Graph optimization method supporting suboptimal data training diffusion generation

Abstract

The invention discloses a graph optimization method supporting suboptimal data training diffusion generation. The method comprises the steps that a multi-server and multi-user reality system in a general computing network environment is converted into a virtual graph, the virtual graph comprises a plurality of user nodes, a plurality of edge server nodes and edges between the user nodes and the edge server nodes, and when the user nodes process own tasks, the edge server nodes and the edges of the user nodes are distributed in the virtual graph; the local execution cost of the user node is determined, and when the user node unloads the task of the user node to the edge server node for processing, the transmission cost for unloading the execution cost and the execution cost of the edge server are determined. The technical problems that in an existing graph form modeling network optimization method, network optimization is complex and difficult to solve, convergence of an optimal solution is not guaranteed, dependence on a large-scale optimal solution sample is high, and combination of optimality, high efficiency and flexibility is lacked are solved.

Description

Technical Field

[0001] The present invention relates to multiple technical fields such as diffusion generation models, network optimization, graph optimization, and computation offloading, and in particular to a graph optimization method that supports diffusion generation of suboptimal data training. Background Art

[0002] With the transformation of specialized services in the fields of communication, perception, and computing to integrated and intelligent inter-sensory computing networks, a large number of new application services have emerged. The number of terminals and hosts, sensors and collectors, and the amount of data they generate are growing exponentially. The services and tasks involved in scenarios such as the Internet of Vehicles, drone networks, and the Internet of Things have put forward higher requirements on the performance, efficiency, and reliability of the network. The network system and computing infrastructure need to achieve optimization goals in specific indicators such as power consumption, transmission throughput, computing efficiency, and processing latency. For the problem of computing offloading, since the local hardware (storage, power) and logic (computing, reasoning) resources of terminal devices are usually limited, it is necessary to use computing offloading technology to transfer computing-intensive and delay-sensitive tasks from the terminal to servers with more sufficient resources for execution and return execution results, thereby improving the service quality of computing tasks.

[0003] Existing network optimization methods can be roughly divided into two categories. The first category is numerical optimization methods, which can generally guarantee theoretical optimality, but the disadvantage is that the operation efficiency is low due to the requirement for iteration or recursion, and the flexibility to deal with different scenarios is also poor, because such methods often require expert knowledge to make new designs; the second category is deep learning methods, which have the advantage of high time and space efficiency, because single-step reasoning directly outputs the solution, but this is essentially learning the mapping function from input parameters to output solutions, and there is no theoretical guarantee for the convergence of the optimal solution. In addition, although the neural network model is universal, most of the existing deep learning algorithms rely on large-scale optimal data sets, which are difficult to prepare for complex optimization problems, so the flexibility is also limited.

[0004] However, modeling network optimization problems in the form of graphs can support a variety of network optimization problems such as multi-server multi-user computing offloading, routing scheduling, and computing task scheduling, achieving a wider range of problem coverage. At the same time, suboptimal solutions to such complex optimization problems can often be obtained efficiently, and the quality of these suboptimal solutions is acceptable compared to the optimal solution in terms of the objective optimization function. Therefore, the suboptimal solution samples contain rich optimization target knowledge that can be captured by deep learning models.

[0005] In summary, network optimization problems in reality are complex and difficult to solve, and existing methods have limitations such as no guarantee of convergence to the optimal solution and strong dependence on large-scale optimal solution samples. They lack optimality, efficiency and flexibility. It is necessary to design graph optimization methods that meet these requirements for complex task goals of synaesthesia network optimization. Summary of the invention

[0006] The embodiment of the present invention provides a graph optimization method that supports diffusion generation of suboptimal data training, so as to at least solve the limitations of existing graph modeling network optimization methods, such as the complexity and difficulty of network optimization, the lack of guarantee of convergence of the optimal solution, the strong dependence on large-scale optimal solution samples, and the lack of technical problems of optimality, efficiency and flexibility.

[0007] According to one aspect of an embodiment of the present invention, a graph optimization method that supports diffusion generation of suboptimal data training is provided. The method may include: converting a multi-server multi-user real system in a synaesthesia computing network environment into a virtual graph, wherein the virtual graph includes multiple user nodes, multiple edge server nodes, and edges between each user node and each edge server node; when a user node processes its own task, the local execution cost of the user node is determined; when the user node offloads its own task to an edge server node for processing, the transmission cost of the offload execution cost and the edge server execution cost are determined; modeling is performed on the virtual graph based on the local execution cost, the transmission cost of the offload execution cost, and the edge server execution cost to obtain a graph optimization problem, wherein the graph optimization problem includes two variables, the offload decision and the computing resource allocation, and the graph optimization problem is equivalent to the total cost minimization objective of the real system; in the graph optimization problem, the computing resource allocation is randomly heuristically initialized to obtain To the initialized computing resource allocation, based on the initialized computing resource allocation, the transmission cost of the offloading execution cost and the edge server execution cost, the offloading execution cost is determined; based on the local execution cost and the offloading execution cost, a network flow graph is constructed, wherein the network flow graph includes: edge server nodes, virtual edges, real edges, source points, sinks, and user nodes, and the inflow and outflow traffic of the edge server nodes and the user nodes in the network flow graph are equal; in the network flow graph, the minimum cost maximum flow algorithm is used to select the best edge for each user node from multiple edges connected to each user node; the local execution cost, the transmission cost of the offloading execution cost and the edge server cost of the network flow graph corresponding to the initialized computing resource allocation and the best edge of each user node are input into the diffusion model of the graph optimization to obtain the target edge and target computing resource allocation of each user node.

[0008] Optionally, modeling is performed on a virtual graph based on the local execution cost, the transmission cost of the offloaded execution cost, and the edge server execution cost, and the expression of the graph optimization problem is obtained as follows:

[0009]

[0010] in, is the local execution cost, α i ∈[0,1],α i It is used to represent the cost weight of the computing task corresponding to the i-th connection in the virtual graph. i is used to represent the i-th connection line connecting the user node to the edge server node in the virtual graph. The value range of i is 1 to L. represents the transmission cost of the offload execution cost, Used to represent the edge server execution cost, A o represents the computing resource allocation when edge nodes allocate computing resources to user nodes, D o Indicates whether the user node assigns tasks to the edge server, E j represents the edge from the user node offloading tasks to the edge server node, U α represents the αth user node in the virtual graph, D j Indicates that the tasks of the user node are completely executed locally or the tasks of the user node are completely offloaded to an edge server node for execution. The value range of α is 1 to M. D j A j It is used to indicate that the amount of computing resources occupied by tasks offloaded from different user nodes to the same edge server node does not exceed the total available amount, S β Represents the βth edge server node in the virtual graph, and the value range of β is 1 to K.

[0011] Optionally, the expression for the local execution cost is:

[0012]

[0013] in, Indicates the local computation time of the task of the user node of the i-th connection line, The task-local computing power consumption of the user node of the i-th connection line.

[0014] Optionally, the expression of the transmission cost of the offload execution cost is:

[0015]

[0016] Among them, I i It is used to indicate the task data volume of the user node corresponding to the i-th connection line. It is used to represent the transmission rate of the user node on the i-th connection line offloading the task to the edge server node, P t It is used to represent the transmission power of the user node on the i-th connection line to offload the task to the edge server node.

[0017] Optionally, the expression of the edge server execution cost is:

[0018]

[0019] Among them, V i It is used to represent the task computing requirements of the edge server node of the i-th connection line, P e is the execution power of the edge server node, and F represents the available computing resources of the edge server node.

[0020] Optionally, based on the initialized computing resource allocation, the transmission cost of the offloading execution cost, and the edge server execution cost, an expression for determining the offloading execution cost is:

[0021]

[0022] Among them, C is used to represent the offloading execution cost, A i Indicates the allocation of computing resources after initialization.

[0023] Optionally, the initialized computing resource allocation and the local execution cost, the transmission cost of the offloaded execution cost, and the edge server cost of the network flow graph corresponding to the optimal edge of each user node are input into the diffusion model of graph optimization to obtain the target edge and target computing resource allocation of each user node, including: parameterizing and denoising the local execution cost, the transmission cost of the offloaded execution cost, and the edge server cost to obtain a parameterized result; inputting the parameterized result into the diffusion model to obtain a generated solution; parameterizing and sampling the generated solution to obtain the target edge and target computing resource allocation of each user node.

[0024] Beneficial effects of the present invention:

[0025] The present invention proposes a graph optimization method that supports diffusion generation of suboptimal data training, obtains a network flow graph equivalent to the virtual graph through modeling, and determines the suboptimal computing resource allocation and unloading decision of the network flow graph, realizes theoretical modeling of learning suboptimal solutions to generate optimal solutions based on the theoretical definition of parameter solution space, obtains the optimal computing resource allocation and unloading decision corresponding to the suboptimal computing resource allocation and unloading decision, solves the limitations of existing graph modeling network optimization methods such as complex and difficult to solve network optimization, no guarantee of convergence of optimal solutions, strong dependence on large-scale optimal solution samples, and lacks technical problems of both optimality, high efficiency and flexibility, achieves the equivalent graph optimization problem of network optimization problems obtained through graph modeling, and uses efficient suboptimal solution production tools to prepare sufficient suboptimal solutions, realizes theoretical modeling of learning suboptimal solutions to generate optimal solutions based on the theoretical definition of parameter solution space, uses DGM to implement GODM model with GNN as the model backbone, and can achieve the technical effect of real-time generation of optimal solutions on various network optimization objectives by learning efficiently prepared suboptimal solutions. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of this application. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0027] Figure 1 is a flow chart of a graph optimization method supporting diffusion generation of suboptimal data training according to an embodiment of the present invention;

[0028] Figure 2 is a modeling diagram of a multi-server multi-user computing offloading scenario according to an embodiment of the present invention;

[0029] Figure 3 is a network flow graph of random heuristics and minimum cost maximum flow according to an embodiment of the present invention;

[0030] Figure 4 is a schematic diagram of a diffusion generation model for graph optimization according to an embodiment of the present invention. DETAILED DESCRIPTION

[0031] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only embodiments of a part of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0032] It should be noted that the terms "first", "second", etc. in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects and to describe a specific order or sequence. It should be understood that the terms used in this way can be interchanged where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0033] Example 1

[0034] According to an embodiment of the present invention, a graph optimization method supporting diffusion generation of suboptimal data training is provided. It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system comprising at least one set of computer executable instructions, and although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in an order different from that shown here.

[0035] Figure 1 is a flowchart of a graph optimization method supporting diffusion generation of suboptimal data training according to an embodiment of the present invention. Figure 1 As shown, the method may include the following steps:

[0036] Step S101, converting a multi-server multi-user real system in a synaesthesia network environment into a virtual graph, wherein the virtual graph includes multiple user nodes, multiple edge server nodes, and edges between each user node and each edge server node. When a user node processes its own task, the local execution cost of the user node is determined. When the user node offloads its own task to the edge server node for processing, the transmission cost of the offloaded execution cost and the edge server execution cost are determined.

[0037] In the technical solution provided in the above step S101 of the present invention, Figure 2 is a multi-server multi-user computing offloading scenario modeling diagram according to an embodiment of the present invention, such as Figure 2 As shown, the real system is converted into a virtual graph. There are K edge servers: S = {S1,...,S K}, there are M mobile users: U={U1,...,U M}, the available connection between each mobile user and the edge server in the network is represented by a directed edge ε = {E j (U α ,S β )|if U αWith S β There is an available connection between them}. Assume L = |ε|, considering the limited wireless coverage in real scenarios, each mobile user only has available connections with some edge servers. At the same time, we can also know that the edges of the same mobile user connecting to different edge servers correspond to the same actual computing task. In the multi-server multi-user computing offloading problem, there is a directed acyclic graph. The edge features include the cost of offloading the user task on the edge to the edge server on the edge. This cost is calculated based on parameters such as communication status and computing task information. In a round of computing offloading, each mobile user has a computing task, which is either executed completely locally or completely offloaded to an edge server for execution. Therefore, the total cost of a task is divided into three parts (local execution cost, transmission cost of offloading execution cost, and edge server execution cost).

[0038] Step S102, modeling is performed on a virtual graph based on the local execution cost, the transmission cost of the offloaded execution cost, and the edge server execution cost to obtain a graph optimization problem, wherein the graph optimization problem includes two variables: offloading decision and computing resource allocation, and the graph optimization problem is equivalent to the total cost minimization objective of the real system.

[0039] In the technical solution provided in step S102 of the present invention, the problem is modeled on the virtual graph to obtain a graph optimization problem equivalent to the original total computational cost minimization objective, so as to offload the decision and computing resource allocation The two variables are optimization variables, which correspond to edge selection and edge weight assignment in the virtual graph respectively.

[0040] Step S103, in the graph optimization problem, the computing resource allocation is randomly heuristically initialized to obtain the initialized computing resource allocation, and the offloading execution cost is determined based on the initialized computing resource allocation, the transmission cost of the offloading execution cost, and the edge server execution cost.

[0041] In the technical solution provided in step S103 of the present invention, the cost (triplet) of the edge feature is calculated to represent the flow cost of the edge in the network flow graph. The complete calculation of this cost requires the random heuristic initialization of the edge weights of the original virtual graph. That is, the allocation of computing resources According to the computing resources V required for each task i Perform random heuristic initialization to obtain the initialized computing resource allocation, calculate the initialized computing resource allocation, the transmission cost of the offloading execution cost, and the edge server execution cost to obtain the offloading execution cost.

[0042] Step S104, constructing a network flow graph based on the local execution cost and the offload execution cost, wherein the network flow graph includes: edge server nodes, virtual edges, real edges, source points, sink points, and user nodes. In the network flow graph, the inflow and outflow traffic of the edge server nodes and the user nodes are equal.

[0043] In the technical solution provided in the above step S104 of the present invention, a network flow graph is constructed according to the local execution cost and the offload execution cost. Figure 3 is a network flow diagram of random heuristics and minimum cost maximum flow according to an embodiment of the present invention, Figure 3 The cost of local execution Directly used as the traffic cost of the corresponding edge (dashed edge) The cost of offloading execution (the solid line traffic cost) is represented by a two-tuple. The first variable is the traffic upper limit (0 or 1) and the traffic is a non-negative integer; the second variable is the cost per unit traffic.

[0044] Step S105 , in the network flow graph, the optimal edge is selected for each user node from multiple edges connected to each user node by using the minimum cost maximum flow algorithm.

[0045] In the technical solution provided in the above step S105 of the present invention, the minimum cost maximum flow algorithm is a solution for obtaining the minimum total cost while ensuring the maximum flow from the source point to the sink point. Figure 2 The network flow graph ensures that the upper limit of the flow into each user node is always 1, so the maximum flow is constantly equal to the number of users, which ensures that all tasks must be executed; then the outflow path of each user node includes a dotted edge directly pointing to the sink and a solid edge pointing to several edge server nodes. The cost of the dotted edge is the local execution cost, and the cost of the solid edge is the corresponding offload execution cost. The minimum cost maximum flow algorithm will only select one of these edges for a user node. The minimum cost process ensures the initialization back The optimal (unloading decision and computing resource allocation Both variables are optimized variables). Therefore, although the random heuristic initialization It will introduce suboptimality, but the existing minimum cost maximum flow algorithm with polynomial complexity can be used on the network flow graph to obtain the edge selection result. .

[0046] Step S106, input the initialized computing resource allocation and the local execution cost of the network flow graph corresponding to the optimal edge of each user node, the transmission cost of the offloaded execution cost and the edge server cost into the diffusion model of the graph optimization to obtain the target edge and target computing resource allocation of each user node.

[0047] In the technical solution provided in step S106 of the present invention, the initialized computing resource allocation and the local execution cost of the network flow graph corresponding to the optimal edge of each user node, the transmission cost of the offload execution cost and the edge server cost are input into Figure 3 The optimal solution for the initialized computing resource allocation and the optimal edge of each user node is solved, and the target edge and target computing resource allocation of each user node are obtained.

[0048] The above method of this embodiment is further introduced below.

[0049] As an optional implementation method, step S102, based on the local execution cost, the transmission cost of the offload execution cost and the edge server execution cost, a model is built on the virtual graph, and the expression of the graph optimization problem is obtained as follows:

[0050]

[0051] in, is the local execution cost, α i ∈[0,1],α i It is used to represent the cost weight of the computing task corresponding to the i-th connection in the virtual graph. i is used to represent the i-th connection line connecting the user node to the edge server node in the virtual graph. The value range of i is 1 to L. represents the transmission cost of the offload execution cost, Used to represent the edge server execution cost, A o represents the computing resource allocation when edge nodes allocate computing resources to user nodes, D o Indicates whether the user node assigns tasks to the edge server, E j represents the edge from the user node offloading tasks to the edge server node, U α represents the αth user node in the virtual graph, D j Indicates that the tasks of the user node are completely executed locally or the tasks of the user node are completely offloaded to an edge server node for execution. The value range of α is 1 to M. j A k It is used to indicate that the amount of computing resources occupied by tasks offloaded from different user nodes to the same edge server node does not exceed the total available amount, S β Represents the βth edge server node in the virtual graph, and the value range of β is 1 to K.

[0052] In this embodiment, a model is built on a virtual graph according to the local execution cost, the transmission cost of the offloaded execution cost, and the edge server execution cost to obtain a graph optimization problem.

[0053] As an optional implementation method, the expression of the local execution cost is:

[0054]

[0055] in, Indicates the local computation time of the task of the user node of the i-th connection line, The task-local computing power consumption of the user node of the i-th connection line.

[0056] In this embodiment, if Figure 1 As shown in the virtual diagram in i ∈[0,1] is the cost weight of the computing task corresponding to the ith connection. If it approaches 1, it emphasizes the latency cost, otherwise it emphasizes the power consumption cost. (second) is the local computing time of the task, which is determined by the task computing demand V corresponding to the connection. i (CPU Cycles) divided by the available computing resources of the mobile user device Get; set as the local computing power consumption of the task Where κ is the known hardware parameter of the computing unit, and the task calculation demand V corresponding to the connection is i And the amount of task data I i Calculated according to the generally accepted formula. Then the local execution cost is The expression is:

[0057]

[0058] As an optional implementation mode, the expression of the transmission cost of the offload execution cost is:

[0059]

[0060] Among them, I i It is used to indicate the task data volume of the user node corresponding to the i-th connection line. It is used to represent the transmission rate of the user node on the i-th connection line offloading the task to the edge server node, P t It is used to represent the transmission power of the user node on the i-th connection line to offload the task to the edge server node.

[0061] In this embodiment, if Figure 1 As shown in the virtual figure in the figure, the transmission cost of the offloading execution cost is referred to the common orthogonal multiple access in practice, and according to the channel gain h i , transmission power P t And additive Gaussian white noise N0 calculates the signal-to-noise ratio of each edge corresponding to the wireless connection Based on the fixed bandwidth B of each wireless connection, the transmission rate of the uplink is calculated according to the formula The transmission cost of offloading the execution cost can be calculated as:

[0062]

[0063] As an optional implementation mode, the expression of the edge server execution cost is:

[0064]

[0065] Among them, V i It is used to represent the task computing requirements of the edge server node of the i-th connection line, P e is the execution power of the edge server node, and F represents the available computing resources of the edge server node.

[0066] In this embodiment, for the offloaded edge server execution cost, let P e is the execution power of the server, then the edge server execution cost can be calculated as:

[0067]

[0068] As an optional implementation manner, in step S103, based on the initialized computing resource allocation, the transmission cost of the offloading execution cost and the edge server execution cost, the expression of the offloading execution cost is determined as:

[0069]

[0070] Among them, C is used to represent the offloading execution cost, A i Indicates the allocation of computing resources after initialization.

[0071] In this embodiment, the initialized computing resource allocation, the transmission cost of the offloading execution cost, and the edge server execution cost are calculated to obtain the offloading execution cost.

[0072] As an optional implementation method, step S106, inputs the initialized computing resource allocation and the local execution cost, the transmission cost of the offloaded execution cost, and the edge server cost of the network flow graph corresponding to the optimal edge of each user node into the diffusion model of graph optimization to obtain the target edge and target computing resource allocation of each user node, including: parameterizing and denoising the local execution cost, the transmission cost of the offloaded execution cost, and the edge server cost to obtain a parameterized result; inputting the parameterized result into the diffusion model to obtain a generated solution; parameterizing and sampling the generated solution to obtain the target edge and target computing resource allocation of each user node.

[0073] In this embodiment, if Figure 4As shown in Figure 1, the offloading decision (edge ​​selection) and computing resource allocation (edge ​​weight) are considered as discrete solutions (0 and 1) and continuous solutions, respectively. The parameterized solution space defines a continuous vector (or matrix) in the discrete solution space and the continuous solution space. To describe the spatial distribution of high-quality solutions for a given input x:

[0074]

[0075] Among them, N represents the optimization variable dimension, is the input parameter space, Represents the feasible solution space determined by x, the optimal solution y in this distribution * has the highest probability, and the better solution always has a higher probability than the worse solution. Such parameterization results are not unique. There are almost infinite parameterization results that can basically correctly describe the distribution of high-quality solutions and make the optimal solution reachable with high probability. These parameterization results can be described as θ′~p(Θ x ). The generative model can learn such a high-quality solution distribution very well, and the randomness of its generated results has almost no negative impact under the setting that the parameterized results are not unique, so it can infer the correct high-quality solution distribution and perform sampling to achieve the optimal solution.

[0076] For the learning objective of the diffusion generation model, the target computing resource allocation first introduces its forward noise addition process. Here That is, the virtual graph and all its features described in step 1. The noisy solution in the forward noise addition process is represented as y t ,t∈{1,2,…,T}, use q(y t |y t-1 ) represents a step in the forward noise addition process, that is, a single-step noise addition with a fixed noise distribution. The forward noise addition process treats the original y as y0 and performs T-step noise addition:

[0077]

[0078] Thus, we can obtain pure noise y with known distribution (such as Gaussian distribution) T , where the diffusion generation model Φ learns the inverse denoising process. Represents the complete noise reduction process, where As a condition. Multiple single-step noise reduction According to the multiplication rule, we can get the value of p(y T ) to p(y0):

[0079]

[0080] Thus, the noise solution y can be sampled from a known noise distributionT And gradually reduce the noise to y0. Although the denoising process can ensure that the final solution is consistent with the distribution of the original solution, there will still be additional random noise introduced, which ensures a certain randomness of the generated solution, so that the original solution distribution can be sampled through repeated generation or parallel generation to achieve the optimal solution.

[0081] For discrete solutions, corresponding to edge selection in this problem, the learning objective of the diffusion generation model Φ is:

[0082]

[0083] in, is the estimated true solution, p Φ (y t-1 |y t ) is the single-step noise distribution predicted by model Φ, yes Determine the actual noise distribution from step t to step t-1, is an explicit representation of the model Φ predicted by y t and is the estimated true solution distribution conditional on .

[0084] For continuous solutions, corresponding to the edge weight assignment in this problem, the learning objective of the diffusion generation model Φ is:

[0085]

[0086] in, represents the target noise for model inference, β t represents the noise factor, which satisfies This ensures that the noise addition result is completely noise-free, where In the specific performance, the learning goal of the model is θ′~p(Θ x ),therefore, Figure 4 The forward denoising process and the reverse denoising process are both θ′ t express.

[0087] In an embodiment of the present invention, a multi-server multi-user real system in a synaesthesia computing network environment is converted into a virtual graph, wherein the virtual graph includes multiple user nodes, multiple edge server nodes, and edges between each user node and each edge server node. When the user node processes its own task, the local execution cost of the user node is determined. When the user node offloads its own task to the edge server node for processing, the transmission cost of the offload execution cost and the edge server execution cost are determined. Modeling is performed on the virtual graph based on the local execution cost, the transmission cost of the offload execution cost, and the edge server execution cost to obtain a graph optimization problem, wherein the graph optimization problem includes two variables: offload decision and computing resource allocation, and the graph optimization problem is equivalent to the total cost minimization objective of the real system. In the graph optimization problem, the computing resource allocation is randomly heuristically initialized to obtain the initialized computing resource allocation, and the offload execution cost is determined based on the initialized computing resource allocation, the transmission cost of the offload execution cost, and the edge server execution cost. Based on the local execution cost and the offload execution cost, a network flow graph is constructed, wherein the network flow graph includes: edge server nodes, virtual edges, real edges, source points, sink points, user Nodes, the inflow and outflow traffic of edge server nodes and user nodes in the network flow graph are equal; in the network flow graph, the minimum cost maximum flow algorithm is used to select the best edge for each user node from multiple edges connected to each user node; the local execution cost of the network flow graph corresponding to the initialized computing resource allocation and the optimal edge of each user node, the transmission cost of the offloaded execution cost and the edge server cost are input into the diffusion model of graph optimization to obtain the target edge and target computing resource allocation of each user node, which solves the limitations of the existing graph modeling network optimization methods, such as the complexity and difficulty of network optimization, the lack of guarantee of convergence of the optimal solution, the strong dependence on large-scale optimal solution samples, and the lack of technical problems of both optimality, efficiency and flexibility, and achieves the equivalent graph optimization problem of network optimization problem obtained through graph modeling, and uses efficient suboptimal solution production tools to prepare a sufficient number of suboptimal solutions. Based on the theoretical definition of parameter solution space, the theoretical modeling of learning suboptimal solutions to generate optimal solutions is realized, and the GODM model is realized with DGM and GNN as the model backbone. The technical effect of real-time generation of optimal solutions can be achieved on various network optimization objectives by learning efficiently prepared suboptimal solutions.

[0088] The serial numbers of the above embodiments of the present invention are only for description and do not represent the advantages or disadvantages of the embodiments.

[0089] In the above embodiments of the present invention, the description of each embodiment has its own emphasis. For parts that are not described in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0090] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. Among them, the device embodiments described above are only schematic. For example, the division of units can be a logical function division. There may be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of units or modules, which can be electrical or other forms.

[0091] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed over multiple units. Some or all of the units may be selected according to actual needs to achieve the purpose of the present embodiment.

[0092] In addition, each functional unit in each embodiment of the present invention may be integrated into a first processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0093] The above are only preferred embodiments of the present invention. It should be pointed out that, for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A graph optimization method supporting diffusion generation of suboptimal data training, characterized in that: include: The multi-server multi-user real system in the synaesthesia computing network environment is converted into a virtual graph, wherein the virtual graph includes multiple user nodes, multiple edge server nodes, and edges between each user node and each edge server node. When the user node processes its own task, the local execution cost of the user node is determined. When the user node offloads its own task to the edge server node for processing, the transmission cost of the offloaded execution cost and the edge server execution cost are determined. Based on the local execution cost, the transmission cost of the offloaded execution cost and the edge server execution cost, a graph optimization problem is obtained. The graph optimization problem includes two variables: offload decision and computing resource allocation. The graph optimization problem is equivalent to the total cost minimization objective of the real system. In the graph optimization problem, the computing resource allocation is initialized randomly heuristically to obtain the initialized computing resource allocation, and the offloading execution cost is determined based on the initialized computing resource allocation, the transmission cost of the offloading execution cost, and the edge server execution cost; Based on the local execution cost and the offload execution cost, a network flow graph is constructed, wherein the network flow graph includes: edge server nodes, virtual edges, real edges, source nodes, sink nodes, and user nodes, and the inflow and outflow traffic of the edge server nodes and the user nodes in the network flow graph are equal; In the network flow graph, the minimum cost maximum flow algorithm is used to select the best edge for each user node from the multiple edges connected to each user node; The initialized computing resource allocation and the local execution cost of the network flow graph corresponding to the optimal edge of each user node, the transmission cost of the offloaded execution cost, and the edge server cost are input into the diffusion model of graph optimization to obtain the target edge and target computing resource allocation of each user node.

2. The method according to claim 1, characterized in that Based on the local execution cost, the transmission cost of the offloaded execution cost, and the edge server execution cost, the model is built on the virtual graph, and the expression of the graph optimization problem is obtained as follows: s.t.C1: C2: C3: C4: in, is the local execution cost, α i ∈[0,1],α i It is used to represent the cost weight of the computing task corresponding to the i-th connection in the virtual graph. i is used to represent the i-th connection line connecting the user node to the edge server node in the virtual graph. The value range of i is 1 to L. represents the transmission cost of the offload execution cost, Used to represent the edge server execution cost, A i represents the computing resource allocation when edge nodes allocate computing resources to user nodes, D i Indicates whether the user node assigns tasks to the edge server, E j represents the edge from the user node offloading tasks to the edge server node, U α represents the αth user node in the virtual graph, D j Indicates that the tasks of the user node are completely executed locally or the tasks of the user node are completely offloaded to an edge server node for execution. The value range of α is 1 to M. D j A j It is used to indicate that the amount of computing resources occupied by tasks offloaded from different user nodes to the same edge server node does not exceed the total available amount, S β Represents the βth edge server node in the virtual graph, and the value range of β is 1 to K.

3. The method according to claim 2, characterized in that The expression of local execution cost is: in, Indicates the local computation time of the task of the user node of the i-th connection line, The task-local computing power consumption of the user node of the i-th connection line.

4. The method according to claim 2, characterized in that: The expression of the transmission cost of offloading the execution cost is: Among them, I i It is used to indicate the task data volume of the user node corresponding to the i-th connection line. It is used to represent the transmission rate of the user node on the i-th connection line offloading the task to the edge server node, P t It is used to represent the transmission power of the user node on the i-th connection line to offload the task to the edge server node.

5. The method according to claim 2, characterized in that: The expression of edge server execution cost is: Among them, V i It is used to represent the task computing requirements of the edge server node of the i-th connection line, P e is the execution power of the edge server node, and F represents the available computing resources of the edge server node.

6. The method according to claim 1, characterized in that Based on the initialization of computing resource allocation, the transmission cost of the offloading execution cost and the edge server execution cost, the expression for determining the offloading execution cost is: Among them, C is used to represent the offloading execution cost, A i Indicates the allocation of computing resources after initialization.

7. The method according to claim 1, characterized in that The initialized computing resource allocation and the local execution cost of the network flow graph corresponding to the optimal edge of each user node, the transmission cost of the offloaded execution cost, and the edge server cost are input into the diffusion model of graph optimization to obtain the target edge and target computing resource allocation of each user node, including: The local execution cost, the transmission cost of the offloaded execution cost, and the edge server cost are parameterized and denoised to obtain a parameterized result; The parameterized results are input into the diffusion model to obtain the generated solution; The generated solution is parameterized and sampled to obtain the target edge and target computing resource allocation for each user node.

8. A processor, characterized in that: The processor is used to run a program, wherein the program executes the method according to any one of claims 1 to 7 when running.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium includes a stored program, wherein when the program is executed, the device where the computer-readable storage medium is located is controlled to execute the method according to any one of claims 1 to 7.

Images