A MIMO-NOMA resource joint optimization method based on graph reinforcement learning

Abstract

This invention discloses a MIMO-NOMA resource joint optimization method based on graph reinforcement learning, comprising the following steps: Step 1: Establishing the basic framework of the MIMO-NOMA resource joint allocation system, defining the environment and constraints for user relationship modeling, grouping, and resource allocation; Step 2: Constructing a hierarchical joint resource allocation network based on graph reinforcement learning; Step 3: Performing offline training and online inference on the graph reinforcement learning network; Step 4: Completing system performance verification based on the resource allocation scheme. The beneficial effects of this invention are: overcoming the limitations of traditional resource allocation methods in expressing complex user relationships, improving the accuracy of resource allocation decisions, and thus improving system efficiency, rate, and spectral efficiency. It effectively enhances the algorithm's adaptability to complex channel environments and dynamic user relationships, enabling the system to maintain high performance in scenarios with severe multi-user interference and variable channel conditions, and significantly improving the convergence speed and stability of resource allocation.

Description

Technical Field

[0001] This invention belongs to the field of wireless communication technology, specifically relating to a MIMO-NOMA resource joint optimization method based on graph reinforcement learning. Background Technology

[0002] With the widespread deployment of 5G mobile communication systems and the rapid development of 6G mobile communication technology, future wireless communication networks are evolving towards high spectral efficiency, massive connectivity, ultra-low latency, and high reliability. Traditional orthogonal multiple access (OMA) technology, requiring different users to occupy orthogonal time-frequency resources, effectively avoids inter-user interference, but its spectral efficiency and system access capacity are increasingly insufficient to meet future communication demands in high-density, multi-service scenarios. Non-orthogonal multiple access (NOMA), by multiplexing multiple users in the power domain and using serial interference cancellation (SIC) technology to achieve signal separation, can significantly improve spectral resource utilization and system access capacity, and is therefore considered one of the important candidate key technologies for future wireless communication. Meanwhile, multiple-input multiple-output (MIMO) technology, by configuring multiple antennas at the transmitter and receiver, achieves spatial multiplexing and diversity gain, effectively improving system capacity and link reliability. Combining MIMO and NOMA to form a MIMO-NOMA system can simultaneously leverage the dual advantages of spatial and power domain multiplexing, showing significant potential in improving network throughput, increasing spectral efficiency, and supporting massive connectivity.

[0003] However, improving the performance of MIMO-NOMA systems depends not only on the physical layer transmission mechanism itself, but also on a rational and efficient radio resource management strategy. In practical systems, user grouping determines which users share the same resource block, subcarrier allocation determines the way spectrum resources are used, and power allocation directly affects the interference level between users, the feasibility of SiC decoding, and the system throughput performance. There is a significant coupling relationship among these three: user grouping affects subsequent subcarrier and power configuration space, subcarrier allocation affects the interference pattern between users, and power allocation, in turn, determines the effectiveness of grouping and spectrum reuse. Therefore, how to achieve joint optimization of user grouping, subcarrier allocation, and power allocation in MIMO-NOMA scenarios has become an important research problem in this field.

[0004] In terms of traditional optimization methods, existing research has largely employed convex approximations and heuristic search to solve single or partially joint problems such as user grouping, sub-channel allocation, or power allocation. These methods can usually obtain relatively optimal solutions under specific assumptions and have a clear theoretical basis. However, due to the significant combinatorial optimization characteristics of the MIMO-NOMA resource allocation problem, traditional optimization algorithms often face problems such as high computational complexity, long solution time, and insufficient real-time performance when the number of users, subcarriers, and constraints increase, making them difficult to adapt to dynamic wireless environments.

[0005] In recent years, with the successful application of deep reinforcement learning (DRL) in sequential decision-making and high-dimensional control problems, more and more research has begun to explore its application in the field of wireless resource management. DRL can autonomously learn resource allocation strategies through interaction with the environment, avoiding precise problem modeling, and thus exhibits good adaptability in high-dimensional, dynamic, and complex coupled scenarios. Existing literature shows that DQN and its improved models are suitable for discrete resource allocation problems, while actor-critic structures such as DDPG are more suitable for continuous power control problems. Especially in MIMO-NOMA resource scheduling scenarios, existing research has integrated user grouping, sub-channel allocation, and power allocation into a unified deep reinforcement learning framework and verified the effectiveness of hierarchical decision structures in improving system throughput and spectral efficiency.

[0006] On the other hand, with the development of Graph Neural Networks (GNNs) and Graph Reinforcement Learning (GRL), more and more scholars are beginning to focus on the relationship modeling problem in wireless networks. Unlike traditional vector or matrix state representations, graph structures can more naturally characterize the topological relationships, interference coupling, and dependency structures between users, links, and resources. Existing research has shown that in the wireless resource allocation problem, many states that were originally constructed in matrix form can be transformed into graph state representations, and then feature extraction can be performed using GNNs. In particular, graph reinforcement learning methods can embed graph encoders into actor-critic structures, better handling node relationships, arrangement changes, and topological dependencies, and have good generalization ability and scalability in complex network scenarios.

[0007] Despite significant progress in MIMO-NOMA joint resource allocation, existing research still faces several challenges and limitations, specifically in the following aspects: (1) Shallow clustering is difficult to capture complex relationships: Most existing methods use clustering algorithms such as K-means or enhanced K-means. Although enhanced K-means has improved upon the traditional random initialization clustering method and considers channel gain and channel correlation to some extent, it still relies on manually designed shallow features and distance metrics, making it difficult to fully explore the more complex higher-order correlations, interference coupling relationships, and SIC compatibility relationships among users. For systems like MIMO-NOMA where user relationships are highly structured, relying solely on shallow clustering often fails to yield globally better grouping results.

[0008] (2) Insufficient state representation capability: In reinforcement learning modeling, many existing methods design environmental states as simple vectors or matrices. Although this facilitates network input, such representations often fail to reflect the relationships between users and the system structure. For example, in MIMO-NOMA systems, users not only have differences in channel gain, but also channel correlation, interference caused by shared resources, and decoding dependencies related to the SIC order. This information is essentially relational and graph-structured. If ordinary vector concatenation is still used to input the network, it may lead to insufficient state representation capability, thereby limiting the learning effect of subsequent resource allocation strategies.

[0009] (3) Simple front-end coding method: Although existing studies have verified that dueling DQN is suitable for handling discrete subcarrier / subchannel allocation and DDPG is suitable for handling continuous power control, most front-end state coding methods are still relatively simple and have not fully utilized the ability of graph structure learning to model the relationship features of wireless networks.

[0010] A prior art method for multi-dimensional resource allocation in NOMA multi-beam satellite communication systems based on reinforcement learning (publication number: CN117833997A) is disclosed, which uses reinforcement learning to achieve joint optimization of user grouping, channel allocation, and power allocation. However, this technical solution does not use graph reinforcement learning to construct a user relationship graph, making it impossible to perform structured modeling of interference and correlation between users, and thus difficult to achieve efficient and accurate MIMO-NOMA joint resource allocation. Summary of the Invention

[0011] This invention proposes a resource allocation method based on graph reinforcement learning, aiming to maximize spectral efficiency in MIMO-NOMA systems. It overcomes the problems of strong variable coupling and difficulty in modeling user relationships in traditional resource allocation, and proposes a hierarchical resource allocation scheme based on graph neural networks and reinforcement learning. The system constructed by this scheme can significantly enhance the model's ability to express channel characteristics, and ultimately improve spectral efficiency.

[0012] To achieve the above objectives, this application proposes a MIMO-NOMA resource joint optimization method based on graph reinforcement learning, including: Step 1: Establish the basic framework of the MIMO-NOMA resource joint allocation system, define the environment and constraints for user relationship modeling, grouping and resource allocation, and complete the user relationship and system scenario modeling, transmitter hierarchical resource allocation architecture design and joint optimization problem modeling; Step 2: Construct a hierarchical joint resource allocation network based on graph reinforcement learning to achieve graph attention encoding, hierarchical decision-making, and constraint output; Step 3: Perform offline training and online inference on the graph reinforcement learning network to obtain the resource allocation strategy; Step 4: Perform system performance verification based on the resource allocation scheme and output the optimized resource allocation results.

[0013] Preferably, step 1 further includes user relationship and system scenario modeling, modeling the user set as a user relationship graph, the mathematical representation of which is: in, This is the node feature matrix (U is the number of users, F is the feature dimension). It is an adjacency matrix used to represent the connection relationships between users.

[0014] Preferably, step 1 further includes a hierarchical resource allocation architecture design for the transmitter, and the signal vector of the system receiver is represented as: in, Let M be the channel matrix, W be the number of base station antennas, P be the resource allocation matrix, P be the power allocation diagonal matrix, and n be additive white Gaussian noise.

[0015] Preferably, step 1 further includes modeling a joint optimization problem, wherein the joint resource allocation optimization problem is expressed as: in, For system and rate, For user grouping schemes, Assign an indicator matrix to the subcarrier. This is the power allocation matrix.

[0016] Preferably, the hierarchical joint resource allocation network in step 2 includes a graph attention encoding layer and a hierarchical decision layer; the hierarchical decision layer is divided into two parallel sub-modules: The GAT grouping submodule is used to group users based on user embedding features. The dual-branch decision submodule is used to perform resource allocation decisions based on user grouping results and group-level graph features; The attention feature extraction of the graph attention network satisfies: in, Let W be the attention coefficient between users i and j, and W be the feature transformation matrix. For the original characteristics of user i, User characteristics weighted by attention.

[0017] Preferably, the dual-branch decision submodule includes a constraint output layer, and the constraint mapping satisfies: in, For the actor network of Graph-DDPG, For the value network of dueling DQN, To assign a threshold, This is an indicator function.

[0018] Preferably, the offline training in step 3 uses the Huber loss function, expressed as: in, This represents the true optimal resource allocation scheme. For network output, This is a hyperparameter (usually set to 1.0).

[0019] Preferably, in step 3, the online inference inputs the measured channel matrix and user relationship features into the trained graph reinforcement learning network, and outputs a joint scheme of user grouping, subcarrier allocation and power allocation.

[0020] Preferably, the dual-branch decision submodule performs discrete subcarrier allocation and continuous power allocation in parallel based on the output of the GAT grouping submodule.

[0021] Preferably, steps 1 to 4 are executed sequentially to achieve joint optimization of user grouping, subcarrier allocation, and power allocation in the MIMO-NOMA system through user relationship graph modeling, hierarchical decision-making parallel processing, offline training, and online inference.

[0022] (1) A user relationship graph model for MIMO-NOMA was constructed: Considering the complex channel correlation, interference relationship and SIC compatibility among users, this paper models the user set as a graph structure, carefully designs node features and edge features, and extends the traditional matrix state to a more expressive graph state representation, providing a unified and reliable theoretical basis for the effective development of subsequent graph learning and graph reinforcement learning, and realizing the accurate characterization of complex relationships among users.

[0023] (2) A user grouping method based on graph neural network (GAT) is proposed: In view of the problem that the traditional enhanced K-means grouping method has limited ability to express complex user relationships and insufficient grouping accuracy, this paper introduces graph attention network to deeply encode the constructed user relationship graph. Through the attention mechanism, important relationships between users are adaptively captured, and high-order user embedding representations are learned. The limitations of traditional grouping methods are abandoned, and user grouping is completed based on the learned graph embedding features, thereby significantly improving the quality of user grouping and creating better preconditions for the efficient allocation of resources such as subcarriers and power in the future.

[0024] (3) A dueling DQN subcarrier allocation method based on graph feature enhancement was designed: Based on the user grouping results, this paper integrates the group-level graph features extracted by GAT with the current subcarrier occupancy status, channel quality and interference information of each user, and uses them together as input features of dueling DQN. This makes up for the defect of traditional subcarrier allocation methods that ignore user structure correlation, effectively improves the effectiveness, stability and adaptability of discrete subcarrier allocation decision, and realizes reasonable scheduling of subcarrier resources.

[0025] (4) A continuous power allocation method based on Graph-DDPG was designed: In view of the continuous action characteristics of power allocation and the problem that traditional power allocation methods are difficult to perceive the structural relationship between users and have limited optimization effect, this paper embeds a graph encoder in the actor-critic framework of deep deterministic policy gradient, so that the power control policy can accurately perceive the structural relationship and coupling information between users, make full use of user association features to optimize the power allocation policy, thereby improving the accuracy and efficiency of continuous power optimization and realizing the efficient utilization of system power resources.

[0026] The final optimized resource allocation scheme is obtained based on the deep learning algorithm trained above.

[0027] Compared with existing technologies, the beneficial effects of this invention, employing the above technical solutions, are as follows: This invention combines graph reinforcement learning, graph attention networks, deep reinforcement learning, and hierarchical joint optimization to construct a MIMO-NOMA joint resource allocation scheme. This scheme overcomes the limitations of traditional resource allocation methods in representing complex user relationships, improves the accuracy of resource allocation decisions, and thus enhances system performance, rate, and spectral efficiency. Furthermore, this invention integrates graph feature encoding and hierarchical optimization mechanisms, effectively improving the algorithm's adaptability to complex channel environments and dynamic user relationships. This enables the system to maintain high performance in scenarios with severe multi-user interference and variable channel conditions, significantly improving the convergence speed and stability of resource allocation. Attached Figure Description

[0028] Figure 1 This is a schematic diagram of the architecture of a MIMO-NOMA hybrid precoding communication system. Detailed Implementation

[0029] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0030] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The following description of at least one exemplary embodiment is merely illustrative and is in no way intended to limit the present invention or its application or use. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0031] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0032] Example 1: A MIMO-NOMA resource joint optimization method based on graph reinforcement learning, comprising the following steps: Step 1: Establish the basic framework of the MIMO-NOMA resource joint allocation system, define the environment and constraints for user relationship modeling, grouping and resource allocation, and complete the user relationship and system scenario modeling, transmitter hierarchical resource allocation architecture design and joint optimization problem modeling; Step 2: Construct a hierarchical joint resource allocation network based on graph reinforcement learning to achieve graph attention encoding, hierarchical decision-making, and constraint output; Step 3: Perform offline training and online inference on the graph reinforcement learning network to obtain the resource allocation strategy; Step 4: Perform system performance verification based on the resource allocation scheme and output the optimized resource allocation results.

[0033] Example 2: This example is based on the previous example, and the specific implementation further elaborates on the present invention. A MIMO-NOMA resource joint optimization method based on graph reinforcement learning includes the following steps: Step 1: Establish the basic framework of the MIMO-NOMA joint resource allocation system, defining the environment and constraints for user relationship modeling, grouping, and resource allocation. Considering channel correlation, interference relationships, and SIC compatibility among users, as well as the characteristics of GAT, dueling DQN, and Graph-DDPG, ensure that each module can work collaboratively and seamlessly. This system framework provides the foundation for subsequent user relationship graph model construction and the design of a hierarchical joint resource allocation method based on graph reinforcement learning. (1) User Relationship and System Scenario Modeling To address the characteristics of densely distributed users, strong channel correlation, and inter-group interference coupling in MIMO-NOMA systems, this paper employs a graph structure to characterize the relationships between users. Considering scenarios where the base station is equipped with a large-scale antenna array and the user end has a single antenna divided into multiple groups, a user relationship graph for MIMO-NOMA is constructed. , where the node set For all users in the system, edge set This model characterizes the channel gain strength and interference coupling degree among users. After modeling the user set as a graph structure, node and edge features are designed: node features encompass the user's effective channel gain, service rate requirements, and interference level; edge features quantify the channel correlation and SIC compatibility among users, extending the traditional matrix state representation to a more expressive graph state representation. The channel state information and user association features generated based on this graph model provide a unified input foundation for the feature encoding of the subsequent graph attention network. The specific mathematical representation of the user relationship graph is as follows: in, This is the node feature matrix (U is the number of users, F is the feature dimension). It is an adjacency matrix used to represent the connection relationships between users.

[0034] (2) Layered resource allocation architecture for the transmitter To address the resource allocation requirements of MIMO-NOMA systems, a hierarchical joint optimization architecture is designed. This architecture consists of three core modules from top to bottom: user grouping, subcarrier allocation, and power allocation. These modules correspond to three algorithm models: GAT, dueling DQN, and Graph-DDPG, respectively, ensuring that each module is functionally independent yet logically connected.

[0035] User grouping module (GAT driven): Based on the constructed user relationship graph, GAT is introduced to encode the features of nodes and edges, learn the high-order embedding representation of users, and adaptively select key related users through the attention mechanism to complete user grouping, thus solving the problem that the traditional K-means grouping method has limited ability to express complex user relationships.

[0036] Subcarrier allocation module (dueling DQN driver): Based on the user grouping results, the group-level graph features extracted by GAT are fused with subcarrier occupancy status, channel quality, and interference information, and used as input features for dueling DQN to realize dynamic allocation decisions for discrete subcarrier resources, thereby improving the effectiveness and stability of allocation.

[0037] Power Allocation Module (Graph-DDPG Driver): Targeting the continuous operation characteristics of power allocation, a graph encoder is embedded in the actor-critic framework of DDPG, enabling the power control strategy to perceive the structural relationships and coupling information between users, thereby achieving precise optimization of continuous power resources.

[0038] After the input signal at the transmitting end is processed by the hierarchical module, user grouping, subcarrier allocation, and power allocation are completed sequentially, ultimately generating a resource allocation scheme that satisfies system constraints. The signal vector at the system receiving end can be represented as: in, Let M be the channel matrix, W be the number of base station antennas, P be the resource allocation matrix, P be the power allocation diagonal matrix, and n be additive white Gaussian noise.

[0039] (3) Modeling of joint optimization problem The core objective of the MIMO-NOMA hierarchical resource allocation system studied in this paper is to maximize the system's sum rate and spectral efficiency under user packet constraints, subcarrier allocation constraints, power allocation constraints, and SiC decoding constraints. Considering the interference relationships between users and channel correlations, this joint resource allocation optimization problem can be formulated as follows: in, For system and rate, For user grouping schemes, Assign an indicator matrix to the subcarrier. Let be the power allocation matrix. Since this optimization problem involves the coupling of discrete subcarrier allocation and continuous power allocation, and includes non-convex interference terms caused by user association, direct solution is extremely difficult. Therefore, this invention proposes a hierarchical joint resource allocation method based on a graph reinforcement learning framework for approximate solution.

[0040] Step 2: Construct a hierarchical joint resource allocation network based on graph reinforcement learning This scheme constructs a graph reinforcement learning-enhanced hierarchical network, using user relationship graphs and channel state information as network inputs to adaptively generate user grouping schemes, subcarrier allocation strategies, and continuous power allocation vectors, thereby improving the overall sum rate and spectral efficiency of the MIMO-NOMA system. The network consists of a graph attention coding layer, a hierarchical decision layer (including a duelingDQN subcarrier allocation header and a Graph-DDPG power allocation header), and a constraint output layer. First, the graph attention coding layer separates the real and imaginary parts of the complex-valued channel matrix, and combines user node / edge features as input to the graph attention network to extract high-order correlation features between users and group-level graph features, providing a unified feature representation for subsequent hierarchical resource allocation. The hierarchical decision layer is divided into two parallel sub-modules: The GAT grouping submodule: The user embedding features output from the graph attention coding layer are input into the GAT grouper. Users are aggregated and associated through attention weights to generate the optimal user grouping scheme, laying the foundation for subcarrier and power allocation.

[0041] The dual-branch decision submodule, based on user grouping results and group-level graph features, outputs discrete subcarrier allocation strategies from the dueling DQN subcarrier allocation header and continuous power allocation vectors from the Graph-DDPG power allocation header, achieving collaborative optimization of hierarchical decision-making. The dilated convolutional feature extraction logic, adapted to local feature mining of the channel matrix, is replaced in this scheme with global correlation modeling using a graph attention mechanism. The specific attention feature extraction formula is as follows: in, Let W be the attention coefficient between users i and j, and W be the feature transformation matrix. For the original characteristics of user i, Features weighted by attention.

[0042] After processing by the hierarchical decision layer, the constraint output layer introduces SIC decoding constraints and total resource constraints to ensure that the output user packet, subcarrier allocation, and power allocation schemes meet the actual operating conditions of the system. For continuous constraints in power allocation, a power vector satisfying the total power limit is output through the actor network of Graph-DDPG; for binary constraints in subcarrier allocation, legal allocation is achieved through the softmax output of dueling DQN and threshold truncation. The specific constraint mapping formula is as follows: in, For the actor network of Graph-DDPG, For the value network of dueling DQN, To assign a threshold, This is an indicator function.

[0043] Step 3: Offline Training and Online Inference of Graph Reinforcement Learning Networks (1) Offline training phase Network parameters are optimized in batches through offline training, constructing a sample dataset containing user relationship graphs, channel states, and optimal resource allocation schemes. For each training batch, a joint loss function is designed with the goal of maximizing system performance and rate. This loss function includes cross-entropy loss for subcarrier allocation and mean squared error loss for power allocation, while also introducing an attention regularization term to enhance the model's ability to perceive user relationships. This invention selects Huber loss as the network's total loss function, combining the advantages of cross-entropy loss and mean squared error loss to avoid gradient explosion during training and improve model convergence stability. The Huber loss function expression is: in, This represents the true optimal resource allocation scheme. For network output, This is a hyperparameter (usually set to 1.0). The total loss function is minimized using the backpropagation algorithm, and the network parameters of GAT, Dueling DQN, and Graph-DDPG are updated until the model converges.

[0044] (2) Online reasoning stage After offline training, the trained graph reinforcement learning network model is saved. During online testing, the measured channel matrix and user relationship features are input into the model, and the network adaptively outputs a user grouping scheme, subcarrier allocation strategy, and continuous power allocation vector. Subsequently, based on the output resource allocation scheme and combined with the SIC decoding process of the MIMO-NOMA system, the system's sum rate and spectral efficiency are calculated. The performance advantages of the proposed method are verified by comparing it with traditional resource allocation methods.

[0045] Step 4: System Performance Verification Through the above steps, a low-complexity, high-spectral-efficiency MIMO-NOMA graph reinforcement learning hierarchical joint resource allocation system was constructed. In complex wireless communication scenarios, this system can accurately perceive channel correlations and interference relationships among users, effectively solving the problems of low grouping accuracy and poor decision robustness in traditional resource allocation methods. It achieves high real-time dynamic resource scheduling while strongly guaranteeing the system's sum rate, spectral efficiency, and user fairness.

Claims

1. A MIMO-NOMA resource joint optimization method based on graph reinforcement learning, characterized in that, Includes the following steps: Step 1: Establish the basic framework of the MIMO-NOMA resource joint allocation system, and define the environment and constraints for user relationship modeling, grouping, and resource allocation; Step 2: Construct a hierarchical joint resource allocation network based on graph reinforcement learning; Step 3: Perform offline training and online inference on the graph reinforcement learning network; Step 4: Complete system performance verification based on the resource allocation scheme.

2. The method of claim 1, wherein, Step 1 includes user relationship and system scenario modeling, modeling the user set as a user relationship graph, the mathematical representation of which is: wherein, is a node feature matrix (U is the number of users, and F is the feature dimension), is an adjacency matrix, which is used to represent the connection relationship between users.

3. The method of claim 1, wherein, Step 1 includes the design of a hierarchical resource allocation architecture for the transmitter, and the signal vector representation of the system receiver is as follows: in, Let M be the channel matrix, W be the number of base station antennas, P be the resource allocation matrix, P be the power allocation diagonal matrix, and n be additive white Gaussian noise.

4. The method according to claim 1, characterized in that, Step 1 includes modeling a joint optimization problem, which is expressed as follows: wherein, is a system and rate, is a user grouping scheme, is a subcarrier allocation indication matrix, is a power allocation matrix.

5. The method of claim 1, wherein, The hierarchical joint resource allocation network in step 2 includes a graph attention encoding layer and a hierarchical decision layer; the hierarchical decision layer is divided into two parallel sub-modules: The GAT grouping submodule is used to group users based on user embedding features. The dual-branch decision submodule is used to perform resource allocation decisions based on user grouping results and group-level graph features; The attention feature extraction of the graph attention network satisfies: wherein, is the attention coefficient between users i and j, W is the feature transformation matrix, is the original feature of user i, is the attention weighted user feature.

6. The method of claim 5, wherein, The dual-branch decision submodule includes a constraint output layer, and the constraint mapping satisfies: in, For the actor network of Graph-DDPG, For the value network of dueling DQN, To assign a threshold, This is an indicator function.

7. The method of claim 6, wherein, In step 3, offline training uses the Huber loss function, expressed as follows: wherein, is the true optimal resource allocation scheme, is the network output, is the hyperparameter.

8. The method of claim 1, wherein, In step 3, the online inference inputs the measured channel matrix and user relationship features into the trained graph reinforcement learning network, and outputs a joint scheme of user grouping, subcarrier allocation and power allocation.

9. The method of claim 5, wherein, The dual-branch decision submodule performs discrete subcarrier allocation and continuous power allocation in parallel based on the output of the GAT grouping submodule.

10. The method according to claim 1, characterized in that, Steps 1 to 4 are executed sequentially to achieve joint optimization of user grouping, subcarrier allocation, and power allocation in the MIMO-NOMA system through user relationship graph modeling, hierarchical decision-making parallel processing, offline training, and online inference.

Images