Cluster collaborative intelligent decision-making method and system based on large language model
By adopting a cluster collaborative intelligent decision-making method based on a large language model, the problems of poor real-time performance and fragmentation of multiple modules in cluster collaborative decision-making in dynamic environments are solved. This method enables efficient and low-error-rate generation and execution of cluster action instructions, and supports unified decision-making for discrete tasks and continuous control.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- INSPUR TIANYUAN COMM INFORMATION SYST CO LTD
- Filing Date
- 2026-02-11
- Publication Date
- 2026-06-05
AI Technical Summary
Traditional optimization algorithms and reinforcement learning methods suffer from poor real-time performance, fragmentation of multiple modules, and difficulty in collaboration among heterogeneous agents when making collaborative decisions in dynamic environments.
A cluster collaborative intelligent decision-making method based on a large language model is adopted. Through environmental information textification, multi-level prompt word construction, instruction generation, multi-level instruction verification, and instruction execution and status management, cluster action instructions are generated and verified to achieve real-time collaborative decision-making for cluster actions.
It improves decision-making efficiency, shortens training time, reduces error rate, enhances scenario adaptability and compatibility with the unified decision-making framework, and supports discrete task allocation and continuous control.
Smart Images
Figure CN122151502A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of intelligent decision-making and automatic control technology, specifically to a cluster collaborative intelligent decision-making method and system based on a large language model. Background Technology
[0002] Traditional optimization algorithms, such as the Kuhn-Munkres algorithm, solve task allocation through matrix modeling, but are only applicable to static discrete targets and cannot generate continuous spatial control instructions; reinforcement learning methods (such as mean field MARL) require separate training of models for heterogeneous agents, have difficulty converging in large-scale collaborative scenarios, and have high deployment costs; rule expert systems rely on predefined priority rules and have poor adaptability to dynamic environments.
[0003] Overcoming the shortcomings of poor real-time performance, fragmented modules, and difficulties in collaboration among heterogeneous intelligent agents in dynamic environments is a technical problem that needs to be solved. Summary of the Invention
[0004] The technical objective of this invention is to address the above-mentioned shortcomings by providing a cluster collaborative intelligent decision-making method and system based on a large language model, in order to overcome the technical problems of poor real-time performance, fragmentation of multiple modules, and difficulty in collaboration among heterogeneous intelligent agents in dynamic environments.
[0005] In a first aspect, the present invention provides a cluster collaborative intelligent decision-making method based on a large language model, comprising the following steps:
[0006] Environmental information textification: Collect environmental information of the cluster's operating environment and convert it into structured environmental text that is adapted to the large language model. The environmental information includes the ID, type, location, and status information of each unit, including friendly units and enemy units.
[0007] Multi-level cue word construction: Construct multi-level cue words that include role shaping, agent information, text specifications and cue information. In this process, the large language model is defined as the cluster decision-making center through role shaping, the action space of the unit is declared through agent information, the format of the large language model's command output is constrained through text specifications, and dynamic tactical principles are injected into the large language model through cue information.
[0008] Command generation: The structured environment text and multi-level prompts are input into the large language model. The large language model performs semantic understanding and reasoning, parses task requirements, matches dynamic tactical principles, generates corresponding cluster action commands, and outputs the cluster action commands in a predetermined format based on text specifications. Among them, the cluster action commands include discrete task allocation and continuous space control commands.
[0009] Multi-level instruction verification: The cluster action instructions output by the large language model are parsed and verified, and errors are corrected based on the verification results to obtain legal and executable cluster action instructions.
[0010] Command execution and state management: Convert legally executable cluster action commands into action signals that can be responded to by the underlying execution units, manage the execution process of cluster action commands through a state machine, and feed back the unit state update information to the environmental situation textification step to form a closed loop.
[0011] As a preferred approach, when textualizing environmental information, the environmental information is converted into JSON format for output, and the format of the large language model's instruction output is constrained to JSON format by the text specification.
[0012] As a preferred approach, multi-level instruction verification includes the following operations:
[0013] Level 1 verification: Parse the cluster action commands output by the large language model. If parsing fails, the large language model is required to regenerate the commands. If parsing succeeds, proceed to Level 2 verification.
[0014] Level 2 verification: Check the parameter range of the cluster action command. If a parameter is out of bounds, automatically correct it to the closest valid value and perform Level 3 verification.
[0015] Level 3 verification: Check the unit ID. If an illegal unit ID is found, ignore the corresponding action instruction and log it.
[0016] Level 4 verification: Perform a feasibility check on the actions in the instruction. If the action is within the unit action space and meets the current environmental conditions, the action is executable; otherwise, the action is deemed infeasible and the instruction is ignored.
[0017] Level 5 verification: The instructions are comprehensively verified. After confirming that there are no other errors in the instructions, a legal and executable cluster action instruction is obtained.
[0018] As a preferred option, the actions defined in the action space include Move, Attack, and Wait, and the dynamic tactical principles include encirclement and flank cooperation.
[0019] Instruction execution and status management include the following operations:
[0020] Command conversion: For movement commands, the target coordinates are converted into unit heading angles using the heading angle calculation formula;
[0021] State machine management: Three states are set: Move, Attack, and Wait. The state machine switches states based on the progress of instruction execution and environmental changes. In the Move state, the unit moves towards the target position according to the converted heading angle. When it reaches a threshold distance from the target position... When the target disappears, the unit switches to the Wait state. In the Attack state, the unit continuously tracks the target and performs attack actions until the target disappears. Then, the unit switches to the Wait state. In the Wait state, the unit is in standby mode and ready to respond to new commands at any time.
[0022] Secondly, the present invention provides a cluster collaborative intelligent decision-making system based on a large language model, including an environmental information textification module, a multi-level prompt word construction module, an instruction generation module, an instruction multi-level verification module, and an instruction execution and status management module.
[0023] The environmental information textification module is used to perform the following operations: collect environmental information of the cluster's operating environment, and convert the environmental information into structured environmental text adapted to the large language model. The environmental information includes the ID, type, location, and status information of each unit, and the units include friendly units and enemy units.
[0024] The multi-level prompt word construction module is used to perform the following operations: construct multi-level prompt words that include role shaping, agent information, text specifications and prompt information. In this module, the large language model is defined as the cluster decision-making center through role shaping, the action space of the unit is declared through agent information, the format of the large language model's instruction output is constrained through text specifications, and dynamic tactical principles are injected into the large language model through prompt information.
[0025] The instruction generation module is used to perform the following operations: inputting structured environment text and multi-level prompt words into the large language model, the large language model performs semantic understanding and reasoning, parses task requirements, matches dynamic tactical principles, generates corresponding cluster action instructions, and outputs the cluster action instructions in a predetermined format based on text specifications. Among them, the cluster action instructions include discrete task allocation and continuous space control instructions.
[0026] The multi-level instruction verification module is used to perform the following operations: parse and verify the cluster action instructions output by the large language model, and correct errors based on the verification results to obtain legal and executable cluster action instructions;
[0027] The instruction execution and status management module is used to perform the following operations: convert legally executable cluster action instructions into action signals that can be responded to by the underlying execution units, manage the execution process of cluster action instructions through a state machine, and feed back the unit status update information to the environmental situation textification step to form a closed loop.
[0028] As a preferred option, the environmental information textification module is used to convert environmental information into JSON format for output, and the format of the large language model instruction output is constrained to JSON format by the text specification.
[0029] As a preferred embodiment, the multi-level instruction verification module is used to perform the following operations:
[0030] Level 1 verification: Parse the cluster action commands output by the large language model. If parsing fails, the large language model is required to regenerate the commands. If parsing succeeds, proceed to Level 2 verification.
[0031] Level 2 verification: Check the parameter range of the cluster action command. If a parameter is out of bounds, automatically correct it to the closest valid value and perform Level 3 verification.
[0032] Level 3 verification: Check the unit ID. If an illegal unit ID is found, ignore the corresponding action instruction and log it.
[0033] Level 4 verification: Perform a feasibility check on the actions in the instruction. If the action is within the unit action space and meets the current environmental conditions, the action is executable; otherwise, the action is deemed infeasible and the instruction is ignored.
[0034] Level 5 verification: The instructions are comprehensively verified. After confirming that there are no other errors in the instructions, a legal and executable cluster action instruction is obtained.
[0035] As a preferred option, the actions defined in the action space include Move, Attack, and Wait, and the dynamic tactical principles include encirclement and flank cooperation.
[0036] The instruction execution and status management module is used to perform the following operations:
[0037] Command conversion: For movement commands, the target coordinates are converted into unit heading angles using the heading angle calculation formula;
[0038] State machine management: Three states are set: Move, Attack, and Wait. The state machine switches states based on the progress of instruction execution and environmental changes. In the Move state, the unit moves towards the target position according to the converted heading angle. When it reaches a threshold distance from the target position... When the target disappears, the unit switches to the Wait state. In the Attack state, the unit continuously tracks the target and performs attack actions until the target disappears. Then, the unit switches to the Wait state. In the Wait state, the unit is in standby mode and ready to respond to new commands at any time.
[0039] The cluster collaborative intelligent decision-making method and system based on a large language model of the present invention have the following advantages:
[0040] 1. Improved decision-making efficiency: 98% shorter training time compared to reinforcement learning (30-second deployment vs. 2000 training iterations).
[0041] 2. Low error rate: Instruction generation error rate ≤ 0.78% (DeepSeek-R1);
[0042] 3. Scenario adaptability: No retraining is required when the unit configuration changes;
[0043] 4. Unified decision-making framework: compatible with discrete task allocation and continuous control (e.g., generating Move(11,[x,y,z]) instructions). Attached Figure Description
[0044] To more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0045] The invention will be further described below with reference to the accompanying drawings.
[0046] Figure 1 This is a flowchart of a cluster collaborative intelligent decision-making method based on a large language model, as shown in Example 1. Detailed Implementation
[0047] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, so that those skilled in the art can better understand and implement the present invention. However, the embodiments are not intended to limit the present invention. In the absence of conflict, the embodiments of the present invention and the technical features in the embodiments can be combined with each other.
[0048] This invention provides a cluster collaborative intelligent decision-making method and system based on a large language model, which is used to solve the technical problems of how to overcome the defects of poor real-time performance, fragmentation of multiple modules, and difficulty in collaboration of heterogeneous intelligent agents in dynamic environments.
[0049] Example 1:
[0050] This invention discloses a cluster collaborative intelligent decision-making method based on a large language model, comprising five steps: environmental information textification, multi-level prompt word construction, instruction generation, multi-level instruction verification, and instruction execution and state management.
[0051] Step S100 Environmental Information Textification: Collect environmental information of the cluster's operating environment and convert it into structured environmental text adapted to the large language model. The environmental information includes the ID, type, location, and status information of each unit, which includes friendly units and enemy units.
[0052] When texturing environmental information, the environmental information is converted into JSON format for output. The format of the large language model's instruction output is constrained to JSON format by the text specification, for example: {"Our unit": [{"id":14, "type":"PlaneA","position":[362.41,355.88,14.78]}], "Enemy unit": [{"id":-6, "type":"APS","position":[500.00,500.00,0.00]}]}.
[0053] Step S200: Multi-level prompt word construction: Construct multi-level prompt words that include role shaping, agent information, text specifications, and prompt information. In this process, the large language model is defined as the cluster decision-making center through role shaping, the action space of the unit is declared through agent information (e.g., "Act Space": ["Move(unit ID, target coordinates)", "Attack(unit ID, target ID)"]) through agent information, the format of the large language model's instruction output is constrained through text specifications (JSON template), and dynamic tactical principles (such as encirclement and flanking cooperation) are injected into the large language model through prompt information.
[0054] Step S300 Instruction Generation: The structured environment text and multi-level prompts are input into the large language model. The large language model performs semantic understanding and reasoning, parses task requirements, matches dynamic tactical principles, generates corresponding cluster action instructions, and outputs the cluster action instructions in a predetermined format based on text specifications. Among them, the cluster action instructions include discrete task allocation and continuous space control instructions. For example, the instruction is: {"task_allocation": ["Move(11, [500,500,10])", "Attack(31, -6)"]}.
[0055] Step S400: Multi-level verification of instructions: The cluster action instructions output by the large language model are parsed and verified, and errors are corrected based on the verification results to obtain legal and executable cluster action instructions.
[0056] As a specific implementation of multi-level instruction verification, this step includes the following operations:
[0057] Level 1 verification: Parse the cluster action commands output by the large language model. If parsing fails, the large language model is required to regenerate the commands. If parsing succeeds, proceed to Level 2 verification.
[0058] Level 2 verification: Check the parameter range of the cluster action command. If a parameter is out of bounds, automatically correct it to the closest valid value and perform Level 3 verification.
[0059] Level 3 verification: Check the unit ID. If an illegal unit ID is found, ignore the corresponding action instruction and log it.
[0060] Level 4 verification: Perform a feasibility check on the actions in the instruction. If the action is within the unit action space and meets the current environmental conditions, the action is executable; otherwise, the action is deemed infeasible and the instruction is ignored.
[0061] Level 5 verification: The instructions are comprehensively verified. After confirming that there are no other errors in the instructions, a legal and executable cluster action instruction is obtained.
[0062] Step S500: Instruction Execution and Status Management: Convert legally executable cluster action instructions into action signals that can be responded to by the underlying execution units, manage the execution process of cluster action instructions through a state machine, and feed back the unit status update information to the environmental situation textualization step to form a closed loop.
[0063] In this embodiment, the actions defined in the action space include Move, Attack, and Wait, and the dynamic tactical principles include encirclement and flanking cooperation. Correspondingly, instruction execution and state management include the following operations:
[0064] (1) Command conversion: For movement commands, the target coordinates are converted into unit heading angles using the heading angle calculation formula;
[0065] (2) State machine management: Three states are set: Move, Attack, and Wait. The state machine switches states according to the progress of instruction execution and environmental changes. In the Move state, the unit moves towards the target position according to the converted heading angle. When it reaches the threshold distance from the target position... When the target disappears, the unit switches to the Wait state. In the Attack state, the unit continuously tracks the target and performs attack actions until the target disappears. Then, the unit switches to the Wait state. In the Wait state, the unit is in standby mode and ready to respond to new commands at any time.
[0066] The unit's position update law in three-dimensional space is calculated based on a kinematic model and used to continuously adjust the trajectory and height. The calculation formula is expressed as follows:
[0067] ,
[0068] ,
[0069] ,
[0070] in, , and Units exist The three-dimensional coordinates in the Cartesian coordinate system at all times. , and Units exist The three-dimensional coordinates in the Cartesian coordinate system at all times. Indicates the sampling time interval. Unit exist Instantaneous velocity at a given moment Units exist The heading angle at any moment, Units exist The pitch angle at any given moment;
[0071] The formula for determining the attack trigger condition is expressed as follows:
[0072]
[0073] The attack trigger condition judgment formula is used to determine whether a unit has entered an attack state, realizing condition switching based on distance and angle. Units exist Time and Goal Horizontal distance in the XY plane Units The attack horizontal distance threshold, Units exist Time and Goal Azimuth deviation in the XY plane Units The attack azimuth deviation threshold, Units The previous moment towards the target The moment of launching the attack, Units Attack cooldown time, express Time unit The number of attack resources available;
[0074] The formula for calculating the command-navigation angle conversion is:
[0075] ,
[0076] ,
[0077] The command-navigation angle conversion formula is used to convert the target coordinates into the desired heading angle that needs to be adjusted per unit, ensuring the accuracy of command execution. express Time unit The expected heading angle express Time unit The expected pitch angle, , and Units The current three-dimensional coordinates, , and Units exist The three-dimensional coordinates of the time in the spatial rectangular coordinate system.
[0078] This embodiment's method constructs a COA-LLM framework, combining environmental situation textualization, multi-level instruction parsing, and prompt word engineering to achieve real-time collaborative decision-making for clusters in dynamic environments. This method integrates target allocation, path planning, and risk assessment into a unified semantic-driven framework, utilizing LLM to parse natural language instructions and generate control commands. It overcomes the limitation of traditional optimization algorithms that only support discrete task allocation, achieving a unification of discrete task and continuous space decision-making. Experiments show that this method significantly improves decision-making efficiency and adaptability in simulation environments, with strong input flexibility and output interpretability.
[0079] Example 2:
[0080] This invention discloses a cluster collaborative intelligent decision-making system based on a large language model, comprising an environmental information textification module, a multi-level prompt word construction module, an instruction generation module, an instruction multi-level verification module, and an instruction execution and status management module.
[0081] The environmental information textification module is used to perform the following operations: collect environmental information of the cluster's operating environment, and convert the environmental information into structured environmental text adapted to the large language model. The environmental information includes the ID, type, location, and status information of each unit, and the units include friendly units and enemy units.
[0082] When texturing environmental information, the environmental information is converted into JSON format for output. The format of the large language model's instruction output is constrained to JSON format by the text specification, for example: {"Our unit": [{"id":14, "type":"PlaneA","position":[362.41,355.88,14.78]}], "Enemy unit": [{"id":-6, "type":"APS","position":[500.00,500.00,0.00]}]}.
[0083] The multi-level prompt word construction module is used to perform the following operations: construct multi-level prompt words containing role shaping, agent information, text specifications, and prompt information. In this module, the large language model is defined as the cluster decision-making center through role shaping, the action space of the unit is declared through agent information (e.g., "Act Space": ["Move(unit ID, target coordinates)", "Attack(unit ID, target ID)"]) through agent information, the format of the large language model's instruction output is constrained through text specifications (JSON template), and dynamic tactical principles (such as encirclement and flanking cooperation) are injected into the large language model through prompt information.
[0084] The instruction generation module performs the following operations: inputting structured environment text and multi-level prompts into the large language model, which performs semantic understanding and reasoning, parses task requirements, matches dynamic tactical principles, generates corresponding cluster action instructions, and outputs the cluster action instructions in a predetermined format based on text specifications. Among these, the cluster action instructions include discrete task allocation and continuous space control instructions, for example, the instruction is: {"task_allocation": ["Move(11,[500,500,10])", "Attack(31, -6)"]}.
[0085] The multi-level instruction verification module is used to perform the following operations: parse and verify the cluster action instructions output by the large language model, and correct errors based on the verification results to obtain legal and executable cluster action instructions.
[0086] As a specific implementation of the instruction multi-level verification module, this module is used to perform the following operations:
[0087] Level 1 verification: Parse the cluster action commands output by the large language model. If parsing fails, the large language model is required to regenerate the commands. If parsing succeeds, proceed to Level 2 verification.
[0088] Level 2 verification: Check the parameter range of the cluster action command. If a parameter is out of bounds, automatically correct it to the closest valid value and perform Level 3 verification.
[0089] Level 3 verification: Check the unit ID. If an illegal unit ID is found, ignore the corresponding action instruction and log it.
[0090] Level 4 verification: Perform a feasibility check on the actions in the instruction. If the action is within the unit action space and meets the current environmental conditions, the action is executable; otherwise, the action is deemed infeasible and the instruction is ignored.
[0091] Level 5 verification: The instructions are comprehensively verified. After confirming that there are no other errors in the instructions, a legal and executable cluster action instruction is obtained.
[0092] The instruction execution and status management module is used to perform the following operations: convert legally executable cluster action instructions into action signals that can be responded to by the underlying execution units, manage the execution process of cluster action instructions through a state machine, and feed back the unit status update information to the environmental situation textification step to form a closed loop.
[0093] In this embodiment, the actions defined in the action space include Move, Attack, and Wait, and the dynamic tactical principles include encirclement and flanking cooperation. Correspondingly, instruction execution and state management include the following operations:
[0094] (1) Command conversion: For movement commands, the target coordinates are converted into unit heading angles using the heading angle calculation formula;
[0095] (2) State machine management: Three states are set: Move, Attack, and Wait. The state machine switches states according to the progress of instruction execution and environmental changes. In the Move state, the unit moves towards the target position according to the converted heading angle. When it reaches the threshold distance from the target position... When the target disappears, the unit switches to the Wait state. In the Attack state, the unit continuously tracks the target and performs attack actions until the target disappears. Then, the unit switches to the Wait state. In the Wait state, the unit is in standby mode and ready to respond to new commands at any time.
[0096] The unit's position update law in three-dimensional space is calculated based on a kinematic model and used to continuously adjust the trajectory and height. The calculation formula is expressed as follows:
[0097] ,
[0098] ,
[0099] ,
[0100] in, , and Units exist The three-dimensional coordinates in the Cartesian coordinate system at all times. , and Units exist The three-dimensional coordinates in the Cartesian coordinate system at all times. Indicates the sampling time interval. Unit exist Instantaneous velocity at a given moment Units exist The heading angle at any moment, Units exist The pitch angle at any given moment;
[0101] The formula for determining the attack trigger condition is expressed as follows:
[0102]
[0103] The attack trigger condition judgment formula is used to determine whether a unit has entered an attack state, realizing condition switching based on distance and angle. Units exist Time and Goal Horizontal distance in the XY plane Units The attack horizontal distance threshold, Units exist Time and Goal Azimuth deviation in the XY plane Units The attack azimuth deviation threshold, Units The previous moment towards the target The moment of launching the attack, Units Attack cooldown time, express Time unit The number of attack resources available;
[0104] The formula for calculating the command-navigation angle conversion is:
[0105] ,
[0106] ,
[0107] The command-navigation angle conversion formula is used to convert the target coordinates into the desired heading angle that needs to be adjusted per unit, ensuring the accuracy of command execution. express Time unit The expected heading angle express Time unit The expected pitch angle, , and Units The current three-dimensional coordinates, , and Units exist The three-dimensional coordinates of the time in the spatial rectangular coordinate system.
[0108] The system in this embodiment can execute the method disclosed in Embodiment 1 to achieve cluster collaborative intelligent decision-making.
[0109] The above provides a detailed description of the cluster collaborative intelligent decision-making method and system based on a large language model provided by this invention. Specific examples have been used to illustrate the principles and implementation methods of this invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of this invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this invention. Therefore, the content of this specification should not be construed as a limitation of this invention.
Claims
1. A cluster collaborative intelligent decision-making method based on a large language model, characterized in that, Includes the following steps: Environmental information textification: Collect environmental information of the cluster's operating environment and convert it into structured environmental text that is adapted to the large language model. The environmental information includes the ID, type, location, and status information of each unit, including friendly units and enemy units. Multi-level cue word construction: Construct multi-level cue words that include role shaping, agent information, text specifications and cue information. In this process, the large language model is defined as the cluster decision-making center through role shaping, the action space of the unit is declared through agent information, the format of the large language model's command output is constrained through text specifications, and dynamic tactical principles are injected into the large language model through cue information. Command generation: The structured environment text and multi-level prompts are input into the large language model. The large language model performs semantic understanding and reasoning, parses task requirements, matches dynamic tactical principles, generates corresponding cluster action commands, and outputs the cluster action commands in a predetermined format based on text specifications. Among them, the cluster action commands include discrete task allocation and continuous space control commands. Multi-level instruction verification: The cluster action instructions output by the large language model are parsed and verified, and errors are corrected based on the verification results to obtain legal and executable cluster action instructions. Command execution and state management: Convert legally executable cluster action commands into action signals that can be responded to by the underlying execution units, manage the execution process of cluster action commands through a state machine, and feed back the unit state update information to the environmental situation textification step to form a closed loop.
2. The cluster collaborative intelligent decision-making method based on a large language model according to claim 1, characterized in that, When environmental information is textualized, it is converted into JSON format for output. The format of the large language model's instructions is constrained to be JSON format by the text specification.
3. The cluster collaborative intelligent decision-making method based on a large language model according to claim 1, characterized in that, Multi-level instruction verification includes the following operations: Level 1 verification: Parse the cluster action commands output by the large language model. If parsing fails, the large language model is required to regenerate the commands. If parsing succeeds, proceed to Level 2 verification. Level 2 verification: Check the parameter range of the cluster action command. If a parameter is out of bounds, automatically correct it to the closest valid value and perform Level 3 verification. Level 3 verification: Check the unit ID. If an illegal unit ID is found, ignore the corresponding action instruction and log it. Level 4 verification: Perform a feasibility check on the actions in the instruction. If the action is within the unit action space and meets the current environmental conditions, the action is executable; otherwise, the action is deemed infeasible and the instruction is ignored. Level 5 verification: The instructions are comprehensively verified. After confirming that there are no other errors in the instructions, a legal and executable cluster action instruction is obtained.
4. The cluster collaborative intelligent decision-making method based on a large language model according to claim 1, characterized in that, Actions defined in the action space include Move, Attack, and Wait; dynamic tactical principles include encirclement and flanking cooperation. Instruction execution and status management include the following operations: Command conversion: For movement commands, the target coordinates are converted into unit heading angles using the heading angle calculation formula; State machine management: Three states are set: Move, Attack, and Wait. The state machine switches states based on the progress of instruction execution and environmental changes. In the Move state, the unit moves towards the target position according to the converted heading angle. When it reaches a threshold distance from the target position... When the target disappears, the unit switches to the Wait state. In the Attack state, the unit continuously tracks the target and performs attack actions until the target disappears. Then, the unit switches to the Wait state. In the Wait state, the unit is in standby mode and ready to respond to new commands at any time.
5. A cluster collaborative intelligent decision-making system based on a large language model, characterized in that, It includes an environmental information textification module, a multi-level prompt word construction module, an instruction generation module, a multi-level instruction verification module, and an instruction execution and status management module; The environmental information textification module is used to perform the following operations: collect environmental information of the cluster's operating environment, and convert the environmental information into structured environmental text adapted to the large language model. The environmental information includes the ID, type, location, and status information of each unit, and the units include friendly units and enemy units. The multi-level prompt word construction module is used to perform the following operations: construct multi-level prompt words that include role shaping, agent information, text specifications and prompt information. In this module, the large language model is defined as the cluster decision-making center through role shaping, the action space of the unit is declared through agent information, the format of the large language model's instruction output is constrained through text specifications, and dynamic tactical principles are injected into the large language model through prompt information. The instruction generation module is used to perform the following operations: inputting structured environment text and multi-level prompt words into the large language model, the large language model performs semantic understanding and reasoning, parses task requirements, matches dynamic tactical principles, generates corresponding cluster action instructions, and outputs the cluster action instructions in a predetermined format based on text specifications. Among them, the cluster action instructions include discrete task allocation and continuous space control instructions. The multi-level instruction verification module is used to perform the following operations: parse and verify the cluster action instructions output by the large language model, and correct errors based on the verification results to obtain legal and executable cluster action instructions; The instruction execution and status management module is used to perform the following operations: converting legally executable cluster action instructions into action signals that can be responded to by the underlying execution units, managing the execution process of cluster action instructions through a state machine, and feeding back unit status update information to the environmental situation textification step to form a closed loop.
6. The cluster collaborative intelligent decision-making system based on a large language model according to claim 5, characterized in that, The Environmental Information Textification module is used to convert environmental information into JSON format for output. The format of the large language model's instructions is constrained to be JSON format by the text specification.
7. The cluster collaborative intelligent decision-making system based on a large language model according to claim 5, characterized in that, The multi-level instruction verification module is used to perform the following operations: Level 1 verification: Parse the cluster action commands output by the large language model. If parsing fails, the large language model is required to regenerate the commands. If parsing succeeds, proceed to Level 2 verification. Level 2 verification: Check the parameter range of the cluster action command. If a parameter is out of bounds, automatically correct it to the closest valid value and perform Level 3 verification. Level 3 verification: Check the unit ID. If an illegal unit ID is found, ignore the corresponding action instruction and log it. Level 4 verification: Perform a feasibility check on the actions in the instruction. If the action is within the unit action space and meets the current environmental conditions, the action is executable; otherwise, the action is deemed infeasible and the instruction is ignored. Level 5 verification: The instructions are comprehensively verified. After confirming that there are no other errors in the instructions, a legal and executable cluster action instruction is obtained.
8. The cluster collaborative intelligent decision-making system based on a large language model according to claim 5, characterized in that, Actions defined in the action space include Move, Attack, and Wait; dynamic tactical principles include encirclement and flanking cooperation. The instruction execution and status management module is used to perform the following operations: Command conversion: For movement commands, the target coordinates are converted into unit heading angles using the heading angle calculation formula; State machine management: Three states are set: Move, Attack, and Wait. The state machine switches states based on the progress of instruction execution and environmental changes. In the Move state, the unit moves towards the target position according to the converted heading angle. When it reaches a threshold distance from the target position... When the target disappears, the unit switches to the Wait state. In the Attack state, the unit continuously tracks the target and performs attack actions until the target disappears. Then, the unit switches to the Wait state. In the Wait state, the unit is in standby mode and ready to respond to new commands at any time.