Unmanned aerial vehicle cluster maritime penetration decision-making method, system, equipment and medium
By combining a large language model and a multi-agent reinforcement learning model, a hierarchical decision-making architecture is constructed, which resolves the conflict between strategic and tactical decision-making in UAV swarm maritime penetration. This enables efficient and reliable generation of maritime penetration strategies and rapid adversarial adaptation, making it suitable for UAV swarm maritime penetration missions.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- XIAN TECH UNIV
- Filing Date
- 2026-01-13
- Publication Date
- 2026-05-08
AI Technical Summary
Existing UAV swarm maritime penetration path planning technologies suffer from conflicts in strategic and tactical decision-making, making it difficult to cope with the complex and dynamic modern naval warfare environment. They also lack prior military knowledge, have low training efficiency, limited computing resources, cannot achieve efficient edge deployment, and have poor adaptability to adversarial situations.
Long-term strategic plans are generated using a large language model (LLM) and instantaneous tactical strategies are generated using a multi-agent reinforcement learning model (MARL). A hierarchical decision-making architecture of "large and small model collaboration" is constructed. By combining knowledge distillation and meta-learning, the decoupling of strategy and tactics and online adaptive optimization are achieved. The lightweight model is deployed on the edge computing platform of UAVs.
It enables efficient and reliable decision-making for UAV swarms in complex naval warfare environments, allowing them to quickly adapt to changes in enemy strategy, ensure tactical survivability and strategic optimization, and meet the requirements of real-time performance and robustness.
Smart Images
Figure CN121995932A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of intelligent decision-making technology, and in particular to a method, system, device and medium for decision-making regarding unmanned aerial vehicle (UAV) swarms for maritime penetration. Background Technology
[0002] Currently, in the planning of maritime penetration paths for UAV swarms, the cooperative path planning method based on multi-agent reinforcement learning is widely adopted and is considered a mainstream cutting-edge technology for solving such complex decision-making problems. The core architecture of this technical solution is typically as follows: The maritime penetration mission is modeled as a partially observable Markov decision process. Each UAV in the system acts as an agent, aiming to learn a cooperative strategy through interaction with the environment (trial and error) to maximize the cumulative reward for completing the penetration mission. It usually includes three modules: a situational awareness module, which uses convolutional neural networks or recurrent neural networks to process the individual's local observation information (such as its own position and sensor data); a cooperative decision-making module, which uses a centralized training and distributed execution framework, such as the MADDPG algorithm, to utilize global information during training to guide the optimization of individual strategies and learn swarm cooperative behavior; and a path execution module, which translates the actions (such as heading and speed commands) output by the decision-making module into specific control signals.
[0003] However, the algorithmic framework of the aforementioned scheme employs a "homogeneous single network" at the decision-making architecture level to handle multi-scale spatiotemporal problems. Specifically, it uses the same neural network to handle both long-term mission planning (such as selecting a penetration corridor) and instantaneous tactical maneuvers (such as evading missiles). This coupled architecture requires network parameters to be traded between conflicting objectives, often leading to "short-term optimal rather than long-term optimal" decisions. For example, the agent may learn overly conservative tactics, deviating significantly from the optimal strategic route to avoid every instantaneous threat; or conversely, it may ignore immediate, deadly threats in order to remain faithful to the macroscopic route. This results in a misalignment between strategic and tactical decisions, limiting its practical application effectiveness in unmanned aerial vehicle (UAV) swarm maritime penetration missions. Summary of the Invention
[0004] The purpose of this invention is to provide a method, system, device, and medium for decision-making regarding unmanned aerial vehicle (UAV) swarms at sea, which can solve the aforementioned technical problems.
[0005] This invention provides a method for decision-making regarding unmanned aerial vehicle (UAV) swarm penetration at sea, comprising the following steps: To acquire raw battlefield environment information and the mission objectives of the drone swarm; Based on battlefield environment information and mission objectives, the large language model retrieves and reasons from a pre-set military knowledge base to generate long-term strategic plans and corresponding strategic reward constraints. Based on the long-term strategic plan, a simulation of the UAV swarm's maritime penetration was conducted. Using a multi-agent reinforcement learning model, the instantaneous tactical strategy of the UAV swarm was determined based on strategic reward constraints and battlefield environment information during the UAV swarm's penetration during the simulation. The simulation aims to assess the performance of long-term strategic plans and instantaneous tactical strategies in combat against new enemy tactical tactics and to identify the threats faced. Based on the performance of long-term strategic plans and instantaneous tactical strategies in combat and the threats faced, the long-term strategic plan is adjusted using a large language model. Then, the process of unmanned aerial vehicle swarms penetrating at sea in different battlefield environments is simulated based on the adjusted long-term strategic plan to obtain multiple instantaneous tactical strategies. Based on the adjusted long-term strategic plan and the best-performing instantaneous tactical strategy among multiple instantaneous tactical strategies in combat, a maritime penetration strategy for drone swarms is determined.
[0006] Furthermore, the process of generating long-term strategic plans and corresponding strategic reward constraints by retrieving and reasoning from a pre-set military knowledge base based on battlefield environment information and mission objectives using a large language model includes: Based on battlefield environment information and mission objectives, the large language model retrieves similar battle examples and corresponding tactical clauses from the military knowledge base. Based on the retrieved battle examples and corresponding tactical clauses, it infers long-term strategic plans and corresponding strategic reward constraints.
[0007] Furthermore, the long-term strategic plan includes multiple penetration paths, grouping suggestions and action sequences for drone swarms, and a plan for the use of the electromagnetic spectrum; The strategic reward constraint refers to additional rewards that conform to military principles, based on the achievement of mission objectives by the drone swarm.
[0008] Furthermore, the determination of the maritime penetration strategy for the drone swarm based on the adjusted long-term strategic plan and the best-performing instantaneous tactical strategy among multiple instantaneous tactical strategies in combat includes: Based on the adjusted long-term strategic plan and the best-performing instantaneous tactical strategy among multiple instantaneous tactical strategies in combat, the multi-agent reinforcement learning model is adjusted. Based on the knowledge distillation method, a large language model and an adjusted multi-agent reinforcement learning model are used as teacher models to train student models. By using trained student models, we can determine maritime penetration strategies for different drone swarms in various battlefield environments.
[0009] Furthermore, the trained student model is deployed on each drone in the drone swarm.
[0010] This invention also provides a decision-making system for unmanned aerial vehicle (UAV) swarm maritime penetration, comprising the following modules: The data acquisition module is used to acquire raw battlefield environment information and the mission objectives of the drone swarm; The strategy generation module is used to retrieve and reason from a pre-set military knowledge base based on battlefield environment information and mission objectives using a large language model, in order to generate long-term strategic plans and corresponding strategic reward constraints. The tactical generation module is used to simulate the maritime penetration of UAV swarms based on long-term strategic plans. It also uses a multi-agent reinforcement learning model to determine the instantaneous tactical strategies of the UAV swarms based on strategic reward constraints and battlefield environment information during the penetration process. The strategic assessment module is used to simulate the process of UAV swarms penetrating the sea under the new tactical strategies of the enemy, through long-term strategic plans and instantaneous tactical strategies, in order to evaluate the performance of long-term strategic plans and instantaneous tactical strategies in combat and identify the threats faced. The strategic adjustment module is used to adjust the long-term strategic plan based on the performance of the long-term strategic plan and the instantaneous tactical strategy in combat and the threats faced, through a large language model. It also simulates the maritime penetration process of the UAV swarm in different battlefield environments based on the adjusted long-term strategic plan to obtain multiple instantaneous tactical strategies. The intelligent decision-making module is used to determine the maritime penetration strategy of the drone swarm based on the adjusted long-term strategic plan and the best-performing instantaneous tactical strategy among multiple instantaneous tactical strategies in combat.
[0011] Embodiments of the present invention also provide a computer device, including: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to execute the above-described unmanned aerial vehicle (UAV) swarm maritime penetration decision-making method.
[0012] Embodiments of the present invention also provide a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the above-described method for decision-making regarding unmanned aerial vehicle (UAV) swarm maritime penetration.
[0013] The UAV swarm maritime penetration decision-making method provided by this invention has at least the following beneficial effects: This invention designs a hierarchical decision-making architecture of "large and small model collaboration." First, a large language model with military cognitive capabilities (i.e., the large model) is used to generate long-term strategic plans (such as penetration corridor selection and troop deployment timing) and corresponding strategic reward constraints. Then, a multi-agent reinforcement learning model (i.e., the small model) generates instantaneous tactical strategies (such as real-time obstacle avoidance and interference evasion). This decouples long-term mission planning from real-time maneuver control, achieving a high degree of unity between macro-strategic optimization and micro-survivability, overcoming the scale conflict defects of single models. Simultaneously, by evaluating the performance of long-term strategic plans and instantaneous tactical strategies in combat and identifying and optimizing the threats faced, the final generated maritime penetration strategy can adapt to changes in enemy strategies and maintain a decision-making advantage. Therefore, the decision-making method of this scheme can effectively guarantee the practical application effectiveness of UAV swarm maritime penetration missions. Attached Figure Description
[0014] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this invention, illustrate exemplary embodiments of the invention and are used to explain the invention, but do not constitute an undue limitation of the invention. In the drawings:
[0015] Figure 1 This is a schematic diagram of a pathfinding algorithm using the A* algorithm provided by the present invention. Figure 2 A schematic diagram of the spatiotemporal coupling degree of strategy and tactics provided for this invention; Figure 3 A schematic diagram illustrating a hierarchical decision-making model based on size, as provided by this invention; Figure 4 A flowchart illustrating a decision-making method for unmanned aerial vehicle (UAV) swarming at sea provided by this invention. Figure 1 ; Figure 5 A flowchart illustrating a decision-making method for unmanned aerial vehicle (UAV) swarming at sea provided by this invention. Figure 2 . Detailed Implementation
[0016] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.
[0017] Currently, the planning of maritime penetration paths for drone swarms mainly relies on traditional deterministic algorithms and single-architecture intelligent algorithms. While existing solutions can accomplish mission planning to a certain extent, they suffer from the following significant technical shortcomings when facing the complex, dynamic, and highly contested environment of modern naval warfare:
[0018] In terms of situational awareness and threat modeling, existing technologies suffer from staticity and limitations. Traditional path planning algorithms (such as the A* algorithm) typically simplify enemy threats (such as radar detection range and air defense fire zones) into a static, binary obstacle map. Figure 1 As shown, its path planning can only avoid static threat areas and cannot cope with real-time tactical behaviors such as enemy radar switching on / off, ship maneuvers, and dynamic changes in the electromagnetic spectrum. This leads to a significant reduction in the survival probability of the planned path during actual execution. In addition, the perception modules of existing systems are mostly designed for single data sources (such as processing only geographic information or only radar signals), lacking the ability to fuse multi-source heterogeneous information such as satellite imagery, electronic reconnaissance, and communications intelligence, making it difficult to construct a comprehensive and accurate real-time battlefield situation map.
[0019] In terms of decision-making architecture, existing intelligent planning methods based on reinforcement learning or deep reinforcement learning suffer from the problem of "spatiotemporal scale coupling." For example... Figure 2 As shown, a single neural network model needs to be responsible for both long-term strategic planning (such as the overall trajectory of the entire penetration route) and short-term tactical maneuvers (such as real-time evasion of incoming missiles), which often leads to decision conflicts. For example, a drastic maneuver to avoid a momentary threat may cause the cluster to deviate from the predetermined strategic route or even enter other unknown high-threat areas. This architecture struggles to achieve a good balance between the long-term optimization of strategic objectives and the short-term survival of tactical actions.
[0020] In terms of adaptability and intelligence, existing systems are rigid. Most rule-based or traditional machine learning methods rely on pre-defined and fixed decision-making logic. Once an adversary defense system adopts new and unforeseen tactics (such as novel jamming patterns, decoy strategies, or dynamic fire coordination), existing planning systems cannot learn and adjust their strategies online, making them highly vulnerable to suppression and deception. The root cause lies in the lack of mechanisms to incorporate prior knowledge of the military domain (such as tactical doctrines and historical battle examples) and the ability to evolve strategies in continuous confrontation with high-level adversaries.
[0021] At the engineering implementation level, there is a significant contradiction between the computing power requirements of existing high-performance intelligent algorithms and the limited computing resources of UAV platforms. Large-parameter models capable of handling complex planning problems (such as large neural networks) typically cannot be deployed on resource-constrained UAV edge computing nodes and must instead rely on centralized computing at remote command centers. This introduces unacceptable communication latency, and communication link interruptions in environments with strong electromagnetic interference can even cause the cluster to "disable," severely violating the autonomy and robustness requirements of penetration missions.
[0022] Therefore, existing UAV swarm maritime penetration path planning technology is insufficient to meet the core requirements of modern intelligent naval warfare for dynamic adaptability, decision-making intelligence, cross-modal fusion, and edge practicality, and a new solution is urgently needed.
[0023] While existing MARL models can solve such complex decision-making problems, they have several fundamental flaws, stemming from the inherent limitations of their algorithmic framework, as follows: At the knowledge utilization level, this approach suffers from the limitation of "learning from scratch," resulting in low learning efficiency and a lack of tactical rationality in the strategy. The MARL model relies entirely on extensive trial and error in a simulated environment to accumulate experience from zero, with its policy network's initial weights being randomly generated. This leads to the model requiring millions of simulation iterations to converge to a usable policy, resulting in extremely high training costs. More seriously, this learning method lacks the embedding and utilization of prior military knowledge (such as tactical regulations, historical battle examples, and electromagnetic spectrum characteristics). Therefore, while the trained strategy may mathematically optimize the reward function, it often exhibits behaviors that violate basic tactical principles and contradict military common sense (such as traversing high-risk areas in unreasonable dense formations), casting doubt on the reliability and interpretability of its decision-making results. The fundamental reason lies in the fact that the standard reinforcement learning paradigm itself lacks an effective mechanism for introducing and utilizing structured prior knowledge.
[0024] At the decision-making architecture level, this scheme employs a "homogeneous single network" to handle multi-scale spatiotemporal problems, leading to a misalignment between strategic and tactical decisions. Existing schemes use the same neural network to handle both long-term mission planning (such as selecting penetration corridors) and instantaneous tactical maneuvers (such as missile evasion). This coupled architecture forces network parameters to trade off between conflicting objectives, often resulting in "short-term optimal rather than long-term optimal" decisions. For example, the agent may learn overly conservative tactics, deviating significantly from the optimal strategic route to avoid every instantaneous threat; or conversely, ignoring immediate, deadly threats in order to remain faithful to the macroscopic route. This deficiency stems from the inherent design of the MARL model architecture, whose network structure does not explicitly model and separate the different time scales inherent in the decision-making task.
[0025] At the adversarial adaptation level, this approach suffers from significant policy degradation risk, making it ill-equipped to cope with the policy evolution of adversary agents. Under a self-game-based training mechanism, the MARL agent's policy may converge to a narrow local optimum, i.e., a "vulnerable" policy optimized specifically for the particular adversary policies encountered during training. Once it encounters a real-world adversarial defense system with machine learning capabilities and the ability to dynamically adjust tactics (whose policy distribution differs from that during training), the performance of the existing approach will significantly degrade. The underlying reason is that standard adversarial training lacks a "policy brain" with strong generalization and meta-learning capabilities to quickly evaluate new adversary policies and generate corresponding countermeasures, resulting in an adaptation speed far slower than the adversary's rate of change.
[0026] At the engineering deployment level, this solution suffers from a fundamental contradiction between computational complexity and the requirements of distributed deployment. To achieve satisfactory collaborative decision-making performance, existing MARL models typically require deep and large-scale parameter networks, which places computational demands far exceeding the capacity of current UAV-borne computing platforms. Therefore, in practical applications, a "cloud-based decision-making - terminal execution" model is often necessary, where perception data is transmitted back to the command center for centralized decision-making before instructions are issued to the cluster. This model is extremely risky in the intense electromagnetic warfare environment of modern naval warfare, where delays, interruptions, or deception of communication links can directly lead to the loss of cluster combat capabilities. This engineering bottleneck stems from the end-to-end reinforcement learning model's over-reliance on parameter scale, while the model itself lacks a fundamentally lightweight design for resource-constrained edge computing environments.
[0027] In conclusion, although the scheme is theoretically feasible, its inherent defects in knowledge utilization, decision-making structure, adversarial adaptability, and engineering implementation limit its practical application effectiveness in real, highly dynamic, and highly adversarial UAV swarm maritime penetration missions.
[0028] This invention aims to fundamentally solve the aforementioned technical bottlenecks encountered when applied in real battlefield environments. The purpose of this invention is not to make partial improvements to existing methods, but to provide a completely new architectural paradigm, specifically designed to solve the following five interrelated technical problems: First, addressing the problems of existing reinforcement learning models lacking prior military knowledge, exhibiting low training efficiency, and insufficient decision-making rationality, this invention aims to solve the key technical problem of constructing an intelligent decision-making framework that integrates military domain knowledge with the cognitive capabilities of a large model. The traditional MARL method's "zero-based learning" model cannot utilize rich battle examples, doctrines, and tactical rules, resulting in a "black box" decision-making process with slow convergence. The primary objective of this invention is to innovatively introduce a large model (LLM) with military cognitive capabilities as a "strategic brain." By encoding and reasoning about structured knowledge (such as historical battle examples, tactical manuals, and electromagnetic spectrum characteristics), it provides UAV swarms with initial feasible solutions and strategic constraints that conform to military principles, thereby significantly improving the initial quality, interpretability, and training efficiency of decision-making.
[0029] Secondly, addressing the challenge of coordinating strategic planning and tactical actions across multiple scales in a single decision-making network, this invention aims to solve the problem of seamless decoupling and coordination between long-term mission planning and real-time maneuver control. Existing solutions suffer from decision-making conflicts due to the coupling of strategy and tactics. The core objective of this invention is to propose a hierarchical decision-making architecture based on "coordination between large and small models," such as... Figure 3 As shown, by decoupling tasks, the large model is responsible for hourly-level macro-strategic planning (such as the selection of penetration corridors and the timing of troop deployment), while the lightweight small model is responsible for second-level tactical action optimization (such as real-time obstacle avoidance and interference avoidance). This ensures that the decision-making system achieves a high degree of unity between macro-strategic optimization and micro-survivability, overcoming the scale conflict defects of a single model.
[0030] Third, addressing the shortcomings of existing systems in slow strategy adaptation and easy degradation when facing intelligent and dynamic adversaries, this invention aims to solve the problem of how to achieve continuous online evolution and adversarial adaptability of decision-making strategies in highly adversarial environments. Existing MARL relies on static self-games, which is difficult to cope with unknown tactics. The purpose of this invention is to construct a dynamic adversarial evolution mechanism by utilizing the strong generalization and meta-learning capabilities of large models. That is, the large model quickly infers and generates a set of candidate strategies to deal with new threats, and then the small model quickly verifies and fine-tunes them in a simulation environment, forming an online learning closed loop of "observation-evaluation-adjustment," enabling the system to continuously adapt to changes in the enemy's strategies and maintain a decision-making advantage.
[0031] Fourth, addressing the engineering bottleneck of existing intelligent algorithms' high computing power requirements and difficulty in deployment at the resource-constrained edge of UAVs, this invention aims to solve the problem of how to efficiently deploy high-performance intelligent decision-making capabilities at tactical edge nodes. Cloud-based decision-making models cannot meet the real-time and robustness requirements of highly adversarial environments. The ultimate goal of this invention is to compress and transfer the strategic cognitive capabilities inherent in large models to lightweight small models through technological innovations such as knowledge distillation and model quantization. This allows the small models ultimately deployed on UAV onboard computing platforms to enjoy the strategic vision of large models while meeting the requirements of low power consumption and low latency real-time computing, truly realizing the practical application of artificial intelligence decision-making capabilities at the edge.
[0032] In summary, this invention aims to systematically solve the full-chain technical challenges from knowledge utilization, decision coordination, adversarial adaptation to engineering deployment through a novel architecture based on large and small model collaboration, providing an efficient, reliable and practical solution for intelligent maritime penetration by UAV swarms.
[0033] The technical solutions provided by the various embodiments of the present invention will be described in detail below with reference to the accompanying drawings.
[0034] One embodiment of the present invention relates to a decision-making method for unmanned aerial vehicle (UAV) swarms to penetrate maritime defenses. The specific process of the UAV swarms maritime penetration decision-making method in this embodiment can be as follows: Figure 4 As shown, it includes: Step 401: Obtain the original battlefield environment information and the mission objectives of the drone swarm.
[0035] Step 402: Based on battlefield environment information and mission objectives, the large language model performs retrieval and reasoning from a pre-set military knowledge base to generate long-term strategic plans and corresponding strategic reward constraints.
[0036] Step 403: Simulate the maritime penetration of the UAV swarm based on the long-term strategic plan, and determine the instantaneous tactical strategy of the UAV swarm by using a multi-agent reinforcement learning model, based on strategic reward constraints and battlefield environment information of the UAV swarm during penetration in the simulation process.
[0037] Step 404: Simulate the process of the drone swarm penetrating the sea under the enemy's new tactical strategies through long-term strategic plans and instantaneous tactical strategies, in order to evaluate the performance of long-term strategic plans and instantaneous tactical strategies in combat and identify the threats faced.
[0038] Step 405: Based on the performance of long-term strategic plans and instantaneous tactical strategies in combat and the threats faced, the long-term strategic plan is adjusted using a large language model. Then, the process of the UAV swarm penetrating the sea in different battlefield environments is simulated based on the adjusted long-term strategic plan to obtain multiple instantaneous tactical strategies.
[0039] Step 406: Based on the adjusted long-term strategic plan and the best-performing instantaneous tactical strategy among multiple instantaneous tactical strategies in combat, determine the maritime penetration strategy for the drone swarm.
[0040] The following is a detailed description of the implementation details of the UAV swarm maritime penetration decision-making method in this embodiment. The following content is only for the convenience of understanding and is not necessary for implementing this solution.
[0041] Military knowledge injection and initial strategy generation are carried out in steps 401 and 402 (offline preparation phase).
[0042] Objective: To address the issue of "learning from scratch" by injecting prior military knowledge into the system and improving the initial strategy at the starting point; Execution entity: Large Language Model (LLM) in the cloud or on-premises; Input: Battlefield environment information (map, threat distribution, weather), mission objectives (starting point, ending point, mission type), structured military knowledge base; Output: Macro-strategic plan, tactical behavior rule set, initial strategy constraints.
[0043] Specifically, it includes the following sub-steps: 1. Situational awareness and information coding: Input: The system receives raw battlefield environment information (satellite imagery, radar detection, electromagnetic spectrum scan results) and converts it into a structured text description (Prompt) that can be understood by LLM through a preprocessing module.
[0044] Example: "Mission area: XX sea area in the East China Sea; Red team defense: 3 radar sites (coordinates...), 2 air defense missile sites (coordinates...), medium electromagnetic interference intensity; Blue team cluster: 10 drones, starting point A, target point B; Mission objective: covert penetration, maximize survivability."
[0045] 2. Military knowledge retrieval and reasoning: Input: After receiving situational information, the LLM retrieves and performs correlation reasoning from its embedded military knowledge base (battle examples, regulations, tactical manuals, electromagnetic characteristics database). That is, the large language model retrieves similar battle examples and corresponding tactical clauses from the military knowledge base based on battlefield environment information and mission objectives, and performs reasoning based on the retrieved battle examples and corresponding tactical clauses to obtain long-term strategic plans and corresponding strategic reward constraints.
[0046] Example: LLM will connect it to electronic suppression battles like the "Beca Valley Air Battle" or tactical clauses like "radar evasion maneuvers," forming a chain of reasoning.
[0047] 3. Generation of macro-strategic plans: Steps: Based on the reasoning results, LLM generates several high-quality macro-strategic alternatives.
[0048] Output content: Path corridors: 2-3 main penetration paths are recommended; Force deployment: Cluster grouping suggestions (e.g., decoy aircraft group, reconnaissance aircraft group, attack aircraft group), action sequence (decoy aircraft go first at T+0, main attack group follows at T+5); Resource allocation: Suggested electromagnetic spectrum usage scheme (when to remain silent, when to actively jam).
[0049] 4. Generation of tactical rules and strategy constraints: Steps: LLM decomposes the macro-level scheme into specific, actionable tactical rules and reward function shaping, providing prior guidance for training small models.
[0050] Output content: Tactical rules: "Terrain cover should be prioritized," "Low-altitude flight should be conducted at the edge of radar detection," and "When encountering tracking radar, jamming should be released immediately and evasive maneuvers should be performed." Reward function constraints: In addition to standard RL rewards (such as survival and arrival), reward items that conform to military principles should be added (such as +rewarding the use of terrain cover and -penalizing the behavior of staying in threatening airspace for a long time).
[0051] In step 403, hierarchical collaborative decision-making and simulation verification are performed (offline training phase).
[0052] Objective: To resolve the scale conflict between strategy and tactics, and to verify and fine-tune the initial scheme generated by LLM; Execution subject: LLM (strategic layer) + lightweight MARL model (tactical layer) + simulation environment; Input: strategic scheme and rules generated by LLM, high-fidelity military simulation environment; Output: trained tactical layer mini-model, verified strategic scheme.
[0053] Specifically, it includes the following sub-steps: 1. Issuance of strategic instructions: Steps: LLM translates the macro-strategic plans it generates into high-level instructions that can be executed by tactical mini-models.
[0054] Example: The LLM issues an instruction to a swarm of drones: "Within 10 minutes, maneuver along corridor α at an altitude of <100m and a speed of 200 knots to waypoint P1."
[0055] 2. Execution of tactical actions: Steps: The lightweight MARL model receives strategic instructions from the LLM and real-time observations of the current environment (friendly locations, enemy threat status, obstacles), and outputs specific control instructions (acceleration, heading, pitch angle).
[0056] Role: The MARL model is responsible for performing second-level optimizations within the "framework" of the LLM plan, achieving real-time obstacle avoidance, formation maintenance, and interference avoidance.
[0057] 3. Simulation environment interaction and verification: Steps: Tactical actions are executed in a high-fidelity simulation environment, which then returns the next state and reward. This verifies the feasibility of the LLM scheme.
[0058] Loop: The MARL model learns quickly how to efficiently complete strategic instructions through traditional RL training (such as PPO and QMIX algorithms) and by utilizing the reward function constraints provided by LLM.
[0059] Dynamic adversarial evolution and meta-learning (online learning phase) are carried out in steps 404 and 405.
[0060] Objective: To address the issues of slow strategy adaptation and easy degradation, and to achieve continuous evolution in adversarial environments; Execution entity: LLM (meta-learner) + simulated adversarial environment + evaluation module; Input: enemy's new tactics, current strategy performance evaluation; Output: evolved new strategy, updated tactical model.
[0061] Specifically, it includes the following sub-steps: 1. Performance Assessment and Threat Identification: Steps: The evaluation module continuously monitors the performance of the small model in simulated adversarial scenarios. When the win rate decreases or the loss rate exceeds a threshold, it is determined that the current strategy may have been adapted by the adversary or a new threat has emerged.
[0062] Example: The enemy has been identified as employing a new "mobile radar decoy" tactic, resulting in a 20% decrease in our penetration success rate.
[0063] 2. Meta-reasoning and candidate strategy generation: Steps: Pass descriptions of failed battles and emerging threats to the LLM. The LLM acts as a meta-learner, quickly inferring enemy tactical weaknesses and generating a new set of candidate countermeasures.
[0064] Example: An LLM might suggest: "Try dividing the cluster into more smaller groups to penetrate from multiple directions simultaneously, saturating its command system." or "Change the electromagnetic silence strategy to implement concentrated jamming at specific times."
[0065] 3. Rapid verification and fine-tuning: Steps: Inject the new strategies generated by LLM as new strategic instructions into the simulation environment. Command small models to rapidly test these new strategies in multiple parallel simulation instances.
[0066] Steps: Select the best-performing strategy and perform rapid fine-tuning on the small model in that strategy direction to update its network parameters. That is, adjust the multi-agent reinforcement learning model based on the adjusted long-term strategic options and the best-performing instantaneous tactical strategy among multiple instantaneous tactical strategies in combat.
[0067] 4. Strategy Updates and Closed-Loop Formation: Steps: Deploy the fine-tuned new policy model to the simulation or even the actual system to complete a policy upgrade. This forms an online learning closed loop of "observation-evaluation-adjustment (LLM inference)-verification-update".
[0068] The final strategy can be obtained by following the steps above.
[0069] In some embodiments, edge deployment and online inference (engineering implementation phase) address the issues of high computing power requirements and difficulty in deployment, thereby enabling the practical application of high-performance intelligent decision-making at the edge.
[0070] Execution entities: knowledge distillation toolchain, UAV onboard computing platform; Input: fully trained and evolved tactical layer small model (teacher model); Output: deployable lightweight model (student model).
[0071] Specifically, it includes the following sub-steps: 1. Knowledge distillation: Steps: The strategic vision of the large model (LLM) and the tactical capabilities of the trained small model (MARL) are used together as "teachers" and distilled into a smaller, more efficient neural network "student". That is, based on the knowledge distillation method, the large language model and the adjusted multi-agent reinforcement learning model are used as teacher models to train the student model.
[0072] Method: The student model mimics the input-output mapping (action distribution) of the teacher model, retaining its core decision-making logic, but significantly reducing model parameters and computational load.
[0073] 2. Model quantization: Steps: Quantize the distilled model to convert the model parameters from 32-bit floating-point numbers to 8-bit integers, further compressing the model size and improving computation speed.
[0074] 3. Edge deployment and real-time inference: Steps: Deploy the final lightweight, quantized model to an onboard embedded platform for the drone (such as Jetson Orin NX).
[0075] Operational Mode: In actual combat, the UAV relies on a local small model for completely offline real-time inference (millisecond-level response) to generate tactical actions, no longer depending on cloud communication, thus meeting the real-time and robustness requirements under intense confrontation. The LLM, on the other hand, continuously evolves its strategy in the rear, distilling and updating the front end with new strategies through intermittent communication when necessary.
[0076] This will help determine the maritime penetration strategies for different drone swarms in different battlefield environments.
[0077] Based on the above embodiments, the UAV swarm maritime penetration decision-making method of the present invention can... Figure 5 The four phases shown are centered on the collaboration and iterative optimization of large and small models. A smart decision-making framework is constructed that decouples "strategy" and "tactics," combines offline and online operations, and coordinates "cloud" and "edge." A large language model acts as the "strategic brain," infused with prior military knowledge, for macro-level planning and meta-reasoning; a lightweight reinforcement learning model acts as the "tactical cerebellum," responsible for real-time maneuver control. Both are continuously iterated and optimized through a dynamic adversarial evolution mechanism, and ultimately, knowledge distillation compresses the capabilities of the large model into a small model, achieving efficient edge deployment.
[0078] The following specific example illustrates how to achieve maritime penetration decision-making for drone swarms through these stages: Background: The blue team's drone swarm attacks the red team's island defense system.
[0079] Blue Team (Attacker): A swarm of 12 low-cost unmanned aerial vehicles (UAVs) divided into 3 groups (4 UAVs each), with electromagnetic sensing and jamming capabilities; Red (Defender): An island base equipped with long-range early warning radar, air defense missile sites, and electronic warfare equipment.
[0080] Mission objective: The blue team cluster starts from point A on the mothership and must break through the red team's defenses within 45 minutes, reach point B (the red team's command post), and execute a simulated attack, while maximizing the cluster's survivability.
[0081] Environment: Nighttime, sea state 3, with scattered small islands and reefs that can be used for terrain cover.
[0082] Phase 1: Military knowledge injection and initial strategy generation (pre-war cloud planning).
[0083] 1. Situational awareness and information coding: The system receives intelligence from multiple sources, including satellites and reconnaissance aircraft, and generates a general battlefield situation map. The structured Prompt received by the LLM is as follows:
[0084] "
Mission Assumptions
Red Force Threats
Environment
Mission Constraints
[0085] 2. Military knowledge retrieval and reasoning: LLM retrieves its knowledge base and performs the following inference: Case study relevance: "This scenario is similar to the 'Beka Valley,' requiring priority to suppress / evade air defense radars."
[0086] The regulation states: "The Electronic Warfare Regulations stipulate that a strategy of 'low-altitude penetration + intermittent silence' should be adopted when dealing with early warning radars."
[0087] Tactical options: "The terrain of the archipelago can be used for covert maneuvers to avoid the radar main lobe."
[0088] Electromagnetic characteristics: "'Eye' radar operates in the S-band, and cluster jamming requires concentrated power in this frequency band."
[0089] 3. Generation of macro-strategic plans: LLM generates two alternative strategic options: Option 1 (Covered Penetration): The entire swarm will fly at extremely low altitudes, using the shadows of islands and reefs to infiltrate in three routes. Electromagnetic silence throughout.
[0090] Option 2 (Suppression and Breakthrough): Groups 1 and 2 act as "decoy / jamming" groups, advancing forward and actively launching jamming attacks to attract the enemy's attention; Group 3, the "attack" group, utilizes electromagnetic chaos to break through the enemy's defenses at high speed along the shortest path.
[0091] LLM recommends Option 2, and provides a detailed plan: "T+0: The cluster departs from point A. T+10: Groups 1 and 2 climb to an altitude of 500 meters, activate jamming, and feign a main attack. T+15: Group 3 maintains an altitude of 50 meters and stealthily passes through the passage on the west side of the XX reef group. T+30: Groups 1 and 2 jettison the jamming pods, rapidly descend to ultra-low altitude, and retreat. T+40: Group 3 arrives at point B."
[0092] 4. Generation of tactical rules and strategy constraints: LLM reward function constraints: Rewards: Utilize islands and reefs for cover, maintain an angle >30° with the threatening radar, or successfully jam the radar. Penalties: Prolonged exposure within the radar's maximum detection range, flight altitude exceeding 100 meters (Scenario 3), and being locked onto by radar. Phase Two: Hierarchical Collaborative Decision Making and Simulation Verification (Pre-war Simulation and Deduction).
[0093] 1. Issuance of strategic instructions: LLM translates Option 2 into high-level commands and issues them to the three drone teams in the simulation environment: Group 1 & 2: Instructions: Proceed from path α, climb to 500m at T=600s, activate 1.3GHz jamming, continue for 900s, then retreat along path β.
[0094] Group 3: Instructions: Proceed along path γ (low-altitude concealed path), maintain an altitude of <50m throughout, remain electromagnetically silent, and proceed at full speed towards point B.
[0095] 2. Execution and verification of tactical actions: The lightweight MARL model takes over the real-time control of each drone.
[0096] As Group 1 flew toward path α, the small model detected in real time a Red Team patrol helicopter that was not covered by the initial LLM intelligence.
[0097] The small model immediately makes a tactical-level decision: without deviating from the general direction of the LLM strategic instructions, it orders Group 1 to make a short right turn maneuver, using nearby clouds to avoid visual reconnaissance by helicopters, and then return to the original path.
[0098] The entire scheme was run 10 times in the simulation with a success rate of 80%. The evaluation found that due to the Red Team's fast reaction speed, the retreat path β of Group1&2 had a high probability of being intercepted.
[0099] Phase 3: Dynamic adversarial evolution and meta-learning (in-battle strategy adjustment).
[0100] 1. Performance Assessment and Threat Identification: The assessment module determined that the initial withdrawal path was high-risk and needed optimization. Meanwhile, the Red Team's electronic warfare system, "Fog," adopted a new frequency-hopping mode, reducing the Blue Team's jamming effectiveness (a new type of threat).
[0101] 2. Meta-reasoning and candidate strategy generation: The LLM receives failed data and new information about "the enemy using frequency hopping" and performs rapid inference: "The enemy's electronic warfare capabilities have been enhanced. Recommendations: 1) Change the jamming strategy to use reactive jamming instead of continuous wave jamming. 2) The decoy group's withdrawal path should be more varied, employing a 'random snake maneuver' to evade predictive fire. 3) The attack group can depart 5 minutes earlier, taking advantage of the earlier window of opportunity when the decoy group draws fire."
[0102] 3. Rapid verification and fine-tuning: The simulation environment was used to rapidly test three new policy variants generated by LLM in parallel. The results showed that the combination of "responsive disturbance + serpentine maneuver" was the most effective, increasing the success rate to 90%.
[0103] The small model was rapidly fine-tuned (through online learning) under the guidance of the reward function of the new strategy, and better mastered the timing and amplitude of the "snake maneuver".
[0104] 4. Strategy Update: The finely tuned new strategy model is then transmitted via data link and over-the-air (OTA) updates to the drone swarm actually performing the mission.
[0105] Phase Four: Edge Deployment and Online Inference (Practical Edge Execution).
[0106] Throughout the mission, the lightweight mini-models deployed on each UAV (which have been knowledge distilled and quantized) run independently on the onboard computer.
[0107] Real-world scenario: Group 3 is flying at extremely low altitude. Onboard sensors suddenly detect a temporarily activated man-portable air defense system (MANPADS) (a completely unknown and unpredictable tactical threat).
[0108] The cloud-based LLM cannot respond in real time.
[0109] Edge decision-making: Lightweight, small onboard models perform local real-time inference immediately (milliseconds): Threat detected (infrared signal source).
[0110] Based on its trained tactical knowledge (derived from the LLM reward function: being locked on -> significant penalty; releasing chaff -> reward; maneuvering evasion -> reward), it immediately generates the optimal tactical action: "release infrared decoy chaff + perform a dive turn maneuver with maximum overload".
[0111] This invention addresses a series of core problems in existing UAV swarm decision-making methods based on multi-agent reinforcement learning (MARL), such as low training efficiency, lack of interpretability of decisions, strategic and tactical coupling conflicts, insufficient adaptability in adversarial environments, and difficulty in deployment at resource-constrained edge environments. It proposes an innovative intelligent decision-making framework based on the collaboration of large and small models.
[0112] Compared to the traditional MARL paradigm, which relies entirely on starting from scratch and engaging in massive trial-and-error "self-play," the core contribution of this invention lies in achieving a paradigm shift from purely "data-driven" to a "fusion of knowledge-driven and data-driven approaches." By introducing a large-scale language model (LLM) with prior knowledge in the military domain as a strategic-level cognitive core, the system can generate initial strategic plans and behavioral constraints that conform to military principles, thereby significantly improving the starting point quality and convergence speed of reinforcement learning training and effectively overcoming the inherent defects of the "zero-foundation learning" model.
[0113] In terms of architectural design, this invention cleverly decouples the multi-scale decision-making challenges of strategic planning and tactical control through a hierarchical decision-making mechanism. The LLM (Layered Model) handles macro-level, hourly task planning, while the lightweight model handles real-time, second-level maneuver control. Their collaboration ensures a high degree of unity between global strategic optimization and local survivability. Furthermore, thanks to the powerful generalization and reasoning capabilities of the LLM, the system constructs a dynamic adversarial evolutionary closed loop, capable of rapidly generating and validating new strategies against unknown threats. This gives the system the crucial ability to continuously evolve online in highly adversarial environments, fundamentally improving its adversarial robustness.
[0114] Addressing the bottlenecks in engineering applications, this invention utilizes techniques such as knowledge distillation and model quantization to efficiently transfer the strategic cognitive capabilities inherent in large models to lightweight models. This successfully solves the challenge of deploying high-performance intelligent algorithms with low power consumption and low latency at tactical edge nodes, providing a feasible technical path for realizing the practical application of artificial intelligence decision-making capabilities on end platforms.
[0115] In summary, this invention effectively bridges the gaps between data-driven and knowledge-driven approaches, strategy and tactics, offline learning and online adaptation, and cloud intelligence and edge computing through a systematic solution, providing a new theoretical framework and technical approach for building an efficient, reliable, interpretable, and practically deployable intelligent decision-making system for UAV swarms.
[0116] The steps of the various methods described above are only for clarity. In practice, they can be combined into one step or some steps can be split into multiple steps. As long as they include the same logical relationship, they are all within the protection scope of this invention. Adding insignificant modifications or introducing insignificant designs to the algorithm or process, without changing the core design of the algorithm and process, are also within the protection scope of this invention.
[0117] Another embodiment of the present invention relates to a decision-making system for maritime penetration of unmanned aerial vehicle (UAV) swarms. The implementation details of this UAV swarm maritime penetration decision-making system are described below. The following implementation details are provided for ease of understanding and are not essential for implementing this solution. The UAV swarm maritime penetration decision-making system of this embodiment includes: The data acquisition module is used to acquire raw battlefield environment information and the mission objectives of the drone swarm; The strategy generation module is used to retrieve and reason from a pre-set military knowledge base based on battlefield environment information and mission objectives using a large language model, in order to generate long-term strategic plans and corresponding strategic reward constraints. The tactical generation module is used to simulate the maritime penetration of UAV swarms based on long-term strategic plans. It also uses a multi-agent reinforcement learning model to determine the instantaneous tactical strategies of the UAV swarms based on strategic reward constraints and battlefield environment information during the penetration process. The strategic assessment module is used to simulate the process of UAV swarms penetrating the sea under the new tactical strategies of the enemy, through long-term strategic plans and instantaneous tactical strategies, in order to evaluate the performance of long-term strategic plans and instantaneous tactical strategies in combat and identify the threats faced. The strategic adjustment module is used to adjust the long-term strategic plan based on the performance of the long-term strategic plan and the instantaneous tactical strategy in combat and the threats faced, through a large language model. It also simulates the maritime penetration process of the UAV swarm in different battlefield environments based on the adjusted long-term strategic plan to obtain multiple instantaneous tactical strategies. The intelligent decision-making module is used to determine the maritime penetration strategy of the drone swarm based on the adjusted long-term strategic plan and the best-performing instantaneous tactical strategy among multiple instantaneous tactical strategies in combat.
[0118] It is not difficult to see that this embodiment is a system embodiment corresponding to the above method embodiments, and this embodiment can be implemented in conjunction with the above method embodiments. The relevant technical details and technical effects mentioned in the above embodiments are still valid in this embodiment, and will not be repeated here to reduce repetition. Accordingly, the relevant technical details mentioned in this embodiment can also be applied to the above embodiments.
[0119] It is worth mentioning that all modules involved in this embodiment are logical modules. In practical applications, a logical unit can be a physical unit, a part of a physical unit, or a combination of multiple physical units. Furthermore, to highlight the innovative aspects of this invention, this embodiment does not introduce units that are not closely related to solving the technical problem proposed by this invention; however, this does not mean that other units are absent from this embodiment.
[0120] Another embodiment of the present invention relates to a computer device, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the UAV swarm maritime penetration decision-making method of the above embodiments.
[0121] The memory and processor are connected via a bus, which can include any number of interconnecting buses and bridges, connecting various circuits of one or more processors and memories. The bus can also connect various other circuits, such as peripheral devices, voltage regulators, and power management circuits, which are well known in the art and will not be described further herein. The bus interface provides an interface between the bus and the transceiver. The transceiver can be a single element or multiple elements, such as multiple receivers and transmitters, providing a unit for communicating with various other devices over a transmission medium. Data processed by the processor is transmitted over the wireless medium via an antenna, which further receives data and transmits it to the processor.
[0122] The processor manages the bus and general processing, and also provides various functions, including timing, peripheral interfaces, voltage regulation, power management, and other control functions. Memory is used to store data used by the processor during operation.
[0123] Another embodiment of the present invention relates to a computer-readable storage medium storing a computer program. When executed by a processor, the computer program implements the method embodiments described above.
[0124] That is, those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware. This program is stored in a storage medium and includes several instructions to cause a device (which may be a microcontroller, chip, etc.) or processor to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.
[0125] Those skilled in the art will understand that the above embodiments are specific embodiments for implementing the present invention, and in practical applications, various changes can be made to them in form and detail without departing from the spirit and scope of the present invention.
Claims
1. A decision-making method for unmanned aerial vehicle (UAV) swarm penetration at sea, characterized in that, The method includes: To acquire raw battlefield environment information and the mission objectives of the drone swarm; Based on battlefield environment information and mission objectives, the large language model retrieves and reasons from a pre-set military knowledge base to generate long-term strategic plans and corresponding strategic reward constraints. Based on the long-term strategic plan, a simulation of the UAV swarm's maritime penetration was conducted. Using a multi-agent reinforcement learning model, the instantaneous tactical strategy of the UAV swarm was determined based on strategic reward constraints and battlefield environment information during the UAV swarm's penetration during the simulation. The simulation aims to assess the performance of long-term strategic plans and instantaneous tactical strategies in combat against new enemy tactical tactics and to identify the threats faced. Based on the performance of long-term strategic plans and instantaneous tactical strategies in combat and the threats faced, the long-term strategic plan is adjusted using a large language model. Then, the process of unmanned aerial vehicle swarms penetrating at sea in different battlefield environments is simulated based on the adjusted long-term strategic plan to obtain multiple instantaneous tactical strategies. Based on the adjusted long-term strategic plan and the best-performing instantaneous tactical strategy among multiple instantaneous tactical strategies in combat, a maritime penetration strategy for drone swarms is determined.
2. The unmanned aerial vehicle (UAV) swarm maritime penetration decision-making method according to claim 1, characterized in that, The process involves using a large language model to retrieve and reason from a pre-set military knowledge base based on battlefield environment information and mission objectives to generate long-term strategic plans and corresponding strategic reward constraints, including: Based on battlefield environment information and mission objectives, the large language model retrieves similar battle examples and corresponding tactical clauses from the military knowledge base. Based on the retrieved battle examples and corresponding tactical clauses, it infers long-term strategic plans and corresponding strategic reward constraints.
3. The unmanned aerial vehicle (UAV) swarm maritime penetration decision-making method according to claim 2, characterized in that, The long-term strategic plan includes multiple penetration routes, grouping suggestions and action sequences for drone swarms, and a plan for the use of the electromagnetic spectrum. The strategic reward constraint refers to additional rewards that conform to military principles, based on the achievement of mission objectives by the drone swarm.
4. The unmanned aerial vehicle (UAV) swarm maritime penetration decision-making method according to claim 1, characterized in that, The method for determining the maritime penetration strategy for drone swarms, based on the adjusted long-term strategic plan and the best-performing instantaneous tactical strategy among multiple instantaneous tactical strategies in combat, includes: Based on the adjusted long-term strategic plan and the best-performing instantaneous tactical strategy among multiple instantaneous tactical strategies in combat, the multi-agent reinforcement learning model is adjusted. Based on the knowledge distillation method, a large language model and an adjusted multi-agent reinforcement learning model are used as teacher models to train student models. By using trained student models, we can determine maritime penetration strategies for different drone swarms in various battlefield environments.
5. The unmanned aerial vehicle (UAV) swarm maritime penetration decision-making method according to claim 4, characterized in that, The trained student model is deployed on each drone in the drone swarm.
6. A decision-making system for unmanned aerial vehicle (UAV) swarm maritime penetration, characterized in that, The system includes: The data acquisition module is used to acquire raw battlefield environment information and the mission objectives of the drone swarm; The strategy generation module is used to retrieve and reason from a pre-set military knowledge base based on battlefield environment information and mission objectives using a large language model, in order to generate long-term strategic plans and corresponding strategic reward constraints. The tactical generation module is used to simulate the maritime penetration of UAV swarms based on long-term strategic plans. It also uses a multi-agent reinforcement learning model to determine the instantaneous tactical strategies of the UAV swarms based on strategic reward constraints and battlefield environment information during the penetration process. The strategic assessment module is used to simulate the process of UAV swarms penetrating the sea under the new tactical strategies of the enemy, through long-term strategic plans and instantaneous tactical strategies, in order to evaluate the performance of long-term strategic plans and instantaneous tactical strategies in combat and identify the threats faced. The strategic adjustment module is used to adjust the long-term strategic plan based on the performance of the long-term strategic plan and the instantaneous tactical strategy in combat and the threats faced, through a large language model. It also simulates the maritime penetration process of the UAV swarm in different battlefield environments based on the adjusted long-term strategic plan to obtain multiple instantaneous tactical strategies. The intelligent decision-making module is used to determine the maritime penetration strategy of the drone swarm based on the adjusted long-term strategic plan and the best-performing instantaneous tactical strategy among multiple instantaneous tactical strategies in combat.
7. A computer device, characterized in that, include: At least one processor; And a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to execute the UAV swarm maritime penetration decision-making method as described in any one of claims 1 to 5.
8. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the decision-making method for unmanned aerial vehicle swarm maritime penetration as described in any one of claims 1 to 5.