Communication network topology planning method and system

Abstract

The present invention provides a communication network topology planning method and system, comprising: modeling network planning problem instance parameters as a network planning graph signal X; and establishing a conversion relationship between a network topology solution and a graph adjacency matrix; inputting the obtained network planning signal X into a communication network planning problem solver based on a graph diffusion generation model, and combining the conversion relationship to obtain an optimal solution C of the network planning problem. ★ The present invention is a novel generative sampling solution that can obtain multiple high-quality network topology feasible solutions at once with low computational complexity, effectively improving the efficiency of solving network planning problems.

Description

Technical Field

[0001] The present invention relates to the field of network communication technology, and in particular to a communication network topology planning method and system. Background Art

[0002] In recent years, multimedia content service providers have been required to invest billions of dollars annually in network operating costs to build Software-Defined Networks (SDNs) to optimize centralized network topologies and traffic transmission to support global traffic demands exceeding hundreds of terabits per second. By leveraging customizable switching network components such as optical circuit switches or wireless transmission channels, content service providers can adaptively adjust communication links, such as optical fiber or wireless networks, through network planning, meeting ever-changing traffic demands while reducing network operating costs.

[0003] Communication network planning is a key step in supporting the demand for low-latency and high-bandwidth video traffic transmission services while optimizing overall network costs, which include both operating and construction costs. Given the network's transmission traffic demand, physical link status, and network service requirements, network planning requires decision-making, optimization, and updating of network modules involving network topology (physical and network layer node locations and link connections), corresponding link capacity, and routing strategies. In actual online video streaming scenarios, the traffic demand between sites and the availability of physical links between sites often exhibit random and time-varying characteristics. Therefore, when the system status of the content service provider's transmission network changes significantly, the network topology needs to be quickly and adaptively re-optimized and updated. The difficulty in solving this challenge lies in how to determine the optimal or suboptimal network topology within a limited computing time.

[0004] However, network planning is a typical mixed integer programming (MIP) problem, where the search space for optimization variables grows exponentially with the number of network links. Consequently, commercial solvers such as CPLEX, Gurobi, and SCIP typically take a long time to solve and are not scalable to large networks. Furthermore, it is impractical to reduce computational complexity by manually adjusting heuristic methods with expert experience at each planning step.

[0005] At present, how to design an efficient solution to network planning problems and obtain the optimal / suboptimal network topology solution under the constraints of limited computing time and computing resources is a major challenge facing current communication network planning tasks. Summary of the Invention

[0006] In view of the defects in the prior art, an object of the present invention is to provide a communication network topology planning method and system.

[0007] According to one aspect of the present invention, a communication network topology planning method is provided, comprising:

[0008] Model the network planning problem instance parameters as the network planning graph signal X; and establish the network topology solution The conversion relationship between the graph adjacency matrix and the graph;

[0009] The obtained network planning signal X is input into the communication network planning problem solver based on the graph diffusion generation model, and the optimal solution C of the network planning problem is obtained by combining the conversion relationship. ＊ .

[0010] Preferably, modeling the network planning problem instance parameters as a network planning graph signal X includes:

[0011] Obtain network planning problem instance parameters, including: IP layer topology site set V, IP link set ε, physical link set Physical link design capacity S u , IP layer site traffic requirements and the network overall cost function U(C);

[0012] The network planning problem instance parameters are modeled as a network planning graph signal X = [X0, ..., X i ,…,X N-1 ] T ,The specific modeling process is: Element X i Depend on The three vectors are concatenated in sequence, where the vector Contains IP layer network site v i Outbound traffic requirements to all other IP layer network sites and other IP site-to-site v i The inbound traffic demand, vector Contains the current IP layer network site v i Related physical links Design capacity S u Related information, vector Contains the network planning problem optimization objective U(C) and the IP layer network site v i Parameters related to the available IP layer and physical layer network links for the connection.

[0013] Preferably, the establishment of a network topology solution The conversion relationship between the graph adjacency matrix and the graph includes:

[0014] Assuming that the link capacities in both directions of any IP chain are equal, the relationship between the IP link capacity value and the elements in the adjacency matrix A is expressed as:

[0015]

[0016] Among them, z norm Represents a scaling factor, which changes the value A e =A ij =A ji Normalized to the interval [0,1].

[0017] Preferably, the obtained network planning signal X is input into a communication network planning problem solver based on a graph diffusion generation model, and the optimal solution C of the network planning problem is obtained by combining the conversion relationship. ＊ ,include:

[0018] Use an offline trained graph diffusion generation model score function s θ , get the score function s θ The output value of

[0019] Based on the network planning graph signal X and the score function, M graph adjacency matrix samples are generated by parallel sampling, and the generated adjacency matrix samples are combined with the conversion relationship to convert into M network topology solutions.

[0020] Will As multiple initial solutions to the network planning problem, we add variable search space constraints on the original network planning problem, perform local adjustment and optimization, and obtain the optimal network topology solution C that meets feasibility. ＊ .

[0021] Preferably, the graph diffusion generation model score function s θ , a general deep graph neural network based on the GraphAttention mechanism is used to build the model, with the model parameter θ and the input being the original or noisy graph representation and time step t, where A t is the adjacency matrix after noise addition. The noise addition process is determined by the forward noise addition process of the general diffusion generation model. The output is the graph representation of the diffusion generation process. Gradient of the log probability density function with respect to the adjacency matrix The fitted value of

[0022] Preferably, the diffusion-generated score function s based on the deep graph neural network θ The offline training process is as follows:

[0023] Collect existing communication network planning problem instance parameters and network topology solutions obtained by any solution method;

[0024] Convert all problem instance parameters into network planning diagram signals;

[0025] Collect all network planning map signals and corresponding network topology solutions to construct an offline training set;

[0026] Establish the scoring function s θ The training loss function of parameter θ is:

[0027]

[0028] Among them, λ(t) is a positive weight function, σ(t) represents the variance of the Gaussian distribution of the noise addition process, ∈ is a standard Gaussian noise sample, and T is the maximum value of the set time step t.

[0029] Preferably, the generating M graph adjacency matrix samples by parallel sampling based on the network planning graph signal X and the scoring function includes:

[0030] Sample M N×N-dimensional initial graph adjacency matrices from random noise Where N is the number of IP layer network sites, and the reverse stochastic partial differential equation SDE solution time coordinate set representing the diffusion generation process is obtained Arrange in descending order;

[0031] The parallel sampling generation process starts from Starting from the largest time coordinate, for all In parallel, Solve the reverse SDE at all time coordinates in and obtain the adjacency matrix at each time coordinate The reverse SDE formula is expressed as:

[0032]

[0033] Among them, the function f(A t ,t) and g(t) are the outputs of the offset coefficient and diffusion coefficient of the reverse diffusion generation process respectively, w is the random noise, E(X,A t ,y) is the energy function in the back-diffusion generation process, y is the artificially defined adjacency matrix sample expected attribute, and the scalar γ is the weight coefficient of the energy function;

[0034] Finally, the reverse SDE solution at time coordinate t = 0 is Adjacency matrix samples generated as a parallel sampling generation process.

[0035] Preferably, the energy function E(X,A t ,y) is specifically expressed as,

[0036] E(X,A t ,y)=||f(A t ,X)-y||2

[0037] Where f(A t ,X) represents a predictor used to evaluate the specific properties of the graph adjacency matrix sample, and the input is the graph signal X of the network planning problem and the corresponding adjacency matrix sample A t , the output is sample A t The predicted scalar or vector value for a specific attribute; y is the expected attribute label value of the sample in the artificially defined adjacency matrix.

[0038] Preferably, the generated adjacency matrix samples are combined with the conversion relationship to convert into M network topology solutions include:

[0039]

[0040] Among them, the scalar δ is the preset IP link bandwidth capacity unit adjustment interval, The operator returns the largest integer value less than or equal to x.

[0041] Preferably, the As multiple initial solutions to the network planning problem, we add variable search space constraints on the original network planning problem, perform local adjustment and optimization, and obtain the optimal network topology solution C that meets feasibility. ＊ ,include:

[0042] The initial topological solution of any network planning problem By adding an initial solution to the original network planning problem The corresponding variable search space constraints form a local adjustment optimization problem for network planning; the local adjustment optimization problem is solved by any commercial mixed integer programming problem solver, such as Gurobi, SCIP software, to obtain a suboptimal solution after local adjustment

[0043] For M topological initial solutions Perform the above local optimization operations to obtain no more than M suboptimal solutions that meet all the constraints of the optimization problem. Then, calculate all suboptimal solutions The corresponding optimization target value, select the suboptimal solution with the minimum or maximum target value As the optimal network topology solution C ＊ .

[0044] Preferably, for any network planning problem, the initial topology solution The constraint condition of the variable search space constrains the optimization variable search space of the original network planning problem to the initial solution. within the neighborhood of ;

[0045] The variable search space constraints are specifically:

[0046]

[0047] in, is the optimization variable of the original network planning problem, and the variable μ ≥ 1 is the relaxation factor for constructing the local search space of the optimization variable.

[0048] According to a second aspect of the present invention, there is provided a communication network topology planning system, comprising:

[0049] Communication network planning problem graph signal modeling module: Model the network planning problem instance parameters as network planning graph signal X; and establish the network topology solution The conversion relationship between the graph adjacency matrix and the graph;

[0050] Solving module based on graph diffusion generation model: input the obtained network planning signal X into the communication network planning problem solver based on graph diffusion generation model, and combine the conversion relationship to obtain the optimal solution C of the network planning problem ＊ .

[0051] Compared with the prior art, the embodiments of the present invention have at least one of the following beneficial effects:

[0052] The communication network topology planning method in the embodiment of the present invention, as an auxiliary module of the existing communication network design and planning scheme, can improve the solution speed of network planning problems without sacrificing solution performance.

[0053] The communication network topology planning method in the embodiment of the present invention can use hardware computing chips such as GPU or TPU to accelerate the parallel sampling and generation process of graph adjacency matrix samples in the online generation module, greatly reducing the computational complexity of the solution to the network planning optimization problem, and helping to perform real-time network design planning and adjustment for communication networks containing large-scale communication sites, reducing the operation and construction costs of the communication network, and improving the overall service quality of communication network traffic transmission tasks.

[0054] The graph signal modeling method for the communication network planning problem in the embodiment of the present invention converts the solution process of the network topology planning problem into a graph sampling generation process. The signal modeling method can effectively alleviate the problem of graph signal data dimension growth under the large-scale IP network sites and link numbers, and improve the scalability of the method to the network scale.

[0055] The offline training process of the diffusion generative model in this embodiment of the present invention can learn the graph structure probability distribution of high-quality network topology solutions from existing experience in solving network planning problems, and effectively train a graph diffusion generative model scoring function that can extract the characteristics of this distribution. The scoring function used can be constructed using any neural network model architecture, and the parameter training loss function is simple, making the training process stable.

[0056] The controllable online topology sample generation process in the embodiments of the present invention, a key step in solving network planning problems, can rapidly generate graph adjacency matrix samples corresponding to high-performance network topology solutions based on graph signals corresponding to network planning problem parameters. This generation process can be performed in parallel for multiple different network planning problems, solving them simultaneously. Furthermore, by presetting an energy function with custom target attributes, the generation process can controllably generate graph adjacency matrices with predefined properties.

[0057] The online topology feasible solution fine-tuning process in this embodiment of the present invention, as an auxiliary process for the controllable generation of online topology samples, can locally adjust the sampled topology solutions to convert them into feasible topology solutions for the network planning problem and select the optimal solution from them. The local adjustment method employed is simple in principle and can be used with any commercial mixed-integer programming problem-solving software to ensure the feasibility of the proposed method in obtaining topological solutions. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Other features, objects and advantages of the present invention will become more apparent upon reading the detailed description of non-limiting embodiments with reference to the following drawings:

[0059] Figure 1 is a flow chart of a communication network topology planning method in one embodiment of the present invention;

[0060] Figure 2 This is a framework diagram of a communication network topology planning system in another embodiment of the present invention. DETAILED DESCRIPTION

[0061] The present invention will be described in detail below with reference to specific embodiments. The following embodiments will help those skilled in the art to further understand the present invention, but are not intended to limit the present invention in any form. It should be noted that, for those skilled in the art, several variations and improvements can be made without departing from the scope of the present invention. These all fall within the scope of protection of the present invention.

[0062] To facilitate understanding of the present invention, we first briefly introduce the background and mathematical form of the communication network planning problem. The definition of communication network planning is as follows:

[0063] Define the IP site in the communication network IP layer topology as v i∈V, V is the network IP layer topology site set, the total number of sites is N; define edge e i,j ∈ε is the IP layer topology site v i and site v j The IP links between them, ε is the set of all IP links. S represents the set of physical links available in the physical layer of the communication network. u Indicates the physical link in the physical layer The maximum design capacity, Δ u Indicates that the physical link An IP link in the IP layer topology is i,j It can be composed of multiple physical layer links u. Definition Contains the network site v in the IP layer i Outbound traffic requirements to all other network IP sites and other network IP site-to-site v i Inbound traffic demand, For IP link e i,j The allocated bearer traffic demand The transmission flow value. For all IP links between IP layer sites i,j Design link capacity Set, U(C) is the overall network cost function that meets the IP link design capacity C, including IP layer cost and physical layer cost.

[0064] The optimization goal of the network planning problem is to optimize the IP link design capacity C while meeting the traffic requirements of all IP sites, so as to minimize the overall network cost U(C). U(C) can be any artificially designed cost function. The mathematical optimization problem model P1 of the network planning problem can be expressed as:

[0065] P1:min C,D(e,ω) U(C)

[0066]

[0067] Among them, the constraints Ensures communication network IP site v i The flow conservation of site v i The inbound traffic of site v is always equal to i Outbound traffic plus the destination for site v i The flow value; constraint C e ≥∑ ω∈Ω D(e,ω) ensures that the designed IP link capacity can carry the total traffic allocated to the link; the constraint condition is The capacity of the IP link passing through the physical link u should be subject to its designed maximum available capacity S of the physical link u restrictions.

[0068] Based on the above background, the present invention provides an embodiment, see Figure 1 , a method for solving communication network planning problems based on a deep graph diffusion generation model, including:

[0069] S100, modeling the network planning problem instance parameters as a network planning graph signal X; and establishing a network topology solution The conversion relationship between the graph adjacency matrix and the graph;

[0070] S200, input the obtained network planning signal X into the communication network planning problem solver based on the graph diffusion generation model, and obtain the optimal solution C of the network planning problem by combining the conversion relationship. ＊ .

[0071] The above embodiment, as a new type of generative sampling solution, can obtain multiple high-quality network topology feasible solutions at one time with low computational complexity, effectively improving the efficiency of solving network planning problems.

[0072] In order to obtain the optimal solution to the network planning problem more quickly, the processing of the network planning problem instance parameters needs to be faster and more cautious. Therefore, in a preferred embodiment of the present invention, step S100 can adopt the following steps:

[0073] S101, obtain problem instance parameters, generally the IP layer topology site set V, IP link set ε, physical link set, physical link design capacity S u , IP layer site traffic requirements And the network overall cost function U(C). Model these problem instance parameters as a network planning graph signal X=[X0,…,X i ,…,X N ] T , where X i Depend on The three vectors are concatenated in sequence, that is, (·||·) is a vector concatenation operation. Vector Contains network IP site v i Outbound traffic requirements to all other network IP sites and other network IP site-to-site v i The inbound traffic demand, vector Contains the physical links associated with the current site Capacity constraint information of the site, such as i The minimum transmission capacity to other IP sites, that is Δi,j For site v i and site v j The set of physical links on the shortest path between them, N is the total number of IP sites in the network; vector Contains site v i Related network planning problems optimize target-related parameters, such as site v i Unit cabling cost or unit bandwidth capacity operating cost to other sites.

[0074] S102, establish network topology solution The conversion relationship between the graph adjacency matrix and the link capacity in both directions of any IP chain is equal. The relationship between the IP link capacity value and the elements in the adjacency matrix A is expressed as:

[0075]

[0076] Among them, z norm Represents a scaling factor, which changes the value A e =A ij =A ji Normalized to the interval [0,1].

[0077] The graph signal modeling method for the communication network planning problem in the above embodiment converts the solution process of the network topology planning problem into a graph sampling generation process. The signal modeling method can effectively alleviate the problem of graph signal data dimension growth under large-scale IP network sites and link numbers, and improve the scalability of the method to network scale.

[0078] In order to obtain multiple high-quality network topology feasible solutions at once with low computational complexity and improve the efficiency of solving network planning problems, in a preferred embodiment of the present invention, the process of S200 is divided into an offline part and an online solution part. Specifically, S200 in this embodiment can adopt the following steps:

[0079] S201, uses an offline trained deep graph neural network-based diffusion generation score function s θ , get the score function s θ The output value of

[0080] S202, based on the network planning graph signal X and the score function, parallel sampling generates M graph adjacency matrix samples, and converts the generated adjacency matrix samples into M network topology solutions in combination with the conversion relationship.

[0081] S203, will As multiple initial solutions to the network planning problem, we add variable search space constraints on the original network planning problem, perform local adjustment and optimization, and obtain the optimal network topology solution C that meets feasibility.★ .

[0082] Furthermore, in a preferred embodiment, a training process of a diffusion-generated score function based on a deep graph neural network is provided:

[0083] First, we collect the existing communication network planning problem instance parameters and high-quality network topology solutions obtained by any solution method. The problem instance parameters include the IP layer topology site set V, the IP link set ε, the physical link set physical link set, and the physical link design capacity S. u , IP layer site traffic requirements And the overall network cost function U(C), the network topology solution is the IP link design capacity C.

[0084] Then, through step S100, an offline training set of graph signals is constructed.

[0085] Finally, use the offline s to score function s θ The neural network model parameters are θ for training and updating to obtain the optimal parameter values.

[0086] In order to make the obtained model parameter θ more accurate, some specific embodiments further provide a better training process.

[0087] S2011, set the positive weight function λ(t) of the diffusion generation model, the mean α(t) and variance σ(t) of the Gaussian distribution of the noise process, the standard Gaussian noise ∈, and the maximum value T of the diffusion noise process time step t;

[0088] S2012, from offline training set Sample (X, A) in the middle, sample the diffusion noise process time step t from the interval 0, T];

[0089] S2013, the sampled adjacency matrix sample A is noised. The noise adding process is represented as A after noise addition from a Gaussian random probability distribution. t Sampling of values:

[0090] p(A t |A)=N(A t |α(t)A0,σ 2 (t)I),

[0091] Here, I is the identity matrix of the same dimension as A, and N(x|·,·) represents a Gaussian random distribution.

[0092] 2014, using noised samples Training score function s θ Parameter θ, the training loss function is:

[0093]

[0094] S2015, repeat steps S2012-S2014 until the parameter θ training converges.

[0095] The offline training process of the diffusion generative model in the above-mentioned embodiment can learn the graph structure probability distribution of high-quality network topology solutions from existing experience in solving network planning problems, and effectively train a graph diffusion generative model scoring function that can extract the characteristics of this distribution. The scoring function used can be constructed using any neural network model architecture, and the parameter training loss function is simple, making the training process stable.

[0096] It is worth noting that the present invention does not restrict the type of diffusion generative model deployed. Diffusion generative models based on continuous partial differential equations or discrete denoising diffusion probability generative models can be deployed. Furthermore, the present invention does not restrict the method for constructing the scoring function of the diffusion generative model deployed. Score function prediction models based on any neural network architecture can be deployed.

[0097] After training, the score function s is obtained θ The neural network model parameter is θ, and the input is the original or noisy graph representation and time step t, where A t is the adjacency matrix after noise addition. The noise addition process is determined by the forward noise addition process of the general diffusion generation model. The output is the graph representation of the diffusion generation process. Gradient of the log probability density function with respect to the adjacency matrix The fitted value of

[0098] Furthermore, in a preferred embodiment, S202 may adopt the following steps:

[0099] S2021: Sample M initial graph adjacency matrices with the same dimension as A from random noise And obtain the time coordinate set of the solution of the inverse stochastic partial differential equation (SDE) representing the diffusion generation process Arrange in descending order;

[0100] S2022: Parallel sampling generation process from Starting from the largest time coordinate, for all In parallel, Solve the reverse SDE at all time coordinates in and obtain the adjacency matrix at each time coordinate The reverse SDE formula is expressed as:

[0101]

[0102] Among them, the function f(A t ,t) and g(t) are the outputs of the offset coefficient and diffusion coefficient of the reverse diffusion generation process respectively, w is the random noise, E(X,A t ,y) is the energy function in the back diffusion generation process, y is the artificially defined adjacency matrix sample expected attribute, and the scalar γ is the weight coefficient of the energy function.

[0103] S2023: Solve the reverse SDE at time coordinate t=0 As the adjacency matrix samples generated by the controllable graph diffusion generation process, and converted into M network topology initial solutions

[0104] In a specific embodiment, the energy function used in S2022 may be E(X,A t ,y), including a predictor f(A t ,X), the input is the graph signal X of the network planning problem and the corresponding adjacency matrix sample A t , the output is sample A t For the predicted scalar or vector value of a specific attribute, the energy function is specifically expressed as E(X,A t ,y)=||f(A t ,X)-y|| 2 , where y is the expected attribute label value of the adjacency matrix sample defined artificially.

[0105] The present invention does not limit the specific formula form of the deployed energy function. In other embodiments, any function form that can output an adjacency matrix to generate sample attributes can be arbitrarily adopted.

[0106] In a specific implementation, the conversion method of the adjacency matrix sample and the network topology initial solution in S2023 can adopt the following formula:

[0107]

[0108] Among them, the scalar δ is the preset IP link bandwidth capacity unit adjustment interval, The operator returns the largest integer value less than or equal to x.

[0109] The present invention does not limit the conversion method of the deployed adjacency matrix samples and the network topology initial solution. In other embodiments, any conversion formula can be used.

[0110] The controllable online topology sample generation process in the above-mentioned embodiment, as a key step in solving network planning problems, can rapidly generate graph adjacency matrix samples corresponding to high-performance network topology solutions based on graph signals corresponding to network planning problem parameters. This generation process can be performed in parallel for multiple different network planning problems, i.e., solving multiple different network planning problems simultaneously. Furthermore, by presetting an energy function with custom target attributes, the generation process can controllably generate a graph adjacency matrix with predefined attributes.

[0111] Furthermore, in a preferred embodiment, S23 As multiple initial solutions to the network planning problem, we add variable search space constraints on the original network planning problem P1, perform local adjustment and optimization, and obtain the optimal network topology solution C that meets feasibility. ＊ This can be achieved by taking the following steps:

[0112] S2031, for any network planning problem topology initial solution generated by online topology sample controllable generation module By adding an initial solution to the original network planning problem The corresponding variable search space constraints form the network planning local adjustment optimization problem P2, and the specific formula is as follows:

[0113] P2:min C,D(e, ω)U(C)

[0114]

[0115] in, is the optimization variable of the original network planning problem, and the variable μ≥1 is the relaxation factor for constructing the local search space of the optimization variable. The variable search space constraint constrains the optimization variable search space of the original network planning problem to the initial solution. within the neighborhood range.

[0116] Solve the optimization problem P2 using any commercial mixed integer programming solver, such as Gurobi and SCIP software, and obtain a suboptimal solution after local adjustment.

[0117] S2032: Initial solution for M topologies Execute the local optimization operations described in S2031-S2032 to obtain no more than M suboptimal solutions that meet all the constraints of the optimization problem. Then, calculate all suboptimal solutions The corresponding optimization target value, select the suboptimal solution with the minimum or maximum target value As the optimal network topology solution C ＊ .

[0118] The online topology feasible solution fine-tuning process described in the above embodiment, as an auxiliary process for the controllable generation of online topology samples, can locally adjust the sampled topology solutions to convert them into feasible topology solutions for the network planning problem, and then select the optimal solution from among them. The local adjustment method employed is simple in principle and can be used with any commercial mixed-integer programming problem-solving software to ensure the feasibility of the proposed method in obtaining topological solutions.

[0119] Based on the same inventive concept, in another embodiment of the present invention, a communication network topology planning system is provided, such as Figure 2 As shown, it includes two modules: a network planning problem graph signal modeling module S1 and a solution module S2 based on a graph diffusion generation model.

[0120] Network planning problem graph signal modeling module S1: Model the network planning problem instance parameters as network planning graph signal X; and establish the network topology solution The conversion relationship between the graph adjacency matrix and the graph;

[0121] Solving module S2 based on graph diffusion generation model: inputs the obtained network planning signal X into the communication network planning problem solver based on graph diffusion generation model, and obtains the optimal solution C of the network planning problem by combining the conversion relationship. ★ .

[0122] The solution module based on the graph diffusion generation model includes an offline training phase and an online generative solution phase.

[0123] The offline training phase corresponds to an offline training module S21 for the diffusion generative model. An offline training set is constructed using a solution module based on the graph diffusion generative model. This module first constructs a diffusion generative scoring function based on a deep graph neural network to approximate the gradient of the log probability density function of the graph signal with respect to the adjacency matrix during the diffusion generative process. The model parameters of the scoring function are then trained using the offline training set obtained by the solution module based on the graph diffusion generative model.

[0124] The online generative solution phase involves the online controllable topology sample generation module S22 and the online topology feasible solution fine-tuning module S23. Given any given instance of the network planning problem to be solved, the graph signal corresponding to that instance is input into the online controllable topology sample generation module S22 as a scoring function. Parallel sampling generates M graph adjacency matrix samples, which are then converted into M network topology solutions. The online topology feasible solution fine-tuning module S23 adjusts and optimizes the network topology solution to obtain the optimal solution.

[0125] This embodiment assists in finding the optimal network topology for communication network design and planning solutions, utilizing a generative solution approach to effectively improve the speed of solving network planning problems without sacrificing performance. This facilitates real-time network design and planning adjustments for communication networks encompassing large-scale communication sites, reducing operational and construction costs, and improving the overall quality of service for communication network traffic transmission tasks.

[0126] The specific implementation techniques of the modules / units in the above examples of the present invention may refer to the corresponding steps of the generative solution method for the telecommunication network topology planning problem in the above embodiments, which will not be described in detail here.

[0127] The above describes specific embodiments of the present invention. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art may make various modifications or variations within the scope of the claims without affecting the essence of the present invention. The above preferred features may be used in any combination as long as they do not conflict with each other.

Claims

1. A communication network topology planning method, characterized in that: include: Model the network planning problem instance parameters as a network planning graph signal X; and establish a network topology solution The conversion relationship between the graph adjacency matrix A; The obtained network planning signal X is input into the communication network planning problem solver based on the graph diffusion generation model, and the optimal solution C of the network planning problem is obtained by combining the conversion relationship. ★ ; The network planning problem instance parameters are modeled as a network planning graph signal X, including: Obtain network planning problem instance parameters, including: IP layer topology site set V, IP link set ε, physical link set Physical link design capacity S u , IP layer site traffic requirements and the network overall cost function U(C); The network planning problem instance parameters are modeled as a network planning graph signal X = [X0, ..., X i ,…,X N-1 ] T , the specific modeling process is: element X i Depend on The three vectors are concatenated in sequence, where the vector Contains IP layer network site v i Outbound traffic requirements to all other IP layer network sites and other IP site-to-site v i The inbound traffic demand, vector Contains the current IP layer network site v i Related physical links Design capacity S u Related information, vector Contains the network planning problem optimization objective U(C) and the IP layer network site v i Parameters related to the available IP layer and physical layer network links; The establishment of network topology solution The conversion relationship between the graph adjacency matrix A includes: Assuming that the link capacities in both directions of any IP chain are equal, the network topology solution C is converted into the adjacency matrix A as follows: Among them, A ij is an element in the adjacency matrix A, is the IP link capacity value in the topology solution, z norm Represents a scaling factor, which changes the value A e =A ij =A ji Normalized to the interval [0,1]; Conversely, the conversion relationship of the adjacency matrix A to the network topology solution C is expressed as: Among them, the scalar δ is the preset IP link bandwidth capacity unit adjustment interval, The operator returns the largest integer value less than or equal to x; The obtained network planning signal X is input into the communication network planning problem solver based on the graph diffusion generation model, and the optimal solution C of the network planning problem is obtained by combining the conversion relationship. ★ ,include: Use an offline trained graph diffusion generation model score function s θ , get the score function s θ The output value of Based on the network planning graph signal X and the score function, M graph adjacency matrix samples are generated by parallel sampling, and the generated adjacency matrix samples are combined with the conversion relationship to convert into M network topology solutions. Will As multiple initial solutions to the network planning problem, we add variable search space constraints on the original network planning problem, perform local adjustment and optimization, and obtain the optimal network topology solution C that meets feasibility. ★ ; The graph diffusion generative model score function s θ , a deep graph neural network based on the graph attention mechanism is constructed, with a model parameter of θ and an input of the original or noisy graph representation and time step t, where A t is the adjacency matrix after noise addition. The noise addition process is determined by the forward noise addition process of the general diffusion generation model. The output is the graph representation of the diffusion generation process. Gradient of the log probability density function with respect to the adjacency matrix The fitted value of The generating M graph adjacency matrix samples by parallel sampling based on the network planning graph signal X and the score function includes: Sample M N×N-dimensional initial graph adjacency matrices from random noise Where N is the number of IP layer network sites, and the time coordinate set of the reverse stochastic partial differential equation SDE solution representing the graph diffusion generation process is obtained Arrange in descending order; The parallel sampling generation process starts from Starting from the largest time coordinate, for all In parallel, Solve the reverse SDE at all time coordinates in and obtain the adjacency matrix at each time coordinate The reverse SDE formula is expressed as: Among them, the function f(A t ,t) and g(t) are the outputs of the offset coefficient and diffusion coefficient of the reverse diffusion generation process respectively, w is the random noise, E(X,A t ,y) is the energy function in the back-diffusion generation process, y is the artificially defined adjacency matrix sample expected attribute, and the scalar γ is the weight coefficient of the energy function; Solve the reverse SDE at time coordinate t = 0 As the adjacency matrix samples generated by the parallel sampling generation process, based on the network topology solution The conversion relationship between the graph adjacency matrix A and the M adjacency matrix samples generated by parallel sampling is converted into M network topology solutions Wherein, the energy function E(X,A t ,y) is specifically expressed as, E(X,A t ,y)=||f(A t ,X)-y|| 2 Where f(A t ,X) represents a predictor used to evaluate the specific properties of the graph adjacency matrix sample, and the input is the graph signal X of the network planning problem and the corresponding adjacency matrix sample A t , the output is sample A t The predicted scalar or vector value for a specific attribute; y is the expected attribute label value of the sample in the artificially defined adjacency matrix.

2. A communication network topology planning method according to claim 1, characterized in that: The graph diffusion generative model score function s θ The offline training process is as follows: Collect existing communication network planning problem instance parameters and network topology solutions obtained by any solution method; Convert all problem instance parameters into network planning diagram signals; Collect all network planning map signals and corresponding network topology solutions to construct an offline training set; Establish the scoring function s θ The training loss function of parameter θ is: Among them, λ(t) is a positive weight function, σ(t) represents the variance of the Gaussian distribution of the noise addition process, ∈ is a standard Gaussian noise sample, and T is the maximum value of the set time step t.

3. A communication network topology planning method according to claim 1, characterized in that: The said As multiple initial solutions to the network planning problem, we add variable search space constraints on the original network planning problem, perform local adjustment and optimization, and obtain the optimal network topology solution C that meets feasibility. ★ ,include: The initial topological solution of any network planning problem By adding an initial solution to the original network planning problem The corresponding variable search space constraints form a local adjustment optimization problem for network planning; the local adjustment optimization problem is solved by any commercial mixed integer programming problem solver to obtain a suboptimal solution after local adjustment. For M topological initial solutions Perform the above local optimization operations to obtain no more than M suboptimal solutions that meet all the constraints of the optimization problem. Then, calculate all suboptimal solutions The corresponding optimization target value, select the suboptimal solution with the minimum or maximum target value As the optimal network topology solution C ★ .

4. A communication network topology planning method according to claim 3, characterized in that: For any network planning problem, the initial topology solution The constraint condition of the variable search space constrains the optimization variable search space of the original network planning problem to the initial solution. within the neighborhood of ; The variable search space constraints are specifically: in, is the optimization variable of the original network planning problem, and the variable μ ≥ 1 is the relaxation factor for constructing the local search space of the optimization variable.

5. A communication network topology planning system, configured to execute a communication network topology planning method according to any one of claims 1 to 4, characterized in that: include: Communication network planning problem graph signal modeling module: Modeling network planning problem instance parameters as network planning graph signal X; and establish a network topology solution The conversion relationship between the graph adjacency matrix and the graph; Solving module based on graph diffusion generation model: input the obtained network planning signal X into the communication network planning problem solver based on graph diffusion generation model, and combine the conversion relationship to obtain the optimal solution C of the network planning problem ★ .

Images