An online multitask optimization method for resource scheduling of a communication network base station

Abstract

The application discloses an online multi-task optimization method for communication network base station resource scheduling, solves the problem of continuous evolution of the relationship between multiple base station scheduling tasks such as spectrum allocation, power control, user access and load balancing in the dynamic change of network environment. Through the online multi-task relationship matrix tracking mechanism, the dynamic influence of user distribution change, traffic fluctuation and channel quality fading on the correlation between each scheduling task is monitored in real time, and new network load mode and user behavior characteristics are continuously adapted. A two-stage online learning framework is adopted, first, the task relationship matrix is dynamically updated based on real-time base station monitoring data, and then the matrix is used to guide the parameter update of the multi-task scheduling model, realizing intelligent collaborative allocation of base station resources. The method significantly improves the base station resource utilization efficiency and user service quality in the open network environment, and is especially suitable for 5G / 6G base station, small base station dense deployment, mobile edge computing and other communication network scenarios that need real-time resource scheduling optimization.

Description

Technical Field

[0001] This invention relates to an online multi-task optimization method for communication network base station resource scheduling, belonging to the field of communication network resource management and intelligent scheduling technology. Background Technology

[0002] With the rapid development of 5G and the upcoming 6G communication technologies, the number of base stations, service types, and user scale in cellular communication networks are increasing dramatically, making the dynamic scheduling and optimization of network resources increasingly complex. In actual operation, base station scheduling tasks not only include multiple interrelated sub-tasks such as spectrum allocation, power control, user access, and load balancing, but also need to cope with the continuous dynamic changes in factors such as user distribution, service traffic, and channel quality. Traditional base station resource scheduling methods often assume that the correlation between tasks is static and rely on offline trained models for scheduling decisions. However, in an open and dynamic network environment, the relationships between tasks change significantly over time. For example, changes in user behavior patterns, fluctuations in service peaks, and random fading of the channel environment all lead to continuous evolution of task coordination relationships. This dynamic change makes it difficult for scheduling methods based on static assumptions to adapt to new network conditions in a timely manner, resulting in reduced resource utilization and decreased user service quality.

[0003] On the other hand, existing multi-task learning methods often fail to efficiently capture dynamic changes in task relationships when faced with real-time task feedback in the form of data streams. Furthermore, cross-task information sharing is inefficient in the case of sparse feedback, further limiting the applicability of scheduling systems in complex real-world environments. In addition, emerging communication scenarios such as 5G / 6G networks and dense deployments of small base stations present complex network topologies, high user mobility, and diverse service types, further exacerbating the complexity of base station resource scheduling. Traditional methods struggle to maintain stable performance in such open and dynamic environments. Therefore, there is an urgent need for an online learning method capable of tracking changes in task relationships in real time and adaptively optimizing multi-task scheduling strategies to enhance the effectiveness of base station resource scheduling and improve resource utilization and user experience. Summary of the Invention

[0004] Objective: To address the shortcomings of existing communication network base station resource scheduling methods in adapting to continuously changing task relationships, low efficiency in cross-task information sharing, and insufficient resource utilization in open and dynamic environments, this invention proposes an online multi-task adaptive optimization method for communication network base station resource scheduling. This method can track the correlation between multiple tasks in real time in a dynamic network environment and guide the updating of the scheduling model based on these dynamic task relationships. It achieves intelligent collaborative optimization of various resource allocation tasks such as spectrum, power, access, and load, thereby improving resource utilization efficiency and user experience quality.

[0005] Technical Solution: An online multi-task optimization method for base station resource scheduling in communication networks is proposed to adapt in real time to the dynamic impact of factors such as changes in user distribution, fluctuations in traffic, and channel quality fading on the correlation between different scheduling tasks. It is used to achieve real-time collaborative optimization of various scheduling tasks in communication networks, including spectrum allocation, power control, user access, and load balancing. Specifically, initialization is performed in the offline prediction phase to learn a high-performance initial multi-task scheduling model. Then, in the online adaptation phase, the system dynamically tracks the task relationship matrix: at each time step, the system explicitly learns and updates a task relationship matrix. This matrix is ​​used to quantify the similarity and correlation between different scheduling tasks at the current time. This learning process is achieved by minimizing the structural difference between the current scheduling model and an ideal optimal model, and a regularization term is introduced to ensure the stability and robustness of the matrix. Finally, the model is updated online using the task relationship matrix: after obtaining the dynamically updated task relationship matrix, the system uses it to guide the parameter update of the multi-task scheduling model. Specifically, this method integrates the task relationship matrix into the update rules of the online mirror gradient descent algorithm. When a task receives feedback and calculates its gradient, this gradient information is used to update other related tasks based on the weights of the task relationship matrix. Tasks with stronger relationships receive greater weight for information sharing. This mechanism significantly improves sample utilization efficiency. Compared to existing methods, this invention can accurately capture the dynamic changes in the task relationship matrix and adjust the information sharing method among different tasks accordingly. This allows for real-time online improvement of system adaptability and resource scheduling performance in dynamic and complex communication scenarios such as 5G / 6G base stations, dense deployment of small base stations, and mobile edge computing.

[0006] To ensure good performance of the initial multi-task scheduling model, the specific steps of the offline prediction stage initialization training method are as follows:

[0007] Step 100: Offline collection of historical data on the operation of communication network base stations. ,in This represents the total number of samples in the offline dataset. This represents the feature vector collected by the base station's real-time monitoring module, including but not limited to user spatial distribution characteristics, traffic statistics, channel quality indicators (such as SINR and CQI), and base station hardware status parameters. The corresponding scheduling decision result label (such as spectrum allocation scheme number, power control parameters, user access policy, load balancing scheme, etc.) indicates the scheduling decision result label. For a set of tags.

[0008] Step 101, Select a multi-task scheduling model ,in This represents the parameter space of the multi-task scheduling model. Represents the input feature space, This represents the output in the real number space. Multi-task scheduling models can include multi-task linear regression models, multi-task neural network models, or multi-task deep models based on attention mechanisms.

[0009] Step 102, Select the loss function It is used to measure the difference between the predicted scheduling result and the optimal reference schedule. The squared loss function, weighted cross-entropy loss function, Huber loss function, etc. can be selected.

[0010] Step 103: Using the base station operation history dataset collected in step 100, minimize the loss function to obtain the offline initial model parameters using the model selected in step 101 and the loss function selected in step 102. .

[0011] To track the dynamic changes of the inter-task relationship matrix in real time, the specific steps for dynamically tracking the inter-task relationship matrix are as follows:

[0012] Step 200, Initialize the task relationship matrix (Identity matrix), where This indicates the number of scheduled tasks, used to represent the initial assumption of independence between tasks.

[0013] Step 201, at each time step Input characteristics are collected in real time from the base station operation monitoring system. And some feedback information of the currently scheduled tasks.

[0014] Step 2011, based on the current model parameters Comparison of model parameters with theoretical optimal Calculate the structured Bregman divergence ,in Operators that represent the trace operation of a matrix. This represents the task relationship matrix at the current moment.

[0015] Step 2012: Construct the optimization objective function ,in Represents the task relationship matrix. and The regularization coefficients λ and α are used to prevent overfitting and maintain matrix invertibility, respectively. The parameters of the regularization coefficients λ and α are set according to the statistical characteristics of the offline collected historical dataset of communication network base station operation, including variance and correlation coefficient.

[0016] Step 2013: Solve the optimization problem in step 2012 using convex optimization methods (such as projective gradient descent or Bregman projection) to obtain a new task relation matrix. And ensure its symmetry and positive semidefiniteness.

[0017] Step 2014: Store the updated task relationship matrix and pass it to the model update module to guide the next stage of the multi-task model optimization process.

[0018] To utilize the task relationship matrix for cross-task collaborative optimization, the specific steps for online model updating using the task relationship matrix are as follows:

[0019] Step 300: At each time step t, collect the feedback gradients of the currently active scheduled tasks. The gradient is calculated only for tasks that receive feedback; methods for obtaining feedback on currently active tasks may include: direct collection based on real-time monitoring of base stations, aggregated uploading based on edge computing nodes, and unified distribution based on centralized scheduling of the core network.

[0020] Step 301, convert the task relationship matrix By incorporating the online mirror gradient descent (OMD) update formula, a structured Bregman divergence is defined: ,in These are the parameters for the multi-task scheduling model.

[0021] Step 302, solve the following optimization problem to update the multi-task model parameters: ,in Let the loss function be for the current round. This is the learning rate for the current round. Here, represents the parameters of the multi-task scheduling model, and represents the parameter space of the multi-task scheduling model. This is the task relationship matrix.

[0022] Step 303, based on mathematical derivation, yields the explicit update form: ,in, For adaptive step size: ,in, Let G denote the Frobenius norm, and G be the upper bound of the gradient. The number of tasks.

[0023] Step 304: When the relationship between tasks changes slowly, the step size... Decrease the step size to ensure the stability of the update; when the relationship changes drastically, the step size should be reduced. Increase the size to improve the model's response speed.

[0024] Step 305, update the model parameters It is applied to the base station resource scheduling decision module to generate the spectrum allocation, power control, user access and load balancing scheme for the next time step, and continues to execute steps 200 to 305 in the next iteration to achieve online adaptive optimization.

[0025] The multi-task scheduling models available in step 101 include: multi-task linear regression model, multi-task generalized linear model, shared parameter neural network model, multi-task deep neural network model based on attention mechanism, etc.

[0026] The loss functions available in step 102 include: squared loss function, weighted cross-entropy loss function, Huber loss function, logistic loss function, etc.

[0027] The methods for obtaining feedback on currently active tasks in step 300 may include: direct collection based on real-time monitoring of base stations, aggregated uploading based on edge computing nodes, and unified distribution based on centralized scheduling of the core network.

[0028] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the online multi-task optimization method for resource scheduling of communication network base stations as described above.

[0029] A computer-readable storage medium storing a computer program that performs the online multitasking optimization method for scheduling base station resources in a communication network as described above.

[0030] Beneficial Effects: Compared with existing communication network base station scheduling optimization methods, this invention designs a two-stage online multi-task optimization framework that combines dynamic task relationship tracking and structure-aware updates. This framework can learn the structural correlations between tasks in real time and guide cross-task information sharing even under rapidly changing user distribution, service traffic, and channel quality. In environments with sparse feedback or frequent task changes, this invention maintains high resource utilization efficiency and excellent user service quality, making it particularly suitable for scenarios such as 5G / 6G, large-scale dense deployment of small base stations, and mobile edge computing. Attached Figure Description

[0031] Figure 1 This is a flowchart illustrating the overall process of the online multi-task adaptive optimization method for communication network base station resource scheduling according to an embodiment of the present invention.

[0032] Figure 2 This is a flowchart illustrating the dynamic tracking task relationship matrix and the method for online updating of a multi-task model using the task relationship matrix, according to an embodiment of the present invention. Detailed Implementation

[0033] The present invention will be further illustrated below with reference to specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. After reading the present invention, any modifications of the present invention in various equivalent forms by those skilled in the art will fall within the scope defined by the appended claims.

[0034] This embodiment proposes an online multi-task adaptive optimization method for communication network base station resource scheduling. The system mainly includes a data acquisition module, a task relationship matrix dynamic tracking module, a multi-task model online update module, and a scheduling decision execution module.

[0035] like Figure 1 As shown, the parameters of the multi-task scheduling model are first initialized offline. and task relationship matrix The initialization process includes the collection of historical base station operation data, model structure, loss function selection, and initial parameter training.

[0036] like Figure 2 As shown, the process of dynamically tracking the task relationship matrix during the online adaptation phase is as follows: First, the task relationship matrix is ​​initialized; then, at each time step, characteristic data such as user distribution, service traffic, and channel quality are collected in real time from the base station monitoring system, and the current model parameters are calculated. Comparison of model parameters with theoretical optimal The structural differences are addressed by minimizing the structured Bregman divergence and regularization term, and then using a convex optimization algorithm to solve and update the task relation matrix. This matrix reflects the structural similarity between different scheduling tasks, such as spectrum allocation, power control, user access, and load balancing. After obtaining the updated task relationship matrix, the model is updated online using this matrix: feedback gradients of currently active tasks are collected, and the task relationship matrix is ​​incorporated into the online mirror gradient descent update rule to achieve structure-aware gradient weighted propagation. When the correlation between tasks is high, gradient information will be shared more extensively; when the correlation is low, information transmission is suppressed. An adaptive step size mechanism automatically adjusts the update magnitude according to the rate of environmental change to achieve a balance between stability and response speed.

[0037] Finally, the updated multi-task scheduling model generates the base station resource allocation scheme for the next time step and executes it in the network scheduling system, entering the next iteration, thereby achieving adaptive optimization of resource scheduling.

[0038] The specific steps for initializing the training method in the offline prediction phase are as follows:

[0039] Step 100: Offline collection of historical data on the operation of communication network base stations. ,in This represents the total number of samples in the offline dataset. This represents the feature vector collected by the base station's real-time monitoring module, including but not limited to user spatial distribution characteristics, traffic statistics, channel quality indicators (such as SINR and CQI), and base station hardware status parameters. The corresponding scheduling decision result label (such as spectrum allocation scheme number, power control parameters, user access policy, etc.) is indicated. For a set of tags.

[0040] Step 101, Select a multi-task scheduling model , where represents the parameter space of the multi-task scheduling model. Represents the input feature space, This represents the output in the real number space. Multi-task scheduling models can include multi-task linear regression models, multi-task neural network models, or multi-task deep models based on attention mechanisms.

[0041] Step 102, Select the loss function It is used to measure the difference between the predicted scheduling result and the optimal reference schedule. The squared loss function, weighted cross-entropy loss function, Huber loss function, etc. can be selected.

[0042] Step 103: Using the base station operation history dataset collected in step 100, minimize the loss function to obtain the offline initial model parameters using the model selected in step 101 and the loss function selected in step 102.

[0043] .

[0044] The specific steps for dynamically tracking the relationship matrix between tasks are as follows:

[0045] Step 200, Initialize the task relationship matrix (Identity matrix), where This indicates the number of scheduled tasks, used to represent the initial assumption of independence between tasks.

[0046] Step 201, at each time step Input characteristics are collected in real time from the base station operation monitoring system. And some feedback information of the currently scheduled tasks.

[0047] Step 2011, based on the current model parameters Comparison of model parameters with theoretical optimal Calculate the structured Bregman divergence ,in Operators that represent the trace operation of a matrix. This represents the task relationship matrix at the current moment.

[0048] Step 2012: Construct the optimization objective function ,in Represents the task relationship matrix. and These are regularization coefficients, used to prevent overfitting and maintain matrix invertibility, respectively.

[0049] Step 2013: Solve the optimization problem in step 2012 using convex optimization methods (such as projective gradient descent or Bregman projection) to obtain a new task relation matrix. And ensure its symmetry and positive semidefiniteness.

[0050] Step 2014: Store the updated task relationship matrix and pass it to the model update module to guide the next stage of the multi-task model optimization process.

[0051] The specific steps for updating the model online using the task relationship matrix are as follows:

[0052] Step 300, at each time step Collect feedback gradients of currently active scheduled tasks. The gradient is calculated only for the task that receives feedback.

[0053] Step 301, convert the task relationship matrix By incorporating the online mirror gradient descent (OMD) update formula, a structured Bregman divergence is defined: ,in These are the parameters for the multi-task scheduling model.

[0054] Step 302, solve the following optimization problem to update the multi-task model parameters: ,in Let the loss function be for the current round. This is the learning rate for the current round. Here, represents the parameters of the multi-task scheduling model, and represents the parameter space of the multi-task scheduling model. This is the task relationship matrix.

[0055] Step 303, based on mathematical derivation, yields the explicit update form: ,in, For adaptive step size: ,in, Describing the Frobenius norm, This is the upper bound of the gradient. The number of tasks.

[0056] Step 304: When the relationship between tasks changes slowly, the step size... Decrease the step size to ensure the stability of the update; when the relationship changes drastically, the step size should be reduced. Increase the size to improve the model's response speed.

[0057] Step 305, update the model parameters It is applied to the base station resource scheduling decision module to generate the spectrum allocation, power control, user access and load balancing scheme for the next time step, and continues to execute steps 200 to 305 in the next iteration to achieve online adaptive optimization.

[0058] Obviously, those skilled in the art should understand that the steps of the online multi-task adaptive optimization method for communication network base station resource scheduling described in the above embodiments of the present invention can be implemented using general-purpose computing devices. They can be centralized on a single computing device or distributed across a network of multiple computing devices. Optionally, they can be implemented using computer-executable program code, thereby storing them in a storage device for execution by the computing device. Furthermore, in some cases, the steps shown or described can be performed in a different order than presented here, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. Thus, the embodiments of the present invention are not limited to any particular hardware and software combination.

Claims

1. An online multi-task optimization method for base station resource scheduling in communication networks, characterized in that, It includes an offline prediction phase and an online adaptation phase. First, in the offline prediction phase, historical base station operation data is collected to initialize the parameters of the multi-task scheduling model and the task relationship matrix. Then, in the online adaptation phase, the task relationship matrix is ​​dynamically tracked: at each time step, a task relationship matrix is ​​explicitly learned and updated through an online optimization process. This matrix is ​​used to quantify the similarity and correlation between different scheduling tasks at the current moment. The task relationship matrix is ​​updated by minimizing the structured Bregman divergence between the current scheduling model and the theoretically optimal model, and a regularization term is introduced to ensure the stability and robustness of the matrix. Then, the task relationship matrix is ​​used for model update: the dynamically updated task relationship matrix is ​​integrated into the online mirror gradient descent algorithm. When a task receives feedback, its gradient information is propagated among tasks according to the weights of the task relationship matrix. Tasks with high correlation share greater weights, thereby achieving cross-task information sharing and collaborative optimization in feedback-sparse scenarios. Finally, the updated multi-task scheduling model is applied to base station resource allocation decisions to generate real-time spectrum allocation schemes, power control parameters, user access strategies, and load balancing schemes, achieving adaptive optimization of communication network resources.

2. The online multi-task optimization method for communication network base station resource scheduling according to claim 1, characterized in that, The specific steps of the offline prediction phase initialization training method are as follows: Step 100: Offline collection of historical data on the operation of communication network base stations. ,in This represents the total number of samples in the offline dataset. This represents the feature vector collected by the base station real-time monitoring module. This indicates the label representing the corresponding scheduling decision result. A collection of tags; Step 101, Select a multi-task scheduling model ,in This represents the parameter space of the multi-task scheduling model. Represents the input feature space, Indicates the output of real numbers; Step 102, Select the loss function It is used to measure the difference between the predicted scheduling result and the optimal reference schedule; Step 103: Using the base station operation history dataset collected in step 100, minimize the loss function to obtain the offline initial model parameters using the model selected in step 101 and the loss function selected in step 102. ,in This represents the parameter space of the multi-task scheduling model.

3. The online multi-task optimization method for communication network base station resource scheduling according to claim 1, characterized in that, The specific steps for dynamically tracking the relationship matrix between tasks are as follows: Step 200, Initialize the task relationship matrix ,in This indicates the number of scheduled tasks, used to represent the initial assumption of independence between tasks; Step 201, at each time step Input characteristics are collected in real time from the base station operation monitoring system. and some feedback information of the currently scheduled task; Step 2011, based on the current model parameters Comparison of model parameters with theoretical optimal Calculate the structured Bregman divergence ,in Operators that represent the trace operation of a matrix. This represents the task relationship matrix at the current moment. Step 2012: Construct the optimization objective function ,in Represents the task relationship matrix. and These are regularization coefficients, used to prevent overfitting and maintain matrix invertibility, respectively. Step 2013: Solve the optimization problem in step 2012 using convex optimization methods to obtain a new task relationship matrix. And ensure its symmetry and positive semidefiniteness; Step 2014: Store the updated task relationship matrix and pass it to the model update module to guide the next stage of the multi-task model optimization process.

4. The online multi-task optimization method for communication network base station resource scheduling according to claim 3, characterized in that, The specific steps for updating the model online using the task relationship matrix are as follows: Step 300, at each time step Collect feedback gradients of currently active scheduled tasks. The gradient is calculated only for the task that receives feedback; Step 301, convert the task relationship matrix By incorporating the online mirror gradient descent (OMD) update formula, a structured Bregman divergence is defined: ,in These are parameters for the multi-task scheduling model; Step 302: Solve the following optimization problem to update the parameters of the multi-task scheduling model. : ,in Let the loss function be for the current round. This is the learning rate for the current round. For parameters of the multi-task scheduling model, This represents the parameter space of the multi-task scheduling model. This is a task relationship matrix; Step 303, based on mathematical derivation, yields the explicit update form: ,in, The learning rate for the current round: ,in, Describing the Frobenius norm, This is the upper bound of the gradient. The number of scheduled tasks, where For the current round's time, , They are respectively time, The optimal comparison model parameters based on the time-theory; Step 304: When the relationship between tasks changes slowly, the step size... Decrease the step size to ensure the stability of the update; when the relationship changes drastically, the step size should be reduced. Increase the size to improve the model's response speed; Step 305, update the model parameters It is applied to the base station resource scheduling decision module to generate the spectrum allocation, power control, user access and load balancing scheme for the next time step, and continues to execute steps 200 to 305 in the next iteration to achieve online adaptive optimization.

5. The online multi-task optimization method for communication network base station resource scheduling according to claim 3, characterized in that, The regularization coefficients λ and α, as well as the step size, are set based on the statistical characteristics of the offline collected historical dataset of communication network base station operation. The statistical characteristics include variance and correlation coefficient.

6. The online multi-task optimization method for communication network base station resource scheduling according to claim 2, characterized in that, The multi-task scheduling models available in step 101 include: multi-task linear regression model, multi-task generalized linear model, shared parameter neural network model, and multi-task deep neural network model based on attention mechanism.

7. The online multi-task optimization method for communication network base station resource scheduling according to claim 2, characterized in that, The loss functions available in step 102 include: squared loss function, weighted cross-entropy loss function, Huber loss function, and logistic loss function.

8. The online multi-task optimization method for communication network base station resource scheduling according to claim 4, characterized in that, The methods for obtaining feedback on currently active tasks in step 300 include: direct collection based on real-time monitoring of base stations, aggregated uploading based on edge computing nodes, and unified distribution based on centralized scheduling of the core network.

9. A computer device, characterized in that: The computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the online multi-task optimization method for resource scheduling of communication network base stations as described in any one of claims 1-8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When executed by a processor, the computer program implements the steps of the online multi-task optimization method for scheduling base station resources in a communication network as described in any one of claims 1-8.

Images