A joint grant and beamforming method for a cell-free system and related devices

Abstract

This application provides a joint licensing and beamforming method and related equipment for a cellular system, relating to the field of wireless communication technology. This method introduces an end-to-end joint learning optimization framework consisting of a user-base station licensing learner and a beamforming learner into the central processing unit. This enables the synchronous generation of licensing decisions that satisfy system discrete constraints and beamforming vectors that comply with base station transmit power limits directly from downlink channel state information. This effectively overcomes the performance loss caused by decoupling in traditional staged methods. It not only strictly guarantees the hard constraints determined by fronthaul capacity, such as each user being served by a maximum number of designated base stations and each base station serving a maximum number of users, but also significantly improves spectral efficiency and algorithm real-time performance in cellular systems, achieving efficient collaborative optimization of discrete and continuous variables.

Description

Technical Field

[0001] This application relates to the field of wireless communication technology, and in particular to a joint licensing and beamforming method for non-cellular systems and related equipment. Background Technology

[0002] Cellular wireless communication systems, through the collaborative service of numerous distributed base stations to user equipment, effectively eliminate the cell boundary effect in traditional cellular networks, significantly improving spectral efficiency, energy efficiency, and user fairness, and are considered one of the key enabling architectures for future 6G networks. In this system, jointly optimizing the licensing relationship between users and base stations and the downlink beamforming vector is the core mechanism for achieving performance gains. However, this problem simultaneously involves strongly coupled discrete and continuous variables, constituting a typical mixed-integer nonlinear programming problem, belonging to the category of nondeterministic polynomial-hard problems, making it difficult to find the global optimal solution within a reasonable time. Existing technologies typically employ a phased heuristic strategy, first determining the user-base station licensing relationship through channel gain sorting or greedy algorithms, and then optimizing beamforming separately under a fixed license. However, this decoupling process ignores the strong coupling characteristics between the two, leading to a significant degrade in system performance. To improve performance, some studies have attempted to relax discrete variables into continuous variables before joint optimization, such as using continuous convex approximations or alternating direction multiplier methods. However, these methods not only introduce performance losses due to relaxation but also have high computational complexity and slow convergence, making it difficult to meet the real-time requirements of millisecond-level decision-making in practical systems. Summary of the Invention

[0003] This application provides a joint licensing and beamforming method and related equipment for a cellular-free system to solve one or more technical problems existing in the prior art, and at least provides a beneficial option or creates conditions that enable efficient joint optimization of discrete licensing decision and continuous beamforming, significantly improving spectral efficiency and algorithm real-time performance while strictly meeting the system's hard constraints.

[0004] On the one hand, this application provides a joint licensing and beamforming method for a cellular system, the cellular system including a central processing unit, multiple base stations and multiple users, wherein all base stations are connected to the central processing unit, and the base stations provide cooperative transmission services to the corresponding users according to dynamic licensing relationships; The method includes the following steps: The channel state information of the downlink is obtained at the central processing unit; The channel state information is input into a pre-trained user-base station grant learner, which outputs discrete variables representing the grant relationship between the user and the base station. The discrete variables are used to represent the grant decision and satisfy preset discrete constraints. The channel state information and the discrete variables are input together into a pre-trained beamforming learner, which outputs a beamforming vector that satisfies the base station transmit power constraint. The user-base station grant learner and the beamforming learner constitute an end-to-end joint learning optimization framework for jointly optimizing discrete grant decisions and continuous beamforming vectors.

[0005] Furthermore, the discrete variable is a binary matrix, where the position of the element with a value of 1 corresponds to the authorized user-base station pair; the user-base station authorization learner outputs the support set of the binary matrix, which consists of the indices of all user-base station pairs with an element value of 1, and is used to characterize the non-zero structure of the discrete variable.

[0006] Furthermore, the user-base station authorization learner adopts an encoder-decoder neural network architecture, wherein: The encoder is configured to first separate the complex channel state information into real and imaginary parts and perform vectorization processing, and then map it into a channel embedding vector; The decoder is configured to generate index elements in the support set one by one in an autoregressive manner, generating a user-base station pair index as the current action each time, and updating the internal state based on the generated action sequence until a complete support set is generated.

[0007] Furthermore, when the decoder generates each action, the following operations are performed: Based on the context state of the current decoding step and the channel embedding vector, calculate the conditional probability distribution of all unselected user-base station pairs as candidate actions; Based on the discrete constraints, identify and exclude candidate actions that would lead to constraint violations, and force their corresponding conditional probabilities to zero. The next action is selected from the remaining legal candidate actions based on the modified probability distribution, thereby dynamically constructing an action space that satisfies all discrete constraints during the decoding process.

[0008] Furthermore, the discrete constraint conditions include at least one of the following: (1) Each user is served by a maximum of N base stations, where N is a positive integer determined based on the system's fronthaul link capacity; (2) Each base station can serve a maximum of M users, where M is a positive integer determined based on the system's fronthaul link capacity; (3) The total number of authorized relationships does not exceed the upper limit set by the system; In any decoding step, if the addition of a candidate user-base station pair would cause any of the discrete constraints to be broken, then the candidate action is determined to be illegal and removed from the set of possible actions.

[0009] Furthermore, the beamforming learner includes a channel embedding layer, a multi-layer graph message passing update module, and a beam vector output layer connected in sequence; The channel embedding layer is used to decompose complex channel state information into real and imaginary parts and then vectorize it. The graph message passing update module iteratively transmits aggregated information between base station nodes and user nodes based on the user-base station connection topology defined by the discrete variables. The number of iterations is pre-configured according to system performance requirements. After generating the initial beamforming vector, the beam vector output layer performs power normalization processing on the sub-vector corresponding to each base station, so that the transmit power of each base station does not exceed its maximum allowable power threshold.

[0010] Furthermore, the user-base station authorization learner is optimized using a policy gradient-based reinforcement learning algorithm during the offline training phase, and its reward signal is the system weighted sum rate. The beamforming learner employs supervised or unsupervised learning methods during offline training. It maximizes the system and rate objective functions through a stochastic gradient ascent algorithm and is trained collaboratively with the user-base station grant learner through a weighted joint loss function. The weights of the reinforcement learning strategy gradient loss and the beamforming learning loss weights are dynamically adjusted according to system performance requirements to achieve end-to-end performance optimization.

[0011] On the other hand, this application provides a cellular-free system, which adopts a wireless communication architecture, including a central processing unit, multiple base stations and multiple users, with all base stations connected to the central processing unit; The central processing unit is used to execute the aforementioned joint licensing and beamforming method for non-cellular systems, and is responsible for centralized channel state information acquisition, user-base station dynamic association decision-making, and downlink beamforming vector calculation. Each user is served by multiple base stations, and each base station can serve multiple users simultaneously, forming a many-to-many cooperative transmission relationship.

[0012] On the other hand, this application provides an electronic device including a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the aforementioned joint licensing and beamforming method for non-cellular systems.

[0013] On the other hand, this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the aforementioned joint licensing and beamforming method for non-cellular systems.

[0014] The beneficial effects of this application are as follows: This application provides a joint licensing and beamforming method for a cellular system. This method introduces an end-to-end joint learning optimization framework consisting of a user-base station licensing learner and a beamforming learner into the central processing unit. It can directly and synchronously generate licensing decisions that satisfy the system's discrete constraints and beamforming vectors that comply with base station transmit power limits from downlink channel state information. This effectively overcomes the performance loss caused by decoupling processing in traditional staged methods. It not only strictly guarantees the hard constraints determined by fronthaul capacity, such as each user being served by a maximum number of designated base stations and each base station serving a maximum number of users, but also significantly improves the spectral efficiency and algorithm real-time performance in a cellular system, achieving efficient collaborative optimization of discrete and continuous variables. This application also provides related equipment for the above method. The beneficial effects of the related equipment are similar to those of the above method and will not be elaborated here.

[0015] Other features and advantages of this application will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the application. The objectives and other advantages of this application may be realized and obtained by means of the structures particularly pointed out in the description, claims and drawings. Attached Figure Description

[0016] The accompanying drawings are provided to further understand the technical solutions of the present invention and constitute a part of the specification. They are used together with the embodiments of the present invention to explain the technical solutions of the present invention, and do not constitute a limitation on the technical solutions of the present invention.

[0017] Figure 1 This is a flowchart of the joint licensing and beamforming method for the non-cellular system provided in this application; Figure 2 This is a schematic diagram of the method for the joint user-base station licensing learner and beamforming learner provided in this application; Figure 3 This is a schematic diagram of the user-base station authorization learner provided in this application; Figure 4 This application provides A schematic diagram of the system and speed performance under certain conditions; Figure 5 This application provides A schematic diagram of the system and speed performance under certain conditions; Figure 6 This is a schematic diagram comparing the inference latency of various methods provided in this application. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0019] The present application will be further described below with reference to the accompanying drawings and specific embodiments. The described embodiments should not be considered as limitations on the present application, and all other embodiments obtained by those skilled in the art without inventive effort are within the scope of protection of the present application.

[0020] In the following description, references are made to “some embodiments,” which describe a subset of all possible embodiments. However, it is understood that “some embodiments” may be the same subset or different subsets of all possible embodiments and may be combined with each other without conflict.

[0021] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0022] Cellular wireless communication systems, as a key candidate architecture for future 6G networks, are based on the principle of abandoning the traditional cell-centric approach of cellular networks. Instead, they employ a large number of distributed, low-power access points, under the unified coordination of a central processing unit, to collaboratively transmit and receive data for user equipment. This architecture fundamentally eliminates cell edge effects, significantly improving the system's spectral efficiency, energy efficiency, and fairness among users, making it particularly suitable for scenarios with high-density access and high data rate requirements. In cellular systems, each user is typically served simultaneously by multiple geographically dispersed base stations, and each base station may also provide signals to multiple users simultaneously, thus forming a complex many-to-many service relationship.

[0023] To achieve efficient collaboration, the system must address two tightly coupled core issues: first, determining the user-base station licensing relationship, i.e., deciding which base stations serve which users; and second, designing the downlink beamforming vector, i.e., calculating the optimal precoding weights for each base station for the users it serves. These two issues together determine the overall system performance and inherently possess a mixed-integer nonlinear programming mathematical structure. The licensing relationship is a discrete variable, while beamforming is a continuous variable; their high coupling makes finding the global optimum computationally a nondeterministic polynomial problem.

[0024] To address this challenge, existing technologies primarily employ a phased heuristic strategy. A typical approach involves first pre-determining the licensing relationship between users and base stations based on metrics such as channel gain, distance, or signal-to-interference-plus-noise ratio (SNR). This is done through ranking, thresholding, or greedy algorithms. For example, only a few base stations with the strongest channels might be allowed to serve a particular user, or each base station might be limited to serving a fixed number of users. Based on this fixed licensing relationship, a beamforming optimization problem is then solved separately, typically modeled as a convex optimization problem and solved using methods such as semidefinite relaxation, second-order cone programming, or the minimum mean square error criterion. However, this decoupled approach ignores the inherent strong coupling between licensing decisions and beamforming. In reality, the optimal licensing relationship often depends on the specific beamforming scheme, and vice versa. Because the phased approach cannot coordinate the adjustment of both, system performance falls far below the theoretical upper limit, especially under dense user conditions or complex channel conditions, where performance losses are even more significant.

[0025] To alleviate these problems, some studies have attempted to relax discrete grant variables into continuous variables, thereby transforming the original mixed-integer problem into a differentiable continuous optimization problem. For example, continuous convex approximation, a weighted minimum mean square error framework combined with alternating optimization, or the introduction of auxiliary variables using the alternating direction multiplier method for joint solution are employed. While these methods improve performance to some extent, they still have significant drawbacks. On the one hand, variable relaxation itself introduces model distortion, and the resulting continuous solution needs subsequent thresholding or quantization before it can be used for actual granting, inevitably causing performance loss. On the other hand, these iterative optimization-based methods typically have high computational complexity and slow convergence speed, making it difficult to meet the stringent latency requirements of millisecond-level real-time decision-making in cellular-free systems, especially in massive MIMO and large-user scenarios, where their feasibility is further limited.

[0026] In recent years, with the development of artificial intelligence, deep learning-based learning optimization methods have been introduced into the field of wireless resource management, demonstrating great potential for low-latency inference. Some works have attempted to construct neural networks that directly map channel state information to beamforming vectors, but they face fundamental obstacles when the problem involves discrete grant decisions. Neural networks inherently rely on gradient backpropagation for training, and discrete outputs cannot provide effective gradient signals.

[0027] To address this, researchers often employ pass-through estimators or the Gumbel-Softmax technique to approximate discrete operations with differentiability. However, these approximation methods are prone to introducing incorrect gradient directions, leading to training instability, convergence difficulties, and a lack of assurance that the generated licensing relationships strictly satisfy the system's hard constraints. For example, in practical non-cellular systems, due to limited fronthaul link capacity, it is typically stipulated that each user can be served by a maximum of N base stations, and each base station can serve a maximum of M users. Existing learning methods often incorporate these constraints as soft penalty terms into the loss function, failing to absolutely guarantee constraint satisfaction during the inference phase. This can lead to infeasible licensing schemes and affect system reliability.

[0028] In summary, current joint licensing and beamforming technologies in non-cellular systems still face three major challenges: traditional phased methods suffer from suboptimal performance due to decoupling; continuous relaxation optimization methods, while capable of joint solutions, are computationally complex and incur performance losses; and existing deep learning methods struggle to effectively address the coexistence of discrete decision-making and rigid combination constraints.

[0029] To address the aforementioned issues, this application provides a joint licensing and beamforming method and related equipment for cellular-free systems. It constructs an end-to-end joint learning optimization framework, composed of a user-base station licensing learner and a beamforming learner. Under the unified scheduling of the central processing unit, it can directly and synchronously generate licensing decisions that satisfy system discrete constraints and beamforming vectors that comply with base station transmit power limits from downlink channel state information. The licensing learner is specifically designed to output discrete variables that strictly adhere to hard combined constraints such as each user being served by a maximum number of designated base stations and each base station serving a maximum number of users, avoiding infeasible solutions caused by relaxed approximations or soft constraints in traditional methods. The beamforming learner, using channel state information and the established licensing relationship as joint input, efficiently generates continuous precoding vectors that satisfy power constraints. The entire framework aligns the two learners on the optimization objective through joint training, thereby achieving efficient collaboration between discrete licensing and continuous beamforming during the inference phase. This overcomes the performance loss of staged heuristic strategies and avoids the latency bottleneck of high-complexity iterative optimization, significantly improving system spectral efficiency and real-time decision-making capabilities.

[0030] First, the joint licensing and beamforming method for a cellular system provided in this application will be described in detail below with reference to the accompanying drawings. The cellular system includes a central processing unit, multiple base stations, and multiple users, wherein all base stations are connected to the central processing unit, and the base stations provide cooperative transmission services to the corresponding users according to dynamic licensing relationships.

[0031] Reference Figure 1 The implementation process of the joint licensing and beamforming method for non-cellular systems provided in this application embodiment includes, but is not limited to, the following steps.

[0032] Step S110: Obtain downlink channel state information at the central processing unit.

[0033] Step S110 provides the basic input data for the entire joint optimization process. In a cellular-free system, all base stations are connected to the central processing unit, which coordinates resource allocation and signal processing. Downlink channel state information reflects the wireless propagation characteristics between each base station and each user, including key parameters such as path loss, shadowing fading, and small-scale fading, and is the core basis for determining licensing relationships and beamforming strategies. After the central processing unit centrally acquires complete channel state information, it can globally grasp the network topology and channel quality distribution, thus laying the information foundation for subsequent intelligent decision-making. Without accurate and complete channel state information, any licensing or precoding scheme will deviate from the actual channel conditions, leading to a severe degradation in system performance. Therefore, this step is not only a prerequisite for method implementation but also a crucial step in ensuring effective inference by the learner.

[0034] Step S120: Input the channel state information into the pre-trained user-base station grant learner and output discrete variables representing the grant relationship between the user and the base station.

[0035] Among them, discrete variables are used to characterize the authorization decision and satisfy the preset discrete constraints.

[0036] In step S120, a specially designed user-base station grant learner directly infers the discrete service relationships that conform to the system's physical constraints from the channel state information. This learner has been pre-trained offline and has learned how to select the optimal set of serving base stations for each user under a given channel environment. Its output is a set of discrete variables, each explicitly indicating whether a base station is authorized to serve a particular user, thus forming a dynamic many-to-many cooperative topology.

[0037] Crucially, the learner is explicitly constrained in its structure and training objectives, ensuring that its output strictly meets hard discrete constraints determined by system configuration parameters such as fronthaul link capacity. For example, each user can be served by a maximum of N base stations, and each base station can serve a maximum of M users. This ability to directly generate feasible licensing schemes avoids the performance loss and infeasibility risks associated with traditional methods that rely on relaxation, thresholding, or post-processing corrections, achieving accurate modeling and efficient execution of discrete decisions.

[0038] Step S130: The channel state information and discrete variables are input into the pre-trained beamforming learner, and the beamforming vector that satisfies the base station transmit power constraint is output.

[0039] In step S130, based on the established licensing relationship and the original channel state information, a downlink beamforming vector that meets the actual hardware constraints is further generated. The beamforming learner is also pre-trained and can quickly calculate the complex precoding weights for each base station for the users it serves, using both channel state information and licensing decisions as input. The beamforming vector composed of these weights must not only maximize the received signal strength of the target user and suppress interference to other users, but also strictly adhere to the total transmit power limit of each base station to meet the actual capabilities of hardware such as RF power amplifiers.

[0040] Since the licensing relationship is fixed and feasible in the previous step, the beamforming learner can focus on optimizing continuous variables under a defined service topology, thereby improving the efficiency of signal space utilization. This step works in conjunction with step S120 to achieve a seamless connection between discrete licensing and continuous precoding within an end-to-end framework. This ensures physical feasibility while significantly reducing online computational complexity, enabling the system to complete high-quality joint resource allocation within milliseconds.

[0041] In some embodiments of this application, the discrete variable is embodied as a binary matrix. The rows of this matrix correspond to multiple users, and the columns correspond to multiple base stations. Each element in the matrix takes a value of 0 or 1. An element with a value of 1 explicitly indicates that the corresponding base station is authorized to provide cooperative transmission services to the corresponding user, while an element with a value of 0 indicates that it is not authorized. This binary matrix form intuitively and accurately depicts the dynamically changing user-base station service relationship in a cellular-free system, providing a clear topological basis for subsequent beamforming. Since the fronthaul link capacity is limited in actual systems, it is typically required that each user be served by only a small number of base stations, and each base station can only serve a limited number of users. Therefore, this binary matrix is ​​highly sparsity, meaning that the vast majority of elements are 0, with only a few key positions being 1.

[0042] To efficiently represent and process this sparse structure, the user-base station grant learner does not directly output the complete binary matrix, but instead outputs its support set. The support set is the set of indices of all user-base station pairs with an element value of 1 in the matrix, recording which specific users and which specific base stations have established grant connections. By outputting the support set instead of the full matrix, not only is the dimensionality and storage overhead of the data representation significantly reduced, but it also naturally aligns with the nature of combinatorial optimization problems, allowing the learner to focus on key decision variables. More importantly, the generation process of the support set can be ensured during the training phase through carefully designed loss functions and constraint mechanisms, strictly satisfying the discrete constraints specified by the system configuration. For example, the number of base stations associated with each user does not exceed a preset limit, and the number of users served by each base station does not exceed its capacity limit. This support set-based output method preserves the accuracy of discrete grant decisions while improving the computational efficiency and scalability of the algorithm, providing a feasible path for real-time joint optimization in large-scale non-cellular systems.

[0043] In some embodiments of this application, the user-base station grant learner employs an encoder-decoder neural network architecture. This design aims to efficiently process complex channel state information and generate discrete grant decisions that satisfy combinatorial constraints. The encoder's role is to extract and represent the structured features of the raw downlink channel state information. Since wireless channels are inherently composed of complex numbers, the encoder first separates each complex value into its corresponding real and imaginary parts, thus transforming the complex matrix into a real tensor. Subsequently, it flattens this tensor into a one-dimensional input sequence through vectorization operations. Based on this, the encoder uses a multilayer perceptron or attention mechanism to map this sequence into a set of high-dimensional channel embedding vectors. These embedding vectors can fully capture the spatial correlation, channel strength, and interference coupling relationship between the user and the base station, providing a semantically rich and compact contextual representation for subsequent grant decisions.

[0044] The decoder is responsible for progressively constructing the support set in an autoregressive manner based on the channel embedding output by the encoder. The support set represents the set of indices for all authorized user-base station pairs, and its generation process is modeled as a sequence decision problem. At each step, the decoder selects a user-base station pair index as the current action, which corresponds to a position in the binary grant matrix where an element has a value of 1. When generating each new index, the decoder not only refers to the global channel embedding but also dynamically fuses the internal state formed by the previously generated action sequences, thereby ensuring logical consistency and constraint compliance between the current decision and existing grants.

[0045] For example, if a user has already been allocated the maximum allowed number of base stations, no new base stations will be allocated to them in subsequent steps. This autoregressive mechanism allows the decoder to directly output a feasible support set that satisfies the hard discrete constraints without relying on relaxation or approximation. The entire process continues until a complete and legal authorization structure is generated, ensuring both strict satisfaction of the combinatorial constraints and effective exploration and optimization of the high-dimensional discrete space.

[0046] In some embodiments of this application, the following operations are performed when the decoder generates each action.

[0047] Step S210: Based on the context state and channel embedding vector of the current decoding step, calculate the conditional probability distribution of all unselected user-base station pairs as candidate actions.

[0048] In step S210, the decoder is provided with comprehensive candidate decision-making criteria for the current generation step. During the autoregressive generation of the support set, the decoder maintains an internal context state at each step. This state encodes historical information of the selected user-base station pairs and, combined with the global channel embedding vector extracted by the encoder, characterizes the environment for the current grant decision. Based on this joint representation, the decoder calculates the conditional probability distribution of all unselected user-base station pairs as the next action using an attention mechanism or a scoring network. This distribution reflects the expected contribution of each potential connection to system performance under the current granted structure and channel conditions, thus providing data-driven prioritization for subsequent selections. This step ensures that the decoding process not only relies on local heuristics but also fully utilizes global channel information for intelligent trade-offs.

[0049] Step S220: Based on the discrete constraints, identify and exclude candidate actions that would lead to constraint violations, and forcibly set their corresponding conditional probabilities to zero.

[0050] In step S220, the system's hard discrete constraints are embedded into the decision-making process in real time to ensure that the generated authorization scheme is always feasible. In cellular-free systems, due to limitations in fronthaul link capacity and base station processing capabilities, there are usually strict combinatorial constraints, such as each user can be served by a maximum of N base stations, and each base station can serve a maximum of M users. Before each action selection, the decoder dynamically checks each candidate user-base station pair: if selecting the pair would cause any user or base station to exceed its service limit, the candidate action is considered illegal. For all such candidate actions that would lead to constraint violations, their conditional probabilities calculated in the previous step are forcibly set to zero. This operation effectively eliminates the infeasible solution space, avoiding the uncertainty and performance loss caused by relying on post-processing corrections or soft penalty terms in traditional methods, allowing the learner to explore only within the legal region from the source.

[0051] Step S230: Select the next action from the remaining legal candidate actions according to the modified probability distribution, thereby dynamically constructing an action space that satisfies all discrete constraints during the decoding process.

[0052] In step S230, after excluding illegal candidates, actual action sampling is performed from the remaining legal actions, thereby gradually constructing a complete and compliant support set. After clearing the probability of candidates that violate the constraints, the probability distribution of the remaining legal candidate actions is renormalized, forming a corrected action space. Based on this corrected distribution, the decoder uses a deterministic strategy, such as selecting the action with the highest probability, or a random strategy, such as sampling by probability, to select the next user-base station pair index as the current output. This mechanism ensures that each newly added licensing relationship not only has high channel quality potential but also strictly satisfies the discrete constraints specified by the system configuration. By dynamically maintaining and updating the action space during the decoding process, the entire generation process can achieve efficient and accurate combinatorial optimization without sacrificing feasibility, ultimately outputting a user-base station licensing structure that is both high-performance and physically implementable.

[0053] In some embodiments of this application, discrete constraints are explicitly set to reflect the physical and resource limitations of cellular-free systems in actual deployment. The first constraint stipulates that each user can be served by a maximum of N base stations, where N is a positive integer determined by the capacity of the system's fronthaul link. Since the fronthaul link is responsible for transmitting baseband signals from the central processing unit to each distributed base station, its bandwidth and load capacity are limited. Allowing a single user to be served by too many base stations would cause a surge in fronthaul data, exceeding the link's carrying capacity. Therefore, limiting the maximum number of base stations served by each user effectively controls fronthaul overhead and ensures system stability.

[0054] The second constraint stipulates that each base station can serve a maximum of M users, where M is also a positive integer determined based on the fronthaul link capacity. This limitation stems from the signal processing capabilities of the base station and the fronthaul uplink backhaul burden. When a base station simultaneously generates beamforming signals for a large number of users, not only does the computational complexity increase significantly, but the amount of data it feeds back to or receives from the central processing unit also grows rapidly. By setting a service user limit M for each base station, it is possible to avoid local node overload while ensuring service quality, thus ensuring load balancing and efficient operation of the entire distributed network.

[0055] The third constraint further controls the scale of global authorization relationships, meaning the total number of authorized user-base station pairs must not exceed a system-preset upper limit. This upper limit comprehensively considers the total fronthaul bandwidth of the entire network, the computing resources of the central processing unit, and real-time scheduling capabilities, aiming to prevent excessively dense cooperative topologies that could lead to resource contention or processing bottlenecks. These three constraints together constitute a strict boundary of the feasible solution space. In any decoding step, if the addition of a candidate user-base station pair would cause any one of these constraints to be broken—for example, causing a user's number of serving base stations to exceed N, a base station's number of serving users to exceed M, or the total number of authorized pairs to exceed the global upper limit—then the candidate action is immediately deemed illegal and completely removed from the current set of feasible actions. This dynamic legality check mechanism ensures that the decoder makes decisions only within physically achievable limits at each step of the autoregressive generation process, thereby outputting an authorization scheme that fully satisfies the system engineering constraints and avoiding the generation of infeasible solutions and the resulting performance losses or deployment risks.

[0056] In some embodiments of this application, the beamforming learner adopts a graph-oriented deep neural network architecture, which consists of a channel embedding layer, a multi-layer graph message passing update module, and a beam vector output layer connected in sequence, so as to achieve efficient generation of beamforming vectors under complex cooperative topologies.

[0057] (1) The channel embedding layer is used to decompose complex channel state information into real and imaginary parts and vectorize them.

[0058] Specifically, the channel embedding layer preprocesses and represents the features of the raw downlink channel state information. Since wireless channels are inherently composed of complex numbers, this layer first decomposes each complex numerical channel coefficient into its corresponding real and imaginary parts, thus transforming the complex matrix into a real-valued tensor. Subsequently, it flattens this tensor into a one-dimensional or two-dimensional input format suitable for neural network processing through vectorization operations. This processing not only preserves complete channel phase and amplitude information but also provides a structured input foundation for subsequent graph neural network modules, enabling the model to effectively perceive the spatial coupling relationship between the user and the base station.

[0059] (2) The graph message transmission update module is based on the user-base station connection topology defined by discrete variables. It iteratively transmits aggregated information between base station nodes and user nodes. The number of iterations is pre-configured according to the system performance requirements.

[0060] Specifically, this module models base stations and users as two types of nodes in a graph, with user-base station pairs (each with a discrete variable value of 1) forming edges, thus creating a sparse but semantically clear bipartite graph topology. On this graph, the module repeatedly exchanges local neighborhood information, such as channel strength, interference levels, and current license status, between base station and user nodes through a multi-layered iterative message-passing mechanism. Each update integrates the embedding features of neighboring nodes and their own state, gradually refining a more globally consistent node representation. The number of iterations is pre-configured based on the system's trade-off between performance and computational overhead, ensuring sufficient information diffusion to capture long-distance dependencies while avoiding increased latency due to excessive computation. This graph-based modeling approach naturally aligns with the collaborative nature of cellular-free systems, enabling beamforming decisions to accurately focus on the actual connection pairs participating in the service, significantly improving interference coordination capabilities and signal synthesis efficiency.

[0061] (3) After the beam vector output layer generates the initial beamforming vector, it performs power normalization processing on the sub-vector corresponding to each base station so that the transmit power of each base station does not exceed its maximum allowable power threshold.

[0062] Specifically, this layer first generates initial complex precoding weights corresponding to all authorized users based on the embedding of each base station node, combining them to form a complete beamforming vector. Then, for each base station's corresponding sub-vector, power normalization is performed, i.e., the Euclidean norm of the sub-vector is calculated and scaled to not exceed the maximum permissible transmit power threshold of that base station. This normalization step is crucial, ensuring that the transmit power of all base stations strictly complies with the physical limitations of the RF front-end devices and system energy efficiency requirements, preventing signal distortion or equipment damage due to power exceeding limits. By embedding power constraints into the network output layer, the beamforming learner can directly learn feasible solutions during end-to-end training without additional post-processing or projection operations, thus maintaining high inference efficiency while ensuring physical compliance.

[0063] In some embodiments of this application, the user-base station grant learner and the beamforming learner constitute an end-to-end joint learning optimization framework for jointly optimizing discrete grant decisions and continuous beamforming vectors. This breaks the limitation of the separation between discrete decision-making and continuous optimization in traditional phased design, and realizes the collaborative generation of grant relationships and beamforming vectors and the improvement of global performance.

[0064] This framework models the entire resource allocation process of a cellular-free system as a unified, differentiable, or optimizable mapping, forming a complete path from the initial channel state information input to the final transmitted signal output. Through joint training, the two learners are aligned on the optimization objective, enabling grant decisions to consider not only local channel quality but also its impact on subsequent beamforming performance. Furthermore, the beamforming module can perform accurate precoding based on a realistic and feasible grant topology. This end-to-end architecture significantly improves the overall spectral efficiency of the system and achieves millisecond-level low-latency decision-making during the inference phase, meeting the dual requirements of high throughput and real-time performance for future 6G networks.

[0065] In some embodiments of this application, the user-base station grant learner is optimized using a policy gradient-based reinforcement learning algorithm during the offline training phase, and its reward signal is the system weighted sum rate.

[0066] Since the license output is a binary support set, belonging to a discrete action space that is discontinuous and non-differentiable, supervised learning lacks explicit labels, while reinforcement learning is naturally suitable for such sequential decision-making problems. This learner views the entire license generation process as a strategy for an agent to take a series of actions in the channel environment; its policy network is the decoder structure, and the environmental feedback is reflected in the system weighted sum rate, a key performance indicator. The weighted sum rate comprehensively reflects the achievable data rate for all users and can be weighted according to service priority, serving as a reward signal to guide policy updates. Through policy gradient methods, such as REINFORCE or its low-variance variants, the learner can adjust policy parameters based on reward levels, gradually learning to select user-base station connection combinations that maximize system capacity while satisfying hard discrete constraints.

[0067] In some embodiments of this application, the beamforming learner adopts supervised or unsupervised learning methods during the offline training phase. It maximizes the system and rate objective functions through the stochastic gradient ascent algorithm and is co-trained with the user-base station licensing learner through a weighted joint loss function. The weights of the reinforcement learning strategy gradient loss and the beamforming learning loss weights are dynamically adjusted according to the system performance requirements to achieve end-to-end performance optimization.

[0068] Supervised learning leverages high-quality beamforming vectors generated by traditional optimization algorithms such as WMMSE as labels for imitation learning, while unsupervised learning directly uses the system and rate as objective functions, eliminating the need for external solvers and offering greater scalability. Regardless of the approach, the training objective of the beamforming learner is always linked to the overall system performance. More importantly, the two learners are not trained independently but are optimized end-to-end through a joint loss function, a weighted combination of the reinforcement learning policy gradient loss and the beamforming learning loss. The reinforcement learning loss drives authorization decisions towards high-reward regions, while the beamforming loss ensures the effectiveness of continuously pre-encoded vectors. The weights of both are dynamically adjusted based on system performance requirements; for example, initially focusing on exploring diverse authorization structures, and later enhancing beam accuracy to converge to a high-performance solution. This dynamic weighting mechanism allows the entire framework to achieve a balance between discrete and continuous optimization, ultimately achieving globally optimal or near-optimal end-to-end performance.

[0069] In some embodiments of this application, taking a downlink non-cellular communication system as an example, the system includes indivual Antenna base station and Each user has a single antenna. All base stations are connected to a central processing unit for joint signal processing. Define a binary user-base station grant variable. , among which when Time indicates the first The user is from the first Individual base station services, when Time indicates the first The user is not allowed by the first Individual base station services.

[0070] Further define the first The base station and the first The channel between user equipment is ( (Representing the complex field). At this time, the... Signal received by a user It can be expressed as the following formula (1): (1); In formula (1), Indicates the first The user's expected signal transmission is satisfied. ; It is the first Beamforming vectors for each user; It has a mean of zero and a variance of Additive white Gaussian noise; yes The conjugate transpose of is used to calculate the matched filter gain of the signal in the channel. In formula (1), the first term refers to the . The useful signals received by a user equipment from all the base stations serving it; the second item refers to the useful signals received by the user equipment from all the base stations serving it. A user device receives a message from another user. ( Multi-user interference (MUI).

[0071] Accordingly, all The sum of users and rates It can be expressed as the following formula (2): (2); Furthermore, the joint user-base station licensing and beamforming design problem for the non-cellular system can be expressed as the following formula (3): (3); In formula (3), It is a column vector of all user-base station license variables; It is the concatenation of all beamforming vectors; It is the maximum transmission power of the base station.

[0072] In this optimization problem, constraint one states that each user can be at most […]. Each base station can serve a maximum of [number] base stations; constraint two indicates that each base station can serve [number] base stations at most. The first two constraints are due to the limited capacity of the fronthaul link. The third constraint states that the base station's transmit power cannot exceed the maximum transmission power. This problem involves a mix of discrete and continuous variables and contains highly coupled discrete constraints.

[0073] In some embodiments of this application, discrete variables are introduced. The concept of a support set is introduced to simplify the solution. The support set for discrete variables is defined as follows: Its relationship with discrete variables It has a one-to-one correspondence, expressed as the following formula (4): (4); Introducing discrete variables The support set has the following advantages: Most importantly, discrete variables It is a sparse discrete variable, meaning that most of its elements are 0. According to the optimization problem in formula (3), under constraints one and two, the variable contains at most 1 elements. Therefore, a support set is introduced. It can simplify the solution of neural networks, reducing the number of variables that need to be inferred by the neural network. The number was reduced to indivual.

[0074] Reference Figure 2 This application designs two learners to jointly design user-base station granting and beamforming, namely a user-base station granting learner and a beamforming learner. First, the user-base station granting learner receives channel data. Input, output discrete variable support set , is represented as: ,in Then, the beamforming learner receives... and The input is the beamforming result, and the output is represented as... Based on the above modeling, the original problem (3), i.e., formula (3), can be reformulated as the following formula (5): (5); In formula (5), among all possible user-base station grant learners and beamforming learner In the meantime, and in conjunction with the above formula (4), we find the optimal pair of functions that maximize the system and speed. maximize.

[0075] It is worth noting that the three constraints of the optimization problem in formula (3) can be satisfied by the special design of the user-base station learner and the beamforming learner.

[0076] In some embodiments of this application, reference is made to Figure 3 User-Base Station Authorization Learner It consists of an encoder and a decoder.

[0077] Specifically, the encoder first converts the complex channel into a real matrix, and then separates the real part according to the following formula (6). and the virtual part : (6); Then, the encoder passes through an embedding layer, Each update layer is embedded. ,in User and base stations Inter-channel The embedding. Specifically, the encoder obtains the initial embedding through an embedding layer. It satisfies the following formula (7): (7); in and These are learnable parameters. After that, the... Each update layer will Transform into Its update method satisfies the following formula (8): (8); in , and It consists of three nonlinear mappings, each consisting of a fully connected linear layer and a Rectified Linear Unit (ReLU) activation function.

[0078] The decoder receives the encoder's output. and channel As input, output support set probability distribution From constraints one and two in the optimization problem of formula (3), it can be seen that the support set Maximum contents There are 1 element, so it can be further divided into 1 element and 2 elements. Represented as ,in .

[0079] In order to satisfy constraints one and two in the optimization problem of formula (3), the decoder adopts an ordered sequence decoding mechanism, and its solution method satisfies the following formula (9): (9); In formula (9), in the first... During the step, the decoder outputs , Defined as The output method is to first use a system state embedding network to describe the current system state, satisfying the following formula (10): (10); In formula (10), Representing the Step-by-step system state embedding; It is a non-linear mapping, typically consisting of a linear layer and a ReLU activation function. Indicates the preceding The embedded representation after the selected action satisfies the following formula (11): (11); in, and It consists of two nonlinear mappings, consisting of a linear fully connected layer and a ReLU activation function.

[0080] Then, the decoder outputs the conditional probability through an attention network. Its output method satisfies the following formula (12): (12); In formula (12), Indicates the decoder's... When generating support set elements, candidate user-base station pairs The unnormalized attention score is used to calculate the selection. As the first An intermediate quantity of the conditional probability of an authorized connection; It is the hyperbolic tangent function. It is an adjustable hyperparameter. , , and These are learnable parameters; It consists of two types of points: one is all the elements in the original support set, i.e. Secondly, adding this element to form a new support set would violate either constraint one or constraint two.

[0081] The specification definition satisfies the following formula (13): (13); in, Representative with Discrete variables as support set Then, the conditional probability is obtained by the normalization function, satisfying the following formula (14): (14); From formulas (12), (13), and (14), it can be seen that for all elements that violate the constraints, the corresponding conditional probability is... It tends to 0, thus ensuring that discrete constraints are satisfied.

[0082] During training, after obtaining the corresponding conditional probability from formula (14), the user-base station grant learner samples the points from this probability. ,Right now And form a new support set, namely During online deployment, this step is directly selected by a greedy process. .

[0083] In some embodiments of this application, a beamforming learner Consists of an embedding layer, It consists of a message passing update layer and an output layer.

[0084] First, the embedding layer will channel and support set The initial embedding is transformed into the following formula (15): (15); In formula (15), and These are learnable parameters.

[0085] Then, the Each update layer will Transform into The method satisfies the following formula (16): (16); In formula (16), , and It consists of three nonlinear mappings, each consisting of a linear fully connected layer and a ReLU activation function.

[0086] Furthermore, the output layer first performs a dimensionality transformation on the depth features obtained from the update layer. This is done by using a linear layer to transform the depth features. Transform into A dimensional vector that satisfies the following formula (17): (17); In formula (17) and These are learnable parameters.

[0087] Furthermore, to satisfy constraint three in problem (3), the base stations exceeding the power limit are normalized using the following formula (18): (18); Finally, the real number variable Convert to complex variables .Will The former Dimensional variables as The real part will After As The imaginary part is used to obtain the final beamforming. .

[0088] In some embodiments of this application, an example of a cellular-free communication system is provided, wherein Individual base station services Number of users. Number of base station antennas: Each base station can serve a maximum of Each user equipment. All base stations and users are in Rectangular area on a plane (unit: meter) Uniformly distributed within the interior. Large-scale fading coefficients are based on the path loss model. (Unit: decibels) generated, where The distance between the base station and the user is in meters. The small-scale Rayleigh fading coefficient follows a complex Gaussian distribution with a mean of 0 and a variance of 1. The background noise power is -100 dBm.

[0089] Specifically, consider two typical scenarios: 1) In the cooperative beamforming mode, each base station only shares channel state information and does not share user data; 2) This corresponds to the joint transmission mode, in which base stations simultaneously share channel information and user data.

[0090] Secondly, configure the user-base station authorizer. The encoder contains One update layer; beamformer Include One update layer; embedding dimension is Hyperparameters Set to 8. Performance was evaluated on 1024 samples, including user, rate, and inference time.

[0091] Furthermore, a user-base station authorization learner is proposed. and beamforming learner An algorithm for joint training. Based on conditional probability The joint training of the original problem (3) can be expressed as the following optimization problem, satisfying the following formula (19): (19); in, and Represent and The learnable parameters are explained below. and Training methods.

[0092] Policy gradient (PG) is used for The update is performed in a manner that satisfies the following formula (20): (20); because right Directly differentiable, the objective function can be directly calculated with respect to... The gradient of satisfies the following formula (21): (twenty one); Therefore, the mini-batch stochastic gradient ascent algorithm (Adam optimizer) is used to update the function in order to maximize the objective function. It should be noted that equations (20) and (21) should be in negative gradient form in the optimizer to adapt to the update mechanism in the Adam optimizer.

[0093] Reference Figure 4 , Figure 4 This application provides The diagram illustrates the system and rate performance under cooperative beamforming mode. It can be seen that, under cooperative beamforming mode, the proposed framework significantly outperforms all comparative methods under different transmit power budgets. In particular, compared to methods based on pass-through estimators and GS approximations, this application directly models discrete decisions, avoiding any gradient mismatch or approximation error issues, thus achieving better system performance. Furthermore, iterative optimization-based methods can only solve the relaxed continuous optimization problem and converge to a stationary point, making it difficult to fully exploit the performance potential of mixed integer problems. Additionally, the performance of the proposed method is significantly higher than that of stepwise optimization-based methods. It is worth noting that as transmit power increases, the system gradually enters an interference-constrained region, making efficient interference management crucial. Under this high-interference environment, this application demonstrates excellent robustness, significantly outperforming various heuristic, iterative optimization, and existing deep learning methods.

[0094] Reference Figure 5 , Figure 5 This application provides The diagram illustrates the system and rate performance under the joint transmission mode. Further considering the joint transmission mode, the user-base station granting strategy is more complex, and there is strong coupling between user-base station correlation variables, rendering iterative optimization methods inapplicable. For example... Figure 5 As shown, this application maintains a significant advantage in the joint transmission mode. In contrast, the performance of methods based on pass-through estimators and GS approximations degrades more severely, highlighting their fundamental limitation in handling strongly coupled problems: they are hampered by severe gradient mismatch. Furthermore, traditional stepwise optimization methods, unable to handle complex joint granting logic, perform far worse than this application.

[0095] Reference Figure 6 , Figure 6The inference latency comparison of various methods is presented. Thanks to the efficient inference characteristics of the feedforward neural network architecture, the decision generation speed of this application is significantly faster than the method based on iterative optimization, and the latency is similar to that of traditional step-by-step and heuristic methods, demonstrating good potential for real-time deployment.

[0096] Furthermore, this application provides a cellular-free system. The cellular-free system employs a wireless communication architecture including a central processing unit, multiple base stations, and multiple users, with all base stations connected to the central processing unit. The core of this architecture design lies in improving resource allocation efficiency and service quality through centralized control, enabling the entire network to flexibly respond to complex and ever-changing wireless environments. The central processing unit, as the "brain" of the system, is not only responsible for coordinating the operations between base stations but also undertakes critical data processing and optimization tasks, ensuring information flow and collaborative work between different components, thereby providing users with a more stable and efficient wireless connection experience.

[0097] The central processing unit is used to execute the aforementioned joint licensing and beamforming method for non-cellular systems, and is responsible for centralized channel state information acquisition, user-base station dynamic association decision-making, and downlink beamforming vector calculation.

[0098] The central processing unit (CPU) can make optimization decisions based on a global perspective, rather than being limited to the local optimum of a single base station. By collecting and analyzing channel state information from each base station, the CPU can more accurately assess the current network condition and make the best user-base station pairing selection accordingly. Furthermore, for beamforming vector calculation, the CPU also relies on this comprehensive data to generate the most favorable radiation pattern for signal transmission, improving spectral efficiency and quality of service, thereby meeting the diverse needs of high-density access scenarios.

[0099] In this system, each user is served by multiple base stations, and each base station can serve multiple users simultaneously, forming a many-to-many cooperative transmission relationship. This design makes full use of spatial resources, improving system capacity and coverage. For users, simultaneous service from multiple base stations means that even if one base station is overloaded or encounters obstacles blocking the signal, other base stations can still guarantee communication quality, significantly improving the stability and reliability of the user experience. For base stations, being able to serve multiple users simultaneously maximizes infrastructure utilization efficiency and reduces resource idleness. This many-to-many cooperative mode not only enhances the system's anti-interference capability but also promotes improved frequency reuse efficiency, making it one of the key technologies for realizing future high-speed, low-latency wireless communication.

[0100] Furthermore, embodiments of this application provide an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the aforementioned joint licensing and beamforming method for non-cellular systems.

[0101] Furthermore, embodiments of this application provide a computer-readable storage medium storing a computer program that, when executed by a processor, implements the aforementioned joint licensing and beamforming method for non-cellular systems.

[0102] In summary, the joint licensing and beamforming method and related equipment for cellular-free systems provided in this application have the following technical effects.

[0103] This method introduces a support set to sparsely model discrete grant variables, transforming the high-dimensional combinatorial optimization problem into a sequence generation task, significantly reducing the search space complexity. Simultaneously, the user-base station grant learner employs an autoregressive decoding architecture and a dynamic masking mechanism, explicitly embedding system hard constraints in each action selection step. By setting the attention score of illegal candidate actions to negative infinity, it ensures that the generated grant relationships always meet the fronthaul link capacity limit, fundamentally avoiding infeasible solutions. The beamforming learner, based on a graph neural network structure, performs differentiated embedding initialization in conjunction with the grant topology and aggregates global channel information through a message passing mechanism, enabling the precoding vector to accurately coordinate multi-base station signal synthesis and interference suppression.

[0104] Furthermore, the two learners form an end-to-end joint optimization framework, collaboratively trained with system and rate as unified objectives: the authorization learner optimizes discrete decisions using a policy gradient algorithm, while the beamforming learner directly maximizes the rate through differentiable paths, and introduces a power normalization operation at the output layer to strictly satisfy the maximum transmit power constraints of each base station. In typical simulation scenarios, this method can output high-performance, physically feasible resource allocation schemes within milliseconds, outperforming traditional step-by-step or heuristic methods in terms of system and rate, constraint satisfaction rate, and computational efficiency. This provides an efficient and robust intelligent decision-making paradigm for future high-density non-cellular networks.

[0105] It should be noted that in all specific embodiments of this application, all data processing activities related to user identity or personal characteristics, such as user information, user behavior data, historical data, and location information, will be conducted in accordance with the principles of legality, legitimacy, and necessity. All data collection, use, storage, and processing will be subject to compliance with applicable national and regional laws, regulations, and industry standards, and informed consent from users will be obtained in a clear and explicit manner before processing. For the processing of sensitive personal information, separate consent from users will be obtained through prominent means such as pop-up prompts and independent confirmation pages. If any processing conflicts with laws and regulations, the laws and regulations will prevail, and necessary data processing will only be carried out within the scope permitted by laws and regulations, ensuring that all data-based applications, analyses, and technical implementations are conducted within the scope permitted by laws and regulations.

[0106] In some alternative embodiments, the functions / operations mentioned in the block diagrams may not occur in the order shown in the operation diagrams. For example, depending on the functions / operations involved, two consecutively shown blocks may actually be executed substantially simultaneously, or the blocks may sometimes be executed in reverse order. Furthermore, the embodiments presented and described in the flowcharts of this application are provided by way of example to provide a more comprehensive understanding of the technology. The disclosed methods are not limited to the operations and logic flows presented herein. Alternative embodiments are contemplated in which the order of various operations is changed and sub-operations described as part of a larger operation are executed independently.

[0107] Furthermore, although this application is described in the context of functional modules, it should be understood that, unless otherwise stated, one or more of the functions and / or features may be integrated into a single physical device and / or software module, or one or more functions and / or features may be implemented in a separate physical device or software module. It is also understood that a detailed discussion of the actual implementation of each module is unnecessary for understanding this application. Rather, given the properties, functions, and internal relationships of the various functional modules in the apparatus disclosed herein, the actual implementation of the module will be understood within the scope of ordinary skill of an engineer. Therefore, those skilled in the art can implement the application set forth in the claims using ordinary skill. It is also understood that the specific concepts disclosed are merely illustrative and are not intended to limit the scope of this application, which is determined by the full scope of the appended claims and their equivalents.

[0108] If a function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several programs to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0109] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequential list of executable programs for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, a program execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can retrieve and execute a program from or in conjunction with such a program execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can mean any means that can contain, store, communicate, propagate, or transmit a program for use by or in conjunction with a program execution system, apparatus, or device.

[0110] More specific examples (a non-exhaustive list) of computer-readable media include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Additionally, computer-readable media can even be paper or other suitable media on which programs can be printed, for example, by optically scanning the paper or other media, then editing, interpreting, or, if necessary, processing it in a suitable manner to obtain the program electronically, and then storing it in computer memory.

[0111] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable program execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0112] In the foregoing description of this specification, the reference to terms such as "one embodiment / implementation," "another embodiment / implementation," or "certain embodiments / implementations," etc., indicates that a specific feature, structure, material, or characteristic described in connection with an embodiment or example is included in an embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0113] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

[0114] The above is a detailed description of the preferred embodiments of the present invention. However, the present invention is not limited to the embodiments. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the present invention. All such equivalent modifications or substitutions are included within the scope defined by the claims of the present invention.

Claims

1. A joint licensing and beamforming method for a cellular-free system, characterized in that, The non-cellular system includes a central processing unit, multiple base stations, and multiple users, wherein all base stations are connected to the central processing unit, and the base stations provide cooperative transmission services to the corresponding users according to dynamic authorization relationships. The method includes the following steps: The channel state information of the downlink is obtained at the central processing unit; The channel state information is input into a pre-trained user-base station grant learner, which outputs discrete variables representing the grant relationship between the user and the base station. The discrete variables are used to represent the grant decision and satisfy preset discrete constraints. The channel state information and the discrete variables are input together into a pre-trained beamforming learner, which outputs a beamforming vector that satisfies the base station transmit power constraint. The user-base station grant learner and the beamforming learner constitute an end-to-end joint learning optimization framework for jointly optimizing discrete grant decisions and continuous beamforming vectors.

2. The joint licensing and beamforming method for cellular-free systems according to claim 1, characterized in that, The discrete variable is a binary matrix, where the position of the element with a value of 1 corresponds to the authorized user-base station pair; the user-base station authorization learner outputs the support set of the binary matrix, which consists of the indices of all user-base station pairs with an element value of 1, and is used to characterize the non-zero structure of the discrete variable.

3. The joint licensing and beamforming method for non-cellular systems according to claim 2, characterized in that, The user-base station authorization learner adopts an encoder-decoder neural network architecture, wherein: The encoder is configured to first separate the complex channel state information into real and imaginary parts and perform vectorization processing, and then map it into a channel embedding vector; The decoder is configured to generate index elements in the support set one by one in an autoregressive manner, generating a user-base station pair index as the current action each time, and updating the internal state based on the generated action sequence until a complete support set is generated.

4. The joint licensing and beamforming method for non-cellular systems according to claim 3, characterized in that, When the decoder generates each action, the following operations are performed: Based on the context state of the current decoding step and the channel embedding vector, calculate the conditional probability distribution of all unselected user-base station pairs as candidate actions; Based on the discrete constraints, identify and exclude candidate actions that would lead to constraint violations, and force their corresponding conditional probabilities to zero. The next action is selected from the remaining legal candidate actions based on the modified probability distribution, thereby dynamically constructing an action space that satisfies all discrete constraints during the decoding process.

5. The joint licensing and beamforming method for cellular-free systems according to claim 1, characterized in that, The discrete constraints include at least one of the following: (1) Each user is served by a maximum of N base stations, where N is a positive integer determined based on the system's fronthaul link capacity; (2) Each base station can serve a maximum of M users, where M is a positive integer determined based on the system's fronthaul link capacity; (3) The total number of authorized relationships does not exceed the upper limit set by the system; In any decoding step, if the addition of a candidate user-base station pair would cause any of the discrete constraints to be broken, then the candidate action is determined to be illegal and removed from the set of possible actions.

6. The joint licensing and beamforming method for cellular-free systems according to claim 1, characterized in that, The beamforming learner includes a channel embedding layer, a multi-layer graph message passing update module, and a beam vector output layer connected in sequence. The channel embedding layer is used to decompose complex channel state information into real and imaginary parts and then vectorize it. The graph message passing update module iteratively transmits aggregated information between base station nodes and user nodes based on the user-base station connection topology defined by the discrete variables. The number of iterations is pre-configured according to system performance requirements. After generating the initial beamforming vector, the beam vector output layer performs power normalization processing on the sub-vector corresponding to each base station, so that the transmit power of each base station does not exceed its maximum allowable power threshold.

7. The joint licensing and beamforming method for cellular-free systems according to claim 1, characterized in that, The user-base station grant learner is optimized using a policy gradient-based reinforcement learning algorithm during the offline training phase, and its reward signal is the system weighted sum rate. The beamforming learner employs supervised or unsupervised learning methods during offline training. It maximizes the system and rate objective functions through a stochastic gradient ascent algorithm and is trained collaboratively with the user-base station grant learner through a weighted joint loss function. The weights of the reinforcement learning strategy gradient loss and the beamforming learning loss weights are dynamically adjusted according to system performance requirements to achieve end-to-end performance optimization.

8. A cellless system, characterized in that, The non-cellular system adopts a wireless communication architecture, including a central processing unit, multiple base stations and multiple users, with all base stations connected to the central processing unit; The central processing unit is used to execute the joint licensing and beamforming method for non-cellular systems as described in any one of claims 1 to 7, and is responsible for centralized channel state information acquisition, user-base station dynamic association decision-making, and downlink beamforming vector calculation; Each user is served by multiple base stations, and each base station can serve multiple users simultaneously, forming a many-to-many cooperative transmission relationship.

9. An electronic device, characterized in that, The electronic device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the joint licensing and beamforming method for non-cellular systems as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the joint licensing and beamforming method for non-cellular systems as described in any one of claims 1 to 7.

Images