the game scenario, and the simulator will deduce the environmental state of each frame.

[0047] Furthermore, through such a continuous cyclical process, multiple embodied agents combine the powerful reasoning capabilities of large language models to make effective decisions and collaborate in complex game environments, thereby achieving the goal of collaborative game.

[0048] The multi-agent collaborative game decision-making method based on a large language model of the present invention has at least the following beneficial technical effects:

[0049] 1. Intelligent Opponent: The large language model can act as an intelligent opponent with intelligence up to the human level, possessing game strategies comparable to human players, increasing the challenge of the game and allowing players to enjoy a more competitive game experience;

[0050] 2. Real-time decision support: In game scenarios, players usually need to make decisions in a short period of time. Large language models can quickly generate strategy suggestions, hints or explanations to help players make wise decisions and improve the game experience.

[0051] 3. Large-scale knowledge base: Large language models can access a wide range of knowledge resources, including text and data on the Internet, thus possessing extensive background knowledge, which helps intelligent agents to better consider tactical, military, scientific and other factors in the decision-making process;

[0052] 4. Powerful language processing capabilities: The large language model has powerful natural language processing capabilities, capable of understanding and processing complex natural language commands and dialogues. This allows the multiple embodied intelligent agents played by the large language model to better interact with players and understand their intentions and requirements.

[0053] 5. Learning and Improvement: Large language models can continuously learn and improve their game strategies, accumulating experience through interaction with a large number of players, thereby providing a better gaming experience.

[0054] This multi-agent collaborative game decision-making method based on a large language model completely abandons the reinforcement learning paradigm based on reward-based inverse optimization models in multi-agent collaborative game scenarios, and completely breaks free from the various constraints mentioned above that are dependent on traditional reinforcement learning. To our knowledge, this is a novel invention that applies a large language model to multi-agent collaborative game decision-making scenarios for the first time.

[0055] Of course, product instances implementing this invention in any intelligent collaborative game scenario do not necessarily need to achieve all the advantages and significance described above at the same time. Attached Figure Description

[0056] Figure 1 The accompanying figure illustrates the abstract of a multi-agent collaborative game decision-making method based on a large language model.

[0057] Figure 2 This provides a framework for the overall operational process of a multi-embodied intelligent agent collaborative game decision-making method based on a large language model, which can serve as a detailed summary diagram.

[0058] Figure 3 This is a flowchart illustrating the operation of the reasoning and decision-making module based on a large-model dialogue.

[0059] Figure 4 This is a flowchart illustrating the operation of an intelligent agent communication module based on a large language model for dialogue.

[0060] Figure 4 In the process: 1. Given a communication initiating agent A and a requested communication agent B; 2. Using an automated prompt template and the known information of the communication initiating agent, construct a complete communication request from the communication initiating agent A in natural language; 3. The requested communication agent B, represented by GPT-4 (or other large language models), responds to the complete communication request from the communication initiating agent A in natural language. Detailed Implementation

[0061] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. It should be noted that the embodiments described below are merely some embodiments of the present invention, and not embodiments compatible with all intelligent agent cooperative game scenarios. Referring to the embodiments demonstrated in this invention, all other embodiments obtained by those skilled in the art without any inventive step are also within the protection scope of this invention.

[0062] To address the design challenges of intelligent game theory algorithms in multi-agent cooperative game scenarios, this invention provides a multi-agent cooperative game decision-making method based on a large language model. Please refer to the appendix of the specification. Figure 2 As shown.

[0063] Furthermore, this invention provides a specific embodiment: based on the two-player land, sea, and air force command game of *Command & Conquer: Red Alert 3*, focusing on land warfare, a typical multi-agent cooperative combat scenario between the red and blue sides. The red side comprises multiple unit types: tanks, infantry, submarines, patrol aircraft, reconnaissance aircraft, etc., aiming to destroy the blue side's defensive barracks. Assume that at time t, the agent acting as the blue side's reconnaissance aircraft detects the enemy barracks target and engages in a cooperative game decision-making process with the agent acting as a friendly tank.

[0064] The embodiments of the present invention are given for the purposes of illustration and description, and are not intended to be exhaustive or to limit the invention to the forms disclosed. Those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features.

[0065] Based on the aforementioned game combat scenarios, the method flow of this invention will be described in detail, and the specific method includes:

[0066] First, the method inputs are the embodied intelligent agent's visualized environment observations, including real-time frames and streaming data of RGB images. Specifically, in this embodiment, by monitoring the real-time first-person perspective of the game observed by the red team's reconnaissance aircraft, the embodied visualized environment state observation information representing the red team's reconnaissance aircraft intelligent agent as a first-person perspective is recorded.

[0067] Furthermore, the designed environmental feature extraction process will extract and analyze the observable environmental features of the agent based on the aforementioned monitored environmental state observation information. These features include the agent's own attributes (position, orientation, speed, etc.), the states (attributes, positions, etc.) of other agents on both sides, and environmental elements such as terrain within the visible range. Specifically, in this example, based on the images acquired by the reconnaissance aircraft agent, 3D point cloud analysis is performed to obtain a set of points with three-dimensional coordinates. Then, through object instance masking segmentation, the entity features of the enemy barracks and fortresses are used as the segmentation index to obtain the specific coordinates, height, and other parameterized information of the enemy target.

[0068] Furthermore, by constructing an agent memory storage and update mechanism, the real-time understanding and memory of the cooperative game environment by an agent based on a large language model is simulated. Specifically, in this embodiment, the red reconnaissance aircraft agent records its own attributes and state, including: its own coordinates, ammunition quantity, ammunition power, fuel quantity, flight speed, altitude, and orientation; records the enemy's state: the location coordinates of the blue team's barracks fortress, the location coordinates of other blue team forces recorded in historical observations, etc.; records the communication history between agents, including the dialogue history with friendly tanks, infantry, patrol aircraft, etc.; records the action history, including historical decision-making actions for a period of time before time t, including the coordinates of the blue team's anti-tank force detected at time t-3, the coordinates of the blue team's patrol aircraft detected at time t-2, and the blue team's barracks fortress detected at time t-1, etc.

[0069] Furthermore, the large language model acts as an intelligent agent to make high-level action decisions. That is, the intelligent agent, played by the large language model, makes the next high-level action decision based on its understanding and memory of the real-time collaborative game environment, using natural language as the carrier.

[0070] First, the formulation of text prompts for agent reasoning and decision-making includes the prompt instruction header description, task progress description, state description (including: self-state description, observable environment description, enemy and friendly agent state description, self-stored scene memory description, action history, and dialogue history), and instruction tail description (used to guide the large language model to make action decisions or communication requests), etc., which correspond to the appendix to the instruction manual. Figure 3 Input part of the Zhongda language model.

[0071] Secondly, the decision-making reasoning process based on large language model dialogue is divided into two cases: making high-level decision actions and initiating inter-agent communication requests. If a high-level decision action is made, the design and management of inter-agent communication are skipped and the low-level action planning of the agent is entered; if an inter-agent communication request is initiated, the design and management of inter-agent communication are carried out.

[0072] Specifically, in this embodiment, after the red team's reconnaissance aircraft discovers the blue team's barracks fortification, it generates a decision-making request in natural language form using an automated prompt template. The reconnaissance aircraft's intelligent agent is played by a large language model, which processes the decision-making request and provides feedback to make the next planning decision. The entire dialogue decision-making process based on the large language model is described in the appendix to the specification. Figure 3 As shown. The large language model acts as an intelligent agent, analyzing and making action decisions, which fall into two different categories: 1. Initiating a communication request: It determines that the patrol aircraft's ammunition is insufficient to destroy the red force's fortified camp target, requiring the deployment of friendly tank units and infantry groups for coordinated attacks; 2. Making a high-level action decision: It determines that the patrol aircraft is fully capable of destroying the target, and directly issues an attack command. The above corresponds to the appendix to the instruction manual. Figure 3 Feedback section of the Zhongda language model.

[0073] Furthermore, regarding the design and management of inter-agent communication, it is necessary to control when, how, and the specific content of communication between agents. Then, a large language model is used to simulate the agent reasoning about the next high-level action using natural language. Specifically, in this embodiment, a communication request text in natural language is first constructed for the blue reconnaissance aircraft, including a prompt instruction header description, task progress description, and state description (including: its own state description, observable environment description, enemy and friendly agent state descriptions, its stored scene memory description, action history, dialogue history, and sub-task requests to other agents). This text is then input into the large language model to obtain feedback. The above corresponds to the appendix to the specification. Figure 4 The large language model input section. Based on the above request, the large language model acts as the communicating intelligent agent (tank and infantry group) to make high-level decision-making actions using natural language. The above corresponds to the appendix to the instruction manual. Figure 4 Feedback section of the Zhongda language model.

[0074] Furthermore, for the agent's decision-making actions, a heuristic low-level action planning method is used to map the high-level action decisions in natural language form to the low-level action command encoding in the game. The execution process of this low-level action planning method includes: receiving and parsing high-level action decisions, action decomposition mapping, and action sequence generation. Specifically, in this embodiment, the input to the low-level action planning method is: the high-level decision actions in natural language form of the tank and infantry group agents that have received support requests. For the tank and infantry group agents, the low-level action planning method parses out the sub-commands in their high-level natural language decisions: move to area F, approach the target point (4179384, 1371781), and attack the blue square circular barracks fortress. The above actions are encoded into a one-hot encoding sequence in the agent's low-level action space, where each encoding in the sequence maps to a unique action type in the action space.

[0075] Furthermore, through a cross-platform command transmission scheme, the underlying action command encoding sequence from the previous step is uploaded to the multi-agent collaborative game environment and enters the next round of cyclic decision-making in the next time step. Specifically, in this embodiment, the one-hot action encoding sequence obtained in the previous step is adapted to a cross-platform transmission format, such as using ProtocolBuffer sequence structured data; then, TFRecord is used to encapsulate the serialized command actions into a cross-platform transmission format; subsequently, using a Python development package based on the *Command & Conquer: Red Alert 3* game, the underlying action command encoding sequence is converted into a format compatible with the game platform; finally, the action command information is uploaded to the target game platform via network transmission or other data transmission methods for a new round of environment simulation in the collaborative game.

[0076] Furthermore, by repeating the above process, we can advance the simulation of multi-agent collaborative game scenarios and conduct a continuous collaborative game process between the opposing and friendly sides.

[0077] To reiterate, the above embodiments are merely some examples of the present invention, and not embodiments compatible with all intelligent agent cooperative game scenarios. All other embodiments obtained by those skilled in the art without any inventive step, referring to the embodiments demonstrated in this invention, are also within the scope of protection of this invention.

Claims

1. A method for multi-agent coordination game decision-making based on a large language model, characterized in that, Includes the following steps: Step 1: Given a real-time collaborative game environment and multiple embodied agents played by a large language model, use the current visual environmental state observations of the multiple embodied agents as the input of this decision-making method. The multiple embodied agents include RGB images and streaming data. Step 2: In the real-time collaborative game environment, based on the current visual environmental state observation of multiple agents, analyze and extract the observable environmental features of the agents. The observable environmental features of the agents include unit attributes, the states of other agents on both sides, and terrain environmental elements within the visible range. Step 3: Based on the extracted observable environmental features of the agent, a memory storage and update mechanism is constructed through a large language model to simulate the agent's understanding and memory of the real-time collaborative game environment. Step 4: Based on the agent's understanding and memory of the real-time collaborative game environment, the large language model uses natural language as a carrier to make high-level action decisions, which are divided into two cases: when the agent makes a decision action, the agent's low-level action planning is carried out; and when the agent initiates a communication request, information communication between agents is carried out. Step 5: Design and manage the communication process between intelligent agents, including controlling when and how agents communicate, and designing the specific content of the communication; after the communication is completed, the large language model continues to make high-level action decisions. Step Six: Make decision actions for the intelligent agent. The intelligent agent carries out heuristic low-level action planning, mapping the high-level action decisions in natural language to the low-level action instruction codes in the game. Step 7: Design a cross-platform instruction transmission scheme to upload the underlying action instruction code to the multi-agent collaborative game environment. The agents perform actions according to the instructions and enter the next moment of the collaborative game environment simulation.

2. The method of claim 1, wherein the method further comprises: Real-time collaborative game environment refers to a simulated environment in which a series of intelligent agents engage in collaborative game reasoning and decision-making, and the environmental deduction process in which various roles controlled by multiple embodied intelligent agents complete various goals and tasks; large language model refers to a series of generative language models based on artificial intelligence deep learning, which have rich world knowledge and powerful reasoning ability.

3. The method of claim 1, wherein the method further comprises: In step one, the visual environment state observation information of the multi-agent at the current time includes: taking the i-th agent as the first perspective, the RGB image P i of the image set when the game picture is taken as the carrier, or the real-time observation data represented by JSON, XML text when the streaming data is taken as the carrier.

4. The method of claim 1, wherein, Step two, based on the visualized environmental state observation of multiple agents, includes the following steps: The process of analyzing and extracting environmental features, specifically the feature extraction process for the game-based adversarial environment using streaming data, includes the following steps: Information location and capture: For game environments based on streaming data, adaptive template generation is used to capture key information in the environment; Information alignment: By comparing the adaptively generated template with the actual streaming data, the information in the streaming data is aligned with the game environment to support the agent's understanding of the state and features of the game environment.

5. The method of claim 1, wherein, In step two, based on the visualized environmental state observation of multiple agents, the environmental feature analysis and extraction process, specifically the feature extraction process for the game adversarial environment using RGB images as the carrier, includes the following steps: RGB image input to 3D point cloud mapping: Through 3D point cloud, the RGB image information observed by each agent is converted into a set of points with three-dimensional coordinates; Object instance mask segmentation: Using an instance segmentation model, objects in a 3D point cloud are segmented and identified, and a unique identifier is assigned to each instance; Object Relationship Extraction and Recognition: Based on the above 3D coordinate point set and instance mask segmentation results, the coordinates, size, and relative positions and distances between objects are calculated, supporting the agent's understanding of object layout and relationships in the environment.

6. The method of claim 1, wherein the method further comprises: In step three, a memory storage and update mechanism is constructed using a large language model to simulate the agent's understanding and memory of the real-time cooperative game environment. This mechanism is specifically responsible for the following: Record overall task progress: Record the completion status of each subtask in a multi-task game scenario; Self-state: Recording changes in one's own state; Observe the state of other friendly and enemy agents: track the observable states of other friendly and enemy agents; Scene Reproduction Memory: Responsible for storing and remembering historical information about all objects in the environment that the agent acquires through its own observation and communication with other agents, so as to assist the agent in making long-term plans and decisions.

7. The method of claim 1, wherein the method further comprises: In step four, the large language model uses natural language as a medium to make high-level action decisions, specifically including: Develop text prompts for agent reasoning and decision-making: Design natural language prompts for simulating agent decision-making in large language models, including prompt instruction header description, task progress description, state description, and instruction tail description; Decision-making reasoning based on large language model dialogue: The constructed text prompts are input into the large language model, the following output of the model is obtained, and the model makes subsequent high-level formal action decisions or initiates communication requests to other agents; if the model makes a high-level action decision, the design and management of inter-agent communication is skipped and the high-level action decision is entered; if the model makes a communication request, the design and management of inter-agent communication is entered.

8. The method of claim 1, wherein, Step five involves designing and managing the communication process between intelligent agents, specifically including the following steps: Design intelligent agent communication text prompts: Design prompts for use in large language models to simulate intelligent agent communication, including prompt instruction header description, task progress description, status description, and instruction tail description; Inter-agent communication and feedback: The constructed complete communication text prompt is input into the large language model, which acts as the agent to communicate and provides a text description of the subsequent action arrangement in natural language form; Inter-agent communication decision processing: If an agent makes its own action decision, it enters the agent's heuristic underlying action planning; if an agent makes an inter-agent communication request, it executes inter-agent communication design and management, and continues the inter-agent communication process.

9. The method of claim 1, wherein, In step six, the agent makes a decision-making action, and the agent unfolds a heuristic underlying action planning process, as follows: Receiving and parsing high-level action decisions: receiving action decisions for the design and management of inter-agent communication or in a high-level form, outputting high-level action decisions in natural language, parsing and extracting key instruction information; Action decomposition and mapping: Decompose the key instruction information of high-level action decision into a series of specific low-level action codes and map them to the one-hot action code of the intelligent agent action space in the game scene. Action sequence generation: Based on the above one-hot action encoding, identify and create an ordered sequence of the above action set, and output the underlying action instruction encoding of the multi-agent in the game environment.

10. The method of claim 1, wherein, In step seven, the cross-platform instruction transmission scheme uploads the underlying action instruction encoding to the multi-agent collaborative game environment. Specific methods include: Action instruction code acquisition: First, the underlying action instruction code is obtained from the heuristic underlying action planning, which corresponds to the specific decision action that the agent will take in the game environment; Action instruction encoding formatting: Serializes action instructions and adapts them to cross-platform transmission formats, structuring the data; converts action mapping vectors into a format compatible with the game platform; and transmits the formatted action instruction encoding to the game environment of the intelligent agent game scenario. Action command decoding and execution: In this agent-cooperative game scenario, the action commands that can be recognized by the game environment are parsed from the formatted action command code and executed, the environment is updated and the game scenario enters the next time step.

Images