Satellite Edge Computing Offloading and Resource Allocation Method for Multi-Agent Collaboration

Abstract

A method for satellite edge computing offloading and resource allocation with multi-agent collaboration, which relates to the field of satellite communication and edge computing technology. In order to optimize the resource allocation problem under the offloading strategy, in the present invention, each device can independently generate a task offloading decision without prior knowledge of other devices. After obtaining the task offloading decision, the optimal resource allocation is obtained based on the Lagrangian method. Experimental results show that the proposed algorithm can better adapt to the changes in channel conditions and the processing capabilities of edge nodes, and can effectively reduce the task delay and task discard rate.

Description

Technical Field

[0001] The present invention relates to the technical fields of satellite communication and edge computing. Background Art

[0002] Limited by economic costs and technical levels, the coverage of the current 5G communication network is limited and cannot achieve true full-area coverage. Especially in harsh environments such as deserts, polar regions, and oceans where it is difficult to establish base stations, there is a lack of network coverage, and the ground communication network is vulnerable to natural disasters. Due to its wide coverage and powerful communication, satellite communication has become a powerful supplement to 6G.

[0003] In recent years, the computing power of intelligent devices has been significantly improved, but it will also face higher computing requirements, such as augmented reality (AR), virtual reality (VR), etc. In the transparent transmission mode of traditional non-terrestrial networks, satellites act as relay nodes between ground gateways and user devices. The service mode of requesting services from the ground cloud center through satellite relay occupies a large amount of bandwidth, resulting in unbearable response delays and energy consumption. Inspired by terrestrial multi-access edge computing (MEC), satellite edge computing sinks the caching and computing functions to LEO satellites, and a computing platform composed of multiple LEO satellites provides computing offloading services for users. The wide coverage of satellites can provide network services for users in remote areas. Deploying MEC servers on LEO satellites can reduce the number of link transmissions between satellites and the ground, data transmission congestion, and user response delays.

[0004] In recent years, deep reinforcement learning has been widely applied in satellite edge computing. C. Qiu et al. proposed a software-defined satellite-ground network that jointly manages and orchestrates network, caching, and computing resources. Experimental results show that the network converges quickly, and the joint optimization strategy is superior to the single optimization strategy. N. Cheng et al. proposed an edge-cloud cooperation architecture where users can offload tasks to the edge cloud or relay them to the cloud center. Y. Wang proposed an actor-critic deep reinforcement network to jointly optimize task offloading and resource allocation. Unmanned aerial vehicle (UAV)-assisted edge computing is proposed in the literature to provide computing offloading services for ground users. It is worth noting that the proposed algorithm serves only one user per time slot. F. Xu et al. proposed a space-air-ground-sea multi-dimensional integration framework. A joint task offloading and resource allocation strategy is proposed, but the drawback is that they only considered the offloading strategy between multiple users and a single multi-access edge computing (MEC) server. N. Chen proposed a multi-agent collaborative computing offloading strategy that takes into account the uplink and downlink transmission rates and the computing capabilities of edge computing servers. The network training is accelerated based on expert samples and priority strategies, but they did not consider the possible queue delay that may lead to task failure. M. Tang proposed a distributed multi-agent cooperation model considering task queuing time and maximum tolerable delay. However, they did not consider the impact of time-varying channel transmission rate and resource allocation on system performance.

[0005] Based on the above analysis, we consider the task offloading scenario of multi-user multi-edge computing satellites, which not only takes into account the time-varying channel transmission rate but also the impact of task queue delay and the dynamic load of edge computing satellites, which is closer to the actual situation. Tasks with processing delays exceeding the maximum tolerable delay will be discarded. We propose a decentralized multi-agent collaborative task offloading and resource allocation algorithm. Devices generate offloading decisions and optimal resource allocations in a distributed manner based on their observed states. We first propose a multi-user multi-edge computing satellite multi-agent collaborative task offloading model with weighted delay as the optimization objective. Each agent can independently generate offloading decisions according to its own state without prior knowledge of other agents. We jointly consider the time-varying channel transmission rate, queue delay, and the dynamic load of edge computing satellites to minimize the system delay. Considering the non-convexity of the optimization problem, we decouple the original optimization problem into task offloading and resource allocation sub-problems. A deep reinforcement learning algorithm is proposed to obtain the offloading decision. After fixing the task offloading decision, we prove that the resource allocation sub-problem is a convex optimization problem and can be solved by the Lagrange multiplier method. Summary of the Invention

[0006] The present invention aims to optimize the resource allocation problem under the offloading strategy, and thus proposes a multi-agent collaborative satellite edge computing offloading and resource allocation method.

[0007] Method for task offloading and resource allocation among multiple agents in satellite edge computing, which includes the following steps:

[0008] Step 1, mobile users generate tasks at a certain probability in each time slot, and for the generated tasks, choose to calculate locally or offload them to the edge computing satellite for calculation, as the user offloading strategy.

[0009] Step 2, according to the user offloading strategy obtained in Step 1, obtain the utility function of each user under the corresponding offloading strategy.

[0010] Step 3, construct the utility function of the entire system according to the utility function of each user under the corresponding offloading strategy obtained in Step 2;

[0011] Step 4, convert the optimization problem of the utility function of the entire system constructed in Step 3 into a task offloading sub-problem and a resource allocation sub-problem;

[0012] Step 5, for the user's task offloading sub-problem, use the method of deep reinforcement learning to handle the user task offloading and obtain the task offloading strategy;

[0013] Step 6, after obtaining the task offloading strategy in Step 5, for the calculation resource allocation of the satellite, construct a Lagrangian function to obtain the optimal resource strategy for the satellite to allocate to users;

[0014] Step 7, ground users offload tasks to the edge computing satellite based on the task offloading strategy in Step 5. At the same time, the edge computing satellite performs resource allocation according to the optimal resource strategy obtained in Step 6, and completes one multi-agent collaborative satellite edge computing offloading and resource allocation.

[0015] Furthermore, in Step 1, the user task The choice of either local calculation or task offloading to the edge computing satellite is based on the user's own quality of experience QoE requirements; among them, B u (t) represents the data size of the task; D u (t) represents the number of CPU cycles required to complete the task; τ u (t) represents the maximum tolerable delay of the task;

[0016] The offloading strategy can be expressed as:

[0017]

[0018]

[0019] Among them, x u (t) = 1 represents local offloading, y un(t) = 1 indicates that the user offloads the task to satellite n; n is a positive integer; the offloading strategy of each user is binary. The constraint conditions of the offloading strategy are shown in formulas (1) and (2).

[0020] Furthermore, in step one, in the selection of the user offloading strategy: when the user selects local computing, considering the impact of the task queue delay, the task completion time slot or discard time slot can be expressed as:

[0021]

[0022] where: represents the waiting delay of the current task computing queue; represents the processing delay of the current task; the first term of formula (3) does not consider the task completion time slot with the maximum tolerable delay, and the second term represents the task discard time slot when the maximum tolerable delay of the task is reached.

[0023] Furthermore, the time slot required for local computing is expressed as:

[0024]

[0025] Furthermore, in step one, when the user selects to offload to the edge computing satellite for computing, considering the impact of the task transmission queue delay, the user transmission completion time slot or discard time slot is expressed as:

[0026]

[0027] where: represents the waiting delay of the current task transmission queue; represents the transmission delay of the current task; the first term of formula (5) does not consider the task transmission completion time slot without the impact of the maximum tolerable delay, and the second term represents the task discard time slot when the maximum tolerable delay of the task is reached.

[0028] Furthermore, when the user selects to offload to the edge computing satellite for computing, the set of users served by satellite n is expressed as:

[0029]

[0030] where: represents the size of the task of user u arriving at satellite n; represents the queue length of user u in the satellite at time t - 1; S n (t) mainly counts the users to be served by the satellite in the current time slot.

[0031] Furthermore, the number of data bits of the queue of user u that satellite n can process in the current time slot is expressed as:

[0032]

[0033] Wherein: represents the computing resources allocated by the satellite to the user. represents the number of CPU cycles required for the task; δ represents the duration of the time slot.

[0034] Furthermore, the queue update of the satellite is expressed as:

[0035]

[0036] Wherein: represents the number of tasks discarded in the current time slot.

[0037] Furthermore, the time slot when the task starts to be processed at the satellite side can be expressed as:

[0038]

[0039] Wherein, represents the completion time of the task. That is, it is necessary to wait for the previous tasks in the queue to be processed before the current task can be processed.

[0040] Since the users served by the satellite are time-varying, the following conditions are satisfied.

[0041]

[0042] Then the completion time of the task selected by the user for computing offloading is expressed as:

[0043]

[0044] Wherein: I(k∈K) = 1 represents the indication factor.

[0045] Furthermore, in step three, the utility function for constructing the entire system (i.e., Figure 1 the simulation environment composed of the user satellites in

[0046]

[0047] Wherein, C1 represents that the user can only select one offloading strategy; C2, C3 represent that the offloading strategy is binary. C4 represents that the number of users served by the satellite cannot exceed its maximum computing capacity; C5 represents that the computing capacity allocated to the user is positive.

[0048] The present invention has the following features and significant improvements:

[0049] 1. The present invention migrates the edge computing platform to LEO satellites to provide full-coverage computing offloading services for ground users.

[0050] 2. Compared with the traditional offloading scheme, the present invention takes into account the impact of queue delay and dynamic environment changes, which is more in line with the real scenario.

[0051] 3. It is very important to obtain a reasonable offloading decision under resource-constrained conditions. The present invention considers the impact of dynamic satellite load on system delay and task discard ratio.

[0052] 4. The present invention proposes a decentralized multi-agent collaborative task offloading and resource allocation algorithm (MATORA). Devices generate offloading decisions and optimal resource allocations in a distributed manner based on the states they observe. Each agent can independently generate offloading decisions according to its own state without prior knowledge of other agents. We jointly consider the time-varying channel transmission rate, queue delay, and dynamic load of edge computing satellites to minimize system delay. Brief Description of the Drawings

[0053] Figure 1 is a schematic diagram of task offloading for multi-user multi-edge computing satellites;

[0054] Figure 2 is a schematic diagram of decoupling of the optimization problem;

[0055] Figure 3 is a schematic diagram of the MATORA algorithm system;

[0056] Figure 4 is a loss schematic diagram of the MATORA algorithm;

[0057] Figure 5 is a schematic diagram of the change of task discard rate and average task delay with the number of users. Detailed Embodiment

[0058] Detailed Embodiment 1. The following further describes in detail a method for satellite edge computing offloading and resource allocation with multi-agent collaboration according to the present invention in conjunction with the drawings of the specification Figures 1-5 of the present invention.

[0059]

[0060]

[0061] Among them, x u (t)=1 represents local offloading, and y un (t)=1 represents that user u offloads the task to satellite n. It should be noted that the offloading strategy of each user is binary. The constraint conditions of the offloading strategy are shown in formulas (1) and (2).

[0062] I. Communication Model

[0063] Ground users offload computing tasks to computing satellites in an orthogonal manner. Therefore, the uplink data rate between user u and satellite n can be expressed as:

[0064]

[0065] where B is the bandwidth of satellite n, P u is the transmit power of user u, h un (t) is the channel gain between the user and the satellite, and σ 2 is the background noise power.

[0066] Considering that the downlink transmission rate of LEO satellites is much greater than the uplink transmission rate, and the calculation results are generally relatively small, the result return time is ignored. The uplink transmission time of the task can be expressed as:

[0067]

[0068] II. Calculation Model

[0069] When the user selects local computing, the processing time slot of the current task can be expressed as:

[0070]

[0071] where D u (t) represents the number of CPU cycles required for the task. f u represents the computing power of user u.

[0072] Considering the influence of queue delay, the waiting time slot of the local computing queue can be expressed as:

[0073]

[0074] where represents the time slot when the previous task in the queue is completed.

[0075] The task completion time slot or discard time slot can be expressed as:

[0076]

[0077] where represents the waiting delay of the current task computing queue. represents the processing delay of the current task. The first term in formula (3) does not consider the task completion time slot with the maximum tolerable delay, and the second term represents the task discard time slot when the maximum tolerable delay of the task is reached.

[0078] The time slot required for local computing can be expressed as:

[0079]

[0080] When the user selects computing offloading, considering the impact of the task transmission queue delay, the time slot when the user's transmission delay is completed or the discarded time slot can be expressed as:

[0081]

[0082] Among them, represents the waiting delay of the current task transmission queue. represents the transmission delay of the current task. The first term of formula (9) is the time slot when the task transmission is completed without considering the impact of the maximum tolerable delay, and the second term represents the task discard time slot when the maximum tolerable delay of the task is reached.

[0083] When the user selects computing offloading, the set of users served by satellite n can be expressed as:

[0084]

[0085] Among them, represents the size of the task of user u arriving at satellite n. represents the queue length of user u in the satellite at time t - 1. S n (t) mainly counts the users to be served by the satellite in the current time slot.

[0086] The data bits of the queue of user u that satellite n can process in the current time slot can be expressed as:

[0087]

[0088] Among them, represents the computing resources allocated by the satellite to the user. represents the number of CPU cycles required for this task. δ represents the length of the time slot.

[0089] The queue update of the satellite can be updated to:

[0090]

[0091] Among them, represents the number of tasks discarded in the current time slot.

[0092] The start time slot of the task at the satellite end can be expressed as:

[0093]

[0094] Among them, represents the completion time of the task. That is, it is necessary to wait for the previous tasks in the queue to be processed before the current task can be processed.

[0095] Since the users of satellite services are time-varying, Meet the following conditions.

[0096]

[0097] The completion time of the task selected by the user for computing offloading can be expressed as:

[0098]

[0099] Where, Represents the indication factor. Due to data transmission, it may cause the satellite to receive the offloaded task only after multiple time slots.

[0100] III. System Modeling

[0101] Each user selects at most one satellite for task offloading, and a utility function for the entire system is constructed based on task offloading. The minimization problem of the present invention is modeled as follows:

[0102]

[0103] Where, C1 represents that the user can only select one offloading strategy. C2 and C3 represent that the offloading strategy is binary. C4 represents that the users of satellite services cannot exceed its maximum computing capacity. C5 represents that the computing capacity allocated to the user is positive.

[0104] In this part, we regard each device as an independent agent. Each agent can make an offloading decision independently without considering other agents. The time-varying environment makes it difficult to obtain the state transition probability. We propose a task offloading and resource allocation strategy based on distributed DQN. Among them, the neural network is trained on the edge computing satellite, and the user will periodically download the latest network parameters from the edge computing satellite. Each agent outputs an offloading decision according to the state table and the latest network parameters. We can further obtain the optimal resource allocation based on the offloading strategy. Considering that the offloading decision is a discrete variable and the resource allocation variable is a continuous variable, the optimization problem is a non-convex mixed integer problem with high complexity. To solve this problem, we decompose the original optimization problem into task offloading and resource allocation sub-problems.

[0105] IV. Resource Allocation Problem

[0106] According to formula (12), we can obtain the queue length of the current satellite for user u. The processing time slot of the queue in the edge computing satellite is only related to the computing capacity allocated to the queue in the current time slot. Represents the computing capacity allocated by the edge computing satellite u to user u, λ un (t) represents the percentage of computing resources allocated in the current time slot t.

[0107] To minimize the processing time of the edge computing queue, it is necessary to allocate optimal computing resources to users. The computing resource allocation problem of the edge computing satellite can be described as:

[0108]

[0109] It can be seen that the constraint condition of formula (17) is convex and its Hessian matrix is positive definite. Therefore, the resource allocation function is a convex optimization problem. The resource allocation problem can be solved by constructing a Lagrangian function to obtain the optimal computing resource allocation, which can be expressed as:

[0110]

[0111] V. Task Offloading Problem

[0112] Considering the uncertainty of the time-varying channel transmission rate and the load of the edge computing satellite, to reduce the algorithm complexity, we model the task offloading problem as an MDP process. Define three elements of reinforcement learning for the MATORA algorithm: state, action, and reward.

[0113] (1) State

[0114] Reinforcement learning aims to continuously learn from historical information. The RL-Agent can output the optimal offloading decision. Therefore, the definition of the state of the RL-Agent seriously affects the offloading decision. We consider the task attributes, queue delay, wireless network environment, and the load of the edge computing satellite, and define the state of user u at time slot t as.

[0115]

[0116] (2) Action

[0117] In our offloading scenario, all mobile devices will be regarded as RL-Agents. The mobile devices with generated tasks will make offloading decisions according to the observed state. The offloading decision mainly includes two aspects. First, it is necessary to judge whether the task is local computing or offloading computing. Second, it is to select the edge computing server to be offloaded. Therefore, the action space of user u at time slot t is defined as.

[0118] a u (t)=(x u (t),y u1 (t),...,y un (t)) (20)

[0119] (3) Reward

[0120] In practice, the reward function should be positively correlated with the objective function. The optimization objective is to minimize the overall system latency, while reinforcement learning aims to maximize the long-term reward. Therefore, we define the reward function as follows:

[0121]

[0122] Suppose π u is the policy mapping from state to action for user u. The goal of the reinforcement learning algorithm is to maximize the expected cumulative discounted reward, that is, to minimize the system latency. Our optimization goal is to find the optimal offloading policy for each user u which can be expressed as:

[0123]

[0124] where γ represents the discount factor.

[0125] VI. MATORA Algorithm

[0126] As described in Table 1, the MATORA algorithm decouples the original optimization problem into a task offloading sub-problem and a resource allocation sub-problem. For the task offloading sub-problem, considering the time-varying channel environment and dynamic edge computing load, we use deep reinforcement learning to obtain the offloading decisions of users. For the resource allocation sub-problem, convex optimization can be used to solve the resource allocation.

[0127] Table 1: Task Offloading and Resource Allocation Methods among Multiple Agents in Satellite Edge Computing

[0128]

[0129]

[0130] VII. Simulation Experiments

[0131] The present invention will be described in detail below in combination with simulation experiments.

[0132] The simulation scenario parameter settings and neural network parameter configurations are shown in Table 2. Considering the time-varying channel environment, queue delay, and dynamic load of the edge computing satellite, we verify the rationality of the MATORA algorithm based on the simulation parameters. The task delay and task discard rate are used as evaluation indicators.

[0133] Table 2 Simulation Parameters

[0134]

[0135]

[0136] Figure 1Schematic diagram of task offloading for a multi - user multi - edge - computing satellite in the satellite edge - computing offloading and resource allocation method for multi - agent cooperation of the present invention;

[0137] Figure 1 As can be seen, the local device has a computing queue and a transmission queue, and the edge - computing satellite maintains the computing queue offloaded by users. The number of users served by the satellite varies in each time slot, which is more in line with the actual situation.

[0138] Figure 2 Decoupling schematic diagram of the satellite edge - computing offloading and resource allocation method for multi - agent cooperation of the present invention;

[0139] Figure 2 It can be seen that we decouple the original optimization problem into a task - offloading and a resource - allocation sub - problem. This is mainly because the task - offloading variable is a binary integer while the resource - allocation variable is continuous. Therefore, this problem is a mixed - integer non - linear programming problem with a high computational complexity.

[0140] Figure 3 System diagram of the MATORA algorithm for the satellite edge - computing offloading and resource allocation method for multi - agent cooperation of the present invention;

[0141] We regard each device as an independent agent. Each agent makes an offloading decision independently to minimize the system overhead based on the observed state, without considering other agents. The time - varying environment makes it difficult to obtain the state - transition probability. We propose a task - offloading and resource - allocation strategy based on distributed DQN, as Figure 3 shown. There are a total of U users, where the neural network is trained on the edge - computing satellite, and the users will periodically download the latest network parameters from the edge - computing satellite. Each agent outputs an offloading decision based on the state table and the latest network parameters. We can further obtain the optimal resource allocation based on the offloading strategy.

[0142] Figure 4 Loss diagram of the MATORA algorithm for the satellite edge - computing offloading and resource allocation method for multi - agent cooperation of the present invention;

[0143] Figure 4 The abscissa of [] is the number of iterations, and the ordinate is the loss value. We first verify the convergence of the MATORA algorithm. From Figure 4 it can be seen that the policy - generating network of the MATORA algorithm converges very quickly. This means that the policy - generating network can output effective offloading decisions.

[0144] Figure 5 Schematic diagram of the variation of the task - discard rate and the average task delay with the number of users in the satellite edge - computing offloading and resource allocation method for multi - agent cooperation;

[0145] Figure 5 The abscissa is the number of iterations, and the ordinate is the task discard ratio and the average delay.

[0146] The proposed MATORA algorithm takes into account the dynamic changes of the channel and the dynamic load of the edge computing satellite, and can make more reasonable task offloading decisions. Figure 5 It shows that the proposed algorithm is superior to other algorithms and can effectively reduce the task discard rate and the average task processing delay. In the local processing mode, the task discard rate and the average task processing delay fluctuate around a fixed value. This is mainly because the difference in processing capabilities and task sizes among users is not significant. In the random offloading mode, the computing power and channel conditions are not considered. The task discard rate and the average task processing delay in the random offloading mode are higher than those of DRTO and MATORA. Compared with the DRTO algorithm, the MATORA algorithm fully considers the time-varying channel conditions and the dynamic load of the edge computing satellite, and the task discard ratio and the average processing delay of the MATORA algorithm are the smallest. The performance of the MATORA algorithm increases with the increase of users. Because the MATORA algorithm not only considers the dynamic load of the edge computing satellite, but also fully considers the dynamic channel conditions between the user and the edge computing satellite.

[0147] As described above, it is only used to illustrate the technical solution of the present invention and not to limit it. The present invention should not be limited to the content disclosed in this embodiment and the drawings. Any modification made within the spirit and principle of the present invention is within the protection scope of the present invention.

Claims

1. A method for satellite edge computing offloading and resource allocation with multi-agent collaboration, characterized by: It includes the following steps: Step 1, the mobile user generates tasks with a certain probability in each time slot, and for the generated tasks, it selects to calculate locally or offload to the edge computing satellite for calculation as the user offloading strategy; Step 2, according to the user offloading strategy obtained in Step 1, obtain the utility function of each user under the corresponding offloading strategy; Step 3, construct the utility function of the entire system according to the utility function of each user under the corresponding offloading strategy obtained in Step 2; Step 4, convert the optimization problem of the utility function of the entire system constructed in Step 3 into a task offloading sub-problem and a resource allocation sub-problem; Step 5, for the task offloading sub-problem of the user, adopt the method of deep reinforcement learning to handle the user task offloading and obtain the task offloading strategy; Step 6, after obtaining the task offloading strategy in Step 5, for the computing resource allocation of the satellite, construct the Lagrangian function to obtain the optimal resource strategy allocated by the satellite to the user; Step 7, the ground user offloads tasks to the edge computing satellite based on the task offloading strategy in Step 5. At the same time, the edge computing satellite performs resource allocation according to the optimal resource strategy obtained in Step 6 to complete a multi-agent collaborative satellite-edge computing offloading and resource allocation; In Step 1, in the selection of the user offloading strategy: when the user selects local calculation, considering the influence of the task queue delay, the time slot for task completion or discard is expressed as: Wherein: Indicates the waiting delay of the current task calculation queue; Indicates the processing delay of the current task; the first term of formula (1) does not consider the task completion time slot of the maximum tolerable delay, and the second term indicates the task discard time slot when the maximum tolerable delay of the task is reached; τ u (t) represents the maximum tolerable delay of the task; The time slot required for local calculation is expressed as: In Step 1, when the user selects to offload to the edge computing satellite for calculation, considering the influence of the task transmission queue delay, the time slot for the user to complete transmission or discard is expressed as: Wherein: Indicates the waiting delay of the current task transmission queue; Indicates the transmission delay of the current task; the first term of formula (3) represents the task transmission completion time slot without considering the impact of the maximum tolerable delay, and the second term represents the task discard time slot when the maximum tolerable delay of the task is reached.

2. The method for satellite edge computing offloading and resource allocation with multi-agent collaboration according to claim 1, wherein In step one, the user task The selection of either local computing or task offloading to the edge computing satellite is based on the user's own quality of experience (QoE) requirements; among them, B u (t) represents the data size of the task; D u (t) represents the number of CPU cycles required to complete the task; The offloading strategy is expressed as: Among them, x u (t) = 1 indicates local offloading, and y un (t) = 1 indicates that the user offloads the task to satellite n; n is a positive integer; the offloading strategy of each user is binary; the constraint conditions of the offloading strategy are shown in formulas (4) and (5).

3. The method for satellite edge computing offloading and resource allocation with multi-agent collaboration according to claim 1, wherein When the user selects to offload to the edge computing satellite for calculation, the set of users served by satellite n is expressed as: Wherein: represents the size of the task of user u arriving at satellite n; represents the queue length of user u in the satellite at time t-1; mainly counts the users to be served by the satellite in the current time slot.

4. The multi-agent collaborative satellite edge computing offloading and resource allocation method according to claim 3, wherein The number of data bits of the queue of user u that can be processed by satellite n in the current time slot is expressed as: Wherein: represents the computing resources allocated by the satellite to the user; represents the number of CPU cycles required for the task; δ represents the duration of the time slot.

5. The method for satellite edge computing offloading and resource allocation with multi-agent collaboration according to claim 4, wherein The queue update of the satellite is expressed as: Wherein: represents the number of tasks discarded in the current time slot.

6. The method for satellite edge computing offloading and resource allocation with multi-agent collaboration according to claim 5, wherein The time slot when the task starts to be processed at the satellite end is expressed as: Among them, represents the completion time of the task; that is, the current task starts to be processed only after the previous tasks in the queue are processed; Since the users of satellite services are time-varying, meet the following conditions; Then the task completion time for the user to select computing offloading is expressed as: Wherein: represents an indication factor.

7. The method for satellite edge computing offloading and resource allocation with multi-agent collaboration according to claim 1, characterized in that In Step 3, the construction of the utility function of the entire system is expressed as: C2:x u (t) ∈ {0, 1} C3:y un (t) ∈ {0, 1} Among them, C1 represents that the user can only select one offloading strategy; C2 and C3 represent that the offloading strategy is binary; C4 represents that the users served by the satellite cannot exceed its maximum computing capacity; C5 represents that the computing capacity allocated to the user is positive.

Images