Wireless strategy optimization method and device

Abstract

A wireless policy optimization method comprises: obtaining a graph representation of a wireless communication network, the graph representation comprising a plurality of nodes and an edge connecting the plurality of nodes, the plurality of nodes comprising a first node and a second node, the first node and the second node being connected through the edge; taking the graph representation as the input of the graph neural network, and outputting a final feature vector of an edge; guiding optimization of a wireless strategy based on the final feature vector of the edge; the structure of the graph neural network is constructed based on Taylor expansion guidance of matrix inversion operation, and the problem that the generalization performance of a traditional GNN is poor is solved.

Description

A wireless strategy optimization method and device Technical Field

[0001] The present application relates to the field of communication technology, and in particular to a method and device for optimizing a wireless strategy. Background Art

[0002] Wireless policy optimization (such as resource allocation, channel estimation, and signal detection) is a critical issue in wireless communications. Taking precoding as an example, precoding in wireless communication systems can effectively improve system energy and spectral efficiency. Although existing literature has proposed iterative algorithms to solve non-convex precoding optimization problems, the resulting numerical solutions are sensitive to practical system imperfections (such as channel estimation errors). Another drawback of traditional optimization methods is the high computational complexity when the problem scale is large, which limits their application in networks with real-time requirements. Existing literature has also proposed using deep neural networks (DNNs) or graph neural networks (GNNs) to solve wireless policy optimization problems. However, typical DNNs often lack dimensionality generalization. When the dimensionality of the problem changes with the number of participating terminal devices and network devices, DNN methods require retraining the neural network model, which incurs significant computational overhead. While traditional GNNs can achieve dimensionality generalization when solving wireless policy optimization problems, they indiscriminately consider the interactions between different edges on the graph, resulting in poor performance.

[0003] Summary of the Invention

[0004] The embodiments of the present application provide a method and device for optimizing a wireless strategy, which solves the problem of poor performance of traditional GNN.

[0005] In the first aspect, the present application provides a method for optimizing a wireless policy, comprising: obtaining a graph representation of a wireless communication network, the graph representation comprising multiple nodes and edges connecting the multiple nodes, wherein the multiple nodes include at least a first node and a second node, and the edge connects the first node and the second node; using the graph representation as the input of a graph neural network, updating the first eigenvector of the edge through a multi-layer feature update network in the graph neural network, and outputting the final first eigenvector of the edge; obtaining a solution to the wireless policy optimization problem based on the final first eigenvector of the edge; wherein the first eigenvector of the edge output by each layer of the feature update network in the multi-layer feature update network is related to a first weight parameter, and the first weight parameter is determined based on the initial first eigenvector of the first adjacent edge of the edge and the first eigenvector of the first adjacent edge output by the feature update network of the previous layer of the feature update network of each layer, and the first adjacent edge is an edge that shares a first node with the edge.

[0006] The graph neural network in the wireless strategy optimization method provided in the embodiment of the present application has good scalability and generalization performance, which solves the problem that traditional GNN has limited expression ability and can only learn element-by-element functions when the problem scale is large.

[0007] In one possible implementation, a data processing network is further provided between adjacent feature update networks in a multi-layer feature update network; the data processing network is used to obtain the second feature vector of the edge based on the first feature vector of the edge, the first feature vector of the first adjacent edge, and the first weight parameter output by the feature update network of the previous layer of the feature update network of each layer; the second feature vector of the edge is used as the input of the feature update network of each layer, and the feature update network of each layer obtains the updated first feature vector of the edge based on the second feature vector of the edge, the second feature vector of the first adjacent edge of the edge, and the second feature vector of the second adjacent edge of the edge; the parameters of the feature update network of each layer are updated during the training process of the graph neural network, and the second adjacent edge is the edge that shares the second node with the edge.

[0008] In another possible implementation, each layer of feature update network obtains the updated first eigenvector of the edge based on the first eigenvector of the edge, the first eigenvector of the first adjacent edge of the edge, and the first eigenvector of the second adjacent edge of the edge output by the feature update network of the previous layer of feature update network; the second adjacent edge of the edge is the edge that shares the second node with the edge; the parameters of the feature update network of each layer are related to the first weight parameter.

[0009] In another possible implementation, the initial first eigenvectors of the edges of the graph representation are determined based on a channel matrix of the wireless communication network, and the final first eigenvectors of the edges guide the determination of precoding of the wireless communication network.

[0010] In another possible implementation, the initial first eigenvectors of the edges of the graph representation are determined based on pilot signals of the wireless communication network, and the final first eigenvectors of the edges guide determination of channel estimation of the wireless communication network.

[0011] In another possible implementation, the initial first eigenvectors of the edges of the graph representation are determined based on a channel matrix of the wireless communication network, and the final first eigenvectors of the edges guide the determination of a MIMO detector of the wireless communication network.

[0012] In one possible implementation, the first node and the second node are nodes of different categories.

[0013] In one example, the graph neural network is obtained based on supervised training or unsupervised training.

[0014] In the second aspect, the present application provides a wireless policy optimization device including an acquisition module, an inference module and an optimization module: wherein the acquisition module is used to obtain a graph representation of the wireless communication network, the graph representation including multiple nodes and edges connecting the multiple nodes, wherein the multiple nodes include at least a first node and a second node, and the edge connects the first node and the second node; the inference module is used to take the graph representation as the input of the graph neural network, update the first eigenvector of the edge through the multi-layer feature update network in the graph neural network, and output the final first eigenvector of the edge; the optimization module is used to obtain a solution to the wireless policy optimization problem based on the final first eigenvector of the edge; wherein the first eigenvector of the edge output by each layer of the feature update network in the multi-layer feature update network is related to a first weight parameter, and the first weight parameter is determined based on the initial first eigenvector of the first adjacent edge of the edge and the first eigenvector of the first adjacent edge output by the feature update network of the previous layer of the feature update network of each layer, and the first adjacent edge is an edge that shares the first node with the edge.

[0015] In one possible implementation, a data processing network is further provided between adjacent feature update networks in a multi-layer feature update network; the data processing network is used to obtain the second feature vector of the edge based on the first feature vector of the edge, the first feature vector of the first adjacent edge, and the first weight parameter output by the feature update network of the previous layer of the feature update network of each layer; the second feature vector of the edge is used as the input of the feature update network of each layer, and the feature update network of each layer obtains the updated first feature vector of the edge based on the second feature vector of the edge, the second feature vector of the first adjacent edge of the edge, and the second feature vector of the second adjacent edge of the edge; the parameters of the feature update network of each layer are updated during the training process of the graph neural network, and the second adjacent edge is the edge that shares the second node with the edge.

[0016] In another possible implementation, each layer of feature update network obtains the updated first eigenvector of the edge based on the first eigenvector of the edge, the first eigenvector of the first adjacent edge of the edge, and the first eigenvector of the second adjacent edge of the edge output by the feature update network of the previous layer of feature update network; the second adjacent edge of the edge is the edge that shares the second node with the edge; the parameters of the feature update network of each layer are related to the first weight parameter.

[0017] In another possible implementation, the initial first eigenvectors of the edges of the graph representation are determined based on a channel matrix of the wireless communication network, and the final first eigenvectors of the edges guide the determination of precoding of the wireless communication network.

[0018] In another possible implementation, the initial first eigenvectors of the edges of the graph representation are determined based on pilot signals of the wireless communication network, and the final first eigenvectors of the edges guide determination of channel estimation of the wireless communication network.

[0019] In another possible implementation, the initial first eigenvectors of the edges of the graph representation are determined based on a channel matrix of the wireless communication network, and the final first eigenvectors of the edges guide the determination of a MIMO detector of the wireless communication network.

[0020] In one possible implementation, the first node and the second node are nodes of different categories.

[0021] In another possible implementation, the graph neural network is obtained based on supervised training or unsupervised training.

[0022] In a third aspect, the present application provides an electronic device comprising a memory and a processor, wherein the memory stores executable code, and the processor executes the executable code to implement the method provided in the first aspect of the present application.

[0023] In a fourth aspect, the present application provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed in a computer, the computer is caused to execute the method provided in the first aspect of the present application.

[0024] In a fifth aspect, the present application provides a computer program or a computer program product, wherein the computer program or computer program product includes instructions, and when the instructions are executed, the method provided in the first aspect of the present application is implemented. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] FIG1 is a schematic diagram of the structure of a fully connected neural network;

[0026] Figure 2 is a schematic diagram of a loss function optimization;

[0027] Figure 3 is a schematic diagram of gradient back propagation;

[0028] Figure 4 is a schematic diagram of edge feature update of a traditional GNN;

[0029] FIG5 is a schematic diagram of a wireless communication network;

[0030] FIG6 is a schematic diagram of another wireless communication network;

[0031] Figure 7 is a schematic diagram of edge feature update of a model GNN;

[0032] FIG8 is a flow chart of a method for optimizing a wireless strategy according to an embodiment of the present application;

[0033] FIG9 is a schematic structural diagram of a wireless strategy optimization device;

[0034] FIG10 is a schematic diagram of the structure of a computing device. DETAILED DESCRIPTION

[0035] The technical solution of the present application is further described in detail below through the accompanying drawings and examples.

[0036] Machine learning is an important research direction in the field of artificial intelligence, and research on neural networks has been a hot topic in recent years. The following uses a fully connected neural network as an example to briefly introduce the neural network training process that may be involved in the embodiments of this application.

[0037] A fully connected neural network, also known as a multilayer perceptron (MLP), is shown in Figure 1. An MLP consists of an input layer (left), an output layer (right), and multiple hidden layers (center). Each layer contains several nodes, called neurons. Neurons in adjacent layers are connected to each other.

[0038] Considering two adjacent layers of neurons, the output h of the neurons in the next layer is the weighted sum of all the neurons x in the previous layer connected to it and then passes through the activation function. It can be expressed as a matrix

[0039] h=f(wx+b)

[0040] Where w is the weight matrix, b is the bias vector, and f is the activation function. The output of the neural network can be recursively expressed as

[0041] y=f n (w n f n-1 (…)+b n )

[0042] Simply put, a neural network can be understood as a mapping relationship from an input data set to an output data set. Neural networks are usually randomly initialized, and the process of obtaining this mapping relationship from random w and b using existing data is called neural network training.

[0043] The specific training method is to use the loss function to evaluate the output of the neural network and backpropagate the error. Through the gradient descent method, w and b can be iteratively optimized until the loss function reaches the minimum value (see Figure 2).

[0044] The process of gradient descent can be expressed as: Where θ is the parameter to be optimized (such as w and b), L is the loss function, and η is the learning rate, which controls the step size of the gradient descent.

[0045] The back propagation process uses the chain rule for partial derivatives, that is, the gradient of the previous layer parameters can be recursively calculated from the gradient of the next layer parameters (see Figure 3). The formula can be expressed as where w ijis the weight of node j connecting to node i, s i is the weighted sum of the inputs to node i.

[0046] Deep neural networks (DNNs) are used to learn resource allocation strategies, including power control, link scheduling, and precoding. Deep learning methods can jointly optimize multiple strategies, and the learned strategies are robust to channel estimation errors and have low computational complexity. However, training DNNs requires a high amount of time, samples, and training parameters. For traditional fully connected neural networks (FNNs), even with a sufficiently high training complexity, training performance may still be poor. Even if a DNN can achieve good performance on a training set, it cannot guarantee its ability to generalize across various parameters. Because the wireless environment changes dynamically over time, it is important to improve training efficiency by enhancing the generalization ability of the DNN or reducing the training complexity of the DNN. By reducing training complexity, the DNN can be retrained using newly collected samples when parameters that cannot be generalized well change significantly.

[0047] One way to improve learning efficiency is to incorporate prior knowledge when designing the DNN structure. This can reduce the hypothesis space the DNN needs to search for optimal parameters. For example, by leveraging the permutation equivariance property common to many wireless tasks, permutation-equivariant neural networks or graph neural networks (GNNs) can be designed to learn power allocation or precoding strategies. While the training complexity of these DNNs can be significantly reduced compared to unstructured FNNs, they still require a large number of samples to learn policies, even for moderately large problems. These policies involve complex, nonlinear operations such as matrix inversion. GNN learning performance also decreases with increasing sample size, suggesting that GNNs cannot generalize to different problem sizes. To improve learning efficiency by simplifying the mappings that the DNN must learn, a mathematical model can be introduced into the DNN structure. This can be an iterative algorithm or a mathematical expression. For example, deep unrolling methods use DNNs to mimic the update process of an iterative algorithm, where only a portion of the algorithm's operations or parameters need to be learned.

[0048] We will use the multi-user precoding solution as an example to introduce the solution of the embodiment of the present application. Of course, the solution provided by the embodiment of the present application can also be used to solve other problems (such as channel estimation, MIMO (multiple-in multipleout) detection, etc.), which will be introduced in the embodiment.

[0049] The system model of the single-cell multi-user precoding problem considered in the embodiment of the present application is as follows: Consider a base station with N antennas, which will serve K single-antenna users within its coverage area. The channel between the nth base station antenna and the kth user is denoted as h nk , then the channel from all base station antennas to user k is h k =[h 1k ,h 2k ,…,h Nk ], and the channel between the base station antenna and all users is H = [h1,h2,…,h K ]. The optimal precoding matrix V can be modeled as:

[0050] P1:

[0051] stTr(V H V)≤P max

[0052] in, is the throughput of the kth user, v k =[v 1k ,v 2k ,…,v Nk ] is the precoding vector used by the base station to send data to user k. All precoding vectors are combined to form the precoding matrix V = [v1,v2,…,v K ], P max is the maximum transmit power of the base station, (·) H represents the conjugate transpose operation, Tr(·) represents the trace operation, is the noise power.

[0053] Edge aggregation graph neural network can be used to model and solve the precoding problem. As shown in Figure 4, consider the precoding solution problem with N=4 and K=3. Figure 4 includes two types of nodes: base station antenna nodes and user nodes, and the edge features (or edge representation) can be obtained through multi-layer edge aggregation GNN. The output of the layer is recorded as It can be obtained by the following two steps.

[0054] Aggregation: For the adjacent edges connected to the edge (n, k) by the same vertex, their outputs in the previous layer are aggregated by an aggregation function. The aggregated output of the adjacent edges connected to (n, k) by user k and antenna n is:

[0055]

[0056] Among them PL u (·) and PL b(·) represents the aggregation and pooling function of the adjacent edges (n, k) connected by user k and antenna n, respectively. and Contains trainable parameters and A function for extracting useful information from the adjacent edges connecting user k and antenna n.

[0057] Combine: To get the output of edge (n,k), use a combine function to combine the edge (n,k) in the first Layer output and aggregate output:

[0058]

[0059] Where CB(·) is the parameter containing trainable The combination function of .

[0060] In order to ensure that the permutation of similar vertices (such as users or antennas) does not affect the output of GNN, and It is the same for different edges, and the pooling function PL u (·) and PL b (·) satisfies the commutative law, such as summing or maximizing. For the sake of simplicity, assume that q u (·) and q b (·) are the same, and PL u (·) and PL b (·) is the same. Therefore, the subscripts u and b can be omitted. At the same time, it is assumed that q(·), CB(·), PL(·) and the trainable parameters remain the same between different layers, so the above three formulas and the following Superscript Can be omitted.

[0061] By substituting formula (1) into formula (2), we can get the formula for updating edge (n, k) in the first The update equation of the layer is:

[0062]

[0063] When q(·) is a linear function and CB(·) is a function composed of a linear equation and an activation function σ(·), Equation (3) degenerates into:

[0064]

[0065] When the summation function is used as the pooling function, that is, PL(·) = ∑(·), Equation (4) can be written as

[0066]

[0067] The input of the GNN shown in Figure 4 (i.e. the initial value of the edge feature ) is the channel matrix H, the output (i.e. the edge features after L rounds of iteration ) is the learned precoding matrix The input-output relationship of GNN is recorded as in are all the trainable parameters in the edge-aggregated GNN.

[0068] However, when GNN uses sum pooling (i.e. Find the average pooling (i.e. ) or find the maximum pooling (i.e. ) function (GNNs using these functions are called traditional GNNs), then GNNs can only learn element-by-element functions when N and K are large. We call the function V = f(H) an element-by-element function. If v nk Depends only on h nk At this time, the traditional GNN can only learn the element-by-element precoding strategy. However, under high SNR conditions, the optimal precoding strategy is not an element-by-element function of the channel, so the traditional GNN cannot learn the optimal precoding strategy at this time.

[0069] Another approach involves combining the model with a deep unrolled network (DNN) to learn precoding strategies that maximize energy efficiency. This deep unrolled network consists of multiple layers, each of which outputs an updated precoding strategy. Each layer corresponds to one iteration of the WMMSE algorithm (i.e., the mathematical model of the WMMSE algorithm is incorporated into the DNN), requiring only a small number of parameters or operations to be trained. Numerical results demonstrate that the proposed method achieves near-optimal performance in learning spectrally efficient precoding strategies.

[0070] However, the application of this solution is limited by the depth of the algorithm developed. For example, the WMMSE algorithm can only solve problems with the maximum data rate-dependent utility function under power constraints. It cannot solve problems with other constraints (such as data rate constraints). Furthermore, due to the introduction of an iterative structure, this technology has long training and inference times. This method also has poor scalability. Models trained for a given number of base station antennas and users cannot be directly applied to scenarios with different numbers of base station antennas and users.

[0071] This embodiment of the present application proposes a model-driven GNN structure (hereinafter referred to as model GNN) with good scalability and generalization performance. It addresses the limited expressiveness of traditional GNNs and their ability to only learn element-by-element functions for large-scale problems. The model-driven GNN structure proposed in this embodiment of the application can be used to solve problems such as precoding matrix solution, channel estimation, and MIMO detector design.

[0072] The solution provided in the embodiments of this application can be applied to wireless communication systems such as 5G, satellite communications, and short-range communications. The system architecture is shown in Figure 5. A wireless communication system can be composed of cells, each of which contains a base station (BS), which provides communication services to multiple mobile stations (MS). The wireless communication system can also perform point-to-point communication, such as communication between multiple terminals.

[0073] It should be noted that the wireless communication systems mentioned in the embodiments of the present application include but are not limited to: narrowband Internet of Things (NB-IoT) systems, global system for mobile communications (GSM), enhanced data rate for GSM evolution (EDGE), wideband code division multiple access (WCDMA), code division multiple access 2000 (CDMA2000), time division synchronous code division multiple access (TD-SCDMA), long term evolution (LTE), and the three major application scenarios of the next generation 5G mobile communication system, eMBB, URLLC and eMTC.

[0074] Another typical application scenario of the embodiments of the present application is a wireless smart home or smart factory scenario. In a wireless smart home scenario, different smart home products (such as speakers, TVs, refrigerators, kitchen appliances, sweeping robots, etc.) are connected via a wireless network. Figure 6 is a schematic diagram of a mobile phone and a smart TV connected wirelessly for wireless screen projection. In a smart factory scenario, different devices (such as smart robots, lathes, transport vehicles, etc.) are connected via a wireless network.

[0075] The distributed nodes involved in the embodiments of the present application may include various handheld devices with wireless communication capabilities, vehicle-mounted devices, wearable devices, computing devices, or other processing devices connected to a wireless modem. The distributed nodes may also be referred to as terminals, and may also be subscriber units, cellular phones, smart phones, wireless data cards, personal digital assistants (PDAs), tablet computers, wireless modems, handheld devices (handsets), laptop computers, machine type communication (MTC) terminals, etc.

[0076] The following is a detailed introduction to the GNN network structure involved in the embodiments of this application.

[0077] In wireless communication systems, numerous algorithms rely on inversion (or pseudo-inversion) operations, such as precoding matrix calculation, channel estimation, and MIMO detection algorithms. Learning these algorithms using neural networks requires sufficient expressive power to accurately approximate matrix inversion. Therefore, we design the GNN architecture based on matrix inversion.

[0078] Let the pseudo-inverse of matrix H be H + . H + The first-order Taylor expansion at H0 is:

[0079]

[0080] The result of Taylor approximation can be used as H + is an approximation of Closer to H + In order to obtain a more accurate H + Approximation of H can be iteratively updated + , where in each iteration (such as In the iterations), use The result of the step approximation is expanded, that is

[0081]

[0082] in It is H + Approximation, D (0) Random initial point.

[0083] It can be proved that the iterative equation derived from the first-order Taylor expansion is a compression mapping. Through more iterative steps, Can more accurately approximate H + .

[0084] remember in Therefore, formula (7) can be written as The elements in the nth row and kth column of are:

[0085]

[0086] This iterative equation can be viewed as an update equation for an edge-aggregated GNN that does not contain trainable parameters, where is the edge (n,k) in the The output of the layer, and the aggregation function and the combination function are:

[0087]

[0088] It can be seen that the weight multiplied by the adjacent edge information (such as Weight ) is different for different edges. Weight Can reflect the information on each edge for updating the approximate pseudo-inverse At the same time, it can be seen that the aggregated output of the side information adjacent to user k (i.e., in formula (1) ) is not included in the output of the aggregate function because this information is already in the calculation of the matrix Specifically, The kth column of That is the information gathered to user k, where is the kth column of the matrix H.

[0089] Based on the above derivation, two implementation methods of the GNN model can be obtained.

[0090] The two-step implementation of the GNN model. Layer edge (n,k) operation:

[0091] Step 1: Inspired by formula (8), based on Output of the layer Get Vector

[0092] Step 2: Obtained through a GNN layer The update equation of GNN is

[0093]

[0094] in S, P, and Q are trainable parameter matrices. Their dimensions are independent of N and K. Therefore, the parameter matrices trained under one set of N and K values ​​can be used in other N and K values.

[0095] Based on the above two-step implementation, we can derive the one-step implementation of the GNN model. P=[P0,P1],Q=[Q0,Q1], then formula (11) can be further written as:

[0096]

[0097] in

[0098]

[0099] In the above formula, for the sake of simplicity, the weight matrix is ​​also ignored Superscript In (12), item (a) aggregates the information of the adjacent edges connected to user k. Item (b) aggregates the information of the adjacent edges connected to antenna n, and item (c) aggregates the information of the edges that are not adjacent to edge (n, k). In order to reduce the computational complexity without losing the representation ability of GNN, item (c) can be ignored because the information of non-adjacent edges can be aggregated by stacking multiple GNN layers. Comparing the update formula (5) of the traditional GNN with the update formula (12) of the model GNN, it can be seen that the weights are taken into account in the update operation of the model GNN. For the precoding strategy, it characterizes the correlation between the channel of user i and the precoding vector of user j, so it has stronger expression ability than traditional GNN.

[0100] Therefore, the one-step implementation of the model GNN is Layer edge (n,k) operation:

[0101] Step 1: Output of the layer The input update equation is the following GNN

[0102]

[0103] in

[0104] As shown in Figure 7, the update diagram of edge (1,1) in the one-step implementation of the GNN model. In layer operations, edge features By its own upper layer characteristics The previous layer features of the edge connected to user node UE1 The previous layer feature of the edge connected to the base station antenna node AN1 Decide.

[0105] Edge (n,k) in the The output feature of the layer can be a vector rather than a scalar, that is, It contains elements. hour, So J0=2. When , the output of GNN is the learned strategy. Taking precoding as an example, its output is the learned precoding vector, that is, So J L =2. When hour, Can be greater than 2, and The larger the value, the stronger the representation ability of the model GNN. The input-output relationship of GNN is recorded as where θ V Contains all trainable parameters in the model GNN.

[0106] After obtaining the implementation of the model neural network, the parameter matrix can be trained based on the training data. The training method can be either unsupervised learning or supervised learning. Taking precoding strategy learning as an example

[0107] Unsupervised learning: based on the output of all edges of the model GNN Get the precoding matrix Calculate the total throughput of the system Find the parameter θ V The gradient of θ is updated using the gradient descent method described above. V .

[0108] Supervised learning: In supervised learning, the label V corresponding to the channel H (i.e., the target precoding matrix) will be provided * , the precoding matrix based on the model GNN output and label V * Calculate the loss function (such as mean square error) and find the parameter θ for the loss function V The gradient of θ is updated using the gradient descent method described above. V .

[0109] This application implements a method for optimizing wireless strategies, and its implementation is introduced below.

[0110] FIG8 is a flowchart of a wireless policy optimization method provided in an embodiment of the present application. The method is applied to the wireless communication scenario shown in FIG4 or FIG5. The method can be executed by any device, equipment, platform or device cluster with computing capabilities, such as the terminal and / or base station in FIG5, or the mobile phone and / or television shown in FIG6. The present application does not specifically limit the specific computing device for executing the method, and a suitable computing device can be selected as needed. As shown in FIG8, the wireless policy optimization method includes at least steps S801-S803.

[0111] In step S801 , a graph representation of a wireless communication network is obtained.

[0112] Exemplarily, the graph representation includes multiple nodes and multiple edges, where the edges are used to connect different nodes among the multiple nodes. The feature vectors of the edges in the graph representation can be determined based on the initial input parameters of the wireless strategy optimization task of the wireless communication network. For example, in the precoding task, the initial feature vector of the edge in the graph representation is the channel vector; in the channel estimation task, the initial feature vector of the edge in the graph representation is the pilot signal; in the MIMO detection task, the initial feature vector of the edge in the graph representation is the channel matrix, etc.

[0113] In another example, the graph representation can also include nodes of different categories and edges connecting nodes of different categories, such as first-category nodes, second-category nodes, and edges connecting first-category nodes and second-category nodes, as shown in Figure 4 or Figure 7. The initial eigenvector of the edge in the graph representation is the initial state parameter of the optimization task. For example, in the precoding task, the initial eigenvector of the edge in the graph representation is the channel vector; in the channel estimation task, the initial eigenvector of the edge in the graph representation is the pilot signal; and in the MIMO detection task, the initial eigenvector of the edge in the graph representation is the channel matrix.

[0114] In step S802, the graph representation is used as the input of the graph neural network, and the feature vector of the edge is updated through the multi-layer feature update network in the graph neural network, and the final feature vector of the edge is output.

[0115] The structure of the graph neural network can be found in the description of the structure of the model graph neural network above. The first weight parameter can be found in the above

[0116] In step S803 , the optimization of the wireless strategy is guided based on the final feature vector of the edge.

[0117] By optimizing the wireless strategy of the wireless communication network, the system energy efficiency and spectrum efficiency of the wireless communication network are effectively improved.

[0118] The model GNN provided in the embodiments of the present application can be used for strategy optimization of wireless communication networks, such as precoding, channel estimation, MIMO detection, etc.

[0119] For example, the model GNN is applied to precoding, for example, in a scenario where a base station with N antennas sends signals to K single-antenna users through a precoding matrix:

[0120] Input: channel vector,

[0121] Model GNN operation: two-step implementation or one-step implementation, L-layer update operation.

[0122] Output: precoding vector,

[0123] Training: Unsupervised learning or supervised learning, as described above.

[0124] An example of applying the model GNN to channel estimation is: a base station with N antennas transmits a pilot signal to K single-antenna users. The pilot signal is where N pl is the number of pilot signals sent on each antenna. The received signal is From this formula, we can see that the estimated channel can be written as In order to perform more accurate channel estimation, the model GNN can be used to learn a channel estimation result related to the pseudo-inverse.

[0125] The output of the model GNN at each layer is The output is X + The update equation of GNN is shown in formula (11). The difference from the first embodiment is that the weight The calculation is: where x j is the column vector consisting of the j-th row of matrix X, is a matrix The kth column of .

[0126] The model GNN contains L layers. Its input is X and its output is D (L) The channel estimation obtained by using the output is The model GNN can be trained in a supervised manner, and the training goal is to minimize the mean square error between the GNN output and the actual channel.

[0127] The model GNN can also perform joint optimization of joint channel estimation and related tasks (such as precoding). Precoding is related to the pseudo-inverse of the estimated channel matrix, that is,

[0128] The output of the GNN model for precoding with pseudo-inverse of the channel matrix at layer l is The output is Y +The update equation of GNN is shown in formula (11). The difference from the above is that: the weight The calculation is: where y j is the column vector formed by the j-th row of matrix Y, is a matrix The kth column of .

[0129] The model GNN contains L layers. Its input is Y and its output is D (L) The channel estimation obtained by using the output is The GNN model can be trained in a supervised or unsupervised manner. When using unsupervised training, the training loss function is the same as the loss function used above.

[0130] An example of applying the model GNN to MIMO detection is: Consider using the model GNN provided in the embodiment of the present application to design a combiner at the base station receiving end in uplink transmission. K single-antenna users transmit signals to an N-antenna base station. The transmitted signal is The channel matrix is The received signal can be expressed as y = H T x+n, where n is the variance noise.

[0131] To recover the signal at the transmitter, a combiner can be used at the receiver. Common combiners include the least squares combiner (V LS =H H (HH H ) -1 ) and MMSE combiner

[0132] In order to learn a combiner related to the pseudo-inverse of the channel matrix, the model GNN provided in the embodiment of the present application can be used. The update equation of the model GNN is shown in formula (11).

[0133] The model GNN contains L layers. Its input is H and its output is the learned combiner The model GNN can be trained in a supervised manner, and the training goal is to minimize the mean square error between the GNN received signal and the transmitted signal.

[0134] The model GNN provided in the embodiments of the present application solves the problem of poor generalization performance of traditional GNNs and can be applied to wireless strategy optimization in wireless communication systems, such as precoding, channel estimation, and MIMO detection, to obtain precoding matrices, channel estimation, and MIMO detectors with higher performance at lower complexity.

[0135] Taking precoding as an example, the beneficial effects of the model GNN provided in the embodiment of the present application are illustrated through simulation.

[0136] Consider a scenario where a base station with N antennas serves K single-antenna users. The power consumption of each antenna is P c =17.6W, the constant power consumption of each base station P0 = 43.3W, and the power amplifier efficiency 1 / ρ = 0.311.

[0137] Consider using an unsupervised approach to train DNN. The inputs in the training and test samples are generated by independent Rayleigh distribution channels, that is, the elements in H follow a complex Gaussian distribution. For each test sample, a label is also generated by a numerical algorithm as a baseline method for evaluating learning performance. Problem P1 can be solved by the weighted minimum mean squared error (WMMSE) algorithm. The number of iterations is set to 100, so that the algorithm can generate labels within an acceptable time frame. The algorithm may not converge after iterations and can only find a suboptimal solution. In the simulation, a training set containing 200,000 samples is generated (although the actual samples used for training may be fewer), and another 100 samples are generated as a test set.

[0138] The hyperparameters of GNN are shown in Table 1. In this table, the hyperparameters of each layer (such as The number of hidden nodes in the layer is the dimension of each edge output, that is, It is independent of K and N. Therefore, these hyperparameters can be used to learn policies under different K and N scenarios. Furthermore, considering using more hidden layers to learn the most energy-efficient policies, GNNs can better meet QoS constraints. The activation function of each GNN layer is σ(X) = X / |X|. GNNs are trained using the Adam algorithm.

[0139] Table 1 GNN hyperparameters

[0140]

[0141] The training performance of DNN is obtained by averaging the training results of each DNN five times to eliminate the randomness caused by sample selection and initial weights. In each training, N randomly selected samples from the training set are used. tr The DNN is trained on 100 samples and tested on 100 samples. All simulation results are obtained on a computer equipped with a 14-core Intel i9-9940X CPU, an Nvidia RTX 2080Ti GPU and 64GB of memory.

[0142] The following table shows the optimal precoding performance for problem P1 using both the model GNN and traditional GNN. For a deep unfolding network, the WMMSE algorithm is expanded into multiple layers, with each layer learning only a small number of parameters or operations. To minimize training time, the number of layers in the simulation was set to 5. The simulation used the spectral efficiency ratio (SE ratio) to evaluate learning performance. This is the ratio of the spectral efficiency achieved by each method to that achieved by the WMMSE algorithm.

[0143] Table 2 compares the spectral efficiency and inference time of various DNNs and the WMMSE algorithm for different N, K, and signal-to-noise ratios. The inference time reflects the computational complexity of the inference phase. The DNN was trained using 1000 samples. The inference time of the GNN model and the traditional GNN was obtained on a GPU. Since the code for the deep unrolled network and the WMMSE algorithm can only be implemented on a CPU, the table also evaluates the inference time of the GNN model on the CPU, making its computational complexity comparable with these two methods. It can be seen that the GNN model achieves a spectral efficiency of nearly or exceeding 95% for all N and K. In contrast, the traditional GNN has a lower spectral efficiency, especially when N and K are large and the signal-to-noise ratio is high, resulting in poor generalization performance. The inference time of the GNN model on the GPU increases only slightly with problem size, but is still lower than that of the traditional GNN. The inference time of the GNN model on the CPU increases with problem size, but remains significantly lower than that of the deep unrolled network and the WMMSE algorithm, which contain computationally complex operations (such as numerous matrix multiplications across multiple iterations).

[0144] Table 2

[0145]

[0146] Based on the same concept as the aforementioned embodiment of the wireless policy optimization method, an embodiment of the present application also provides a wireless policy optimization device 900, which includes units or modules for implementing each step in the wireless policy optimization method shown in Figures 4-8.

[0147] FIG9 is a schematic diagram of the structure of a wireless policy optimization device provided by an embodiment of the present application. The device is applied to a computing device. As shown in FIG9, the wireless policy optimization device 900 includes at least: an acquisition module 901, an inference module 902, and an optimization module 903: wherein the acquisition module 901 is used to obtain a graph representation of a wireless communication network, the graph representation including multiple nodes and edges connecting the multiple nodes, wherein the multiple nodes include at least a first node and a second node, and the edge connects the first node and the second node; the inference module 902 is used to use the graph representation as the input of a graph neural network, update the first feature vector of the edge through a multi-layer feature update network in the graph neural network, and output the final first feature vector of the edge; the optimization module 903 is used to obtain a solution to the wireless policy optimization problem based on the final first feature vector of the edge; wherein the first feature vector of the edge output by each layer of the feature update network in the multi-layer feature update network is related to a first weight parameter, and the first weight parameter is determined based on the initial first feature vector of the first adjacent edge of the edge and the first feature vector of the first adjacent edge output by the feature update network of the previous layer of the feature update network of each layer, and the first adjacent edge is an edge that shares the first node with the edge.

[0148] In one possible implementation, a data processing network is further provided between adjacent feature update networks in a multi-layer feature update network; the data processing network is used to obtain the second feature vector of the edge based on the first feature vector of the edge, the first feature vector of the first adjacent edge, and the first weight parameter output by the feature update network of the previous layer of the feature update network of each layer; the second feature vector of the edge is used as the input of the feature update network of each layer, and the feature update network of each layer obtains the updated first feature vector of the edge based on the second feature vector of the edge, the second feature vector of the first adjacent edge of the edge, and the second feature vector of the second adjacent edge of the edge; the parameters of the feature update network of each layer are updated during the training process of the graph neural network, and the second adjacent edge is the edge that shares the second node with the edge.

[0149] In another possible implementation, each layer of feature update network obtains the updated first eigenvector of the edge based on the first eigenvector of the edge, the first eigenvector of the first adjacent edge of the edge, and the first eigenvector of the second adjacent edge of the edge output by the feature update network of the previous layer of feature update network; the second adjacent edge of the edge is the edge that shares the second node with the edge; the parameters of the feature update network of each layer are related to the first weight parameter.

[0150] In another possible implementation, the initial first eigenvectors of the edges of the graph representation are determined based on a channel matrix of the wireless communication network, and the final first eigenvectors of the edges guide the determination of precoding of the wireless communication network.

[0151] In another possible implementation, the initial first eigenvectors of the edges of the graph representation are determined based on pilot signals of the wireless communication network, and the final first eigenvectors of the edges guide determination of channel estimation of the wireless communication network.

[0152] In another possible implementation, the initial first eigenvectors of the edges of the graph representation are determined based on a channel matrix of the wireless communication network, and the final first eigenvectors of the edges guide the determination of a MIMO detector of the wireless communication network.

[0153] In one possible implementation, the first node and the second node are nodes of different categories.

[0154] In another possible implementation, the graph neural network is obtained based on supervised training or unsupervised training.

[0155] According to the wireless policy optimization device 900 of the embodiment of the present application, it can correspond to the execution of the method described in the embodiment of the present application, and the above-mentioned and other operations and / or functions of each module in the wireless policy optimization device 900 are respectively for implementing the corresponding processes of each method in Figures 4-8. For the sake of brevity, they will not be repeated here.

[0156] An embodiment of the present application also provides a computing device, including at least one processor, a memory, and a communication interface, wherein the processor is configured to execute the method described in Figures 1-8.

[0157] FIG10 is a schematic diagram of the structure of a computing device provided in an embodiment of the present application.

[0158] As shown in Figure 10 , computing device 1000 includes at least one processor 1001, memory 1002, and communication interface 1004. Processor 1001, memory 1002, and communication interface 1003 are communicatively connected, either via a wired connection (e.g., a bus) or a wireless connection. Communication interface 1003 is configured to receive data sent by other devices. Memory 1002 stores computer instructions, which processor 1001 executes to implement the wireless policy optimization method described in the aforementioned method embodiment.

[0159] It should be understood that in the embodiment of the present application, the processor 1001 may be a central processing unit (CPU), or may be other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field programmable gate arrays (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor, etc.

[0160] The memory 1002 may include a read-only memory and a random access memory, and provides instructions and data to the processor 1001. The memory 1002 may also include a nonvolatile random access memory.

[0161] The memory 1002 may be a volatile memory or a non-volatile memory, or may include both volatile and non-volatile memories. The non-volatile memory may be a read-only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPROM), or a flash memory. The volatile memory may be a random access memory (RAM), which is used as an external cache. By way of example and not limitation, many forms of RAM are available, such as static RAM (SRAM), dynamic random access memory (DRAM), synchronous DRAM (SDRAM), double data rate synchronous dynamic random access memory (DDR SDRAM), enhanced synchronous dynamic random access memory (ESDRAM), synchronous link DRAM (SLDRAM), and direct rambus RAM (DR RAM).

[0162] It should be understood that the computing device 1000 according to the embodiment of the present application can execute the method shown in Figures 1-8 in the embodiment of the present application. The detailed description of the implementation of the method is given above and will not be repeated here for the sake of brevity.

[0163] An embodiment of the present application provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the above-mentioned method is implemented.

[0164] An embodiment of the present application provides a chip, which includes at least one processor and an interface, wherein the at least one processor determines program instructions or data through the interface; the at least one processor is used to execute the program instructions to implement the above-mentioned method.

[0165] An embodiment of the present application provides a computer program or a computer program product, which includes instructions. When the instructions are executed, the computer is caused to perform the above-mentioned method.

[0166] Those skilled in the art should further appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the composition and steps of each example have been generally described in terms of function in the above description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art may use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0167] The steps of the methods or algorithms described in conjunction with the embodiments disclosed herein may be implemented using hardware, a software module executed by a processor, or a combination of the two. The software module may be placed in a random access memory (RAM), a memory, a read-only memory (ROM), an electrically programmable ROM, an electrically erasable programmable ROM, a register, a hard disk, a removable disk, a CD-ROM, or any other form of storage medium known in the art.

[0168] The specific implementation methods described above further illustrate the purpose, technical solutions and beneficial effects of this application. It should be understood that the above description is only the specific implementation methods of this application and is not intended to limit the scope of protection of this application. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of this application should be included in the scope of protection of this application.

Claims

1. A wireless strategy optimization method, characterized in that: include: Obtaining a graph representation of a wireless communication network, the graph representation comprising a plurality of nodes and edges connecting the plurality of nodes, wherein the plurality of nodes comprises at least a first node and a second node, and the edge connects the first node and the second node; Using the graph representation as an input to a graph neural network, updating the first eigenvector of the edge through a multi-layer feature update network in the graph neural network, and outputting a final first eigenvector of the edge; Obtaining a solution to the wireless strategy optimization problem based on the final first eigenvector of the edge; In which, the first eigenvector of the edge output by each layer of the feature update network in the multi-layer feature update network is related to a first weight parameter, and the first weight parameter is determined based on the initial first eigenvector of the first adjacent edge of the edge and the first eigenvector of the first adjacent edge output by the previous layer of the feature update network of each layer of the feature update network, and the first adjacent edge is an edge that shares the first node with the edge.

2. The method according to claim 1, characterized in that A data processing network is further provided between adjacent feature update networks in the multi-layer feature update network; The data processing network is configured to obtain a second eigenvector of the edge based on the first eigenvector of the edge output by the feature update network of the previous layer of the feature update network of each layer, the first eigenvector of the first adjacent edge, and the first weight parameter; The second eigenvector of the edge is used as an input to the feature update network of each layer, and the feature update network of each layer obtains an updated first eigenvector of the edge based on the second eigenvector of the edge, the second eigenvector of the first adjacent edge of the edge, and the second eigenvector of the second adjacent edge of the edge; The parameters of the feature update network of each layer are updated during the training process of the graph neural network, and the second adjacent edge is an edge that shares the second node with the edge.

3. The method according to claim 1, characterized in that Each layer of the feature update network obtains an updated first feature vector of the edge based on the first feature vector of the edge output by the feature update network of the previous layer of the feature update network, the first feature vector of the first adjacent edge of the edge, and the first feature vector of the second adjacent edge of the edge; the second adjacent edge of the edge is an edge that shares the second node with the edge; The parameters of the feature update network of each layer are related to the first weight parameter.

4. The method according to any one of claims 1 to 3, characterized in that Initial first eigenvectors of edges of the graph representation are determined based on a channel matrix of the wireless communication network, and final first eigenvectors of the edges guide determination of precoding of the wireless communication network.

5. The method according to any one of claims 1 to 3, characterized in that Initial first eigenvectors of edges of the graph representation are determined based on pilot signals of the wireless communication network, and final first eigenvectors of the edges guide determination of channel estimates of the wireless communication network.

6. The method according to any one of claims 1 to 3, characterized in that Initial first eigenvectors of edges of the graph representation are determined based on a channel matrix of the wireless communication network, and final first eigenvectors of the edges guide determination of a MIMO detector of the wireless communication network.

7. The method according to any one of claims 1 to 6, characterized in that The first node and the second node are nodes of different categories.

8. The method according to any one of claims 1 to 7, characterized in that The graph neural network is obtained based on supervised training or unsupervised training.

9. A wireless strategy optimization device, characterized in that: include: an acquisition module, configured to acquire a graph representation of a wireless communication network, the graph representation comprising a plurality of nodes and edges connecting the plurality of nodes, wherein the plurality of nodes comprises at least a first node and a second node, and the edge connects the first node and the second node; an inference module, configured to use the graph representation as input to a graph neural network, update the first eigenvector of the edge through a multi-layer feature update network in the graph neural network, and output a final first eigenvector of the edge; an optimization module, configured to obtain a solution to a wireless policy optimization problem based on the final first eigenvector of the edge; In which, the first eigenvector of the edge output by each layer of the feature update network in the multi-layer feature update network is related to a first weight parameter, and the first weight parameter is determined based on the initial first eigenvector of the first adjacent edge of the edge and the first eigenvector of the first adjacent edge output by the previous layer of the feature update network of each layer of the feature update network, and the first adjacent edge is an edge that shares the first node with the edge.

10. The optimization device according to claim 9, characterized in that A data processing network is further provided between adjacent feature update networks in the multi-layer feature update network; The data processing network is configured to obtain a second eigenvector of the edge based on the first eigenvector of the edge output by the feature update network of the previous layer of the feature update network of each layer, the first eigenvector of the first adjacent edge, and the first weight parameter; The second eigenvector of the edge is used as an input to the feature update network of each layer, and the feature update network of each layer obtains an updated first eigenvector of the edge based on the second eigenvector of the edge, the second eigenvector of the first adjacent edge of the edge, and the second eigenvector of the second adjacent edge of the edge; The parameters of the feature update network of each layer are updated during the training process of the graph neural network, and the second adjacent edge is an edge that shares the second node with the edge.

11. The optimization device according to claim 9, characterized in that Each layer of the feature update network obtains an updated first feature vector of the edge based on the first feature vector of the edge output by the feature update network of the previous layer of the feature update network, the first feature vector of the first adjacent edge of the edge, and the first feature vector of the second adjacent edge of the edge; the second adjacent edge of the edge is an edge that shares the second node with the edge; The parameters of the feature update network of each layer are related to the first weight parameter.

12. The optimization device according to any one of claims 9 to 11, characterized in that: Initial first eigenvectors of edges of the graph representation are determined based on a channel matrix of the wireless communication network, and final first eigenvectors of the edges guide determination of precoding of the wireless communication network.

13. The optimization device according to any one of claims 9 to 11, characterized in that: Initial first eigenvectors of edges of the graph representation are determined based on pilot signals of the wireless communication network, and final first eigenvectors of the edges guide determination of channel estimates of the wireless communication network.

14. The optimization device according to any one of claims 9 to 11, characterized in that: Initial first eigenvectors of edges of the graph representation are determined based on a channel matrix of the wireless communication network, and final first eigenvectors of the edges guide determination of a MIMO detector of the wireless communication network.

15. The optimization device according to any one of claims 9 to 14, characterized in that: The first node and the second node are nodes of different categories.

16. The optimization device according to any one of claims 9 to 15, characterized in that: The graph neural network is obtained based on supervised training or unsupervised training.

17. An electronic device comprising a memory and a processor, characterized in that: The memory stores executable code, and the processor executes the executable code to implement the method according to any one of claims 1 to 8.

18. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed in a computer, the computer is caused to execute the method according to any one of claims 1 to 8.

19. A computer program or computer program product, characterized in that The computer program or computer program product comprises instructions, and when the instructions are executed, the method according to any one of claims 1 to 8 is implemented.

Images