Task allocation method for deadline of auxiliary task in unmanned aerial vehicle group

By adopting a cutoff priority-based attention calculation method in the drone population, combined with the graph neural network and attention mechanism, the task allocation problem under the characteristics of auxiliary tasks is solved, and the task completion rate and system survival rate are improved. A more efficient task allocation strategy is realized.

CN118966647APending Publication Date: 2024-11-15SOUTHEAST UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411005425.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-25
Publication Date
2024-11-15

AI Technical Summary

Technical Problem

In the highly dynamic game confrontation scenario, the drone group task allocation algorithm fails to effectively consider the cut-off time characteristics of auxiliary tasks, resulting in uncertain task priority and system stability affected, making it difficult to maximize task completion rate and system survival rate.

Method used

Using the attention calculation method based on cutoff priority, through task time coding, cutoff priority calculation and multi-agent decision learning, combined with graph neural network and attention mechanism, model relationships and agent characteristics between tasks, calculate tasks and intelligence. The priority of the body is designed to design task allocation strategies to maximize the effectiveness of auxiliary tasks.

Benefits of technology

The task completion rate and system survival rate are improved, emergency tasks can be better screened out, and the utility value of the drone population is improved, ensuring that auxiliary tasks are completed before the deadline, thereby maximizing task utility.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118966647A_ABST
    Figure CN118966647A_ABST
Patent Text Reader

Abstract

The invention provides a task allocation method for auxiliary task deadline characteristics of an unmanned aerial vehicle group in a high-dynamic game scene. The method aims to solve the influence of overtime disappearance of the auxiliary task on the system stability. The auxiliary task is completed before the deadline, so that the completion of other tasks can be accelerated. In a high-dynamic game scene, the action range of the auxiliary tasks is uncertain, the cascade diffusion influence between the auxiliary tasks is also uncertain, and the complexity of the problem is increased. The method comprises the steps of firstly performing time coding on a task to generate a feature vector; then calculating a task cut-off priority and updating the features of the intelligent agent; and finally, task allocation is carried out by applying multi-agent system decision learning. Compared with the prior art, the method has the advantages that the task completion rate and the survival rate of the system in a complex game environment are effectively improved through calculation of time coding and cut-off priority in combination with an attention mechanism.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of multi-task allocation in a drone group in an actual high-dynamic game scenario, and in particular to a task allocation method for auxiliary task deadlines in a drone group. Background Art

[0002] In actual confrontation scenarios, auxiliary tasks such as reconnaissance, patrolling, and supply can have a beneficial impact on other tasks within a certain range after completion, helping other main and auxiliary tasks to be completed. While helping to complete the main task, auxiliary tasks can also assist in the completion of other auxiliary tasks, and their auxiliary capabilities can be regarded as one of their characteristics. Executing auxiliary tasks can reduce the execution time of all other types of tasks around them. However, since the execution of auxiliary tasks also takes time, the final evaluation indicator of the system is the total time required to complete all main tasks. Therefore, when choosing to execute auxiliary tasks, it is necessary to comprehensively consider auxiliary tasks that can shorten the total execution time of the task. At the same time, auxiliary tasks not only affect the main task, but may also have an impact on other auxiliary tasks, so the cascading effect between auxiliary tasks needs to be considered. The scope of influence of auxiliary tasks is an uncertain attribute that may vary in size. Therefore, it is necessary to combine graph neural network technology to model the relationship between tasks as a graph and use the uncertainty of the graph to represent the uncertainty of auxiliary tasks.

[0003] At the same time, tasks do not necessarily exist in the environment in advance, but may arrive periodically and have a certain deadline. These tasks generally have a fixed period, and the task information of the next period can be obtained in advance in the current period. After each batch of tasks arrives, if they cannot be completed before the deadline, they will fail and disappear.

[0004] Previous task allocation algorithms did not take into account the deadline characteristics of auxiliary tasks. There are two challenges in designing targeted algorithms: 1) Since tasks have deadline characteristics, the system needs to complete as many tasks as possible within the specified time. Therefore, the system needs to calculate the priority of executing tasks, that is, the priority of the tasks. However, since the auxiliary tasks are uncertain, this priority is also uncertain, which increases the dimension of the understanding space; 2) Since the tasks have deadlines, when the tasks disappear due to timeout, it will cause disturbances to the original task allocation strategy. Since the auxiliary tasks will also disappear, this disturbance is also uncertain, which increases the complexity of understanding.

[0005] In response to the above challenges, we hope to propose a method for calculating attention based on deadline priority for the scenario where auxiliary tasks have deadline characteristics. This method first considers the uncertain impact of the deadline of auxiliary tasks on other tasks and updates the feature vector representation of all tasks; then, based on the deadline characteristics of the tasks, the priority order of the tasks for the agent is calculated and used as the basis for task allocation. This algorithm not only takes into account the uncertainty of auxiliary tasks, but also the uncertain impact of the deadline characteristics of tasks, ultimately achieving a higher task completion rate and system survival rate. Summary of the invention

[0006] Technical problem: The purpose of the present invention is to propose a new task allocation method for the deadline of auxiliary tasks in drone groups in highly dynamic game confrontation scenarios. The algorithm first considers the game confrontation scenario. Since the task has a deadline, the agents in the multi-agent system of the drone group need to reasonably allocate tasks to complete the task as much as possible before the task times out, especially for auxiliary tasks. Because the auxiliary capabilities of such tasks are uncertain, the multi-agent system needs to determine whether the task can have a high auxiliary utility before the timeout. If so, it needs to be completed as soon as possible. If not, the system needs to abandon the task in time. In summary, the multi-agent system needs to combine the uncertainty of the auxiliary capabilities of the auxiliary tasks and the deadline characteristics to design a reasonable task allocation strategy to ensure that the auxiliary tasks can maximize their utility before the timeout, so as to achieve the purpose of maximizing the task completion rate and system survival rate.

[0007] Technical solution: For the task allocation of auxiliary task deadlines in drone groups in game confrontation scenarios, the present invention proposes an attention algorithm based on cutoff priority. The main technical solutions of the task allocation algorithm are as follows:

[0008] The multi-agent system of the drone swarm first performs time encoding processing on the tasks and calculates the average risk of the tasks. The deterministic features of the tasks, such as the task deadline, are directly transmitted through the fully connected network without any changes to maintain their constant properties; the average risk, task scope, etc. are used as uncertain features and are sequentially input into the FC, GRU, and FC network layers, and the network output is used as a new feature to finally obtain a new feature vector representation of the task. The task set is modeled as a directed graph to represent the relationship between the main task and the auxiliary task, and the attention constraint is set to limit the attention relationship to a specific task set. Then, the multi-head attention technique is used to calculate the attention scores between all existing tasks and the next batch of task nodes. For the arrival and deadline attributes of the task, a negative weight network is built to convert the batch distance and the task deadline into an attention bias, which is used to correct the original attention score so that the system pays more attention to the tasks that are close to the current and about to expire. Finally, the attention score is used to aggregate the features of the neighboring nodes through weighted summation to obtain the new features of the task node.

[0009] For the task deadline priority, the drone swarm multi-agent system first embeds the features of the agents and also calculates the average risk of the agents. The deterministic features such as the agent's position and ability are directly transmitted through the FC network; the average risk of the uncertain features is sequentially input into the FC, GRU, and FC layers to obtain a new representation as a new feature. A multi-head attention mechanism is introduced to obtain the original attention level of the agent in each attention head for the current task or the next task based on the feature vector of the agent. The concept of additional time is introduced to measure the time distance from the current time to the task deadline, and it is sent to the negative weight network to obtain the attention bias of the deadline. Then, the attention bias of the deadline is used to adjust the original attention level to obtain the final attention level.

[0010] The decision learning phase of the multi-agent system sets an attention weight threshold based on the number of existing tasks and the next batch of tasks, limiting the agent's observation tasks to the high-priority task set. At the same time, the current time, all task arrival times, and deadline attributes are added to each agent's observation. Finally, the decision learning phase of the multi-agent system penalizes task timeouts in the reward function and uses this to guide the agent's decision-making.

[0011] The specific technical solutions are as follows:

[0012] A task allocation method for auxiliary task deadlines in a drone group is characterized in that the tasks of the drone group in a high-dynamic game scenario are composed of main tasks and auxiliary tasks. The completion of auxiliary tasks before the deadline can accelerate the completion of other tasks. The specific steps are as follows: first, task time encoding processing is performed, and a new feature vector representation is calculated for each task when considering the deadline and uncertain auxiliary relationship of the task; then, task deadline priority calculation is performed, and the attention priority of the agent to the task is corrected according to the deadline of the task; finally, according to this priority, a multi-agent decision algorithm is used to calculate the task allocation strategy for each agent.

[0013] As a further improvement of the present invention, the task set in a high-dynamic game scenario consists of a main task set and an auxiliary task set, each main task being represented by task type, task location, task size, task arrival time and task deadline; each auxiliary task is additionally represented by an auxiliary range and an auxiliary coefficient.

[0014] As a further improvement of the present invention, the tasks in the high-dynamic game scene arrive in batches, and after a period of time, the tasks of the current batch will disappear and the tasks of the next batch will arrive.

[0015] As a further improvement of the present invention, the method includes: in the task time encoding stage, the system considers the uncertain characteristics of the tasks and the uncertain auxiliary relationships between tasks, and obtains a new feature vector for each task; in the task deadline priority stage, by comprehensively considering the uncertainty characteristics of the agents and their attention levels when performing various tasks, and considering the deadline time characteristics of the tasks, the priority of the agents to allocate tasks is determined, and new feature vectors of the agents are generated; in the multi-agent system decision learning stage, according to the attention weights of the agents to the tasks, high-priority tasks are screened out for observation learning, the agent state observation space is set, and a reward function suitable for the deadline characteristics is designed, which is applied to the multi-agent reinforcement learning algorithm to help learning and guide agent decision-making.

[0016] As a further improvement of the present invention, the task time encoding stage includes: embedding task features, calculating attention scores between tasks, calculating deadline attention bias and correcting attention scores, and using new attention scores to update task feature vectors.

[0017] As a further improvement of the present invention, the task feature embedding is specifically as follows: first, the average risk of the area where the task is located is calculated as a new attribute of the task feature vector. For the main task, the deterministic features are the task type, task size, and task deadline, and the uncertain feature is the average risk; for the auxiliary task, the deterministic features are the task type, task size, task deadline, auxiliary coefficient, etc., and the uncertain features are the auxiliary range and the average risk; for the uncertain features of the task, they are sequentially passed through the FC layer, the GRU network layer, and the FC layer, and the output of the network layer is used as the new feature of the task to replace the original uncertain feature; for the deterministic features, they are directly passed through the FC network layer, and finally the new features obtained by the GRU layer are spliced ​​with the original deterministic features to obtain a new feature vector representation of the task;

[0018] Specifically, the attention score between the tasks is calculated by using the graph neural network technology to calculate the attention coefficient between the two nodes; when calculating the attention score of the task node, the attention score calculation is performed only when the two tasks are existing tasks or tasks of the next batch, and the variable t is defined neibor , calculated as follows:

[0019]

[0020] The multi-agent system only needs to consider t neibor =0 task, t neibor Tasks > 0 are ignored;

[0021] The calculation of the deadline attention bias and correction of the attention score is specifically as follows: constructing a negative weight network structure ct(·), and the input is the task deadline t ddl , use the obtained result to correct the original attention score, that is:

[0022]

[0023] This step only calculates t neibor =0 task.

[0024] As a further improvement of the present invention, the task deadline priority calculation stage includes: embedding agent features, calculating the original attention score of the agent for the task, calculating the deadline attention bias, correcting the original attention score using the attention bias, and updating the agent feature vector according to the attention weight.

[0025] As a further improvement of the present invention, the agent feature embedding is specifically as follows: first, the average risk of the agent is calculated, and it is successively input into the FC layer, the GRU layer, and the FC layer to obtain a new representation of the average risk, and other deterministic features of the agent are passed through the FC layer, and finally the agent obtains a new feature vector;

[0026] Specifically, the original attention score of the computing agent for the task is to consider only the existing tasks and the next batch of tasks, adopt a multi-head attention mechanism, calculate the attention score of the agent for the task, and reflect the priority of the task;

[0027] The calculation deadline attention bias is specifically defined as follows: Where currTime represents the current time, represents equivalent distance, additional time Represents the degree of adequacy of agent a to complete task t. The multi-agent system gives priority to tasks whose additional time is greater than 0 but small. Build a negative weight network dt γ (·),Will Input, get the cutoff attention bias

[0028] The method of using attention bias to correct the original attention score is as follows: using the arrival time bias bt γ (Δt arr ) and cut-off time bias Correct the original attention score to get the final attention score: After normalizing the attention scores, the attention weights of all attention heads are concatenated and then linearly transformed to obtain the final attention weight.

[0029] As a further improvement of the present invention, an attention weight threshold w is set in the decision learning stage of the multi-agent system. td =1 / cnt, where cnt represents the total number of existing tasks and the next batch of tasks in the high-dynamic game scenario. If the agent's attention weight for the task is greater than w td , then the task is included in the high priority task set Medium; also sets a lower bound on the number of high priority tasks and upper bound |A| is the total number of agents, which limits the number of tasks in the high-priority task set: Limit the agent's observation tasks to the set middle.

[0030] As a further improvement of the present invention, the following reward function is designed in the hierarchical reinforcement learning stage:

[0031]

[0032] Among them, fac τ represents the reward coefficient, t type Indicates the task type, t state Indicates the task status, t size Indicates the task size.

[0033] Beneficial effects:

[0034] (1) Improving the task completion rate In the task allocation method for auxiliary task deadlines in drone groups, since the uncertain auxiliary relationship between tasks is modeled and the deadline attention bias is used to calculate the task priority, this algorithm can better screen out the most urgent and important tasks for each intelligent agent, so the task completion rate obtained by the algorithm is improved.

[0035] (2) Improving the system survival rate The task allocation method based on the deadline of auxiliary tasks in the drone group integrates the average risk into the characteristics of the agent and the task through the recurrent neural network layer, and also uses the deadline attention bias to fuse the features, thereby improving the survival rate of the drone group in various scenarios.

[0036] (3) Improving the utility function value The task allocation method based on the deadline of auxiliary tasks in the drone group can better take into account the relationship between the uncertainty of auxiliary tasks and the deadline. The system can better decide on more prioritized and more urgent tasks, thereby improving the utility value achieved by the drone group. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 It is the detailed method flow of the task time coding processing phase.

[0038] Figure 2 It is the detailed method flow of task deadline priority calculation.

[0039] Figure 3 It is the decision-making learning method process of multi-agent system.

[0040] Figure 4 It is the main principle diagram of the method of the present invention. DETAILED DESCRIPTION

[0041] The present invention will be further explained below in conjunction with the accompanying drawings and specific embodiments. It should be understood that the following specific embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention.

[0042] As shown in the figure, the present invention provides a task allocation method for auxiliary task deadlines in a drone group, aiming to solve the impact of auxiliary task timeout on system stability. The completion of auxiliary tasks before the deadline can accelerate the completion of other tasks. In a highly dynamic game scenario, the scope of action of auxiliary tasks is uncertain, and the cascading diffusion effect between auxiliary tasks is also uncertain, which increases the complexity of the problem. The present invention first time-encodes the tasks and generates feature vectors; then calculates the task deadline priority and updates the agent features; and finally applies multi-agent system decision learning to perform task allocation. Specifically including:

[0043] (1) In a highly dynamic game confrontation scenario, since the tasks have deadlines, the UAV multi-agent system needs to reasonably allocate tasks to complete the tasks as much as possible before the task timeout. The system combines the uncertainty of the auxiliary task's auxiliary capabilities and the deadline characteristics to design a task allocation strategy to ensure that the auxiliary tasks can maximize their utility before the timeout, so as to maximize the task completion rate and system survival rate. This method is mainly divided into three stages: task time encoding, task deadline priority calculation, and multi-agent system decision learning.

[0044] (2) Task time encoding stage: The drone swarm multi-agent system takes the task risk as a new feature, and models the relationship between tasks in a graph. It uses the attention algorithm to calculate the attention score between tasks, and then corrects the attention score through the deadline attention bias, so that the multi-agent system pays more attention to tasks with earlier deadlines, and finally obtains a new feature vector for the task. At the same time, in order to reduce the scale of the algorithm, the algorithm only considers existing tasks and the next batch of tasks. The method is divided into four steps. The first step is to embed the task features, that is, embed the environmental risk of the task into the feature vector of the task itself; the second step is to calculate the attention score between all existing tasks and the next batch of tasks; the third step is to calculate the deadline encoding of the task, and use this encoding as the attention bias to correct the attention score; the fourth step is to calculate a new feature vector for each task based on the corrected attention score.

[0045] a) Task feature embedding: First, calculate the average risk of the task; for uncertain features such as the average risk of the task, send them to the FC, GRU, and FC network layers in sequence, and use the output of the network layer as the new feature of the task to replace the original uncertain feature; for deterministic features, directly pass them through the fully connected network, and finally concatenate the new features obtained by the GRU layer with the original deterministic features to obtain the new feature vector representation of the task.

[0046] b) Calculate the attention score between tasks: In this step, define the variable t neibor , calculated as follows:

[0047]

[0048] This variable is used to measure whether the task is an existing task or a task in the next batch. neibor = 0, the system needs to consider the task, t neibor >0, the system does not need to consider this task. Then use graph neural network technology to calculate t neibor =0.

[0049] c) Calculate the attention bias of the deadline and correct the attention score: Construct the negative weight network ct(·) and input the task deadline t ddl , get the attention bias, and then use the attention bias to correct the original attention score obtained in the previous step:

[0050]

[0051] d) Use the new attention score to update the task feature vector.

[0052] (3) Task deadline priority calculation stage: In highly dynamic game scenarios, tasks will always arrive and disappear, that is, tasks have arrival time and deadline attributes. Therefore, the UAV agent needs to first select high-priority tasks that bring high benefits, low execution costs and close deadlines for priority execution. Task priority is represented by attention weight. The higher the attention weight of an agent for a task, the higher the priority of this task for this agent. Specifically, the calculation of the attention weight of a task for an agent is divided into the following steps: 1) pass the agent through the GRU layer to process the uncertain features of the agent; 2) calculate the original attention score of the agent for the task; 3) calculate the attention bias based on attributes such as deadline; 4) use the attention bias to correct the attention score; 5) update the feature vector of the agent based on the attention weight. The final attention weight obtained is the priority of the task for the agent.

[0053] a) Agent feature embedding: Calculate the average risk of the agent. For the uncertain environmental risk feature of the agent, pass it through the FC layer, GRU layer, and FC layer in sequence to obtain a new representation of the average risk, which is then spliced ​​onto the original feature.

[0054] b) Calculate the original attention score of the agent for the task: The system uses a multi-head attention mechanism to calculate the agent’s attention score for t neibor =0. The higher the score, the higher the priority of the task for the agent.

[0055] c) Calculate the deadline attention bias: Since the tasks have time constraints, the multi-agent system needs to prioritize high-priority tasks that are time-sensitive, high-reward, and low-cost. Based on this consideration, for a task, if the time cost of completing it is very high and the task is about to be due, then the agent may give up the task, that is, lower the priority of the task. Define the concept of extra time: represents the time left from the current time when the agent starts to execute the task immediately to the time when the task is completed. Perform a simple transformation, and will Input to the negative weight network dt γ (·), get the attention bias

[0056] d) Use attention bias to correct the original attention score: Use the arrival time attention bias and the deadline attention bias to correct the original attention score and get the final attention score: The attention scores are then normalized, and then the attention weights of all attention heads are concatenated, and then a linear transformation is performed to obtain the final result.

[0057] e) Update the agent’s feature vector based on the attention weights.

[0058] (4) Multi-agent system decision learning stage: setting an attention weight threshold wt the =1 / cnt, where cnt represents the total number of existing tasks and the next batch of tasks in the environment. If the agent's attention weight for the task is greater than wt thd , then the task is included in the high priority task set In; at the same time, it is the threshold wt thd Setting the lower bound and the previous Limit the number of tasks in the high-priority task set: Limit the agent's observation tasks to the set At the same time, in the upper-level reinforcement learning, each agent observation needs to increase the current time currTime and the arrival time t of all tasks arr and the deadline t ddl ; In the lower-level reinforcement learning, the observation of a single agent needs to add the current time of the environment currTime and the arrival time t of the assigned task arr and the deadline t ddl ; In addition, a reward function is designed The multi-agent system is encouraged to avoid the main task timeout as much as possible; and then the hierarchical reinforcement learning method is used to guide the agent's decision-making.

Claims

1. A task allocation method for auxiliary task deadlines in a drone group, characterized by: The tasks of a drone swarm in a highly dynamic game scenario consist of main tasks and auxiliary tasks. The completion of auxiliary tasks before the deadline can accelerate the completion of other tasks. The specific steps are as follows: first, task time encoding is performed. When considering the deadline of the task and the uncertain auxiliary relationship, a new feature vector representation is calculated for each task; then, the task deadline priority is calculated, and the attention priority of the agent to the task is corrected according to the deadline of the task; finally, based on this priority, a multi-agent decision-making algorithm is used to calculate the task allocation strategy for each agent.

2. The method for allocating tasks according to the deadline of auxiliary tasks in a drone group according to claim 1 is characterized by: The task set in the high-dynamic game scenario consists of a main task set and an auxiliary task set. Each main task is represented by the task type, task location, task size, task arrival time and task deadline; each auxiliary task is additionally represented by an auxiliary range and an auxiliary coefficient.

3. The method for allocating tasks according to the deadline of auxiliary tasks in a drone group according to claim 1 is characterized by: Tasks in highly dynamic game scenarios arrive in batches, and after a period of time, the tasks of the current batch will disappear and the next batch of tasks will arrive.

4. The method for allocating tasks according to claim 1, wherein: The method includes: a task time encoding stage, in which the system considers the uncertain characteristics of the tasks and the uncertain auxiliary relationship between the tasks, and obtains a new feature vector for each task; a task deadline priority stage, by comprehensively considering the uncertainty characteristics of the agents and their attention levels when performing various tasks, and considering the arrival deadline characteristics of the tasks, the priority of the agents to allocate tasks is determined, and a new feature vector of the agents is generated; a multi-agent system decision learning stage, according to the attention weights of the agents to the tasks, high-priority tasks are screened for observation learning, the agent state observation space is set, and a reward function suitable for the deadline characteristics is designed, which is applied to the multi-agent reinforcement learning algorithm to help learning and guide the agent decision.

5. The method for allocating tasks according to claim 4, wherein: The task time encoding stage includes: embedding task features, calculating the attention scores between tasks, calculating the deadline attention bias and correcting the attention scores, and using the new attention scores to update the task feature vector.

6. The method for allocating tasks according to claim 5, wherein: The task feature embedding is specifically as follows: first, the average risk of the area where the task is located is calculated as a new attribute of the task feature vector. For the main task, the deterministic features are the task type, task size, and task deadline, and the uncertain feature is the average risk; for the auxiliary task, the deterministic features are the task type, task size, task deadline, auxiliary coefficient, etc., and the uncertain features are the auxiliary range and the average risk; for the uncertain features of the task, they are sequentially passed through the FC layer, the GRU network layer, and the FC layer, and the output of the network layer is used as the new feature of the task to replace the original uncertain feature; for the deterministic features, they are directly passed through the FC network layer, and finally the new features obtained by the GRU layer are concatenated with the original deterministic features to obtain a new feature vector representation of the task; Specifically, the attention score between the tasks is calculated by using the graph neural network technology to calculate the attention coefficient between the two nodes; when calculating the attention score of the task node, the attention score calculation is performed only when the two tasks are existing tasks or tasks of the next batch, and the variable t is defined neibor , calculated as follows: The multi-agent system only needs to consider t neibor =0 task, t neibor Tasks > 0 are ignored; The calculation of the deadline attention bias and correction of the attention score is specifically as follows: constructing a negative weight network structure ct(·), and the input is the task deadline t ddl , use the obtained result to correct the original attention score, that is: This step only calculates t neibor =0 task.

7. The method for allocating tasks according to claim 4, wherein: The task deadline priority calculation stage includes: embedding the agent features, calculating the original attention score of the agent for the task, calculating the deadline attention bias, correcting the original attention score using the attention bias, and updating the agent feature vector according to the attention weight.

8. The method for allocating tasks according to the deadline of auxiliary tasks in a drone group according to claim 7 is characterized by: Specifically, the agent feature embedding is to first calculate the average risk of the agent, and input it into the FC layer, the GRU layer, and the FC layer in sequence to obtain a new representation of the average risk, and pass other deterministic features of the agent through the FC layer, and finally the agent obtains a new feature vector; Specifically, the original attention score of the computing agent for the task is to consider only the existing tasks and the next batch of tasks, adopt a multi-head attention mechanism, calculate the attention score of the agent for the task, and reflect the priority of the task; The calculation deadline attention bias is specifically defined as follows: Where currTime represents the current time, represents equivalent distance, additional time Represents the degree of adequacy of agent a to complete task t. The multi-agent system gives priority to tasks whose additional time is greater than 0 but small; build a negative weight network dt γ (·),Will Input, get the cutoff attention bias The method of using attention bias to correct the original attention score is as follows: using the arrival time bias bt γ (Δt arr ) and cut-off time bias Correct the original attention score to get the final attention score: After normalizing the attention scores, the attention weights of all attention heads are concatenated and then linearly transformed to obtain the final attention weight.

9. The method for allocating tasks according to claim 4, wherein: In the decision-making learning phase of the multi-agent system, an attention weight threshold w is set td =1 / cnt, where cnt represents the total number of existing tasks and the next batch of tasks in the high-dynamic game scenario. If the agent's attention weight for the task is greater than w td , then the task is included in the high priority task set Medium; also sets a lower bound on the number of high priority tasks and upper bound |A| is the total number of agents, which limits the number of tasks in the high-priority task set: Limit the agent's observation tasks to the set middle.

10. The method for allocating tasks according to claim 9 for the deadline of auxiliary tasks in a drone group, characterized in that: The following reward function is designed in the reinforcement learning stage: Among them, fac τ represents the reward coefficient, t type Indicates the task type, t start Indicates the task status, t size Indicates the task size.