Task offloading method, system and apparatus based on task priority

Abstract

The present application belongs to the technical field of task offloading, and particularly relates to a task offloading method, system and device based on task priority. An aerospace-terrestrial integrated system model is constructed, which comprises a user device, a UAV, a base station and a satellite. The computing tasks generated by the user device have different priorities. Each time slot outputs the trajectory and transmission power of each UAV by a multi-agent deep deterministic policy gradient network. The output is used as a fixed constraint, and a double-layer alternating iterative optimization is used to solve the offloading decision and computing resource allocation, wherein the first layer determines the offloading target based on the best response, and the second layer allocates resources according to the task weight ratio. The decision is executed, and the actual cost is fed back to the reinforcement learning network to update the parameters. Through the decoupling optimization and priority guarantee mechanism, the present application reduces the system weighted time delay and energy consumption overhead, and effectively improves the service quality and resource utilization efficiency of differentiated tasks in a dynamic environment.

Description

Technical Field

[0001] This invention belongs to the field of task unloading technology, specifically relating to task unloading methods, systems, and devices based on task priority. Background Technology

[0002] Driven by the booming development of Internet of Things (IoT) technology, a large number of computationally intensive and latency-sensitive applications are emerging. As a core supporting technology for future 6G, the Space-Air-Ground Integrated Network (SAGIN) is gradually gaining attention. By integrating satellites, unmanned aerial vehicles (UAVs) and ground nodes, it can provide wide-area coverage and computation offloading services for remote IoT devices.

[0003] However, SAGIN also faces many challenges. First, the computing power, coverage, and latency characteristics of UAVs, satellites, and ground nodes in SAGIN differ. Selecting the appropriate offloading node based on task requirements is crucial for reducing system latency and energy consumption. Second, SAGIN auxiliary server (MEC) systems typically cater to multiple service scenarios, each with varying latency, energy consumption, and computing resource requirements. Ensuring timely processing of high-priority tasks while multiple user devices (UEs) share limited resources is a significant factor affecting system service quality.

[0004] Existing research struggles to adapt to dynamically changing network environments. When dealing with multiple decision variables simultaneously, single-layer reinforcement learning methods are prone to problems such as excessively high action space dimensionality, unstable training, and difficulty in policy convergence, making it difficult to effectively balance system latency reduction, energy consumption optimization, and service guarantees for different task priorities. Summary of the Invention

[0005] The purpose of this invention is to provide a task unloading method, system, and apparatus based on task priority.

[0006] The task unloading method based on task priority includes the following steps: S1. Based on the attributes of four types of entities—user equipment, drones, base stations, and satellites—and their communication links, computing resources, and computing tasks, an integrated air-space-ground model is constructed, in which computing tasks generated by user equipment have different priorities. S2. At the beginning of each discrete time slot, obtain the current environmental state from the integrated air-space-ground model. The current environmental state includes the position and computing task attributes of all user equipment, the current position of all UAVs, and the trajectory and power execution results of the previous time slot. Input the environmental state into the pre-trained multi-agent deep deterministic policy gradient network to obtain the flight trajectory adjustment amount and transmission power value of each UAV in the current time slot. S3. Using the flight trajectory adjustment and transmission power value of each UAV in the current time slot as fixed constraints, and keeping these fixed constraints unchanged, a two-layer alternating iterative optimization process is adopted to solve the task offloading target node of each user equipment in the current time slot and the computing resource allocation scheme of each task offloading target node. S4. Perform flight and signal transmission according to the flight trajectory adjustment amount and transmission power value of each UAV in the current time slot, control each user equipment to send the computing task to the corresponding UAV, base station or satellite according to the task offloading target node of each user equipment in the current time slot, and control each task offloading target node to allocate computing resources to the computing task according to the computing resource allocation scheme of each task offloading target node.

[0007] S3 employs a two-layer alternating iterative optimization process, including: the first layer determines the offloading decision for each user device based on the best iterative response; the second layer determines the allocation of computing resources for the user devices served by the offloading target node for each task based on the priority of the computing tasks; the two layers are executed alternately until the overall cost of the current time slot converges.

[0008] The first layer determines the offloading decision for each user device based on iterative optimal response, specifically including: For any user equipment, keep the current offloading decisions of all other user equipment unchanged, assume that the computing tasks of the user equipment are offloaded to each task offloading target node, calculate the comprehensive cost of the user equipment under each assumption, and select the task offloading target node with the minimum comprehensive cost as the best response offloading decision for the user equipment. The uninstallation decisions for all user devices are updated sequentially in a preset order, and this process is repeated iteratively until the uninstallation decisions for all user devices no longer change or the preset maximum number of iterations is reached.

[0009] The second layer determines the allocation of computing resources for each task's offloading target node to the user devices it serves based on the priority of the computing tasks. Specifically, this includes: For any task offloading target node, determine the set of all user devices currently assigned to that task offloading target node; Based on the priority and data volume of each user device's computing task, a resource allocation weight is constructed for that user device. The resource allocation weight is positively correlated with the priority of the user device's computing task and with the data volume of the computing task. The task will be unloaded from the target node and all computing resources will be allocated according to the resource allocation weight of each user device.

[0010] The resource allocation weights are constructed as follows: ,in, For user equipment Resource allocation weights, For the size of the task data, To calculate task priority, This is a priority adjustment factor.

[0011] The process of alternating execution at two layers until the overall cost convergence in the current time slot specifically includes: Initialize the computing resource allocation scheme for each task's unloading target node; In each iteration, the first layer is executed to determine the offloading decision for each user device based on the best response of the iteration. Under the condition of fixing the current computing resource allocation, the offloading decision for all user devices is updated. Then, under the condition that the updated unloading decision is fixed, the second layer is executed to determine the allocation of computing resources of each task unloading target node to the user devices it serves based on the priority of computing tasks, and the allocation of computing resources of each task unloading target node to the user devices it serves is updated. Calculate the overall cost in the current round. If the decrease in overall cost relative to the previous round is less than a preset threshold, terminate the iteration; otherwise, continue to the next round of iteration.

[0012] In S2, the multi-agent deep deterministic policy gradient network is trained with the goal of minimizing the overall cost, which includes the weighted sum of the transmission latency, computation latency, transmission energy consumption, and computation energy consumption of all computation tasks, and the weighting coefficients are determined by the priority of each computation task.

[0013] S4 also includes: At the end of the current time slot, the total cost actually generated by this decision is recorded, and this total cost is fed back as a reward signal to the multi-agent deep deterministic policy gradient network to update the network parameters.

[0014] A task uninstallation system based on task priority is used to implement task uninstallation methods based on task priority, including: The model building module constructs an integrated air-space-ground model based on the attributes of four types of entities: user equipment, drones, base stations, and satellites, as well as their communication links, computing resources, and computing tasks. In this model, the computing tasks generated by user equipment have different priorities. The UAV parameter adjustment module obtains the current environmental state from the air-ground integrated model at the beginning of each discrete time slot. The current environmental state includes the position and computing task attributes of all user equipment, the current position of all UAVs, and the trajectory and power execution results of the previous time slot. The environmental state is then input into a pre-trained multi-agent deep deterministic policy gradient network to obtain the flight trajectory adjustment amount and transmission power value of each UAV in the current time slot. The resource allocation scheme acquisition module takes the flight trajectory adjustment amount and transmission power value of each UAV in the current time slot as fixed constraints. Under the condition of keeping the fixed constraints unchanged, it adopts two-layer alternating iterative optimization processing to solve the task offloading target node of each user equipment in the current time slot and the computational resource allocation scheme of each task offloading target node. The unloading task execution module performs flight and signal transmission according to the flight trajectory adjustment amount and transmission power value of each UAV in the current time slot. It controls each user equipment to send the computing task to the corresponding UAV, base station or satellite according to the task unloading target node of each user equipment in the current time slot. It also controls each task unloading target node to allocate computing resources to the computing task according to the computing resource allocation scheme of each task unloading target node.

[0015] A task-priority-based task unloading device includes a processor and a memory, wherein the processor implements the task-priority-based task unloading method when executing a computer program stored in the memory.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention systematically decomposes the original coupled optimization problem into multiple sub-problems, forming a clear hierarchical relationship between them. Trajectory and power control serve as fundamental constraints, while unloading decisions and resource allocation are jointly solved through alternating iterative optimization under these constraints. Through a decoupling strategy, continuous and discrete decision variables are processed separately, significantly reducing the action dimensions required for the reinforcement learning agent to output. This makes the training process more stable and efficient, and significantly improves the convergence speed.

[0017] This invention introduces a multi-agent deep deterministic policy gradient network, enabling drones to autonomously adjust their flight trajectories to optimize coverage based on the actual distribution of user devices, adaptively adjust transmission power to balance transmission rate and energy consumption based on task load, and automatically avoid collision risks based on the real-time location of other drones, thereby maintaining stable service performance in fully dynamic scenarios.

[0018] This application systematically introduces a task prioritization mechanism at three levels: the level of building an integrated air-space-ground model, the level of computing resource allocation, and the level of comprehensive cost assessment. This ensures that high-priority tasks always receive service quality commensurate with their urgency in a competitive environment where multiple user devices share limited resources.

[0019] Since the unloading decision determines the load distribution of user equipment on the target node of each task, and the resource allocation, in turn, affects the computing cost of each user equipment on each candidate node, thus affecting the optimal unloading selection, this scheme designs a two-layer alternating iterative optimization process, so that the unloading decision and resource allocation can reach a coordinated and stable optimization state within a limited number of iterations, effectively overcoming the technical defects of independent optimization that are prone to getting trapped in local suboptimal solutions. Attached Figure Description

[0020] Figure 1 This is an example of an application scenario for the present invention; Figure 2 Example of associating user equipment with drone coverage area; Figure 3 A comparison of the overall costs of different uninstallation methods; Figure 4 The relationship between the maximum launch power and overall cost of UAVs with different unloading methods. Detailed Implementation

[0021] To further understand the content of this invention, the invention will be described in detail with reference to the embodiments.

[0022] This invention relates to a task unloading method based on task priority, applicable to... Figure 1 The scenario shown includes the following steps: S1. Based on the attributes of four types of entities—user equipment, drones, base stations, and satellites—and their communication links, computing resources, and computing tasks, an integrated air-space-ground model is constructed, in which computing tasks generated by user equipment have different priorities.

[0023] This invention addresses a Space-Air-Ground Integrated Network (SAGIN) assisted mobile edge computing (MEC) model, which consists of... User Equipment (UE) It consists of a relay drone (UAV) and two edge devices—a satellite and a base station—equipped with mobile edge computing (MEC) servers. The UE set and the UAV set are denoted as follows: and . No. The task attributes of each UE are ,in Indicates the size of the task data. This indicates the priority of the calculation task. UE tasks are divided into three categories: low priority, medium priority, and high priority. The larger the value, the more sensitive the task is to latency and the more urgent the service demand. The entire time period is divided into... A time slot of equal length, denoted as A UE task is randomly generated for each time slot. Considering the limited local computing power of the UE, a UAV needs to be deployed to provide MEC services for it. The binary service-related variable for a UE can be defined as follows: When the first The UE is from the first When providing a UAV service, ,otherwise .

[0024] Assuming each UE is served by at most one UAV, therefore . No. The first UAV and the first The three-dimensional Cartesian coordinates of each UE are respectively represented as follows: and ,in It is the first The UAV is at a fixed height, and the UE is located on the ground.

[0025] In actual deployment, to avoid collisions and conserve energy, the flight trajectories of UAVs are carefully planned. Each UAV provides MEC services to ground UEs only within a sub-area of ​​its coverage, and these sub-areas do not overlap. In this SAGIN-assisted MEC system, due to limitations in size, weight, and power consumption, UAVs can only provide limited computing and communication resources. In contrast, edge devices typically have more powerful MEC servers with richer computing and communication resources.

[0026] Furthermore, the priority levels of user equipment include at least three categories: low priority, medium priority, and high priority. High priority tasks correspond to disaster emergency communication or emergency service tasks, while low priority tasks correspond to routine data collection or inspection tasks.

[0027] In each time slot within, no. The distance that the UAVs travel along the horizontal flight direction is This direction is determined by the angle. Definition. Then, the next time slot... The location of each UAV is To avoid collisions between any two UAVs, the Euclidean distance between UAVs must satisfy the constraint. ,in, Indicates except Another drone, Indicate its coordinates, This indicates the minimum safe distance.

[0028] The horizontal distance between the UE and the UAV can be defined as: The maximum elevation angle of the UAV is expressed as... Therefore, the coverage radius of a UAV can be expressed as If the first The tasks generated by the first UE need to be offloaded to the second UE. If processing is performed on a UAV, the UE must be located within the service area covered by the UAV. .

[0029] Transmission from UE to UAV: No. The UE and the first The distance between UAVs can be expressed as: To avoid transmission interference, Orthogonal Frequency Division Multiple Access (OFDMA) is assumed to support communication between the UE and UAV. The communication link is treated as a line-of-sight link; therefore, the channel is modeled using a free-space path loss model. The UE and the first The channel gain between the UAVs is ,in This is the channel power gain at a reference distance of 1 m. During task offloading, assume the uplink bandwidth... They are evenly distributed to each connected UE. At this time, the... The UE and the first The uplink transmission rate between UAVs can be expressed as: ,in It is UAV in time slot Number of UEs serving the service It is the first The transmit power of each UE This is the Gaussian white noise power of each UAV. The UE and the first The transmission delay and energy consumption between UAVs can be calculated using the following formula. , ,in It is the first The receiving power of each UAV.

[0030] UE task processing model: After receiving a task from the UE, the UAV decides how to handle it. For the first... The task of each UE is defined. The set of target nodes for candidate task unloading is used to process computational tasks. If the location index... for This indicates that the computing task is offloaded to the base station for processing. This indicates that the data will be unloaded onto a satellite for processing; other... The value indicates that the processing and computation tasks are performed locally on the UAV. The binary decision variables for each UE are ,in When, it indicates the first Each UE's task is processed and computed locally on the UAV. This indicates that computational tasks are being processed at the base station. This indicates that computational tasks are performed on a satellite. Furthermore, it is assumed that each task can only be performed at one location. . No. The latency and power consumption of the three processing modes for each UE task are as follows: UAV processing mode: Task In the The time required to complete the computation task by performing local processing on a UAV can be expressed as: ,in Indicates from the first The UAV is assigned to the first Computational resources per UE This is the number of CPU cycles required to process one bit of data. The first UAV processing The computational energy consumption of a UE task is defined as follows: ,in This is the CPU effective switched capacitor coefficient for each UAV. During the local processing of the UAV... The total latency of each UE is Total energy consumption is .

[0031] Base station processing mode: Task By relay number Each UAV transmits data to the base station for processing. The time required to complete the task includes both transmission and processing time. It is the fixed location of the base station, the first The location between a UAV and a base station can be represented as: . No. The channel gain between a UAV and a base station can be defined as... The transmission rate is ,in It is the bandwidth allocated to the base station. It is the base station in the time slot Number of UEs serving the service UAV In the time slot The transmission power, This is the maximum transmission power of each UAV. This is the Gaussian white noise power of the base station. (UAV) In UE The transmission delay and transmission power consumption between the base station and the ground station can be expressed as follows: and The computational latency for the base station to process UE tasks can be expressed as: ,in Is the base station assigned to the first The computing resources of the UE are offloaded to the base station. The total latency of each UE is Total energy consumption is .

[0032] Satellite processing mode: Mission By relay number The UAV transmits data to the satellite for processing. The time required to complete the task includes both transmission and processing time. The satellite's position is... , No. The position between the UAV and the satellite is The transmission rate is ,in It is the bandwidth allocated to the satellite. It is the satellite in the time slot Number of UEs serving the service It is the satellite's Gaussian white noise power. It is the first Channel gain between the UAV and the satellite. The UAV in the first The transmission delay between each UE and the satellite is Due to the considerable distance between the UAV and the satellite, the propagation delay cannot be ignored. . No. The UAV in the first The transmission energy consumption between each UE and the satellite is The computational latency for satellite processing of UE tasks can be expressed as: ,in The satellite is allocated to the first The computing resources of the UE. Offloaded to the satellite. The total latency of each UE is Total energy consumption is .

[0033] Problem statement: No. Each UE is at the task offloading target node The average cost per point is By weighting the average weighted latency and the average power consumption of the UE into a comprehensive cost, then the time slot... The system cost is in and These are the weighting factors for latency and energy consumption, respectively. Success rate is expressed as... UAV trajectory control Transmit power control Task unloading decision With computing resource allocation The joint optimization problem, the integrated air-space-ground model, can be represented as: .

[0034] Constraint C1 is the flight area restriction for the UAV; constraint C2 is the distance restriction between UAVs to prevent collisions; constraints C3 and C4 are the flight distance and azimuth constraints for the UAV; constraints C5 and C6 are the service association constraints between the UE and the UAV; constraints C7 and C8 are the task offloading decision constraints for the UE; constraint C9 requires the UE to be located within the service area covered by the UAV; constraint C10 is the transmit power constraint for the UAV; and constraint C11 is the computational resource constraint for the task offloading target node.

[0035] S2. At the beginning of each discrete time slot, obtain the current environmental state from the integrated air-space-ground model. The current environmental state includes the location and computing task attributes of all user equipment, the current location of all UAVs, and the trajectory and power execution results of the previous time slot. Input the environmental state into the pre-trained multi-agent deep deterministic policy gradient network to obtain the flight trajectory adjustment amount and transmission power value of each UAV in the current time slot.

[0036] The multi-agent deep deterministic policy gradient network is trained with the goal of minimizing the overall cost, which includes the weighted sum of transmission latency, computation latency, transmission energy consumption, and computation energy consumption of all computation tasks, and the weighting coefficients are determined by the priority of each computation task.

[0037] The multi-agent deep deterministic policy gradient network treats each drone as an agent. The state space of each agent includes the position coordinates of the drone, the position coordinates of each user device, and the task attributes of each user device in the current time slot. The action space of each agent includes the horizontal movement distance, horizontal movement azimuth angle, and transmission power of the drone. The reward function of each agent is defined as the negative of the comprehensive cost and includes out-of-bounds penalty, collision penalty, and uncovered penalty.

[0038] Furthermore, each UAV is treated as an intelligent agent, and the environment model is established as a state transition process. .

[0039] State space: Composed of the position coordinates of the UAV and UE, as well as the attributes of the UE's computational tasks. The state space can be defined as: .

[0040] Action space: The travel distance, azimuth angle, and transmission power of a UAV are defined as the UAV's movement space. The action is denoted as: Reward function: The first A UAV in environmental condition Next action The UAV will then receive environmental rewards, which help it improve its strategy. The goal of deep reinforcement learning is to maximize the cumulative reward. Therefore, to align with the goal of the optimization problem, the reward function can be defined as follows: ,in It refers to the number of UAVs that have crossed the boundary. It represents the number of collisions between UAVs. It's a penalty for crossing the line. It's a collision penalty. This is a penalty when a UE is not covered by a UAV.

[0041] This process is divided into two phases: centralized training and distributed execution. During training, each agent's Critic network collects the states and actions of all agents and generates a Q-value. During execution, each agent's Actor network makes decisions and acquires actions based solely on its observed local states.

[0042] The deterministic policies of the M agents in the Actor policy network and the target policy network can be expressed as follows: and The parameters of the deterministic policy in the Actor policy network and the Target policy network can be expressed as: For the first The cumulative expected reward for an agent can be expressed as: ,in This represents the experience buffer, which stores the experiences of all agents. , The set of rewards for all agents, with done as the termination condition; This is a reward discount factor. For deterministic strategies... The policy gradient is: ,in The Critic network outputs a condensed action-value function based on the states and actions of all agents as input, used to evaluate the output policy of the Actor network. The Critic policy network is updated by minimizing the loss function. , represented as: Target value for: ,in Let M be the set of actions of M agents; Indicates that there is a delayed update parameter Deterministic strategy set The target network. Delay parameters. It is possible Update, in which These are parameters for soft updates.

[0043] The specific process is as follows: 1. Initialize the Actor policy network and target policy network weight and Experience buffer pool Random noise And the step counter.

[0044] 2. Initialize the environment state at the beginning of each training round. .

[0045] 3. Within each time step, each UAV outputs an action based on the current policy network, and adds random noise to explore the policy space: Input the status and actions of all UAVs into the environment. Perform a two-layer alternating iterative optimization to generate UE task offloading decisions and computational resource allocation results.

[0046] 4. Generate rewards for performing UAV actions and the state at the next moment If a UAV flies out of bounds, collides with another UAV, or if a UAV is not covered, a corresponding penalty will be imposed, and the current state, action, reward, and next state will be recorded and stored in the experience buffer pool. .

[0047] 5. Once the data in the experience buffer reaches the preset quantity, a small batch of data is randomly selected for training. First, the loss function of the Critic network is calculated based on the target value, and the parameters of the Critic network are updated. Then, the gradient descent method is used to update the parameters of the Actor policy network. Finally, the soft update mechanism is used to update the parameters of the target policy network.

[0048] 6. Repeat steps 3 to 5 until the current training round ends. Complete all training rounds in sequence until the UAV policy network converges or the preset number of training rounds is reached.

[0049] S3. Using the flight trajectory adjustment and transmission power value of each UAV in the current time slot as fixed constraints, and keeping these fixed constraints unchanged, a two-layer alternating iterative optimization process is adopted to solve the task offloading target node of each user device in the current time slot and the computing resource allocation scheme of each task offloading target node.

[0050] The process employs a two-layer alternating iterative optimization, which includes: the first layer determines the offloading decision for each user device based on the best iterative response; the second layer determines the allocation of computing resources for the user devices served by the offloading target node for each task based on the priority of the computing tasks; the two layers are executed alternately until the overall cost of the current time slot converges.

[0051] The first layer determines the offloading decision for each user device based on iterative optimal response, specifically including: For any user equipment, keep the current offloading decisions of all other user equipment unchanged, assume that the computing tasks of the user equipment are offloaded to each task offloading target node, calculate the comprehensive cost of the user equipment under each assumption, and select the task offloading target node with the minimum comprehensive cost as the best response offloading decision for the user equipment. The uninstallation decisions for all user devices are updated sequentially in a preset order, and this process is repeated iteratively until the uninstallation decisions for all user devices no longer change or the preset maximum number of iterations is reached.

[0052] Furthermore, in the SAGIN-assisted MEC system, the offloading decisions of each UE are interconnected. Specifically, although each UE can independently select a task processing node among associated UAVs, base stations, and satellites, its selection results affect the communication and computational overhead of other UEs. When multiple UEs select the same task offloading target node, the computational resources of that node need to be shared among multiple UEs; when multiple UEs complete task forwarding through the same transmission link, the corresponding link bandwidth will also be shared. Based on this, the UE task offloading decision problem is modeled as a non-cooperative decision-making process under resource contention conditions, and an iterative mechanism based on optimal response is adopted to enable UEs to compare costs among different offloading targets, thereby obtaining a stable offloading strategy.

[0053] Let the first The offloading decision for each UE is Its value is an integer. In the time slot The internal values ​​remain unchanged. A value of M+1 indicates offloading to the base station, a value of M+2 indicates offloading to the satellite, and other values ​​indicate processing on the UAV itself. Let the offloading decision combination for all UEs be denoted as... Except for the first The offloading decision combination for UEs other than the one UE is: Decision variables can be used to determine the first... The average cost of a UE is expressed as: The above unloading problem is modeled as a non-cooperative game: Each UE aims to minimize its own cost by selecting the optimal response strategy from its set of available offloading strategies. For a given combination of strategies for other UEs... , No. The optimal response for each UE is: This involves each UE evaluating the costs of three offloading options: UAV, base station, and satellite, and selecting the option with the lowest cost. The resulting offloading decision combination... Satisfying the Nash equilibrium condition, satisfying No single UE can unilaterally change its offloading decision and further reduce its costs.

[0054] The first-level process is as follows: First, initialize the task offloading selection variables for each UE. For any UE, calculate the cost of executing its task on each candidate processing node, and determine the current optimal response strategy for that UE based on the principle of minimum cost. After calculating the current optimal response strategy for the UE, determine whether the current optimal response strategy is consistent with its previous offloading strategy. If they are inconsistent, update the UE's offloading strategy to the current optimal response strategy; if they are consistent, keep the UE's original offloading strategy unchanged. Iterate through each UE in a preset order, and perform the above optimal response calculation and strategy update process on each covered UE. After one round of iteration, if no UE offloading strategy has changed in this round, the UE offloading decision process is considered to have reached a stable state; otherwise, continue to the next iteration until a stable state is reached or the preset maximum number of iterations is reached. Finally, based on the offloading decision variables of each UE in the stable state, generate the UE task processing decision result to determine whether the task of each UE is processed locally in the UAV or offloaded to the base station or satellite for processing.

[0055] The second layer determines the allocation of computing resources for each task's offloading target node to the user devices it serves based on the priority of the computing tasks. Specifically, this includes: For any task offloading target node, determine the set of all user devices currently assigned to that task offloading target node; Based on the priority and data volume of each user device's computing task, a resource allocation weight is constructed for that user device. The resource allocation weight is positively correlated with the priority of the user device's computing task and with the data volume of the computing task. The task will be unloaded from the target node and all computing resources will be allocated according to the resource allocation weight of each user device.

[0056] In the second layer, computational resources are allocated to UAVs, base stations, and satellites for the task computation process. With fixed UAV trajectories, transmit power, and offload decisions, the computational resource allocation subproblem can be constructed as a convex optimization function. To reduce real-time solution complexity and emphasize the importance of high-priority tasks, a closed-loop resource allocation method based on task weights is adopted. This method maps UE priority and task data volume to resource allocation weights, and proportionally allocates computational resources to UAVs, base stations, and satellites according to these weights.

[0057] The specific algorithm flow is as follows: For any task unloading target node, let its available computing resources be... For the first node assigned to the unloading target node of this task Each UE constructs resource allocation weights: ,in, For the first Resource allocation weights for each user device For the size of the task data, To calculate task priority, As a priority adjustment factor, in one optional implementation, The range of values ​​is In order to strike a compromise between prioritizing high-priority tasks and maintaining fairness in resource allocation, The value is 1.5.

[0058] For the task unloading target node The sum of the weights of all UEs on the platform is Unload the task from the target node. The total computing resources are allocated to each UE according to the UE resource allocation weight ratio. .

[0059] Combining the solutions from the two layers mentioned above, this invention proposes an alternating iterative optimization approach.

[0060] The process of alternating execution at two layers until the overall cost convergence in the current time slot specifically includes: Initialize the computing resource allocation scheme for each task's unloading target node; In each iteration, the first layer is executed to determine the offloading decision for each user device based on the best response of the iteration. Under the condition of fixing the current computing resource allocation, the offloading decision for all user devices is updated. Then, under the condition that the updated unloading decision is fixed, the second layer is executed to determine the allocation of computing resources of each task unloading target node to the user devices it serves based on the priority of computing tasks, and the allocation of computing resources of each task unloading target node to the user devices it serves is updated. Calculate the overall cost in the current round. If the decrease in overall cost relative to the previous round is less than a preset threshold, terminate the iteration; otherwise, continue to the next round of iteration.

[0061] The specific process is as follows: After obtaining the UAV's trajectory and transmission power, initialize the computational resource allocation. During the iteration process, when the overall cost does not meet the preset convergence condition, perform the following two layers of optimization sequentially: The first layer executes an offload decision method based on the iterative best response under a fixed computational resource allocation; the second layer executes a proportional resource allocation method based on task weights under a fixed offload decision. After each iteration, determine whether the overall cost meets the preset convergence condition. If not, return to the first layer and continue to the next iteration; if it does, stop the iteration. After the iteration ends, output the final UE task offload decision and computational resource allocation results.

[0062] S4. Perform flight and signal transmission according to the flight trajectory adjustment amount and transmission power value of each UAV in the current time slot, control each user equipment to send the computing task to the corresponding UAV, base station or satellite according to the task offloading target node of each user equipment in the current time slot, and control each task offloading target node to allocate computing resources to the computing task according to the computing resource allocation scheme of each task offloading target node.

[0063] S4 also includes: At the end of the current time slot, the total cost actually generated by this decision is recorded, and this total cost is fed back as a reward signal to the multi-agent deep deterministic policy gradient network to update the network parameters.

[0064] Total cost = latency cost + energy cost, where: Latency costs include: Transmission latency: The time it takes for a user device to transmit task data to the target node for task unloading; Computational latency: The time it takes for the target node to process the target task using its CPU after the target task is unloaded; Propagation delay: the round-trip time for a signal to travel from the drone to the satellite; Energy costs include: Transmission energy consumption: The electrical energy consumed by user equipment or drones when sending / forwarding data; Computational energy consumption: The electrical energy consumed when the target task unloads the target node to perform CPU calculations.

[0065] The smaller the value, the better the solution. It is a unified metric connecting the four decision variables: "how the drone flies, how much power it uses, who it transmits the task to, and how much computing power it allocates." The ultimate goal of all optimization actions is to minimize this value.

[0066] like Figure 2As shown, 10 user equipments are randomly distributed across two hotspot areas, and two drones are randomly positioned. Each drone, after flight, is almost at the center of a hotspot, enabling it to efficiently offload computational tasks. According to the strategy proposed in this invention, user equipment computational tasks are offloaded to target nodes for processing, resulting in lower system costs.

[0067] from Figure 3 As can be seen, the overall cost of the proposed method and other strategies increases with the increase in workload. Under the fixed transmit power strategy, although the UAV's energy consumption increases, the communication rate is improved, which reduces transmission latency to some extent. The two offset each other, resulting in an overall cost slightly higher than the algorithm in this paper. The UAV local processing strategy has the highest overall cost because user equipment tasks cannot be processed using satellites and base stations, leading to concentrated computational load and high latency.

[0068] Figure 4 This demonstrates the relationship between the maximum transmit power of a UAV and its overall cost. The UAV-local processing strategy has a higher overall cost compared to other strategies because it can only utilize the UAV's local computing resources, regardless of the UAV's transmit power. The method proposed in this invention has a lower overall cost compared to other strategies.

[0069] A task uninstallation system based on task priority is used to implement task uninstallation methods based on task priority, including: The model building module constructs an integrated air-space-ground model based on the attributes of four types of entities: user equipment, drones, base stations, and satellites, as well as their communication links, computing resources, and computing tasks. In this model, the computing tasks generated by user equipment have different priorities. The UAV parameter adjustment module obtains the current environmental state from the air-ground integrated model at the beginning of each discrete time slot. The current environmental state includes the position and computing task attributes of all user equipment, the current position of all UAVs, and the trajectory and power execution results of the previous time slot. The environmental state is then input into a pre-trained multi-agent deep deterministic policy gradient network to obtain the flight trajectory adjustment amount and transmission power value of each UAV in the current time slot. The resource allocation scheme acquisition module takes the flight trajectory adjustment amount and transmission power value of each UAV in the current time slot as fixed constraints. Under the condition of keeping the fixed constraints unchanged, it adopts two-layer alternating iterative optimization processing to solve the task offloading target node of each user equipment in the current time slot and the computational resource allocation scheme of each task offloading target node. The unloading task execution module performs flight and signal transmission according to the flight trajectory adjustment amount and transmission power value of each UAV in the current time slot. It controls each user equipment to send the computing task to the corresponding UAV, base station or satellite according to the task unloading target node of each user equipment in the current time slot. It also controls each task unloading target node to allocate computing resources to the computing task according to the computing resource allocation scheme of each task unloading target node.

[0070] A task-priority-based task unloading device includes a processor and a memory, wherein the processor implements the task-priority-based task unloading method when executing a computer program stored in the memory.

Claims

1. A task unloading method based on task priority, characterized in that, Includes the following steps: S1. Based on the attributes of four types of entities—user equipment, drones, base stations, and satellites—and their communication links, computing resources, and computing tasks, an integrated air-space-ground model is constructed, in which computing tasks generated by user equipment have different priorities. S2. At the beginning of each discrete time slot, obtain the current environmental state from the integrated air-space-ground model. The current environmental state includes the position and computing task attributes of all user equipment, the current position of all UAVs, and the trajectory and power execution results of the previous time slot. Input the environmental state into the pre-trained multi-agent deep deterministic policy gradient network to obtain the flight trajectory adjustment amount and transmission power value of each UAV in the current time slot. S3. Using the flight trajectory adjustment and transmission power value of each UAV in the current time slot as fixed constraints, and keeping these fixed constraints unchanged, a two-layer alternating iterative optimization process is adopted to solve the task offloading target node of each user equipment in the current time slot and the computing resource allocation scheme of each task offloading target node. S4. Perform flight and signal transmission according to the flight trajectory adjustment amount and transmission power value of each UAV in the current time slot, control each user equipment to send the computing task to the corresponding UAV, base station or satellite according to the task offloading target node of each user equipment in the current time slot, and control each task offloading target node to allocate computing resources to the computing task according to the computing resource allocation scheme of each task offloading target node.

2. The method according to claim 1, characterized in that, S3 employs a two-layer alternating iterative optimization process, including: the first layer determines the offloading decision for each user device based on the best iterative response; the second layer determines the allocation of computing resources for the user devices served by the offloading target node for each task based on the priority of the computing tasks; the two layers are executed alternately until the overall cost of the current time slot converges.

3. The method according to claim 2, characterized in that, The first layer determines the offloading decision for each user device based on iterative optimal response, specifically including: For any user equipment, keep the current offloading decisions of all other user equipment unchanged, assume that the computing tasks of the user equipment are offloaded to each task offloading target node, calculate the comprehensive cost of the user equipment under each assumption, and select the task offloading target node with the minimum comprehensive cost as the best response offloading decision for the user equipment. The uninstallation decisions for all user devices are updated sequentially in a preset order, and this process is repeated iteratively until the uninstallation decisions for all user devices no longer change or the preset maximum number of iterations is reached.

4. The method according to claim 2, characterized in that, The second layer determines the allocation of computing resources for the user devices served by the target node for each task based on the priority of the computing tasks. Specifically, this includes: For any task offloading target node, determine the set of all user devices currently assigned to that task offloading target node; Based on the priority and data volume of each user device's computing task, a resource allocation weight is constructed for that user device. The resource allocation weight is positively correlated with the priority of the user device's computing task and with the data volume of the computing task. The task will be unloaded from the target node and all computing resources will be allocated according to the resource allocation weights of each user device.

5. The method according to claim 4, characterized in that, The resource allocation weights are constructed as follows: ,in, for Resource allocation weights for each user device For the size of the task data, To calculate task priority, This is a priority adjustment factor.

6. The method according to claim 2, characterized in that, The process of alternating execution at two layers until the overall cost convergence in the current time slot specifically includes: Initialize the computing resource allocation scheme for each task's unloading target node; In each iteration, the first layer is executed to determine the offloading decision for each user device based on the best response of the iteration. Under the condition of fixing the current computing resource allocation, the offloading decision for all user devices is updated. Then, under the condition that the updated unloading decision is fixed, the second layer is executed to determine the allocation of computing resources of each task unloading target node to the user devices it serves based on the priority of computing tasks, and the allocation of computing resources of each task unloading target node to the user devices it serves is updated. Calculate the overall cost in the current round. If the decrease in overall cost relative to the previous round is less than a preset threshold, terminate the iteration; otherwise, continue to the next round of iteration.

7. The method according to claim 1, characterized in that, In S2, the multi-agent deep deterministic policy gradient network is trained with the goal of minimizing the overall cost, which includes the weighted sum of the transmission latency, computation latency, transmission energy consumption, and computation energy consumption of all computation tasks, and the weighting coefficients are determined by the priority of each computation task.

8. The method according to claim 1, characterized in that, S4 also includes: At the end of the current time slot, the total cost actually generated by this decision is recorded, and this total cost is fed back as a reward signal to the multi-agent deep deterministic policy gradient network to update the network parameters.

9. A task unloading system based on task priority, used to implement the task unloading method based on task priority according to any one of claims 1-8, characterized in that, include: The model building module constructs an integrated air-space-ground model based on the attributes of four types of entities: user equipment, drones, base stations, and satellites, as well as their communication links, computing resources, and computing tasks. In this model, the computing tasks generated by user equipment have different priorities. The UAV parameter adjustment module obtains the current environmental state from the air-ground integrated model at the beginning of each discrete time slot. The current environmental state includes the position and computing task attributes of all user equipment, the current position of all UAVs, and the trajectory and power execution results of the previous time slot. The environmental state is then input into a pre-trained multi-agent deep deterministic policy gradient network to obtain the flight trajectory adjustment amount and transmission power value of each UAV in the current time slot. The resource allocation scheme acquisition module takes the flight trajectory adjustment amount and transmission power value of each UAV in the current time slot as fixed constraints. Under the condition of keeping the fixed constraints unchanged, it adopts two-layer alternating iterative optimization processing to solve the task offloading target node of each user equipment in the current time slot and the computational resource allocation scheme of each task offloading target node. The unloading task execution module performs flight and signal transmission according to the flight trajectory adjustment amount and transmission power value of each UAV in the current time slot. It controls each user equipment to send the computing task to the corresponding UAV, base station or satellite according to the task unloading target node of each user equipment in the current time slot. It also controls each task unloading target node to allocate computing resources to the computing task according to the computing resource allocation scheme of each task unloading target node.

10. A task unloading device based on task priority, characterized in that, It includes a processor and a memory, wherein the processor implements the task offloading method based on task priority as described in any one of claims 1-8 when executing a computer program stored in the memory.

Images