A space-air-ground integrated network reasoning efficiency optimization method based on submode optimization

Abstract

The application discloses a space-air-ground integrated network reasoning efficiency optimization method based on submode optimization and belongs to the fields of artificial intelligence and wireless communication technology in the space-air-ground integrated network.The application adopts: ①building an edge cooperative reasoning system in the space-air-ground integrated network, proposing several cooperative modes for reasoning tasks on vehicles, and obtaining reasoning time delays of reasoning tasks in different cooperative modes; ②constructing a resource allocation model and a terminal association model of the edge cooperative reasoning system, and obtaining total time delays experienced by vehicles in different allocation modes; ③constructing a resource allocation model of the edge cooperative reasoning system, proposing a target function and constraint conditions, and forming an optimization problem; and ④decomposing the optimization problem into two subproblems, alternately solving the two subproblems until an optimal reasoning efficiency optimization strategy is obtained.The application realizes efficient resource allocation with minimized total time delay of the system, has lower reasoning time delay and higher resource utilization efficiency.

Description

TECHNICAL FIELD

[0001] The application relates to the field of artificial intelligence and wireless communication technology in a sky-ground integrated network, and in particular to a sky-ground integrated network reasoning efficiency optimization method based on submodular optimization. BACKGROUND

[0002] The fifth generation (5G) mobile communication network aims to support enhanced mobile broadband (eMBB), massive machine type communication (mMTC) and ultra-reliable low latency communication (URLLC). With the commercialization of 5G technology, the research community has turned its attention to the sixth generation (6G) network, which is expected to provide ubiquitous coverage and ultra-wideband access anytime, anywhere, which is crucial for future Internet of Things (IoT), remote coverage, emergency communication and ecological remote sensing applications. However, the current 5G network architecture is not sufficient to support these advanced services, especially the demand for ubiquitous ultra-wideband access, artificial intelligence services and numerous Internet of Things applications. To meet this challenge, the sky-ground integrated network (SAGIN) as a promising architecture can meet the needs of these services and applications. At the same time, the rapid development of artificial intelligence (AI) technology has led to an increase in model inference demand. AI, with its ability to simulate and extend human intelligence, has become a driving force for new technologies and industrial revolutions. In the context of 6G networks, AI applications are ready to become ubiquitous, and as AI technology develops, the demand for model inference continues to grow. Although SAGIN has great potential, research on collaborative inference within the network is still in its infancy. Resource allocation in SAGIN is crucial for efficient completion of inference tasks, which are characterized by the need for high-quality completion within limited resources and time. SUMMARY

[0003] The application provides a sky-ground integrated network reasoning efficiency optimization method based on submodular optimization, which solves the problem that the existing technology cannot complete tasks with high quality within limited resources and time.

[0004] The application provides a sky-ground integrated network reasoning efficiency optimization method based on submodular optimization, which includes the following steps:

[0005] S1, an edge cooperative inference system for inferring the sky-ground integrated network is built, wherein the sky-ground integrated network includes vehicles, unmanned aerial vehicles, satellites and base stations, the edge cooperative inference system is built based on the vehicles, unmanned aerial vehicles, satellites and base stations in the sky-ground integrated network, a plurality of cooperation modes are proposed for inference tasks on the vehicles, and inference task inference time delays under different cooperation modes are obtained;

[0006] S2, based on the edge cooperative reasoning system, constructing a resource allocation model of the edge cooperative reasoning system and a terminal association model, and obtaining the total time delay experienced by the vehicle under different allocation modes;

[0007] S3, constructing a resource allocation model of the edge cooperative reasoning system, and proposing a target function and a constraint condition, wherein the target function is defined as:

[0008]

[0009] wherein W represents a task offloading decision variable, I represents a model partitioning decision variable, Z represents a resource allocation decision variable, T j represents the total time delay experienced by vehicle j;

[0010] The constraint condition is the limitation of computing resources and the limitation of the number of vehicles that can be associated with the unmanned aerial vehicle, the satellite and the base station;

[0011] S4, the optimization problem is decomposed into sub-problems P1 and P2, i.e. the joint terminal association and resource allocation problem and the model partitioning problem, the joint terminal association and resource allocation problem proposes a low complexity association strategy based on submodular optimization, and the model partitioning problem determines the best split point by exhaustive search, and the joint terminal association and resource allocation problem and the model partitioning problem are solved alternately until the optimal reasoning efficiency optimization strategy is obtained.

[0012] Preferably, the S1 step specifically comprises:

[0013] S11, in the edge cooperative reasoning system, including the vehicle, the unmanned aerial vehicle, the satellite, the base station and the reasoning task on the vehicle, obtaining the data of the vehicle, the unmanned aerial vehicle, the satellite and the base station and the neural network parameters;

[0014] The S11 step specifically comprises: there are J vehicles in the edge cooperative reasoning system, the index of the vehicle is defined as J∈{1, 2, …, J}, the base station includes U base stations, the index of the base station is defined as U∈{1, 2, …, U}, there are K unmanned aerial vehicles in the edge cooperative reasoning system, the index of the unmanned aerial vehicle is defined as K∈{1, 2, …, K}, and a satellite S; each vehicle in the network in the edge cooperative reasoning system has a reasoning task A jTo be executed, ground vehicles through wireless channel with communication range of base station, unmanned aerial vehicle or satellite to cooperate reasoning; the base station allocates N computing units to execute the task of the ground vehicle, the index of its computing unit is N∈{1,2,…,N};The unmanned aerial vehicle allocates L computing units to execute the task of the ground vehicle, the index of its computing unit is L∈{1,2,…,L};The satellite allocates M computing units to execute the task of the ground vehicle, the index of its computing unit is M∈{1,2,…,M};The index of the neural network layer in the reasoning task is i∈{0,1,2,…,I}, I is the total number of neural network layers;

[0015] S12, build task offloading decision variable, the task offloading decision variable indicates that the reasoning task of the vehicle is completed locally or cooperates with other terminals and servers to reason;

[0016] The S12 step is specifically:

[0017] The vehicle j∈J decides to execute its reasoning task locally, and defines As a task offloading decision; The vehicle j completes its reasoning task A j , Otherwise Considering that the reasoning performance of the vehicle itself is insufficient, the vehicle j∈J cooperates with the base station u∈U to complete the reasoning task;

[0018] Define As a task offloading decision, the vehicle j and the base station u cooperate to complete the reasoning task, then Otherwise If Define the model split point between the vehicle j and the base station u as

[0019] The vehicle j considers whether the reasoning task is allocated to UAV, which refers to unmanned aerial vehicle, and defines As a task offloading decision, the vehicle j and the UAV k cooperate to complete the reasoning task, then Otherwise Define the split point of the model between the vehicle j and the UAV k as

[0020] After receiving the offloaded cooperative reasoning task from the vehicle associated with the UAV k, the UAV k decides to process the task locally or offload the UAV k to the satellite according to the computing capacity available to the UAV k; Define As a terminal association strategy, the vehicle j completes the reasoning task by cooperating with the UAV k and the satellite s through the relay, then Otherwise Define the split point of the model at this time as

[0021] S13, constructing resource allocation decision variables, the resource allocation decision variables representing the unmanned aerial vehicles, the satellites, the computing units on the base stations being allocated to the reasoning tasks of the vehicles;

[0022] The S13 step specifically includes:

[0023] Define The decision variable for the computing unit, the computing unit n on the base station u being allocated to the vehicle j, is Otherwise

[0024] Define The decision variable for the computing unit, the computing unit l on the UAV k being allocated to the vehicle j, is Otherwise

[0025] Define The decision variable for the computing unit, the computing unit m on the satellite s being allocated to the vehicle j, is Otherwise

[0026] S14, the data of the vehicles, the unmanned aerial vehicles, the satellites, the base stations and the neural network parameters, based on the task offloading decision variables and the resource allocation decision variables, calculating the offloading waiting time of the reasoning tasks on the vehicle j in different allocation modes in the edge cooperative reasoning communication system;

[0027] The communication rate of the vehicle j to the base station u is The offloading transmission time of the vehicle j offloading part of the reasoning tasks to the base station u at the model split point is Wherein, is the output of the model layer;

[0028] The communication rate of the vehicle j to the UAV k is The transmission delay experienced by the vehicle j when offloading part of the reasoning tasks to the UAV k at the model split point is Wherein, is the output of the model layer.

[0029] Preferably, the S2 step specifically includes:

[0030] Based on the resource allocation decision variable, the task offloading decision variable, the vehicle, the UAV, the satellite, the base station data and the neural network parameters, the total delay experienced by the vehicle under different allocation modes in the edge cooperation inference system is calculated;

[0031] The vehicle j decides to execute its inference task locally, i.e. Then the total delay experienced by the vehicle j to complete the inference task is:

[0032]

[0033] Where ξ j is the computing power of vehicle j, α j is the number of CPU cycles required to process one bit of data;

[0034] When the base station u completes the inference task, the time required is:

[0035]

[0036] Where is the layer After the inference task A j The amount of data to be processed, the total delay experienced by the vehicle j and the base station u to complete the cooperative inference task is:

[0037]

[0038] Define The computing unit allocation decision variable, if the computing unit l on the UAV k is allocated to the vehicle j, then Otherwise

[0039] When the UAV k completes the inference task, the time required is:

[0040]

[0041] Where is the layer After the inference task A j The amount of data to be processed, the total delay experienced by the vehicle j, the UAV k to complete the cooperative inference task is:

[0042]

[0043] Define The computing unit allocation decision variable, if the computing unit m on the satellite s is allocated to the vehicle j, then Otherwise

[0044] The time required for the satellite s to complete the inference task is:

[0045]

[0046] When the inference task of the vehicle j is offloaded to the satellite, the total delay experienced by the vehicle j is:

[0047]

[0048] Preferably, the S3 step is specifically:

[0049] The objective function is specifically:

[0050]

[0051] The constraint conditions include:

[0052] Indicates the constraints of computing resources; Indicates that each vehicle can only choose one cooperative inference mode;

[0053] Indicates that each computing unit can only be assigned to at most one vehicle in an allocation period;

[0054] Indicates that the number of vehicles that the base station, the drone, and the satellite can associate with is limited;

[0055] Indicates the resource allocation strategy and terminal association strategy;

[0056] Indicates the model partition point;

[0057] Wherein, Indicates the upper limit of the number of vehicles accessed by the base station, the drone, and the satellite, N max , L max , M max Indicates the upper limit of the computing resources of the base station, the drone, and the satellite.

[0058] Preferably, the S4 step is specifically:

[0059] S41, to the problem The sub-problem P1 is given as:

[0060]

[0061] The constraint conditions include:

[0062] represents the constraint of computing resources; represents that each vehicle can only select one cooperative reasoning mode;

[0063] represents that each computing unit can only be allocated to at most one vehicle in one allocation period;

[0064] represents that the number of vehicles that can be associated with the base station, the unmanned aerial vehicle, and the satellite is limited;

[0065] represents the resource allocation strategy and the terminal association strategy;

[0066] wherein the objective function D j (w,z)=T0-T j aims to minimize the delay cost, and T0 is the task waiting time;

[0067] The sub-problem P2 is given as:

[0068]

[0069] The constraint conditions include:

[0070] represents the model segmentation point;

[0071] S42, for the joint terminal association and resource allocation problem, it is converted into a submodular problem for solving;

[0072] The sub-problem P1 is converted into a submodular problem, and the computing resource allocation and terminal association strategy are defined as

[0073]

[0074] wherein represents the action of allocating the computing resource l on the unmanned aerial vehicle k to the vehicle j, represents the action of allocating the computing resource n on the base station u to the vehicle j, represents the action of associating the base station u with the vehicle j, represents the action of associating the unmanned aerial vehicle k with the vehicle j; the basic set A can be divided into V disjoint sets (V'=UV), that is, for any v≠v', wherein the basic set B can be divided into disjoint sets the basic set C can be divided into U disjoint sets, wherein the basic set D can be divided into K disjoint sets;

[0075] Define a pair E = (F, G), where F = {A, B, C, D} is a global set, A set of independent subsets of F is used to replace the original constraints, defined as

[0076]

[0077] Where X is the set of final solutions; prove that the set family E = (F, G) is a pseudo-matrix, and the delayed reduction function Is a monotone submodular function on X∈G; on this basis, by replacing the constraints in the problem with set pairs, the subproblem P1 is expressed as a matroid-constrained monotone submodular maximization problem:

[0078]

[0079] stX∈G

[0080] initialization is an empty set, where v=1,...,V′, u=1,...,U,k=1,...,K, and define X={A,B,C,D} and Y X = F; in each iteration, the new element with the highest marginal gain is added Add to Collection X X and from the candidate action set Y X Remove the candidate action set Y X is the set of all currently selectable strategies; In the case of , the computing units on other base stations and drones are assigned to the action of the vehicle j, and the actions of other base stations and drones associated with the vehicle j are removed from the candidate action set, that is: Y = YA v′ ,v′=1,...,V′

[0081] v′≠(u-1)V+1,(u-1)V+2,...,uV,

[0082]

[0083] and

[0084]

[0085] Repeat this iteration until the candidate action set Y X Empty or marginal gain Δ D (a|X) is zero; wherein the marginal gain refers to: in a submodular function, when a new element is added to a set, the contribution of the element to the increase in the function value; in the present invention, the marginal gain Δ D(a|X) = D(a∪{X})-D(X);

[0086] In considering the satellite, the resource accumulation-based redistribution method is checked by a loop whether each vehicle can be connected to the satellite through the UAV, and the specific steps are as follows:

[0087] Initialize X, Y x=F and the global optimal allocation strategy; when the conditions of and are met, enter the loop to calculate the inference delay; if a suitable satellite access strategy is found and the inference delay can be reduced, update the global optimal allocation strategy; after updating, allocate the corresponding computing resources, and mark the allocation as used; if the resource allocation of the satellite is better than other options, select to connect with the satellite through the UAV, and update the computing capacity and delay of the vehicle;

[0088] S43, model segmentation algorithm to place the model,

[0089] The S43 step specifically includes:

[0090] The optimal split point is determined by traversal search, and the model segmentation algorithm is proposed as follows in the case of vehicle and base station cooperation inference:

[0091] Input: total number of DNN model layers I, bandwidth B, computing capacity of base station u Computing capacity of vehicle j ξ j , output data volume of each layer Optimal segmentation point i best ;

[0092] 5. Initialize the minimum time of vehicle j best as local computing time

[0093] 6. Assume the initial optimal segmentation layer i best =I;

[0094] 7. For each layer i∈{1,2,…,I} of the neural network in the inference task, perform the following operations:

[0095] Calculate the transmission time of the current layer Processing time on base station u and processing time on vehicle j

[0096] If the total time of executing the inference task , update the minimum time as and set the current layer i as i best ;

[0097] 7. After the loop ends, return the optimal segmentation point ibest ;

[0098] S44. Solve the two sub-problems alternately until the optimal strategy is obtained.

[0099] Compared with the prior art, the present invention adopts the above technical solution and has the following technical effects:

[0100] (1) This method minimizes the total system latency by optimizing the inference efficiency in an integrated space-ground network;

[0101] (2) This method is based on low-complexity submodular optimization and has lower inference latency and higher resource utilization efficiency compared to traditional methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0102] Figure 1 This is a diagram of an integrated air-space-ground collaborative reasoning network architecture.

[0103] Figure 2 This is a flowchart of the method for optimizing the inference efficiency of the space-ground integrated network based on submodule optimization.

[0104] Figure 3 The embodiment of the present invention provides a system minimum delay after optimizing the inference efficiency of an air-space-ground integrated network based on submodule optimization. DETAILED DESCRIPTION

[0105] The technical solutions in the embodiments of the present invention will be described clearly and completely below. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention. The experimental methods described in the following examples are all conventional methods unless otherwise specified.

[0106] According to a method for optimizing the inference efficiency of an air-ground integrated network based on submodule optimization provided by the present invention, the edge cooperative inference 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 mobile edge computing communication system in the integrated sky-ground network consists of ground vehicles, drones, base stations, and a LEO satellite. The goal is to minimize the total system delay while meeting the number of associated vehicles and drones. Figure 3 This embodiment provides the minimum system delay after optimizing the inference efficiency of the integrated air-space-ground network based on sub-module optimization.

[0107] like Figure 2 As shown, an embodiment of the present invention provides a flowchart of a method for optimizing the efficiency of an air-ground integrated network reasoning based on submodule optimization, the method comprising the following steps:

[0108] Step S1 specifically comprises:

[0109] S11, constructing an edge cooperative reasoning system based on space-ground-air integrated network, including vehicles, unmanned aerial vehicles, satellites, base stations and reasoning tasks on vehicles, obtaining data of vehicles, unmanned aerial vehicles, satellites and base stations and neural network parameters;

[0110] S12, constructing a task offloading decision variable for representing that the reasoning task of the vehicle is completed locally or cooperatively reasoned with other terminals and servers;

[0111] S13, constructing a resource allocation decision variable for representing that the computing unit on the unmanned aerial vehicle, satellite or base station is allocated to the reasoning task of the vehicle;

[0112] S14, based on the task offloading decision variable and the resource allocation decision variable, the data of the vehicle, the unmanned aerial vehicle, the satellite and the base station and the neural network parameters, calculating the offloading waiting time of the reasoning task in different allocation modes in the edge cooperative reasoning communication system;

[0113] In further embodiments, considering the cooperative reasoning problem in the space-ground-air integrated network, there are J vehicles in the system, the index of the vehicle is defined as J∈{1,2,…,J}, U base stations, the index of the base station is defined as U∈{1,2,…,U}, K unmanned aerial vehicles, the index of the unmanned aerial vehicle is defined as K∈{1,2,…,K}, and a satellite S. Each vehicle in the considered network has a reasoning task A j To execute, the ground vehicle cooperates with the base station, unmanned aerial vehicle or satellite within the communication range through the wireless channel for reasoning. It is assumed that the base station allocates N computing units to execute the task of the ground vehicle, and the index of the computing unit is N∈{1,2,…,N}. It is assumed that the unmanned aerial vehicle allocates L computing units to execute the task of the ground vehicle, and the index of the computing unit is L∈{1,2,…,L}. The satellite allocates M computing units to execute the task of the ground vehicle, and the index of the computing unit is M∈{1,2,…,M}. The index of the neural network layer in the reasoning task is i∈{0,1,2,…,I}, and I is the total number of neural network layers. It is assumed that all computing units have the same computing capacity, i.e. ξ rounds of CPU per second, and the neural network is at most split into two parts for processing.

[0114] If the vehicle j∈J decides to execute its reasoning task locally, define as the terminal association strategy, if the vehicle j completes its reasoning task A j locally, then otherwise Considering that the reasoning performance of the vehicle itself is insufficient, the vehicle j∈J cooperates with the base station u∈U to complete the reasoning task. Define For the terminal association policy, if vehicle j and base station u cooperate to complete the inference task, then Otherwise If Define the model split point between vehicle j and base station u at this time as

[0115] Assume the distance between vehicle j and base station u is The path loss model can be expressed as:

[0116]

[0117] where P j is the transmit power at vehicle j; is the received power of base station u at a distance ; λ is the path loss exponent; G j and G u are the antenna gains at vehicle j and base station u, respectively.

[0118] The communication rate of vehicle j to base station u can be calculated as

[0119]

[0120] where σ 2 is the noise power, is the average interference power when vehicle j and base station u communicate.

[0121] Based on the communication rate, the offloading transmission time of vehicle j to offload part of the inference task to base station u at the model split point is where is the output of the m-th layer of the model.

[0122] Secondly, vehicle j considers whether the inference task is assigned to UAV, define For the terminal association policy, if vehicle j and UAV k cooperate to complete the inference task, then Otherwise If Define the model split point between vehicle j and UAV k as

[0123] Define [x j , y j ] T and as the horizontal coordinates of ground vehicle j and UAV k, and the distance between ground vehicle j and UAV k is

[0124]

[0125] ​Then, by employing the free-space path loss model, the channel gain between device j and UAV k is:

[0126]

[0127] where denotes the channel gain at a reference distance d0= 1 m, and 0 is the path loss exponent. The communication rate at vehicle j is computed as:

[0128]

[0129] where, P j is the uplink transmission power of j, is the achievable channel gain between vehicle j and UAV k, 0 2 is the noise power, is the average interference power when vehicle j communicates with UAV k. According to the communication rate, the transmission delay experienced by vehicle j when offloading part of the inference task to UAV k at the model partition point is where, is the output of the m-th layer of the model.

[0130] Upon receiving the offloaded cooperative inference task from the vehicle associated with UAV k, UAV k decides to process the task locally at UAV k or offload UAV k to a satellite according to the computing capability available at UAV k; define as the terminal association policy, vehicle j completes the inference task in cooperation with UAV k relaying and satellite s, then otherwise define the partition point of the model at this time as

[0131] The available mmWave backhaul link capacity between UAV k and satellite s can be expressed as:

[0132]

[0133] where is the mmWave bandwidth between UAV k and satellite, is the transmit power of UAV k, and are the antenna gains of the transmitter and receiver, L r is the attenuation factor, E s is the noise temperature, a is the Boltzmann constant, f c mm is the mmWave carrier frequency, reflects the distance between UAV k and satellite s. The transmission delay between UAV k and satellite is expressed as

[0134]

[0135] wherein, is the output of the model’s layer.

[0136] To solve the computing resource allocation problem, define as a decision variable for the allocation of computing units, if computing unit n at base station u is allocated to vehicle j, then otherwise

[0137] define as a decision variable for the allocation of computing units, if computing unit l at UAV k is allocated to vehicle j, then otherwise

[0138] define as a decision variable for the allocation of computing units, if computing unit m at satellite s is allocated to vehicle j, then otherwise

[0139] Step S2 specifically comprises:

[0140] Based on the task offloading decision variables and the resource allocation decision variables, and the data of the vehicles, UAVs, satellites and base stations, calculate the total delay experienced by the vehicles under different allocation modes in the edge cooperation inference system;

[0141] In further embodiments, first, if vehicle j decides to execute its inference task locally (i.e. ), then the total delay experienced by vehicle j in completing the inference task is:

[0142]

[0143] wherein ξ j is the computing capability of vehicle j, and a j is the number of CPU cycles required to process one bit of data.

[0144] When the inference task A j needs to process the amount of data

[0145]

[0146] wherein is the amount of data layer j of the inference task A

[0147]

[0148] Definition Assign decision variables for computing units, if computing unit l on UAV k is assigned to vehicle j, then Otherwise

[0149] The time taken by UAV k to complete the inference task is:

[0150]

[0151] Where is the layer Post-inference task A j The amount of data to be processed. The total delay experienced by vehicle j to complete the cooperative inference task is:

[0152]

[0153] Definition Assign decision variables for computing units, if computing unit m on satellite s is assigned to vehicle j, then Otherwise

[0154] The time taken by satellite s to complete the inference task is:

[0155]

[0156] Then, when the inference task of vehicle j is offloaded to the satellite, the total delay experienced by vehicle j is:

[0157]

[0158] Step S3 specifically comprises:

[0159] Based on the goal of minimizing the total delay experienced by the vehicle, a resource allocation model of the edge cooperative inference system is constructed, and the proposed objective function is the minimization of the total delay of the system while ensuring that the associated quantity constraints are met. The objective function is defined as:

[0160]

[0161] Where W represents the task offloading decision variable, I represents the model partitioning decision variable, Z represents the resource allocation decision variable, T j represents the total delay of the system, specifically:

[0162]

[0163] The constraint conditions include:

[0164] indicates the constraint of computing resources. indicates that each vehicle can only select one cooperative reasoning mode.

[0165] indicates that each computing unit can only be allocated to at most one vehicle in one allocation period.

[0166] indicates that the number of vehicles that can be associated by the base station, unmanned aerial vehicle, and satellite is limited.

[0167] indicates the resource allocation strategy and terminal association strategy.

[0168] Finally, indicates the model partition point.

[0169] Step S4 specifically includes:

[0170] S41, based on the target of minimizing the total system delay, the original problem is decomposed into two sub-problems, sub-problem one is the joint terminal association and resource allocation problem, and sub-problem two is the model partition problem;

[0171] S42, for sub-problem one, it is converted into a submodular problem for solving;

[0172] S43, for sub-problem two, a model partition algorithm is proposed to place the model;

[0173] S44, the two sub-problems are solved alternately until the optimal strategy is obtained;

[0174] According to one aspect of the present application, step S41 specifically includes:

[0175] To solve the problem of association strategy and resource allocation, the problem The sub-problem P1 is given as:

[0176]

[0177] The constraint conditions include:

[0178] indicates the constraint of computing resources. indicates that each vehicle can only select one cooperative reasoning mode.

[0179] indicates that each computing unit can only be allocated to at most one vehicle in one allocation period.

[0180] indicates that the number of vehicles that can be associated by the base station, unmanned aerial vehicle, and satellite is limited.

[0181] indicates a resource allocation strategy, a terminal association strategy.

[0182] where the objective function D j (w,z) = T0-T j aiming to minimize the delay cost, T0 is the task waiting time.

[0183] To solve the model placement problem, a sub-problem P2 is given:

[0184]

[0185] The constraint conditions include:

[0186] indicates a model partition point.

[0187] According to an aspect of the present application, step S42 is specifically:

[0188] In order to convert the sub-problem P1 into a submodular problem, the computing resource allocation and the terminal association strategy are defined as

[0189]

[0190] where indicates the action of allocating computing resource l on UAV k to vehicle j, indicates the action of allocating computing resource n on base station u to vehicle j, indicates the action of associating base station u with vehicle j, indicates the action of associating UAV k with vehicle j. The basis set A can be divided into V disjoint sets (V' = UV), that is, for any v≠v', where Similarly, the basis set B can be divided into disjoint sets The basis set C can be divided into U disjoint sets, where Similarly, the basis set D can be divided into K disjoint sets.

[0191] Define a pair E = (F, G), where F = {A, B, C, D} is a global set, is a set of independent subsets of F used to replace the original constraint condition, and is defined as

[0192]

[0193] where X is a set of final solutions. It is not difficult to prove that the set family E = (F, G) is a matroid, and the delay reduction function is a monotone submodular function on X∈G. On this basis, by replacing the constraints in the problem with set pairs, the subproblem P1 can be expressed as a matroid-constrained monotone submodular maximization problem:

[0194]

[0195] stX∈G

[0196] The embodiment of the present invention proposes a submodular optimization algorithm to solve this problem. First, initialize is an empty set, where v=1,...,V′, u=1,...,U,k=1,...,K, and define X={A,B,C,D} and Y X = F. In each iteration, the new element with the highest marginal gain is added Add to Collection X X and from the candidate action set Y X Remove the candidate action set Y X is the set of all currently selectable strategies. In the case of , the action of allocating computing units on other base stations and drones to vehicle j, as well as the action of associating other base stations and drones with vehicle j should be removed from the candidate action set, that is:

[0197]

[0198] v'≠(u-1)V+1,(u-1)V+2,...,uV,

[0199]

[0200] and

[0201]

[0202] Repeat this iteration until the candidate action set Y X Empty or marginal gain Δ D (a|X) is zero. Wherein the marginal gain refers to: in a submodular function, when a new element is added to a set, the contribution of the element to the increase in the function value. In the present invention, the marginal gain Δ D (a|X)=D(a∪{X})-D(X).

[0203] When considering satellites, the embodiment of the present invention further proposes a reallocation method based on resource accumulation.

[0204] This method goes through a loop to check for each vehicle whether it can connect to the satellite via the drone.

[0205] First, initialize X, Y X = F and the global optimal allocation strategy. When the condition of and is met, enter the loop to calculate the inference delay. If a suitable satellite access strategy is found and the inference delay can be reduced, update the global optimal allocation strategy. After updating, allocate the corresponding computing resources and mark the allocation as used. If the resource allocation of the satellite is more optimal than other options, select to connect with the satellite through the unmanned aerial vehicle and update the computing power and delay of the vehicle.

[0206] According to an aspect of the present application, step S43 is specifically:

[0207] Step S43 specifically includes:

[0208] For sub-problem two, the best split point is determined by exhaustive search. In the case where the vehicle cooperates with the base station for inference, the present application proposes a model segmentation algorithm as follows:

[0209] Input: total number of layers of DNN model I, bandwidth B, computing power of base station u Computing power of vehicle j ξ j Output data volume of each layer Optimal segmentation point i best .

[0210] 1. Initialize the minimum time of vehicle j as local computing time

[0211] 2. Assume the initial best segmentation layer i best = I.

[0212] 3. For each layer i of the neural network in the inference task, execute the following operations:

[0213] Calculate the transmission time of the current layer Processing time on base station u And processing time on vehicle j

[0214] If the total time of executing the inference task , update the minimum time as And set i best as the current layer i.

[0215] 4. After the loop ends, return the best segmentation point i best .

[0216] S44, solve the two sub-problems alternately until the optimal strategy is obtained;

[0217] It should be pointed out that the above description is only for the preferred embodiments of the present application and is not intended to limit the present application. Although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art will appreciate that modifications can be made to the technical solutions described in the foregoing embodiments, or some of the technical features thereof can be replaced equivalently, without departing from the spirit and principle of the present application. Any modifications, equivalent replacements, improvements, and the like made within the spirit and principle of the present application should be included in the protection scope of the present application.

Claims

1. A method for optimizing the inference efficiency of an air-ground integrated network based on submodular optimization, characterized in that: The following steps are involved: S1. Building an edge cooperative reasoning system for reasoning on the integrated air-space-ground network, wherein the integrated air-space-ground network includes vehicles, drones, satellites, and base stations. The edge cooperative reasoning system is built based on the vehicles, drones, satellites, and base stations in the integrated air-space-ground network. Several cooperative modes are proposed for reasoning tasks on the vehicles, and the inference latency of the reasoning tasks under different cooperative modes is obtained. S2. Based on the edge cooperative reasoning system, construct a resource allocation model and a terminal association model of the edge cooperative reasoning system, and obtain the total delay experienced by the vehicle under different allocation models; S3. Construct a resource allocation model for the edge cooperative reasoning system, and propose an objective function and constraints. Based on the objective function and the constraints, an optimization problem is formed, where the objective function is defined as: Where W represents the task offloading decision variable, I represents the model segmentation decision variable, Z represents the resource allocation decision variable, and T j represents the total delay experienced by vehicle j; The constraints are the limitation of computing resources and the limit of the number of vehicles that can be associated with drones, satellites and base stations; S4. Decompose the optimization problem into sub-problems P1 and P2, namely, the joint terminal association and resource allocation problem and the model segmentation problem. A low-complexity association strategy based on sub-module optimization is proposed for the joint terminal association and resource allocation problem. The model segmentation problem is solved by determining the optimal splitting point through traversal search. The joint terminal association and resource allocation problem and the model segmentation problem are solved alternately until the optimal inference efficiency optimization strategy is obtained.

2. The method for optimizing the efficiency of space-ground integrated network reasoning based on submodule optimization according to claim 1, characterized in that: The S1 step is specifically as follows: S11. In the edge cooperative reasoning system, including the vehicle, the drone, the satellite, the base station, and the reasoning task on the vehicle, the data of the vehicle, the drone, the satellite, the base station, and the neural network parameters are obtained; The step S11 is specifically as follows: there are J vehicles in the edge cooperative reasoning system, the index of the vehicle is defined as J∈{1,2,…,J}, the base stations include U base stations, the index of the base stations is defined as U∈{1,2,…,U}, there are K drones in the edge cooperative reasoning system, the index of the drone is defined as K∈{1,2,…,K}, and a satellite S; each vehicle in the network in the edge cooperative reasoning system has an inference task A j To perform, the ground vehicle conducts cooperative reasoning with a base station, drone or satellite within the communication range through a wireless channel; the base station allocates N computing units to perform the tasks of the ground vehicle, and the index of its computing units is N∈{1,2,…,N}; the drone allocates L computing units to perform the tasks of the ground vehicle, and the index of its computing units is L∈{1,2,…,L}; the satellite allocates M computing units to perform the tasks of the ground vehicle, and the index of its computing units is M∈{1,2,…,M}; the index of the neural network layer in the reasoning task is i∈{0,1,2,…,I}, where I is the total number of layers of the neural network; S12. Constructing a task offloading decision variable, wherein the task offloading decision variable indicates whether the vehicle's reasoning task is completed locally or in cooperative reasoning with other terminals and servers; The S12 step is specifically as follows: The vehicle j∈J decides to perform its reasoning task locally, defining It is a task offloading decision; the vehicle j completes its reasoning task A locally j ,but otherwise Considering the insufficient reasoning performance of the vehicle itself, the vehicle j∈J cooperates with the base station u∈U to complete the reasoning task; definition For task offloading decision, the vehicle j and the base station u cooperate to complete the reasoning task, then otherwise like The model segmentation point between the vehicle j and the base station u is defined as The vehicle j considers whether the reasoning task is assigned to the UAV, where the UAV refers to an unmanned aerial vehicle, and defines For task offloading decision, the vehicle j and the UAVk cooperate to complete the reasoning task, then otherwise The split point of the model between the vehicle j and the UAV k is defined as After receiving the offloaded cooperative reasoning task from the vehicle associated with the UAVk, the UAVk decides to process the task locally at the UAVk or offload the UAVk to the satellite based on the computing power available at the UAVk; defining For the terminal association strategy, the vehicle j completes the reasoning task through the UAV k relay and the satellite s. otherwise Define the split point of the model at this time as S13. Construct a resource allocation decision variable, where the resource allocation decision variable indicates that the computing units on the UAV, the satellite, and the base station are allocated to the reasoning task of the vehicle; The S13 step is specifically as follows: definition Assign decision variables to the computing units, the computing unit n on the base station u is assigned to the vehicle j, then otherwise definition Assign decision variables to the computation units, and the computation unit l on the UAVk is assigned to the vehicle j, then otherwise definition Assign decision variables to the computing units. The computing unit m on the satellite s is assigned to the vehicle j. Then otherwise S14, calculating the offloading waiting delay of the inference task on the vehicle j under different allocation modes in the edge cooperative inference communication system based on the data of the vehicle, the drone, the satellite, the base station, and the neural network parameters; The communication rate from vehicle j to base station u is At the model split point The unloading transmission time for vehicle j to offload part of the reasoning task to base station u is in, For model The output of the layer; The communication rate from vehicle j to UAV k is The vehicle j is at the model segmentation point The transmission delay experienced when offloading part of the reasoning task to the UAVk is in, For model The output of the layer.

3. The method for optimizing the inference efficiency of an integrated space-ground network based on submodule optimization according to claim 2 is characterized in that: The S2 step is specifically as follows: Calculate the total delay experienced by the vehicle in different allocation modes in the edge cooperative reasoning system based on the resource allocation decision variable, the task offloading decision variable, the data of the vehicle, the drone, the satellite, the base station, and the neural network parameters; The vehicle j decides to perform its reasoning task locally, i.e. Then the total delay experienced by vehicle j to complete the reasoning task is: where ξ j is the computing power of vehicle j, α j is the number of CPU cycles required to process one bit of data; When , the time required for the base station u to complete the inference task is: in For layer Post-reasoning Task A j The amount of data that needs to be processed, the total delay experienced by the vehicle j and the base station u to complete the cooperative reasoning task is: definition Assign decision variables to the computation units, and the computation unit l on the UAVk is assigned to the vehicle j, then otherwise When , the time required for the UAVk to complete the reasoning task is: in For layer Post-reasoning Task A j The amount of data that needs to be processed, the total delay experienced by the vehicle j and the UAV k to complete the cooperative reasoning task is: definition Assign decision variables to the computing units. The computing unit m on the satellite s is assigned to the vehicle j. Then otherwise When , the time required for the satellite s to complete the inference task is: When the inference task of vehicle j is offloaded to the satellite, the total delay experienced by vehicle j is:

4. The method for optimizing the inference efficiency of an integrated space-ground network based on submodule optimization according to claim 1, characterized in that: The S3 step is specifically as follows: The objective function is specifically: The constraints include: Represents the constraints of computing resources; Indicates that each vehicle can only choose one cooperative reasoning mode; Indicates that each computing unit can be assigned to at most one vehicle in one allocation cycle; Indicates that the base station, the drone, and the satellite can associate with a limited number of vehicles; Indicates resource allocation strategy and terminal association strategy; Indicates the model split point; in, N represents the upper limit of the number of vehicles that can be accessed by the base station, the drone, and the satellite. max , L max , M max Indicates the upper limit of the computing resources of the base station, the drone, and the satellite.

5. The method for optimizing the inference efficiency of an integrated space-ground network based on submodule optimization according to claim 1, characterized in that: The S4 step is specifically as follows: S41. Questions The sub-problem P1 is given as: Constraints include: Represents the constraints of computing resources; Indicates that each vehicle can only choose one cooperative reasoning mode; Indicates that each computing unit can be assigned to at most one vehicle in one allocation cycle; Indicates that the base station, the drone, and the satellite can associate with a limited number of vehicles; Indicates resource allocation strategy and terminal association strategy; Among them, the objective function D j (w,z)=T0-T j Aims to minimize delay cost, T0 is the task waiting time; Given the subproblem P2: Constraints include: Indicates the model split point; S42, converting the joint terminal association and resource allocation problem into a submodule problem and solving it; The subproblem P1 is transformed into a submodular problem, and the computing resource allocation and terminal association strategy are defined as in represents the action of allocating the computing resource l on the drone k to the vehicle j, The action indicating that the computing resource n on the base station u is allocated to the vehicle j, represents the action of associating the base station u with the vehicle j, represents the action associated with the drone k and the vehicle j; the basis set A can be divided into V disjoint sets (V′=UV), that is, for any v≠v′, in The basis set B can be divided into disjoint sets The basis set C can be divided into U disjoint sets, where The basis set D can be divided into K disjoint sets; Define a pair E = (F, G), where F = {A, B, C, D} is a global set, A set of independent subsets of F is used to replace the original constraints, defined as Where X is the set of final solutions; prove that the set family E = (F, G) is a pseudo-matrix, and the delayed reduction function Is a monotone submodular function on X∈G; on this basis, by replacing the constraints in the problem with set pairs, the subproblem P1 is expressed as a matroid-constrained monotone submodular maximization problem: stX∈G Initialize A v , B v , C j , D j is an empty set, where and define X = {A, B, C, D} and At each iteration, the new element with the highest marginal gain is added Add to the set X x and remove from the candidate action set Y x, the candidate action set Y x is the set of all currently selectable strategies; In the case of , the computing units on other base stations and drones are assigned to the action of the vehicle j, and the actions of other base stations and drones associated with the vehicle j are removed from the candidate action set, that is: and Repeat this iteration until the candidate action set Y x is empty or the marginal gain Δ D (a|X) is zero; wherein the marginal gain refers to: in a submodular function, when a new element is added to a set, the contribution of the element to the increase in the function value; wherein the marginal gain Δ D (a|X)=D(a∪{X})-D(X); When considering satellites, a resource accumulation-based reallocation method is used to check whether each vehicle can connect to the satellite through the drone through a loop. The specific steps are as follows: Initialize X, And the global optimal allocation strategy; when as well as When the conditions are met, the inference latency is calculated in a loop. If a suitable satellite access strategy is found that can reduce the inference latency, the global optimal allocation strategy is updated. After the update, the corresponding computing resources are allocated and the allocation is marked as used. If the satellite resource allocation is better than other options, the drone is selected to connect to the satellite and the vehicle's computing power and latency are updated. S43, model segmentation algorithm to place the model, The S43 step specifically includes: The optimal split point is determined by traversal search. Considering the cooperative reasoning between vehicles and base stations, the model segmentation algorithm is proposed as follows: Input: total number of layers of the DNN model I, bandwidth B, computing power of base station u The computing power of vehicle j ξ j , the amount of output data per layer Optimal split point i best ; ① Minimum time to initialize vehicle j Calculate time locally ② Assume that the initial optimal segmentation layer i best =I; ③For each layer i∈{1,2,…,I} of the neural network in the inference task, perform the following operations: Calculate the transmission time of the current layer Processing time at base station u and the processing time on vehicle j If the total time to perform the inference task The minimum update time is And i best Set to current layer i; ④After the cycle ends, return to the optimal segmentation point i best ; S44. Solve the two sub-problems alternately until the optimal strategy is obtained.

Images