An edge computing task offloading method based on high-altitude communication platform assistance

Abstract

The application discloses an edge computing task unloading method based on high-altitude communication platform assistance, and the method comprises the following steps: acquiring computing resource and communication node information; constructing a communication network model according to the computing resource and the communication node information; wherein the communication network model comprises a ground communication network and a high-altitude communication network; performing unloading calculation in the communication network model by using a deep deterministic policy gradient algorithm to obtain an unloading decision and a computing resource allocation scheme; and performing edge computing task unloading according to the unloading decision and the computing resource allocation scheme. Through reasonable scheduling of computing resources, the application assists vehicles to complete computing tasks, and through coexistence of edge servers and parked vehicles, the unloading efficiency and communication stability are ensured, and the service quality is greatly improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computing resource utilization, and particularly relates to an edge computing task offloading method based on high-altitude communication platform assistance. BACKGROUND

[0002] The main features of the new era data transmission task are large data volume and low delay. In the face of the situation of scarce computing resources, insufficient servers, unstable communication conditions, and few ground communication nodes in the suburban area of the city, the computing resources are difficult to fully utilize due to the long transmission distance of the cloud server, and only vehicles in close proximity can quickly offload tasks and get instant feedback. From the perspective of data transmission, the communication capacity of the backbone network is too limited compared to so many task requests, further leading to core network congestion and excessive average transmission delay. In fact, as a supplement to cloud computing, mobile edge computing is considered an ideal method to solve the above problems. By deploying edge servers in the vicinity, transmission delay and computing delay are reduced, and fast response is achieved.

[0003] However, the scarcity of computing resources and the integrity of roadside units in the suburbs will hinder the implementation effect of vehicle edge computing. First, compared with the urban core area, the server deployment density in the suburbs is low, and the computing resources that can be provided are relatively limited, which has a great impact on computing delay. Second, compared with the bustling urban core area, the deployment density and maintenance frequency of communication equipment in the suburbs are relatively low, and are easily damaged by natural disasters and bad weather, which also has a great impact on transmission delay. Therefore, the traditional edge computing server is difficult to meet the computing requirements, and new computing nodes are urgently needed to supplement.

[0004] Therefore, the prior art still needs to be improved and developed. SUMMARY

[0005] The technical problem to be solved by the present application is to provide an edge computing task offloading method based on high-altitude communication platform assistance, aiming at solving the problem that the traditional edge computing server is difficult to meet the computing requirements and new computing nodes are urgently needed to supplement.

[0006] The technical scheme adopted by the present application to solve the technical problem is as follows:

[0007] In a first aspect, the present application provides an edge computing task offloading method based on high-altitude communication platform assistance, wherein the method comprises:

[0008] obtaining computing resource and communication node information;

[0009] constructing a communication network model according to the computing resource and communication node information; wherein the communication network model comprises a ground communication network and a high-altitude communication network;

[0010] The deep deterministic policy gradient algorithm is used in the communication network model to perform offloading calculation, to obtain an offloading decision and a calculation resource allocation scheme;

[0011] According to the offloading decision and the calculation resource allocation scheme, edge computing task offloading is performed.

[0012] In an implementation manner, the obtaining of the calculation resource and the communication node information comprises:

[0013] obtaining position information, priority and communication capability of the communication node; wherein the communication node comprises a server node, a road side unit node and a high altitude communication platform node, and the priority of the server node and the road side unit node is higher than the priority of the high altitude communication platform node; the server node comprises a physical edge server node and a virtual edge server node; the communication capability comprises calculation capability, residual energy and communication bandwidth;

[0014] According to the position information and the priority, the communication node information is obtained;

[0015] According to the communication capability, the calculation resource is obtained.

[0016] In an implementation manner, the constructing of the communication network model according to the calculation resource and the communication node information comprises:

[0017] According to the communication node information, communication distances between a plurality of road side unit nodes and server nodes are obtained;

[0018] The communication distances are calculated by using a multivariate shortest path algorithm to obtain ground shortest communication paths;

[0019] According to the ground shortest communication paths, the ground communication network is obtained.

[0020] In an implementation manner, the constructing of the communication network model according to the calculation resource and the communication node information comprises:

[0021] According to the communication node information, communication distances between a plurality of high altitude communication platform nodes and server nodes are obtained;

[0022] The communication distances are calculated by using a multivariate shortest path algorithm to obtain high altitude shortest communication paths;

[0023] According to the high altitude shortest communication paths, a high altitude communication network is obtained.

[0024] The ground shortest communication paths are filled with the high altitude shortest communication paths to obtain shortest communication paths.

[0025] According to the shortest communication path and the computing resource, the communication network model is obtained.

[0026] In an implementation manner, the offloading calculation in the communication network model by using the deep deterministic policy gradient algorithm obtains an offloading decision and a computing resource allocation scheme, and the method comprises the following steps of:

[0027] An offloading task in the communication network model is obtained; wherein, the task comprises a starting node and a target node;

[0028] If the starting node and the target node are the same, the task offloading is performed in a local communication node;

[0029] If the starting node and the target node are different, a communication path and a computing resource of the offloading task in the communication network model are obtained;

[0030] According to the communication path and the computing resource, a first state parameter of the offloading task is obtained, and the first state parameter is input into the deep deterministic policy gradient algorithm for offloading calculation to obtain the offloading decision and the computing resource allocation scheme; wherein, the first state parameter comprises a sum of time delays between the starting node and the target node, a computing capability and a residual energy.

[0031] In an implementation manner, the deep deterministic policy gradient algorithm comprises a main executor network, a main critic network, a target executor network and a target critic network; wherein, the main executor network is used to obtain a first offloading decision according to the first state parameter, and obtain a second state parameter according to the first offloading decision; the main critic network is used to obtain a first evaluation value according to the first state parameter and the first offloading decision; the target executor network is used to obtain a second offloading decision according to the second state parameter; and the target critic network is used to obtain a second evaluation value according to the second state parameter and the second offloading decision.

[0032] In an implementation manner, the deep deterministic policy gradient algorithm is used to calculate a mean square error value according to the first evaluation value and the second evaluation value, and update network parameters of the main executor network and the main critic network according to the mean square error value.

[0033] In a second aspect, an embodiment of the present application further provides an edge computing task offloading device based on a high-altitude communication platform assistance, wherein the device comprises:

[0034] A computing resource and communication node information acquisition module is used to acquire computing resource and communication node information;

[0035] The modeling module is configured to construct a communication network model according to the computing resource and communication node information, wherein the communication network model comprises a ground communication network and a high-altitude communication network.

[0036] The computing module is configured to perform offloading calculation in the communication network model by using a deep deterministic policy gradient algorithm to obtain an offloading decision and a computing resource allocation scheme.

[0037] The offloading module is configured to perform edge computing task offloading according to the offloading decision and the computing resource allocation scheme.

[0038] In a third aspect, an embodiment of the present application further provides an intelligent terminal, wherein the intelligent terminal comprises a memory, a processor, and an edge computing task offloading program based on a high-altitude communication platform assisted, which is stored in the memory and executable on the processor, and the processor executes the edge computing task offloading program based on a high-altitude communication platform assisted to implement the steps of the edge computing task offloading method based on a high-altitude communication platform assisted as described in any one of the above.

[0039] In a fourth aspect, an embodiment of the present application further provides a computer readable storage medium, wherein the computer readable storage medium stores an edge computing task offloading program based on a high-altitude communication platform assisted, and the edge computing task offloading program based on a high-altitude communication platform assisted is executable on a processor to implement the steps of the edge computing task offloading method based on a high-altitude communication platform assisted as described in any one of the above.

[0040] Advantages: Compared with the prior art, the present application provides an edge computing task offloading method based on a high-altitude communication platform assisted. First, the computing resource and communication node information are obtained, and a communication network model is constructed, and by supplementing roadside unit nodes and high-altitude communication platform nodes into a traditional edge computing server network, the aggregated stopped vehicle resources are realized, and the situation that a ground communication node cannot be connected can be supplemented by using a high-altitude communication platform, the server network is expanded, and it is ensured that the offloaded task can be transmitted to the nearest service node. Then, the deep deterministic policy gradient algorithm is used for offloading calculation to obtain an offloading decision and a computing resource allocation scheme. Through autonomous learning and autonomous updating of parameters of a neural network, an intelligent offloading decision and a computing resource allocation scheme are obtained, so that the task offloading scheme is more reasonable and efficient. BRIEF DESCRIPTION OF DRAWINGS

[0041] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiments or prior art description will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments described in the present application, and other drawings can be obtained by those skilled in the art without creating any creative labor.

[0042] Figure 1 is a flowchart of a method for edge computing task offloading assisted by a high-altitude communication platform according to an embodiment of the present application.

[0043] Figure 2 is a schematic diagram of a system model according to an embodiment of the present application.

[0044] Figure 3 is a schematic diagram of an information transmission process according to an embodiment of the present application.

[0045] Figure 4 is a schematic diagram of a Q-learning state-action Q table according to an embodiment of the present application.

[0046] Figure 5 is a schematic diagram of a deep Q-learning neural network according to an embodiment of the present application.

[0047] Figure 6 is a schematic diagram of a deep deterministic policy gradient algorithm framework according to an embodiment of the present application.

[0048] Figure 7 is an interface diagram for simulating vehicle flow according to an embodiment of the present application.

[0049] Figure 8 is a schematic diagram of a comparison of convergence speeds of DDPG and DQN according to an embodiment of the present application.

[0050] Figure 9 is a schematic diagram of the influence of changes in the number of users in different algorithms on total latency according to an embodiment of the present application.

[0051] Figure 10 is a schematic diagram of the influence of changes in the average number of stopped vehicles in different algorithms on optimization rates according to an embodiment of the present application.

[0052] Figure 11 is a schematic diagram of the influence of changes in the average number of stopped vehicles in optimal solutions of DDPG (deep deterministic policy gradient algorithm) on offloading rates according to an embodiment of the present application.

[0053] Figure 12 is a schematic diagram of a comparison of optimization rates in the case of a ground communication network breakdown according to an embodiment of the present application.

[0054] Figure 13 is a schematic diagram of a comparison of optimization rates in the case of a ground communication network in good condition according to an embodiment of the present application.

[0055] Figure 14 is a principle block diagram of an offloading device for edge computing tasks assisted by a high-altitude communication platform according to an embodiment of the present application.

[0056] Figure 15is an internal structure principle block diagram of the intelligent terminal provided by the embodiment of the present application. DETAILED DESCRIPTION

[0057] For the purpose, technical solutions and effects of the present application to be more clear and explicit, the present application is further described in detail below with reference to the drawings and examples. It should be understood that the specific examples described herein are only used to explain the present application, and are not used to limit the present application.

[0058] Those skilled in the art can understand that, unless specifically stated, the singular forms "a", "an" and "the" used herein also include the plural forms. It should be further understood that the use of the term "include" in the specification of the present application means that the stated features, integers, steps, operations, elements and / or components are present, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or groups thereof. It should be understood that when we say an element is "connected" or "coupled" to another element, it can be directly connected or coupled to the other element, or there can be an intermediate element. In addition, "connected" or "coupled" used herein can include wireless connection or wireless coupling. The phrase "and / or" used herein includes all or any one of the associated listed items and all combinations thereof.

[0059] Those skilled in the art can understand that, unless otherwise defined, all terms (including technical and scientific terms) used herein have the same meaning as that generally understood by those skilled in the art to which the present application belongs. It should also be understood that terms such as those defined in a general dictionary should be understood to have meanings consistent with those in the context of the prior art, and should not be interpreted to have idealized or overly formal meanings unless specifically defined as such.

[0060] The main features of data transmission in the new era are large data volume and low latency. Due to the long transmission distance, the computing resources of the cloud server are difficult to fully utilize, and only vehicles in close proximity can quickly offload tasks and receive immediate feedback. From the perspective of data transmission, the communication capacity of the backbone network is too limited compared to so many task requests, further leading to core network congestion and excessive average transmission delay. In fact, as a supplement to cloud computing, mobile edge computing is considered an ideal method to solve the above problems. By deploying edge servers in nearby areas, transmission delay and computing delay are reduced, enabling fast response. However, in the suburbs, there are two factors that can hinder the implementation effect of vehicle edge computing: the scarcity of computing resources and the integrity of roadside units. First, compared with the urban core area, the server deployment density in the suburbs is low, and the computing resources that can be provided are relatively limited, which has a great impact on computing delay. Second, compared with the bustling urban core area, the deployment density and maintenance frequency of communication equipment in the suburbs are relatively low, and are easily damaged by natural disasters and bad weather, which has a great impact on transmission delay, so additional computing nodes are needed to supplement.

[0061] At the same time, modern cars themselves have low utilization rates: relevant data shows that about 70% of vehicles spend 95% of their time in parking lots, so the present invention proposes to use stopped vehicles for edge computing to increase computing resources in the edge network. At the same time, in order to ensure the stability of data transmission, it is necessary to consider adding high-altitude communication platforms outside local ground nodes to avoid poor transmission. Even if most ground nodes are paralyzed, task offloading can be completed with the assistance of air nodes. Through the above measures, suburban vehicles have more abundant computing resources and more stable communication environment. Different types of devices are connected to each other, which can effectively utilize the computing resources of the edge network, alleviate the congestion of the core network, and provide better service quality for all vehicles.

[0062] The present invention explores the stability of parked vehicles in parking lots to ensure the availability of aggregated resources. On this basis, a high-altitude communication platform assisted parking vehicle edge computing framework is proposed, in which edge servers and parked vehicles may coexist to ensure offloading efficiency and communication stability. Task data will be transmitted between interconnected nodes through multiple segments to the final computing node to reduce time delay.

[0063] Exemplary method

[0064] The present embodiment provides a high-altitude communication platform assisted edge computing task offloading method.

[0065] As shown in Figure 1 , the method comprises the following steps:

[0066] Step S100, obtain computing resource and communication node information;

[0067] Specifically, in the face of the situation of the scarcity of computing resources, the lack of servers, the unstable communication condition, the small number of ground communication nodes and the vulnerability of the nodes in the urban suburb area. The high-altitude communication platform refers to placing wireless base stations on aircrafts that stay in high altitude for a long time to provide telecommunications services. The stopped vehicle edge computing refers to using the idle computing unit cluster of the stopped vehicle itself as a server to process computing tasks. By obtaining the high-altitude communication platform and the stopped vehicle as computing resources and communication node information, the parked vehicle cluster can be used as a virtual server to process computing offloading tasks and enrich computing resources.

[0068] In an implementation manner, the step S100 includes the following steps:

[0069] In step S101, the position information, priority and communication capability of the communication node are obtained; the communication node includes a server node, a roadside unit node and a high-altitude communication platform node, and the priority of the server node and the roadside unit node is higher than that of the high-altitude communication platform node; the server node includes a physical edge server node and a virtual edge server node; the communication capability includes computing capability, residual energy and communication bandwidth.

[0070] Specifically, in the suburban communication environment, there are three ground communication nodes and one high-altitude communication platform node in the embodiment, as shown in the following table: Figure 2

[0071] 1. Physical edge server. The physical server, as the most stable computing resource, plays an indispensable role in the system. The physical server is generally allocated in the regional center to provide convenient computing resources for moving vehicles and effectively reduce the execution delay of tasks. The scarcity of computing resources in the suburb network makes the number of edge servers possibly not more than one, and E is used to represent the physical edge server, and the tuple (f e ,E e ) represents the computing capability and residual energy of the physical edge server, and the affiliated base station endows the physical edge server with communication capability to exchange data and information with other nodes.

[0072] ​2. Roadside units. In all vehicle communication modes, the most common way is vehicle-to-roadside unit (V2I) communication. Compared with vehicle-to-vehicle (V2V), vehicle-to-roadside unit reduces uncertainty and the communication condition is more reliable. Through the interlinking between roadside units, data can be transmitted to servers with powerful computing capabilities. In the designed system, roadside units and server nodes are organized as a ground communication network, and the roadside unit set is represented as R = {1, 2,..., nl}, and the task data is mainly transmitted through the ground communication network. Roadside units do not process task data, only transmit, so the only useful attribute of roadside units is communication capability W r .

[0073] 3. Virtual edge servers. Virtual edge servers are clusters of parking lot vehicles P = {nl+1, nl+2,..., nl+n2}, which aggregate computing resources within the parking lot and represent independent computing nodes externally. As long as the number of internal vehicles exceeds 1, the parking lot can serve as a potential computing node. We use a set to represent the virtual server set, and each computing node in set P aggregated by the parking lot has similar attributes to physical servers, i.e., the tuple (f r ,E r ,W r ) respectively for computing power, remaining energy, and communication resources (communication bandwidth). Unlike physical edge servers, virtual edge servers are more dynamic and are affected by the movement of internal vehicles. When vehicles enter, computing power and remaining energy increase, and when vehicles exit, computing power and remaining energy decrease.

[0074] 4. High-altitude communication platforms. High-altitude communication platforms are some communicable nodes arranged in the air, such as hovering drones and high-altitude near-earth orbit satellites. Wireless base stations are placed on aircraft that stay in the high sky for a long time to provide telecommunications services, which are considered a broadband wireless access method with good potential application value. Data is first sent by the sender, enters the high-altitude platform acceptance range, and the high-altitude platform accepts the data and forwards it to the data receiver. High-altitude communication platforms, due to their placement in the air, can effectively avoid the loss of communication caused by urban buildings and mountains and hills, reduce signal loss during transmission, and are less likely to be damaged by natural disasters (earthquakes, mudslides) and human damage (car accidents). On the premise of ensuring the effect of transmission signals, it can achieve longer transmission distances. Similar to the previous definition, the set H = {nl+n2+1, nl+n2+2,..., nl+n2+n3} is used to represent high-altitude communication platform nodes. The disadvantage of high-altitude platforms relative to roadside units is that they are more expensive and cannot be deployed in large quantities, so high-altitude communication only plays a supporting role in data transmission, and transmission capacity should be the main consideration. Therefore, the main focus is on its data transmission capacity, i.e., bandwidth W h.

[0075] Step S102, obtaining the communication node information according to the position information and the priority;

[0076] Step S103, obtaining the computing resource according to the communication capability.

[0077] Specifically, in the embodiment, the priority of the server node and the road side unit node is higher than that of the high altitude communication platform node, so that it can be ensured that the vehicle generates task data and transmits it to the nearest communication node, and the road side unit and the server node are preferentially transmitted, and if there is no ground communication node meeting the conditions, the high altitude platform node is selected for transmission. By obtaining the communication node information and the computing resource, the communication node is obtained as a computing unit to obtain the basic unit of modeling.

[0078] Step S200, constructing a communication network model according to the computing resource and the communication node information; wherein the communication network model includes a ground communication network and a high altitude communication network;

[0079] Specifically, in the embodiment, the ground communication node and the high altitude communication platform are respectively organized into a two-layer communication network. The ground node is preferentially selected for task transmission, and in the case where it cannot be reached, the high altitude communication platform is used for auxiliary transmission.

[0080] Specifically, in the embodiment, the suburb area is established as a grid type area with multiple horizontal and vertical roads. Considering that the number of road side units in the suburb area itself is small and the communication condition is unstable, the high altitude communication platform is used to assist in completing data transmission. Compared with the previous studies involving one-way roads and cross roads, the embodiment considers a rectangular area composed of rectangular cross roads, and the data transmission is not limited to the narrow range covered by a single ground communication node, but the transmission delay is set as the multi-section transmission delay between communication nodes to be close to reality. In addition, due to the persistence and stability of the vehicles parked in the parking lot, the parked vehicles are clustered to expand the computing performance of the suburb area, and the high altitude communication platform is used to assist the ground node to ensure the communication condition.

[0081] In one implementation manner, the step S200 of the embodiment includes the following steps:

[0082] Step S201, obtaining the communication distance between a plurality of road side unit nodes and server nodes according to the communication node information;

[0083] Specifically, the traffic road direction and parking lot position distribution in the actual scene are very regular, and the road can be modeled as a general north-south and east-west direction. Different roads intersect to form a rectangular block area, and the parking lot is placed in the center of the area, aggregating the stopped vehicle resources. The vehicles generate tasks and send task data to the nearest communicable unit, and the edge server in the center receives an information indicating the task calculation amount and data amount. After updating the calculation resource and communication node status in the current time environment, the system model is established. According to the location information of the communication node in the communication node information, the communication distance between any roadside unit node and server node can be obtained.

[0084] Step S202, calculating the communication distance by using a multi-source shortest path algorithm to obtain a ground shortest communication path;

[0085] Step S203, obtaining the ground communication network according to the ground shortest communication path.

[0086] Specifically, due to the scarcity of communication and calculation devices, it is difficult to transmit data to the final calculation node in one step within the communication range of a single device. Therefore, it is necessary to complete the data transmission task by multiple communication nodes, such as Figure 3 As shown. The roadside unit and the edge server with communication capability are both communication units of the ground communication network and participate in data transmission. Taking node A and node B as an example, first, the direct communication distance between the two nodes (the height of the ground node is 0, so the direct distance is calculated using the Euclidean distance in two-dimensional space) d A,B is calculated. The communication capabilities (maximum communication distance) of node A and node B are respectively judged, and if d A,B is less than the communication capability, it means that it can be directly transmitted (record d A,B ), otherwise, the direct communication distance between the two nodes is set to positive infinity (inf). After calculating the direct communication distance between the ground nodes, the shortest communication distance and the intermediate nodes on the shortest path between any two ground nodes are obtained by using a multi-source shortest path algorithm (floyd algorithm), and the ground communication network is built.

[0087] Step S204, obtaining the communication distance between the high-altitude communication platform node and the server node according to the communication node information;

[0088] Step S205, calculating the communication distance by using a multi-source shortest path algorithm to obtain a high-altitude shortest communication path;

[0089] Step S206, obtaining a high-altitude communication network according to the high-altitude shortest communication path;

[0090] Step S207, filling the ground shortest communication path with the high-altitude shortest communication path to obtain a shortest communication path;

[0091] Specifically, for high-altitude communication platforms, high-altitude communication nodes and server nodes are organized into high-altitude communication networks according to the foregoing steps. Two-dimensional calculation communication distances are used between high-altitude communication platforms and between server nodes, and three-dimensional calculation communication distances are used between high-altitude communication platforms and servers. Thereafter, node pairs such as A and B in the ground network whose distances are still infinite are searched for the nearest high-altitude communication platform of A, and the shortest distance and path are filled in the high-altitude network as an intermediary. After the above steps, the communication network construction is completed with the ground network as the main part and the high-altitude network as the auxiliary part.

[0092] Step S208, obtaining the communication network model according to the shortest communication path and the computing resource.

[0093] Specifically, in this embodiment, the task model, system model, and transmission model of the stopped vehicle edge computing in the stop suburb area are modeled. First, the task data is modeled. Assuming that the number of vehicles in the area within a certain time period is M, the task generated from vehicle i can be represented as a tuple T i =(s i ,k i ,c i ,t i，max ), where s i represents the amount of task data, k i represents the amount of computation (number of CPU cycles) required to complete a unit amount of data by the required computing resource, c i =s i *k i represents the total amount of computation required to complete task i, and t i，max represents the maximum tolerable delay of the task. The task completion time must be less than t i，max to be considered successfully executed, otherwise the task execution fails.

[0094] Specifically, after modeling the nodes on the entire transmission path, the transmission process of the task data can refer to the following formulas (1)-(3):

[0095]

[0096]

[0097]

[0098] wherein formula (1) is used to calculate the transmission capacity of the node, (j, k) being adjacent nodes in the transmission path of task i. W j is the maximum bandwidth of node j, represents the transmission power of the node, where d represents the distance between two direct communication nodes, k is the path loss, v is a constant, hi represents the channel gain, and formulas (2) and (3) calculate the transmission delay and transmission energy consumption, respectively, where s i is the task data volume.

[0099] Step S300, performing offloading calculation in the communication network model by using a deep deterministic policy gradient algorithm to obtain an offloading decision and a computing resource allocation scheme;

[0100] Specifically, the task offloading refers to deciding which specific computing node the task data is handed over to and allocating how much computing resource to process the task. The embodiment designs a reinforcement learning algorithm (deep deterministic policy gradient algorithm) in a continuous space to give a joint optimization solution (offloading decision and resource allocation scheme) to the problem of stopping vehicle edge computing.

[0101] In an implementation manner, the step S300 in the embodiment includes the following steps.

[0102] Step S301, obtaining an offloaded task in the communication network model; wherein the task includes a starting node and a target node;

[0103] Step S302, if the starting node and the target node are the same, performing task offloading at a local communication node;

[0104] Specifically, in terms of binary offloading, there are two choices, local computing and complete offloading. If the starting node and the target node are the same, the offloaded task is local computing. The power consumption of a processor is approximately proportional to the third power of the frequency, and the computing time is inversely proportional to the frequency, so the overall energy consumption is proportional to the square of the frequency. The main variable in the computing process is the amount of allocated computing resource.

[0105] Specifically, in local computing, each vehicle is a separate computing unit (with a computing capability f0) with limited energy e0. The vehicle uses its own computing unit to execute the task, and in this case, there is no need to transmit data to other nodes. For a locally executed task i, the computing delay and energy consumption are shown in formulas (4) and (5)

[0106]

[0107]

[0108] where μ0 is a constant related to the vehicle itself, c i , s i , k i Define a reference task model. In the case of no data transmission, only consists of computing delay, while the energy consumption Also only the computation energy consumption is taken into account.

[0109] In step S303, if the start node and the target node are different, the communication path and the computing resource of the offloading task in the communication network model are obtained.

[0110] Specifically, if the start node and the target node are different, full offloading is performed, that is, after multi-segment transmission, the server allocates computing resources to assist the vehicle to perform the computing task. The resource allocated to the task i is denoted as f i,off , the computation delay and energy consumption are calculated by using similar formulas (4) and (5), where μ cal is a constant related to the offloading node.

[0111]

[0112]

[0113] Unlike local computing, the sum of the delay and the energy consumption includes the transmission process and the computing process, so is the total delay of the task, and represents the total energy consumption of the task.

[0114] Specifically, without loss of generality, the embodiment uses a={r1, r2,..., r M to indicate the offloading decision of M tasks. If r i =0, the task selects local computing, otherwise, it selects a server k, r i =k. The value of r i must be an integer. We use f={f1, f2,..., f M to represent the computing resources allocated to the task. For a moving vehicle, shorter time delay is equal to better service quality, so we intend to obtain the values of r, f to minimize the total delay. The nature of the task itself, energy reserves, and device computing capacity impose constraints on this problem, and the original unconstrained extreme value problem is modeled as a mixed integer nonlinear programming problem. In the following formula (8) is the optimization objective, which pursues the minimization of the total delay of M tasks, and formula (9) requires the total delay of each task to be less than the maximum tolerable delay in its task model. f i cal represents the computing amount of the task i, represents the energy consumed by the task computation. In equation (10), the offloading decision is 0, which means the task chooses local execution, otherwise, it means the task data is offloaded to the corresponding server node. In equation (11), once the task i chooses local execution, the amount of computing resources allocated to it will be less than the maximum computing resources f0of the vehicle, and the energy consumption will also be less than the remaining energy e0of the vehicle. Once the task chooses to be offloaded to the server node, equation (13) shows that the amount of computing resources allocated is positive, and equations (14) and (15) show that for all tasks allocated to a certain server k, the sum of computing capabilities will be less than the maximum computing capability C k , and the sum of energy consumption will also be less than the energy reserve E k of the server.

[0115] Minimize:

[0116] t i ≤t i,max (9)

[0117] r i ∈{0, 1, 2,..., n2}(10)

[0118]

[0119]

[0120] 0<f i cal =f i,off (r i >0) (13)

[0121]

[0122]

[0123] In step S304, a first state parameter of the offloaded task is obtained according to the communication path and the computing resource, and the first state parameter is input into a deep deterministic policy gradient algorithm for offloading calculation to obtain the offloading decision and the computing resource allocation scheme; wherein the first state parameter includes the sum of time delays between the starting node and the target node, the computing capability and the remaining energy.

[0124] Specifically, the deep deterministic policy gradient algorithm comprises a main executor network, a main critic network, a target executor network and a target critic network; the main executor network is configured to obtain a first offloading decision according to the first state parameter, and obtain a second state parameter according to the first offloading decision; the main critic network is configured to obtain a first evaluation value according to the first state parameter and the first offloading decision; the target executor network is configured to obtain a second offloading decision according to the second state parameter; and the target critic network is configured to obtain a second evaluation value according to the second state parameter and the second offloading decision. The deep deterministic policy gradient algorithm adopts an actor-critic framework, and the action selection process comprises two parts: an actor outputs an action according to a probability, and a critic evaluates and scores the output action, and the final evaluation result is the Q value of the action output by the actor in the current environment.

[0125] Specifically, the deep deterministic policy gradient algorithm is configured to calculate a mean square error value according to the first evaluation value and the second evaluation value, and update network parameters of the main executor network and the main critic network according to the mean square error value.

[0126] Specifically, for edge computing of vehicles, dynamic changes cause difficulties in offloading decisions and resource allocation. A general heuristic algorithm always starts from some initial solutions, and imitates the essence of transactions, and cannot completely match the situation of the joint optimization problem in the time-varying vehicle network. In addition, the heuristic algorithm is too dependent on the design of programmers, and lacks strong universality, so reinforcement learning is selected as a solution to the original problem. Reinforcement learning mainly comprises three spaces: a state space, an action space and a reward space. First, the state is the attribute of the environment that can be perceived by the subject, and is also the information set perceived by the agent to help action selection, and there is a one-to-one mapping between the information and the state. More representation parameters not only mean higher precision, but also mean the complexity of establishing a model. Whether it is a physical edge server or a virtual edge server, the computing power and the remaining energy are indexes for evaluating the performance of the server, and greater computing power can allocate more computing capacity to different tasks, and more remaining energy can ensure the timely completion of the number of tasks. Therefore, the computing power and the time delay sum are added to the state space. In addition, the sum of the time delays is an optimization target that we are concerned about, and therefore it is also a state parameter. Thus, the first state parameter is as shown in formula (16)

[0127] s i =(t sum ,c left ,e left ) (16)

[0128] Specifically, the construction of action space depends on the designer's thinking of the problem situation. The solution space of the aforementioned joint optimization problem is regarded as the action space in this design. In other words, the action will contain two parts: offloading decision and resource allocation. The complexity of this setup lies in that the former is a discrete value, representing the binary offloading decision of server selection, while the latter will be selected in an infinite continuous space, so the solution space is semi-continuous. We use a i = (a i1 , a 2i ) to represent the operation on task i, a i1 is the offloading decision list (integer variable), and a 2i is the allocated computing resource list (floating-point variable). At the end of each time step, the agent reacts to the action selected according to the current state s i and moves to the next state s i+1 .

[0129] a i = (a i1 , a 2i ) (17)

[0130] Specifically, the reward in the learning process is the standard for evaluating the rationality of behavior, which directly affects the update of cumulative reward value. Generally speaking, the learning process of the agent is to select appropriate actions while constantly improving the understanding of the environment and pursue the maximization of cumulative return. However, the joint optimization goal in this scheme is to minimize the overall delay of all tasks, so the cumulative return must be negatively correlated with the time delay. In view of this, the single-step reward in this embodiment is set as the reduction rate of local execution delay, as shown in formula (18)

[0131]

[0132] where t loc,i is the total delay of local execution for all tasks, and t total,i is the total delay of all tasks after executing the offloading decision and resource allocation strategy. In the case where the former is greater than the latter, the reward is negative, otherwise, the reward should be a positive number between 0 and 1, representing the relative reduction rate of total execution delay.

[0133] Specifically, the overall framework of the deep deterministic policy gradient algorithm in this embodiment is shown in Figure 6 :

[0134] (1). Establish a fixed-size storage experience pool to store training data in the form of (s, a, s', r, done), s is the first state parameter, that is, the state, a is the first unloading decision, that is, the action, s' is the second state parameter generated after the state s takes the action a, r is the single-step reward obtained by the first unloading decision, and done represents whether the task has been completed (out-of-bound case).

[0135] (2). Set four neural networks, main executor network, main critic network, target executor network, and target critic network. The two executor networks have the same structure, the two critic networks have the same structure, the main network has a higher update frequency, and the network parameters are copied to the target network every certain period. The fixed-target idea is adopted to establish the main network and the target network, making the learning process more stable and ensuring that the updated target and the judgment standard do not come from the same network.

[0136] (3). Specific steps

[0137] The executor in the main network outputs the first unloading decision a according to the first state parameter s (①);

[0138] The corresponding critic gives the first evaluation value Q(s, a) according to the values of s and a (②, ③);

[0139] The target network executor obtains the second state parameter a' according to the second state parameter s' (①);

[0140] The critic in the target network gives the second evaluation value Q'(s', a') (②, ③);

[0141] Calculate the mean square error based on Q(s, a) and Q'(s', a'), which is used to update the critic network parameters in the main network (④);

[0142] The executor network in the main network updates the network parameters based on the value of Q(s, a) (⑤);

[0143] Set a fixed update period to assign the main network parameters to the target network.

[0144] Specifically, the structure of the main executor network in the embodiment is described. The executor network is composed of an input layer, an output layer and three hidden layers, a total of five layers. The number of units in the input layer is the same as the number of state parameters, and the number of units in the output layer is the same as the number of action parameters. The relu function is used as the activation layer to add non-linear properties. The reviewer network is composed of a state s input layer, an action a input layer, an intermediate layer and an output layer. Since the executor network only outputs one action, improper setting of the network initial value will cause the only output action to always be in the boundary region, and the true result cannot be obtained. Therefore, a normalization layer is set to ensure that the output value is within the action space. The learning rate of the executor and the reviewer is designed to be 0.005, the reward discount factor is 0.99, the buffer size is set to 1000, the batch size is 32 (the number of samples used to update the network parameters), the maximum number of steps and the maximum number of rounds are 500 and 200, respectively.

[0145] In an implementation manner, Q-learning and deep Q-learning are value-based reinforcement methods. The Q value is the cumulative reward. These two methods are realized by requiring the agent to find the best action that maximizes the reward with a certain probability at each state. The Q-learning algorithm mainly aims at a limited state space and a limited action space, and stores the state-action pairs in the form of a Q table. After the Q table is established, all actions in the current state are searched in each iteration process to select the action with the maximum reward. If one of the state quantity or the action quantity tends to infinity, it becomes troublesome to store and query the action value pair through the Q table.

[0146] Further, deep Q-learning combines a neural network with Q-learning. The neural network is used to replace the Q table query operation. The number of input units is the same as the number of state parameters, and the number of output units is the same as the number of action spaces. Through original data training and approximation, the neural network learns the structural parameters of the system. The value of each output unit gradually approximates the Q value obtained by taking the action corresponding to the output unit in the current state. In short, input a state and output the Q values corresponding to all possible actions, and select the action with the maximum reward from them. Since the neural network only reads one state value at a time, even if the total number of states tends to infinity, it will not affect the query and storage. However, for an infinite action space, simple deep Q-learning still does not have a good method to handle it. Generally, large-scale actions are discretized, and part of them is selected as possible output actions. For related concepts, refer to Figure 4 and Figure 5 As shown in the formula (1), S represents a state, and a represents an action.

[0147] Step S400, according to the offloading decision and the computing resource allocation scheme, performing edge computing task offloading.

[0148] Specifically, the embodiment uses the DDPG (Deep Deterministic Policy Gradient) algorithm to apply deep reinforcement learning to edge computing for parking, and obtains the optimal offloading decision and resource allocation scheme according to the resource environment and communication environment. Compared with other schemes such as deep Q learning and particle swarm algorithm, the DDPG algorithm shows excellent performance in the design.

[0149] For example, the scheme adopts an intelligent driver model (IDM) to simulate road traffic conditions. The intelligent driver model is characterized in that the speed is controlled according to the distance between the vehicle and the front vehicle, and when the distance is below a certain level, the vehicle will decelerate at a certain acceleration, which is similar to the real road conditions. The simulated traffic conditions are as shown in Figure 7

[0150] Table 1. System parameter settings

[0151] Structural parameters Values Transmission power and computing power 5w Number of physical edge servers 1 Computing power of physical edge servers 10 GHz / sec Energy reserve of physical edge servers 5 kW Number of virtual edge servers 2 Computing power of a single vehicle 1 GHz / sec Energy reserve of a single vehicle 1 kW Width of the area 3 km Maximum communication distance of a road side unit 1 km Number of road junctions 4 Background noise 174 dBm / Hz

[0152] The embodiment compares the task execution conditions with different algorithms. The relevant algorithms and introductions are described below.

[0153] DDPG: Deep Deterministic Policy Gradient method, a joint optimization problem solving method used in this paper;

[0154] DQN: Deep Q learning, a common reinforcement learning method;

[0155] PSO: Particle swarm algorithm, a heuristic algorithm for solving optimization problems;

[0156] Random: Randomly select offloading strategy and resource allocation scheme;

[0157] Local: All vehicles select locally generated tasks for execution;

[0158] Figure 8 is a comparison of the convergence time of the DDPG and DQN algorithms. After setting the maximum training round to 200 and the maximum training step to 500, it is found that compared with the DQN algorithm, the DDPG can converge in a relatively short round, because it only outputs one action, rather than selecting the action with the maximum return from all actions according to the greedy algorithm. The horizontal axis is the training round.

[0159] Figure 9 ​is set to 5, the total execution delay of different methods is compared by changing the number of users (mobile vehicles) from 4 to 20, the horizontal axis is the number of users (mobile vehicles), and the vertical axis is the total delay (in seconds). Reinforcement learning can achieve better results than the other three offloading schemes and resource allocation strategies. The PSO algorithm is prone to local optimization and cannot obtain global optimization. The DQN algorithm is discretized for the action space, and in some cases, it cannot obtain the optimal solution. DDPG can obtain the optimal value in the global case and performs better.

[0160] Figure 10 In the case of changing the number of average stopped vehicles in the parking lot to affect the virtual server computing power and energy consumption, the number of mobile vehicles is set to 10, the horizontal axis is the number of average stopped vehicles (single parking lot), and the vertical axis is the optimization rate (relative reduction of total delay relative to the Local scheme). More average number of stopped vehicles means more abundant computing resources, which can effectively alleviate the problem of insufficient computing resources in the suburban area, and the optimization rate is also higher. The DDPG algorithm still performs better than other algorithms.

[0161] Figure 11 In the case of changing the number of average stopped vehicles in the parking lot to affect the server offloading ratio, the number of average stopped vehicles is set to 10, and the offloading ratio is calculated as shown in the following table. Figure 10 The results of the experiment are shown in the following table, the horizontal axis is the number of average stopped vehicles (single parking lot), and the vertical axis is the offloading ratio (the ratio of the number of tasks offloaded to a server to the total number of tasks). As described above, a physical edge server (Mec) and two virtual edge servers (Park1 and Park2) are established, and the offloading ratio is calculated for the three servers, and the remaining is the ratio of local execution (local).

[0162] To verify the communication effect of the high-altitude communication platform, two cases are set up for verification, and the results are shown in Figures 12-13 The two subgraphs are two cases, case 1 Figure 13 is that all roadside unit nodes communicate normally and can only complete communication through the ground network, and case 2 Figure 12 is that all roadside units are faulty, and task offloading cannot directly reach the server node, but only indirect transmission through the high-altitude communication platform. The horizontal axis is the number of users (number of mobile vehicles), and the vertical axis is the optimization rate. The results show that even in the case of all ground nodes being faulty, the high-altitude communication platform can still be used as an intermediary to transmit task data to the server node, which shows the role of deploying a high-altitude communication platform in the suburban network to enhance the stability of the communication network.

[0163] Exemplary apparatus

[0164] As shown in Figure 14As shown in the embodiments, the present application also provides an edge computing task offloading device based on high-altitude communication platform assistance, which comprises:

[0165] a computing resource and communication node information acquisition module 10, configured to acquire computing resource and communication node information;

[0166] a modeling module 20, configured to construct a communication network model according to the computing resource and communication node information; wherein the communication network model comprises a ground communication network and a high-altitude communication network;

[0167] a computing module 30, configured to perform offloading computation in the communication network model by using a deep deterministic policy gradient algorithm to obtain an offloading decision and a computing resource allocation scheme;

[0168] an offloading module 40, configured to perform edge computing task offloading according to the offloading decision and the computing resource allocation scheme

[0169] In an implementation manner, the computing resource and communication node information acquisition module 10 comprises:

[0170] an information acquisition unit, configured to acquire location information, priority and communication capability of a communication node; wherein the communication node comprises a server node, a roadside unit node and a high-altitude communication platform node, and the priority of the server node and the roadside unit node is higher than that of the high-altitude communication platform node; the server node comprises a physical edge server node and a virtual edge server node; and the communication capability comprises computing capability, residual energy and communication bandwidth;

[0171] a communication node information acquisition unit, configured to obtain the communication node information according to the location information and the priority;

[0172] a computing resource acquisition unit, configured to obtain the computing resource according to the communication capability.

[0173] In an implementation manner, the modeling module 20 comprises:

[0174] a ground communication distance acquisition unit, configured to obtain communication distances between a plurality of roadside unit nodes and server nodes according to the communication node information;

[0175] a ground shortest communication path acquisition unit, configured to calculate the communication distances by using a multivariate shortest path algorithm to obtain ground shortest communication paths;

[0176] a ground communication network acquisition unit, configured to obtain the ground communication network according to the ground shortest communication paths.

[0177] The high-altitude communication distance obtaining unit is configured to obtain communication distances between a plurality of high-altitude communication platform nodes and server nodes according to the communication node information.

[0178] The high-altitude shortest communication path obtaining unit is configured to obtain high-altitude shortest communication paths by using a multi-element shortest path algorithm to calculate the communication distances.

[0179] The high-altitude communication network obtaining unit is configured to obtain a high-altitude communication network according to the high-altitude shortest communication paths.

[0180] The shortest communication path obtaining unit is configured to obtain shortest communication paths by using the high-altitude shortest communication paths to fill the ground shortest communication paths.

[0181] The communication network model obtaining unit is configured to obtain the communication network model according to the shortest communication paths and the computing resources.

[0182] In an implementation manner, the computing module 30 includes:

[0183] The offloading task obtaining unit is configured to obtain an offloading task in the communication network model, wherein the task includes a starting node and a target node.

[0184] The local offloading unit is configured to perform task offloading at a local communication node if the starting node and the target node are the same.

[0185] The full offloading unit is configured to obtain a communication path and computing resources of the offloading task in the communication network model if the starting node and the target node are different.

[0186] The computing unit is configured to obtain a first state parameter of the offloading task according to the communication path and the computing resource, input the first state parameter into a deep deterministic policy gradient algorithm for offloading calculation, and obtain the offloading decision and the computing resource allocation scheme; wherein the first state parameter comprises a sum of time delays between the starting node and the target node, a computing capability and a residual energy; the deep deterministic policy gradient algorithm comprises a main executor network, a main critic network, a target executor network and a target critic network; wherein the main executor network is configured to obtain a first offloading decision according to the first state parameter, and obtain a second state parameter according to the first offloading decision; the main critic network is configured to obtain a first evaluation value according to the first state parameter and the first offloading decision; the target executor network is configured to obtain a second offloading decision according to the second state parameter; the target critic network is configured to obtain a second evaluation value according to the second state parameter and the second offloading decision; and the deep deterministic policy gradient algorithm is configured to calculate a mean square error value according to the first evaluation value and the second evaluation value, and update network parameters of the main executor network and the main critic network according to the mean square error value.

[0187] Based on the above-mentioned embodiments, the application further provides an intelligent terminal, a principle block diagram of which can be shown in Figure 15 The intelligent terminal comprises a processor, a memory, a network interface, a display screen and a temperature sensor connected through a system bus. The processor of the intelligent terminal is configured to provide computing and control capabilities. The memory of the intelligent terminal comprises a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operating system and the computer program in the non-volatile storage medium to run. The network interface of the intelligent terminal is configured to communicate with external terminals through network connection. The computer program is executed by the processor to implement an edge computing task offloading method based on a high-altitude communication platform. The display screen of the intelligent terminal can be a liquid crystal display screen or an electronic ink display screen. The temperature sensor of the intelligent terminal is pre-set in the intelligent terminal and is configured to detect the running temperature of internal devices.

[0188] Those skilled in the art can understand that Figure 15 The principle block diagram shown in the above-mentioned embodiments is only a block diagram of part of the structure related to the application scheme and does not constitute a limitation on the intelligent terminal to which the application scheme is applied. The specific intelligent terminal can comprise more or fewer components than those shown in the diagram, or some components can be combined, or have a different component arrangement.

[0189] In one embodiment, an intelligent terminal is provided, which comprises a memory, a processor, and a high-altitude communication platform assisted edge computing task offloading program stored in the memory and executable on the processor. When the processor executes the high-altitude communication platform assisted edge computing task offloading program, the following operation instructions are implemented:

[0190] Obtaining computing resource and communication node information;

[0191] According to the computing resource and communication node information, a communication network model is constructed; wherein the communication network model comprises a ground communication network and a high-altitude communication network;

[0192] In the communication network model, offloading calculation is performed by using a deep deterministic policy gradient algorithm to obtain an offloading decision and a computing resource allocation scheme;

[0193] According to the offloading decision and the computing resource allocation scheme, edge computing task offloading is performed.

[0194] A person of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiments can be completed by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer readable storage medium. When the computer program is executed, it can include the processes of the above-mentioned embodiments. Any reference to memory, storage, operation database or other medium used in the embodiments provided by the present application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration but not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double operation data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0195] In summary, this invention discloses an edge computing task offloading method based on a high-altitude communication platform. The method includes: acquiring computing resource and communication node information; constructing a communication network model based on the computing resource and communication node information; wherein the communication network model includes a ground communication network and a high-altitude communication network; performing offloading calculations in the communication network model using a deep deterministic policy gradient algorithm to obtain offloading decisions and computing resource allocation schemes; and performing edge computing task offloading based on the offloading decisions and computing resource allocation schemes. This invention assists vehicles in completing computing tasks by rationally scheduling computing resources, and by allowing edge servers and parked vehicles to coexist, it ensures offloading efficiency and communication stability, significantly improving service quality.

[0196] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for edge computing task offloading based on high-altitude communication platform assistance, characterized in that, The method comprises: obtaining computing resource and communication node information; constructing a communication network model according to the computing resource and communication node information; wherein the communication network model comprises a ground communication network and an aerial communication network; performing offloading calculation in the communication network model by using a deep deterministic policy gradient algorithm to obtain an offloading decision and a computing resource allocation scheme; performing edge computing task offloading according to the offloading decision and the computing resource allocation scheme; performing offloading calculation in the communication network model by using a deep deterministic policy gradient algorithm to obtain an offloading decision and a computing resource allocation scheme, comprising: obtaining an offloading task in the communication network model; wherein the task comprises a starting node and a target node; if the starting node and the target node are the same, performing task offloading in a local communication node; if the starting node and the target node are different, obtaining a communication path and a computing resource of the offloading task in the communication network model; obtaining a first state parameter of the offloading task according to the communication path and the computing resource, and inputting the first state parameter into a deep deterministic policy gradient algorithm for offloading calculation to obtain the offloading decision and the computing resource allocation scheme; wherein the first state parameter comprises the sum of time delays, computing capacity and residual energy between the starting node and the target node; the deep deterministic policy gradient algorithm comprises a main executor network, a main critic network, a target executor network and a target critic network; wherein the main executor network is used to obtain a first offloading decision according to the first state parameter, and obtain a second state parameter according to the first offloading decision; the main critic network is used to obtain a first evaluation value according to the first state parameter and the first offloading decision; the target executor network is used to obtain a second offloading decision according to the second state parameter; and the target critic network is used to obtain a second evaluation value according to the second state parameter and the second offloading decision; the deep deterministic policy gradient algorithm is used to calculate a mean square error value according to the first evaluation value and the second evaluation value, and update network parameters of the main executor network and the main critic network according to the mean square error value.

2. The method of claim 1, wherein, the obtaining of the computing resource and communication node information comprises: obtaining location information, priority and communication capability of a communication node; wherein the communication node comprises a server node, a roadside unit node and an aerial communication platform node, and the priority of the server node and the roadside unit node is higher than that of the aerial communication platform node; the server node comprises a physical edge server node and a virtual edge server node; and the communication capability comprises computing capacity, residual energy and communication bandwidth; obtaining the communication node information according to the location information and the priority; obtaining the computing resource according to the communication capability. 3.The method of claim 1, wherein, the construction of the communication network model according to the computing resource and the communication node information comprises: obtaining communication distances between a plurality of roadside unit nodes and server nodes according to the communication node information; The communication distance between the high-altitude communication platform nodes and the server nodes is obtained according to the communication node information. The ground communication network is obtained according to the ground shortest communication path.

4. The method of claim 3, wherein, The communication network model is constructed according to the computing resource and communication node information, including: The communication distance between the high-altitude communication platform nodes and the server nodes is obtained according to the communication node information. The high-altitude shortest communication path is obtained by using a multi-element shortest path algorithm to calculate the communication distance. The high-altitude communication network is obtained according to the high-altitude shortest communication path. The shortest communication path is obtained by filling the ground shortest communication path with the high-altitude shortest communication path. The communication network model is obtained according to the shortest communication path and the computing resource.

5. An apparatus for edge computing task offloading based on high-altitude communication platform assistance, comprising: The device comprises: A computing resource and communication node information acquisition module for acquiring computing resource and communication node information; A modeling module for constructing a communication network model according to the computing resource and communication node information; wherein the communication network model comprises a ground communication network and a high-altitude communication network; A computing module for performing offloading calculation in the communication network model using a deep deterministic policy gradient algorithm to obtain an offloading decision and a computing resource allocation scheme; An offloading module for performing edge computing task offloading according to the offloading decision and the computing resource allocation scheme; In the communication network model, the deep deterministic policy gradient algorithm is used for offloading calculation to obtain an offloading decision and a computing resource allocation scheme, including: An offloading task in the communication network model is obtained; wherein the task includes a starting node and a target node; If the starting node and the target node are the same, task offloading is performed in the local communication node; If the starting node and the target node are different, the communication path and the computing resource of the offloading task in the communication network model are obtained; According to the communication path and the computing resource, the first state parameter of the offloading task is obtained, and the first state parameter is input into the deep deterministic policy gradient algorithm for offloading calculation to obtain the offloading decision and the computing resource allocation scheme; wherein the first state parameter includes the sum of the time delay, the computing capacity and the remaining energy between the starting node and the target node; The deep deterministic policy gradient algorithm comprises a main executor network, a main critic network, a target executor network and a target critic network; wherein the main executor network is used to obtain a first offloading decision according to the first state parameter, and a second state parameter is obtained according to the first offloading decision; the main critic network is used to obtain a first evaluation value according to the first state parameter and the first offloading decision; the target executor network is used to obtain a second offloading decision according to the second state parameter; and the target critic network is used to obtain a second evaluation value according to the second state parameter and the second offloading decision; The deep deterministic policy gradient algorithm is used to calculate the mean square error value according to the first evaluation value and the second evaluation value, and update the network parameters of the main executor network and the main critic network according to the mean square error value.

6. A smart terminal, characterized by The intelligent terminal comprises a memory, a processor, and a high-altitude platform assisted edge computing task offloading program stored in the memory and executable on the processor. When the processor executes the high-altitude platform assisted edge computing task offloading program, the steps of the high-altitude platform assisted edge computing task offloading method according to any one of claims 1-4 are implemented.

7. A computer readable storage medium characterized in that, The computer readable storage medium stores a high-altitude platform assisted edge computing task offloading program. When the processor executes the high-altitude platform assisted edge computing task offloading program, the steps of the high-altitude platform assisted edge computing task offloading method according to any one of claims 1-4 are implemented.

Images