A method and system for wireless resource allocation in a MU-MIMO system based on a diffusion model

Abstract

The application discloses a kind of MU-MIMO system wireless resource allocation method and system based on condition diffusion strategy enhancement, for MU-MIMO system downlink transmission, method includes: with the total capacity of the MU-MIMO wireless communication system as target, jointly optimize resource block allocation and transmission power, construct optimization problem;Condition diffusion model is trained based on expert dataset;Multiple candidate solutions are generated using the trained model, and the optimal solution of the optimization problem is obtained by further optimizing the multiple candidate solutions through policy enhancement;By constructing a condition diffusion model, imitation learning is performed on the expert dataset;And introduce resampling mechanism, select the optimal solution from multiple candidate solutions, while ensuring to meet system constraints, significantly improve system throughput and reduce computational complexity.

Description

Technical Field

[0001] This invention belongs to the field of physical layer resource allocation for MU-MIMO communication systems. Specifically, it relates to a wireless resource allocation method and system for MU-MIMO systems based on a diffusion model. Background Technology

[0002] With the advancement of 6G research, multi-user multiple-input multiple-output (MU-MIMO) systems allow concurrent transmission to multiple users on a shared frequency band, significantly improving spectrum efficiency and becoming an important component of 6G technology. However, due to the high-dimensional decision space in this wireless communication scenario, existing resource allocation algorithms often struggle to simultaneously achieve good system performance and low complexity. On one hand, 6G networks extend from the ground to a three-dimensional space-air-ground integrated network, encompassing diverse scenarios such as ultra-reliable low-latency communication, extended reality, and terminal-side AI agent collaboration. Resource allocation requires the coordinated scheduling of multi-dimensional resources such as power, subcarriers, antenna ports, and precoding matrices, resulting in an extremely high-dimensional decision space. On the other hand, to meet the sub-millisecond latency requirements of 6G, resource allocation algorithms must possess rapid real-time response capabilities. Existing solutions often struggle to balance performance and complexity: traditional optimization algorithms based on accurate channel models can approximate optimal system performance, but their computational complexity is too high, often taking several seconds or even minutes to solve, failing to meet real-time requirements; while heuristic algorithms reduce the computational burden, they struggle to handle high-dimensional coupled non-convex optimization problems, leading to performance degradation in spectral efficiency and user fairness. In multi-antenna user scenarios, inter-user interference further complicates algorithm design. Furthermore, the increasingly scarce 6G low-frequency band spectrum resources and the drastic dynamic changes in high-frequency band channels further increase the difficulty of resource allocation strategy design, placing higher demands on the adaptability and robustness of algorithms.

[0003] The rapid development of generative artificial intelligence has provided new approaches to addressing complex high-dimensional optimization problems. In particular, generative diffusion models, with their progressive denoising mechanism, have demonstrated superior performance in high-dimensional data modeling. Unlike traditional generative models such as generative adversarial networks and variational autoencoders, diffusion models, through an iterative approach of gradually adding noise in the forward process and gradually denoising in the reverse process, can precisely characterize high-dimensional policy distributions, capture complex correlation features, and possess good generalization ability and stability. This allows for the rapid generation of optimal or suboptimal solutions that satisfy constraints without the need for complex iterative solutions. This characteristic perfectly addresses the two core requirements of 6G MU-MIMO resource allocation: addressing the modeling challenges of high-dimensional decision spaces while ensuring low algorithm complexity and real-time performance.

[0004] Generative diffusion models achieve refined modeling of high-dimensional policy distributions through a progressive denoising process, providing a new approach for real-time solutions to wireless resource allocation problems. Representing resource allocation policies in this scenario using diffusion models and leveraging the powerful generalization capabilities of generative models to achieve high system performance while maintaining low computational complexity presents a significant new challenge. Summary of the Invention

[0005] To address the problems existing in the prior art, this invention provides a wireless resource allocation method for MU-MIMO systems based on a diffusion model, which achieves high system performance while ensuring low computational complexity.

[0006] To achieve the above objectives, in a first aspect, the present invention provides a method for allocating radio resources in a MU-MIMO system based on a diffusion model, comprising the following steps: With the goal of maximizing the total capacity of the MU-MIMO wireless communication system, an optimization problem is constructed by jointly optimizing resource block allocation and transmission power;

[0007] in, This indicates the maximum carrier aggregation capability of the user terminal. This indicates the maximum aggregation capacity of resource blocks on the user terminal. For indicator functions, Indicates user Minimum communication rate requirement, This indicates the maximum transmission power of the base station; For user collection, The total number of users per antenna. The total number of resource blocks, For carrier set, In resource blocks Upward users The allocated power value, For users The total reachable rate; A conditional diffusion model is trained based on an expert dataset; multiple candidate solutions are generated using the trained model; and the optimal solution to the optimization problem is obtained by further selecting the best solution from the multiple candidate solutions through policy enhancement.

[0008] Furthermore, a conditional diffusion model is trained based on an expert dataset; multiple candidate solutions are generated using the trained model; and policy enhancement is used to further select the best solution from the multiple candidate solutions to obtain the optimal solution to the optimization problem, including: a) Obtaining datasets from optimization theory-based solution methods , For channel conditions, For power allocation matrix; b) Initialize the noise prediction network parameters Set the diffusion step number T and variance scheduling parameters. , The hyperparameter for controlling the diffusion intensity at time step t, where T is the total number of diffusion steps; c) Load from dataset Sampling time steps from a uniform distribution ; d) According to the formula for the forward diffusion process, towards Injection variance is Gaussian noise Obtain the power allocation matrix of injected Gaussian noise. The power allocation matrix to inject Gaussian noise Channel conditions as well as Input noise prediction neural network ; e) Calculate the loss function Update parameters ; f) Repeat steps c)-e) until the loss function is calculated. The signal converges, resulting in the trained noise prediction neural network. g) Sample N samples from a Gaussian distribution and set the current channel state. Input conditional diffusion model; h) According to Perform T reverse generation processes, using the trained noise prediction neural network to generate the corresponding power allocation strategy and obtain N candidate solutions. ; i) Select candidate solutions based on the constraints to obtain a set of feasible solutions. ; j) Calculate the action-state value function of a feasible solution. ; k) Based on Selected power allocation matrix To optimize the optimal solution to the problem.

[0009] Furthermore, the loss function The details are as follows;

[0010] loss function Equivalent to expert strategy distribution Distribution of diffusion strategies The Kullback–Leibler divergence between them.

[0011] Furthermore, the action-state value function is:

[0012] For the power allocation matrix, For users In resource block The achievable transmission rate is It is the set of all resource blocks for all carriers.

[0013] Furthermore, a noise prediction neural network based on a Transformer structure is used, with the input being a noisy power allocation matrix. Channel conditions And the diffusion time step t, specifically: the noisy power allocation matrix. With channel conditions The network is concatenated along the channel dimension, enabling it to simultaneously perceive the current power distribution state and the corresponding channel environment information. The concatenated features are initially encoded using convolutional layers, achieving feature dimensionality enhancement and extraction of local coupling relationships. The diffusion time step t is mapped to high-dimensional temporal features via a temporal embedding module and added to the main features through broadcasting. After normalization, these features are used as input to subsequent networks. The encoded features then enter the main network, composed of multiple Transformer layers. The output convolutional layer maps the features back to the original power distribution dimension, yielding the predicted noise. .

[0014] Furthermore, consider a coverage radius of A single-base station MU-MIMO communication system for circular cellular cells, including Single-antenna user and Each equipment machine, A base station with multiple carriers, the bandwidth of each carrier. They are all the same, and each carrier contains Each resource block, the wireless channel model between the base station and the user adopts a block fading model, and the channel state information is determined by path loss. Shadow decay With small-scale fading describe.

[0015] user In resource block The normalized beamforming vector on is:

[0016] Then the user In resource block The signal-to-interference-plus-noise ratio is:

[0017] Wherein, the numerator represents the user. The expected signal power, the denominator includes the interference power. With noise power , In resource blocks Upward users The allocated power value, user Total reachability The sum of its rates across all resource blocks:

[0018] This represents the bandwidth of each carrier.

[0019] Furthermore, this invention employs a conditional diffusion model to learn the expert policy distribution from the dataset. Among them, channel conditions It determines the power allocation matrix. The prior conditions for the distribution; The conditional diffusion model is trained through forward diffusion and backward generation processes to directly generate... During the forward diffusion process, the power allocation matrix is... Gradually inject Gaussian noise The forward diffusion process forms a Markov chain, directly from... get ; The ideal reverse generation process distribution is obtained based on Bayes' theorem and the forward diffusion process. ; The parameters used are A noise prediction neural network, under given channel conditions Predictive Noise in Mean estimation of the distribution of the reverse generation process:

[0020] in To predict noise, the distribution of the reverse generation process is as follows:

[0021] Let covariance matrix be the variance matrix. As a unit array, This is a hyperparameter for controlling the diffusion intensity at time step t.

[0022] Secondly, the present invention provides a single-base station MU-MIMO communication system, including... One single-antenna user and one equipped root antenna and A base station with multiple carriers, the bandwidth of each carrier. They are all the same, and each carrier contains Resource blocks; resource allocation is performed using the aforementioned MU-MIMO system wireless resource allocation method based on the diffusion model.

[0023] Thirdly, the present invention can also provide a computer device, including a processor and a memory, wherein the memory is used to store a computer executable program, the processor reads the computer executable program from the memory and executes it, and the processor can implement the wireless resource allocation method for MU-MIMO system based on diffusion model described in the present invention when executing the computer executable program.

[0024] A computer-readable storage medium is also provided, in which a computer program is stored. When the computer program is executed by a processor, it can implement the wireless resource allocation method for MU-MIMO systems based on the diffusion model described in this invention.

[0025] Compared with existing technologies, this invention has at least the following advantages: By constructing a conditional diffusion model and performing imitation learning pre-training on a high-quality expert dataset, this invention achieves the modeling of multimodal resource allocation strategies. In the generation stage, the diffusion model gradually generates multiple resource allocation solutions through a reverse denoising sampling process. This process involves large-scale tensor operations and neural network forward inference, which can fully utilize the parallel computing capabilities of modern GPUs to achieve batch parallel generation of multiple candidate resource allocation schemes, thereby significantly improving the generation efficiency of resource allocation schemes. Furthermore, this invention proposes a resampling technique based on the diffusion model, which effectively avoids the violation of constraints and improves the resource allocation effect, even surpassing the expert strategies it imitates. Attached Figure Description

[0026] Figure 1 The downlink transmission scenario of the MU-MIMO system considered in this invention.

[0027] Figure 2 This is a schematic diagram of the CDPE-RA algorithm in this invention.

[0028] Figure 3 This is a diagram of the neural network structure contained in the CDPE-RA algorithm of this invention.

[0029] Figure 4 This graph shows the change in total system throughput as a function of the number of users for the method of this invention and the comparative scheme.

[0030] Figure 5 This is a graph showing the variation of the total system throughput of the method of the present invention with the number of users on different expert datasets.

[0031] Figure 6 This graph shows the change in computation time of the method of the present invention and the comparative scheme as a function of the number of users. Detailed Implementation

[0032] The present invention will now be described in detail with reference to the accompanying drawings.

[0033] This invention considers a coverage radius of A single-base station MU-MIMO communication system in a circular cellular cell, such as Figure 1 As shown. The system includes Single-antenna user and Each equipment Root antenna, A base station with one component carrier (CC). The user set and the carrier set are represented as follows: and .

[0034] Assuming the bandwidth of each carrier They are all the same, and each carrier contains One resource block. Carrier. The Each resource block is represented as .in, This represents the set of all resource blocks across all carriers.

[0035] Considering the generally low mobility of users within a cell, the wireless channel model between the base station and users adopts a block fading model, which assumes that the channel is static within each channel coherence interval. Therefore, the base station and users... In resource block Channel state information It can be represented as:

[0036] That is, channel state information is determined by path loss. Shadow decay With small-scale fading describe.

[0037] Path loss characterizes the attenuation of signal power with propagation distance, and a logarithmic distance model is used:

[0038] in, For users Distance from base station, Indicates the reference distance. This is the path loss index, and its value depends on the scenario (typically, in densely populated urban areas). ), Indicates user Path loss of the channel.

[0039] Shadow decay Caused by obstruction, this invention assumes that it follows a normal distribution in the logarithmic domain:

[0040] in For users Shadow fading in the decibel range, Standard deviation (typical value) Shadow fading and path loss together constitute large-scale fading; they are added in the decibel domain and multiplied in the linear domain.

[0041] Small-scale fading Caused by multipath effects, it is modeled as a cyclic symmetric complex Gaussian random vector:

[0042] in The spatial covariance matrix represents the base station. Channel correlation between root antennas. Under the assumption of antenna independence. , yes An identity matrix of dimension 1.

[0043] For downlink data transmission, this invention employs the maximum ratio transmission (MRT) precoding scheme to achieve multi-user beamforming due to its advantages of low complexity and limited fronthaul overhead. Assuming the base station can obtain perfect channel state information, then the user... In resource block The normalized beamforming vector on can be designed as follows:

[0044] Channel vector The expected value of the square of the second norm, i.e. the average channel power, is used for normalization to ensure that the beam vector meets the power constraint. Represents the field of complex numbers The beamforming vector is a column vector space containing... A column vector of complex elements, corresponding to the base station Root transmitting antenna.

[0045] Then the user In resource block The signal-to-interference-plus-noise ratio (SINR) on the surface can be expressed as:

[0046] Wherein, the numerator represents the user. The expected signal power, the denominator includes the interference power. With noise power , In resource blocks Upward users The allocated power value. According to Shannon's theorem, the user... In resource block The achievable transmission rate is:

[0047] user Total reachability The sum of its rates across all resource blocks:

[0048] This invention aims to maximize the total capacity of the MU-MIMO wireless communication system by jointly optimizing resource block allocation and transmission power. .when When, it indicates to the user Resource blocks were allocated ,when At that time, it indicated that no one had been informed. Allocate resource blocks .

[0049] The optimization problem is modeled as follows:

[0050] in, This indicates the maximum carrier aggregation capability of the user terminal. This indicates the maximum aggregation capacity of resource blocks on the user terminal. For indicator functions, Indicates user Minimum communication rate requirement, This represents the maximum transmission power of the base station. In this problem, constraints 1 and 2 ensure that the number of carriers and resource blocks aggregated for each user does not exceed the maximum aggregation capacity of the user terminal, constraint 3 indicates that the allocation result should meet the minimum rate requirement of each user, and constraint 4 guarantees that the total power allocated to each resource block does not exceed the maximum transmit power of the base station.

[0051] like Figure 2 As shown, the resource allocation algorithm based on the conditional diffusion model (CDPE-RA) proposed in this invention includes two stages: the training stage and the deployment stage.

[0052] During the training phase, expert datasets are generated based on optimization algorithms. Two representative optimization schemes are used to generate two expert datasets.

[0053] (1) The hierarchical matching game based on particle swarm optimization (PSO-HMG) scheme determines the joint user association and frequency resource allocation by constructing a hierarchical matching game, and uses the particle swarm optimization algorithm for power control optimization.

[0054] (2) Greedy scheme, in which the user selects the best RB in turn and distributes the power evenly to the selected RB.

[0055] Each dataset contains 44,000 samples. Specifically, 40,000 samples are used for training, 2,000 for validation, and the remaining 2,000 for testing. The conditional diffusion model is trained using these samples. The deployment phase comprises two sub-processes: Conditional Diffusion-Based Resource Allocation (CDRA) generation and Policy Enhancement (PE). The CDRA process generates multiple candidate solutions using the trained model, and these candidate solutions are then further optimized through the PE process to improve the final system performance and ensure constraint satisfaction.

[0056] (1) Training phase: Each sample in the expert dataset contains equivalent channel information. and the corresponding power allocation matrix The proposed solution employs a conditional diffusion model to learn the expert policy distribution from the dataset. Among them, channel conditions It determines the power allocation matrix. The prior conditions for the distribution.

[0057] The conditional diffusion model is trained through forward diffusion and backward generation processes, enabling it to directly generate... During forward diffusion, the power allocation matrix... Gradually injecting Gaussian noise can be written in the following form:

[0058] in This represents the time step in the diffusion process, where T is the total number of diffusion steps. To control the hyperparameters of diffusion intensity at time step t, It is random Gaussian noise. It is a zero vector. It is a unit array.

[0059] Therefore, the forward diffusion process constitutes a Markov chain, corresponding to the process from... arrive The transition probability distribution can be written as:

[0060] This indicates that the corresponding from arrive The transition probability distribution has a mean of The covariance matrix is The Gaussian distribution.

[0061] Furthermore, based on the properties of Markov chains, it can be directly derived from... get ,Right now

[0062]

[0063] in .

[0064] because ,when Sometimes, and ,in This indicates that the distribution converges. Therefore, it is possible to sample from a Gaussian distribution. It performs T reverse generation processes and outputs the generated power allocation matrix. Based on Bayes' theorem and the formula for the forward diffusion process, the distribution of the ideal reverse generation process can be obtained. The expression for the power allocation matrix, where The prior conditions for this distribution are as follows.

[0065]

[0066] It is important to note that the distribution at each step of the ideal reverse generation process is still a Gaussian distribution, and its covariance matrix is... Completely determined by hyperparameters Decision. In the actual implementation phase, due to It is unavailable, therefore the mean of the ideal reverse generation process distribution cannot be obtained. Therefore, the parameters that can be used are: A noise prediction neural network, under given channel conditions Predictive Noise in The mean estimate of the distribution of the reverse generation process can then be written as:

[0067] in To predict noise.

[0068] Furthermore, the distribution of the reverse generation process can be estimated as follows:

[0069] To train this noise prediction neural network, its loss function is constructed as follows:

[0070] It can be proven that this loss function is equivalent to the expert policy distribution. Distribution of diffusion strategies The Kullback–Leibler (KL) divergence between them, i.e.:

[0071] This invention uses a noise prediction neural network based on the Transformer structure, such as... Figure 3 As shown, the multi-head attention layer is used to capture the long-range dependency between the power allocation matrix and channel conditions. The input to the noise prediction neural network is the noisy power allocation matrix. Channel conditions And the diffusion time step t, the output of the noise prediction neural network is the predicted noise. .like Figure 3 As shown, this invention employs a noise prediction neural network based on the Transformer architecture for noise estimation of noisy power allocation strategies in diffusion models. This network uses a noisy power allocation matrix... Channel conditions And the diffusion time step t is used as input, and the output is the same as... Same-dimensional prediction noise .

[0072] In the data processing process, the noisy power allocation matrix is ​​first... With channel conditions By concatenating features along the channel dimension, the network can simultaneously perceive the current power distribution state and the corresponding channel environment information. The concatenated features are initially encoded through convolutional layers to achieve feature dimensionality enhancement and extraction of local coupling relationships. The diffusion time step t is mapped to high-dimensional temporal features through a temporal embedding module, and then added to the main features via broadcasting. After normalization, it is used as the input to subsequent networks, thereby enabling the model to dynamically perceive different diffusion stages.

[0073] The encoded features are fed into the main network, which consists of a multi-layer Transformer structure. Each layer includes a multi-head attention module and a convolutional enhancement module, both employing residual connections and group normalization. The multi-head attention mechanism captures long-range dependencies between different users and between different resource blocks, enabling global correlation modeling between the power allocation matrix and channel conditions. The convolutional module enhances the expressive power of local structures, improving the ability to characterize the spatial properties of the power distribution. Through the synergistic effect of attention and convolutional structures, a fusion model of global and local features is achieved.

[0074] Finally, the output convolutional layer maps the features back to the original power allocation dimension to obtain the predicted noise. This is used in the reverse denoising process of the diffusion model to gradually recover a high-quality power allocation scheme that meets the constraints.

[0075] The training phase steps of the CDPE-RA scheme are shown in Algorithm 1.

[0076]

[0077] (2) Deployment phase: During the deployment phase, under given channel conditions The CDRA process then uses a trained noise prediction neural network to generate the corresponding power allocation matrix. The matrix is ​​obtained through a multi-round reverse generation process from step t=T to step t=1.

[0078] Specifically, in each step t, It can be calculated using the following formula:

[0079] in It is important to note that, The final result It follows the learned distribution. .

[0080] The aforementioned separate CDRA process is only intended to mimic the distribution of expert strategies. This aims to generate a power allocation matrix and its corresponding RB allocation scheme; however, due to dataset limitations, the result may not represent the original problem. The optimal solution. Due to the stochastic nature of the diffusion process, the resulting power allocation scheme may violate the original problem. The constraints.

[0081] To address these issues, this invention designs a PE process based on language model alignment technology to improve performance and handle constraint violations. It assumes the original problem of maximizing system throughput under imposed constraints. The globally optimal strategy is determined by Indicate, then This can be achieved by using the action-state value function. Distribution of learned diffusion strategies Weighted average, i.e.:

[0082] in Hyperparameters for enhancing the strength of the control strategy, when This indicates that no adjustments are made. The objective of the problem posed by this invention is to maximize the instantaneous total rate of all users without considering long-term returns. Therefore, in this invention, the action-state value function is designed as follows:

[0083] The PE process aims to distribute data through a sampling self-diffusion strategy. The optimal solution of the target distribution is approximated using a resampling method. Specifically, the PE process utilizes the CDRA process to distribute from a diffusion strategy. Generate N candidate solutions Subsequently, the generated candidate solutions are constrained and filtered to obtain a set of feasible solutions. Finally, the process calculates the feasible set. The action-state value of each candidate solution is calculated, and the candidate solution with the highest value is selected as the current channel condition. The resource allocation scheme is as follows. Therefore, the final determined power allocation matrix is... It can be written as: .

[0084] Based on the above discussion, it can be seen that the above formula yields... It is the optimal solution The approximation is given by the given condition, and when N is sufficiently large. The specific steps for the CDPE-RA deployment phase are shown in Algorithm 2.

[0085]

[0086] Numerical simulation and result analysis: This invention evaluates the performance of the proposed CDPE-RA scheme through simulation. In the simulation, users are randomly distributed within a circular area with a radius of 500 meters centered on the base station. The relevant simulation parameters are shown in Table 1.

[0087] Table 1 Simulation Parameters

[0088] Two representative optimization schemes are used to generate two expert datasets.

[0089] (1) The hierarchical matching game based on particle swarm optimization (PSO-HMG) scheme determines the joint user association and frequency resource allocation by constructing a hierarchical matching game, and uses the particle swarm optimization algorithm for power control optimization.

[0090] (2) Greedy scheme, in which the user selects the best RB in turn and distributes the power evenly to the selected RB.

[0091] Each dataset contains 44,000 samples. Specifically, 40,000 samples are used for training, 2,000 samples for validation, and the remaining 2,000 samples for testing.

[0092] To demonstrate the superiority of the proposed scheme, its performance is compared with that of the PSO-HMG scheme, the greedy scheme, and a supervised learning (SL) scheme. For SL, the same network structure as the noise prediction network is used and trained on the PSO-HMG dataset to generate power allocation solutions. Furthermore, to evaluate the effectiveness of the Policy Enhancement (PE) process, the performance of the CDPE-RA scheme is compared with that of the independent CDRA scheme without the PE process.

[0093] During the training phase, the learning rate is set to... The batch size was set to 1024, and the number of training epochs was 3000. The hidden dimension of the noise prediction neural network was set to 512, and the number of layers was... Set it to 4, and the number of attention heads to 8. (This is the correct setting.) The square cosine scheduling method is adopted.

[0094] Figure 4 The system throughput of the proposed CDPE-RA scheme and other comparative schemes is demonstrated. From Figure 4 It can be observed that the throughput of all schemes increases with the number of users. Furthermore, the throughput of the CDPE-RA scheme, trained on the PSO-HMG dataset, even surpasses that of the PSO-HMG scheme itself, reaching the highest level. Although the SL scheme uses the same neural network structure and has the same training dataset as the scheme proposed in this invention, the scheme proposed in this invention is still significantly superior to the SL scheme, demonstrating that generative diffusion models have a clear advantage over traditional supervised learning methods.

[0095] Figure 5 The system throughput comparison between the proposed CDPE-RA scheme and the CDRA scheme without the PE process is presented. The results show that CDPE-RA outperforms CDRA on both datasets, verifying the effectiveness of the constructed PE process.

[0096] Figure 6 The computation time for each scheme is shown under different numbers of users. From Figure 6 As can be seen, the proposed scheme significantly reduces computation time compared to the PSO-HMG scheme. Furthermore, although the greedy scheme and the SL scheme have slightly shorter computation times than the proposed scheme, their corresponding system throughput is significantly lower, indicating that the proposed scheme achieves an effective balance between complexity and performance.

[0097] Based on the above method, the present invention can also provide a single-base station MU-MIMO communication system, including... One single-antenna user and one equipped root antenna and A base station with multiple carriers, the bandwidth of each carrier. They are all the same, and each carrier contains Resource blocks; resource allocation is performed using the aforementioned MU-MIMO system wireless resource allocation method based on the diffusion model.

[0098] The present invention can also provide a computer device, including a processor and a memory, wherein the memory is used to store a computer executable program, the processor reads the computer executable program from the memory and executes it, and the processor can implement the wireless resource allocation method for MU-MIMO system based on diffusion model described in the present invention when executing the computer executable program.

[0099] The present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, can implement the wireless resource allocation method for MU-MIMO systems based on a diffusion model as described in the present invention.

[0100] The computer device may be a laptop, a desktop computer, or a workstation.

[0101] The processor can be a central processing unit (CPU), a digital signal processor (DSP), an application-specific integrated circuit (ASIC), or a field-programmable gate array (FPGA).

[0102] The memory described in this invention can be an internal storage unit of a laptop, desktop computer, or workstation, such as memory or hard disk; or it can be an external storage unit, such as a portable hard disk or flash memory card.

[0103] Computer-readable storage media can include computer storage media and communication media. Computer storage media includes volatile and non-volatile, removable and non-removable media implemented using any method or technology for storing information such as computer-readable instructions, data structures, program modules, or other data. Computer-readable storage media can include: read-only memory (ROM), random access memory (RAM), solid-state drives (SSDs), or optical discs, etc. Random access memory can include resistive random access memory (ReRAM) and dynamic random access memory (DRAM).

[0104] The above content is only for illustrating the technical concept of the present invention and should not be construed as limiting the scope of protection of the present invention. Any modifications made to the technical solution based on the technical concept proposed in this invention shall fall within the scope of protection of the claims of this invention.

Claims

1. A method for allocating radio resources in a MU-MIMO system based on a diffusion model, characterized in that, Includes the following steps: With the goal of maximizing the total capacity of the MU-MIMO wireless communication system, an optimization problem is constructed by jointly optimizing resource block allocation and transmission power; in, This indicates the maximum carrier aggregation capability of the user terminal. This indicates the maximum aggregation capacity of resource blocks on the user terminal. For indicator functions, Indicates user Minimum communication rate requirement, This indicates the maximum transmission power of the base station; For user collection, The total number of users per antenna. The total number of resource blocks, For carrier set, In resource blocks Upward users The allocated power value, For users The total reachable rate; A conditional diffusion model is trained based on an expert dataset; multiple candidate solutions are generated using the trained model; and the optimal solution to the optimization problem is obtained by further selecting the best solution from the multiple candidate solutions through policy enhancement.

2. The wireless resource allocation method for MU-MIMO systems based on a diffusion model according to claim 1, characterized in that, A conditional diffusion model is trained based on an expert dataset; multiple candidate solutions are generated using the trained model; and policy enhancement is used to further select the optimal solution from the multiple candidate solutions to obtain the optimal solution to the optimization problem, including: a) Obtaining datasets from optimization theory-based solution methods , For channel conditions, For power allocation matrix; b) Initialize the noise prediction network parameters Set the diffusion step number T and variance scheduling parameters. , The hyperparameter for controlling the diffusion intensity at time step t, where T is the total number of diffusion steps; c) Load from dataset Sampling time steps from a uniform distribution ; d) According to the formula for the forward diffusion process, towards Injection variance is Gaussian noise Obtain the power allocation matrix of injected Gaussian noise. The power allocation matrix to inject Gaussian noise Channel conditions as well as Input noise prediction neural network ; e) Calculate the loss function Update parameters ; f) Repeat steps c)-e) until the loss function is calculated. The signal converges, resulting in the trained noise prediction neural network. g) Sample N samples from a Gaussian distribution and set the current channel state. Input conditional diffusion model; h) According to Perform T reverse generation processes, using the trained noise prediction neural network to generate the corresponding power allocation matrix to obtain N candidate solutions. ; i) Select candidate solutions based on the constraints to obtain a set of feasible solutions. ; j) Calculate the action-state value function of a feasible solution. ; k) Based on Selected power allocation matrix To optimize the optimal solution to the problem.

3. The wireless resource allocation method for MU-MIMO systems based on a diffusion model according to claim 2, characterized in that, loss function The details are as follows; loss function Equivalent to expert strategy distribution Distribution of diffusion strategies The Kullback–Leibler divergence between them.

4. The wireless resource allocation method for MU-MIMO systems based on a diffusion model according to claim 2, characterized in that, The action-state value function is: For the power allocation matrix, For users In resource block The achievable transmission rate is It is the set of all resource blocks for all carriers.

5. The wireless resource allocation method for MU-MIMO systems based on a diffusion model according to claim 2, characterized in that, A noise prediction neural network based on a Transformer architecture is used, with the input being a power allocation matrix injected with Gaussian noise. Channel conditions And the diffusion time step t, specifically: the power allocation matrix for injecting Gaussian noise. With channel conditions The network is concatenated along the channel dimension, enabling it to simultaneously perceive the current power distribution state and the corresponding channel environment information. The concatenated features are initially encoded through convolutional layers, achieving feature dimensionality enhancement and extraction of local coupling relationships. The diffusion time step t is mapped to high-dimensional temporal features via a temporal embedding module and added to the main features through broadcasting. After normalization, these features are used as input to subsequent networks. The encoded features enter the main network, which consists of multiple Transformer structures. The output convolutional layer maps the features back to the original power distribution dimension, yielding the predicted noise. .

6. The radio resource allocation method for MU-MIMO systems based on a diffusion model according to claim 1, characterized in that, Consider a coverage radius of A single-base station MU-MIMO communication system for circular cellular cells, including Single-antenna user and Each equipment Root antenna, A base station with multiple carriers, the bandwidth of each carrier. They are all the same, and each carrier contains Each resource block, the wireless channel model between the base station and the user adopts a block fading model, and the channel state information is determined by path loss. Shadow decay With small-scale fading describe; user In resource block The normalized beamforming vector on is: Then the user In resource block The signal-to-interference-plus-noise ratio is: Wherein, the numerator represents the user. The expected signal power, the denominator includes the interference power. With noise power , In resource blocks Upward users The allocated power value, user Total reachability The sum of its rates across all resource blocks: This represents the bandwidth of each carrier.

7. The radio resource allocation method for MU-MIMO systems based on a diffusion model according to claim 1, characterized in that, The conditional diffusion model is used to learn the distribution of expert policies from the dataset. Among them, channel conditions It determines the power allocation matrix. The prior conditions for the distribution; The conditional diffusion model is trained through forward diffusion and backward generation processes to directly generate... During the forward diffusion process, the power allocation matrix is... Gradually inject Gaussian noise The forward diffusion process forms a Markov chain, directly from... get ; The ideal reverse generation process distribution is obtained based on Bayes' theorem and the forward diffusion process. ; The parameters used are A noise prediction neural network, under given channel conditions Predictive Noise in Mean estimation of the distribution of the reverse generation process: in To predict noise, the distribution of the reverse generation process is as follows: Let covariance matrix be the variance matrix. As a unit array, This is a hyperparameter for controlling the diffusion intensity at time step t.

8. A single-base station MU-MIMO communication system, characterized in that, include One single-antenna user and one equipped root antenna and A base station with multiple carriers, the bandwidth of each carrier. They are all the same, and each carrier contains Resource blocks; resource allocation is performed using the MU-MIMO system wireless resource allocation method based on the diffusion model as described in any one of claims 1-7.

9. A computer device, characterized in that, It includes a processor and a memory, the memory being used to store a computer-executable program, the processor reading part or all of the computer-executable program from the memory and executing it, and when the processor executes part or all of the computer-executable program, it can implement the wireless resource allocation method for a MU-MIMO system based on a diffusion model as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, Used to store computer-readable programs or instructions, which, when executed by a processor, can implement the wireless resource allocation method for a MU-MIMO system based on a diffusion model as described in any one of claims 1-7.

Images