A method and apparatus for optimizing wireless strategies

Abstract

A method for optimizing a wireless strategy includes: obtaining 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 a first node and a second node, and the first node and the second node are connected by edges; using the graph representation as input to a graph neural network, and outputting the final feature vectors of the edges; and guiding the optimization of the wireless strategy based on the final feature vectors of the edges; wherein the structure of the graph neural network is constructed based on Taylor expansion guided by matrix inversion operation, which solves the problem of poor generalization performance of traditional GNNs.

Description

Technical Field

[0001] This application relates to the field of communication technology, and in particular to a method and apparatus for optimizing wireless strategies. Background Technology

[0002] Wireless policy optimization (such as resource allocation, channel estimation, and signal detection) is a crucial problem in wireless communication. Taking precoding as an example, precoding in wireless communication systems can effectively improve system energy efficiency and spectral efficiency. Although existing literature has proposed iterative algorithms to solve non-convex precoding optimization problems, the obtained numerical solutions are sensitive to defects in real-world systems (such as channel estimation errors). Another drawback of traditional optimization methods is their high computational complexity when the problem size 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, general DNNs often lack dimensionality generalization ability. When the problem dimension changes with the number of participating terminal devices, network devices, etc., DNN methods need to retrain the neural network model, resulting in significant computational overhead. While traditional GNNs can achieve dimensionality generalization when solving wireless policy optimization problems, they indiscriminately consider the mutual influence of different edges on the graph, leading to poor performance. Summary of the Invention

[0003] The embodiments of this application provide a method and apparatus for optimizing wireless strategies, which solves the problem of poor performance of traditional GNNs.

[0004] In a first aspect, this application provides a method for optimizing a wireless strategy, comprising: obtaining 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 edges connect the first node and the second node; using the graph representation as input to a graph neural network, updating the first feature vector of the edges through a multi-layer feature update network in the graph neural network, and outputting the final first feature vector of the edges; obtaining a solution to the wireless strategy optimization problem based on the final first feature vector of the edges; wherein the first feature vector of the edges output by each layer of the feature update network in the multi-layer feature update network is related to a first weight parameter, the first weight parameter being 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 previous layer of the feature update network of each layer, the first adjacent edge being an edge that shares a first node with the edge.

[0005] The graph neural network in the wireless strategy optimization method provided in the embodiments of this application has good scalability and generalization performance, which solves the problem that traditional GNNs have limited expressive power and can only learn element-wise functions when the problem scale is large.

[0006] In one possible implementation, a data processing network is also set up between adjacent feature update networks in the 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 output by the feature update network of the previous layer, the first feature vector of the first adjacent edge, and the first weight parameter. The second feature vector of the edge is used as the input of each layer feature update network. Each layer feature update network 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 each layer feature update network are updated during the training of the graph neural network. The second adjacent edge is the edge that shares the second node with the edge.

[0007] In another possible implementation, each layer of the feature update network obtains the updated first feature vector of the edge based on the first feature vector of the edge output by the previous layer of the feature update network, the first feature vector of the edge's first adjacent edge, and the first feature vector of the edge's second adjacent edge; the second adjacent edge of the edge is the edge that shares a second node with the edge; the parameters of each layer of the feature update network are related to the first weight parameter.

[0008] In another possible implementation, the initial first eigenvector of the edge in the graph representation is determined based on the channel matrix of the wireless communication network, and the final first eigenvector of the edge guides the determination of the precoding of the wireless communication network.

[0009] In another possible implementation, the initial first eigenvector of the edge in the graph representation is determined based on the pilot signal of the wireless communication network, and the final first eigenvector of the edge guides the determination of the channel estimation of the wireless communication network.

[0010] In another possible implementation, the initial first eigenvector of the edge in the graph representation is determined based on the channel matrix of the wireless communication network, and the final first eigenvector of the edge guides the determination of the MIMO detector of the wireless communication network.

[0011] In one possible implementation, the first node and the second node are nodes of different categories.

[0012] In one example, the graph neural network is obtained based on supervised training or unsupervised training.

[0013] Secondly, this application provides a wireless strategy optimization apparatus comprising an acquisition module, an inference module, and an optimization module: wherein the acquisition module is used to acquire 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 edges connect the first node and the second node; the inference module is used to take the graph representation as input to a graph neural network, update the first feature vector of the edges through a multi-layer feature update network in the graph neural network, and output the final first feature vector of the edges; the optimization module is used to obtain a solution to the wireless strategy optimization problem based on the final first feature vector of the edges; wherein the first feature vector of the edges output by each layer of the feature update network in the multi-layer feature update network is related to a first weight parameter, the first weight parameter being 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 previous layer of the feature update network, the first adjacent edge being an edge that shares a first node with the edge.

[0014] In one possible implementation, a data processing network is also set up between adjacent feature update networks in the 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 output by the feature update network of the previous layer, the first feature vector of the first adjacent edge, and the first weight parameter. The second feature vector of the edge is used as the input of each layer feature update network. Each layer feature update network 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 each layer feature update network are updated during the training of the graph neural network. The second adjacent edge is the edge that shares the second node with the edge.

[0015] In another possible implementation, each layer of the feature update network obtains the updated first feature vector of the edge based on the first feature vector of the edge output by the previous layer of the feature update network, the first feature vector of the edge's first adjacent edge, and the first feature vector of the edge's second adjacent edge; the second adjacent edge of the edge is the edge that shares a second node with the edge; the parameters of each layer of the feature update network are related to the first weight parameter.

[0016] In another possible implementation, the initial first eigenvector of the edge in the graph representation is determined based on the channel matrix of the wireless communication network, and the final first eigenvector of the edge guides the determination of the precoding of the wireless communication network.

[0017] In another possible implementation, the initial first eigenvector of the edge in the graph representation is determined based on the pilot signal of the wireless communication network, and the final first eigenvector of the edge guides the determination of the channel estimation of the wireless communication network.

[0018] In another possible implementation, the initial first eigenvector of the edge in the graph representation is determined based on the channel matrix of the wireless communication network, and the final first eigenvector of the edge guides the determination of the MIMO detector of the wireless communication network.

[0019] In one possible implementation, the first node and the second node are nodes of different categories.

[0020] In another possible implementation, the graph neural network is obtained based on supervised training or unsupervised training.

[0021] Thirdly, this application provides an electronic device, including 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 this application.

[0022] Fourthly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed in a computer, causes the computer to perform the method provided in the first aspect of this application.

[0023] Fifthly, this application provides a computer program or computer program product, the computer program or computer program product including instructions that, when executed, implement the method provided in the first aspect of this application. Attached Figure Description

[0024] Figure 1 This is a schematic diagram of the structure of a fully connected neural network;

[0025] Figure 2 This is a schematic diagram of a loss function optimization.

[0026] Figure 3 This is a schematic diagram of gradient backpropagation;

[0027] Figure 4 This is a schematic diagram of edge feature update in a traditional GNN;

[0028] Figure 5 This is a schematic diagram of a wireless communication network;

[0029] Figure 6 This is a schematic diagram of another type of wireless communication network;

[0030] Figure 7 This is a schematic diagram illustrating the edge feature update of a GNN model.

[0031] Figure 8 A schematic flowchart illustrating a wireless strategy optimization method provided in an embodiment of this application;

[0032] Figure 9 A schematic diagram of a wireless strategy optimization device;

[0033] Figure 10 This is a schematic diagram of the structure of a computing device. Detailed Implementation

[0034] The technical solution of this application will be further described in detail below with reference to the accompanying drawings and embodiments.

[0035] 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 is a brief introduction to the neural network training process that may be involved in the embodiments of this application, using a fully connected neural network as an example.

[0036] Fully connected neural networks are also called multilayer perceptrons (MLPs), such as... Figure 1 As shown, an MLP consists of an input layer (left side), an output layer (right side), and multiple hidden layers (middle side). Each layer contains several nodes, called neurons. Neurons in adjacent layers are connected to each other.

[0037] Considering neurons in two adjacent layers, the output h of a neuron in the next layer is the weighted sum of all neurons x connected to it in the previous layer, after passing through an activation function. This can be represented by a matrix as follows:

[0038] h = f(wx + b)

[0039] Where w is the weight matrix, b is the bias vector, and f is the activation function. The output of the neural network can then be recursively expressed as:

[0040] y = f n (w n f n-1 (…)+b n )

[0041] Simply put, a neural network can be understood as a mapping from an input data set to an output data set. Neural networks are typically initialized randomly; the process of obtaining this mapping from random values ​​w and b using existing data is called training the neural network.

[0042] The specific training method involves evaluating the output of the neural network using a loss function and backpropagating the error. Gradient descent is then used to iteratively optimize w and b until the loss function reaches its minimum. (See...) Figure 2 ).

[0043] The gradient descent process can be represented 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 gradient descent.

[0044] The backpropagation process utilizes the chain rule for partial derivatives, meaning the gradient of the parameters in the previous layer can be recursively calculated from the gradient of the parameters in the next layer (see [link]). Figure 3 The formula can be expressed as follows: Where w ij Let s be the weight of the connection between node j and node i. i The weighted sum of the inputs at node i.

[0045] 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 with low computational complexity. However, training DNNs requires significant time, a large number of samples, and a high number of training parameters. For traditional fully connected neural networks (FNNs), even with sufficiently high training complexity, training performance may still be poor. Even if a DNN achieves good performance on the training set, it cannot guarantee its generalization ability for various parameters. Since the wireless environment changes dynamically over time, improving training efficiency by enhancing the generalization ability of DNNs or reducing their training complexity is crucial. By reducing training complexity, DNNs can be retrained using newly acquired samples when parameters that cannot be generalized well change significantly.

[0046] One way to improve learning efficiency is to incorporate prior knowledge when designing the DNN architecture, which can reduce the hypothesis space for the DNN to search for optimal parameters. For example, by leveraging permutation equivariance, which exists in many wireless tasks, permutation equivariant neural networks or graph neural networks (GNNs) can be designed to learn power allocation or precoding policies. While the training complexity of the above DNNs can be significantly reduced compared to unstructured FNNs, even for a medium-sized problem, they still require a large number of samples to learn the policy, which involves complex, nonlinear operations such as matrix inversion. The learning performance of GNNs also decreases with the increase in the number of samples. This also shows that GNNs cannot generalize to different problem sizes. To improve learning efficiency by simplifying the mappings that the DNN needs to learn, mathematical models can be introduced into the DNN architecture, which can be iterative algorithms or mathematical expressions. For example, deep unrolling methods use DNNs to mimic the update process of iterative algorithms, where only a portion of the algorithm's operations or parameters need to be learned.

[0047] We will use a multi-user precoding scheme as an example to introduce the scheme in the embodiments of this application. Of course, the scheme provided in the embodiments of this application can also be used to solve other problems (such as channel estimation, MIMO (multiple-in multiple-out) detection, etc.), which will be described in the embodiments.

[0048] The system model for the single-cell multi-user precoding problem considered in this application embodiment is as follows: Consider a base station with N antennas, which will serve K single-antenna users within its coverage area. Let h be the channel between the nth base station antenna and the kth user. nk Then the channels from all base station antennas to user k are h. k =[h 1k ,h 2k ,…,h Nk The channel between the base station antenna and all users is H = [h1, h2, ..., h]. K The problem of finding the optimal precoding matrix V can be modeled as follows:

[0049] P1:

[0050] stTr(V H V)≤P max

[0051] in, It is the throughput of the kth user, v k =[v 1k ,v 2k ,…,v Nk [v1, v2, ..., v] are the precoding vectors used by the base station when sending data to user k. All precoding vectors are combined to form the precoding matrix V = [v1, v2, ..., v] K ], P max That is the maximum transmission power of the base station. H Tr(·) represents the conjugate transpose operation, and Tr(·) represents the trace operation. This represents noise power.

[0052] Edge-clustered graph neural networks can be used to model and solve precoding problems. For example... Figure 4 As shown, consider the problem of solving a precoding scheme with N=4 and K=3. Figure 4 The system includes two types of nodes: base station antenna nodes and user nodes, and the features (or representations) of edges can be obtained through a multi-layer edge aggregation GNN. The edge (n,k) at the nth... The output of the layer is denoted as It can be obtained through the following two steps.

[0053] Aggregation: For adjacent edges connected to edge (n,k) via 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) via user k and antenna n respectively is:

[0054]

[0055] Among them PL u (·) and PL b (·) denotes the pooling functions output by the adjacent edges connecting user k and antenna n, respectively, for aggregation and (n,k). and Each contains trainable parameters and A function for extracting useful information from the adjacent sides connected to user k and antenna n.

[0056] Associativity: To obtain the output of edge (n,k), an associativity function is used to associate the edge (n,k) at the nth position. Layer output and aggregated output:

[0057]

[0058] Where CB(·) contains trainable parameters Associative functions.

[0059] To ensure that permutations of similar vertices (such as users or antennas) do not affect the output of the GNN, and For different edges that are the same, and the pooling function PL u (·) and PL b (·) satisfies the commutative law, such as for summation or maximization. For the sake of notation, assume q. u (·) and q b (·) are the same, and PL u (·) and PL b (·) are the same. Therefore, the subscripts u and b can be omitted. It is also assumed that q(·), CB(·), PL(·) and the trainable parameters remain the same across different layers; therefore, the above three formulas and the following text... superscript It can be omitted.

[0060] By substituting formula (1) into formula (2), we can obtain the formula used to update edge (n,k) at the nth k ... The update equation for the layer is:

[0061]

[0062] When q(·) is a linear function and CB(·) is a function composed of a linear equation and the activation function σ(·), equation (3) degenerates into:

[0063]

[0064] When the summation function is used as the pooling function, i.e., PL(·)=∑(·), equation (4) can be written as

[0065]

[0066] Figure 4 The input to the GNN shown (i.e., the initial values ​​of the side features) H is the channel matrix, and the output (i.e., the edge features after L iterations) is... ) represents the learned precoding matrix The input-output relationship of a GNN is denoted as... in It represents all trainable parameters in the edge-gathering GNN.

[0067] However, when GNN uses summation pooling in the update equation of equation (3) (i.e. Find the average pooling (i.e.) Or find max pooling (i.e. If a GNN uses a function (and GNNs using these functions are called traditional GNNs), then when N and K are large, the GNN can only learn element-wise functions. We call the function V = f(H) an element-wise function; if v nk Only depends on h nk At this point, traditional GNNs can only learn element-wise precoding strategies. However, under high SNR conditions, the optimal precoding strategy is not an element-wise function of the channel, so traditional GNNs cannot learn the optimal precoding strategy at this time.

[0068] Another approach involves using a deep unfolded network combining the model and a DNN to learn the precoder that maximizes energy efficiency. This deep unfolded network contains multiple layers, each outputting an update to the precoder policy. Each layer corresponds to one iteration of the WMMSE algorithm (i.e., introducing the mathematical model of the WMMSE algorithm into the DNN), where a small number of parameters or operations are trained. Numerical results show that the proposed method achieves near-optimal performance in learning the spectral efficiency-maximizing precoder.

[0069] However, the application of this approach is limited by the depth of the algorithm. For example, the WMMSE algorithm can only solve the problem of maximizing the utility function related to data rate under power constraints. It cannot solve other constraints (such as data rate constraints). In addition, due to the introduction of the iterative structure, the training and inference times of this technique are relatively long. The scalability of this method is also poor; the model trained under a given number of base station antennas and user configurations cannot be directly used in scenarios with other base station antenna counts and user counts.

[0070] This application proposes a model-driven GNN structure (hereinafter also referred to as model GNN), which has good scalability and generalization performance, and solves the problem that traditional GNNs have limited expressive power and can only learn element-wise functions when the problem size is large. The model-driven GNN structure proposed in this application can be used to solve problems such as precoding matrix solving, channel estimation, and MIMO detector design.

[0071] The solution provided in this application can be applied to wireless communication systems such as 5G, satellite communication, and short-range communication. The system architecture is as follows: Figure 5 As shown, a wireless communication system can consist of cells, each containing a base station (BS). The base station provides communication services to multiple mobile stations (MS). Wireless communication systems can also perform point-to-point communication, such as communication between multiple terminals.

[0072] It should be noted that the wireless communication systems mentioned in the embodiments of this application include, but are not limited to: narrowband Internet of Things (NB-IoT), 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-Synchronization Code Division Multiple Access (TD-SCDMA), Long Term Evolution (LTE), and the three major application scenarios of next-generation 5G mobile communication systems: eMBB, URLLC, and eMTC.

[0073] Another typical application scenario of this application is a wireless smart home or smart factory scenario. In a wireless smart home scenario, different smart home products (such as speakers, televisions, refrigerators, kitchen appliances, robot vacuum cleaners, etc.) are connected through a wireless network. Figure 6 This diagram illustrates wireless screen mirroring between a mobile phone and a smart TV. In a smart factory scenario, different devices (such as smart robots, lathes, and transport vehicles) connect via a wireless network.

[0074] The distributed nodes involved in the embodiments of this application may include various handheld devices, vehicle-mounted devices, wearable devices, computing devices, or other processing devices connected to a wireless modem with wireless communication capabilities. The distributed node may also be called a terminal, and may also be a subscriber unit, cellular phone, smartphone, wireless data card, personal digital assistant (PDA) computer, tablet computer, wireless modem, handset, laptop computer, machine type communication (MTC) terminal, etc.

[0075] The GNN network structure involved in the embodiments of this application will be described in detail below.

[0076] In wireless communication systems, there are numerous algorithms related to matrix inversion (or pseudo-inversion), such as precoding matrix calculation, channel estimation, and MIMO detection algorithms. When using neural networks to learn from these operations, the neural network needs sufficient expressive power to learn an approximation of matrix inversion as accurately as possible. Therefore, we start with matrix inversion and design the structure of a GNN.

[0077] The pseudo-inverse of matrix H is denoted as Hi. + H + The first-order Taylor expansion of H0 is:

[0078]

[0079] The result of the Taylor approximation can be used as H + It is an approximation of, and it is greater than Closer to H + To obtain a more accurate H + An approximation of H can be iteratively updated. + In each iteration (such as the ), In the next iteration, the first... The approximate result is used for expansion, that is...

[0080]

[0081] in It is H + Approximation of D (0) Random initial point.

[0082] It can be proven that the iterative equation derived from the first-order Taylor expansion is a compression mapping, and with more iterations, Able to more accurately approximate H + .

[0083] remember in Therefore, equation (7) can be written as The element in the nth row and kth column is:

[0084]

[0085] This iterative equation can be viewed as an update equation for an edge-gathering GNN that does not contain trainable parameters, where Is edge (n,k) at the nth position? The layer outputs, and the aggregation and associativity functions are as follows:

[0086]

[0087] The weights multiplied on the information of adjacent edges can be seen (e.g.) weight The weight varies depending on the edge. It can reflect the information on each edge for updating the approximate pseudoinverse. The importance of this can also be seen. At the same time, the aggregated output of edge information adjacent to user k (i.e., in equation (1)) can be observed. This information is not included in the output of the aggregation function because it has already been used to compute the matrix. It is included. Specifically, The kth column That is, the information gathered for user k, where It is the k-th column of matrix H.

[0088] Based on the above derivation, two implementation methods for the GNN model can be obtained.

[0089] The two-step implementation of the GNN model, for the first step of the GNN model... Layer edge (n,k) operations:

[0090] Step 1: Inspired by equation (8), based on the first... Layer output Get vector

[0091] Step 2: Obtained through a GNN layer The update equation for GNN is:

[0092]

[0093] in S, P, Q are trainable parameter matrices whose dimensions are independent of N and K. Therefore, parameter matrices trained under a set of N and K values ​​can be used under other N and K values.

[0094] Based on the two-step implementation described above, a one-step implementation of the GNN model can be derived. Let... If P = [P0, P1] and Q = [Q0, Q1], then equation (11) can be further written as:

[0095]

[0096] in

[0097]

[0098] In the above formula, the weight matrix is ​​ignored for the sake of simplicity. superscript In equation (12), term (a) aggregates information containing the adjacent edges connected to user k. Item (b) aggregates information containing adjacent edges connected to antenna n, and item (c) aggregates information on edges not adjacent to edge (n,k). To reduce computational complexity without sacrificing the representational power of the GNN, item (c) can be ignored, since information on 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 update operation of the model GNN takes into account the weights. The influence of this on the precoding strategy is that it characterizes the correlation between the channel of user i and the precoding vector of user j, thus having a stronger expressive power than traditional GNN.

[0099] Therefore, the one-step implementation of the GNN model is suitable for the first step of the GNN model. Layer edge (n,k) operations:

[0100] Step 1: Place the first Layer output The input update equation is a GNN with the following formula:

[0101]

[0102] in

[0103] like Figure 7 The diagram shows the update of edge (1,1) in the one-step implementation of the GNN model. In layer operations, edge features Due to its own higher-level characteristics Features of the upper layer of the edge connected to user node UE1 The upper layer features of the edge connected to base station antenna node AN1 Decide.

[0104] Edge (n,k) at the th The output feature of a layer can be a vector rather than a scalar, i.e. It contains Each element. When hour, Therefore, J0 = 2. When At that time, the output of the GNN is the learned policy. Taking precoding as an example, its output is the learned precoding vector, i.e. Therefore J L =2. When hour, It can be greater than 2, and The larger the value, the stronger the representational power of the GNN model. The input-output relationship of a GNN is denoted as... Where θ V It contains all trainable parameters of the GNN model.

[0105] Once the implementation of the neural network model is obtained, the parameter matrix can be trained based on the training data. The training method can be either unsupervised or supervised learning; taking pre-encoding policy learning as an example...

[0106] Unsupervised learning: Based on the output of all edges of the model GNN Obtain the precoding matrix Total throughput of computing system Find the parameter θ V The gradient is calculated, and the gradient descent method described above is used to update the parameter θ. V .

[0107] Supervised learning: In supervised learning, the label (i.e., the target precoding matrix) V corresponding to channel H will be provided. * The precoding matrix based on the output of the model GNN and label V * Calculate the loss function (e.g., mean squared error) and find the parameter θ of the loss function. VThe gradient is calculated, and the gradient descent method described above is used to update the parameter θ. V .

[0108] This application provides a method for optimizing wireless strategies, and its implementation is described below.

[0109] Figure 8 A flowchart illustrating a wireless strategy optimization method provided in this application embodiment, the method being applied to Figure 4 or Figure 5 In the wireless communication scenario shown, this method can be executed by any computing-capable device, equipment, platform, or cluster of devices, for example... Figure 5 Terminals and / or base stations in the middle, or Figure 6 The method can be executed on a mobile phone and / or television. This application does not specifically limit the specific computing device for executing the method; a suitable computing device can be selected as needed. Figure 8 As shown, the method for optimizing the wireless strategy includes at least steps S801-S803.

[0110] In step S801, a graphical representation of the wireless communication network is obtained.

[0111] For example, a graph representation includes multiple nodes and multiple edges, with edges 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 edges in the graph representation is the channel vector; in the channel estimation task, the initial feature vector of the edges in the graph representation is the pilot signal; and in the MIMO detection task, the initial feature vector of the edges in the graph representation is the channel matrix, etc.

[0112] In another example, the graph representation may 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, for example... Figure 4 or Figure 7 As shown, the initial eigenvectors of the edges in the graph representation are the initial state parameters of the optimization task. For example, in the precoding task, the initial eigenvectors of the edges in the graph representation are the channel vectors; in the channel estimation task, the initial eigenvectors of the edges in the graph representation are the pilot signals; and in the MIMO detection task, the initial eigenvectors of the edges in the graph representation are the channel matrix.

[0113] In step S802, the graph representation is used as the input to the graph neural network. The feature vectors of the edges are updated by the multi-layer feature update network in the graph neural network, and the final feature vectors of the edges are output.

[0114] The structure of the graph neural network is described above in the section on the structure of the model graph neural network. The first weight parameter can be found in the section above.

[0115] In step S803, the optimization of the wireless strategy is guided by the final feature vector of the edge.

[0116] By optimizing the wireless strategy of the wireless communication network, the system energy efficiency and spectral efficiency of the wireless communication network are effectively improved.

[0117] The model GNN provided in this application can be used for policy optimization in wireless communication networks, such as precoding, channel estimation, and MIMO detection.

[0118] For example, the GNN model is applied to precoding, such as in a scenario where a base station with N antennas transmits signals to K single-antenna users via a precoding matrix:

[0119] Input: Channel vector,

[0120] Model GNN operations: two-step or one-step implementation, L-layer update operation.

[0121] Output: Precoded vector

[0122] Training: Unsupervised learning or supervised learning, see the description above for details.

[0123] An example of applying the GNN model to channel estimation is as follows: A base station with N antennas transmits pilot signals to K single-antenna users. The pilot signals are... Where N pl This represents the number of pilot signals transmitted on each antenna. The received signal is... From this equation, it can be seen that the channel estimation can be written as: To achieve more accurate channel estimation, a model GNN can be used to learn a channel estimation result related to the pseudo-inverse.

[0124] The output of the GNN model at each layer is The output is X. + The approximation is given by equation (11). The difference from Example 1 is that the weights... The calculation is as follows: Where x j Let be the column vector formed by the j-th row of matrix X. For matrix The kth column.

[0125] The GNN model contains L layers. Its input is X, and its output is D. (L) The channel estimate obtained using the output is... A supervised approach can be used to train the GNN model, with the training objective being to minimize the mean square error between the GNN output and the actual channel.

[0126] The GNN model can also perform joint optimization of channel estimation and related tasks (such as precoding). Precoding is related to the pseudo-inverse of the estimated channel matrix, i.e.

[0127] The output of the GNN precoding model with the pseudo-inverse of the channel matrix at layer l is: The output is Y. + The approximation is given by equation (11). Unlike the previous equation, the weights are... The calculation is as follows: Where y j Let be the column vector formed by the j-th row of matrix Y. For matrix The kth column.

[0128] The GNN model contains L layers. Its input is Y, and its output is D. (L) The channel estimate obtained using the output is... The GNN model can be trained using either supervised or unsupervised methods. When training in an unsupervised manner, the training loss function is the same as the loss function used above.

[0129] An example of applying the GNN model to MIMO detection is as follows: Consider using the GNN model provided in this application to design a combiner at the base station receiver during 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 represented as y = H T x+n, where n is the variance. The noise.

[0130] To recover the signal from the transmitter, a combiner can be used at the receiver. Common combiners include the least squares combiner (V... LS =H H (HH H ) -1 ) with MMSE connector

[0131] To learn the combiner that is correlated with the pseudo-inverse of the channel matrix, the model GNN provided in this application embodiment can be used. The update equation of the model GNN is shown in equation (11).

[0132] The GNN model contains L layers. Its input is H, and its output is the learned combiner. A supervised training method can be used to train the GNN model, with the training objective being to minimize the mean square error between the received and transmitted signals of the GNN.

[0133] The model GNN provided in this application solves the problem of poor generalization performance of traditional GNNs. It 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 lower complexity and better performance.

[0134] The following uses precoding as an example to illustrate the beneficial effects of the GNN model provided in the embodiments of this application through simulation.

[0135] Consider a scenario where an N-antenna base station serves K single-antenna users. The power consumption of each antenna is P. c =17.6W, constant power consumption P0 of each base station =43.3W, power amplifier efficiency 1 / ρ =0.311.

[0136] Consider training the DNN using an unsupervised approach. The inputs in the training and testing samples are generated by independent Rayleigh distribution channels, meaning the elements in H follow a complex Gaussian distribution. For each test sample, labels are also generated using a numerical algorithm as a benchmark method to evaluate learning performance. Problem P1 can be solved using the weighted minimum mean squared error (WMMSE) algorithm. The number of iterations is set to 100 to allow the algorithm to generate labels within an acceptable timeframe. The algorithm may not converge after iterations and may only find suboptimal solutions. In the simulation, a training set of 200,000 samples is generated (although the actual number of samples used for training may be smaller), and another 100 samples are generated as a test set.

[0137] The hyperparameters of GNN are shown in Table 1. In this table, the hyperparameters of each layer (such as the first layer) are listed in Table 1. The number of hidden nodes in a layer is the dimension output by each edge, i.e. It is independent of K and N. Therefore, these hyperparameters can be used for policy learning under different K and N scenarios. Furthermore, consider using more hidden layers to learn the most energy-efficient policy, enabling the GNN to better meet QoS constraints. The activation function of each layer of the GNN is σ(X) = X / ∥X∥. The Adam algorithm is used to train the GNN.

[0138] Table 1 Hyperparameters of GNN

[0139]

[0140] The training performance of a DNN is obtained by averaging the results of training each DNN five times, which eliminates the randomness introduced by sample selection and initial weights. In each training iteration, N samples are randomly selected from the training set. tr The DNN was trained on 100 samples and tested on 100 samples. All simulation results were obtained on a computer equipped with a 14-core Intel i9-9940X CPU, an Nvidia RTX 2080Ti GPU, and 64GB of RAM.

[0141] The performance of optimal precoding for learning problem P1 using Model GNN and traditional GNN is presented below. For deep unfolding networks, the WMMSE algorithm is unfolded into a multi-layer structure, with each layer learning only a small subset of parameters or operations. To keep training time acceptable, the number of layers is set to 5 in the simulation. The spectral efficiency ratio (SE ratio) is used in the simulation to evaluate learning performance; that is, the ratio of the spectral efficiency achieved by each method to that achieved by the WMMSE algorithm.

[0142] Table 2 compares the spectral efficiency and inference time of various DNNs and the WMMSE algorithm under different N, K, and signal-to-noise ratios, where the inference time reflects the computational complexity of the inference stage. 1000 samples were used to train the DNN. The inference times of the GNN model and the traditional GNN were obtained on a GPU. Since the code for the deep unfolded 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 to these two methods. It can be seen that the GNN model achieves a spectral efficiency of nearly or more than 95% on all N and K values. 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 is still significantly lower than that of the deep unfolded network and the WMMSE algorithm, because the latter involves computationally complex operations (such as a large number of matrix multiplications in multiple iterations).

[0143] Table 2

[0144]

[0145] Based on the same concept as the aforementioned embodiment of a wireless strategy optimization method, this application also provides a wireless strategy optimization apparatus 900, which includes components for implementing... Figure 4-8 The units or modules in the various steps of the wireless strategy optimization method shown.

[0146] Figure 9This is a schematic diagram of a wireless strategy optimization device provided in an embodiment of this application. The device is applied to a computing device, such as... Figure 9 As shown, the wireless strategy optimization device 900 includes at least an acquisition module 901, an inference module 902, and an optimization module 903. The acquisition module 901 acquires 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 edges connect the first node and the second node. The inference module 902 takes the graph representation as input to a graph neural network, updates the first feature vector of the edges through a multi-layer feature update network in the graph neural network, and outputs the final first feature vector of the edges. The optimization module 903 obtains a solution to the wireless strategy optimization problem based on the final first feature vector of the edges. The first feature vector of the edges output by each layer of the multi-layer feature update network is related to a first weight parameter, which 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 previous layer of the feature update network. The first adjacent edge is an edge that shares a first node with the edge.

[0147] In one possible implementation, a data processing network is also set up between adjacent feature update networks in the 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 output by the feature update network of the previous layer, the first feature vector of the first adjacent edge, and the first weight parameter. The second feature vector of the edge is used as the input of each layer feature update network. Each layer feature update network 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 each layer feature update network are updated during the training of the graph neural network. The second adjacent edge is the edge that shares the second node with the edge.

[0148] In another possible implementation, each layer of the feature update network obtains the updated first feature vector of the edge based on the first feature vector of the edge output by the previous layer of the feature update network, the first feature vector of the edge's first adjacent edge, and the first feature vector of the edge's second adjacent edge; the second adjacent edge of the edge is the edge that shares a second node with the edge; the parameters of each layer of the feature update network are related to the first weight parameter.

[0149] In another possible implementation, the initial first eigenvector of the edge in the graph representation is determined based on the channel matrix of the wireless communication network, and the final first eigenvector of the edge guides the determination of the precoding of the wireless communication network.

[0150] In another possible implementation, the initial first eigenvector of the edge in the graph representation is determined based on the pilot signal of the wireless communication network, and the final first eigenvector of the edge guides the determination of the channel estimation of the wireless communication network.

[0151] In another possible implementation, the initial first eigenvector of the edge in the graph representation is determined based on the channel matrix of the wireless communication network, and the final first eigenvector of the edge guides the determination of the MIMO detector of the wireless communication network.

[0152] In one possible implementation, the first node and the second node are nodes of different categories.

[0153] In another possible implementation, the graph neural network is obtained based on supervised training or unsupervised training.

[0154] The wireless policy optimization apparatus 900 according to the embodiments of this application can correspond to executing the method described in the embodiments of this application, and the above and other operations and / or functions of each module in the wireless policy optimization apparatus 900 are respectively for implementing Figure 4-8 For the sake of brevity, the corresponding processes of each method in the code will not be elaborated here.

[0155] This application embodiment also provides a computing device, including at least one processor, a memory, and a communication interface, wherein the processor is used to execute... Figure 1-8 The method described.

[0156] Figure 10 A schematic diagram of the structure of a computing device provided in an embodiment of this application.

[0157] like Figure 10 As shown, the computing device 1000 includes at least one processor 1001, a memory 1002, and a communication interface 1004. The processor 1001, memory 1002, and communication interface 1004 are communicatively connected, which can be achieved via a wired (e.g., bus) or wireless connection. The communication interface 1003 is used to receive data sent by other devices; the memory 1002 stores computer instructions, and the processor 1001 executes these computer instructions to perform the wireless strategy optimization method described in the aforementioned method embodiment.

[0158] It should be understood that, in the embodiments of this application, the processor 1001 may be a central processing unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), 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.

[0159] The memory 1002 may include read-only memory and random access memory, and provides instructions and data to the processor 1001. The memory 1002 may also include non-volatile random access memory.

[0160] The memory 1002 can be volatile memory or non-volatile memory, or it can include both. The non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. The volatile memory can be random access memory (RAM), which is used as an external cache. By way of example, but not limitation, many forms of RAM are available, such as static random access memory (SRAM), dynamic random access memory (DRAM), synchronous dynamic random access memory (SDRAM), double data rate synchronous dynamic random access memory (DDR SDRAM), enhanced synchronous dynamic random access memory (ESDRAM), synchronous linked dynamic random access memory (SLDRAM), and direct rambus RAM (DR RAM).

[0161] It should be understood that the computing device 1000 according to the embodiments of this application can execute the implementation of the embodiments of this application. Figure 1-8 The method shown is described in detail above, and will not be repeated here for the sake of brevity.

[0162] Embodiments of this application provide a computer-readable storage medium having a computer program stored thereon, wherein when the computer instructions are executed by a processor, the aforementioned method is implemented.

[0163] An embodiment of this application provides a chip including 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 method mentioned above.

[0164] Embodiments of this application provide a computer program or computer program product that includes instructions that, when executed, cause a computer to perform the methods mentioned above.

[0165] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing 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 can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0166] The steps of the methods or algorithms described in conjunction with the embodiments disclosed herein can be implemented in hardware, processor-executed software modules, or a combination of both. The software modules can be located in random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disks, removable disks, CD-ROMs, or any other form of storage medium known in the art.

[0167] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of this application. It should be understood that the above description is only a specific embodiment of this application and is not intended to limit the scope of protection of this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A method for optimizing a wireless strategy, characterized in that, include: 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 edges connect the first node and the second node; The graph representation is used as input to the graph neural network. The first feature vector of the edge is updated by the multi-layer feature update network in the graph neural network, and the final first feature vector of the edge is output. The solution to the wireless policy optimization problem is obtained based on the final first feature vector of the edge. In the multi-layer feature update network, the first feature vector of the edge output by each layer of the feature update network is related to the first weight parameter. 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 previous layer of the feature update network of each layer of the feature update network. The first adjacent edge is the edge that shares the first node with the edge.

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

3. The method according to claim 1, characterized in that, The feature update network of each layer obtains the 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, 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 the edge that shares the second node with the edge. The parameters of each layer of the feature update network are related to the first weight parameter.

4. The method according to any one of claims 1-3, characterized in that, The initial first feature vector of the edge represented by the graph is determined based on the channel matrix of the wireless communication network, and the final first feature vector of the edge guides the determination of the precoding of the wireless communication network.

5. The method according to any one of claims 1-3, characterized in that, The initial first feature vector of the edge represented by the graph is determined based on the pilot signal of the wireless communication network, and the final first feature vector of the edge guides the determination of the channel estimation of the wireless communication network.

6. The method according to any one of claims 1-3, characterized in that, The initial first feature vector of the edge represented by the graph is determined based on the channel matrix of the wireless communication network, and the final first feature vector of the edge guides the determination of the MIMO detector of the wireless communication network.

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

8. The method according to any one of claims 1-7, 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 is used to acquire 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 edges connect the first node and the second node; The inference module is used to take the graph representation as input to the 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. An optimization module is used to obtain a solution to the wireless policy optimization problem based on the final first feature vector of the edge; In the multi-layer feature update network, the first feature vector of the edge output by each layer of the feature update network is related to the first weight parameter. 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 previous layer of the feature update network of each layer of the feature update network. The first adjacent edge is the edge that shares the first node with the edge.

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

11. The optimization device according to claim 9, characterized in that, The feature update network of each layer obtains the 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, 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 the edge that shares the second node with the edge. The parameters of each layer of the feature update network are related to the first weight parameter.

12. The optimization apparatus according to any one of claims 9-11, characterized in that, The initial first feature vector of the edge represented by the graph is determined based on the channel matrix of the wireless communication network, and the final first feature vector of the edge guides the determination of the precoding of the wireless communication network.

13. The optimization apparatus according to any one of claims 9-11, characterized in that, The initial first feature vector of the edge represented by the graph is determined based on the pilot signal of the wireless communication network, and the final first feature vector of the edge guides the determination of the channel estimation of the wireless communication network.

14. The optimization apparatus according to any one of claims 9-11, characterized in that, The initial first feature vector of the edge represented by the graph is determined based on the channel matrix of the wireless communication network, and the final first feature vector of the edge guides the determination of the MIMO detector of the wireless communication network.

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

16. The optimization apparatus according to any one of claims 9-15, 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 1-8.

18. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed in a computer, it causes the computer to perform the method described in any one of claims 1-8.

19. A computer program product, characterized in that, The computer program product includes instructions that, when executed, implement the method according to any one of claims 1-8.

Images