Air-space-ground integrated network cooperative computing service optimization method based on adversarial learning

Abstract

The invention discloses an air-space-ground integrated network cooperative computing service optimization method based on adversarial learning, and belongs to the technical field of artificial intelligence and wireless communication in an air-space-ground integrated network. The method comprises the following steps: (1) in an edge mobile computing communication system in the sky-ground integrated network, proposing a plurality of allocation modes of computing tasks; (2) obtaining total time delay and energy consumption experienced by the wireless equipment in different distribution modes; (3) proposing an objective function and a constraint condition, and forming an optimization problem based on the objective function and the constraint condition; and (4) finding out an optimal result through a plurality of iterations. According to the invention, the total delay of the system is minimized through the space-air-ground integrated network cooperative computing service optimization method; the genetic algorithm driven by the generative adversarial network is provided, the nonlinear high-complexity problem can be solved more efficiently and accurately, the generative adversarial network can learn features and distribution in the data, the generated solution is closer to the actual situation, and the actual problem can be better solved.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence and wireless communication technology in an integrated sky-ground network, and in particular to an air-ground-ground integrated network collaborative computing service optimization method based on adversarial learning. Background Art

[0002] In the integrated sky-ground scenario, more and more terminals have certain computing capabilities, and the demand for terminal model computing is increasing. However, the computing power of a single terminal is limited after all. In order to optimize resource allocation in the integrated sky-ground network, the connection between various types of terminals and edge servers must be strengthened. For example, it is possible to consider collaborative computing between ground wireless devices and nearby drones to accelerate the completion of computing tasks.

[0003] However, in the existing integrated sky-ground network, the terminal computing tasks only consider a single terminal or two terminals for collaborative computing, while ignoring the rich computing capabilities of other nearby terminal devices and satellites, making resource allocation more difficult.

[0004] Aiming at the problem of insufficient edge computing resources in traditional air-ground-space integrated networks, the present invention proposes a resource allocation model for edge mobile computing communication systems, and obtains the optimal resource allocation strategy by generating a genetic algorithm driven by a confrontational network. Summary of the invention

[0005] Purpose of the invention: To provide an air-ground integrated network collaborative computing service optimization method based on adversarial learning to solve the problems of difficult resource allocation in the prior art.

[0006] The present invention provides an air-ground integrated network collaborative computing service optimization method based on adversarial learning, comprising the following steps:

[0007] S1. The edge mobile computing communication system in the sky-ground integrated network for the sky-ground integrated network collaborative computing service is composed of wireless devices, drones, edge servers and a LEO satellite, wherein the number of the wireless devices is J, and the set of the J wireless devices is defined as J = {1, 2, .., J}, the number of the drones is K, and the set of the K drones is defined as K = {1, 2, .., K}, the drone is abbreviated as UAV, the number of the edge servers is U, and the set of the U edge servers is defined as U = {1, 2, .., U}, the LEO satellite is defined as s, and for each of the wireless devices j∈J, there is a computing task M j , propose several allocation modes of the computing tasks, and based on the edge mobile computing communication system, obtain the unloading waiting delay and transmission energy consumption of the computing tasks under different allocation modes;

[0008] S2. Based on the edge mobile computing communication system, obtain the total delay and energy consumption experienced by the wireless device under different allocation modes;

[0009] S3, constructing a resource allocation model for the edge mobile computing communication system, proposing an objective function and constraints, and forming an optimization problem based on the objective function and the constraints;

[0010] The energy constraints of the wireless device and the drone are respectively as well as

[0011] The optimization objective is defined as:

[0012]

[0013] to minimize the total delay experienced by the wireless device while ensuring that the energy constraints of the wireless device j and the drone k are met;

[0014] Where w represents the binary variable of the association strategy, T j represents the total delay experienced by wireless device j;

[0015] The energy constraints of the wireless device j and the drone k include:

[0016]

[0017] in, is the energy consumption of task offloading from the wireless device j to the edge server u, is the energy consumption of task offloading from the wireless device j to UAV k, E k→k′ is the energy consumption of task offloading from UAV k to UAV k′, E k→s is the energy consumption of task offloading from UAV k to satellite s, is the drone’s hovering energy consumption; Perform computing task M for UAV k j Energy consumption when

[0018] S4, the S4 step includes the following contents:

[0019] S41. Use the GAN-GA algorithm to build a generative adversarial network to minimize the total system delay;

[0020] S42, based on the goal of minimizing the total system delay, defining an evaluation function for calculating performance indicators and calculating fitness;

[0021] S43, select according to fitness, use elite selection + roulette to select the next generation of parent solutions and the optimal solution of the population;

[0022] S44, initialize the first generation of parent solutions, start evolution, use the generator of the generative adversarial network to replace the traditional crossover and mutation operations in the genetic algorithm, and select the next generation of parent solutions;

[0023] S45. Find the best result after several iterations.

[0024] Preferably, the step S1 specifically comprises:

[0025] S11. Acquiring data of the wireless device, the drone, the satellite, and the edge server from the edge mobile computing communication system, specifically:

[0026] Each wireless device j∈J in the considered network has a computation task M j , use M j =α j +A j Indicates that, where α j is the CPU cycle required to process one bit of data, A j is the total input data size of the task;

[0027] S12, constructing a task offloading decision variable, indicating that the computing task of the wireless device is completed locally or in collaboration with other terminals and servers, specifically:

[0028] Defining binary decision variables Represents the computing task M j Compute locally only:

[0029]

[0030] Defining binary decision variables Represents the computing task M j Whether to offload to the edge server u:

[0031]

[0032] Defining binary decision variables Represents the computing task M j Whether to offload to the UAV k:

[0033]

[0034] Defining binary decision variables Represents the computing task M jWhether to unload from the UAV k to UAV k′, where the UAV k′ is the closest UAV to the UAV k:

[0035]

[0036] Defining binary decision variables Indicates that the UAV k offloads the task to the satellite s:

[0037]

[0038] S13, using the data of the wireless device, the drone, the satellite and the edge server to calculate the unloading waiting delay and transmission energy consumption of the computing tasks in different allocation modes in the edge mobile computing communication system;

[0039] The communication rate from the wireless device j to the edge server u is The transmission delay experienced by the wireless device j when offloading the computing task to the edge server u is The energy consumption of task offloading from the wireless device j to the edge server u is

[0040] Among them, P j is the uplink transmission power of the wireless device j;

[0041] The communication rate from the wireless device j to the UAV k is The transmission delay experienced by the wireless device j when offloading the computing task to the UAV k is The energy consumption of task offloading from the wireless device j to the UAV k is

[0042] The communication rate from UAV k to UAV k′ is R k→k′ , the transmission delay between the UAV k and the UAV k′ is The energy consumption of task offloading from UAV k to UAV k′ is Among them, P k→k′ is the uplink transmission power from the UAV k to the UAV k′, J k is the total set of wireless devices that perform collaborative computing between the wireless device j and the UAV k;

[0043] The communication rate from the UAV k to the satellite is R k→s , the transmission delay between the UAV k and the satellite is expressed as The energy consumption of task offloading from the UAV k to the satellite is Among them, P k→s is the uplink transmission power from the UAV k to the satellite.

[0044] Preferably, the step S2 is specifically as follows:

[0045] The wireless device Decided to perform its computation tasks locally, i.e. Then the delay experienced by the wireless device to complete the task is:

[0046]

[0047] where F j is the computing power of the wireless device j, i.e., cycle / s; the local energy consumption of the wireless device j is expressed as:

[0048]

[0049] in is a constant that depends on the chip architecture of the wireless device;

[0050] When , the time required for the edge server u to complete the computing task is:

[0051]

[0052] Among them, α j is the number of CPU cycles required to process one bit of data, is the computing resource of the edge server u allocated to the wireless device j, which is formulated as follows by using weighted proportional allocation:

[0053]

[0054] in is the maximum computing capacity at the edge server u, and the delay experienced by the wireless device j and the edge server u in cooperating to complete the computing task is:

[0055]

[0056] The task of the wireless device j is calculated at the UAV k, and the time required for the UAV k to complete the calculation task is:

[0057]

[0058] in is the computing power of UAV k allocated to wireless device j, formulated as follows using weighted proportional allocation:

[0059]

[0060] in is the maximum computational capability of the UAV k, so The total delay experienced by the wireless device j is given by

[0061]

[0062] When , the total delay experienced by the wireless device j is:

[0063]

[0064] The UAV k is used to perform the computing task T j The energy consumption formula is

[0065]

[0066] The computational delay when the computational task of the wireless device j associated with the UAV k is completed on the satellite, which has renewable energy and ignores the computational energy consumption of the satellite; the total execution delay experienced by the wireless device j when the computational task of the wireless device j is offloaded to the satellite is in is the propagation delay between the UAV k and the satellite.

[0067] Preferably, the S3 step is specifically as follows:

[0068] Indicates that the wireless device can only select one collaborative computing mode; represents a binary decision variable for the wireless device; wherein, is the energy consumption of task offloading from the wireless device j to the edge server u, is the energy consumption of task offloading from the wireless device j to the UAV k, E k→k′ is the energy consumption of task offloading from the UAV k to the UAV k′, E k→s is the energy consumption of mission offloading from the UAV k to the satellite s, is the hovering energy consumption of the UAV; Execute computing task M for the UAV k j energy consumption.

[0069] Preferably, the S4 step is specifically as follows:

[0070] The S41 step is specifically as follows:

[0071] GAN trains a generator and a discriminator, wherein the generator and the discriminator participate in a zero-sum game to generate real samples;

[0072] The generator function is represented as G(z), which is derived from the distribution P z The sampled random noise vector z is taken as input and mapped to the data space, i.e. z~P z (z);

[0073] The discriminator acts as a binary classifier; the function of the discriminator is represented as D(x), receiving the actual data x and the generated result G(z) as input; the discriminator strives to make the output D(G(z)) close to 0 and the output D(x) close to 1;

[0074] The objective function of the discriminator is formulated as:

[0075]

[0076] Wherein x~Pdata represents the distribution of real data. Through iterative training of the generator and the discriminator, when the discriminator cannot distinguish between real data and generated results, the generator successfully deceives the discriminator, and the generated results are closely consistent with the actual data distribution;

[0077] The S42 step is specifically as follows:

[0078] The evaluation function performs a series of complex calculations on the input sample features and generates an evaluation result;

[0079] The fitness function accepts the evaluation result of the evaluation function. Considering the given population j∈{1,2,..,J}, the result value of the evaluation function of the individual wireless device j is T j ; In the context of minimization problems, the fitness function is expressed as

[0080] O j =T max -T j +τ

[0081] Where τ is used to ensure O j >0, very small real number, T max Indicates the maximum tolerable delay of the task;

[0082] The S43 step is specifically as follows:

[0083] Sort the individuals according to the fitness function, select the top 20% of the individuals with the highest fitness as elite individuals, and for the remaining individuals, determine the selection probability according to the proportion of their fitness in the population;

[0084] Random selection with probability is performed from these individuals to fill the remaining selection slots;

[0085] Merge the elite individuals and the individuals selected by roulette, and return the selected next generation parent solution, the index of the optimal solution of the population, the fitness of the optimal solution of the population, and the fitness of all selected individuals;

[0086] The S44 step is specifically as follows:

[0087] Initialize the parent population;

[0088] Enter the iterative process, each iteration includes the following steps:

[0089] Generate offspring using GAN;

[0090] Merge the child with the parent;

[0091] Decode the merged population, that is, convert the floating point encoding into the actual parameter value;

[0092] Evaluate each individual's performance;

[0093] Calculate fitness;

[0094] Perform selection operations to select the parent population of the next generation and record the optimal solution and average solution of the current generation;

[0095] In each iteration, GAN accepts the current parent population as input and generates a new child population.

[0096] Compared with the prior art, the present invention adopts the above technical solution and has the following technical effects:

[0097] (1) This method minimizes the total system delay through an air-ground integrated network collaborative computing service optimization method;

[0098] (2) This method proposes a genetic algorithm driven by a generative adversarial network, which can solve nonlinear and highly complex problems more efficiently and accurately. The generative adversarial network can learn the characteristics and distribution of the data, making the generated solution closer to the actual situation and better able to solve practical problems. BRIEF DESCRIPTION OF THE DRAWINGS

[0099] Figure 1 This is a diagram of an integrated space-ground collaborative computing network architecture;

[0100] Figure 2 Flowchart of the optimization method for collaborative computing services in an integrated air-ground-space network based on adversarial learning. DETAILED DESCRIPTION

[0101] The technical solutions in the embodiments of the present invention are described clearly and completely below. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention. The experimental methods described in the following embodiments are all conventional methods unless otherwise specified.

[0102] According to a method for optimizing collaborative computing services in an air-ground integrated network based on adversarial learning provided by the present invention, the edge mobile computing communication system in the air-ground integrated network of the present invention is as follows: Figure 1 As shown in Figure 2, it is assumed that the edge mobile computing communication system in the sky-ground integrated network consists of wireless devices, drones, edge servers, and a LEO satellite. Its goal is to minimize the total system delay while meeting the energy constraints of wireless devices and drones.

[0103] like Figure 2 As shown, a flowchart of a method for optimizing air-ground integrated network collaborative computing services based on adversarial learning is provided in an embodiment of the present invention, and the method includes the following steps:

[0104] Step S1 specifically includes:

[0105] S11. Build an edge mobile computing communication system based on an integrated air-ground-space network, including wireless devices, drones, satellites, and edge servers, and obtain data from wireless devices, drones, satellites, and edge servers;

[0106] S12, constructing a task offloading decision variable, used to indicate that the computing task of the wireless device is completed locally or in collaboration with other terminals and servers;

[0107] S13, based on the task offloading decision variables, the data of the wireless device, the drone, the satellite and the edge server, calculate the offloading waiting delay and the transmission energy consumption of the computing task under different allocation modes in the edge mobile computing communication system;

[0108] In a further embodiment, an edge mobile computing communication system in a sky-ground integrated network is composed of wireless devices, drones, edge servers, and a LEO satellite (s), wherein the set of wireless devices j is defined as J = {1, 2, .., J}, the set of drones k is defined as K = {1, 2, .., K}, the set of edge servers u is defined as U = {1, 2, .., U}, and each wireless device j∈J in the considered network has a computing task M j , use M j =α j +A j Indicates that, where α jis the CPU cycle required to process one bit of data, A j is the total input data size of the task.

[0109] Defining binary decision variables Represents the computing task M j Compute locally only:

[0110]

[0111] The wireless device first considers offloading the computing task to the nearest edge server u∈U, defining Offloading decision variables for tasks

[0112]

[0113] Assume that the distance between wireless device j and edge server u is d j u , the path loss model is expressed as:

[0114]

[0115] Among them, P j is the transmission power at wireless device j; P u (d j u ) is the distance d j u The received power of the edge server u at location; m is the path loss index; G j and G u are the antenna gains of wireless device j and edge server u respectively; J u It is the total set of wireless devices that collaborate with the edge server u for computing.

[0116] The instantaneous data rate of wireless device j is calculated as:

[0117]

[0118] where σ 2 is the noise power, B u is the available bandwidth between wireless device j and edge server u, is the channel gain between wireless device j and edge server u, is the interference at the edge server u.

[0119] The instantaneous data rate achieved by wireless device j associated with edge server u is:

[0120]

[0121] Among them, |J u| is the total number of wireless devices that offload their computational tasks to edge server u. According to the instantaneous data rate, the transmission delay experienced by wireless device j when offloading its computational tasks to edge server u is

[0122]

[0123] The energy consumption of task offloading from wireless device j to edge server u is

[0124]

[0125] definition is the task offloading decision variable, which indicates whether the task of wireless device j is offloaded to UAV k by using the wireless link:

[0126]

[0127] Define [x j ,y j ] T and are the horizontal coordinates of ground wireless device j and UAV k, h k represents the hovering height of UAV k, and the distance between wireless device j and UAV k is

[0128]

[0129] Among them, J k It is the total set of wireless devices that perform collaborative computing between wireless device j and UAV k.

[0130] Then, by adopting the free space path loss model, the channel gain between wireless device j and UAV k is:

[0131]

[0132] Where g0 represents the channel gain at the reference distance d0 = 1m, and θ is the path loss exponent. Therefore, according to the Shannon formula, the rate of wireless device j is calculated as:

[0133]

[0134] Among them, P j is the uplink transmission power of wireless device j, is the achievable channel gain between wireless device j and UAV k, σ 2 is the noise power, is the interference at the UAV k.

[0135] The instantaneous data rate achieved by wireless device j associated with UAV k is calculated as:

[0136]

[0137] Among them, |J k | is the total number of wireless devices that offload their computational tasks to UAV k, B k is the available bandwidth between wireless device j and UAV k. According to the instantaneous data rate, the transmission delay experienced by wireless device j when offloading the computation task to UAV k is

[0138]

[0139] The energy consumption of task offloading from wireless device j to UAV k is

[0140]

[0141] After receiving the offloaded computational task from its associated wireless device, UAV k decides to process the task locally on its server or offload it to a neighboring UAV or satellite, depending on its available computational power. Define the binary decision variable Represents the computing task M j Whether to unload to UAV k′, where UAV k′ is the closest UAV to UAV k:

[0142]

[0143] Assuming there is a line-of-sight communication link between the UAVs, the achievable channel gain between UAVs k and k′ is

[0144]

[0145] in is the path loss between UAV k and k′, where is the attenuation factor of the LoS link, θ k→k′ for:

[0146]

[0147] Among them, f c is the carrier frequency, c is the speed of light, is the distance between UAVs k and k′. Finally, the achievable data rate between UAVs k and k′ is calculated as:

[0148]

[0149] Among them B k→k′ is the available bandwidth between UAV k and UAV k′.

[0150] The transmission delay between UAV k and UAV k′ is:

[0151]

[0152] The energy consumption of task offloading from UAV k to UAV k′ is:

[0153]

[0154] Defining binary decision variables Indicates that drone k offloads the task to satellite s:

[0155]

[0156] The achievable signal-to-noise ratio between UAV k and satellite s is expressed as:

[0157]

[0158] P k→s is the transmission power of UAV k, and is the antenna gain of the transmitter and receiver, L r is the attenuation factor, t n is the noise temperature, a is the Boltzmann constant, is the millimeter wave carrier frequency, reflects the distance between UAV k and the satellite. The available mmWave backhaul link capacity between UAV k and the satellite is expressed as:

[0159]

[0160] in is the millimeter wave bandwidth between UAV k and the satellite. Therefore, the transmission delay between UAV k and the satellite is expressed as:

[0161]

[0162] The energy consumption of task offloading from UAV k to the satellite is:

[0163]

[0164] Step S2 specifically includes:

[0165] Based on the task offloading decision variables, data from wireless devices, drones, satellites, and edge servers, the total delay and energy consumption experienced by wireless devices in different allocation modes in the edge mobile computing communication system are calculated;

[0166] In a further embodiment, first, if the wireless device Decided to perform its computation tasks locally (i.e. ), then the delay experienced by the wireless device to complete the task is:

[0167]

[0168] where F j is the computational capacity of wireless device j (i.e., cycles / s). Then, the local energy consumption of wireless device j is expressed as:

[0169]

[0170] in is a constant that depends on the chip architecture of the wireless device.

[0171] When , the time required for edge server u to complete the computing task is:

[0172]

[0173] Among them, α j is the number of CPU cycles required to process one bit of data, is the computing resource of edge server u allocated to wireless device j, which is formulated as follows using weighted proportional allocation:

[0174]

[0175] in is the maximum computing capacity at edge server u. The delay experienced by wireless device j and edge server u in completing the computing task is:

[0176]

[0177] Secondly, if the task of wireless device j is calculated at UAV k, the time required for UAV k to complete the calculation task is:

[0178]

[0179] in is the computing power of UAV k allocated to wireless device j, which is formulated as follows using a weighted proportional allocation:

[0180]

[0181] in is the maximum computational capacity at UAV k, so The total delay experienced by wireless device j is given by

[0182]

[0183] Same reason The total delay experienced by wireless device j is given by

[0184]

[0185] UAV k is used to perform computing task T j The energy consumption formula is

[0186]

[0187] Finally, the computational delay when the computational task of wireless device j associated with UAV k is completed on the satellite. Assuming that the satellite has renewable energy, the computational energy consumption of the satellite is ignored. Then, when the computational task of wireless device j is offloaded to the satellite, the total execution delay experienced by wireless device j is

[0188]

[0189] in is the propagation delay between UAV k and the satellite.

[0190] Step S3 is specifically as follows:

[0191] Based on the goal of minimizing the total delay experienced by wireless devices, a resource allocation model for edge mobile computing communication systems is constructed, and the objective function is proposed to minimize the total delay of the system while ensuring that the energy constraints of wireless devices and drones are met.

[0192] In a further embodiment, in order to minimize the total delay experienced by the wireless device while ensuring that the wireless device and drones The energy constraint is thus defined as:

[0193]

[0194] Where w represents the binary variable of the association strategy, T j Indicates the total system delay, specifically:

[0195]

[0196] The constraints include:

[0197] Represents the energy constraints of both the wireless device and the drone.

[0198] Indicates that the wireless device can only select one collaborative computing mode. represents the binary decision variable of the wireless device. is the energy consumption of task offloading from wireless device j to edge server u, is the energy consumption of task offloading from wireless device j to UAV k, E k→k′ is the energy consumption of task offloading from UAV k to UAV k′, E k→s is the energy consumption of task offloading from UAV k to satellite s, is the drone’s hovering energy consumption; Perform computing task M for UAV k j energy consumption.

[0199] Step S4 is specifically as follows:

[0200] S41. Based on the goal of minimizing the total system delay, the GAN-GA algorithm is used to construct a generative adversarial network.

[0201] S42, based on the goal of minimizing the total system delay, defining an evaluation function for calculating performance indicators and calculating a fitness function;

[0202] S43. Based on the goal of minimizing the total system delay and the fitness function, elite selection and roulette are used to select the next generation parent solution and the population optimal solution.

[0203] S44, initialize the first generation of parent solutions, start evolution, use the generator of the generative adversarial network to replace the traditional crossover and mutation operations in the genetic algorithm, and select the next generation of parent solutions;

[0204] S45 and step S44 are repeated until the desired optimal solution of the system tends to be stable.

[0205] Step S41 is specifically as follows:

[0206] GANs require training of two basic components, the generator and the discriminator. These components participate in a zero-sum game where the goal is to generate realistic samples.

[0207] Generator: The main goal of the generator is to generate synthetic samples that are very similar to real data. Its function is represented by G(z). It will be drawn from the distribution P z The sampled random noise vector z is taken as input and mapped to the data space, i.e. z~P z (z). The purpose is to ensure that G(z) is very similar to the real data.

[0208] Discriminator: The discriminator is a binary classifier whose function is represented as D(x). It receives the actual data x and the generated result G(z) as input. The main goal is to distinguish the samples generated by the generator from the real data. It strives to make the output D(G(z)) close to 0 and the output D(x) close to 1.

[0209] The objective function of the discriminator is formulated as

[0210]

[0211] Where x~Pdata represents the distribution of real data. Through iterative training of the generator and the discriminator, when the discriminator cannot distinguish between the real data and the generated results, it means that the generator has successfully deceived the discriminator. This shows that the generated results are closely consistent with the actual data distribution.

[0212] According to one aspect of the present application, step S42 is specifically:

[0213] The evaluation function performs a series of complex calculations on the input sample features and generates an evaluation result.

[0214] The fitness function accepts the evaluation result of the evaluation function. Considering the given population j∈{1,2,..,J}, the result value of the evaluation function of individual j is T j In the context of minimization problems, the fitness function is expressed as

[0215] O j =T max -T j +τ

[0216] Where τ is used to ensure O j >0, very small real number, T max Indicates the maximum tolerable delay of a task.

[0217] Step S43 is specifically as follows:

[0218] First, the individuals are sorted according to the fitness function, and the top 20% of the individuals with the highest fitness are selected as elite individuals. For the remaining individuals, the selection probability is determined according to the proportion of their fitness in the population. Then, random selection is performed from these individuals with probability to fill the remaining selection vacancies. Finally, the elite individuals and the individuals selected by roulette are merged, and the next generation parent solution, the index of the optimal solution of the population, the fitness of the optimal solution of the population, and the fitness of all selected individuals are returned.

[0219] Step S44 is specifically as follows:

[0220] Initialize the parent population.

[0221] Enter the iterative process, each iteration includes the following steps:

[0222] Generate offspring using GAN.

[0223] Merges the child with the parent.

[0224] Decode the merged population, that is, convert the floating point encoding into the actual parameter value.

[0225] Evaluate the performance of each individual.

[0226] Calculate fitness.

[0227] Perform a selection operation to select the parent population for the next generation and record the optimal solution and average solution of the current generation.

[0228] In each iteration, GAN accepts the current parent population as input and generates a new child population.

[0229] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art can still modify the technical solutions described in the aforementioned embodiments or replace some of the technical features therein by equivalents. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for optimizing space-ground integrated network collaborative computing services based on adversarial learning, characterized in that: The following steps are involved: S1. The edge mobile computing communication system in the sky-ground integrated network for the sky-ground integrated network collaborative computing service is composed of wireless devices, drones, edge servers and a LEO satellite, wherein the number of the wireless devices is J, and the set of the J wireless devices is defined as J = {1, 2, .., J}, the number of the drones is K, and the set of the K drones is defined as K = {1, 2, .., K}, the drone is abbreviated as UAV, the number of the edge servers is U, and the set of the U edge servers is defined as U = {1, 2, .., U}, the LEO satellite is defined as s, and for each of the wireless devices j∈J, there is a computing task M j , propose several allocation modes of the computing tasks, and based on the edge mobile computing communication system, obtain the unloading waiting delay and transmission energy consumption of the computing tasks under different allocation modes; S2. Based on the edge mobile computing communication system, obtain the total delay and energy consumption experienced by the wireless device under different allocation modes; S3, constructing a resource allocation model for the edge mobile computing communication system, proposing an objective function and constraints, and forming an optimization problem based on the objective function and the constraints; The energy constraints of the wireless device and the drone are respectively as well as The optimization objective is defined as: to minimize the total delay experienced by the wireless device while ensuring that the energy constraints of the wireless device j and the drone k are met; Where w represents the binary variable of the association strategy, T j represents the total delay experienced by wireless device j; The energy constraints of the wireless device j and the drone k include: in, is the energy consumption of task offloading from the wireless device j to the edge server u, is the energy consumption of task offloading from the wireless device j to UAV k, E k→k′ is the energy consumption of task offloading from UAVk to UAVk′, E k→s is the energy consumption of the task offloading from UAVk to satellite s, is the drone’s hovering energy consumption; Perform computing task M for UAVk j Energy consumption when S4, the S4 step includes the following contents: S41. Use the GAN-GA algorithm to build a generative adversarial network to minimize the total system delay; S42, based on the goal of minimizing the total system delay, defining an evaluation function for calculating performance indicators and calculating fitness; S43, select according to fitness, use elite selection + roulette to select the next generation of parent solutions and the optimal solution of the population; S44, initialize the first generation of parent solutions, start evolution, use the generator of the generative adversarial network to replace the traditional crossover and mutation operations in the genetic algorithm, and select the next generation of parent solutions; S45. Find the best result after several iterations.

2. The method for optimizing space-ground integrated network collaborative computing services based on adversarial learning according to claim 1, characterized in that: The S1 step is specifically as follows: S11. Acquiring data of the wireless device, the drone, the satellite, and the edge server from the edge mobile computing communication system, specifically: Each wireless device j∈J in the considered network has a computation task M j , use M j =α j +A j Indicates that, where α j is the CPU cycle required to process one bit of data, A j is the total input data size of the task; S12, constructing a task offloading decision variable, indicating that the computing task of the wireless device is completed locally or in collaboration with other terminals and servers, specifically: Defining binary decision variables Represents the computing task M j Compute locally only: Defining binary decision variables Represents the computing task M j Whether to offload to the edge server u: Defining binary decision variables Represents the computing task M j Whether to offload to the UAVk: Defining binary decision variables Represents the computing task M j Whether to unload from the UAVk to UAVk′, where the UAVk′ is the closest UAV to the UAVk: Defining binary decision variables Indicates that the UAVk offloads the task to the satellite s: S13, using the data of the wireless device, the drone, the satellite and the edge server to calculate the unloading waiting delay and transmission energy consumption of the computing tasks in different allocation modes in the edge mobile computing communication system; The communication rate from the wireless device j to the edge server u is The transmission delay experienced by the wireless device j when offloading the computing task to the edge server u is The energy consumption of task offloading from the wireless device j to the edge server u is Among them, P j is the uplink transmission power of the wireless device j; The communication rate from the wireless device j to the UAV k is The transmission delay experienced by the wireless device j when offloading the computing task to the UAV k is The energy consumption of task offloading from the wireless device j to the UAV k is The communication rate from UAVk to UAVk′ is R k→k′ , the transmission delay between the UAVk and the UAVk′ is The energy consumption of task offloading from UAV k to UAV k′ is Among them, P k→k′ is the uplink transmission power from the UAVk to the UAVk′, J k is the total set of wireless devices that perform collaborative computing between the wireless device j and the UAV k; The communication rate from the UAVk to the satellite is R k→s , the transmission delay between the UAVk and the satellite is expressed as The energy consumption of task offloading from the UAV k to the satellite is Among them, P k→s is the uplink transmission power from the UAVk to the satellite.

3. The method for optimizing space-ground integrated network collaborative computing services based on adversarial learning according to claim 1, characterized in that: The S2 step is specifically as follows: The wireless device Decided to perform its computation tasks locally, i.e. Then the delay experienced by the wireless device to complete the task is: where F j is the computing power of the wireless device j, i.e., cycle / s; the local energy consumption of the wireless device j is expressed as: in is a constant that depends on the chip architecture of the wireless device; When , the time required for the edge server u to complete the computing task is: Among them, α j is the number of CPU cycles required to process one bit of data, is the computing resource of the edge server u allocated to the wireless device j, which is formulated as follows by using weighted proportional allocation: in is the maximum computing capacity at the edge server u, and the delay experienced by the wireless device j and the edge server u in cooperating to complete the computing task is: The task of the wireless device j is calculated at the UAVk, and the time required for the UAVk to complete the calculation task is: in is the computing capacity of the UAV k allocated to the wireless device j, which is formulated as follows using a weighted proportional allocation: in is the maximum computing capability of the UAVk, so The total delay experienced by the wireless device j is given by When , the total delay experienced by the wireless device j is: The UAV k is used to perform the computing task T j The energy consumption formula is The computational delay when the computational task of the wireless device j associated with the UAV k is completed on the satellite, the satellite has renewable energy, and the computational energy consumption of the satellite is ignored; when the computational task of the wireless device j is offloaded to the satellite, the total execution delay experienced by the wireless device j is and k∈K, where is the propagation delay between the UAV k and the satellite.

4. The method for optimizing space-ground integrated network collaborative computing services based on adversarial learning according to claim 1, characterized in that: The S3 step is specifically as follows: Indicates that the wireless device can only select one collaborative computing mode; u∈U represents the binary decision variable of the wireless device; wherein, is the energy consumption of task offloading from the wireless device j to the edge server u, is the energy consumption of task offloading from the wireless device j to the UAV k, E k→k′ is the energy consumption of task offloading from the UAVk to the UAVk′, E k→s is the energy consumption of the mission offloading from the UAVk to the satellite s, is the hovering energy consumption of the UAV; Execute computing task M for the UAVk j energy consumption.

5. The method for optimizing space-ground integrated network collaborative computing services based on adversarial learning according to claim 1, characterized in that: The S4 step is specifically as follows: The S41 step is specifically as follows: GAN trains a generator and a discriminator, wherein the generator and the discriminator participate in a zero-sum game to generate real samples; The generator function is represented as G(z), which is derived from the distribution P z The sampled random noise vector z is taken as input and mapped to the data space, i.e. z~P z (z); The discriminator acts as a binary classifier; the function of the discriminator is represented as D(x), receiving the actual data x and the generated result G(z) as input; the discriminator strives to make the output D(G(z)) close to 0 and the output D(x) close to 1; The objective function of the discriminator is formulated as: Wherein x~Pdata represents the distribution of real data. Through iterative training of the generator and the discriminator, when the discriminator cannot distinguish between real data and generated results, the generator successfully deceives the discriminator, and the generated results are closely consistent with the actual data distribution; The S42 step is specifically as follows: The evaluation function performs a series of complex calculations on the input sample features and generates an evaluation result; The fitness function accepts the evaluation result of the evaluation function. Considering the given population j∈{1,2,..,J}, the result value of the evaluation function of the individual wireless device j is T j ; In the context of minimization problems, the fitness function is expressed as O j =T max -T j +τ Where τ is used to ensure O j >0, very small real number, T max Indicates the maximum tolerable delay of the task; The S43 step is specifically as follows: Sort the individuals according to the fitness function, select the top 20% of the individuals with the highest fitness as elite individuals, and for the remaining individuals, determine the selection probability according to the proportion of their fitness in the population; Random selection with probability is performed from these individuals to fill the remaining selection slots; Merge the elite individuals and the individuals selected by roulette, and return the selected next generation parent solution, the index of the optimal solution of the population, the fitness of the optimal solution of the population, and the fitness of all selected individuals; The S44 step is specifically as follows: Initialize the parent population; Enter the iterative process, each iteration includes the following steps: Generate offspring using GAN; Merge the child with the parent; Decode the merged population, that is, convert the floating point encoding into the actual parameter value; Evaluate each individual's performance; Calculate fitness; Perform selection operations to select the parent population of the next generation and record the optimal solution and average solution of the current generation; In each iteration, GAN accepts the current parent population as input and generates a new child population.

Images