Wireless resource allocation method and device, computer equipment and readable storage medium

Abstract

The embodiment of the invention provides a wireless resource allocation method and device, computer equipment and a readable storage medium. The method comprises the following steps: acquiring channel state data between a plurality of target base stations and a plurality of target terminals; inputting the channel state data into a model to obtain a corresponding wireless resource allocation result; the first model is obtained by adjusting a pre-trained second model according to target model parameters; the target model parameter is a value of a model parameter variable when a total loss value is minimized, wherein the model parameter variable is obtained after the second model carries out wireless resource allocation on multiple pieces of sample channel state data in a sample null space in combination with a target function; the objective function comprises a parameter optimization function including a function relationship between a total loss variable and a model parameter variable while minimizing the total loss value as an optimization objective, and a constraint function for defining an equality relationship between the first loss and the second loss; in this way, the real-time performance and accuracy of wireless resource allocation can be improved.

Description

Technical Field

[0001] The present application relates to the field of communication technology, and in particular to a wireless resource allocation method, apparatus, computer equipment and readable storage medium. Background Art

[0002] Wireless resource allocation refers to the process of rationally scheduling and optimizing the configuration of limited spectrum, power, time and other resources in wireless communication systems in order to maximize network performance and user experience. In the face of dynamically changing wireless environments and user needs, efficient resource allocation can improve data transmission rates, expand network capacity, and ensure service quality and fairness. It is one of the key technologies for optimizing modern mobile communication networks. With the popularity of mobile devices and the surge in data traffic, efficient wireless resource allocation is crucial to ensuring network stability and improving user experience.

[0003] However, in a complex wireless communication environment, resource allocation faces many challenges. First, spectrum resources are limited and need to be balanced among many users and applications. Second, user needs and network conditions are constantly changing, requiring resource allocation to quickly adapt to these changes. In addition, factors such as network energy consumption, coverage, and service quality also need to be considered.

[0004] In related technologies, in order to optimize resource allocation and improve overall system performance, power control and channel allocation are generally optimized through medium- and long-term network measurement information. Specifically, the network environment can be continuously monitored, and data on user behavior, traffic patterns, channel conditions, etc. can be collected, and the analysis results of network performance can be obtained through these data. Finally, based on these rich measurement data, a more accurate model is built to predict the future network status, thereby achieving more intelligent and dynamic resource management decisions. However, this traditional method has a high computational complexity and is difficult to adapt to the rapidly changing network environment, resulting in low real-time and accuracy of wireless resource allocation. Summary of the invention

[0005] The main purpose of the embodiments of the present application is to propose a wireless resource allocation method, apparatus, computer equipment and readable storage medium, which can improve the real-time and accuracy of wireless resource allocation.

[0006] To achieve the above object, a first aspect of an embodiment of the present application proposes a method for allocating wireless resources, the method comprising:

[0007] Acquiring channel state data between a plurality of target base stations and a plurality of target terminals;

[0008] Inputting the channel state data into a pre-trained first model to obtain wireless resource allocation results of the multiple target base stations to the multiple target terminals;

[0009] The first model is obtained by adjusting the pre-trained second model according to the target model parameters; the target model parameters are the values ​​of the model parameter variables when the second model minimizes the total loss value obtained after allocating wireless resources for multiple sample channel state data in the sample null space in combination with the objective function;

[0010] The sample null space is determined after the second model obtained by adjusting parameters according to the historical stage processes a plurality of historical sample channel state data to obtain an input data matrix, and the objective function includes a parameter optimization function and a constraint function;

[0011] The parameter optimization function includes a functional relationship between a total loss variable and the model parameter variable while minimizing the total loss value as an optimization goal;

[0012] The constraint function is used to limit the equality relationship between the first loss and the second loss; the first loss is based on the intermediate model corresponding to the current value of the model parameter variable, and is calculated after wireless resources are allocated to multiple historical sample channel state data in different historical stages; the second loss is based on the second model corresponding to the historical model parameter, and is calculated after wireless resources are allocated to the multiple historical sample channel state data.

[0013] Accordingly, a second aspect of an embodiment of the present application proposes a wireless resource allocation device, the device comprising:

[0014] An acquisition module, used to acquire channel state data between multiple target base stations and multiple target terminals;

[0015] An input module is used to input the channel state data into a pre-trained first model to obtain the wireless resource allocation results of the multiple target base stations to the multiple target terminals; wherein the first model is obtained by adjusting the pre-trained second model according to the target model parameters; the target model parameters are the values ​​of the model parameter variables when the second model is determined in the sample null space in combination with the objective function to minimize the total loss value after performing wireless resource allocation for multiple sample channel state data; the sample null space is determined after the second model obtained by adjusting the parameters according to the historical stage processes the multiple historical sample channel state data to obtain the input data matrix, and the objective function includes a parameter optimization function and a constraint function; the parameter optimization function includes a functional relationship between the total loss variable and the model parameter variable while minimizing the total loss value as the optimization goal; the constraint function is used to limit the equality relationship between the first loss and the second loss; the first loss is calculated after performing wireless resource allocation for multiple historical sample channel state data at different historical stages based on the intermediate model corresponding to the current value of the model parameter variable; the second loss is calculated after performing wireless resource allocation for the multiple historical sample channel state data based on the second model corresponding to the historical model parameters.

[0016] In some implementations, the wireless resource allocation apparatus further includes a training module, configured to:

[0017] Acquire a sample training set corresponding to the current stage, wherein the sample training set includes a plurality of sample channel state data;

[0018] Taking minimizing the total loss value as the optimization goal, and based on the functional relationship between the total loss variable and the model parameter variable, a parameter optimization function is constructed; wherein the total loss variable is obtained by accumulating multiple current sample losses calculated after wireless resource allocation for multiple sample channel state data based on the intermediate model corresponding to the current value of the model parameter variable;

[0019] After allocating wireless resources to a plurality of historical sample channel state data at different historical stages based on the intermediate model corresponding to the current value of the model parameter variable, a first loss is calculated;

[0020] After allocating wireless resources to the plurality of historical sample channel state data based on a second model corresponding to the historical model parameters, a second loss is calculated;

[0021] constructing a constraint function based on an equality relationship between the first loss and the second loss;

[0022] Constructing an objective function based on the constraint function and the parameter optimization function;

[0023] In combination with the objective function, determining in the sample null space the target model parameters for minimizing the total loss value obtained after the second model performs wireless resource allocation for the plurality of sample channel state data;

[0024] The parameters of the second model are adjusted based on the target model parameters to obtain the first model.

[0025] In some embodiments, the training module is further used to:

[0026] Acquire at least one resource allocation target for wireless resource allocation in the current stage; wherein the resource allocation target includes a sum rate, a communication delay rate, and an allocation balance rate;

[0027] Performing wireless resource allocation on each sample channel state data in turn through the intermediate model corresponding to the value of the current model parameter variable to obtain a first wireless resource allocation result;

[0028] According to the at least one resource allocation target and in combination with the first wireless resource allocation result, calculating at least one sub-loss after wireless resource allocation is performed on each sample channel state data;

[0029] Based on the sum of the at least one sub-loss, a current sample loss corresponding to the value of the current model parameter variable when the intermediate model processes each sample channel state data is obtained.

[0030] In some embodiments, the training module is further used to:

[0031] Processing the corresponding plurality of historical channel state data in each network layer by the second model to obtain an input data matrix corresponding to each network layer;

[0032] In each of the network layers, low-rank approximation processing is performed on the input data matrix to obtain an updated input data matrix;

[0033] Based on the updated input data matrix, the sample null space corresponding to each network layer is determined.

[0034] In some embodiments, the training module is further used to:

[0035] Processing the plurality of historical channel state data according to the historical model parameters corresponding to each network layer through the second model to obtain an input data matrix corresponding to each network layer;

[0036] Repeat the process in the next network layer of the second model, processing the multiple historical channel state data according to the next historical model parameters corresponding to the next network layer, and obtaining the next input data matrix corresponding to the next network layer, until the input data matrix of the last network layer of the intermediate model is obtained.

[0037] In some embodiments, the training module is further used to:

[0038] In each of the network layers, singular value decomposition is performed on the corresponding input data matrix to obtain a left singular vector matrix, a singular value diagonal matrix, and a transpose of a right singular vector matrix of the input data matrix;

[0039] Obtaining a preset extraction value, and extracting the same number of matrix columns as the preset extraction value from the left singular vector matrix to obtain an updated left singular vector matrix;

[0040] Extracting the same number of singular values ​​as the preset extraction values ​​from the singular value diagonal matrix to obtain an updated singular value diagonal matrix;

[0041] Extracting the same number of matrix rows as the preset extraction value from the transpose of the right singular vector matrix to obtain the updated transpose of the right singular vector matrix;

[0042] An updated input data matrix is ​​obtained based on the transpose of the updated left singular vector matrix, the updated singular value diagonal matrix and the updated right singular vector matrix.

[0043] In some implementations, the wireless resource allocation apparatus further includes a disabling module, configured to:

[0044] Obtaining multiple scaling parameters corresponding to multiple network layers of the second model;

[0045] Acquire multiple scaling parameter absolute values ​​corresponding to the multiple scaling parameters, and determine multiple target scaling parameters based on the order of magnitude of the multiple scaling parameter absolute values;

[0046] A plurality of target neurons corresponding to the plurality of target scaling parameters are determined, and the plurality of target neurons are temporarily disabled in the plurality of network layers.

[0047] Correspondingly, the third aspect of the embodiments of the present application proposes a computer device, which includes a memory and a processor, the memory stores a computer program, and the processor implements the wireless resource allocation method described in any one of the embodiments of the first aspect of the present application when executing the computer program.

[0048] Correspondingly, the fourth aspect of the embodiments of the present application proposes a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, it implements the wireless resource allocation method described in any one of the embodiments of the second aspect of the present application.

[0049] In an embodiment of the present application, channel state data between multiple target base stations and multiple target terminals are obtained; the channel state data is input into a pre-trained first model to obtain wireless resource allocation results of multiple target base stations to multiple target terminals; wherein the first model is obtained by adjusting the pre-trained second model according to the target model parameters; the target model parameters are the values ​​of model parameter variables when the second model is determined in the sample null space in combination with the objective function to minimize the total loss value after wireless resource allocation for multiple sample channel state data; the sample null space is determined after the second model obtained by adjusting the parameters according to the historical stage processes the input data matrix of multiple historical sample channel state data, and the objective function includes a parameter optimization function and a constraint function; the parameter optimization function includes a functional relationship between the total loss variable and the model parameter variable while minimizing the total loss value as the optimization goal; the constraint function is used to limit the equality relationship between the first loss and the second loss; the first loss is calculated after wireless resource allocation for multiple historical sample channel state data of different historical stages based on the intermediate model corresponding to the current value of the model parameter variable; the second loss is calculated after wireless resource allocation for the multiple historical sample channel state data based on the second model corresponding to the historical model parameters. In this way, the periodicity and repetitiveness of channel distribution can be fully utilized, and according to the multiple sample channel state data of the current stage, the second model that has been trained and adjusted in advance and has good performance can be quickly trained, so that the model does not need to frequently learn the adapted distribution, thereby improving the efficiency and adaptability of model training. On this basis, after the second model obtained by adjusting the parameters according to the historical stage processes the multiple historical channel state data to obtain the input data matrix, the sample zero space is determined, and the optimization is performed in the sample zero space in the current stage, so that the model can smoothly transition between different stages, avoid the training problem of catastrophic forgetting of historically learned knowledge, and enable the model to efficiently adapt to the dynamically changing wireless environment while retaining the memory of historically learned knowledge, thereby improving the accuracy of resource allocation. At the same time, by constructing an objective function to optimize the model parameters, the model can take into account the long-term and short-term goals of resource allocation during the optimization process, making resource allocation more accurate and better meeting the needs of network performance and user experience. Compared with traditional methods, the method of the present application does not rely on complex numerical optimization techniques, but uses the powerful learning ability and parallel computing characteristics of deep learning, which enables the model to quickly and accurately allocate channel resources, thereby improving the real-time and accuracy of resource allocation. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 It is a schematic diagram of the architecture of the wireless resource allocation system provided in an embodiment of the present application;

[0051] Figure 2 is a flow chart of a wireless resource allocation method provided in an embodiment of the present application;

[0052] Figure 3 is a model training flow chart provided in an embodiment of the present application;

[0053] Figure 4 is a diagram of channel distribution changes at different stages provided by an embodiment of the present application;

[0054] Figure 5 is a performance comparison result diagram of the method of the present application and other learning methods provided in the embodiments of the present application;

[0055] Figure 6 This is a graph showing the training time comparison between the method of the present application and other learning methods provided in the embodiments of the present application;

[0056] Figure 7 It is a functional module diagram of a wireless resource allocation device provided in an embodiment of the present application;

[0057] Figure 8 It is a schematic diagram of the hardware structure of the computer device provided in the embodiment of the present application. DETAILED DESCRIPTION

[0058] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0059] It should be noted that, although the functional modules are divided in the device schematic diagram and the logical order is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than the module division in the device or the order in the flowchart. The terms "first", "second", etc. in the specification, claims and the above drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence.

[0060] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art to which this application belongs. The terms used herein are only for the purpose of describing the embodiments of this application and are not intended to limit this application.

[0061] Wireless resource allocation refers to the process of rationally scheduling and optimizing the configuration of limited spectrum, power, time and other resources in wireless communication systems in order to maximize network performance and user experience. In the face of dynamically changing wireless environments and user needs, efficient resource allocation can improve data transmission rates, expand network capacity, and ensure service quality and fairness. It is one of the key technologies for optimizing modern mobile communication networks. With the popularity of mobile devices and the surge in data traffic, efficient wireless resource allocation is crucial to ensuring network stability and improving user experience.

[0062] However, in a complex wireless communication environment, resource allocation faces many challenges. First, spectrum resources are limited and need to be balanced among many users and applications. Second, user needs and network conditions are constantly changing, requiring resource allocation to quickly adapt to these changes. In addition, factors such as network energy consumption, coverage, and service quality also need to be considered.

[0063] In related technologies, in order to optimize resource allocation and improve overall system performance, power control and channel allocation are generally optimized through medium- and long-term network measurement information. Specifically, the network environment can be continuously monitored, and data on user behavior, traffic patterns, channel conditions, etc. can be collected, and the analysis results of network performance can be obtained through these data. Finally, based on these rich measurement data, a more accurate model is built to predict the future network status, thereby achieving more intelligent and dynamic resource management decisions. However, this traditional method has a high computational complexity and is difficult to adapt to the rapidly changing network environment, resulting in low real-time and accuracy of wireless resource allocation.

[0064] Based on this, the embodiments of the present application provide a wireless resource allocation method, apparatus, computer equipment and readable storage medium, which can improve the real-time and accuracy of resource allocation.

[0065] The wireless resource allocation method, apparatus, computer device and readable storage medium provided in the embodiments of the present application are specifically described through the following embodiments. First, the wireless resource allocation system in the embodiments of the present application is described.

[0066] Please refer to Figure 1 In some implementations, an embodiment of the present application provides a wireless resource allocation system, including a terminal 11 and a server 12.

[0067] Exemplarily, the terminal 11 may be a smart phone, a tablet computer, an Internet of Things device, etc., and the server 12 may be a base station, an edge computing server, a cloud computing center server, etc.

[0068] In some implementations, the terminal 11 may include a communication module, a data processing module, and a resource allocation request module. Specifically, the communication module may be responsible for wireless communication with the base station, receiving downlink signals from the base station, such as channel state information (CSI), resource allocation instructions, etc.; and sending uplink signals, such as its own business demand information, feedback information, etc. For example, in a 5G communication system, the communication module of the terminal 11 supports the New Radio (NR) technology, which can achieve high-speed, low-latency data transmission.

[0069] Furthermore, the data processing module can pre-process the received channel state data, such as channel estimation, noise elimination, etc., to improve the accuracy and reliability of the data. At the same time, the data processing module can generate corresponding resource demand information, such as the required bandwidth, delay and other parameters, according to its own business needs, such as video playback, file downloading, etc.

[0070] Furthermore, the resource allocation request module may integrate the processed channel state data and its own resource demand information to form a resource allocation request, and send it to the server 12 through the communication module.

[0071] In some implementations, the server 12 may be a computer device that may include a data receiving and parsing module, a resource allocation model (such as a first model, a second model, etc.), a resource allocation decision module, an instruction generation and sending module, etc.

[0072] Specifically, the data receiving and parsing module can receive resource allocation requests from multiple terminals 11, and parse the channel state data and resource demand information in the request to extract key parameters. For example, the channel quality indicator (CQI) of each terminal 11, the required service type (such as voice, data, etc.), and other information are parsed to provide a basis for subsequent resource allocation.

[0073] Exemplarily, the resource allocation decision module can be used to allocate wireless resources according to the parsed channel state data and resource demand information. The model can be a deep learning model based on continuous learning, which can quickly adapt to dynamically changing wireless environments by optimizing in the sample null space to avoid catastrophic forgetting, while retaining the memory of historically learned knowledge to improve the accuracy and efficiency of resource allocation.

[0074] Furthermore, the resource allocation decision module can generate specific resource allocation decisions based on the output results of the model, such as the spectrum resources, power resources, time resources, etc. allocated to each terminal 11. Afterwards, the resource allocation decision can be converted into specific instructions, such as the frequency band, power level, time slice, etc. of resource allocation, through the instruction generation and sending module, and sent to the corresponding terminal 11 through the communication module. At the same time, some control instructions can also be sent as needed, such as adjusting the transmission power of the terminal 11, switching the frequency band, etc., to optimize the network performance.

[0075] The wireless resource allocation system can achieve efficient and accurate wireless resource allocation, improving the overall performance and user experience of the wireless communication network.

[0076] The wireless resource allocation method in the embodiment of the present application can be illustrated by the following embodiment.

[0077] It should be noted that in each specific implementation of the present application, when it comes to the need to perform relevant processing based on data related to user identity or characteristics such as user information, user behavior data, user historical data, and user location information, the user's permission or consent will be obtained first. Moreover, the collection, use, and processing of these data will comply with relevant laws, regulations, and standards. In addition, when the embodiment of the present application needs to obtain the user's sensitive personal information, the user's separate permission or separate consent will be obtained through a pop-up window or by jumping to a confirmation page. After clearly obtaining the user's separate permission or separate consent, the necessary user-related data for enabling the normal operation of the embodiment of the present application will be obtained.

[0078] In the embodiment of the present application, the wireless resource allocation device will be described from the perspective of the wireless resource allocation device, which can be integrated into a computer device. Figure 2 , Figure 2 This is a flowchart of the steps of the wireless resource allocation method provided in an embodiment of the present application. In the embodiment of the present application, the wireless resource allocation device is specifically integrated in a terminal or a server as an example. When the processor on the terminal or the server executes the program instructions corresponding to the wireless resource allocation method, the specific process is as follows:

[0079] Step 101: Acquire channel state data between multiple target base stations and multiple target terminals.

[0080] In some implementations, in order to ensure that the model can learn and predict based on the latest and most realistic channel state data, channel state data between a target base station and multiple target terminals may be obtained to achieve efficient, intelligent, and adaptable wireless resource allocation.

[0081] The target base station can be a fixed device in the wireless communication network, which is responsible for sending and receiving wireless signals with user terminals within a certain coverage area. The target base station is the infrastructure of the wireless network, usually installed at a high place (such as a rooftop, iron tower, etc.) to ensure a wide signal coverage range.

[0082] Among them, the target terminal can be a device directly used by the user to access the wireless communication network for voice calls, data transmission, Internet access and other operations. The target terminal can communicate with the base station through a built-in wireless communication module (such as Wi-Fi, 4G / 5G, etc.) to send and receive wireless signals to the target base station.

[0083] The channel state data may be optimization variables for resource allocation and channel data at the current stage, so that the first model can adjust the resource allocation scheme according to the channel characteristics at the current stage.

[0084] In some implementations, the channel state data includes, in addition to the channel data, optimization variables for resource allocation, such as spectrum, power, time allocation, and other parameters, which are the targets of optimization by the first model. The specific data categories included in the channel state data can be adjusted according to actual conditions, and the embodiments of the present application do not impose specific restrictions on this.

[0085] Exemplarily, the optimization variables of resource allocation can be determined by the target base station according to the design and operation strategy of the network, such as the allocation of spectrum resources (frequency selection), power control (transmit power adjustment), time allocation (time slot allocation), etc. Furthermore, the target terminal (such as a mobile phone, tablet, etc.) can measure the channel quality between the base station and the target terminal through the built-in wireless communication module. The measurement results include signal-to-noise ratio, channel gain, signal strength, etc., and the measured channel status data is fed back to the target base station through the uplink. The system can centrally collect and process the channel status data from different target base stations and target terminals to provide decision support for resource allocation.

[0086] By acquiring the channel status data, the channel distribution of the entire network can be understood more accurately, thereby facilitating subsequent wireless resource allocation through the first model to optimize network performance and user experience.

[0087] Step 102, inputting the channel state data into the pre-trained first model to obtain the wireless resource allocation results of multiple target base stations to multiple target terminals;

[0088] The first model is obtained by adjusting the pre-trained second model according to the target model parameters; the target model parameters are the values ​​of the model parameter variables when the total loss value is minimized after the second model performs wireless resource allocation for multiple sample channel state data in the sample null space in combination with the objective function;

[0089] The sample null space is determined by adjusting the parameters of the second model according to the historical stage, and then processing the multiple historical sample channel state data to obtain the input data matrix, and the objective function includes a parameter optimization function and a constraint function;

[0090] The parameter optimization function takes minimizing the total loss value as the optimization goal and contains the functional relationship between the total loss variable and the model parameter variable;

[0091] The constraint function is used to limit the equality relationship between the first loss and the second loss; the first loss is calculated based on the intermediate model corresponding to the value of the current model parameter variable, and wireless resources are allocated to multiple historical sample channel state data at different historical stages; the second loss is calculated based on the second model corresponding to the historical model parameter, and wireless resources are allocated to multiple historical sample channel state data.

[0092] In some embodiments, in order to enable the first model to make intelligent decisions based on real-time channel conditions, channel state data can be input into the pre-trained first model to allocate wireless resources, thereby improving the overall performance of the network and user experience.

[0093] Among them, the first model can be a model that is trained by the second model for the channel distribution of the current stage and has the ability to well distribute channel state data of different historical stages and the current stage.

[0094] The wireless resource allocation result may be a result output by the first model, and is used to indicate a specific allocation amount of wireless resources (such as frequency, power, and time) allocated between multiple target base stations and multiple target terminals.

[0095] Among them, the second model can be the basic model of the first model. After the parameters of the second model are optimized in the historical stage, it has the ability to efficiently and accurately distribute the historical channel state data of all historical stages. By training the second model with the sample channel state data of the current stage, the first model that adapts to the channel state of the current stage can be obtained.

[0096] Among them, the target model parameters can be the model parameters obtained after training the sample channel state data of the current stage during the continuous learning process. The target model parameters can enable the first model to achieve the optimal resource allocation effect when processing the data of the current stage, while retaining the memory of the historical stage data.

[0097] Among them, the objective function can be used to guide the optimization process of the model, including parameter optimization function and constraint function.

[0098] Among them, the sample null space can be the null space corresponding to the weight matrix of each network layer of the second model after the second model is trained. The update operation of the model parameters in the current stage is limited to the sample null space, and the second model can only be adjusted in the direction that will not change the prediction results of historical data, thereby effectively avoiding the catastrophic forgetting problem, that is, the model will not lose the memory of the previous stage data when learning the data of the new stage.

[0099] Among them, the sample channel state data can be used to train the second model. Both the sample channel state data and the channel state data correspond to the current stage, that is, the channel distribution of the two is the same, so that after the second model adjusts the model parameters based on the sample channel state data, it has the ability to efficiently and accurately allocate wireless resources according to the channel state data.

[0100] Among them, the total loss value can be an indicator for users to measure the overall performance of the model during the entire training process, and is used to quantify the allocation capability of the model for wireless resource allocation.

[0101] Among them, the value of the model parameter variable can be the specific numerical value of the adjustable parameter in the intermediate model. During the model training process, the value of the model parameter variable can be adjusted in real time to enable the intermediate model to achieve better training effect.

[0102] The historical channel state data may be data input into the second model for wireless resource allocation after the second model is trained in the previous historical stage of the current stage.

[0103] The input data matrix may be a matrix composed of all input data in the forward propagation process in each network layer of the second model, and each network layer corresponds to an input data matrix.

[0104] The parameter optimization function may be an optimization problem, the goal of which is to continuously optimize the values ​​of the model parameter variables under the constraints of the constraint function and to minimize the total loss value corresponding to all sample channel state data as much as possible.

[0105] Among them, the constraint function can be a constraint condition defined in the process of optimizing the parameter optimization function, which is used to limit the first loss obtained by the intermediate model processing the sample channel state data of each historical stage when training the second model in the current stage to be equal to the second loss obtained by the second model processing the historical sample channel state data corresponding to each historical stage, so as to ensure that the intermediate model can still accurately and efficiently allocate wireless resources according to the historical sample channel state data while learning new data, and will not forget old knowledge.

[0106] The total loss variable may be a measure representing the loss of the entire model, and the total loss variable may be obtained by adding together multiple current sample losses obtained by predicting multiple sample channel state data by the intermediate model.

[0107] Among them, the model parameter variables can be the weights and biases in the intermediate model, and the model parameter variables are continuously adjusted during the training process to minimize the loss function while satisfying the constraint function.

[0108] Among them, the first loss can be the intermediate model corresponding to the value of the model parameter variable in the current training state, and the loss after wireless resource allocation for the historical sample channel state data corresponding to each historical stage.

[0109] Among them, the second loss can be the loss after wireless resource allocation for historical sample channel state data corresponding to each historical stage based on the second model trained in the previous historical stage of the current stage.

[0110] Among them, the intermediate model can refer to a model version under the configuration of the current model parameter variables, which is located in the conversion process from the second model to the first model. The intermediate model does not refer to a fixed model. As long as the model parameter variables of the second model are adjusted to obtain the first model, the model involved can be called an intermediate model.

[0111] The historical sample channel state data may be sample channel state data for training the model in the historical stage corresponding to the second model and the historical stage before the second model. Each historical stage corresponds to a different model and different historical sample channel state data.

[0112] The historical model parameters may be parameter configurations of the second model in the corresponding historical stage.

[0113] For example, assume that there is a wireless communication system, which includes three target base stations (base station 1, base station 2, base station 3) and six target terminals (terminal A, terminal B, terminal C, terminal D, terminal E, terminal F). These base stations and terminals are distributed in a certain geographical area, and each base station is responsible for providing wireless communication services to the terminals within a certain range.

[0114] In some implementations, the channel state data includes not only the channel gain, noise level, interference, etc., but also optimization variables for resource allocation, such as spectrum, power, time allocation, etc. These data can be obtained through channel estimation and network management systems.

[0115] Exemplarily, if the channel state data includes the following information:

[0116] The channel gain from base station 1 to terminal A is 0.8, the noise level is 0.1, and the interference level is 0.2; the channel gain from base station 1 to terminal B is 0.7, the noise level is 0.2, and the interference level is 0.1; the channel gain from base station 2 to terminal C is 0.9, the noise level is 0.1, and the interference level is 0.3; the channel gain from base station 2 to terminal D is 0.6, the noise level is 0.3, and the interference level is 0.2; the channel gain from base station 3 to terminal E is 0.7, the noise level is 0.2, and the interference level is 0.1; the channel gain from base station 3 to terminal F is 0.8, the noise level is 0.1, and the interference level is 0.2.

[0117] The channel state data input into the first model also includes resource allocation optimization variables such as spectrum resources, power resources, and time resources. Then, the above channel state data and resource allocation optimization variables can be integrated into an input matrix, each row represents the channel state data and resource allocation optimization variables from a base station to a terminal, and each column represents a feature. The integrated channel state data is input into the pre-trained first model to obtain the wireless resource allocation result. Assume that the wireless resource allocation result is a matrix, in which each row represents the resource allocation of a target base station to a target terminal, and each column represents a resource allocation variable (such as spectrum allocation, power allocation, time allocation). Exemplarily, according to the output matrix corresponding to the wireless resource allocation result, the following specific allocation results can be obtained:

[0118] Base station 1 allocates 3MHz of spectrum resources, 0.4W of power resources and 3ms of time resources to terminal A; base station 1 allocates 2MHz of spectrum resources, 0.3W of power resources and 2ms of time resources to terminal B; base station 2 allocates 4MHz of spectrum resources, 0.5W of power resources and 4ms of time resources to terminal C; base station 2 allocates 3MHz of spectrum resources, 0.2W of power resources and 3ms of time resources to terminal D; base station 3 allocates 2MHz of spectrum resources, 0.4W of power resources and 2ms of time resources to terminal E; base station 3 allocates 3MHz of spectrum resources, 0.3W of power resources and 3ms of time resources to terminal F.

[0119] In this way, the pre-trained first model can quickly and accurately allocate wireless resources to multiple target terminals for multiple target base stations based on the channel characteristics and resource allocation optimization variables at the current stage, thereby optimizing network performance and user experience.

[0120] Furthermore, the first model can be obtained by training a pre-trained second model, and the second model has been trained on some historical channel state data, and can efficiently and accurately allocate wireless resources to all historical sample channel state data in the historical stage. In order to enable the first model to still have the ability to efficiently and accurately allocate wireless resources to all historical sample channel state data corresponding to all historical stages, the second model can be used to process multiple historical channel state data in the previous historical stage of the current stage to obtain an input data matrix, and based on the input data matrix, the sample null space of the previous historical stage is determined. By limiting the parameter update during the training process of the current stage in the sample null space of the previous stage, catastrophic forgetting of the trained first model can be effectively avoided.

[0121] Exemplarily, the objective function can be used to find a set of parameters W at the current stage t, so that the sample training set D at the current stage t The total loss value on the historical sample training set D is minimized, while satisfying the requirement for all historical stages q q The historical sample channel state data x in q , the current intermediate model processes the historical sample channel state data x of each historical stage q q The first loss l(x q ,W), and the second model processes x q The second loss l(x q ,W t-1 ) are equal. For example, the objective function is of the following form:

[0122]

[0123] Among them, W t represents the parameters of the first model after the second model is trained to obtain the first model; l(x,W) is the loss value obtained by the intermediate model corresponding to the value of the model parameter variable in the current stage after processing the channel state data of each sample in the sample training set, l(x q ,W) is the first loss obtained by the intermediate model after processing the channel state data of each historical sample in the historical sample training set, l(x q ,W t-1 ) is the second loss obtained after the second model processes each historical sample channel state data in the historical sample training set; D t represents the sample training set of the current stage t, x q represents the historical sample channel state data in the historical stage q, W t-1 is the model parameter of the second model, D qRepresents the historical sample training set of the historical stage, the current stage is stage t, the second model is stage t-1, the model used to train the second model is stage t-2, and so on.

[0124] In some embodiments, the parameter optimization function is used to find a set of model parameters W at the current stage t so as to minimize the total loss value on the sample training set at the current stage. Exemplarily, the parameter optimization function may be in the form of:

[0125]

[0126] in, Represents the sample training set D for the current stage t t The sum of multiple losses corresponding to all sample channel state data in , which means the total loss value variable, argmin W Indicates the search for The model parameter W reaches its minimum value.

[0127] In some embodiments, the constraint function is used to ensure that in the process of training the second model, the intermediate model corresponding to the model parameter variable adjusted in real time, when processing the historical sample channel state data of all historical stages q, obtains a first loss equal to the second loss of the second model when processing the historical sample channel state data of all historical stages q, so as to avoid the intermediate model from causing a decrease in the performance of historical data when optimizing the current stage data. Exemplarily, the form of the constraint function can be as follows:

[0128]

[0129] Among them, l(x q ,W) is the first loss obtained by the intermediate model after processing the channel state data of each historical sample in the historical sample training set, l(x q ,W t-1 ) is the second loss obtained after the second model processes each historical sample channel state data in the historical sample training set.

[0130] In some embodiments, the first loss may be less than or equal to the second loss to ensure that the performance of the intermediate model and the first model on historical data is at least equivalent to that of the second model. That is:

[0131]

[0132] By training the sample training set of the current stage with the second model obtained through the training in the previous historical stage of the current stage, the periodicity and repetitive characteristics of the channel distribution can be fully utilized to avoid the model from frequently re-learning the distribution that has been adapted, and at the same time, the model can quickly adapt to the channel distribution of the current stage. Specifically, as a basic model, the second model has been parameter optimized in the historical stage and can process the historical channel state data efficiently and accurately. In order to adapt to the channel distribution of the current stage, the second model is trained using the sample channel state data of the current stage to obtain the first model, so that it can achieve the optimal resource allocation effect when processing the data of the current stage, while retaining the memory of the historical stage data.

[0133] In this process, the objective function plays a key role, which includes the parameter optimization function and the constraint function. The parameter optimization function is responsible for minimizing the total loss value of the current stage when the intermediate model processes the sample channel state data, while the constraint function ensures that the performance of the intermediate model when processing the historical sample channel state data is equivalent to or better than the second model, that is, the first loss is equal to the second loss, or the first loss is less than the second loss. By limiting the parameter update of the current stage to the sample null space, the intermediate model is only allowed to adjust in a direction that will not change the prediction results of the historical data, thereby effectively avoiding catastrophic forgetting.

[0134] Furthermore, in the specific implementation, the input data integrating the channel state data and the resource allocation optimization variables is input into the pre-trained first model, and the first model outputs the wireless resource allocation result, indicating how to allocate wireless resources between multiple target base stations and multiple target terminals. In this way, the first model can quickly and accurately allocate wireless resources according to the channel characteristics and resource allocation optimization variables at the current stage, thereby optimizing network performance and user experience.

[0135] Please refer to Figure 3 In some implementations, in order to enable the intermediate model to keep the memory of old data while continuously learning new data, the second model can be trained based on the sample training set corresponding to the current stage, thereby achieving effective resource allocation in a dynamically changing wireless environment. For example, the first model can be trained in the following way:

[0136] Step 201, obtaining a sample training set corresponding to the current stage, wherein the sample training set includes a plurality of sample channel state data;

[0137] Step 202, taking minimization of the total loss value as the optimization goal, and constructing a parameter optimization function based on the functional relationship between the total loss variable and the model parameter variable; wherein the total loss variable is obtained by accumulating multiple current sample losses calculated after wireless resource allocation for multiple sample channel state data based on the intermediate model corresponding to the value of the current model parameter variable;

[0138] Step 203, based on the intermediate model corresponding to the value of the current model parameter variable, after performing wireless resource allocation on a plurality of historical sample channel state data at different historical stages, a first loss is calculated;

[0139] Step 204, after allocating wireless resources to a plurality of historical sample channel state data based on a second model corresponding to the historical model parameters, a second loss is calculated;

[0140] Step 205, constructing a constraint function based on the equality relationship between the first loss and the second loss;

[0141] Step 206, constructing an objective function based on the constraint function and the parameter optimization function;

[0142] Step 207, combining the objective function, determining in the sample null space the target model parameters for minimizing the total loss value obtained after the second model performs wireless resource allocation for the plurality of sample channel state data;

[0143] Step 208: Adjust the parameters of the second model based on the target model parameters to obtain the first model.

[0144] The current stage may be a specific period in the multi-stage training process in which the second model is being trained to adapt to new channel states. In the current stage, a new sample training set is collected for adjusting model parameters to optimize wireless resource allocation, and the sample channel state data used to train the second model and the channel state data later input into the first model are both in the current stage.

[0145] The sample training set is a set of channel data used to guide the second model to learn to effectively allocate wireless resources at the current stage. At the current stage, the sample training set includes a plurality of sample channel state data to reflect the characteristics of the wireless communication environment at the current stage, such as channel gain, noise level, interference level, etc.

[0146] The current sample loss may be a loss value calculated for a single sample channel state data during the model training process.

[0147] Please refer to Figure 4For example, when the historical stage is stage 1, channel distribution A can correspond to data set D1, and D1 is used to train the model. After the channel distribution changes, the model needs to be retrained according to the data set of the corresponding stage. For example, when the channel distribution A of stage 1 changes to channel distribution B in the current stage (stage t), the data set D corresponding to stage t needs to be used. t The second model is trained to obtain the first model. When the channel distribution B changes to the channel distribution C, it is necessary to obtain the data set D corresponding to the channel distribution C. T The first model is trained, and so on, which are not listed one by one in the embodiments of the present application.

[0148] Exemplarily, each sample channel state data can be input into the second model (after the model parameters are adjusted during the training process, here is the intermediate model), and the allocated sum rate is calculated according to the first wireless resource allocation result output by the second model as the current sample loss. Specifically, the current sample loss l(x,W) is in the following form:

[0149] l(x,W)=-SumRate(x,W);

[0150] Wherein, x represents the sample channel state data, and W represents the value of the current model parameter variable.

[0151] In some implementations, after the model corresponding to the value of the same model parameter processes all sample channel state data included in the sample training set, all current sample losses obtained are added together to obtain a total loss value (total loss variable). For example, if the sample training set contains 5 sample channel state data, then the 5 current sample losses after the model processes the 5 sample channel state data can be added together to obtain the current total loss value.

[0152] Furthermore, the parameter optimization function can be in the form of:

[0153]

[0154] in, Represents the sample training set D for the current stage t t The sum of multiple losses corresponding to all sample channel state data in , which means the total loss value variable, argmin W Indicates the search for The model parameter W reaches its minimum value.

[0155] Furthermore, the corresponding first loss can be obtained after the historical sample channel state data contained in the historical sample training set corresponding to each historical stage is processed by the intermediate model. At the same time, the historical sample channel state data contained in the historical sample training set corresponding to each historical stage is processed by the second model to obtain the corresponding second loss. The constraint function can constrain the first loss to be less than or equal to the second loss. It should be noted that when the first loss of the intermediate model and the second loss of the second model establish an equal relationship, the first loss and the second loss are both calculated based on the same historical sample channel state data. For example, the first loss is the loss corresponding to the intermediate model processing the historical sample channel state data 1, and the second loss is the loss corresponding to the second model processing the historical sample channel state data 1. This ensures that during the optimization process, the performance of the intermediate model on the historical data is at least equivalent to that of the second model, thereby effectively preventing the forgetting of old knowledge.

[0156] In some implementations, the first loss, the second loss, and the current sample loss all correspond to the same loss function formula. For example, the first loss, the second loss, and the current sample loss are all calculated by the sum rate after the wireless resource allocation, etc., to maintain the consistency of the model parameter evaluation and ensure that the performance comparison of the model is accurate. Since the form of the loss function has been expanded when introducing the current sample loss in the previous text, please refer to the above text for details, and will not be listed here one by one.

[0157] In some embodiments, the constraint function is used to ensure that in the process of training the second model, the intermediate model corresponding to the model parameter variable adjusted in real time, when processing the historical sample channel state data of all historical stages q, obtains a first loss equal to the second loss of the second model when processing the historical sample channel state data of all historical stages q, so as to avoid the intermediate model from causing a decrease in the performance of historical data when optimizing the current stage data. Exemplarily, the form of the constraint function can be as follows:

[0158]

[0159] Among them, l(x q ,W) is the first loss obtained by the intermediate model after processing the channel state data of each historical sample in the historical sample training set, l(x q ,W t-1 ) is the second loss obtained after the second model processes each historical sample channel state data in the historical sample training set.

[0160] In some implementations, based on the constraint function and the parameter optimization function, an objective function may be constructed:

[0161] For example, the objective function is in the following form:

[0162]

[0163] Among them, W t represents the parameters of the first model after the second model is trained to obtain the first model; l(x,W) is the loss value obtained by the intermediate model corresponding to the value of the model parameter variable in the current stage after processing the channel state data of each sample in the sample training set, l(x q ,W) is the first loss obtained by the intermediate model after processing the channel state data of each historical sample in the historical sample training set, l(x q ,W t-1 ) is the second loss obtained after the second model processes each historical sample channel state data in the historical sample training set; D t represents the sample training set of the current stage t, x q represents the historical sample channel state data in the historical stage q, W t-1 is the model parameter of the second model, D q Represents the historical sample training set of the historical stage, the current stage is stage t, the second model is stage t-1, the model used to train the second model is stage t-2, and so on.

[0164] Specifically, the objective function can be used to determine a set of model parameters W in the sample null space corresponding to each network layer at the current stage t, so that the sample training set D at the current stage t The total loss value on the historical sample training set D is minimized, while satisfying the requirement for all historical stages q q The historical sample channel state data x in q , the current intermediate model processes the historical sample channel state data x of each historical stage q q The first loss l(x q ,W), and the second model processes x q The second loss l(x q ,W t-1 ) to ensure that the performance of historical data remains unchanged while minimizing the total loss value of the current stage.

[0165] In some implementations, in the linear layer of the intermediate model, there may be the following computational expression:

[0166] WX=Y;

[0167] Where W is the value of the model parameter variable in the current stage, X is the input data matrix, and Y is the output result of the intermediate model. To ensure that the learning of the sample training set corresponding to the channel state in the current stage by the intermediate model does not affect the performance of the previous stage, the model parameter update ΔW can be restricted to the sample null space of the data X in the previous historical stage of the current stage, that is, it satisfies:

[0168] ΔWX=0,(W+ΔW)X=WX;

[0169] By updating the model parameters of the current stage in the sample null space, the intermediate model will not change the prediction results of historical data when training new data, thereby effectively avoiding catastrophic forgetting.

[0170] Furthermore, when calculating the sample null space corresponding to each network layer, after each training phase, the neural network model optimized in this phase can be used to train all data samples (number n) in the current phase. t ) performs a complete forward propagation. During the forward propagation process, the input data matrix of each layer is recorded. Taking the network layer as the lth layer as an example, the input data matrix is ​​a dimension of O l-1 ×n t The matrix, where O l-1 is the output dimension of the neuron in the l-1th layer, that is, the number of features output by the l-1th layer, n t is the number of sample channel state data at the current stage.

[0171] In some embodiments, after each training phase (e.g., after all data in the current phase have been trained), the input data matrix X of each network layer in the model can be calculated and saved as the data of the previous phase. The input data matrix will be used to limit the parameter update of each network layer in the next phase of training so that it remains in the null space of X. In this way, effective memory of historical information can be achieved between various training phases, while providing a more reliable basis for learning data in the current phase.

[0172] Furthermore, after optimizing and obtaining the target model parameters (that is, the values ​​of the model parameter variables that satisfy the first loss equal to the second loss and the minimum total loss value when the sample null space parameters are updated), the parameters of the second model can be adjusted based on the target parameters to obtain a first model that can better adapt to the current stage data while maintaining memory of historical data. In this way, continuous learning and optimization of wireless resource allocation can be achieved in the ever-changing channel state.

[0173] Through the above methods, not only the model's learning ability and speed of the channel data in the current stage are improved, but also its memory of historical data is enhanced, so that the system can maintain robustness and efficiency in complex and changeable wireless communication environments, significantly improving the overall performance and service quality of the system.

[0174] In some embodiments, the method may further include:

[0175] (A.3.1) obtaining at least one resource allocation target for wireless resource allocation in the current stage; wherein the resource allocation target includes a sum rate, a communication delay rate, and an allocation balance rate;

[0176] (A.3.2) performing wireless resource allocation on each sample channel state data in turn using the intermediate model corresponding to the value of the current model parameter variable to obtain a first wireless resource allocation result;

[0177] (A.3.3) calculating at least one sub-loss after wireless resource allocation for each sample channel state data according to at least one resource allocation target and in combination with the first wireless resource allocation result;

[0178] (A.3.4) Based on the sum of at least one sub-loss, the current sample loss when the intermediate model corresponding to the value of the current model parameter variable processes each sample channel state data is obtained.

[0179] The resource allocation target may refer to a specific performance indicator or standard that is desired to be optimized during the wireless resource allocation process. These targets may include but are not limited to system capacity (ie, total throughput), communication delay rate, allocation balance rate, and the like.

[0180] The sum rate may be the sum of the total transmission rates that all target terminals can obtain in a given wireless communication network.

[0181] The communication delay rate may be the time required for a data packet to be sent and received.

[0182] The allocation balance rate may be a fair allocation of resources among different target terminals. A high allocation balance rate means that resources are allocated more evenly and no target terminal is over-prioritized or neglected.

[0183] The first wireless resource allocation result may be a resource allocation scheme calculated by the current intermediate model for each target terminal according to the input sample channel state data. For example, the intermediate model may allocate a specific frequency bandwidth, power level and transmission time to each target terminal.

[0184] The current sample loss may be a loss value calculated based on the resource allocation result of the intermediate model for a single sample channel state data and at least one resource allocation target. For example, if the goal is to maximize the sum rate, the current sample loss may be a negative value of the sum rate value. The current sample losses corresponding to all sample channel state data of the sample training set are added together to obtain a total loss value.

[0185] In some implementations, taking the resource allocation target as the sum rate as an example, each sample channel state data is input into the intermediate model, and the allocated sum rate is calculated according to the first wireless resource allocation result output by the intermediate model as the current sample loss. Specifically, the current sample loss l(x, W) corresponding to any sample channel state data is in the form of:

[0186] l(x,W)=-SumRate(x,W);

[0187] Wherein, x represents the sample channel state data, and W represents the value of the current model parameter variable.

[0188] In some implementations, there may be multiple resource allocation targets at the current stage. For example, in addition to the sum rate, the current sample loss may also be calculated by the communication delay rate or the allocation balance rate. Further, the current sample loss may be obtained by adding the sub-loss corresponding to the sum rate, the sub-loss corresponding to the communication delay rate, and the sub-loss corresponding to the allocation balance rate. The specific form of calculating the current sample loss is flexibly set according to the resource allocation target, and the embodiments of the present application do not impose specific restrictions on this.

[0189] Through the above methods, it can not only ensure that the model will not forget historical data while optimizing the current stage data, but also be able to flexibly adapt to multiple resource allocation goals in the current stage to achieve efficient and fair allocation of resources, thereby continuously improving network performance and user experience in a dynamically changing wireless communication environment.

[0190] In some implementations, in order to ensure that the output of the model remains unchanged for historical data when the second model is learning in the new stage (current stage), the update of the model parameters can be limited to the sample null space of the previous historical stage to effectively avoid catastrophic forgetting, allowing the model to continuously learn new knowledge without affecting the results of the model's prediction of historical data. For example, before step 207, it can also include:

[0191] (B.1) processing the corresponding multiple historical channel state data in each network layer through the second model to obtain an input data matrix corresponding to each network layer;

[0192] (B.2) In each network layer, the input data matrix is ​​subjected to low-rank approximation processing to obtain an updated input data matrix;

[0193] (B.3) Based on the updated input data matrix, determine the sample null space corresponding to each network layer.

[0194] The input data matrix may refer to a matrix form of all sample channel state data received by each network layer in the neural network.

[0195] Furthermore, after each training phase, the neural network model optimized in this phase can be used to train all data samples (number n) in the current phase. t ) performs a complete forward propagation. During this process, the input data matrix X of each layer is recorded. These input data matrices will be used to limit the parameter update of each layer of the network in the next stage of training, so that it remains in the null space of X. In this way, effective memory of historical information can be achieved between each training stage, while providing a more reliable foundation for learning data in the current stage. Taking the network layer as the lth layer as an example, the input data matrix is ​​a dimension of O l-1 ×n t The matrix, where O l-1 is the output dimension of the neuron in the l-1th layer, that is, the number of features output by the l-1th layer, n t is the number of sample channel state data at the current stage.

[0196] Furthermore, the input data matrix of each network layer is processed with low rank approximation to obtain an updated input data matrix, so as to remove unimportant features and retain the most important information, thereby simplifying the intermediate model and reducing the number of parameters.

[0197] Furthermore, when calculating the sample null space corresponding to each network layer, the input data matrix X of each network layer in the model can be calculated after each training stage (such as after all data in the current stage are trained), and it can be saved as the data of the previous stage. The input data matrix will be used to limit the parameter update of each network layer in the next stage of training, so that it remains in the null space of X. In this way, effective memory of historical information can be achieved between various training stages, while providing a more reliable foundation for learning data in the current stage.

[0198] In some implementations, in the linear layer of the intermediate model, there may be the following computational expression:

[0199] WX=Y;

[0200] Where W is the value of the model parameter variable in the current stage, X is the input data matrix, and Y is the output result of the intermediate model. To ensure that the learning of the sample training set corresponding to the channel state in the current stage by the intermediate model does not affect the performance of the previous stage, the model parameter update ΔW can be restricted to the sample null space of the data X in the previous historical stage of the current stage, that is, it satisfies:

[0201] ΔWX=0,(W+ΔW)X=WX;

[0202] By updating the model parameters of the current stage in the sample null space, the model can maintain the memory of historical data during the learning process of the new stage, achieve continuous learning and optimize wireless resource allocation, and improve the robustness of the model.

[0203] In some implementations, in order to provide a basis for subsequent low-rank approximation processing and sample null space determination, after each stage of training, the sample channel state data of the corresponding stage can be processed by the trained model, and the input data matrix of each layer can be obtained layer by layer, so that in the subsequent new stage of training, the model can maintain the memory of historical data and avoid catastrophic forgetting. For example, (B.1) may include:

[0204] (B.1.1) Processing the plurality of historical channel state data according to the historical model parameters corresponding to each network layer through the second model to obtain an input data matrix corresponding to each network layer;

[0205] (B.1.2) Repeat in the next network layer of the second model, process multiple historical channel state data according to the next historical model parameters corresponding to the next network layer, and obtain the next input data matrix corresponding to the next network layer, until the input data matrix of the last network layer of the intermediate model is obtained.

[0206] In some embodiments, each layer of the model receives the output of the previous layer as its input and generates a new feature representation for use by the next layer. The input data matrix of each network layer contains the information that needs to be processed by the network at that layer and is a key component in the model learning process. The parameter update of each network layer is based on the input data matrix of the layer and the gradient of the loss function. Therefore, each layer requires an input data matrix to guide the update of its parameters.

[0207] Exemplarily, if the second model includes an input layer, two hidden layers (hidden layer 1 and hidden layer 2) and an output layer (this is just a simple example, the model structure may be more complicated in actual situations), the second model is used to process historical channel state data (it can also be sample historical channel state data, and the input data is determined according to actual conditions) to obtain the input data matrix of each layer. Historical sample channel state data can be received through the input layer. Assuming that these data have 3 features (for example, channel gain, noise level and interference level), there are a total of 100 data, then the input data matrix X1 of the input layer is a 3×100 matrix.

[0208] Hidden layer 1 receives the output of the input layer as its input. Assume that hidden layer 1 has 4 neurons, each of which is connected to all 3 features of the input layer. The second model processes these input data using its historical parameters in hidden layer 1 to generate the output of hidden layer 1. These outputs are then used as inputs to hidden layer 2. Therefore, the input data matrix X2 of hidden layer 1 will be a 4×100 matrix. Hidden layer 2 repeats the above process and receives the output of hidden layer 1 as its input. Assume that hidden layer 2 also has 4 neurons and processes the output from hidden layer 1. The second model processes these input data using its historical parameters in hidden layer 2 to generate the output of hidden layer 2. These outputs are finally used as inputs to the output layer. Therefore, the input data matrix X3 of hidden layer 2 will also be a 4×100 matrix.

[0209] In the above way, the input data matrix of each layer can be obtained layer by layer. These matrices will be used to determine the subsequent sample null space to ensure that the model can maintain the memory of historical data and avoid catastrophic forgetting in the new stage of training. The input data matrix of each layer not only guides the update of the parameters of this layer, but also provides the necessary information for the next layer, thereby achieving effective learning and memory retention in the entire network.

[0210] In some implementations, in order to reduce the complexity and amount of calculation of parameter updates in subsequent model training while retaining the most important information, the dimension of the input data matrix can be reduced through low-rank approximation processing to effectively avoid the problems of large amount of calculation and long time consumption during model training caused by high-dimensional data, while ensuring that the model's memory of historical data is maintained, thereby improving the adaptability and robustness of the model in a dynamic wireless environment. For example, (B.2) may include:

[0211] (B.2.1) In each network layer, perform singular value decomposition on the corresponding input data matrix to obtain the left singular vector matrix, the singular value diagonal matrix, and the transpose of the right singular vector matrix of the input data matrix;

[0212] (B.2.2) obtaining a preset extraction value, and extracting the same number of matrix columns as the preset extraction value from the left singular vector matrix to obtain an updated left singular vector matrix;

[0213] (B.2.3) extracting the same number of singular values ​​as the preset extracted values ​​from the singular value diagonal matrix to obtain an updated singular value diagonal matrix;

[0214] (B.2.4) extracting the same number of matrix rows as the preset extraction value from the transpose of the right singular vector matrix to obtain the updated transpose of the right singular vector matrix;

[0215] (B.2.5) Based on the transpose of the updated left singular vector matrix, the updated singular value diagonal matrix and the updated right singular vector matrix, an updated input data matrix is ​​obtained.

[0216] The left singular vector matrix can be obtained by performing singular value decomposition (SVD) on the input data matrix (denoted by A), and A=U∑V T , where U is the left singular vector matrix. The column vectors of the left singular vector matrix are orthogonal, and these column vectors are the matrix A multiplied by its transpose (i.e. AA T ), the dimension of the left singular vector matrix is ​​the same as the number of rows of the input data matrix A, and the number of columns is equal to the rank of A.

[0217] The singular value diagonal matrix may be a diagonal matrix Σ obtained by performing singular value decomposition on the input data A. The elements on the diagonal of the singular value diagonal matrix are the singular values ​​of the input data matrix, which are sorted from large to small. These singular values ​​reflect the importance or energy of each principal component.

[0218] The transpose of the right singular vector matrix can be the transpose V of another orthogonal matrix V obtained by performing singular value decomposition on the input data A. T , the column vectors of the right singular vector matrix are orthogonal, and these column vectors are the input data matrix A transposed and right multiplied by A (i.e. AA T ). The dimensions of the transpose of the right singular vector matrix are the same as the number of columns of the input data matrix A, and the number of rows is equal to the rank of the input data matrix.

[0219] The preset extraction value may be a pre-set value for extracting a specific number of columns, singular values ​​or rows in the matrix after singular value decomposition. The preset extraction value may be determined according to the needs of the actual application. For example, when reducing the dimension of the data, the value may be set according to how many main components are retained. For example, if the three most important components in the input data matrix are retained, the preset extraction value is set to 3.

[0220] The updated left singular vector matrix may be obtained by extracting a corresponding number of columns from the left singular vector matrix U according to a preset extraction value. Assuming that the preset extraction value is k, the updated left singular vector matrix is ​​a matrix composed of the first k columns of U, and its dimension is the number of rows of the input data matrix A multiplied by k.

[0221] The updated singular value diagonal matrix may be a singular value matrix Σ in which a corresponding number of singular values ​​are extracted according to a preset extraction value. If the preset extraction value is k, the updated singular value diagonal matrix is ​​a k×k diagonal matrix, and the elements on the diagonal are the largest k singular values ​​in the original singular value diagonal matrix.

[0222] The updated transpose of the right singular vector matrix can be based on the preset extraction value, and the transpose V of the right singular vector matrix is T Extract the corresponding number of rows from . Assuming the preset extraction value is k, then the transpose of the updated right singular vector matrix It is a matrix consisting of the first k rows of , whose dimension is k times the number of columns of the input data matrix A.

[0223] In some embodiments, the preset extraction values ​​of each network layer may be the same or different, and the size of the preset extraction values ​​may be flexibly adjusted according to the actual situation. For example, if the first hidden layer may need to capture the basic features of the input data, when more singular values ​​need to be retained to obtain richer information, the first 80% of the singular values ​​may be retained. For another example, if the third hidden layer is close to the output layer, more abstract features may be required, and only the first 30% of the singular values ​​may be retained, and so on. In actual operation, the preset extraction values ​​can be adjusted based on experimental results to find the best balance between information retention and computational efficiency. The preset extraction values ​​of each layer can be set according to the contribution of the layer to the final output and the complexity of the data of the layer to ensure the performance and computational efficiency of the entire network.

[0224] For example, taking the second model as an example, it has been trained to process channel state data in a wireless communication system and perform resource allocation. At this time, when the second model needs to be processed with a low-rank approximation, for the network layer A in the second model, if the network layer A corresponds to a d×n input data matrix X, where s is the feature dimension and n is the number of samples. Performing singular value decomposition on X can be expressed as: X=UΣV T . Where U is a d×d left singular vector matrix, Σ is a d×n singular value diagonal matrix, V T is the transpose of the n×n right singular vector matrix.

[0225] Furthermore, if the preset extraction value k is set to 3, that is, the first k largest singular values ​​and their corresponding singular vectors are retained, the first k columns are extracted from U to form a new d×k left singular vector matrix U k . Extract the first k largest singular values ​​from Σ to form a new k×k singular value diagonal matrix Σ k . From V T Extract the first k rows from the matrix to form the transpose of a new k×n right singular vector matrix Combine the updated left singular vector matrix, the updated singular value diagonal matrix, and the transpose of the updated right singular vector matrix to obtain the updated input data matrix

[0226] In this way, the computational complexity can be reduced without losing important information, while retaining important information and expanding the dimension of the null space, which not only improves the freedom of parameter updates in subsequent stages, but also enhances the adaptability and stability of the model to new environments.

[0227] In some embodiments, as training progresses, the null space dimension of the input data decreases rapidly, resulting in a limited feasible space for parameter updates, which in turn leads to a decrease in model performance. In order to enable the model to maintain robustness and efficiency in a complex and changeable wireless communication environment, the degree of freedom of parameter updates can be increased through structural pruning to reduce the complexity and computational complexity of the model, improve the training efficiency of the model, and enable the model to quickly adapt to the new channel environment. Exemplarily, the wireless resource allocation method may also include:

[0228] (C.1) obtaining multiple scaling parameters corresponding to multiple network layers of the second model;

[0229] (C.2) obtaining a plurality of scaling parameter absolute values ​​corresponding to the plurality of scaling parameters, and determining a plurality of target scaling parameters based on a magnitude order of the plurality of scaling parameter absolute values;

[0230] (C.3) Determine multiple target neurons corresponding to multiple target scaling parameters, and temporarily disable the corresponding multiple target neurons in multiple network layers.

[0231] The scaling parameter may be a learnable parameter used to adjust the scale of the normalized features. The scaling parameter may be denoted by γ, which usually works together with the bias parameter β to restore the normalized data distribution.

[0232] The absolute value of the scaling parameter may be an absolute value of the scaling parameter, which is used to measure the degree of its influence on the output of the neuron. A larger absolute value means that the neuron has a greater contribution to the model output; conversely, a smaller absolute value means that its contribution is lower.

[0233] The target scaling parameter may be a scaling parameter whose absolute value is less than a scaling threshold during the selective disabling process, and the neuron corresponding to the target scaling parameter will be temporarily disabled to optimize the model performance.

[0234] Among them, the target neuron can be a neuron with a target scaling parameter, that is, a neuron with a smaller absolute value of the scaling parameter and considered to contribute less to the overall performance of the model. During the pruning process, the target neuron and all its related parameters will be temporarily set to zero, making the target neuron completely ineffective in the calculation.

[0235] Exemplarily, all network layers of the second model can be traversed, and for each network layer including a BN layer, the scaling parameters of the layer can be extracted. In each network layer, the absolute value of each scaling parameter is calculated, and all absolute values ​​of scaling parameters in the network layer are sorted in order from small to large.

[0236] Further, a scaling threshold or scaling ratio can be determined, and the scaling parameter with the smallest absolute value can be selected as the target scaling parameter. For example, the top 10% of scaling parameters with the lowest absolute values ​​can be selected as the target scaling parameters, and the target neuron corresponding to the target scaling parameter can be determined, and all relevant parameters of the target neuron and its corresponding BN layer (including γ and β) are set to zero to ensure that these parameters no longer produce any output when the model is trained in the current stage. Further, the scaling threshold or scaling ratio can be set according to actual conditions, for example, the scaling ratio can be 20%, 30%, etc., and the embodiments of the present application do not impose specific restrictions on this.

[0237] In some implementations, the disabled neurons may be disabled at the current stage or may be continuously disabled, which is specifically set according to the actual situation.

[0238] Through the above method, the lower limit of the dimension of the null space is increased, which provides a larger optimization space for the subsequent stages. At the same time, it also improves the learning ability of the second model on new data, maintains the model's memory of historical data, avoids the problem of catastrophic forgetting, and helps the model maintain robustness and efficiency in complex and changing environments.

[0239] Please refer to Figure 5 and Figure 6 In some implementations, the wireless resource allocation method of the present application can be evaluated by testing. For example, for a downlink scenario including 10 target base stations and 10 target terminals, with the downlink transmission and rate optimization as the goal, it is assumed that the channel goes through 3 different distribution stages, and the channel characteristics of each stage are independent and vary significantly. The maximum transmit power of the target base station is set to 1W, and the weight of each target terminal is the same to ensure the fairness of resource allocation.

[0240] Figure 5 The performance comparison results of the wireless resource allocation method of the present application and other learning methods are shown. Figure 5It can be seen that compared with the memory-based learning method, the performance of the method of the present application is improved by nearly 1 / 3. Compared with the transfer learning method, the performance of the method of the present application is more significantly improved. This shows that the method proposed in the present application has stronger adaptability and optimization effect in a dynamic environment.

[0241] Figure 6 The results of the training time comparison between the wireless resource allocation method of the present application and other learning methods are shown. Figure 6 It can be seen that the training time of the memory-based learning method increases exponentially when the number of stages is large. However, the training time of the wireless resource allocation method proposed in this application only increases linearly, which fully demonstrates the significant advantage of the method proposed in this application in terms of training efficiency.

[0242] In an embodiment of the present application, channel state data between multiple target base stations and multiple target terminals are obtained; the channel state data is input into a pre-trained first model to obtain wireless resource allocation results of multiple target base stations to multiple target terminals; wherein the first model is obtained by adjusting the pre-trained second model according to the target model parameters; the target model parameters are the values ​​of model parameter variables when the second model is determined in the sample null space in combination with the objective function to minimize the total loss value after wireless resource allocation for multiple sample channel state data; the sample null space is determined after the second model obtained by adjusting the parameters according to the historical stage processes the input data matrix of multiple historical sample channel state data, and the objective function includes a parameter optimization function and a constraint function; the parameter optimization function includes a functional relationship between the total loss variable and the model parameter variable while minimizing the total loss value as the optimization goal; the constraint function is used to limit the equality relationship between the first loss and the second loss; the first loss is calculated after wireless resource allocation for multiple historical sample channel state data of different historical stages based on the intermediate model corresponding to the current value of the model parameter variable; the second loss is calculated after wireless resource allocation for the multiple historical sample channel state data based on the second model corresponding to the historical model parameters. In this way, the periodicity and repetitiveness of channel distribution can be fully utilized, and according to the multiple sample channel state data of the current stage, the second model that has been trained and adjusted in advance and has good performance can be quickly trained, so that the model does not need to frequently learn the adapted distribution, thereby improving the efficiency and adaptability of model training. On this basis, after the second model obtained by adjusting the parameters according to the historical stage processes the multiple historical channel state data to obtain the input data matrix, the sample zero space is determined, and the optimization is performed in the sample zero space in the current stage, so that the model can smoothly transition between different stages, avoid the training problem of catastrophic forgetting of historically learned knowledge, and enable the model to efficiently adapt to the dynamically changing wireless environment while retaining the memory of historically learned knowledge, thereby improving the accuracy of resource allocation. At the same time, by constructing an objective function to optimize the model parameters, the model can take into account the long-term and short-term goals of resource allocation during the optimization process, making resource allocation more accurate and better meeting the needs of network performance and user experience. Compared with traditional methods, the method of the present application does not rely on complex numerical optimization techniques, but uses the powerful learning ability and parallel computing characteristics of deep learning, which enables the model to quickly and accurately allocate channel resources, thereby improving the real-time and accuracy of resource allocation.

[0243] See also Figure 7 The embodiment of the present application further provides a wireless resource allocation device, which can implement the above wireless resource allocation method, and the wireless resource allocation device includes:

[0244] An acquisition module 71 is used to acquire channel state data between multiple target base stations and multiple target terminals;

[0245] The input module 72 is used to input the channel state data into the pre-trained first model to obtain the wireless resource allocation results of multiple target base stations to multiple target terminals; wherein the first model is obtained by adjusting the pre-trained second model according to the target model parameters; the target model parameters are the values ​​of the model parameter variables when the second model is determined in the sample null space in combination with the objective function to minimize the total loss value after wireless resource allocation for multiple sample channel state data; the sample null space is determined after the second model obtained by adjusting the parameters according to the historical stage processes the multiple historical sample channel state data to obtain the input data matrix, and the objective function includes a parameter optimization function and a constraint function; the parameter optimization function includes a functional relationship between the total loss variable and the model parameter variable while taking the minimization of the total loss value as the optimization goal; the constraint function is used to limit the equality relationship between the first loss and the second loss; the first loss is calculated based on the intermediate model corresponding to the value of the current model parameter variable, and wireless resource allocation is performed on multiple historical sample channel state data at different historical stages; the second loss is calculated based on the second model corresponding to the historical model parameters, and wireless resource allocation is performed on multiple historical sample channel state data.

[0246] The specific implementation of the wireless resource allocation device is basically the same as the specific implementation of the wireless resource allocation method described above, and will not be repeated here. On the premise of meeting the requirements of the embodiment of the present application, the wireless resource allocation device can also be provided with other functional modules to implement the wireless resource allocation method in the above embodiment.

[0247] The embodiment of the present application also provides a computer device, the computer device includes a memory and a processor, the memory stores a computer program, and the processor implements the above wireless resource allocation method when executing the computer program. The computer device can be any intelligent terminal including a tablet computer, a car computer, etc.

[0248] See also Figure 8 , Figure 8 The hardware structure of a computer device according to another embodiment is shown, and the computer device includes:

[0249] The processor 81 may be implemented by a general-purpose CPU (Central Processing Unit), a microprocessor, an application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of the present application;

[0250] The memory 82 can be implemented in the form of a read-only memory (ROM), a static storage device, a dynamic storage device, or a random access memory (RAM). The memory 82 can store an operating system and other applications. When the technical solution provided in the embodiment of this specification is implemented by software or firmware, the relevant program code is stored in the memory 82, and the processor 81 is called to execute the wireless resource allocation method of the embodiment of the present application;

[0251] Input / output interface 83, used to implement information input and output;

[0252] Communication interface 84, used to realize communication interaction between the device and other devices, which can be realized by wired mode (such as USB, network cable, etc.) or wireless mode (such as mobile network, WI FI, Bluetooth, etc.);

[0253] A bus 85 that transmits information between the various components of the device (e.g., the processor 81, the memory 82, the input / output interface 83, and the communication interface 84);

[0254] The processor 81 , the memory 82 , the input / output interface 83 and the communication interface 84 are connected to each other in communication within the device via a bus 85 .

[0255] An embodiment of the present application also provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, the above-mentioned wireless resource allocation method is implemented.

[0256] The memory, as a non-transient computer-readable storage medium, can be used to store non-transient software programs and non-transient computer executable programs. In addition, the memory may include a high-speed random access memory, and may also include a non-transient memory, such as at least one disk storage device, a flash memory device, or other non-transient solid-state storage device. In some embodiments, the memory may optionally include a memory remotely disposed relative to the processor, and these remote memories may be connected to the processor via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0257] The embodiments described in the embodiments of the present application are intended to more clearly illustrate the technical solutions of the embodiments of the present application and do not constitute a limitation on the technical solutions provided in the embodiments of the present application. Those skilled in the art will appreciate that with the evolution of technology and the emergence of new application scenarios, the technical solutions provided in the embodiments of the present application are also applicable to similar technical problems.

[0258] Those skilled in the art will appreciate that the technical solutions shown in the figures do not constitute a limitation on the embodiments of the present application, and may include more or fewer steps than shown in the figures, or a combination of certain steps, or different steps.

[0259] The device embodiments described above are merely illustrative, and the units described as separate components may or may not be physically separated, that is, they may be located in one place or distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0260] Those skilled in the art will appreciate that all or some of the steps in the methods disclosed above, and the functional modules / units in the systems and devices may be implemented as software, firmware, hardware, or a suitable combination thereof.

[0261] The terms "first", "second", "third", "fourth", etc. (if any) in the specification of the present application and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments of the present application described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any of their variations are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device comprising a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0262] It should be understood that in the present application, "at least one (item)" and "several" refer to one or more, and "plurality" refers to two or more. "And / or" is used to describe the association relationship of associated objects, indicating that three relationships may exist. For example, "A and / or B" can mean: only A exists, only B exists, and A and B exist at the same time, where A and B can be singular or plural. The character " / " generally indicates that the previous and subsequent associated objects are in an "or" relationship. "At least one of the following" or similar expressions refers to any combination of these items, including any combination of single or plural items. For example, at least one of a, b or c can mean: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, c can be single or multiple.

[0263] In the several embodiments provided in the present application, it should be understood that the disclosed systems and methods can be implemented in other ways. For example, the system embodiments described above are merely schematic. For example, the division of the above units is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be an indirect coupling or communication connection through some interfaces, devices or units, which can be electrical, mechanical or other forms.

[0264] The units described above as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0265] In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0266] If the integrated unit is implemented in the form of 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 the present application is essentially or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including multiple instructions to enable a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of various embodiments of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (Read-Only Memory, referred to as ROM), random access memory (Random Access Memory, referred to as RAM), disk or optical disk and other media that can store programs.

[0267] The preferred embodiments of the present invention are described above with reference to the accompanying drawings, but the scope of the rights of the present invention is not limited thereto. Any modification, equivalent substitution and improvement made by a person skilled in the art without departing from the scope and essence of the present invention should be within the scope of the rights of the present invention.

Claims

1. A wireless resource allocation method, characterized in that: The method comprises: Acquiring channel state data between a plurality of target base stations and a plurality of target terminals; Inputting the channel state data into a pre-trained first model to obtain wireless resource allocation results of the multiple target base stations to the multiple target terminals; The first model is obtained by adjusting the pre-trained second model according to the target model parameters; the target model parameters are the values ​​of the model parameter variables when the second model minimizes the total loss value obtained after allocating wireless resources for multiple sample channel state data in the sample null space in combination with the objective function; The sample null space is determined after the second model obtained by adjusting parameters according to the historical stage processes a plurality of historical sample channel state data to obtain an input data matrix, and the objective function includes a parameter optimization function and a constraint function; The parameter optimization function includes a functional relationship between a total loss variable and the model parameter variable while minimizing the total loss value as an optimization goal; The constraint function is used to limit the equality relationship between the first loss and the second loss; the first loss is based on the intermediate model corresponding to the current value of the model parameter variable, and is calculated after wireless resources are allocated to multiple historical sample channel state data in different historical stages; the second loss is based on the second model corresponding to the historical model parameter, and is calculated after wireless resources are allocated to the multiple historical sample channel state data.

2. The wireless resource allocation method according to claim 1, characterized in that: The first model is trained in the following way: Acquire a sample training set corresponding to the current stage, wherein the sample training set includes a plurality of sample channel state data; Taking minimizing the total loss value as the optimization goal, and based on the functional relationship between the total loss variable and the model parameter variable, a parameter optimization function is constructed; wherein the total loss variable is obtained by accumulating multiple current sample losses calculated after wireless resource allocation for multiple sample channel state data based on the intermediate model corresponding to the current value of the model parameter variable; After allocating wireless resources to a plurality of historical sample channel state data at different historical stages based on the intermediate model corresponding to the current value of the model parameter variable, a first loss is calculated; After allocating wireless resources to the plurality of historical sample channel state data based on a second model corresponding to the historical model parameters, a second loss is calculated; constructing a constraint function based on an equality relationship between the first loss and the second loss; Constructing an objective function based on the constraint function and the parameter optimization function; In combination with the objective function, determining in the sample null space the target model parameters for minimizing the total loss value obtained after the second model performs wireless resource allocation for the plurality of sample channel state data; The parameters of the second model are adjusted based on the target model parameters to obtain the first model.

3. The wireless resource allocation method according to claim 2, characterized in that: The method further comprises: Acquire at least one resource allocation target for wireless resource allocation in the current stage; wherein the resource allocation target includes a sum rate, a communication delay rate, and an allocation balance rate; Performing wireless resource allocation on each sample channel state data in turn through the intermediate model corresponding to the value of the current model parameter variable to obtain a first wireless resource allocation result; According to the at least one resource allocation target and in combination with the first wireless resource allocation result, calculating at least one sub-loss after wireless resource allocation is performed on each sample channel state data; Based on the sum of the at least one sub-loss, a current sample loss of each sample channel state data processed by the intermediate model corresponding to the current value of the model parameter variable is obtained.

4. The wireless resource allocation method according to claim 2, characterized in that: Before determining, in the sample null space, the target model parameters for minimizing the total loss value obtained after the second model performs wireless resource allocation for the plurality of sample channel state data in combination with the objective function, the method further includes: Processing the corresponding plurality of historical channel state data in each network layer by the second model to obtain an input data matrix corresponding to each network layer; In each of the network layers, low-rank approximation processing is performed on the input data matrix to obtain an updated input data matrix; Based on the updated input data matrix, the sample null space corresponding to each network layer is determined.

5. The wireless resource allocation method according to claim 4, characterized in that: The method of processing the corresponding plurality of historical channel state data in each network layer by the second model to obtain an input data matrix corresponding to each network layer includes: Processing the plurality of historical channel state data according to the historical model parameters corresponding to each network layer through the second model to obtain an input data matrix corresponding to each network layer; Repeat the process in the next network layer of the second model, processing the multiple historical channel state data according to the next historical model parameters corresponding to the next network layer, and obtaining the next input data matrix corresponding to the next network layer, until the input data matrix of the last network layer of the intermediate model is obtained.

6. The wireless resource allocation method according to claim 4, characterized in that: In each of the network layers, the input data matrix is ​​subjected to low-rank approximation processing to obtain an updated input data matrix, including: In each of the network layers, singular value decomposition is performed on the corresponding input data matrix to obtain a left singular vector matrix, a singular value diagonal matrix, and a transpose of a right singular vector matrix of the input data matrix; Obtaining a preset extraction value, and extracting the same number of matrix columns as the preset extraction value from the left singular vector matrix to obtain an updated left singular vector matrix; Extracting the same number of singular values ​​as the preset extraction values ​​from the singular value diagonal matrix to obtain an updated singular value diagonal matrix; Extracting the same number of matrix rows as the preset extraction value from the transpose of the right singular vector matrix to obtain the updated transpose of the right singular vector matrix; An updated input data matrix is ​​obtained based on the transpose of the updated left singular vector matrix, the updated singular value diagonal matrix and the updated right singular vector matrix.

7. The wireless resource allocation method according to claim 4, characterized in that: The method further comprises: Obtaining multiple scaling parameters corresponding to multiple network layers of the second model; Acquire multiple scaling parameter absolute values ​​corresponding to the multiple scaling parameters, and determine multiple target scaling parameters based on the order of magnitude of the multiple scaling parameter absolute values; A plurality of target neurons corresponding to the plurality of target scaling parameters are determined, and the plurality of target neurons are temporarily disabled in the plurality of network layers.

8. A wireless resource allocation device, characterized in that: The device comprises: An acquisition module, used to acquire channel state data between multiple target base stations and multiple target terminals; An input module is used to input the channel state data into a pre-trained first model to obtain the wireless resource allocation results of the multiple target base stations to the multiple target terminals; wherein the first model is obtained by adjusting the pre-trained second model according to the target model parameters; the target model parameters are the values ​​of the model parameter variables when the second model is determined in the sample null space in combination with the objective function to minimize the total loss value after performing wireless resource allocation for multiple sample channel state data; the sample null space is determined after the second model obtained by adjusting the parameters according to the historical stage processes the multiple historical sample channel state data to obtain the input data matrix, and the objective function includes a parameter optimization function and a constraint function; the parameter optimization function includes a functional relationship between the total loss variable and the model parameter variable while minimizing the total loss value as the optimization goal; the constraint function is used to limit the equality relationship between the first loss and the second loss; the first loss is calculated after performing wireless resource allocation for multiple historical sample channel state data at different historical stages based on the intermediate model corresponding to the current value of the model parameter variable; the second loss is calculated after performing wireless resource allocation for the multiple historical sample channel state data based on the second model corresponding to the historical model parameters.

9. A computer device, characterized in that: The computer device includes a memory and a processor, the memory stores a computer program, and the processor implements the wireless resource allocation method according to any one of claims 1 to 7 when executing the computer program.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the wireless resource allocation method according to any one of claims 1 to 7 is implemented.

Images