A method for joint optimization of MU-MIMO frequency selection scheduling and link adaptation

Abstract

The application discloses a kind of MU-MIMO frequency selection scheduling and link adaptive joint optimization method, belong to wireless communication technical field, to solve the problems of existing technology scheduling granularity, action space complexity, optimization target and actual transmission effect disjunction;The application models each resource block RB as independent agent, adopts parameter sharing double-head PPO policy network, respectively completes multiplexing user number determination and user set selection, realizes concurrent scheduling in combination with dynamic mask;At the same time, introduce physical layer link adaptive processing, construct reward closed loop in combination with terminal ACK / NACK feedback, complete network parameter iterative optimization;The application realizes RB level fine frequency selection scheduling, effectively improves system spectral efficiency and actual effective throughput, with excellent environmental adaptability and engineering scalability.

Description

Technical Field

[0001] This invention belongs to the field of wireless communication technology, specifically relating to a joint optimization method of MU-MIMO frequency selection scheduling and link adaptive optimization. Background Technology

[0002] With the evolution of 5G and future 6G networks, MU-MIMO technology plays a core role in improving spectrum efficiency and system capacity. In multi-user scheduling problems, how to efficiently and accurately allocate concurrent user sets to each RB and maximize the effective system throughput under complex spatial interference is a recognized technical challenge in the industry.

[0003] Currently, existing algorithms for MU-MIMO resource scheduling can be mainly divided into: heuristic optimization-based methods and deep reinforcement learning (DRL)-based methods.

[0004] Traditional heuristic optimization methods, due to their relatively low computational complexity and ease of engineering implementation, have long been widely used in the field of MU-MIMO resource scheduling. They quickly obtain feasible solutions for resource allocation through pre-set empirical rules or greedy criteria. A typical example is the patent application CN116318289A, entitled "A Three-Dimensional Resource Allocation Method for MU-MIMO Systems," which proposes a heuristic resource allocation strategy based on a greedy algorithm. This strategy combines users' personalized traffic demands to perform time-frequency resource scheduling to alleviate resource waste. Furthermore, it establishes beam selection criteria based on a comprehensive trade-off between beam matching performance and user selection repetition rate, thereby achieving joint resource allocation across the spatial, temporal, and frequency dimensions.

[0005] However, such heuristic methods based on greedy algorithms rely heavily on human-defined empirical rules. In scenarios where the space occupancy and interference relationships of concurrent users in MU-MIMO systems are complex, they are prone to getting stuck in local optima. They perform poorly when balancing the number of users to be reused with the cancellation of spatial interference, resulting in limited spatial reuse capabilities and low overall spectral efficiency.

[0006] Due to its adaptive learning capabilities in dynamic and complex environments, DRL (Digital Reinforcement Learning) is gradually being introduced into the field of MU-MIMO resource scheduling, aiming to replace traditional iterative search and heuristic rules. A typical example is patent application CN115988667A, entitled "A Multi-User MIMO Resource Scheduling Method Based on Multi-Agent Reinforcement Learning." This patent treats all schedulable "users" as a multi-agent group, using user information and inter-user interference as environmental inputs. It employs a centralized training and decentralized execution strategy to output the optimal sub-user set to maximize the system's total throughput. Another example is patent application CN113261016A, entitled "Single-shot Multi-User Multiple-Input Multiple-Output (MU-MIMO) Resource Pairing Using Deep Q-Network (DQN) Based on Reinforcement Learning." This patent encodes the sequence of MU-MIMO beam combinations into unique values ​​and uses a deep Q-network to directly output one or more optimal beams to assign to the UE, thus avoiding the latency problem of traditional combinatorial search schemes.

[0007] However, the above-mentioned methods based on traditional DRL still have the following significant drawbacks:

[0008] First, the scheduling granularity of existing technologies is usually limited to the full bandwidth level, failing to achieve fine-grained resource allocation at the RB level. This makes it impossible to effectively utilize frequency-selective fading characteristics to achieve differentiated concurrent user configurations, thus limiting further improvements in the overall spectral efficiency of the system. Second, agent modeling lacks scalability. Existing solutions often treat "users" as agents, and when the number of candidate users in the system dynamically increases or decreases, the user-based network structure is prone to failure, lacking scalability for engineering deployment. Furthermore, in terms of action space design, existing solutions either map user combination schemes to independent action bits, causing the action space to explode exponentially with the number of candidate users, increasing the difficulty of network convergence, or they forcibly fix the number of concurrent users, resulting in insufficient scheduling flexibility and an inability to adaptively select the optimal user multiplexing combination.

[0009] Meanwhile, existing technologies, when evaluating the merits of scheduling decisions, often rely on idealized Shannon throughput formulas or simple channel capacity upper bounds as greedy criteria or environmental rewards. They fail to implement strict link adaptive processing at the physical layer and lack a complete physical layer closed-loop feedback mechanism. This leads to a deviation between the theoretical optimization goals of reinforcement learning and the achievable performance of the actual system, resulting in high theoretical indicators but poor practical implementation. Summary of the Invention

[0010] To address the problems mentioned in the background, this invention provides a joint optimization method for MU-MIMO frequency-selective scheduling and link adaptation. This method solves the problems of coarse scheduling granularity, poor scalability of agent modeling, high complexity of action space leading to difficulty in convergence, and disconnect between optimization objectives and actual physical layer transmission effects in existing technologies. It achieves RB-level fine-grained frequency-selective scheduling, improves system spectral efficiency and actual effective throughput, and enhances the engineering scalability and environmental adaptability of the solution.

[0011] To achieve the above objectives, the present invention provides the following technical solution:

[0012] A joint optimization method for MU-MIMO frequency-selective scheduling and link adaptation includes the following steps:

[0013] S1: Architecture setup; Set up a multi-user multiple-input multiple-output (MU-MIMO) system, which includes a base station gNB and K candidate user devices (UEs). The base station is equipped with M transmit antennas, and the system frequency domain resources are divided into N resource blocks RB. The base station and each UE are connected by a frequency-selective fading channel. The base station performs resource scheduling and downlink data transmission based on the channel state information (CSI) reported by the UEs.

[0014] S2: State collection and feature extraction; Based on the CSI reported by the UE within the current transmission time interval (TTI), the base station extracts the channel strength features and spatial beam features of each candidate user on each RB, splices them together to construct the state matrix corresponding to each RB, and uses the state matrix as the input feature of the corresponding RB agent.

[0015] S3: Multi-user scheduling decision optimization of PPO based on dual-head proximal policy; treat each RB as an independent agent and use a parameter-sharing dual-head PPO policy network to execute scheduling decisions;

[0016] Taking the state matrix of a single RB as input, the number of multiplexed users of the RB is first determined through the first output head of the policy network. Then, through the second output head of the policy network combined with the dynamic masking mechanism, users matching the number of multiplexed users are selected from the candidate user set to generate the scheduling user set of the RB. Finally, the scheduling user sets of all RBs are combined to obtain the global scheduling matrix.

[0017] S4: Physical layer precoding and link adaptive processing; Based on the set of scheduled users of each RB, the base station constructs the joint channel matrix of the corresponding RB, performs zero-forcing ZF precoding and power normalization processing to obtain the final precoding matrix, calculates the signal-to-interference-plus-noise ratio (SINR) of each scheduled user on the corresponding RB; maps the SINR of each user on all the RBs allocated to it to RB-level link features, predicts the transmission success probability of candidate MCS through a pre-trained link adaptive neural network based on attention aggregation, and then determines the optimal modulation and coding strategy (MCS) of each user based on the expected throughput maximization criterion, generates and issues scheduling signaling based on the global scheduling matrix and the optimal MCS of each user, and performs downlink data transmission in combination with the final precoding matrix;

[0018] S5: Reward Allocation and Model Update; The base station receives the ACK or NACK information from the UE and calculates the actual throughput of each user; Based on the effective capacity contribution weight of each RB to its scheduled users, the corresponding actual throughput is divided and accumulated to obtain the local reward of each RB agent; The state, scheduling action, and local reward of each RB are constructed into an experience tuple and stored in the trajectory buffer. After sampling, the overall objective function is constructed by combining the pruning policy loss, value network loss, and policy joint entropy regularization term of the PPO algorithm according to the preset weights, and the parameters of the policy network and value network are iteratively updated.

[0019] Compared with the prior art, the beneficial effects of the present invention are:

[0020] 1. This invention models each RB as an independent intelligent agent, replacing the existing modeling method that treats users as intelligent agents. This achieves fine-grained resource scheduling at the RB level, enabling the system to combine the channel quality differences and frequency selective fading characteristics between different RBs to execute differentiated concurrent user configurations and fully exploit frequency selective gains. At the same time, when the number of candidate users in the system dynamically increases or decreases, the model structure can work normally without adjustment, greatly improving the engineering scalability and environmental adaptability of the solution.

[0021] 2. This invention employs a dual-head PPO strategy network, structurally decoupling the two strongly coupled scheduling actions of "determining the number of reused users" and "selecting the specific user set." Combined with a dynamic masking mechanism, it achieves concurrent scheduling of multiple users with variable numbers. This avoids the exponential explosion problem of action space caused by full mapping coding of user combinations, reduces the convergence difficulty of network training, and eliminates the need to forcibly fix the number of concurrent users. This allows the system to adaptively select the optimal user multiplexing combination based on real-time channel conditions and spatial interference, fully leveraging the spatial multiplexing potential of MU-MIMO and improving the system's spectral efficiency.

[0022] 3. This invention introduces a reward mechanism that integrates link adaptation and physical layer closed-loop feedback. After completing the scheduling decision, the user SINR is calculated based on precoding. The transmission success probability of each candidate MCS is predicted through a pre-trained link adaptation network. The expected throughput is calculated in combination with the transport block size to select the optimal MCS. Finally, the reward signal is constructed by combining the ACK / NACK feedback of the terminal. This makes the optimization target of the scheduling decision directly point to the actual effective throughput of the system. It solves the problem of the disconnect between theory and actual performance caused by the optimization of the idealized capacity formula in the prior art, and greatly improves the engineering practical value of the scheduling strategy.

[0023] 4. In this invention, all RB agents adopt a dual-head PPO strategy network with shared parameters, which greatly reduces the computational complexity of network training and ensures the consistency of scheduling decisions of each RB, making it easy to deploy and implement in engineering. It can still maintain stable scheduling performance in dynamic wireless environments with time-varying channel states and complex spatial interference relationships. Attached Figure Description

[0024] Figure 1 This is a flowchart of the RB-level multi-user scheduling and physical layer closed-loop control based on dual-head PPO of the present invention;

[0025] Figure 2 This is a block diagram illustrating the principle of PPO strategy network parameter iteration and model update in this invention.

[0026] Figure 3 This is a schematic diagram of the dual-head PPO strategy network structure of the present invention. Detailed Implementation

[0027] To facilitate understanding of the technical content of this invention by those skilled in the art, the invention will be further described in detail below with reference to the accompanying drawings and specific examples. It should be understood that the specific examples described herein are merely illustrative and not intended to limit the scope of the invention.

[0028] This invention addresses the frequency-selective scheduling scenario of MU-MIMO and proposes a multi-user joint scheduling method with RBs as the decision-making agents. Each RB is modeled as a parameter-sharing agent, and a dual-head policy network is used to determine the number of reused users and the specific user set. Combined with a dynamic masking mechanism, it realizes concurrent scheduling of multi-users with variable numbers of users. At the same time, a reward mechanism that integrates link adaptation and physical layer closed-loop feedback is adopted to enable scheduling decisions to be optimized and updated based on real link feedback.

[0029] The multi-user multiple-input multiple-output (MU-MIMO) system involved in this application includes a base station gNB and K candidate user devices (UEs). The base station is configured with M transmit antennas, and the system frequency domain resources are divided into N resource blocks (RBs). The base station and each UE communicate via a frequency-selective fading channel. The base station performs resource scheduling and downlink data transmission based on the channel state information (CSI) reported by the UEs.

[0030] A joint optimization method for MU-MIMO frequency-selective scheduling and link adaptive scheduling, such as Figure 1 As shown, it includes the following stages:

[0031] 1. State Collection and Feature Extraction Stage: Based on the channel states reported by users, the base station extracts the channel strength features of each candidate user on each RB. Simultaneously, to intuitively characterize spatial interference between multiple users, this stage introduces a pre-defined fixed orthogonal basis matrix (e.g., a fixed beam codebook) to project the user's downlink channel from the antenna domain to the beam domain, thereby extracting the beam energy distribution reflecting its spatial characteristics. Subsequently, the features of each candidate user are concatenated to construct the state matrix of the current resource block. , as input features for multi-agent systems.

[0032] 2. Multi-user scheduling decision-making stage based on dual-head PPO: such as Figure 3 As shown, each RB is treated as an independent agent, and a policy network with shared parameters is used for decision-making. The RB is... The state matrix and local state matrix of each resource block After being input into the shared feature encoder, it outputs in two paths: the first path is the multiplexed user count output header, used to determine the number of users. Number of users reusing a resource block The second path is the user selection output header, used to select from a given number of multiplexed users. Under these conditions, combined with a dynamic masking mechanism, users are selected sequentially from the candidate user set, ultimately forming the first... Set of scheduling users for each resource block .

[0033] 3. Physical Layer Precoding and Link Adaptation Stage: Extract the downlink channel matrix of the scheduled user set, calculate zero-forcing (ZF) precoding, and then obtain the signal-to-interference-plus-noise ratio (SINR) of each selected user on the corresponding RB. Map the SINR of each selected UE on the resource block allocated to it to RB-level link features, and input them into a pre-trained attention-based link adaptive neural network to predict the transmission success probability of candidate MCSs. Then, based on the expected throughput maximization criterion, obtain the optimal MCS. The base station issues scheduling instructions based on the allocation results and the optimal MCS, and executes downlink data transmission.

[0034] 4. Reward Allocation and Model Update Phase: (e.g.) Figure 1 and Figure 2 As shown, the base station receives acknowledgment (ACK) or negative acknowledgment (NACK) information from the terminal and calculates the actual throughput of each user. Then, it calculates the local contribution weight of each RB to the effective capacity of the scheduled users. The actual throughput is then multiplied and summed according to this weight to generate the local reward for each RB. The multi-agent system stores experience tuples containing states, actions, and rewards into an experience trajectory pool. After mini-batch sampling, it calculates the value loss of the value network (Critic), the pruning policy loss of the policy network (Actor), and the joint policy entropy regularization term, and jointly constructs the overall objective function to complete the iterative update of network parameters.

[0035] The specific state collection and feature extraction stage includes the following steps:

[0036] 1.1 Obtain the downlink channel response column vector of the k-th candidate user in the system at the nth RB (RBn) under the current transmission time interval (TTI) t. (where M is the number of base station transmit antennas); and based on the downlink channel response column vector Calculate the channel strength scalar of candidate users on the RB. .

[0037] 1.2 Based on a preset fixed orthogonal basis matrix (Where M is the number of base station transmit antennas, and B is the beam space dimension), calculate the multidimensional spatial projection power vector of candidate users. ,in This represents the conjugate transpose operation of a matrix. This represents the square of the modulus of each element in the vector.

[0038] Subsequently, a normalization operator is used to extract the beam energy distribution column vector characterizing the spatial distribution features of the users. .

[0039] For the k-th candidate user, its scalar form of the channel strength With the beam energy distribution column vector Feature concatenation is performed to construct the local state feature column vector of the user. Specifically, it is expressed as:

[0040] ;

[0041] Subsequently, the local state feature column vectors of all K candidate users on the resource block are... After horizontally stitching the blocks side-by-side, the entire block is transposed to construct its complete state matrix. Specifically, it is expressed as: .

[0042] The multi-user scheduling decision-making stage based on dual-head PPO includes the following steps:

[0043] 2.1 The state matrix of the current resource block The input is fed into the global feature encoder of the PPO network to extract the hidden layer feature tensor, which characterizes the channel quality and spatial interference structure of candidate users in the current resource block. This hidden layer feature tensor is then synchronously input into two parallel action output heads.

[0044] 2.2 Output a feature vector representing the distribution of the number of reuses through the first action output head. The maximum number of reuses is limited by the maximum number of concurrent connections allowed by the system. ; for the feature vector The Softmax activation function is executed to obtain the probability distribution of the reuse quantity, and random sampling is performed based on the probability distribution to determine the actual number of concurrent users for the current resource block. ;

[0045] 2.3 Output the preference score vector for all candidate users through the second action output head. Subsequently, execution The subsequent sampling operation is as follows: In the first sampling, a Softmax operation is performed on all candidate users, and the first user is selected according to the probability distribution. In subsequent sampling, a masking operation is performed on the preference scores of the selected users, setting their corresponding preference scores to a preset minimum value so that the sampling probability of the selected user approaches zero when performing Softmax processing. The Softmax probability distribution is then recalculated from the remaining candidate users, and sampling is performed. This process is repeated until all users are selected. A set of users who generate the current resource block's scheduling user set. The base station combines the scheduling user set of all resource blocks to form a global scheduling matrix for use in the subsequent physical layer precoding and link adaptation phases of the system.

[0046] The physical layer precoding and link adaptation stage includes the following steps:

[0047] 3.1. Based on the set of scheduled users, determine the set of resource block indices allocated to target user k in the current TTI. And the total number of resource blocks occupied by the target user k. ,in For set The number of elements in the middle.

[0048] 3.2 Assume a set of scheduling users Include For each user, extract the downlink channel column vector corresponding to each scheduled user. (in Perform conjugate transpose operations on each column vector and concatenate them row by row to form the joint channel matrix of the resource block. Specifically, it means:

[0049] ;

[0050] in, This represents the set of scheduled users on this RB at time t. Based on the joint channel matrix... Perform ZF precoding to obtain the initial precoding matrix. After power normalization of the initial precoding matrix, the final precoding matrix is ​​obtained. .

[0051] 3.3. The signal-to-interference-plus-noise ratio (SIR) of target user k on resource block n is calculated as follows:

[0052] ;

[0053] in, Noise power; , and These represent the downlink channel vector, transmit power, and precoding vector of user k on that RB, respectively.

[0054] 3.4. Based on a pre-trained link adaptive network, select the optimal MCS for the target user k;

[0055] The attention-based link adaptive neural network aims to establish a mapping relationship between the frequency domain SINR characteristics of the RB set occupied by the target user and the physical layer transmission success rate, and selects the optimal MCS based on the expected throughput maximization criterion. Specifically, it includes the following steps:

[0056] 3.4.1 For target user k, extract its SINR scalar across all assigned RBs. Compare it with the candidate MCS index. ( Feature concatenation is performed to construct the RB-level input vector. .

[0057] 3.4.2. Using Link Feature Coding Network Extract the link feature vector from the input features of each RB: Where d is the output dimension of the link feature encoding network, and the network shares parameters for all RBs occupied by this user. .

[0058] 3.4.3. The scoring network in the frequency domain attention module is used to evaluate the link feature vectors corresponding to each RB to obtain the importance score of each resource block, specifically: ,in, These are the learnable weight parameters for the scoring network.

[0059] 3.4.4. The scores are normalized in the frequency domain using the Softmax function to obtain the importance weights corresponding to each RB. Subsequently, the importance weights are summed with their corresponding link feature vectors to generate the global channel representation vector for the target user k: .

[0060] 3.4.5. Based on the candidate MCS index and the total number of RBs occupied by users Determine the transport block size And calculate the corresponding code block size. and number of code blocks Specifically, the determination of the TBS follows the mapping rules for the downlink shared channel transport block size in the 3GPP TS 38.214 protocol; while the calculation of CBS and C follows the code block segmentation process in the 3GPP TS 38.212 protocol, determining the segmentation parameters based on the selected LDPC basemap type and its corresponding upper limit of code block length. This involves representing the global channel. With code block features After being spliced ​​together, the data is input into the reliability assessment network Ψ to predict the probability of successful transmission under the given channel conditions. .

[0061] 3.4.6. Calculate the expected throughput for each MCS: Select the option that maximizes the desired throughput. The largest candidate MCS as the target user The final modulation and coding strategy is as follows: .

[0062] 3.5. Generate and send scheduling signaling based on the global scheduling matrix and the optimal MCS for each target user; subsequently, combine the final precoding matrix obtained in step 3.2. Perform downlink data transmission to complete the scheduling and transmission within the current TTI.

[0063] The reward allocation and model update phase includes the following steps:

[0064] 4.1 The base station, based on each dispatched user The physical layer acknowledgment information is fed back, defining the transmission indication factor. If the target user If the feedback is ACK, then take If the feedback is NACK, then take... .

[0065] 4.2 Calculate the target resource block For target users Local capacity contribution weights:

[0066] ;

[0067] By combining the transmission indicator factor, transmission block size, and contribution weight, the user is calculated. In resource block Reward slices on The reward for the target resource block is obtained by summing the rewards of the set of users scheduled on that resource block. ; and the trajectory data Store in the trajectory buffer.

[0068] 4.3 After collecting a sampling window of a preset length, the data in the trajectory buffer is divided into multiple batches, and multiple rounds of iterative optimization are performed. The following sub-steps are executed in parallel during each round of optimization:

[0069] 4.3.1 Extracting the scheduling user set under the new and old policy networks. The joint logarithmic probability is used to calculate the importance sampling ratio. The generalized dominance estimate was calculated using the reinforcement learning temporal difference method. Combined with the generalized dominance estimate Calculate the policy loss with pruning mechanism:

[0070] ;

[0071] 4.3.2 Calculation based on return target Value network mean square error loss ;

[0072] 4.3.3 Calculate the joint entropy regularization term of the strategy, which includes the entropy of the population distribution and the entropy of the sequence selection conditions. To encourage movement exploration; among them, Entropy represents the distribution of the number of reused users. The conditional entropy representing the sequence selection process;

[0073] 4.3.4 Constructing the overall objective loss function ,in These represent the value loss weight and the entropy regularization coefficient, respectively. Backpropagation is performed through the optimizer to synchronously update the policy network parameters. With value network parameters .

[0074] The key point of this application is:

[0075] Multi-agent architecture design with RBs as independent agents: This invention configures each RB as an independent agent for multi-agent modeling, replacing the existing network modeling method that treats users as agents. This improvement solves the problems of coarse scheduling granularity, difficulty in combining frequency-selective fading to achieve RB-level differentiated resource allocation, and insufficient model adaptability when the number of candidate users changes dynamically in existing schemes. This achieves fine-grained scheduling at the RB granularity, improves the flexibility of user resource allocation, and enhances the model's scalability and engineering applicability.

[0076] Based on a dual-head PPO network and dynamic masking action space decoupling mechanism: This invention constructs a parameter-sharing dual-head policy network to separately determine the number of multiplexed users and select the user set, and combines dynamic masking to complete user selection. This mechanism overcomes the technical shortcomings of existing DRL schemes when dealing with multi-user combination policies: when using a full mapping coding method, the action space dimension increases sharply with the number of candidate users; when using a fixed concurrent user number method, the resource allocation flexibility is reduced. This reduces the complexity of policy solving and the difficulty of training convergence, enabling the system to adaptively form multi-user multiplexing combinations based on real-time channel conditions and spatial interference relationships.

[0077] This invention introduces a physical layer link-adaptive reward loop: After completing the scheduling decision, it obtains the SINR of each user based on precoding and inputs it into a pre-trained attention-based link-adaptive neural network to predict the transmission success probability of each candidate MCS, thereby calculating the expected throughput to determine the optimal MCS. Finally, it combines ACK / NACK feedback to construct a reward signal. This solves the technical problem in existing schemes where optimization based on idealized Shannon throughput rewards leads to a deviation between the optimization objective and the actual achievable performance of the system. It achieves effective integration between scheduling decisions and actual physical layer transmission results, enabling the scheduling strategy to be optimized based on the system's actual effective throughput.

[0078] The technical effects of this application through the above technical solution are as follows:

[0079] This invention enables fine-grained resource scheduling at the RB level, allowing the system to combine the channel quality differences and frequency-selective fading characteristics between different RBs to execute differentiated concurrent user configurations, thereby more fully leveraging the spatial multiplexing potential of the MU-MIMO system and improving the overall spectrum utilization efficiency and effective throughput of the system.

[0080] The dual-head PPO network architecture ensures resource scheduling flexibility while effectively controlling the increase in spatial complexity caused by multi-user combination decisions. By decoupling the action space, it is beneficial to improve the stability and trainability of the scheduling strategy solution process, enabling the system to adaptively form multi-user multiplexing combinations based on real-time channel conditions and spatial interference relationships, thereby improving the overall scheduling performance in complex and dynamic wireless environments.

[0081] This invention enables scheduling optimization results to more closely approximate actual physical layer transmission performance, reducing deviations caused by optimization based on idealized capacity formulas. By linking scheduling results with actual precoding, MCS selection, and ACK / NACK feedback, the system optimization objective is more directly aligned with the actual effective throughput, thereby enhancing the engineering practical value of the obtained scheduling strategy and avoiding the problem of theoretically superior performance but poor actual transmission benefits.

[0082] In summary, this invention addresses the problems of coarse scheduling granularity, high complexity of action modeling, and the disconnect between optimization objectives and actual physical layer transmission results in existing technologies. It proposes corresponding improvements, enhancing the MU-MIMO frequency-selective scheduling scheme in terms of frequency domain resource utilization, user multiplexing flexibility, and actual effective throughput optimization. Specifically, this method not only more fully leverages frequency selectivity gain and multi-user spatial multiplexing gain, but also maintains good adaptability and operational stability under conditions of dynamic changes in candidate user size, time-varying channel states, and complex spatial interference relationships.

[0083] To verify the feasibility and performance of the method of this invention, this embodiment constructs a single-base station multi-user MU-MIMO downlink transmission scenario based on the NS3 simulation platform, and conducts simulation verification under this scenario. The system simulation parameters and scenario configuration are shown in Table 1.

[0084] In this embodiment, the training samples for the link adaptive neural prediction network are generated by the aforementioned NS3 simulation environment. Specifically, under different channel quality, user distribution, and resource allocation conditions, the downlink data transmission process is simulated, and the SINR sequence, candidate MCS index, and ACK / NACK results fed back from the physical layer for the target user on its allocated RB are collected to construct the link adaptive training dataset. The ACK / NACK results serve as classification labels for network training, used to supervise the network's learning of transmission success probabilities under different channel conditions. The hierarchical structure, input / output dimensions, and main training hyperparameters of the link adaptive neural prediction network are shown in Tables 2 and 3, respectively.

[0085] The scheduling decision-making part is trained using the PPO algorithm, and its main training parameters are shown in Table 4. During the training process, each RB agent executes user reuse decisions based on the real-time state matrix and obtains reward feedback calculated based on the actual throughput of the physical layer from the simulation environment. The scheduling strategy is learned by storing the decision experience in the trajectory buffer and using the pruning objective function of PPO for multiple rounds of iterative updates.

[0086] Table 1: MU-MIMO downlink simulation scenarios and system parameters

[0087]

[0088] Table 2: Link Adaptive Neural Network Structure Parameters

[0089] Table 3: Hyperparameters for training the link-adaptive neural network

[0090] Table 4: Training Hyperparameters of the PPO Scheduling Network

[0091] Those skilled in the art should understand that the above embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention. Any reasonable modifications, equivalent substitutions, or adaptive improvements made based on the technical concepts disclosed in this invention without departing from the essential spirit of the invention should be considered to fall within the scope of protection defined by the claims of this invention.

Claims

1. A joint optimization method for MU-MIMO frequency-selective scheduling and link adaptive operation, characterized in that, Includes the following steps: S1: Architecture settings; A multi-user multiple-input multiple-output (MU-MIMO) system is set up, which includes a base station gNB and K candidate user devices (UEs). The base station is equipped with M transmit antennas, and the system frequency domain resources are divided into N resource blocks RB. The base station and each UE have a frequency-selective fading channel. The base station performs resource scheduling and downlink data transmission based on the channel state information (CSI) reported by the UE. S2: State collection and feature extraction; Based on the CSI reported by the UE within the current transmission time interval (TTI), the base station extracts the channel strength features and spatial beam features of each candidate user on each RB, splices them together to construct the state matrix corresponding to each RB, and uses the state matrix as the input features of the corresponding RB agent. S3: Multi-user scheduling decision optimization of PPO based on dual-head proximal policy; treat each RB as an independent agent and use a parameter-sharing dual-head PPO policy network to execute scheduling decisions; Taking the state matrix of a single RB as input, the number of multiplexed users of the RB is first determined through the first output head of the policy network. Then, through the second output head of the policy network combined with the dynamic masking mechanism, users matching the number of multiplexed users are selected from the candidate user set to generate the scheduling user set of the RB. Finally, the scheduling user sets of all RBs are combined to obtain the global scheduling matrix. S4: Physical layer precoding and link adaptive processing; Based on the set of scheduled users for each RB, the base station constructs a joint channel matrix for the corresponding RB, performs zero-forcing ZF precoding and power normalization to obtain the final precoding matrix, calculates the signal-to-interference-plus-noise ratio (SINR) of each scheduled user on the corresponding RB, maps the SINR of each user on all the RBs allocated to it to RB-level link features, predicts the transmission success probability of candidate MCSs through a pre-trained link adaptive neural network based on attention aggregation, and then determines the optimal modulation and coding scheme (MCS) for each user based on the expected throughput maximization criterion. Based on the global scheduling matrix and the optimal MCS of each user, scheduling signaling is generated and issued, and downlink data transmission is performed in conjunction with the final precoding matrix. S5: Reward Allocation and Model Update; The base station receives the ACK or NACK information from the UE and calculates the actual throughput of each user; Based on the effective capacity contribution weight of each RB to its scheduled users, the corresponding actual throughput is divided and accumulated to obtain the local reward of each RB agent; The state, scheduling action, and local reward of each RB are constructed into an experience tuple and stored in the trajectory buffer. After sampling, the overall objective function is constructed by combining the pruning policy loss, value network loss, and policy joint entropy regularization term of the PPO algorithm according to the preset weights, and the parameters of the policy network and value network are iteratively updated.

2. The MU-MIMO frequency-selective scheduling and link adaptive joint optimization method according to claim 1, characterized in that, S2 specifically includes the following sub-steps: S21: Obtain the downlink channel response column vector of the k-th candidate user on the n-th RB at the current TTI time t. for: ； The channel strength scalar of the candidate user on this RB is calculated. : ； S22: Based on a preset fixed orthogonal basis matrix ,Will The multidimensional spatial projection power vector of the candidate user is calculated by projecting from the antenna domain to the beam domain. After normalization, the beam energy distribution column vector representing the spatial distribution characteristics of users is obtained. ; S23: Scalarize the channel strength of a single candidate user Beam energy distribution column vector By concatenating the features, we obtain the local state feature column vector of the user. ; The local state feature column vectors of all K candidate users on a single RB The state matrix of RB is constructed by transposing the horizontally arranged rows side by side: ； Where B represents the beam space dimension.

3. The MU-MIMO frequency-selective scheduling and link adaptive joint optimization method according to claim 2, characterized in that, The dual-head PPO policy network in S3 includes a shared feature encoder and a first output head and a second output head connected in parallel to the output of the shared feature encoder. The shared feature encoder is used to extract the hidden layer feature tensor of the input state matrix and output the hidden layer feature tensor synchronously to the first output head and the second output head.

4. The MU-MIMO frequency-selective scheduling and link adaptive joint optimization method according to claim 3, characterized in that, In S3, the specific process of determining the number of multiplexed users through the first output header is as follows: The first output head outputs a feature vector representing the distribution of the number of reuses. A Softmax activation function is applied to the feature vector to obtain a probability distribution of the number of reuses. Random sampling is then performed based on this probability distribution to determine the actual number of concurrent reuse users in the current RB. ,in No more than the maximum number of concurrent users allowed by the system .

5. The MU-MIMO frequency-selective scheduling and link adaptive joint optimization method according to claim 4, characterized in that, In S3, the specific process of generating the scheduling user set through the second output header combined with the dynamic masking mechanism is as follows: The second output header outputs a preference score vector for all candidate users, and the process is executed. The sampling process involves several iterations. The first sampling performs a Softmax operation on the preference score vectors of all candidate users, selecting the first user based on the resulting probability distribution. In subsequent sampling rounds, a masking operation is performed on the preference scores of selected users, setting their scores to a preset minimum value so that the sampling probability of these selected users approaches zero during Softmax processing. The Softmax probability distribution is then recalculated from the remaining candidate users, and sampling continues until all users are selected. For each user, generate the current RB's set of scheduled users.

6. The MU-MIMO frequency-selective scheduling and link adaptive joint optimization method according to claim 5, characterized in that, In S4, the construction and precoding of the joint channel matrix specifically include: Extract the downlink channel response column vectors corresponding to each scheduled user in the set of scheduled users of a single RB. After performing conjugate transpose operations on each column vector, the vectors are concatenated row by row to construct the joint channel matrix of the RB: ； Zero-forcing precoding is performed based on the joint channel matrix to obtain an initial precoding matrix. After power normalization of the initial precoding matrix, the final precoding matrix of the RB is obtained. .

7. The MU-MIMO frequency-selective scheduling and link adaptive joint optimization method according to claim 6, characterized in that, In S4, the formula for calculating the signal-to-interference-plus-noise ratio (SINR) of the scheduled user on the corresponding RB is: ； Among them, a n (t) represents the set of scheduled users for the nth RB at time t. The noise power on this RB, , and These are the downlink channel vector, transmit power, and precoding vector for user k on that RB, respectively.

8. The MU-MIMO frequency-selective scheduling and link adaptive joint optimization method according to claim 7, characterized in that, In S4, the specific process of determining the optimal MCS using a pre-trained attention-based link adaptive neural network includes: S41: Map the SINR of the target user on all its assigned RBs to RB-level link feature vectors, input them into a pre-trained link adaptive neural network based on attention aggregation, and output the transmission success probability of the target user under each candidate modulation and coding scheme (MCS). Let m be the number of the candidate MCS. ； S42: Based on the candidate MCSm and the total number of RBs occupied by the target user. Determine the corresponding transport block size And calculate the expected throughput corresponding to the candidate MCS; S43. Select the candidate MCS that maximizes the expected throughput as the optimal modulation and coding strategy for the target user. .

9. The MU-MIMO frequency-selective scheduling and link adaptive joint optimization method according to claim 8, characterized in that, In S5, the calculation and allocation process of local rewards for each RB is as follows: Calculate the local capacity contribution weight of the nth RB to its scheduled target user k. The reward slice is obtained by multiplying the weight by the corresponding user's transport block size. ; The local reward of this RB is obtained by summing the reward slices of all users scheduled for this RB. and the trajectory data corresponding to the RB. Store in the trajectory buffer.

10. The MU-MIMO frequency-selective scheduling and link adaptive joint optimization method according to claim 9, characterized in that, In S5, the specific process of constructing the overall objective function and iteratively updating the network parameters is as follows: The data in the trajectory buffer is sampled in small batches, and the joint logarithmic probability of the scheduling user set under the new and old policy networks is calculated to obtain the importance sampling ratio. The generalized dominance estimate is calculated using the reinforcement learning temporal difference method. The policy loss with pruning mechanism is then calculated by combining the importance sampling ratio with the generalized dominance estimate. Calculate the mean squared error loss of the value network based on the return objective, and the joint entropy regularization term of the strategy that includes the entropy of the distribution of reused users and the entropy of the sequential selection condition; Construct the overall target loss function Where θ represents the policy network parameters and ψ represents the value network parameters. Weighting for value loss. The entropy regularization coefficient is used to perform backpropagation through the optimizer, synchronously updating the policy network parameter θ and the value network parameter ψ.

Images