Wireless network resource management optimization method and device, electronic equipment and storage medium

Abstract

The application provides a wireless network resource management optimization method and device, electronic equipment and storage medium, the method comprises the following steps: determining a wireless network resource management optimization problem, and obtaining an input parameter vector of the wireless network resource management optimization problem; inputting the input parameter vector into a target deep unfolding network to obtain a solution result of the wireless network resource management optimization problem; and performing wireless network resource management optimization according to the solution result; the target deep unfolding network is obtained by training hyperparameters of the target deep unfolding network based on sample input parameter vectors of the wireless network resource management optimization problem and solution results of optimization problems corresponding to the sample input parameter vectors; and the hyperparameters include one or more of an iteration step of an optimization variable, an iteration step of a dual variable and a penalty parameter. The applicability of the optimization problem solution is expanded, and the accuracy of the small sample optimization problem solution is improved.

Description

Technical Field

[0001] This invention relates to the fields of signal processing and wireless communication technology, and in particular to a method, apparatus, electronic device and storage medium for optimizing wireless network resource management. Background Technology

[0002] Resource management and optimization problems in signal processing and wireless communication are often modeled as mathematical optimization problems. Solving these optimization problems often requires designing complex iterative algorithms, resulting in high computational complexity. Ultra-large-scale MIMO (Multiple-Input Multiple-Output) and integrated communication and sensing technologies have become key enabling technologies in 6G. However, with the increase in problem scale (including the increase in the dimensionality of optimization variables and the number of constraints), computational complexity is increasingly becoming a bottleneck for the real-time application of many algorithms. To reduce computational complexity, researchers have proposed the idea of ​​learning optimization. A typical implementation uses deep neural network modeling algorithms and trains them end-to-end based on collected samples. However, because the original optimization algorithm is considered a black box, the resulting algorithm not only lacks interpretability but also requires massive amounts of data to train the constructed deep network. Furthermore, the network's generalization performance is poor, making it difficult to apply to dynamic and time-varying scenarios.

[0003] To address these issues, researchers proposed integrating model information from the algorithm into its design, resulting in a deep unfolding-based learning optimization method. The basic idea of ​​deep unfolding is to treat the iterative steps in the optimization algorithm as layers in a deep neural network, selecting sensitive elements as learnable parameters and using backpropagation for optimization. Because it incorporates model information from the original algorithm, algorithms designed based on deep unfolding not only have good interpretability but also lower computational and sample complexity compared to deep neural network modeling algorithms, while exhibiting better generalization performance. However, existing algorithms based on deep unfolding ultimately rely on interior-point methods for solving the problem, resulting in relatively high network sample complexity. In small sample scenarios, network performance and accuracy in solving the optimization problem are not ideal, limiting their application scope. Summary of the Invention

[0004] To address the problems existing in the prior art, the present invention provides a method, apparatus, electronic device, and storage medium for optimizing wireless network resource management.

[0005] In a first aspect, the present invention provides a method for optimizing wireless network resource management, comprising:

[0006] Identify the wireless network resource management optimization problem and obtain the input parameter vector of the wireless network resource management optimization problem;

[0007] The input parameter vector is input into the target depth unfolded network to obtain the solution to the wireless network resource management optimization problem.

[0008] Based on the solution results, optimize the management of wireless network resources;

[0009] The target depth unfolding network is obtained by training its hyperparameters based on the sample input parameter vector of the wireless network resource management optimization problem and the solution results of the optimization problem corresponding to the sample input parameter vector; the hyperparameters include one or more of the following: the iteration step size of the optimization variable, the iteration step size of the dual variable, and the penalty parameter.

[0010] Optionally, the target depth unfolding network is constructed based on the following method:

[0011] The target wireless network resource management optimization problem used by the target depth unfolding network to solve is determined to be either a smooth optimization problem or a non-smooth optimization problem.

[0012] Based on the type of the target wireless network resource management optimization problem, a corresponding expandable iterative format is constructed, wherein the expandable iterative format includes the hyperparameters;

[0013] Based on the aforementioned expandable iterative format, the target depth unfolded network is constructed.

[0014] Optionally, constructing a corresponding expandable iterative format based on the type of the target wireless network resource management optimization problem includes:

[0015] When the target wireless network resource management optimization problem is a smooth optimization problem, the expandable iterative form corresponding to the target wireless network resource management optimization problem is constructed using the primal dual method or the primal dual method and gradient descent method.

[0016] Optionally, constructing a corresponding expandable iterative format based on the type of the target wireless network resource management optimization problem includes:

[0017] When the target wireless network resource management optimization problem is a non-smooth optimization problem, it is transformed into an alternating optimization problem of a smooth optimization subproblem and a non-smooth optimization subproblem.

[0018] The expandable iterative scheme corresponding to the smooth optimization subproblem is constructed using the primal dual method or the primal dual method and gradient descent method, and the expandable iterative scheme corresponding to the non-smooth optimization subproblem is constructed using the proximity operator method.

[0019] Optionally, constructing the target depth unfolded network based on the unfoldable iterative format includes:

[0020] Based on the expandable iterative format corresponding to the smooth optimization subproblem, the original dual subnetwork is constructed.

[0021] Based on the original dual subnetwork and the expandable iterative format corresponding to the non-smooth optimization subproblem, the target depth expansion network is constructed.

[0022] Optionally, constructing a corresponding expandable iterative format based on the type of the target wireless network resource management optimization problem includes:

[0023] If the target wireless network resource management optimization problem satisfies the first condition, a corresponding expandable iterative format is constructed based on the type of the target wireless network resource management optimization problem;

[0024] The first condition includes:

[0025] For each constraint function in the target wireless network resource management optimization problem, there exist two corresponding convex functions, the difference of which is equal to the constraint function; and...

[0026] For the objective function of the target wireless network resource management optimization problem, there exists a differentiable and convex function and a convex but non-differentiable function, and the linear weighted sum of the differentiable and convex function and the convex but non-differentiable function is equal to the objective function.

[0027] Optionally, the training method of the target deep unfolded network includes:

[0028] Fix all elements in Θ except Θ1, update Θ1 based on the training samples and the loss function, then fix all elements in Θ except Θ2, update Θ2 based on the training samples and the loss function, and so on. For any iteration of Θ... k The updates are all fixed except for Θ. k All other elements remain unchanged, only Θ is updated. k ;or,

[0029] Fix all elements in Θ except Θ1, update Θ1 based on the training samples and the loss function, then fix all elements in Θ except {Θ1,Θ2}, update {Θ1,Θ2} based on the training samples and the loss function, and so on. For any {Θ1,Θ2,...,Θ k The updates of} are all fixed in Θ except for {Θ1,Θ2,...,Θ} k All elements except {Θ1,Θ2,...,Θ} remain unchanged, only {Θ1,Θ2,...,Θ} are updated. k};

[0030] Where Θ represents the set of hyperparameters of the target depth unfolding network, Θ={Θ1,Θ2,...,Θ k,...,Θ K}, Θ k This represents the set of hyperparameters of the k-th subnetwork of the target depth unfolded network, where K represents the number of subnetworks in the target depth unfolded network.

[0031] Secondly, the present invention also provides a wireless network resource management optimization device, comprising:

[0032] The acquisition module is used to determine the wireless network resource management optimization problem and acquire the input parameter vector of the wireless network resource management optimization problem;

[0033] The solution module is used to input the input parameter vector into the target depth unfolded network to obtain the solution result of the wireless network resource management optimization problem;

[0034] The resource management optimization module is used to optimize the management of wireless network resources based on the solution results.

[0035] The target depth unfolding network is obtained by training its hyperparameters based on the sample input parameter vector of the wireless network resource management optimization problem and the solution results of the optimization problem corresponding to the sample input parameter vector; the hyperparameters include one or more of the following: the iteration step size of the optimization variable, the iteration step size of the dual variable, and the penalty parameter.

[0036] Thirdly, the present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor executes the program to implement the wireless network resource management optimization method described in the first aspect above.

[0037] Fourthly, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the wireless network resource management optimization method described in the first aspect above.

[0038] Fifthly, the present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the wireless network resource management optimization method as described above.

[0039] The wireless network resource management optimization method, apparatus, electronic device, and storage medium provided by this invention select scalar parameters such as iteration step size and penalty parameters as network hyperparameters to construct a deep unfolded network to solve optimization problems. This results in low sample complexity for the deep unfolded network, which can be trained to achieve good performance even with small samples, thereby expanding the applicability of optimization problem solving and improving the accuracy of solving optimization problems with small samples. Attached Figure Description

[0040] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0041] Figure 1 A flowchart illustrating the wireless network resource management optimization method provided by the present invention;

[0042] Figure 2 This is a schematic diagram of the original dual subnetwork provided by the present invention;

[0043] Figure 3 This is a schematic diagram of the structure of the composite optimization network provided by the present invention;

[0044] Figure 4 The average reachability performance curves of different algorithms provided in this invention;

[0045] Figure 5 Normalized mean square error performance curves of different algorithms provided by this invention;

[0046] Figure 6 Symbol error rate performance curves of different algorithms provided for this invention;

[0047] Figure 7 Generalization performance curves of different algorithms provided by this invention;

[0048] Figure 8 Performance curves of different algorithms provided in this invention;

[0049] Figure 9 A schematic diagram of the wireless network resource management optimization device provided by the present invention;

[0050] Figure 10 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0051] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0052] Figure 1This is a flowchart illustrating the wireless network resource management optimization method provided by the present invention, as shown below. Figure 1 As shown, the method includes the following steps:

[0053] Step 100: Determine the wireless network resource management optimization problem and obtain the input parameter vector for the wireless network resource management optimization problem.

[0054] Step 101: Input the input parameter vector into the target depth unfolded network to obtain the solution to the wireless network resource management optimization problem.

[0055] Step 102: Based on the solution results, optimize the management of wireless network resources.

[0056] The target depth unfolding network is obtained by training its hyperparameters based on the sample input parameter vector of the wireless network resource management optimization problem and the solution results of the optimization problem corresponding to the sample input parameter vector. The hyperparameters include one or more of the following: the iteration step size of the optimization variable, the iteration step size of the dual variable, and the penalty parameter.

[0057] Specifically, the management and optimization of wireless network resources refers to the management and optimization of various wireless network resources, such as channel configuration and optimization, power control and optimization, etc.

[0058] In the fields of signal processing and wireless communication, when optimizing wireless network resource management, it is often necessary to model the wireless network resource management optimization problem (such as various beamforming or precoding problems) as a mathematical optimization problem. By solving the wireless network resource management optimization problem, the management and optimization of wireless network resources can be carried out based on the solution results.

[0059] This invention provides a general method applicable to solving wireless network resource management optimization problems in the fields of signal processing and wireless communication. By pre-training a deep unfolded network according to different types of wireless network resource management optimization problems, when a certain wireless network resource management optimization problem needs to be solved, the input parameter vector of the optimization problem can be obtained first. The input parameter vector is the known parameter vector of the optimization problem. For example, for multi-user precoding problems (with the optimization objective of maximizing the achievable transmission rate and the user precoding vector as the optimization variable) and multi-user symbol-level precoding problems (with the optimization objective of minimizing the transmission power and the transmission signal vector as the optimization variable), the input parameter vector can include the user channel vector.

[0060] Then, the input parameter vector of the optimization problem is input into the target depth unfolded network used to solve the optimization problem, and the solution to the optimization problem can be obtained (for example, for the multi-user precoding problem mentioned above, the solution can be an optimized precoding vector; for the multi-user symbol-level precoding problem mentioned above, the solution can be an optimized transmitted signal vector). The solution can then be used to optimize the management of wireless network resources.

[0061] Unlike existing deep unfolding networks, the deep unfolding network provided in this invention selects scalar parameters such as iteration step size and penalty parameters as hyperparameters (also known as learnable or trainable parameters). Therefore, a model with good performance can be trained using a small number of samples, resulting in low sample complexity (i.e., good performance with few samples) and good generalization performance. Optionally, the hyperparameters of the target deep unfolding network can be trained using the backpropagation algorithm.

[0062] The optimization variables may include the optimization variables in the original optimization problem and the optimization variables in the optimization problem transformed from the original optimization problem. In some embodiments, the optimization variables can be solved iteratively using the primal-dual method and the gradient descent method, and the iteration step size of the optimization variables is the step size parameter of the gradient descent iteration.

[0063] In some embodiments, dual variables may include dual variables of inequality constraints in the primal optimization problem and dual variables of inequality constraints in an optimization problem transformed from the primal optimization problem.

[0064] In some embodiments, the original optimization problem can be equivalently transformed by introducing optimization variables and equality constraints. For the equality constraints, a penalty parameter and dual variables can be further introduced to obtain the augmented Lagrange penalized form of the equivalent transformed optimization problem. Therefore, in this invention, this penalty parameter can be used as a hyperparameter of the deep unfolded network.

[0065] The wireless network resource management optimization method provided by this invention selects scalar parameters such as iteration step size and penalty parameters as network hyperparameters to construct a deep unfolded network to solve the optimization problem. This makes the sample complexity of the deep unfolded network low, and it can be trained to achieve good performance even with small samples. This expands the applicability of solving optimization problems and improves the accuracy of solving optimization problems with small samples.

[0066] Optionally, the target depth unfolding network is constructed based on the following method:

[0067] The target depth-deployed network is used to solve either a smooth optimization problem or a non-smooth optimization problem for the target wireless network resource management optimization problem.

[0068] Based on the type of target wireless network resource management optimization problem, a corresponding expandable iterative scheme is constructed, which includes hyperparameters;

[0069] Based on the expandable iterative format, a target depth unfolding network is constructed.

[0070] Specifically, the actual wireless network resource management optimization problem may be a smooth optimization problem or a non-smooth optimization problem. When constructing a target depth unfolded network for solving a certain type of wireless network resource management optimization problem, we can first determine whether the wireless network resource management optimization problem is a smooth optimization problem or a non-smooth optimization problem, and then construct an unfoldable iterative format in different ways according to the type of wireless network resource management optimization problem.

[0071] The expandable iterative format includes the network's hyperparameters, such as the iteration step size for the optimization variable, the iteration step size for the dual variable, and penalty parameters. In some embodiments, the expandable iterative format includes the iteration format for the optimization variable and the iteration format for the dual variable, and the corresponding network hyperparameters may include the iteration step size for the optimization variable and the iteration step size for the dual variable.

[0072] Based on the constructed expandable iterative scheme, the forward propagation formula of the target depth expandable network can be obtained, thereby constructing the target depth expandable network.

[0073] By constructing corresponding expandable iterative formats based on the types of target wireless network resource management optimization problems, the construction process of target deep expandable networks can be simplified in a targeted manner, thereby improving the construction efficiency of target deep expandable networks.

[0074] Optionally, based on the type of the target wireless network resource management optimization problem, a corresponding expandable iterative format is constructed, including:

[0075] Given that the target wireless network resource management optimization problem satisfies the first condition, a corresponding expandable iterative format is constructed based on the type of the target wireless network resource management optimization problem;

[0076] The first condition includes:

[0077] For each constraint function in the target wireless network resource management optimization problem, there exist two corresponding convex functions, the difference of which equals the constraint function; and...

[0078] For the objective function of the target wireless network resource management optimization problem, there exists a differentiable and convex function and a convex but non-differentiable function. The linear weighted sum of the differentiable and convex function and the convex but non-differentiable function equals the objective function.

[0079] Specifically, before constructing the expandable iterative format, the present invention first needs to determine whether the target wireless network resource management optimization problem satisfies the first condition. If the first condition is satisfied, the corresponding expandable iterative format can be constructed based on the type of the target wireless network resource management optimization problem. Otherwise, the target wireless network resource management optimization problem can be appropriately transformed (e.g., introducing new optimization variables, constraint functions, objective functions, etc.). If the transformed optimization problem satisfies the first condition, the corresponding expandable iterative format can be constructed.

[0080] For example, consider the following general form of optimization problem:

[0081]

[0082] In the formula, x represents the optimization variable, f(x) and R(x) represent functions related to x, ρ is a constant, f(x) + ρR(x) represents the objective function of the optimization problem, and the optimization problem is to optimize x to minimize the objective function. i (x) represents the constraint function, and the subscript i is used to distinguish multiple different constraint functions. The set of indices representing the constraint functions is a set of positive integers.

[0083] If the optimization problem of equation (1) satisfies the first condition, then it means:

[0084] 1. Each constraint function g i (·) can be written as the difference between two convex functions, i.e., g i (·)=c i (·)-d i (·), otherwise, new optimization variables can be introduced to make the new optimization problem satisfy this condition. Where, c i (·) and d i (·) are all convex functions, and d i (·) may be a zero function (corresponding to g) i (·) represents a convex constraint.

[0085] 2. f(·) is a differentiable and convex function. Otherwise, the new optimization problem can satisfy this condition by introducing a new optimization variable t, a constraint function f(x)≤t, and an objective function t+ρR(x).

[0086] 3. R(·) is a convex but non-differentiable function. For example, typical functions include the classical L1 norm, etc.

[0087] Under the above conditions, the optimization problem of equation (1) can be rewritten as the following optimization problem:

[0088]

[0089] By determining whether the target wireless network resource management optimization problem meets the first condition, and if so, constructing a corresponding expandable iterative format based on the type of the target wireless network resource management optimization problem, the solution accuracy and construction efficiency of deep unfolded networks can be improved.

[0090] Optionally, based on the type of the target wireless network resource management optimization problem, a corresponding expandable iterative format is constructed, including:

[0091] When the target wireless network resource management optimization problem is a smooth optimization problem, the primal dual method or the primal dual method and gradient descent method are used to construct the expandable iterative form corresponding to the target wireless network resource management optimization problem.

[0092] Specifically, if the target wireless network resource management optimization problem is a smooth optimization problem, the primal dual method can be used directly, or the primal dual method combined with the gradient descent method can be used to construct an expandable iterative format, which can simplify the network structure and improve the construction efficiency of deep unfolded networks.

[0093] Specifically, a Lagrange dual function can be constructed by introducing dual variables to address the target wireless network resource management optimization problem, and an iterative format for the dual variables can be constructed based on this dual function.

[0094] In some embodiments, for optimization problems where the optimization variables are relatively easy to derive analytical expressions, the target depth expansion network can be constructed using only the iterative format of the dual variables to simplify the network structure.

[0095] In some embodiments, for optimization problems where the optimization variables do not have analytical expressions or have complex analytical forms, the gradient descent method can be used for multi-step iterations, and the iterative format of the optimization variables can be constructed based on the gradient of the dual function with respect to each optimization variable.

[0096] The following examples illustrate specific application scenarios.

[0097] To address the multi-user precoding problem in MIMO wireless communication, this invention discloses a deep unfolding learning optimization algorithm with low computational and low sample complexity. The algorithm mainly includes the following steps:

[0098] Step 1: Let P represents the set of users served by the base station. b h represents the total transmit power of the base station. u and x u Let represent the channel vector and precoding vector of user u, respectively. Then, the precoding optimization problem can be modeled as:

[0099]

[0100] In the formula, σ 2 This indicates the power of the noise signal in the received signal.

[0101] Step 1.1: Since the optimization problem in equation (3) does not involve non-smooth optimization, there is no need to introduce new optimization variables and equality constraints. However, since this optimization problem does not satisfy the form of the difference between two convex functions, a transformation is still required. Introduce new optimization variables {p} u ,q u Then the optimization problem of equation (3) is equivalent to the following problem:

[0102]

[0103] Step 1.2: Solving the optimization problem of equation (4) is very difficult and requires an iterative approach. Let x u,n and q u,n Describes the optimization variable x u and q u In the nth iteration, x n+1 It can be obtained by solving the optimization problem of equation (5).

[0104]

[0105] Step 2: Since there are no non-smooth constraints, the optimization problem of equation (5) can be directly constructed by using the original dual method and gradient descent method to construct an expandable iterative scheme, as follows.

[0106] Step 2.1: For the optimization problem of equation (5), the corresponding Lagrange function is:

[0107]

[0108] In the formula, λ and υ u and τ u They represent the dual variables corresponding to the inequality constraints, and the mathematical symbol Re represents the real part of the complex number.

[0109] Step 2.2: L({x u ,p u ,q u The gradient calculation for each optimization variable is as follows:

[0110]

[0111] In the formula, I represents the identity matrix.

[0112] Step 2.3: The expandable iterative format for the optimization variables is as follows:

[0113]

[0114] In the formula, a u,n,m,l x represents u,n,m The step size parameter, x, for the l-th gradient descent iteration. u,n,m,l x represents u,n,m The result obtained from the l-th gradient descent iteration, x u,n,m x represents u,n The vector corresponding to the m-th dual iteration, b u,n,m p u,n The iteration step size corresponding to the m-th dual iteration, p u,n,m p u,n The iteration result corresponding to the m-th dual iteration, c u,n,m q u,n The iteration step size corresponding to the m-th dual iteration, q u,n,m q u,n The iteration result corresponding to the m-th dual iteration.

[0115] The expandable iterative format for the dual variable is as follows:

[0116]

[0117] In the formula, α n,m λ n The iteration step size of the m-th iteration, λ n,m+1 λ n The result of the m-th iteration, λ n This indicates that λ corresponds to the optimization variable x. u The vector in the nth iteration, ζ u,n,m Indicates υ u,n The iteration step size of the m-th iteration, υ u,n,m+1 Indicates υ u,n The result of the m-th iteration, υ u,n Indicates υ u Corresponding optimization variable x u The vector in the nth iteration, ξ u,n,m Represents τ u,n The iteration step size τ of the m-th iteration. u,n,m+1 Represents τ u,n The result of the m-th iteration. m = 1, ..., M, where M represents the total number of dual iterations, and l = 1, ..., B, where B represents the total number of gradient descent iterations.

[0118] Step 3: Construct the original dual network, i.e., determine the forward propagation formula and select learnable parameters.

[0119] Step 3.1: Construct the primal dual network based on the calculated gradient vectors. The forward propagation formula for updating the primal variables is shown in equation (8) above, and the forward propagation formula for updating the dual variables is shown in equation (9) above.

[0120] Step 3.2: For the MIMO multi-user precoding optimization problem under consideration, the following parameters can be learned:

[0121]

[0122] In the formula, K represents the optimization variable x u The total number of iterations.

[0123] Step 4: Collect training samples and use techniques such as backpropagation and layer-by-layer training in machine learning to optimize and train the selected learnable parameters.

[0124] Step 4.1: The optimization problem of equation (5) is a convex optimization problem, which can be solved directly using the interior point method. To solve the optimization problem of equation (3) (or the equivalent problem equation (4)), it is necessary to construct the optimization problem of equation (5) based on the obtained new points and solve it. Repeat this process until convergence, and the optimal solution of the optimization problem of equation (3) can be obtained.

[0125] Step 4.2: Sample the channel vector And construct the optimization problem of equation (3), and solve the constructed optimization problem to obtain the optimal solution. but These constitute valid training sample points. Repeating this process yields the training set. Where T represents the total number of training sample points.

[0126] Step 4.3: Define the loss function as mean squared error, and use mini-batch optimization to optimize the learnable parameters.

[0127] Step 5: When faced with a new optimization problem in the same or similar environment (i.e., channel state information is sampled in the same or similar environment), a high-quality approximate solution to the corresponding optimization problem is obtained directly through forward propagation calculation using a trained deep unfolded network (in this embodiment, the original dual network described above).

[0128] Optionally, based on the type of the target wireless network resource management optimization problem, a corresponding expandable iterative format is constructed, including:

[0129] When the target wireless network resource management optimization problem is a non-smooth optimization problem, it is transformed into an alternating optimization problem of a smooth optimization subproblem and a non-smooth optimization subproblem.

[0130] Construct expandable iterative schemes for smooth optimization subproblems using primal-dual methods or a combination of primal-dual methods and gradient descent methods; construct expandable iterative schemes for non-smooth optimization subproblems using the proximity operator method.

[0131] Specifically, if the target wireless network resource management optimization problem is a non-smooth optimization problem, it can first be transformed into an alternating optimization problem of a smooth optimization subproblem and a non-smooth optimization subproblem. For example, new optimization variables can be introduced and augmented Lagrange penalty functions and alternating optimization methods can be used to divide the original optimization problem into two subproblems: non-smooth optimization and smooth optimization.

[0132] Then, for smooth optimization subproblems, expandable iterative schemes can be constructed using the primal-dual method or a combination of the primal-dual method and gradient descent, as described above. For non-smooth optimization subproblems, expandable iterative schemes can be constructed using the proximity operator method.

[0133] By dividing the non-smooth optimization problem into two subproblems, non-smooth optimization and smooth optimization, and solving them alternately, the expansion iterative schemes corresponding to the smooth optimization subproblems can be constructed using the primal dual method or the primal dual method and gradient descent method, while the expansion iterative schemes corresponding to the non-smooth optimization subproblems can be constructed using the proximity operator method. This can effectively improve the construction efficiency of deep expansion networks for solving non-smooth optimization problems.

[0134] Optionally, based on an expandable iterative format, a target depth unfolded network is constructed, including:

[0135] Based on the expandable iterative scheme corresponding to the smooth optimization subproblem, the original dual subnetwork is constructed.

[0136] Based on the original dual subnetwork and the expandable iterative scheme corresponding to the non-smooth optimization subproblem, a target depth expansion network is constructed.

[0137] Specifically, after constructing expandable iterative formats, the forward propagation formula of the network can be determined based on these expandable iterative formats, thereby constructing the target depth expandable network.

[0138] When constructing the target depth unfolded network, we can first determine the forward propagation formula of the original dual subnetwork based on the unfoldable iterative format corresponding to the smooth optimization subproblem, thereby constructing the original dual subnetwork. Then, based on the original dual subnetwork and the unfoldable iterative format corresponding to the non-smooth optimization subproblem, we can determine the forward propagation formula of the final target depth unfolded network and construct the final target depth unfolded network.

[0139] By first constructing the original dual subnetwork, and then constructing the target deep unfolded network based on the unfoldable iterative format corresponding to the original dual subnetwork and the non-smooth optimization subproblem, the construction efficiency of the deep unfolded network can be effectively improved, and the training efficiency of the deep unfolded network can also be improved.

[0140] Taking the optimization problem of the previous formula (2) as an example, the main steps are as follows.

[0141] Step 1: If the optimization problem is a non-smooth optimization problem, introduce new variables and use the augmented Lagrange penalty function and the alternating optimization method to divide the original optimization problem into two sub-problems: non-smooth optimization and smooth optimization.

[0142] Step 1.1: If the function R(·) is not a zero function, we can introduce the optimization variable z and the equality constraint xz, which leads to the following equivalent optimization problem:

[0143]

[0144] By further introducing a penalty parameter κ and a dual variable y to address the equality constraints, the penalty form of the optimization problem in equation (11) can be obtained:

[0145]

[0146] Generally speaking, directly solving the optimization problem of equation (12) is difficult and requires an iterative approach. Let x n Let x represent the nth iteration of the optimization variable x (other variables or parameters can be defined similarly), then x in the (n+1)th iteration... n+1 This can be obtained by solving the following optimization problem:

[0147]

[0148] In the formula, y n κ n Let represent the nth iteration of the dual variable y and the penalty parameter κ, respectively.

[0149] Step 1.2: Solve the optimization problem of equation (13) using the alternating optimization method. Then, the optimization subproblem with respect to the optimization variable x (this problem is a smooth optimization problem) is:

[0150]

[0151] The optimization subproblem concerning the optimization variable z (this is a non-smooth optimization problem) is as follows:

[0152]

[0153] The optimization problem of equation (15) can be equivalently written as the following optimization problem:

[0154]

[0155] Step 1.3: In general, the non-smooth optimization subproblem (16) with respect to variable z has an analytical expression:

[0156]

[0157] In the formula, z * Let prox represent the analytical solution of variable z. h The nearest neighbor operator for function h is:

[0158]

[0159] In the formula, dom h represents the domain of function h.

[0160] Step 2: For smooth optimization subproblems, construct expandable iterative schemes using the primal dual method and gradient descent method; for non-smooth optimization subproblems, construct expandable iterative schemes using the proximity operator method.

[0161] Step 2.1: Since the optimization subproblem of equation (14) is a convex optimization problem, the strong duality theorem holds under certain constraint specifications, so the primal duality method can be used to solve it. For convenience, J(x) and φ are introduced. i (x), that is:

[0162]

[0163] The Lagrange dual function of the optimization subproblem of equation (14) is:

[0164]

[0165] In the formula, μ i To constrain φ i The dual variable of (x)≤0, and L(x,μ) is defined as The mathematical symbol inf represents the infimum of a function, and the mathematical symbol card represents the number of elements in a set.

[0166] The dual problem of the optimization subproblem of equation (14) is defined as:

[0167]

[0168] Step 2.2: Solve the convex optimization problem (14) using the primal dual method. For any given initial dual variables... Construct the primal dual sequence {μ} as follows: m} and {x m}, that is, given μ m ,have:

[0169]

[0170] In the formula, μ m Let x represent the m-th iteration of the dual variable μ. mLet α represent the m-th iteration of the optimization variable x. m,i μ m,i The step size parameter for the m-th iteration, μ m,i Represents vector μ m The i-th component, m = 1, ..., M, where M represents the total number of dual iterations.

[0171] Step 2.3: If the optimization problem If the solution does not have an analytical expression or its form is very complex, then the gradient descent method is used for iterative solution, and multi-step iteration is employed to obtain better performance. The iteration format is as follows:

[0172]

[0173] In the formula, η m,l x represents m The step size parameter, x, for the l-th gradient descent iteration. m,l+1 x represents m The result obtained in the l-th step of gradient descent iteration L(x) represents m,l ,μ m Let l = 1, ..., B, where B represents the total number of gradient descent iterations.

[0174] Step 3: Select scalar parameters such as iteration step size and penalty parameter in the iteration format as learnable parameters, and allow constraints of the same type to share learnable parameters to further reduce sample complexity.

[0175] Step 3.1: Construct the primal dual subnetwork to solve the optimization subproblem of equation (14). The forward propagation formula of the primal dual subnetwork is:

[0176]

[0177] Figure 2 This is a schematic diagram of the original dual subnetwork provided by the present invention, as shown below. Figure 2 As shown, the forward propagation formula of the original dual subnetwork is Equation (24), and the analytical expression of the optimization variable x can be obtained through M dual iterations. * Accordingly, the set of learnable parameters of the original dual subnetwork is denoted as . The definition is as follows:

[0178]

[0179] In the formula, I represents the vector μ m The number of components.

[0180] Step 3.2: Based on the original dual subnetwork described above, a composite optimization network is further constructed (for solving the optimization problem in equation (2)). The forward propagation format of the composite optimization network is (iteration K times):

[0181]

[0182] In the formula, PDSN(·,·) represents the original dual subnetwork, z n This represents the nth iteration of the optimization variable z. The learnable parameters Θ consist of two parts: 1) the set of learnable parameters of the original dual subnetwork. 2) The set of penalty parameters {κ} n |n=1,…,K-1}. The network structure of the composite optimization network is as follows: Figure 3 As shown.

[0183] Step 4: Collect training samples and use techniques such as backpropagation and layer-by-layer training in machine learning to optimize and train the selected learnable parameters Θ.

[0184] Step 4.1: Training samples can be obtained in a variety of ways. One method is to use the previously constructed iterative format and the augmented Lagrange penalty function method to solve the optimization problem (11).

[0185] Step 4.1.1: Alternately solve the optimization subproblems (14) and (15) (or equivalent subproblems (16)). The optimization subproblem (14) can be solved precisely by methods such as the interior point method, and the optimization subproblems (15) or (16) have closed expressions or can be solved precisely by iteration of the nearest neighbor operator.

[0186] Step 4.1.2: Since the augmented Lagrange penalty function method is used, the dual variable y and the penalty parameter κ also need to be updated. The update formula is as follows:

[0187]

[0188] In the formula, β∈(1,∞) and γ∈(0,1) are constants used to control the growth rate of the penalty parameter κ. The convergence condition of the algorithm is that the norm of the difference between two consecutive iterations of the optimization variable x is less than a pre-specified positive real number.

[0189] Step 4.2: Assuming the obtained training set Where T represents the total number of training sample points. For any sample point This represents the input or parameter vector of the optimization problem. Mathematically, a composite optimization network can be viewed as a mapping, denoted as... Accordingly, the training loss function is defined as:

[0190]

[0191] Step 4.3: Based on the defined loss function, use the backpropagation algorithm to train and optimize the parameters Θ.

[0192] Step 5: When faced with a new optimization problem of the same type, a high-quality approximate solution to the corresponding optimization problem can be obtained directly through a trained deep unfolding network (i.e., the composite optimization network mentioned above) via forward propagation with extremely low computational complexity.

[0193] Alternatively, the training methods for the target depth unfolded network include:

[0194] Fix all elements in Θ except Θ1, update Θ1 based on the training samples and the loss function, then fix all elements in Θ except Θ2, update Θ2 based on the training samples and the loss function, and so on. For any iteration of Θ... k The updates are all fixed except for Θ. k All other elements remain unchanged, only Θ is updated. k ;or,

[0195] Fix all elements in Θ except Θ1, update Θ1 based on the training samples and the loss function, then fix all elements in Θ except {Θ1,Θ2}, update {Θ1,Θ2} based on the training samples and the loss function, and so on. For any {Θ1,Θ2,...,Θ k The updates of} are all fixed in Θ except for {Θ1,Θ2,...,Θ} k All elements except {Θ1,Θ2,...,Θ} remain unchanged, only {Θ1,Θ2,...,Θ} are updated. k};

[0196] Where Θ represents the set of hyperparameters of the target depth unfolded network, Θ={Θ1,Θ2,...,Θ k ,...,Θ K}, Θ k Let K represent the set of hyperparameters of the k-th subnetwork of the target depth unfolded network, where K represents the number of subnetworks in the target depth unfolded network.

[0197] Specifically, the simplest way to train a target depth unfolded network is to directly update all elements in Θ each time. This invention provides a layered training method that can improve the efficiency of network training.

[0198] There are two methods for hierarchical training. Taking the composite optimization network training mentioned earlier as an example, the learnable parameters are denoted as... in Let represent the set of learnable parameters for the k-th subnetwork.

[0199] One hierarchical training method is as follows: First, fix all elements in Θ\Θ1, and update the elements in Θ1 based on the training samples and the loss function (this can be done once or multiple times; this invention does not impose any restrictions. The update of other learnable parameter sets is similar and will not be elaborated further below). Then, fix all elements in Θ\Θ2, and update the elements in Θ2 based on the training samples and the loss function. This process is repeated, fixing Θ\Θ... k All elements remain unchanged while only Θ is updated. k Chinese elements.

[0200] The second layered training method: First, fix all elements in Θ\Θ1, and update only the elements in Θ1 based on the training samples and the loss function. Then, fix all elements in Θ\{Θ1,Θ2}, and update only the elements in {Θ1,Θ2} based on the training samples and the loss function. And so on, for each metric k, fix the set Θ\{Θ1,Θ2,…,Θ k In the expression {Θ1,Θ2,…,Θ}, all elements remain unchanged while only {Θ1,Θ2,…,Θ} is updated. k Elements in}

[0201] It should be noted that both of the above-mentioned hierarchical training methods can be used recursively, that is, when training any primal dual subnetwork PSDN... k (i.e., optimize parameter set) The above-mentioned layered training method can be used similarly in all cases.

[0202] The following example illustrates this through another specific application scenario.

[0203] To address the multi-user symbol-level precoding problem in MIMO wireless communication, this invention discloses a deep unfolding learning optimization algorithm with low computational and low sample complexity. The algorithm mainly includes the following steps:

[0204] Step 1: Let h represents the set of users served by the base station. u and ξ u Let represent the channel vector and PSK (Phase Shift Keying) symbol phase of user u, respectively, and let x represent the transmitted signal vector. Then, the precoding optimization problem can be modeled as:

[0205]

[0206] In the formula, F represents the pilot matrix (usually composed of beam training codebooks), ρ is a constant, and K... u Indicates the size of the constellation diagram for user u, γ u This represents the service quality-related index value for user u, j represents the imaginary unit, Im represents the imaginary part of the complex number, and Re represents the real part of the complex number.

[0207] Step 1.1: Since the optimization problem in equation (28) involves non-smooth optimization, a new optimization problem z and equality constraint xz need to be introduced. The corresponding equivalent optimization problem can be written as:

[0208]

[0209] To further address the equality constraints, a penalty parameter κ and a dual variable y are introduced, resulting in the corresponding augmented Lagrange penalty problem:

[0210]

[0211] Step 1.2: Solve the optimization problem of equation (30) using the alternating optimization method, where the smooth optimization subproblem is:

[0212]

[0213] The non-smooth optimization subproblem is:

[0214]

[0215] The optimal solution to the optimization problem of equation (32) is:

[0216] z * =sgn(x+y / κ)⊙max{|x+y / κ|-ρ / κ,0} (33)

[0217] In the formula, sgn(·) represents the complex symbol function, ⊙ represents the Hardmard product, and all the above operations are component-wise.

[0218] Step 2: Since the optimization problem of equation (32) has an optimal solution and the optimization problem of equation (31) is a convex optimization problem, the primal dual method can be directly used to construct an expandable iterative scheme. Consider the nth iteration, κ n This represents the corresponding penalty parameter. Other variables can be defined similarly. The expandable iterative format construction method is as follows.

[0219] Step 2.1: Let x = x R +jx I and Where, x R and x I Let c represent the real and imaginary parts of x, respectively; the other variables are defined similarly. u =tan(π / K) u ), Q R =2Re(F H F)+κ n I, Q I =2Re(F H F) and

[0220] Step 2.2: To obtain a more manageable form, the matrices and vectors are further defined as follows:

[0221]

[0222] Then we can obtain the optimization problem corresponding to the nth iteration:

[0223]

[0224] Step 2.3: The Lagrange function for the optimization problem of equation (35) is:

[0225]

[0226] In the formula, μ represents the inequality constraint. The dual variable.

[0227] Step 3: Construct the original dual subnetwork based on the obtained Lagrange function, i.e., determine the forward propagation formula and select learnable parameters.

[0228] Step 3.1: The original dual update formula is:

[0229]

[0230] In the formula, Let represent the i-th row of matrix A, index n correspond to the n-th primal dual subnetwork in the composite optimization network, index m correspond to the iteration within the primal dual subnetwork, index i correspond to the row vector of matrix A, and α n,m,i μ n,i The step size parameter for the m-th iteration, μ n,m,i Represents vector μ n,m The i-th component, b i Let i represent the i-th component of vector b.

[0231] Step 3.2: The learnable parameters are as follows:

[0232] Θ={α n,m,i ,κ n |1≤n≤K,1≤m≤M,1≤i≤2U} (38)

[0233] Step 4: Collect training samples and use techniques such as backpropagation and layer-by-layer training in machine learning to optimize and train the selected learnable parameters Θ.

[0234] Step 4.1: Note that the optimization problem of equation (28) is a convex optimization problem, which can be solved directly using the interior-point method. Sampling channel vector And construct the optimization problem equation (28), solve this problem to obtain the optimal solution x, then These constitute sample points. Repeating this process yields the training set.

[0235] Step 4.3: Define the loss function as mean squared error, and optimize the learnable parameters Θ using mini-batch processing and backpropagation. Since the number of learnable parameters is small, the backpropagation method can be used directly.

[0236] Step 5: When faced with a new optimization problem under the same or similar environment (i.e., channel state information is sampled under the same or similar environment), a high-quality approximate solution to the corresponding optimization problem is obtained directly through forward propagation.

[0237] To demonstrate the advantages of the deep unfolded learning optimization method proposed in this invention, for two types of precoding (i.e., classic block precoding and symbol-level precoding), this invention provides the performance of different precoding optimization algorithms, including Normalized Mean Square Error (NMSE), Average Achievable Sum-rate (AMSE), and Symbol Error Rate (SER). The channel model used is an FDD (Frequency Division Duplex) massive MIMO channel model with U=3 users. The channel angle extensions for different users are as follows (range [-1, 1]):

[0238] Case 1: [-33 / 64, -27 / 64], [-3 / 64, 3 / 64], [27 / 64, 33 / 64];

[0239] Case 2: [-20 / 64, -14 / 64], [-3 / 64, 3 / 64], [14 / 64, 20 / 64];

[0240] Case 3: [-24 / 64, -13 / 64], [-6 / 64, 6 / 64], [13 / 64, 24 / 64].

[0241] All learning optimization models are trained based on the channel model in Case 1, where D represents the size of the training set when training the learning model. Figures 4-8 All algorithm illustrations in the diagram are suffixed with -AU to indicate that the corresponding algorithm is designed based on the depth unfolding method proposed in this invention. The other algorithms are different baseline algorithms and are designed entirely based on the optimization method.

[0242] Figure 4 The average reachability performance curves of different algorithms provided in this invention are shown below. Figure 4As shown in the figure, the average achievable rate performance curves of the precoding algorithm based entirely on optimization and the deep unfolding precoding algorithm proposed in this invention are displayed. It can be seen that the performance of the precoding algorithm based on the deep unfolding method proposed in this invention is very close to that of the precoding algorithm based entirely on optimization. However, the precoding algorithm based on the deep unfolding method proposed in this invention has extremely low computational complexity, only requiring the first-order information of the objective function and constraint function, and only 3 to 4 iterations are needed.

[0243] Figure 5 The normalized mean square error performance curves of different algorithms provided in this invention are shown below. Figure 5 As shown, with the same number of iterations, the precoding algorithm based on the depth unfolding method proposed in this invention exhibits lower mean square error performance, meaning it approximates the optimal solution more closely. It can also be seen that good performance can be achieved even with a small number of training samples, indicating lower sample complexity.

[0244] Figure 6 The symbol error rate performance curves of different algorithms provided in this invention are shown in the figure. Figure 6 As shown, the precoding algorithm based on the deep unrolling method proposed in this invention has performance very close to that of the precoding algorithm designed entirely based on optimization methods. When the estimated channel state information is inaccurate, the precoding algorithm based on the deep unrolling method proposed in this invention can even achieve better performance.

[0245] Figure 7 The generalization performance curves of different algorithms provided in this invention are shown in the figure. Figure 7 As shown, even though the training model is based on the channel model in Case 1, it still has good symbol error rate performance for the other two channel models. That is, the deep unfolding method proposed in this invention has very good generalization performance.

[0246] Figure 8 The performance curves of different algorithms provided in this invention are shown. One of the important advantages of the algorithm designed based on the depth unfolding idea is its very low computational complexity. Figure 8 This important advantage is visually demonstrated, whether or not Figure 8 Case (a) in the middle is still Figure 8 In case (b) above, the computational complexity of the precoding algorithm based on the deep unfolding method proposed in this invention is significantly lower than that of other algorithms. The advantage of the precoding algorithm designed in this invention in terms of computational complexity comes not only from deep unfolding, but also from the fact that the constructed deep unfolding network only uses the first-order information of the objective function and constraint function, requiring very few iterations, thus resulting in low computational complexity.

[0247] The wireless network resource management optimization device provided by the present invention is described below. The wireless network resource management optimization device described below can be referred to in correspondence with the wireless network resource management optimization method described above.

[0248] Figure 9 This is a schematic diagram of the wireless network resource management optimization device provided by the present invention, as shown below. Figure 9 As shown, the device includes:

[0249] The acquisition module 900 is used to determine the wireless network resource management optimization problem and acquire the input parameter vector of the wireless network resource management optimization problem;

[0250] The solver module 910 is used to input the input parameter vector into the target depth unfolded network to obtain the solution to the wireless network resource management optimization problem;

[0251] The resource management optimization module 920 is used to optimize the management of wireless network resources based on the solution results.

[0252] The target depth unfolding network is obtained by training its hyperparameters based on the sample input parameter vector of the wireless network resource management optimization problem and the solution results of the optimization problem corresponding to the sample input parameter vector. The hyperparameters include one or more of the following: the iteration step size of the optimization variable, the iteration step size of the dual variable, and the penalty parameter.

[0253] Optionally, the target depth unfolding network is constructed based on the following method:

[0254] The target depth-deployed network is used to solve either a smooth optimization problem or a non-smooth optimization problem for the target wireless network resource management optimization problem.

[0255] Based on the type of target wireless network resource management optimization problem, a corresponding expandable iterative scheme is constructed, which includes hyperparameters;

[0256] Based on the expandable iterative format, a target depth unfolding network is constructed.

[0257] Optionally, based on the type of the target wireless network resource management optimization problem, a corresponding expandable iterative format is constructed, including:

[0258] When the target wireless network resource management optimization problem is a smooth optimization problem, the primal dual method or the primal dual method and gradient descent method are used to construct the expandable iterative form corresponding to the target wireless network resource management optimization problem.

[0259] Optionally, based on the type of the target wireless network resource management optimization problem, a corresponding expandable iterative format is constructed, including:

[0260] When the target wireless network resource management optimization problem is a non-smooth optimization problem, it is transformed into an alternating optimization problem of a smooth optimization subproblem and a non-smooth optimization subproblem.

[0261] Construct expandable iterative schemes for smooth optimization subproblems using primal-dual methods or a combination of primal-dual methods and gradient descent methods; construct expandable iterative schemes for non-smooth optimization subproblems using the proximity operator method.

[0262] Optionally, based on an expandable iterative format, a target depth unfolded network is constructed, including:

[0263] Based on the expandable iterative scheme corresponding to the smooth optimization subproblem, the original dual subnetwork is constructed.

[0264] Based on the original dual subnetwork and the expandable iterative scheme corresponding to the non-smooth optimization subproblem, a target depth expansion network is constructed.

[0265] Optionally, based on the type of the target wireless network resource management optimization problem, a corresponding expandable iterative format is constructed, including:

[0266] Given that the target wireless network resource management optimization problem satisfies the first condition, a corresponding expandable iterative format is constructed based on the type of the target wireless network resource management optimization problem;

[0267] The first condition includes:

[0268] For each constraint function in the target wireless network resource management optimization problem, there exist two corresponding convex functions, the difference of which equals the constraint function; and...

[0269] For the objective function of the target wireless network resource management optimization problem, there exists a differentiable and convex function and a convex but non-differentiable function. The linear weighted sum of the differentiable and convex function and the convex but non-differentiable function equals the objective function.

[0270] Alternatively, the training methods for the target depth unfolded network include:

[0271] Fix all elements in Θ except Θ1, update Θ1 based on the training samples and the loss function, then fix all elements in Θ except Θ2, update Θ2 based on the training samples and the loss function, and so on. For any iteration of Θ... k The updates are all fixed except for Θ. k All other elements remain unchanged, only Θ is updated. k ;or,

[0272] Fix all elements in Θ except Θ1, update Θ1 based on the training samples and the loss function, then fix all elements in Θ except {Θ1,Θ2}, update {Θ1,Θ2} based on the training samples and the loss function, and so on. For any {Θ1,Θ2,...,Θ k The updates of} are all fixed in Θ except for {Θ1,Θ2,...,Θ} k All elements except {Θ1,Θ2,...,Θ} remain unchanged, only {Θ1,Θ2,...,Θ} are updated. k};

[0273] Where Θ represents the set of hyperparameters of the target depth unfolded network, Θ={Θ1,Θ2,...,Θ k ,...,Θ K}, Θ k Let K represent the set of hyperparameters of the k-th subnetwork of the target depth unfolded network, where K represents the number of subnetworks in the target depth unfolded network.

[0274] It should be noted that the device provided by the present invention can implement all the method steps implemented in the above method embodiments and can achieve the same technical effect. Therefore, the parts and beneficial effects that are the same as those in the method embodiments will not be described in detail here.

[0275] Figure 10 This is a schematic diagram of the structure of the electronic device provided by the present invention, such as... Figure 10 As shown, the electronic device may include a processor 1010, a communications interface 1020, a memory 1030, and a communication bus 1040, wherein the processor 1010, the communications interface 1020, and the memory 1030 communicate with each other via the communication bus 1040. The processor 1010 can call logical instructions in the memory 1030 to execute any of the wireless network resource management optimization methods provided in the above embodiments.

[0276] Furthermore, the logical instructions in the aforementioned memory 1030 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0277] It should be noted that the electronic device provided by the present invention can implement all the method steps implemented in the above method embodiments and can achieve the same technical effect. Therefore, the parts and beneficial effects that are the same as those in the method embodiments will not be described in detail here.

[0278] On the other hand, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to perform any of the wireless network resource management optimization methods provided in the above embodiments.

[0279] It should be noted that the non-transitory computer-readable storage medium provided by the present invention can implement all the method steps implemented in the above method embodiments and can achieve the same technical effect. Here, the parts that are the same as those in the method embodiments and the beneficial effects will not be described in detail.

[0280] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0281] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0282] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for optimizing wireless network resource management, characterized in that, include: Identify the wireless network resource management optimization problem and obtain the input parameter vector of the wireless network resource management optimization problem; The input parameter vector is input into the target depth unfolded network to obtain the solution to the wireless network resource management optimization problem. Based on the solution results, optimize the management of wireless network resources; The target deep unfolding network is obtained by training its hyperparameters based on the sample input parameter vector of the wireless network resource management optimization problem and the solution results of the optimization problem corresponding to the sample input parameter vector; the hyperparameters include one or more of the following: the iteration step size of the optimization variable, the iteration step size of the dual variable, and the penalty parameter. The target depth unfolding network is constructed based on the following method: The target wireless network resource management optimization problem used by the target depth unfolded network to solve is determined to be either a smooth optimization problem or a non-smooth optimization problem. Based on the type of the target wireless network resource management optimization problem, a corresponding expandable iterative format is constructed, wherein the expandable iterative format includes the hyperparameters; Based on the aforementioned expandable iterative format, the target depth unfolded network is constructed; The construction of a corresponding expandable iterative format based on the type of the target wireless network resource management optimization problem includes: When the target wireless network resource management optimization problem is a non-smooth optimization problem, it is transformed into an alternating optimization problem of a smooth optimization subproblem and a non-smooth optimization subproblem. The expandable iterative scheme corresponding to the smooth optimization subproblem is constructed using the primal dual method or the primal dual method and gradient descent method, and the expandable iterative scheme corresponding to the non-smooth optimization subproblem is constructed using the proximity operator method.

2. The wireless network resource management optimization method according to claim 1, characterized in that, The construction of a corresponding expandable iterative format based on the type of the target wireless network resource management optimization problem includes: When the target wireless network resource management optimization problem is a smooth optimization problem, the expandable iterative form corresponding to the target wireless network resource management optimization problem is constructed using the primal dual method or the primal dual method and gradient descent method.

3. The wireless network resource management optimization method according to claim 1, characterized in that, The construction of the target depth unfolded network based on the unfoldable iterative format includes: Based on the expandable iterative format corresponding to the smooth optimization subproblem, the original dual subnetwork is constructed. Based on the original dual subnetwork and the expandable iterative format corresponding to the non-smooth optimization subproblem, the target depth expansion network is constructed.

4. The wireless network resource management optimization method according to claim 1 or 2, characterized in that, The construction of a corresponding expandable iterative format based on the type of the target wireless network resource management optimization problem includes: If the target wireless network resource management optimization problem satisfies the first condition, a corresponding expandable iterative format is constructed based on the type of the target wireless network resource management optimization problem; The first condition includes: For each constraint function in the target wireless network resource management optimization problem, there exist two corresponding convex functions, the difference of which is equal to the constraint function; and... For the objective function of the target wireless network resource management optimization problem, there exists a differentiable and convex function and a convex but non-differentiable function, and the linear weighted sum of the differentiable and convex function and the convex but non-differentiable function is equal to the objective function.

5. The wireless network resource management optimization method according to claim 1, characterized in that, The training methods for the target depth unfolded network include: fixed Except All other elements are updated based on the training samples and the loss function. Then fix Except All other elements are updated based on the training samples and the loss function. And so on, for any given time The updates are all fixed. Except All other elements remain unchanged, only updated. ;or, fixed Except All other elements are updated based on the training samples and the loss function. Then fix Except All other elements are updated based on the training samples and the loss function. And so on, for any given time The updates are all fixed. Except All other elements remain unchanged, only updated. ; in, This represents the set of hyperparameters of the target depth unfolded network. = , The first depth of the target unfolded network represents the... k The set of hyperparameters of each subnetwork, the K This represents the number of subnetworks in the target depth unfolded network.

6. The wireless network resource management optimization method according to claim 1, characterized in that, The optimization variables include: optimization variables in the original optimization problem, and optimization variables in the optimization problem transformed from the original optimization problem.

7. The wireless network resource management optimization method according to claim 1, characterized in that, The dual variables include: dual variables of inequality constraints in the original optimization problem, and dual variables of inequality constraints in the optimization problem transformed from the original optimization problem.

8. A wireless network resource management and optimization device, characterized in that, include: The acquisition module is used to determine the wireless network resource management optimization problem and acquire the input parameter vector of the wireless network resource management optimization problem; The solution module is used to input the input parameter vector into the target depth unfolded network to obtain the solution result of the wireless network resource management optimization problem; The resource management optimization module is used to optimize the management of wireless network resources based on the solution results. The target deep unfolding network is obtained by training its hyperparameters based on the sample input parameter vector of the wireless network resource management optimization problem and the solution results of the optimization problem corresponding to the sample input parameter vector; the hyperparameters include one or more of the following: the iteration step size of the optimization variable, the iteration step size of the dual variable, and the penalty parameter. The target depth unfolding network is constructed based on the following method: The target wireless network resource management optimization problem used by the target depth unfolded network to solve is determined to be either a smooth optimization problem or a non-smooth optimization problem. Based on the type of the target wireless network resource management optimization problem, a corresponding expandable iterative format is constructed, wherein the expandable iterative format includes the hyperparameters; Based on the aforementioned expandable iterative format, the target depth unfolded network is constructed; The construction of a corresponding expandable iterative format based on the type of the target wireless network resource management optimization problem includes: When the target wireless network resource management optimization problem is a non-smooth optimization problem, it is transformed into an alternating optimization problem of a smooth optimization subproblem and a non-smooth optimization subproblem. The expandable iterative scheme corresponding to the smooth optimization subproblem is constructed using the primal dual method or the primal dual method and gradient descent method, and the expandable iterative scheme corresponding to the non-smooth optimization subproblem is constructed using the proximity operator method.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the program, it implements the wireless network resource management optimization method as described in any one of claims 1 to 7.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the wireless network resource management optimization method as described in any one of claims 1 to 7.

Images