A multi-agent group confrontation decision method and system based on weighted mean field

Abstract

The application discloses a kind of multi-agent group confrontation decision method and system based on weighted average field, the method is by constructing three-dimensional multi-agent group confrontation environment, define agent state and action space, design WMF-MAPOCA algorithm, fusion multi-hot encoding and multi-head attention mechanism realize neighbor information adaptive aggregation, while designing the reward function including target keeping, target approaching, attack success and survival four dimensions, combined with artificial potential field method realizes motion control and obstacle avoidance WMF-MAPOCA algorithm is updated by experience replay buffer and the training of three core networks, effectively alleviate dimension disaster, reduce strategy non-stationarity in large-scale asymmetric group confrontation scene.

Description

Technical Field

[0001] This invention relates to the field of multi-agent control and reinforcement learning technology, specifically to a multi-agent swarm adversarial decision-making method and system based on a weighted average field. Background Technology

[0002] In recent years, unmanned aerial vehicle (UAV) technology has developed rapidly and has been widely used in civilian and military fields such as reconnaissance missions, environmental monitoring, and navigation and positioning. Among them, swarm adversarial scenarios involving multiple UAVs in complex environments have become a research hotspot in the field of intelligent systems due to their dynamic nature, uncertainty, and high requirements for cooperation. Existing multi-agent swarm adversarial solutions are mainly divided into two categories: traditional methods and learning-based methods. Traditional UAV adversarial methods mainly rely on preset rules or optimization algorithms, with core methods including consensus algorithms, model predictive control algorithms, and particle swarm optimization algorithms. The core limitations of traditional methods are: high computational complexity, making it difficult to adapt to large-scale agent cooperation; poor dynamic adaptability, resulting in slow response speed to environmental changes and strategy adjustments; and limited generalization ability, leading to poor portability in adversarial tasks of different scales and scenarios. These shortcomings make it difficult to apply them effectively in large-scale multi-agent swarm adversarial scenarios. To overcome the shortcomings of traditional methods, multi-agent reinforcement learning (MARL) methods, with their autonomous learning and dynamic adaptability, have shown great potential in complex swarm adversarial tasks.

[0003] Despite significant progress made by existing MARL methods, they still face three core challenges in large-scale, dynamic asymmetric swarm adversarial scenarios: the curse of dimensionality, policy non-stationarity, and limited generalization ability.

[0004] An existing multi-aircraft cooperative air combat decision-making method based on MATD3 (Multi-Agent Dual-Delay Deep Deterministic Policy Gradient) and a time-series awareness mechanism (Research on Multi-Aircraft Intelligent Air Combat Strategy Based on MATD3 and Time-Series Awareness Mechanism, Cai Yiming, Han Zhiqiang, Yin Mingxin, Zhen Ziyang, Luo Wei) addresses the short-sightedness and strategy convergence problems in manned / unmanned cooperative combat by designing a heterogeneous decision preference reward function to distinguish between reconnaissance and strike roles, and employing a game theory model to initialize the experience replay pool to accelerate model convergence. However, although this research has achieved significant results in small-scale 3v3 simulations, its technical approach reveals systemic limitations when facing large-scale multi-agent expansion. First, as the cluster size increases, the input dimension grows linearly, leading to an unbearable dimensionality explosion problem. Second, the research assumes zero-delay communication, unlimited bandwidth, and static topology; these idealized conditions are difficult to hold in large-scale dynamic networks and may even lead to tactical coordination failure. Furthermore, the sophisticated heterogeneous reward function relies on a large number of manually tuned weight factors and potential field parameters. The complexity of this reward engineering increases exponentially with the number of role combinations, making it difficult to generalize to large-scale systems with diverse and heterogeneous components such as UAVs, manned aircraft, and early warning aircraft. These limitations indicate that this technology still needs to overcome key bottlenecks such as algorithm efficiency, communication constraints, and scalability in large-scale multi-agent air combat applications. Summary of the Invention

[0005] This invention addresses the problems of existing multi-agent swarm adversarial algorithms by providing a multi-agent swarm adversarial decision-making method and system based on weighted average field (WMF). By deeply integrating WMF theory with the MAPOCA algorithm and introducing multi-hot encoding and multi-head attention mechanisms, it achieves adaptive weighted aggregation of neighbor information. Simultaneously, it fully utilizes the historical decision information of agents, significantly improving decision-making efficiency, collaborative performance, and adversarial win rate in large-scale, asymmetric swarm adversarial scenarios. This method is suitable for large-scale, asymmetric swarm adversarial tasks involving agents such as UAVs in complex 3D environments and can be widely applied to intelligent adversarial scenarios in fields such as national defense, public safety, and emergency response.

[0006] The present invention is achieved by at least one of the following technical solutions.

[0007] A multi-agent swarm adversarial decision-making method based on a weighted average field includes the following steps:

[0008] 1) Collect agent state and action information, and use the WMF-MAPOCA algorithm to fuse and adaptively aggregate observation information and historical action information, outputting target selection probability distribution and action decision strategy; the target selection probability distribution is used to determine the priority attack weight of the agent for each enemy target, and the action decision strategy is used to guide the agent's target locking logic. Both serve as inputs to the artificial potential field method, providing target priority and behavior guidance for the calculation of the desired motion direction; 2) The desired motion direction is calculated using the artificial potential field method, and the pitch and yaw angle control inputs are generated by combining PID control to achieve target tracking and collision avoidance.

[0009] Furthermore, the WMF-MAPOCA algorithm includes an experience replay buffer, a feature encoding module, three core networks, and a weighted average field module; among them, the experience replay buffer is used to store experience tuples generated by the interaction between the agent and the environment; the feature encoding module is used to fuse current observation information and historical decision information to obtain joint information.

[0010] Furthermore, the information aggregation process of the weighted average field module is as follows: (1) Define the neighbor set of the central agent, that is, other agents within the detection range; (2) The features of the central agent are mapped to query vectors and the features of the neighbors are mapped to key vectors and value vectors through a multi-head attention mechanism; (3) Calculate the attention score based on the query vector and key vector, and normalize it to obtain the attention weight; (4) Average pooling is performed based on the attention weights of multiple attention heads and the weighted aggregated neighbor value vector to obtain the final weighted aggregated features of the agent.

[0011] Furthermore, the three core networks are the value network, the baseline network, and the behavior network. The input of the value network is the joint information of the central agent and the aggregated virtual neighbor features. After passing through the feedforward-self-attention-average pooling-residual link, the output is the state value estimate. The training phase adopts the temporal difference objective and updates the parameters through the pruned squared loss. The baseline network shares the WMF module with the value network, but additionally introduces neighbor action information to output action value estimation; the loss function is similar to that of the value network, also using clipped squared error. The behavioral network takes joint information as input and outputs discrete target selection probabilities; it constructs an advantage function using the action value given by the baseline network and updates the parameters through a pruning policy gradient loss.

[0012] Furthermore, the loss function of the value network is:

[0013] in Represents intelligent agents i The value of the value network loss function. Represents intelligent agents i The set of trainable parameters for a value network. This refers to the batch sample size. ( ) represents an intelligent agent i The state value estimate output by the value network. ( ) represents a self-attention module in a value network. Indicates the first c In each sample, the agent i Virtual neighbor characteristics, Indicates the first c In each sample, the agent i The target value.

[0014] Furthermore, in step (4), the reward function of the WMF-MAPOCA algorithm includes four dimensions: target retention reward, target approach reward, attack success reward, and survival reward.

[0015] Furthermore, the artificial potential field method is as follows: The desired direction of motion is:

[0016] in , This is the gain coefficient. A collection of intelligent agents within a safe distance. Let o( represent the desired motion direction vector of agent i at time t). This is used to convert an input vector into a unit vector. This represents the position vector of agent i at the target point at time t. Let represent the position vector of agent i at time t. Let represent the position vector of agent j within the safe distance at time t, where t represents the time step.

[0017] The control input is:

[0018]

[0019] Among them , This is the gain coefficient. Let be the pitch angle control input for agent i at time t. The yaw angle control input for agent i at time t is used to adjust the agent's yaw attitude. Let δi(t) be the current pitch angle of agent i at time t, and let δi(t) be the current yaw angle of agent i at time t. The desired motion direction vector The component in the y-axis direction, The desired motion direction vector The component in the x-axis direction.

[0020] To implement the aforementioned multi-agent swarm adversarial decision-making system based on a weighted average field, the following components are included: Environment Modeling Module: Used to construct 3D adversarial environments, define agent and obstacle parameters, and simulate physical rules; Perception module: Used to collect environmental information, and generate feature vectors by fusing observations and historical actions through multi-hot encoding; The WMF-MAPOCA algorithm module includes an experience replay buffer and three core networks for feature aggregation and network updates. Decision module: used to analyze probability distributions to determine attack targets and generate motion control commands; Motion control module: used to drive the intelligent agent to perform actions, track the desired trajectory, and avoid obstacles; Reward Calculation Module: Used to calculate multi-dimensional reward values ​​in real time, providing supervision signals for the training of the WMF-MAPOCA algorithm.

[0021] A computer device according to the present invention includes a memory and a processor, the memory being electrically connected to the processor, the memory storing a computer program, which, when executed by the processor, causes the processor to implement the method described herein.

[0022] The present invention provides a computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, the processor implements the method described herein.

[0023] Compared with existing technologies, the beneficial effects of the present invention are as follows: 1. Effectively mitigates the curse of dimensionality: By using weighted average field theory, the interaction between the central agent and multiple neighbors is approximated as the interaction with a virtual agent. Combined with multi-head attention mechanism to abstract neighbor information, the computational burden is significantly reduced, and the exponential growth of state and action space is avoided. 2. Reduce policy non-stationarity: Dynamically allocate neighbor attention weights to accurately model the relationships between agents, making policy updates of each agent more targeted, reducing environmental fluctuations caused by mutual policy influence, and improving training stability and convergence speed. Experiments show that the algorithm's convergence speed is more than 30% faster than MAPOCA. 3. Improved generalization ability: The algorithm maintains a high win rate in three different adversarial scenarios (3vs5, 6vs10, 10vs16), and the win rate continues to increase with the scale of the adversarial battle (the win rate reaches 82% in the 10vs16 scenario), demonstrating excellent scenario adaptability and solving the problem of performance degradation of existing algorithms in large-scale scenarios. 4. Enhance information aggregation efficiency: Integrate current observation information with historical action information, and achieve effective extraction and adaptive aggregation of key information through multi-hot encoding and multi-head attention mechanisms, avoiding interference from redundant information and improving decision-making accuracy. Attached Figure Description

[0024] Figure 1 The feature aggregation flowchart of the weighted average field module in this embodiment is shown. Figure 2 This is a diagram of the WMF-MAPOCA algorithm framework for an example. Figure 3 This is a schematic diagram of the intelligent agent motion control and obstacle avoidance logic in an embodiment. Detailed Implementation

[0025] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments 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.

[0026] This embodiment presents a multi-agent swarm adversarial decision-making method based on weighted average field, comprising five core steps: environment modeling and parameter setting, construction of the multi-agent weighted average field post-death credit allocation algorithm (WMF-MAPOCA), reward function design, motion control and obstacle avoidance, and algorithm training and decision execution. The specific steps are as follows: S1. Environment Modeling and Parameter Setting: Construct a 3D virtual environment for multi-agent swarm adversarial competition, and define the attribute parameters of agents and obstacles, including environmental spatial parameters, agent motion parameters, attack parameters, and obstacle parameters.

[0027] In one embodiment, the attributes and parameters of the red team intelligent agent team, the blue team intelligent agent team, and obstacles are clearly defined to provide a basic environment for subsequent algorithm training and decision execution, including: Environmental spatial parameters: Define the length of the three-dimensional environment. ,width ,high ; The agent's motion parameters include: Number of Red Team Agents Number of Blue Square Intelligent Agents Three adversarial scales are set: 3 vs 5, 6 vs 10, and 10 vs 16; the agents maintain a constant speed. ; Health Points (HP): The maximum health of the red team's agents. The maximum health of the blue team's intelligent agent This is used to balance the differences in combat capabilities caused by the asymmetry in quantity; Control input constraints: Pitch angle control input Yaw angle control input ; (where u i is the pitch angle control input for agent i, and uδi is the yaw angle control input for agent i. i is the current pitch angle of agent i, δi is the current yaw angle of agent i, E is the maximum amplitude threshold of pitch angle control input, and Δ is the maximum amplitude threshold of yaw angle control input.

[0028] Attack parameters: Attack range Maximum aiming deviation Each agent can only attack one target per time step; Obstacle parameters: uniformly distributed in the environment There are several cylindrical obstacles, each with a radius of... m, height Obstacles can obstruct the agent's line of sight and hinder movement; Observation range parameter: the observation radius of the agent. It can only perceive information about teammates, enemies, and obstacles within its observation range.

[0029] S2. Define the state and actions of the agent.

[0030] Define state observation: Red side's first Observations of an agent ,in: It indicates its own status information (including its own position, pitch angle, and yaw angle). It indicates the enemy's status information (including enemy ID, health, location, and distance). This indicates the status information of teammates (including their location and distance).

[0031] Define the action space: obtained through sampling from a discrete probability distribution, where the action vector represents the target selection probability distribution. The formula for sampling the discrete probability distribution is:

[0032] in. It is the weighted fusion result of the i-th agent at time t. It is the weight coefficient of the i-th agent for the k-th sampling action. It is the candidate action vector obtained by sampling the k-th discrete probability distribution. It is a discrete probability distribution sampling. This is the kth observation.

[0033] In this embodiment, the action of each red agent is represented as a target selection decision, generating an action probability distribution. ,in Indicates the first The red team agent chooses the first... The probability of a blue agent being the target of an attack.

[0034] S3. Construct the WMF-MAPOCA algorithm.

[0035] The WMF-MAPOCA algorithm framework includes an experience replay buffer and three core networks (value network, baseline network, and behavior network). It introduces a weighted average field module to achieve efficient fusion and adaptive aggregation of observation information and historical action information. The specific structure is as follows: (1) Experience replay buffer: Construct a buffer with a capacity of Experience replay buffer Used to store experience tuples generated by the interaction between the agent and the environment. ,in It is the observation value of the i-th agent at the current moment. It is the state feature (or hidden state) of the i-th agent. It is the action performed by the i-th agent at the current moment. It is the reward value obtained by the i-th agent at the current moment. For the next observation, random sampling batches are used to mitigate sample correlation and improve training stability.

[0036] Feature encoding module: To integrate current observation information with historical decision information, each agent collects a set of enemy targets within its observation range. Combined with the actions of the previous moment Feature vectors are generated through multi-hot encoding. ; to observe the current situation With feature vectors By splicing together, we can obtain the combined information. This provides rich feature inputs for subsequent network processing. It is a multi-hot coding operation that converts a discrete set of inputs into a binary vector. It is the set of enemy targets within the observation range of the i-th agent at the current moment. It is the action executed by the i-th agent in the previous moment.

[0037] (2) Weighted Average Field (WMF) Module: The interaction between the central agent (the agent currently selected as the starting point for information aggregation) and its multiple neighbors is approximated as the interaction with a virtual agent (an abstract equivalent agent generated by the weighted average field module), effectively reducing the input dimensionality and alleviating the curse of dimensionality. The information aggregation process of the weighted average field module is as follows: Neighbor set definition: for the central agent Define the observation range The set of neighboring intelligent agents (Including teammate AI agents); Multi-head attention mechanism: settings Each attention head independently calculates the neighbor weights, thereby improving the feature representation capability; Query-key-value mapping: linking the joint information of the central agent Mapped to query vector Each neighboring intelligent agent The joint information is mapped to the key vector. Sum value vector ,in Let x be a trainable parameter matrix. i =(o i ,z i ) is the joint information of the central agent i, and similarly, It is the joint information of neighboring intelligent agent j; Attention score calculation: ,in This is the query vector dimension, used to mitigate the impact of dimensional differences; Attention weight normalization: Attention weights are obtained by normalization using the Softmax function, achieving adaptive weighting of neighbor information. ; Feature aggregation: Weighted aggregation of neighbor value vectors and concatenation across the head result in aggregated features. :

[0038] in This is the activation function, used to introduce nonlinear feature transformations.

[0039] (3) Network structure and parameter updates of the three core networks (value network, baseline network, and behavior network): The input to the value network is the joint information of the central agent and the aggregated virtual neighbor features. After processing, the output is the state value estimate. The training phase adopts a temporal difference objective and updates the parameters through a pruned squared loss.

[0040] The baseline network shares the same weighted average field module as the value network, but additionally introduces neighbor action information to output action value estimation; the loss function is similar to that of the value network, also using clipped squared error.

[0041] The behavioral network takes joint information as input and outputs discrete target selection probabilities; it constructs an advantage function using the action value given by the baseline network and updates the parameters through a pruning policy gradient loss.

[0042] S4. Reward Function Design: Reward Function The reward is a multi-component, intensive type, and its specific composition is as follows:

[0043] The goal is to maintain the reward A positive reward is given if the objective remains unchanged and the individual survives; otherwise, the reward is zero. Target proximity reward A positive reward is given when the target is within attack range; otherwise, it is 0. Successful attack reward A positive reward is given when the target is defeated; otherwise, the reward is 0. Survival Rewards : A negative reward is given when the user is defeated, otherwise it is 0.

[0044] (5) Motion control and obstacle avoidance: The desired motion direction is calculated by using the artificial potential field method, and the pitch angle and yaw angle control inputs are generated by combining PID control to achieve target tracking and collision avoidance.

[0045] After selecting a target, the red team's drone needs to adjust its flight attitude in real time to quickly approach the locked target while avoiding collisions with teammates or obstacles. This invention uses an improved artificial potential field method to achieve motion control and obstacle avoidance functions, as calculated below: The principles for constructing a potential field include: Attraction field: Using the instantaneous position of the locked target as the attraction center, a directional attraction is applied to the drone, driving it into the attack zone; Repulsive field: Centered on the drone itself, within a preset safe distance range, it generates a repulsive force on nearby teammates and obstacles, with the direction of the force pointing from the threat source towards the drone. In this embodiment, the safe distance is a fixed value, and its specific value is consistent with the radius of the obstacle, ensuring that the drone can maintain the minimum distance in dense environments and prevent rigid collisions.

[0046] Force synthesis: The attractive and repulsive forces are vector-superimposed to form a comprehensive desired motion direction, which is used for subsequent attitude command calculation.

[0047] The desired motion direction is calculated as follows: the UAV calculates a unitized desired vector based on the combined force direction; the attractive and repulsive force weights are experimentally calibrated and then fixed in the control logic to ensure rapid engagement with the enemy while maintaining formation looseness in large-scale group combat scenarios, avoiding the "clustering and blocking" phenomenon. The desired motion direction calculation model is as follows:

[0048] in , This is the gain coefficient. For the set of other intelligent agents within a safe distance, Let o( represent the desired motion direction vector of agent i at time t). This is used to convert an input vector into a unit vector. This represents the position vector of agent i at the target point at time t. Let represent the position vector of agent i at time t. Let represent the position vector of agent j within the safe distance at time t, where t represents the time step.

[0049] The control input is:

[0050]

[0051] in , This is the gain coefficient. Let be the pitch angle control input for agent i at time t. The yaw angle control input for agent i at time t is used to adjust the agent's yaw attitude. Let δi(t) be the current pitch angle of agent i at time t, and let δi(t) be the current yaw angle of agent i at time t. The desired motion direction vector The component in the y-axis direction, The desired motion direction vector The component in the x-axis direction.

[0052] Dynamic update mechanism: The above potential field calculation is repeated within each simulation step to ensure that the UAV can continuously obtain the latest desired direction and adjust its flight trajectory accordingly when the target maneuvers, teammates change positions, or obstacles block the view. The single-step time of the entire calculation process is less than the simulation step, which meets the real-time requirements.

[0053] To implement the aforementioned multi-agent swarm adversarial decision-making system based on a weighted average field, the following components are included: Environment Modeling Module: Used to construct a 3D adversarial environment, define the parameters of intelligent agents and obstacles, and simulate the physical rules of real adversarial scenarios (such as kinematic laws, collision detection, attack mechanisms, etc.). Perception module: Used to collect environmental information, and generate feature vectors by fusing observations and historical actions through multi-hot encoding; The WMF-MAPOCA algorithm module includes an experience replay buffer and three core networks for feature aggregation and network updates. Decision Module: Used to determine the final attack target based on the probability distribution output by the WMF-MAPOCA algorithm module. Combined with motion control and obstacle avoidance logic, it generates motion control commands. The core logic of this module includes: target selection logic (probability sampling), obstacle avoidance decision logic (artificial potential field method), and attack triggering logic (distance and angle judgment). Outputs: Attack target number, pitch angle control input, and yaw angle control input, providing input to the motion control module.

[0054] Motion control module: It receives control commands from the decision module, drives the agent to perform motion and attack actions, ensures that the agent moves along the desired trajectory and avoids collisions. Specifically, it uses a PID controller to adjust the agent's speed and attitude, so that the actual motion trajectory tracks the desired trajectory, and feeds back the agent's actual state (position, speed, attitude) to the perception module to form a closed-loop control.

[0055] Reward Calculation Module: Monitors the interactive behavior of agents in real time (target selection, target approach, attack result, survival status), calculates reward values ​​according to a preset reward function, and outputs the real-time reward value for each agent, providing supervision signals for the network update of the WMF-MAPOCA algorithm module.

[0056] The environment modeling module, built on the Unity simulation engine, is used to construct a visualized 3D environment. It supports real-time rendering of agent states and simulation of interactive logic, and outputs standardized environmental state data, including agent position, speed, health, and obstacle positions, providing input to the perception module. The perception module is mounted on each red agent and collects multi-source information through virtual sensors (simulating cameras, radar, etc.), including its own state, enemy state, teammate state, and obstacle information. It fuses current observation and historical action information through a multi-hot encoding module to generate a high-dimensional feature vector, and outputs a standardized observation feature vector to provide input to the WMF-MAPOCA algorithm module. The WMF-MAPOCA algorithm module receives the feature vector output by the perception module, and achieves adaptive aggregation of neighbor information through a weighted average field module and a multi-head attention mechanism. Through the training and inference of three core networks, it generates an action probability distribution to provide input to the decision module. It is deployed on edge computing units or cloud servers and supports real-time inference.

[0057] Simulation experiments show that, compared with mainstream algorithms such as MAPOCA (Multi-Agent Post-Death Credit Allocation Algorithm), MAAC (Multi-Agent Actor-Attention-Critic Algorithm), MADDPG (Multi-Agent Deep Deterministic Policy Gradient Algorithm), and MATD3 (Multi-Agent Dual-Delay Deep Deterministic Policy Gradient Algorithm), this invention has significant advantages in win rate and survival rate, and can be widely applied in fields such as UAV swarm warfare.

[0058] Table 1 Average win rates of various algorithms

[0059] WMF-MAPOCA leads across all scenarios. Its win rate (0.75→0.82→0.87) is significantly higher than other algorithms across all adversarial scales. It exhibits strong scale adaptability; as the adversarial scale increases, WMF-MAPOCA's win rate continues to rise, while other algorithms' win rates fluctuate. This demonstrates that the WMF module alleviates the curse of dimensionality in multi-agent scenarios, allowing the algorithm to maintain decision-making efficiency in large-scale adversarial situations. MAPOCA is a basic action-conditional policy, but it lacks WMF's "information aggregation" module, therefore its win rate does not improve (and may even slightly decrease) as the scale increases. MAAC's attention mechanism can only optimize local information interaction, and MADDPG's centralized commentator cannot adapt to large-scale scenarios. This table directly verifies the effectiveness of the WMF-MAPOCA algorithm—by reducing the input dimension and optimizing multi-agent interaction, it achieves the technical effect of "higher win rate and stronger adaptability in large-scale adversarial scenarios," which is the key experimental evidence that this patent is superior to existing technologies.

[0060] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, enabling those skilled in the art to better understand and utilize the invention.

Claims

1. A multi-agent swarm adversarial decision-making method based on a weighted average field, characterized in that, Includes the following steps: 1) Collect agent state and action information, and use the WMF-MAPOCA algorithm to fuse and adaptively aggregate observation information and historical action information, outputting target selection probability distribution and action decision strategy; the target selection probability distribution is used to determine the priority attack weight of the agent for each enemy target, and the action decision strategy is used to guide the agent's target locking logic. Both serve as inputs to the artificial potential field method, providing target priority and behavior guidance for the calculation of the desired motion direction; 2) The desired motion direction is calculated using the artificial potential field method, and the pitch and yaw angle control inputs are generated by combining PID control to achieve target tracking and collision avoidance.

2. The multi-agent swarm adversarial decision-making method based on a weighted average field according to claim 1, characterized in that, The WMF-MAPOCA algorithm includes an experience replay buffer, a feature encoding module, three core networks, and a weighted average field module. The experience replay buffer stores experience tuples generated by the agent's interaction with the environment. The feature encoding module fuses current observation information with historical decision information to obtain joint information.

3. A multi-agent swarm adversarial decision-making method based on a weighted average field according to claim 2, characterized in that, The information aggregation process of the weighted average field module is as follows: (1) Define the neighbor set of the central agent, that is, other agents within the detection range; (2) The features of the central agent are mapped to query vectors and the features of the neighbors are mapped to key vectors and value vectors through a multi-head attention mechanism; (3) Calculate the attention score based on the query vector and key vector, and normalize it to obtain the attention weight; (4) Average pooling is performed based on the attention weights of multiple attention heads and the weighted aggregated neighbor value vector to obtain the final weighted aggregated features of the agent.

4. A multi-agent swarm adversarial decision-making method based on a weighted average field according to claim 2, characterized in that, The three core networks are the value network, the baseline network, and the behavior network. The input of the value network is the joint information of the central agent and the aggregated virtual neighbor features. After passing through the feedforward-self-attention-average pooling-residual link, the output is the state value estimate. The training phase adopts the temporal difference objective and updates the parameters through the pruned squared loss. The baseline network shares the WMF module with the value network, but additionally introduces neighbor action information to output action value estimation; the loss function is similar to that of the value network, also using clipped squared error. The behavioral network takes joint information as input and outputs discrete target selection probabilities; it constructs an advantage function using the action value given by the baseline network and updates the parameters through a pruning policy gradient loss.

5. A multi-agent swarm adversarial decision-making method based on a weighted average field according to claim 4, characterized in that, The loss function of the value network is: in Represents intelligent agents i The value of the value network loss function. Represents intelligent agents i The set of trainable parameters for a value network. This refers to the batch sample size. ( ) represents an intelligent agent i The state value estimate output by the value network. ( ) represents the self-attention module in the value network. Indicates the first c In each sample, the agent i Virtual neighbor characteristics, Indicates the first c In each sample, the agent i The target value.

6. A multi-agent swarm adversarial decision-making method based on a weighted average field according to claim 1, characterized in that, In step (4), the reward function of the WMF-MAPOCA algorithm includes four dimensions: target retention reward, target approach reward, attack success reward, and survival reward.

7. A multi-agent swarm adversarial decision-making method based on a weighted average field according to claim 1, characterized in that, The artificial potential field method is as follows: The desired direction of motion is: in , This is the gain coefficient. A collection of intelligent agents within a safe distance. Let o( represent the desired motion direction vector of agent i at time t). This is used to convert an input vector into a unit vector. This represents the position vector of agent i at the target point at time t. Let represent the position vector of agent i at time t. Let represent the position vector of agent j within the safe distance at time t, where t represents the time step; The control input is: Among them , This is the gain coefficient. Let be the pitch angle control input for agent i at time t. The yaw angle control input for agent i at time t is used to adjust the agent's yaw attitude. Let δi(t) be the current pitch angle of agent i at time t, and let δi(t) be the current yaw angle of agent i at time t. The desired motion direction vector The component in the y-axis direction, The desired motion direction vector The component in the x-axis direction.

8. A multi-agent swarm adversarial decision-making system based on a weighted average field as described in claim 1, characterized in that, include: Environment Modeling Module: Used to construct 3D adversarial environments, define agent and obstacle parameters, and simulate physical rules; Perception module: Used to collect environmental information, and generate feature vectors by fusing observations and historical actions through multi-hot encoding; The WMF-MAPOCA algorithm module includes an experience replay buffer and three core networks for feature aggregation and network updates. Decision module: used to analyze probability distributions to determine attack targets and generate motion control commands; Motion control module: used to drive the intelligent agent to perform actions, track the desired trajectory, and avoid obstacles; Reward Calculation Module: Used to calculate multi-dimensional reward values ​​in real time, providing supervision signals for the training of the WMF-MAPOCA algorithm.

9. A computer device comprising a memory and a processor, the memory being electrically connected to the processor, the memory storing a computer program, characterized in that: When the computer program is executed by the processor, it causes the processor to implement the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the processor implements the method as described in any one of claims 1 to 7.

Images