Model generation method and device, related equipment and storage medium

Abstract

The invention discloses a model generation method and device, related equipment and a storage medium. The method is applied to a first network element and comprises the steps of receiving request information of a to-be-demanded model sent by a second network element; the request information at least comprises first information related to the to-be-demanded model application and second information related to performance requirements; determining a first model in preset registration models based on the first information and the second information; generating a second model according to the second information and the first model; training the second model to obtain a trained third model; the third model is used for being sent to the second network element.

Description

Technical Field

[0001] The present application relates to the field of wireless access networks, and in particular to a model generation method, apparatus, related equipment and storage medium. Background Art

[0002] In existing technologies, edge nodes in networks require artificial intelligence (AI) or machine learning (ML) models to provide service capabilities. The services provided by AI / ML models in wireless network management platforms (e.g., non-near-real-time RAN intelligent controllers (Non-RT RICs)) are far from the data source and final destination of edge nodes, resulting in additional transmission delays and bandwidth overhead. Manually building and training a lightweight AI / ML model at an edge node requires extensive code development and data annotation. Furthermore, the edge-side model training capabilities are limited, requiring long training times. The model's performance is limited by its complexity, and conventional training methods yield poor results. Summary of the Invention

[0003] In order to solve the existing technical problems, the embodiments of the present application provide a model generation method, apparatus, related equipment and storage medium.

[0004] The technical solution of the embodiment of the present application is implemented as follows:

[0005] An embodiment of the present application provides a model generation method, applied to a first network element, including:

[0006] receiving request information of the model to be required sent by the second network element; the request information at least including first information related to the application of the model to be required and second information related to performance requirements;

[0007] Determining a first model in a preset registration model based on the first information and the second information;

[0008] generating a second model according to the second information and the first model;

[0009] The second model is trained to obtain a trained third model; the third model is used to send to the second network element.

[0010] In the above solution, the first information includes at least:

[0011] Problem label information; the problem label information represents the task goal of the required model for solving the problem;

[0012] Problem input information; the problem input information represents input variable information used by the demand model to solve the problem;

[0013] Problem output information; the problem output information represents the output variable information of the demand model used to solve the problem.

[0014] In the above solution, the second information includes at least:

[0015] Model calculation amount limit; the model calculation amount limit represents the calculation amount limit of the model to be required;

[0016] Model parameter quantity limitation; the model parameter quantity limitation represents the parameter quantity limitation of the model to be required.

[0017] In the above solution, the first information also includes accuracy requirement information; and determining the first model in the preset registration model based on the first information and the second information includes:

[0018] Performing a text semantic search on the first information in a preset registration model to obtain a plurality of first candidate models;

[0019] Determining a plurality of second candidate models that meet the accuracy requirement information from the plurality of candidate models;

[0020] Sorting the plurality of second candidate models according to the accuracy performance index of each second candidate model to obtain a ranking result of the plurality of second candidate models;

[0021] The first model is determined based on the ranking result and the second information.

[0022] In the above solution, generating the second model according to the second information and the first model includes:

[0023] Calculating a first calculation amount and a first parameter amount of the first model;

[0024] determining a first ratio of the model computation amount limit to the first computation amount and a second ratio of the model parameter amount limit to the first parameter amount;

[0025] Determining whether both the first ratio and the second ratio are greater than a preset threshold;

[0026] If both the first ratio and the second ratio are not greater than the preset threshold, adjusting the first model;

[0027] When both the first ratio and the second ratio are greater than the preset threshold, adjusting the first model is stopped and a second model is generated.

[0028] In the above solution, adjusting the first model includes at least one of the following:

[0029] reducing the number of layers of the first model;

[0030] Reduce the number of neurons in each layer of the first model.

[0031] In the above solution, the method further includes:

[0032] determining a ratio of the first ratio to the second ratio;

[0033] The magnitude of the first model is adjusted based on the magnitude of the ratio.

[0034] In the above solution, the training of the second model to obtain a trained third model includes:

[0035] Obtaining a historical data set based on a training request for the second model;

[0036] Training the second model using the historical data set to obtain a loss function of the second model;

[0037] Parameters of the second model are optimized based on the loss function to obtain a trained third model.

[0038] In the above solution, the loss function includes at least one of the following:

[0039] Mean squared error loss function;

[0040] Negative log-likelihood loss function;

[0041] KL divergence loss function.

[0042] In the above solution, the parameter optimization method includes one of the following:

[0043] Stochastic Gradient Descent (SGD);

[0044] Root Mean Square Propagation (RMSProp);

[0045] Adaptive Gradient (AdaGrad);

[0046] Adaptive Moment Estimation (Adam).

[0047] The embodiment of the present application further provides a model generation method, which is applied to a second network element and includes:

[0048] Sending request information for a model to be required to a first network element; the request information includes at least first information related to the application of the model to be required and second information related to performance requirements; the first information and the second information are used by the first network element to determine a first model from a preset registered model based on the first information and the second information; generating a second model based on the second information and the first model; and training the second model to obtain a trained third model.

[0049] Receive the third model sent by the first network element.

[0050] In the above solution, the first information includes at least:

[0051] Problem label information; the problem label information represents the task goal of the required model for solving the problem;

[0052] Problem input information; the problem input information represents input variable information used by the demand model to solve the problem;

[0053] Problem output information; the problem output information represents the output variable information of the demand model used to solve the problem.

[0054] In the above solution, the second information includes at least:

[0055] Model calculation amount limit; the model calculation amount limit represents the calculation amount limit of the model to be required;

[0056] Model parameter quantity limitation; the model parameter quantity limitation represents the parameter quantity limitation of the model to be required.

[0057] The embodiment of the present application further provides a model generation device, which is provided in a first network element and includes:

[0058] A first receiving unit is configured to receive request information of the model to be required sent by the second network element; the request information at least includes first information related to the application of the model to be required and second information related to performance requirements;

[0059] a determining unit, configured to determine a first model in a preset registered model based on the first information and the second information;

[0060] a generating unit, configured to generate a second model according to the second information and the first model;

[0061] A training unit is used to train the second model to obtain a trained third model; the third model is used to send to the second network element.

[0062] The embodiment of the present application further provides a model generation device, which is provided in a second network element and includes:

[0063] A sending unit, configured to send request information for a model to be required to a first network element; the request information including at least first information related to an application of the model to be required and second information related to performance requirements; the first information and the second information being used by the first network element to determine a first model from a preset registered model based on the first information and the second information; generating a second model based on the second information and the first model; and training the second model to obtain a trained third model.

[0064] The second receiving unit is used to receive the third model sent by the first network element.

[0065] The embodiment of the present application further provides a first network element, comprising: a first communication interface and a first processor; wherein,

[0066] The first communication interface is configured to receive request information of the model to be required sent by the second network element; the request information at least including first information related to the application of the model to be required and second information related to performance requirements;

[0067] The first processor is used to determine a first model in a preset registration model based on the first information and the second information; generate a second model according to the second information and the first model; train the second model to obtain a trained third model; and the third model is used to send to the second network element.

[0068] The embodiment of the present application further provides a second network element, comprising: a second communication interface and a second processor; wherein,

[0069] The second communication interface is used to send request information of the model to be required to the first network element; the request information includes at least first information related to the application of the model to be required and second information related to performance requirements; the first information and the second information are used by the first network element to determine the first model in the preset registration model based on the first information and the second information; generate a second model according to the second information and the first model; train the second model to obtain a trained third model; and receive the third model sent by the first network element.

[0070] The embodiment of the present application further provides a first network element, comprising: a first processor and a first memory for storing a computer program that can be run on the processor,

[0071] The first processor is configured to execute the steps of any one of the above-mentioned methods on the first network element side when running the computer program.

[0072] The embodiment of the present application further provides a second network element, comprising: a second processor and a second memory for storing a computer program that can be run on the processor,

[0073] The second processor is configured to execute the steps of any one of the above-mentioned methods on the second network element side when running the computer program.

[0074] An embodiment of the present application also provides a computer program product, including a computer program, which, when executed by a processor, implements the steps of any of the above-mentioned methods on the first network element side, or implements the steps of any of the above-mentioned methods on the second network element side.

[0075] An embodiment of the present application also provides a storage medium on which a computer program is stored. When the computer program is executed by a processor, the computer program implements the steps of any of the above-mentioned methods on the first network element side, or implements the steps of any of the above-mentioned methods on the second network element side.

[0076] The model generation method, apparatus, related equipment and storage medium provided in the embodiments of the present application are as follows: a first network element receives request information of a model to be required sent by a second network element; the request information includes at least first information related to the application of the model to be required and second information related to performance requirements; a first model is determined in a preset registration model based on the first information and the second information; a second model is generated based on the second information and the first model; the second model is trained to obtain a trained third model; the third model is used to be sent to the second network element. Using the embodiments of the present application, by receiving request information of a model to be required sent by a second network element, including first information related to the application of the model to be required (for example, an AI / ML model) and second information related to performance requirements; a first model is determined in a preset registration model based on the first information and the second information; a second model is generated based on the second information and the first model; the second model is trained to obtain a trained third model; the third model is used to be sent to the second network element. That is, when the first network element (such as a base station or terminal) generates an AI / ML model requirement, it sends a model request to the second network element. The second network element retrieves the existing model based on the model request, and generates a new lightweight model based on the existing model according to the performance requirements in the request. The newly generated lightweight model is trained using the historical data set generated by the inference calculation of the existing model in a real environment, and the trained new model is sent to the second network element for deployment. This solves the problem of missing AI / ML models on the edge side of the wireless network. Compared with the existing AI / ML deployment solution of the network management platform, the AI / ML deployment location can be made closer to the edge data generation source and the actual action location. Compared with other existing edge-side methods of obtaining AI / ML models (such as training from scratch, pruning large models, and quantization), it can effectively reduce human intervention, improve training and deployment speed, and improve model accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0077] Figure 1 A flow chart of a method for generating a model according to an embodiment of the present application;

[0078] Figure 2 A flow chart of another method for generating a model according to an embodiment of the present application;

[0079] Figure 3 A schematic diagram of the process of generating a lightweight model according to an embodiment of the present application;

[0080] Figure 4 A schematic diagram of the candidate model range of the embodiment of the present application;

[0081] Figure 5 A schematic diagram of the process of registering the model and historical data in the embodiment of the present application;

[0082] Figure 6 Schematic diagram of the automatic generation and training process of a lightweight model based on Non-RT RIC in an embodiment of the present application;

[0083] Figure 7 This is a structural diagram of a model generation device according to an embodiment of the present application;

[0084] Figure 8 This is a structural diagram of another model generation device according to an embodiment of the present application;

[0085] Figure 9 This is a schematic diagram of the structure of the first network element in an embodiment of the present application;

[0086] Figure 10 This is a schematic diagram of the structure of the terminal in the embodiment of the present application;

[0087] Figure 11 This is a schematic diagram of the system structure of the model generation embodiment of this application. DETAILED DESCRIPTION

[0088] The present application will be described in further detail below with reference to the accompanying drawings and embodiments.

[0089] Wireless networks are complex, heterogeneous networks characterized by continuous wide-area coverage, high-capacity hotspots, high reliability, low latency, and massive connectivity with low power consumption. Currently, various AI algorithms, such as reinforcement learning and deep learning, are widely used in wireless networks.

[0090] In existing technical solutions, the AI / ML service capabilities of wireless networks are mostly provided by the Near Real-Time RAN Intelligent Controller (Near-RT RIC) platform or the Non-RT RIC platform in the SMO, because these platforms can provide the computing power required for AI / ML services. However, in the AI / ML workflow, the source of its model input data and the final destination of its output data are usually the edge side of the wireless network system, such as base stations and terminals. The AI / ML services in the Near-RT RIC and Non-RT RIC platforms need to calculate the massive data from the edge and then send inference data or instructions to these edge sides, which often causes bandwidth pressure and latency pressure on data transmission.

[0091] Implementing AI / ML service capabilities directly at the point where data is generated and where AI / ML services ultimately serve can effectively alleviate the above pressure. However, the edge side typically has strict limitations on computing resources, storage resources, and device energy consumption, and therefore has strict requirements on model complexity and computational effort. Existing AI / ML models deployed in Near-RT RIC and Non-RT RIC platforms are typically complex models with high precision and cannot meet edge-side deployment requirements.

[0092] There are currently two main ways to deploy AI / ML models on the edge: one is to design, build, and train a small model directly on the edge. This method requires more manual code development and data labeling, and the edge's model training capabilities are limited, which may require a long training time. Therefore, this method has a long development and deployment cycle, and the small model is limited by the size of the parameter space, and cannot achieve the same accuracy as a large model through general training methods. The other method is to compress the complex AI / ML model deployed on the Near-RT RIC or Non-RT RIC platform through pruning, quantization, and other means, and use the compressed model for edge-side inference calculations. This method will also cause a loss in model accuracy.

[0093] Based on the above analysis, edge nodes in the network require AI / ML models to provide service capabilities. However, the shortcomings of existing technologies are:

[0094] (1) The services provided by AI / ML models on wireless network management platforms (e.g., non-RT RIC) are far from the data source and final destination of edge nodes, resulting in additional transmission delay and bandwidth overhead. Therefore, it is necessary to deploy lightweight AI / ML models at edge nodes that meet stringent computing and storage resource, energy consumption, and inference speed constraints.

[0095] (2) Manually building and training a lightweight AI / ML model at the edge node requires more manual code development and data labeling work, and the edge side has limited model training capabilities, which may require a long training time. Therefore, this method has a long development and deployment cycle, less available training data, and limited performance due to model complexity. The general training method has poor model training results.

[0096] (3) The complex AI / ML models deployed on the Near-RT RIC or Non-RT RIC platform are compressed through pruning, quantization, and other means, and the compressed models are used for edge-side inference calculations. This approach will also cause a loss in model accuracy.

[0097] Based on this, in various embodiments of the embodiments of the present application, when an edge-side network element (such as a base station or a terminal) generates an AI / ML model requirement, a model request is sent to a network management unit (such as an SMO or a Near RT RIC). The network management unit retrieves an existing model according to the model request, generates a new lightweight model according to the existing model according to the performance requirements in the request, and uses a historical data set generated by the inference calculation of the existing model in a real environment to train the newly generated lightweight model, and sends the trained new model to the edge side for deployment.

[0098] The embodiment of the present application provides a model generation method, which is applied to a first network element, such as Figure 1 As shown, Figure 1 A flow chart of a method for generating a model according to an embodiment of the present application is provided, including:

[0099] Step 101: Receive request information of a model to be required sent by a second network element; the request information at least includes first information related to the application of the model to be required and second information related to performance requirements;

[0100] Step 102: Determine a first model in a preset registration model based on the first information and the second information;

[0101] Step 103: Generate a second model based on the second information and the first model;

[0102] Step 104: Train the second model to obtain a trained third model; the third model is used to send to the second network element.

[0103] It should be noted that the first network element and the second network element can be determined according to actual conditions and are not limited here. As an example, the first network element can be an edge side network element, such as a base station or a terminal; the second network element can be a network management unit, such as an SMO or a Near RT RIC.

[0104] In step 101, the request information includes at least first information related to the application of the model to be requested and second information related to performance requirements. Both the first and second information can be determined based on specific circumstances and are not limited herein. As an example, the first information may include problem label information, problem input information, problem output information, accuracy requirement information, expected base model information, and wireless scenario feature classification information. The second information may include model computational constraints and model parameter constraints. In practical applications, the problem label information may be referred to as the problem label; the problem input information may be referred to as the problem input; the problem output information may be referred to as the problem output; the accuracy requirement information may be referred to as the model accuracy requirement; the expected base model information may be referred to as the expected base model; and the wireless scenario feature classification information may be referred to as the wireless scenario feature classification. As an example, the problem label indicates the scenario and problem to be solved by the model, i.e., the task objective of the model strategy, such as achieving 10-class image classification or optimal allocation of network slice resources. The problem input indicates the input variable information of the problem model to be solved by the AI ​​model, including but not limited to the variable name and data description (e.g., physical meaning). The problem output is used to indicate the output variable information of the problem model used by the AI ​​model to solve, including but not limited to: the name of the variable, the description of the data (such as the physical meaning). The model accuracy requirement is used to indicate the model accuracy requirement required by the edge side. The expected basic model is used to indicate the name of the expected basic model of the model. The wireless scenario feature classification is used to indicate the type of wireless scenario where the model is deployed. The model computation limit is used to indicate the model computation limit required by the edge side, such as the number of calculations required for a single inference calculation of the model, in units of FLOPs (floating point operations). The computation amount can be calculated based on the edge side network element latency requirement (in seconds) and the computing speed of the computing device (for example, floating point operations per second, FLOPS). The model parameter limit is used to indicate the model parameter limit required by the edge side, such as the number of parameters in the model or the amount of storage space occupied. The parameter amount can be calculated based on the expected available storage space of the edge side network element.

[0105] In practical applications, the request information can be referred to as a model request. The first network element receiving the request information of the required model sent by the second network element can be understood as the edge side network element sending the model request to the wireless network management unit, and the wireless network management unit receiving the model request sent by the edge side network element.

[0106] In step 102, the preset registration model may be determined according to actual conditions, which is not limited here. As an example, the preset registration model may be an existing model.

[0107] The specific determination process of determining the first model in the preset registration model based on the first information and the second information can be determined according to actual conditions and is not limited here. As an example, the first information also includes accuracy requirement information; the determination of the first model in the preset registration model based on the first information and the second information can include performing text semantic retrieval on the first information in the preset registration model to obtain multiple first candidate models; determining multiple second candidate models that meet the accuracy requirement information among the multiple candidate models; sorting according to the accuracy performance index of each second candidate model to obtain the sorting results of the multiple second candidate models; and determining the first model based on the sorting results and the second information.

[0108] In step 103, the specific generation process of generating the second model based on the second information and the first model can be determined according to actual conditions and is not limited here. As an example, generating the second model based on the second information and the first model may include calculating the first calculation amount and the first parameter amount of the first model; determining a first ratio of the model calculation amount limit to the first calculation amount and a second ratio of the model parameter amount limit to the first parameter amount; judging whether the first ratio and the second ratio are both greater than a preset threshold; adjusting the first model if the first ratio and the second ratio are not both greater than the preset threshold; stopping adjusting the first model and generating the second model if the first ratio and the second ratio are both greater than the preset threshold.

[0109] In step 104, the specific training process of training the second model to obtain the trained third model can be determined according to the specific situation and is not limited here. As an example, the training of the second model to obtain the trained third model may include obtaining a historical data set based on a training request for the second model; training the second model using the historical data set to obtain a loss function of the second model; and optimizing the parameters of the second model based on the loss function to obtain a trained third model.

[0110] The third model being sent to the second network element can be understood as model delivery. Specifically, the trained model (structure and parameters) and its description file can be packaged into a format that is easily transmitted and delivered to the edge network element that issued the request. The description file may include model function, model input, model output, model name, wireless scenario feature classification, model computational effort, model parameter quantity, and model accuracy. The model function can be used to indicate the scenario and problem for model application, i.e., the task objective of the model strategy, such as achieving 10-class image classification or optimal allocation of network slice resources. The model input can be used to indicate the input variable information of the problem model that the AI ​​model is used to solve, including but not limited to: variable name, data description (e.g., physical meaning), data type, data collection granularity, data accuracy, data value range, data format, and input dimensionality, such as the input feature map size of a CNN or the number of states in reinforcement learning. The model output can be used to indicate the output variable information of the problem model that the AI ​​model is used to solve, including but not limited to: variable name, data description (e.g., physical meaning), and model output dimensionality, such as the classifier output dimensionality of a CNN or the control action dimensionality of reinforcement learning. The model name can be used to indicate the name of the base model used by the model. The wireless scene feature classification can be used to indicate the type of wireless scene where the model is deployed. The model computational amount can be used to indicate the computational amount of the model, such as the number of computations required for one inference calculation of the model, in units of FLOPs (floating point operations). The model parameter amount can be used to indicate the parameter amount of the model, such as the number of parameters in the model or the size of the storage space occupied. The model accuracy can be used to indicate the accuracy achieved by the model at the end of training on historical data.

[0111] In an embodiment of the present application, request information of a model to be required (for example, an AI / ML model) is received from a second network element, including first information related to the application of the model to be required and second information related to performance requirements; a first model is determined in a preset registration model based on the first information and the second information; a second model is generated according to the second information and the first model; the second model is trained to obtain a trained third model; and the third model is used to send to the second network element. That is, when the first network element (such as a base station or terminal) generates an AI / ML model requirement, it sends a model request to the second network element. The second network element retrieves the existing model based on the model request, and generates a new lightweight model based on the existing model according to the performance requirements in the request. The newly generated lightweight model is trained using the historical data set generated by the inference calculation of the existing model in a real environment, and the trained new model is sent to the second network element for deployment. This solves the problem of missing AI / ML models on the edge side of the wireless network. Compared with the existing AI / ML deployment solution of the network management platform, the AI / ML deployment location can be made closer to the edge data generation source and the actual action location. Compared with other existing edge-side methods of obtaining AI / ML models (such as training from scratch, pruning large models, and quantization), it can effectively reduce human intervention, improve training and deployment speed, and improve model accuracy.

[0112] In one embodiment, the first information includes at least:

[0113] Problem label information; the problem label information represents the task goal of the required model for solving the problem;

[0114] Problem input information; the problem input information represents input variable information used by the demand model to solve the problem;

[0115] Problem output information; the problem output information represents the output variable information of the demand model used to solve the problem.

[0116] In this embodiment, the problem label information represents the task objective of the problem to be solved by the model to be required; the task objective can be determined based on the specific situation and is not limited here. As an example, the task objective can be to achieve 10-class image classification, optimal allocation of network slice resources, etc. In practical applications, the problem label information can be simply referred to as a problem label, which is used to indicate the scenario and problem of model application, that is, the task objective of the model strategy, such as achieving 10-class image classification, optimal allocation of network slice resources, etc.

[0117] The problem input information represents the input variable information used by the model to be required to solve the problem; wherein, the input variable information can be determined according to the specific situation and is not limited here. As an example, the input variable information can include but is not limited to: the name of the variable, the description of the data (such as the physical meaning). In practical applications, the problem input information can be simply referred to as problem input, which is used to indicate the input variable information of the problem model used by the AI ​​model to solve, including but not limited to: the name of the variable, the description of the data (such as the physical meaning).

[0118] The problem output information represents the output variable information of the model to be required for solving the problem; wherein, the output variable information can be determined according to the specific situation and is not limited here. As an example, the output variable information may include but is not limited to: the name of the variable, the description of the data (such as the physical meaning). In practical applications, the problem output information can be simply referred to as problem output, which is used to indicate the output variable information of the model for the problem to be solved by the AI ​​model, including but not limited to: the name of the variable, the description of the data (such as the physical meaning).

[0119] In one embodiment, the second information includes at least:

[0120] Model calculation amount limit; the model calculation amount limit represents the calculation amount limit of the model to be required;

[0121] Model parameter quantity limitation; the model parameter quantity limitation represents the parameter quantity limitation of the model to be required.

[0122] In this embodiment, the model computation limit represents the computation limit of the model to be required; wherein, the computation limit can be determined according to specific circumstances and is not limited here. As an example, the computation limit can include the number of calculations required for one inference calculation of the model, in units of FLOPs (floating point operations). The amount of computation can be calculated based on the edge side network element latency requirement (in seconds, s) and the computing speed of the computing device (for example, the number of floating point operations per second, FLOPS). In actual applications, the model parameter limit is used to indicate the model computation limit required on the edge side, such as the number of calculations required for one inference calculation of the model, in units of FLOPs (floating point operations). The amount of computation can be calculated based on the edge side network element latency requirement (in seconds, s) and the computing speed of the computing device (for example, the number of floating point operations per second, FLOPS).

[0123] The model parameter quantity limit characterizes the parameter quantity limit of the model to be required; wherein, the parameter quantity limit can be determined according to the specific situation and is not limited here. As an example, the parameter quantity limit may include the number of parameters in the model or the size of the storage space occupied. The parameter quantity size can be calculated based on the expected size of the storage space available to the edge side network element. In practical applications, the model parameter quantity limit is used to indicate the model parameter quantity limit required on the edge side, such as the number of parameters in the model or the size of the storage space occupied. The parameter quantity size can be calculated based on the expected size of the storage space available to the edge side network element.

[0124] In one embodiment, the first information further includes accuracy requirement information; and determining the first model from a preset registration model based on the first information and the second information includes:

[0125] Performing a text semantic search on the first information in a preset registration model to obtain a plurality of first candidate models;

[0126] Determining a plurality of second candidate models that meet the accuracy requirement information from the plurality of candidate models;

[0127] Sorting the plurality of second candidate models according to the accuracy performance index of each second candidate model to obtain a ranking result of the plurality of second candidate models;

[0128] The first model is determined based on the ranking result and the second information.

[0129] In this embodiment, the preset registration model can be determined according to specific circumstances and is not limited here. As an example, the preset registration model can be an existing model.

[0130] Performing a text semantic search on the first information in a preset registered model to obtain multiple first candidate models can be understood as performing a text semantic search on the question label information in the first information in a preset registered model to obtain the input and output of the retrieved model, and matching the input and output of the retrieved model with the question input information and question output information in the first information one by one. If the match is successful, it is used as the first candidate model. In actual applications, this content can be understood as a preliminary screening of the model. The wireless network management unit receives a request from the edge-side network element, processes and retrieves the content of the request, and the retrieval method is to perform a text semantic search on the question label in the registered model of the model management unit in the wireless network management unit, and match the input and output of the retrieved model with the input and output variables in the request one by one. If the match is successful, it is used as a candidate model.

[0131] The accuracy requirement information can be determined according to specific circumstances and is not limited here. As an example, the accuracy requirement information may include model accuracy requirements; the model accuracy requirements are used to indicate the model accuracy requirements required on the edge side.

[0132] The specific determination process of determining multiple second candidate models that meet the precision requirement information among the multiple candidate models can be determined according to the specific situation and is not limited here. As an example, determining multiple second candidate models that meet the precision requirement information among the multiple candidate models can be to judge whether the values ​​in the precision requirement information are met among the multiple candidate models. If the values ​​in the precision requirement information are met among the multiple candidate models, multiple second candidate models are determined.

[0133] Sorting the multiple second candidate models based on their accuracy performance indicators yields a ranking result. Determining the first model based on the ranking result and the second information can be understood as selecting the models within the candidate range to select the optimal candidate model or the model with the highest accuracy performance indicator. In practical applications, the models within the candidate range are sorted by their accuracy performance indicators, and the top K models are selected. Character matching is then used to determine whether a model with the same name as the expected base model in the requirements exists among these K models. If so, this model is selected first; otherwise, the model with the highest accuracy performance indicator is selected.

[0134] In one embodiment, generating a second model according to the second information and the first model includes:

[0135] Calculating a first calculation amount and a first parameter amount of the first model;

[0136] determining a first ratio of the model computation amount limit to the first computation amount and a second ratio of the model parameter amount limit to the first parameter amount;

[0137] Determining whether both the first ratio and the second ratio are greater than a preset threshold;

[0138] If both the first ratio and the second ratio are not greater than the preset threshold, adjusting the first model;

[0139] When both the first ratio and the second ratio are greater than the preset threshold, adjusting the first model is stopped and a second model is generated.

[0140] In this embodiment, a first computational cost and a first parameter value of the first model are calculated. The first model may be an existing model, and calculating the first computational cost and the first parameter value of the first model can be understood as calculating the computational cost and the parameter value of the existing model. In practical applications, the computational cost and the parameter value of the existing model are calculated based on the existing model structure.

[0141] Determine a first ratio of the model calculation amount limit to the first calculation amount and a second ratio of the model parameter amount limit to the first parameter amount; wherein the first ratio can be recorded as r c The second ratio can be recorded as r p In practical applications, the computational load and parameter load of the existing model are calculated based on the existing model structure, and the ratio r of the model computational load limit requested by the model to the computational load of the existing model is calculated. c , the ratio of the model parameter limit requested by the model to the existing model parameter amount r p .

[0142] Determine whether the first ratio and the second ratio are both greater than a preset threshold; wherein the preset threshold can be determined according to actual conditions and is not limited here. As an example, the preset threshold can be 1. Determine whether the first ratio and the second ratio are both greater than the preset threshold can be understood as determine whether the first ratio and the second ratio are both greater than 1, that is, r c 、r p Are both greater than 1?

[0143] In the case that the first ratio and the second ratio are not both greater than the preset threshold, adjusting the first model can be understood as c 、r p When the difference is greater than 1, the first model is adjusted continuously.

[0144] When both the first ratio and the second ratio are greater than the preset threshold, the adjustment of the first model is stopped and the second model is generated. c 、r p When both are greater than 1, the adjustment of the first model is stopped and the second model is generated.

[0145] In practical applications, this content can be understood as the model generation process. The management unit adjusts the existing model structure according to the received model request. Before each adjustment, the computational load and parameter amount of the existing model are calculated based on the existing model structure, and the ratio r of the model computational load limit of the model request to the computational load of the existing model is calculated. c , the ratio of the model parameter limit requested by the model to the existing model parameter amount r p If both ratios are greater than 1, stop adjusting.

[0146] In one embodiment, adjusting the first model includes at least one of the following:

[0147] reducing the number of layers of the first model;

[0148] Reduce the number of neurons in each layer of the first model.

[0149] In this embodiment, reducing the number of layers of the first model can be simply referred to as reducing the number of model layers; reducing the number of neurons in each layer of the first model can be simply referred to as reducing the number of neurons in each layer of the model. In practical applications, adjustment methods include but are not limited to: (1) reducing the number of model layers; (2) reducing the number of neurons in each layer of the model. During the adjustment process, only one of these methods can be used, or multiple methods can be used.

[0150] In one embodiment, the method further comprises:

[0151] determining a ratio of the first ratio to the second ratio;

[0152] The magnitude of the first model is adjusted based on the magnitude of the ratio.

[0153] In this embodiment, the first ratio can be recorded as r c The second ratio can be recorded as r p Determining the ratio of the first ratio to the second ratio can be understood as determining r c and r p The ratio of .

[0154] The specific adjustment process of adjusting the amplitude of the first model based on the size of the ratio can be determined according to actual conditions and is not limited here. As an example, the amplitude of the first model adjusted based on the size of the ratio can be to select a larger adjustment amplitude if the ratio is smaller; and to select a smaller adjustment amplitude if the ratio is larger.

[0155] In practice, the magnitude of the adjustment can be measured in proportion to the size of the existing model. The magnitude of each adjustment is adjusted based on the ratio of the two calculated values, with the larger magnitude being chosen for smaller ratios.

[0156] In one embodiment, the training of the second model to obtain a trained third model includes:

[0157] Obtaining a historical data set based on a training request for the second model;

[0158] Training the second model using the historical data set to obtain a loss function of the second model;

[0159] Parameters of the second model are optimized based on the loss function to obtain a trained third model.

[0160] In this embodiment, the second model can be understood as a newly generated model, which can be denoted as S.

[0161] The historical data set can be determined according to actual conditions and is not limited here. As an example, the historical data set can be referred to as historical data. The historical data can be recorded as D T , which can be expressed as

[0162] The second model is trained using the historical data set to obtain a loss function of the second model; wherein the loss function can be determined based on actual conditions and is not limited here. As an example, the loss function may include but is not limited to: mean square error loss function; negative log-likelihood loss function; KL divergence loss function, etc.

[0163] The specific training process of using the historical data set to train the second model and obtain the loss function of the second model can be determined according to the actual situation and is not limited here. As an example, for the newly generated model S, the historical data obtained by running the existing model T in the real environment is used. Training. Divide the historical data into N entries according to each inference calculation. Each data entry contains the model T input data x for each inference calculation of the model. i , Model T output data (e.g., output categories of classification networks, output actions in DQN), outputs of key output layers of model T (For example, the softmax layer output vector of the classification network, the Q network output vector in DQN). During the training process, the model input data x i Input to the newly generated model S, the newly generated model S performs inference calculation and output model S output data Output of the key output layer of model S The new generated model S is calculated As prediction data, there is a model T output The loss is calculated as the true value. Methods for calculating loss include but are not limited to: mean square error loss, negative log-likelihood loss, KL divergence loss, etc.

[0164] The parameters of the second model are optimized based on the loss function to obtain a trained third model. The parameter optimization can be determined according to the actual situation and is not limited here. As an example, the parameter optimization method includes but is not limited to: stochastic gradient descent (SGD), adaptive moment estimation method (ADAM), etc. The parameters of the second model are optimized based on the loss function to obtain the specific parameter optimization process in the trained third model. The specific parameter optimization process can be determined according to the actual situation and is not limited here. As an example, the parameters of the second model are optimized based on the loss function to obtain the trained third model. The parameter update optimization of the newly generated model S according to the calculated loss can be performed. The optimization method includes but is not limited to: stochastic gradient descent (SGD), adaptive moment estimation method (ADAM), etc. When the model accuracy no longer improves or the model accuracy requirement meets the model accuracy requirement in the received model requirement, the training is terminated to obtain the trained third model.

[0165] In one embodiment, the loss function includes at least one of the following:

[0166] Mean squared error loss function;

[0167] Negative log-likelihood loss function;

[0168] KL divergence loss function.

[0169] In this embodiment, the mean square error loss function, the negative log-likelihood loss function, and the KL divergence loss function can all be determined according to actual conditions and are not limited here.

[0170] In one embodiment, the parameter optimization method includes one of the following:

[0171] SGD;

[0172] RMSProp;

[0173] AdaGrad;

[0174] Adam.

[0175] In this embodiment, SGD is understood as stochastic gradient descent; RMSProp is understood as root mean square propagation; AdaGrad is understood as adaptive gradient; and Adam is understood as adaptive matrix estimation. In practical applications, parameter optimization methods include but are not limited to SGD and ADAM.

[0176] Accordingly, the present invention also provides a model generation method, such as Figure 2 As shown, Figure 2 This is a flow chart of another model generation method according to an embodiment of the present application, which is applied to a second network element and includes:

[0177] Step 201: Sending request information for a model to be required to a first network element; the request information includes at least first information related to the application of the model to be required and second information related to performance requirements; the first network element is used to determine a first model from a preset registered model based on the first information and the second information; generate a second model based on the second information and the first model; and train the second model to obtain a trained third model.

[0178] Step 202: Receive the third model sent by the first network element.

[0179] It should be noted that the first network element and the second network element can be determined according to actual conditions and are not limited here. As an example, the first network element can be an edge side network element, such as a base station or a terminal; the second network element can be a network management unit, such as an SMO or a Near RT RIC.

[0180] In step 201, the request information includes at least first information related to the application of the model to be requested and second information related to performance requirements. Both the first and second information can be determined based on specific circumstances and are not limited herein. As an example, the first information may include problem label information, problem input information, problem output information, accuracy requirement information, expected base model information, and wireless scenario feature classification information. The second information may include model computational constraints and model parameter constraints. In practical applications, the problem label information may be referred to as the problem label; the problem input information may be referred to as the problem input; the problem output information may be referred to as the problem output; the accuracy requirement information may be referred to as the model accuracy requirement; the expected base model information may be referred to as the expected base model; and the wireless scenario feature classification information may be referred to as the wireless scenario feature classification. As an example, the problem label indicates the scenario and problem to be solved by the model, i.e., the task objective of the model strategy, such as achieving 10-class image classification or optimal allocation of network slice resources. The problem input indicates the input variable information of the problem model to be solved by the AI ​​model, including but not limited to the variable name and data description (e.g., physical meaning). The problem output is used to indicate the output variable information of the problem model used by the AI ​​model to solve, including but not limited to: the name of the variable, the description of the data (such as the physical meaning). The model accuracy requirement is used to indicate the model accuracy requirement required by the edge side. The expected basic model is used to indicate the name of the expected basic model of the model. The wireless scenario feature classification is used to indicate the type of wireless scenario where the model is deployed. The model computation limit is used to indicate the model computation limit required by the edge side, such as the number of calculations required for a single inference calculation of the model, in units of FLOPs (floating point operations). The computation amount can be calculated based on the edge side network element latency requirement (in seconds) and the computing speed of the computing device (for example, floating point operations per second, FLOPS). The model parameter limit is used to indicate the model parameter limit required by the edge side, such as the number of parameters in the model or the amount of storage space occupied. The parameter amount can be calculated based on the expected available storage space of the edge side network element.

[0181] In practical applications, the request information can be referred to as a model request. The first network element receiving the request information of the required model sent by the second network element can be understood as the edge side network element sending the model request to the wireless network management unit, and the wireless network management unit receiving the model request sent by the edge side network element.

[0182] The preset registration model can be determined according to actual conditions and is not limited here. As an example, the preset registration model can be an existing model.

[0183] The specific determination process of the first network element in determining the first model in the preset registration model based on the first information and the second information can be determined according to actual conditions and is not limited here. As an example, the first information also includes accuracy requirement information; the determination of the first model in the preset registration model based on the first information and the second information can include performing text semantic retrieval of the first information in the preset registration model to obtain multiple first candidate models; determining multiple second candidate models that meet the accuracy requirement information among the multiple candidate models; sorting according to the accuracy performance index of each second candidate model to obtain the sorting results of the multiple second candidate models; and determining the first model based on the sorting results and the second information.

[0184] The specific generation process of generating the second model according to the second information and the first model can be determined according to actual conditions and is not limited here. As an example, generating the second model according to the second information and the first model may include calculating the first calculation amount and the first parameter amount of the first model; determining a first ratio of the model calculation amount limit to the first calculation amount and a second ratio of the model parameter amount limit to the first parameter amount; judging whether the first ratio and the second ratio are both greater than a preset threshold; adjusting the first model if the first ratio and the second ratio are not both greater than the preset threshold; stopping adjusting the first model and generating the second model if the first ratio and the second ratio are both greater than the preset threshold.

[0185] The specific training process of training the second model to obtain the trained third model can be determined according to the specific situation and is not limited here. As an example, the training of the second model to obtain the trained third model may include obtaining a historical data set based on a training request for the second model; training the second model using the historical data set to obtain a loss function of the second model; and optimizing the parameters of the second model based on the loss function to obtain a trained third model.

[0186] In step 202, receiving the third model sent by the first network element can be understood as model delivery. Specifically, the trained model (structure and parameters) and its description file can be packaged into a format that is easy to transmit and delivered to the edge network element that issued the request. The description file may contain model function, model input, model output, model name, wireless scene feature classification, model computation amount, model parameter quantity, and model accuracy. The model function can be used to indicate the scenario and problem of model application, that is, the task objective of the model strategy, such as achieving 10-class image classification and optimal allocation of network slice resources. The model input can be used to indicate the input variable information of the problem model that the AI ​​model is used to solve, including but not limited to: variable name, data description (such as physical meaning), data type, data collection granularity, data accuracy, data value range, data format, input dimension, such as the input feature map size of CNN and the number of states of reinforcement learning. The model output can be used to indicate the output variable information of the problem model that the AI ​​model is used to solve, including but not limited to: variable name, data description (such as physical meaning), model output dimension, such as the classifier output dimension of CNN and the control action dimension of reinforcement learning. The model name can be used to indicate the name of the base model used by the model. The wireless scene feature classification can be used to indicate the type of wireless scene where the model is deployed. The model computational amount can be used to indicate the computational amount of the model, such as the number of computations required for one inference calculation of the model, in units of FLOPs (floating point operations). The model parameter amount can be used to indicate the parameter amount of the model, such as the number of parameters in the model or the size of the storage space occupied. The model accuracy can be used to indicate the accuracy achieved by the model at the end of training on historical data.

[0187] In an embodiment of the present application, request information of a model to be required is sent to a first network element; the request information includes at least first information related to the application of the model to be required and second information related to performance requirements; the first information and the second information are used by the first network element to determine a first model in a preset registration model based on the first information and the second information; a second model is generated according to the second information and the first model; the second model is trained to obtain a trained third model; and the third model sent by the first network element is received. That is, when the first network element (such as a base station or terminal) generates an AI / ML model requirement, it sends a model request to the second network element. The second network element retrieves the existing model based on the model request, and generates a new lightweight model based on the existing model according to the performance requirements in the request. The newly generated lightweight model is trained using the historical data set generated by the inference calculation of the existing model in a real environment, and the trained new model is sent to the second network element for deployment. This solves the problem of missing AI / ML models on the edge side of the wireless network. Compared with the existing AI / ML deployment solution of the network management platform, the AI / ML deployment location can be made closer to the edge data generation source and the actual action location. Compared with other existing edge-side methods of obtaining AI / ML models (such as training from scratch, pruning large models, and quantization), it can effectively reduce human intervention, improve training and deployment speed, and improve model accuracy.

[0188] In one embodiment, the first information includes at least:

[0189] Problem label information; the problem label information represents the task goal of the required model for solving the problem;

[0190] Problem input information; the problem input information represents input variable information used by the demand model to solve the problem;

[0191] Problem output information; the problem output information represents the output variable information of the demand model used to solve the problem.

[0192] In an embodiment of the present application, the problem label information represents the task goal of the model to be required to solve the problem; wherein, the task goal can be determined according to the specific situation and is not limited here. As an example, the task goal can be to achieve 10-class classification of images, optimal allocation of network slice resources, etc. In actual applications, the problem label information can be simply referred to as a problem label, which is used to indicate the scenario and problem of model application, that is, the task goal of the model strategy, such as achieving 10-class classification of images, optimal allocation of network slice resources, etc.

[0193] The problem input information represents the input variable information used by the model to be required to solve the problem; wherein, the input variable information can be determined according to the specific situation and is not limited here. As an example, the input variable information can include but is not limited to: the name of the variable, the description of the data (such as the physical meaning). In practical applications, the problem input information can be simply referred to as problem input, which is used to indicate the input variable information of the problem model used by the AI ​​model to solve, including but not limited to: the name of the variable, the description of the data (such as the physical meaning).

[0194] The problem output information represents the output variable information of the model to be required for solving the problem; wherein, the output variable information can be determined according to the specific situation and is not limited here. As an example, the output variable information may include but is not limited to: the name of the variable, the description of the data (such as the physical meaning). In practical applications, the problem output information can be simply referred to as problem output, which is used to indicate the output variable information of the model for the problem to be solved by the AI ​​model, including but not limited to: the name of the variable, the description of the data (such as the physical meaning).

[0195] In one embodiment, the second information includes at least:

[0196] Model calculation amount limit; the model calculation amount limit represents the calculation amount limit of the model to be required;

[0197] Model parameter quantity limitation; the model parameter quantity limitation represents the parameter quantity limitation of the model to be required.

[0198] In an embodiment of the present application, the model computational amount limit represents the computational amount limit of the model to be required; wherein, the computational amount limit can be determined according to specific circumstances and is not limited here. As an example, the computational amount limit may include the number of calculations required for one inference calculation of the model, in units of FLOPs (floating point operations). The amount of calculation can be calculated based on the edge side network element latency requirement (in seconds, s) and the computing speed of the computing device (for example, the number of floating point operations per second, FLOPS). In practical applications, the model parameter amount limit is used to indicate the model computational amount limit required on the edge side, such as the number of calculations required for one inference calculation of the model, in units of FLOPs (floating point operations). The amount of calculation can be calculated based on the edge side network element latency requirement (in seconds, s) and the computing speed of the computing device (for example, the number of floating point operations per second, FLOPS).

[0199] The model parameter quantity limit characterizes the parameter quantity limit of the model to be required; wherein, the parameter quantity limit can be determined according to the specific situation and is not limited here. As an example, the parameter quantity limit may include the number of parameters in the model or the size of the storage space occupied. The parameter quantity size can be calculated based on the expected size of the storage space available to the edge side network element. In practical applications, the model parameter quantity limit is used to indicate the model parameter quantity limit required on the edge side, such as the number of parameters in the model or the size of the storage space occupied. The parameter quantity size can be calculated based on the expected size of the storage space available to the edge side network element.

[0200] For ease of understanding, the example model generation method of this application is specifically a lightweight model automatic generation and training method, which is applied to edge-side network elements. When edge-side network elements (such as base stations and terminals) generate AI / ML model requirements, they send model requests to network management units (such as SMO and Near RT RIC). The network management unit retrieves the existing model based on the model request, and generates a new lightweight model based on the existing model according to the performance requirements in the request. The newly generated lightweight model is trained using the historical data set generated by the inference calculation of the existing model in a real environment, and the trained new model is sent to the edge side for deployment. The specific process is as follows: Figure 3 shown. Figure 3 A schematic diagram of the process of generating a lightweight model according to an embodiment of the present application.

[0201] Step 1: Model request: The edge side network element sends a model request to the wireless network management unit, and the request content is shown in Table 1. Table 1 is a schematic table of the model request content of the embodiment of the present application.

[0202] Table 1

[0203]

[0204]

[0205] Step 2: Model retrieval: includes two steps: model screening and model selection.

[0206] Step 2.1: Initial model screening: The wireless network management unit receives a request from the edge network element, processes and retrieves the content of the request. The retrieval method is to perform text semantic retrieval on the question label in the registered models of the model management unit in the wireless network management unit, and match the input and output of the retrieved model with the input and output variables in the request one by one. If the match is successful, it will be used as a candidate model. The range of candidate models is as follows: Figure 4 shown. Figure 4 A schematic diagram of the candidate model range of an embodiment of the present application.

[0207] Step 2.2: Model selection: For the models in the candidate range, sort them according to the model accuracy performance index, select the top K models, and use character matching to determine whether there is a model with the same name as the expected basic model in the requirements. If so, this model is selected first. Otherwise, the model with the highest accuracy performance index is selected.

[0208] Step 3: Model generation: The management unit adjusts the existing model structure according to the model request received in step 1. Before each adjustment, the computational load and parameter load of the existing model are calculated based on the existing model structure, and the ratio r of the model computational load limit of the model request to the computational load of the existing model is calculated. c , the ratio of the model parameter limit requested by the model to the existing model parameter amount r p If both ratios are greater than 1, the adjustment is stopped. Adjustment methods include but are not limited to: (1) reducing the number of model layers (2) reducing the number of neurons in each layer of the model. During the adjustment process, only one method or multiple methods can be used. The adjustment range can be measured in proportion to the existing model size. The adjustment range is adjusted according to the size of the two calculated ratios. If the ratio is smaller, a larger adjustment range is selected.

[0209] Step 4: Model training: For the newly generated model S, use the historical data obtained by running the existing model T in the real environment Training. Divide the historical data into N entries according to each inference calculation. Each data entry contains the model T input data x for each inference calculation of the model. i , Model T output data (e.g., output categories of classification networks, output actions in DQN), outputs of key output layers of model T (For example, the softmax layer output vector of the classification network, the Q network output vector in DQN).

[0210] During training, the model is fed with data x i Input to the newly generated model S, the newly generated model S performs inference calculation and output model S output data Output of the key output layer of model S The new generated model S is calculated As prediction data, there is a model T output The loss is calculated as the true value. Methods for calculating loss include, but are not limited to, mean squared error loss, negative log-likelihood loss, and KL divergence loss. Based on the calculated loss, the parameters of the newly generated model S are updated and optimized using methods including, but not limited to, stochastic gradient descent (SGD) and the adaptive moment estimation method (ADAM). Training terminates when model accuracy no longer improves or when the model accuracy meets the model accuracy requirements received in step 1.

[0211] Step 5: Model Delivery: The trained model (structure and parameters) and its description file in step 4 are packaged into a format that is easy to transmit and delivered to the edge network element that issued the request. The contents of the description file are shown in Table 2. Table 2 is a schematic table of the contents of the description file in the embodiment of the present application.

[0212] Table 2

[0213]

[0214]

[0215] Example 1: Implementation of the required model and historical data registration based on Non-RT RIC.

[0216] In this embodiment, the necessary model and data registration process is described to illustrate the source of the model and training data in this solution, such as Figure 5 As shown, Figure 5 A schematic diagram of the process of registering the model and historical data in the embodiment of this application.

[0217] This part of the process can be roughly divided into five steps:

[0218] Step 1: Model Registration: The AI / ML model deployed in the network registers its model information with the AI / ML model management unit within the platform. This includes the model's identifier, the problem it solves, the wireless network scenario it applies to, the model's inputs and outputs, and more.

[0219] Step 2: Model performance monitoring: The model performance monitoring unit in the AI / ML service within the platform monitors the predicted value or control quantity output by the model during operation, and evaluates its accuracy and reliability by comparing it with the monitoring data in the actual environment.

[0220] Step 3: Model performance registration: register the model performance such as accuracy or reliability evaluated in step 2 as the model attribute information to the entry where the model is located in the model management unit, which is used for model selection when a model request is received in this solution.

[0221] Step 4: Model history data registration: During the actual operation of the model, the model input data x generated by each inference calculation of the model is i , model output data (e.g. output categories of classification networks, output actions in DQN), output data of key output layers of the model The data (e.g., the softmax layer output vector of the classification network, the Q network output vector in DQN) are regularly packaged and registered with the data management module in the platform as a historical data set associated with the model. The registration information includes: the corresponding model identifier that generated the data set, the timestamp of each inference data (or the time period when the data set was generated and the time interval between each inference calculation task), etc.

[0222] Step 5: Model historical data storage: After the historical data registration is completed, the historical data set is stored in the database within the platform for training the newly generated model in this solution.

[0223] Example 2: Automatic generation and training of lightweight models based on Non-RT RIC.

[0224] In this embodiment, a lightweight model automatic generation and training method is implemented in an O-RAN architecture using a rApp in a Non-RT RIC as a management unit to implement the overall process and the generation of new models. The management unit uses the AI / ML model management unit in the platform to register the model information of the existing models on the platform. The management unit uses the data management unit in the platform to manage the historical data generated by the model operation. The data warehouse in the platform provides storage capacity for historical data. The management unit uses the AI / ML training service capability in the platform to train new models. The implementation process is as follows: Figure 6 As shown, Figure 6This is a schematic diagram of the automatic generation and training process of a lightweight model based on Non-RT RIC in an embodiment of the present application.

[0225] Step 1: Model request: The base station creates an AI / ML model request based on the needs of the AI / ML service and sends the model request to the SMO where the Non-RT RIC is located through the 01 interface. The content of the request is shown in Table 1.

[0226] Step 2: Model screening and model selection: The management unit in the Non-RT RIC obtains the model request received by the OAM-related function of the SMO through the platform's internal interface. Based on the model request, it performs a text semantic search on the question label in the registered models of the AI / ML model management unit in the RIC platform, and matches the input and output of the retrieved model with the input and output variables in the request one by one. If the match is successful, it will be used as a candidate model. The range of candidate models is as follows: Figure 4 As shown in the figure, the models in the candidate range are sorted by model accuracy performance index, and the top K models are selected. Among these K models, character matching is used to determine whether there is a model with the same name as the expected basic model in the requirement. If so, this model is selected first. Otherwise, the model with the highest accuracy performance index is selected.

[0227] Step 3: Transferring an existing model: After the management unit in the Non-RT RIC selects the AI / ML model, the AI / ML model management unit in the platform informs the rApp where the model is located, and the rApp transfers the model to the management unit.

[0228] Step 4: New model generation: After receiving the existing AI / ML model, the management unit adjusts the existing model structure according to the model request received in step 1. Before each adjustment, the computational load and parameter amount of the existing model are calculated based on the existing model structure, and the ratio r of the model computational load limit of the model request to the computational load of the existing model is calculated. c , the ratio of the model parameter limit requested by the model to the existing model parameter amount r p If both ratios are greater than 1, the adjustment is stopped. Adjustment methods include but are not limited to: (1) reducing the number of model layers (2) reducing the number of neurons in each layer of the model. During the adjustment process, only one method or multiple methods can be used, and the adjustment range can be measured in proportion to the size of the existing model. The adjustment range is adjusted according to the size of the two calculated ratios. If the ratio is smaller, the larger adjustment range is selected. After the adjustment is completed, the new model S is obtained.

[0229] Step 5: Model training request: After the model adjustment is completed, the management unit initiates a model training request to the AI / ML training service within the platform. The training request contains the model to be trained (the newly generated model S) and the training dataset to be used, as well as the training stop condition (for example, the model accuracy requirement determined by the request in step 1, or directly specifying the number of training rounds).

[0230] Step 6: Training dataset request: The AI / ML training service within the platform requests the corresponding data from the data management unit within the platform based on the training dataset requested in the training request.

[0231] Step 7: Data request response and data transmission: The database in the platform retrieves the corresponding historical data set based on the data request received by the data management unit Send a data request response to the AI / ML training service within the platform and pass the historical dataset

[0232] Step 8: Model training: The AI / ML training service inputs the model into the data x i Input to the new generative model S to be trained, the new generative model S performs inference calculation and outputs the output data of the model S Output of the key output layer of model S The new generated model S is calculated As prediction data, there is a model T output The loss is calculated using the true value. Methods for calculating loss include, but are not limited to, mean squared error loss, negative log-likelihood loss, and KL divergence loss. Based on the calculated loss, the parameters of the newly generated model S are updated and optimized using methods including, but not limited to, stochastic gradient descent (SGD) and the adaptive moment estimation method (ADAM). Training terminates when model accuracy no longer improves or when the required accuracy meets the training stop condition specified in the training request.

[0233] Step 9: New model delivery: The AI / ML training service delivers the trained new model to the management unit.

[0234] Step 10: Send the new model and its description file: The management unit packages the trained model (structure and parameters) and its description file into a format that is easy to transmit and sends it to the base station that issued the request. The content of the description file is shown in Table 2.

[0235] In the embodiments of the present application, the services provided by the AI / ML models of the wireless network management platform (e.g., non-RT RIC) are far away from the data generation and final destination of the edge nodes, resulting in additional transmission delays and bandwidth overhead. Therefore, it is necessary to deploy lightweight AI / ML models at the edge nodes that meet strict computing and storage resource, energy consumption, and inference speed constraints.

[0236] In the embodiment of the present application, manually building and training a lightweight AI / ML model at the edge node requires a lot of manual code development and data labeling work, and the ability to train the model on the edge side is limited, which may require a long training time. Therefore, this method has a long development and deployment cycle, little available training data, and the model is limited by complexity, resulting in limited performance. The model training effect obtained by the general training method is poor.

[0237] In an embodiment of the present application, a complex AI / ML model deployed on a Near-RT RIC or Non-RT RIC platform is compressed through pruning, quantization, and other means, and the compressed model is used for edge-side inference calculations. This approach can also cause a loss in model accuracy.

[0238] The embodiments of the present application propose a lightweight model automatic generation and training method to solve the problem of missing AI / ML models on the edge side of wireless networks. Compared with the existing AI / ML deployment solutions of network management platforms, the AI / ML deployment location can be closer to the edge data generation source and actual action location. Compared with other existing edge-side methods of obtaining AI / ML models (building training from scratch, pruning large models, and quantization), it can effectively reduce human intervention, improve training and deployment speed, and improve model accuracy.

[0239] In order to implement the method of the embodiment of the present application, the embodiment of the present application also provides a model generation device, which is set on the first network element of the cloud management platform, such as Figure 7 As shown, Figure 7 Schematic diagram of a model generation device according to an embodiment of the present application; the device 700 includes:

[0240] The first receiving unit 701 is configured to receive request information of a model to be required sent by a second network element; the request information includes at least first information related to the application of the model to be required and second information related to performance requirements;

[0241] A determining unit 702 is configured to determine a first model in a preset registration model based on the first information and the second information;

[0242] A generating unit 703 is configured to generate a second model according to the second information and the first model;

[0243] The training unit 704 is used to train the second model to obtain a trained third model; the third model is used to send to the second network element.

[0244] In one embodiment, the first information includes at least:

[0245] Problem label information; the problem label information represents the task goal of the required model for solving the problem;

[0246] Problem input information; the problem input information represents input variable information used by the demand model to solve the problem;

[0247] Problem output information; the problem output information represents the output variable information of the demand model used to solve the problem.

[0248] In one embodiment, the second information includes at least:

[0249] Model calculation amount limit; the model calculation amount limit represents the calculation amount limit of the model to be required;

[0250] Model parameter quantity limitation; the model parameter quantity limitation represents the parameter quantity limitation of the model to be required.

[0251] In one embodiment, the first information also includes accuracy requirement information; the determination unit 702 is further used to perform text semantic retrieval on the first information in a preset registration model to obtain multiple first candidate models; determine multiple second candidate models that meet the accuracy requirement information among the multiple candidate models; sort the multiple second candidate models according to the accuracy performance index of each second candidate model to obtain a sorting result of the multiple second candidate models; and determine the first model based on the sorting result and the second information.

[0252] In one embodiment, the generation unit 703 is further used to calculate a first computational quantity and a first parameter quantity of the first model; determine a first ratio of the model computational quantity limit to the first computational quantity and a second ratio of the model parameter quantity limit to the first parameter quantity; determine whether the first ratio and the second ratio are both greater than a preset threshold; if the first ratio and the second ratio are not both greater than the preset threshold, adjust the first model; if the first ratio and the second ratio are both greater than the preset threshold, stop adjusting the first model and generate a second model.

[0253] In one embodiment, adjusting the first model includes at least one of the following: reducing the number of layers of the first model; reducing the number of neurons in each layer of the first model.

[0254] In one embodiment, the device 700 further includes an adjustment unit; wherein,

[0255] The determining unit 702 is further configured to determine a ratio between the first ratio and the second ratio;

[0256] The adjustment unit is used to adjust the amplitude of the first model based on the ratio.

[0257] In one embodiment, the training unit 704 is further used to obtain a historical data set based on a training request for the second model; use the historical data set to train the second model to obtain a loss function of the second model; and optimize the parameters of the second model based on the loss function to obtain a trained third model.

[0258] In one embodiment, the loss function includes at least one of the following:

[0259] Mean squared error loss function;

[0260] Negative log-likelihood loss function;

[0261] KL divergence loss function.

[0262] In one embodiment, the parameter optimization method includes one of the following:

[0263] Stochastic Gradient Descent (SGD);

[0264] Root mean square propagation RMSProp;

[0265] Adaptive gradient AdaGrad;

[0266] Adaptive Matrix Estimation Adam.

[0267] In order to implement the method on the second network element side of the embodiment of the present application, the embodiment of the present application also provides a model generation device, which is set on the second network element of the cloud management platform, such as Figure 8 As shown, Figure 8 800 is a schematic diagram of a structure of another model generation device according to an embodiment of the present application. The device 800 includes:

[0268] The second sending unit 801 is configured to send request information for a model to be required to the first network element; the request information includes at least first information related to the application of the model to be required and second information related to performance requirements; the first information and the second information are used by the first network element to determine a first model in a preset registered model based on the first information and the second information; generate a second model based on the second information and the first model; and train the second model to obtain a trained third model.

[0269] The second receiving unit 802 is configured to receive the third model sent by the first network element.

[0270] In one embodiment, the first information includes at least:

[0271] Problem label information; the problem label information represents the task goal of the required model for solving the problem;

[0272] Problem input information; the problem input information represents input variable information used by the demand model to solve the problem;

[0273] Problem output information; the problem output information represents the output variable information of the demand model used to solve the problem.

[0274] In one embodiment, the second information includes at least:

[0275] Model calculation amount limit; the model calculation amount limit represents the calculation amount limit of the model to be required;

[0276] Model parameter quantity limitation; the model parameter quantity limitation represents the parameter quantity limitation of the model to be required.

[0277] It should be noted that the model generation device provided in the above embodiment only uses the division of the above-mentioned program modules as an example to illustrate model generation. In actual applications, the above-mentioned processing can be assigned to different program modules as needed, that is, the internal structure of the device can be divided into different program modules to complete all or part of the above-mentioned processing. In addition, the model generation device provided in the above embodiment and the model generation method embodiment are based on the same concept. The specific implementation process is detailed in the method embodiment and will not be repeated here.

[0278] Based on the hardware implementation of the above-mentioned program module, an embodiment of the present application also provides a first network element, including: a first processor and a first memory for storing a computer program that can be run on the processor, wherein the first processor is used to implement the steps in the model generation method provided in the above-mentioned embodiment when running the computer program.

[0279] Based on the hardware implementation of the above-mentioned program module, an embodiment of the present application also provides a second network element, including: a second processor and a second memory for storing a computer program that can be run on the processor, wherein the second processor is used to implement the steps in the model generation method provided in the above-mentioned embodiment when running the computer program.

[0280] Correspondingly, an embodiment of the present application provides a computer program product, including a computer program, on which a computer program is stored. When the computer program is executed by a processor, the steps in the model generation method provided in the above embodiment are implemented.

[0281] Correspondingly, an embodiment of the present application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the model generation method provided in the above embodiment.

[0282] It should be noted that the description of the above storage medium and device embodiments is similar to the description of the above method embodiments and has similar beneficial effects as the method embodiments. For technical details not disclosed in the storage medium and device embodiments of this application, please refer to the description of the method embodiments of this application for understanding.

[0283] It should be noted that Figure 9 This is a structural diagram of the first network element in the embodiment of the present application, such as Figure 9 As shown, the first network element 900 includes: a first processor 901 and a first memory 903 . Optionally, the first network element 900 may further include a first communication interface 902 .

[0284] It is understood that the first memory 903 can be a volatile memory or a non-volatile memory, or can include both volatile and non-volatile memories. Among them, the non-volatile memory can be a read-only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPROM), a magnetic random access memory (FRAM), a flash memory, a magnetic surface memory, an optical disc, or a compact disc read-only memory (CD-ROM); the magnetic surface memory can be a magnetic disk memory or a magnetic tape memory. The volatile memory can be a random access memory (RAM), which is used as an external cache. By way of example and not limitation, many forms of RAM are available, such as static random access memory (SRAM), synchronous static random access memory (SSRAM), dynamic random access memory (DRAM), synchronous dynamic random access memory (SDRAM), double data rate synchronous dynamic random access memory (DDRSDRAM), enhanced synchronous dynamic random access memory (ESDRAM), synchronous link dynamic random access memory (SLDRAM), and direct rambus random access memory (DRRAM).The first memory 903 described in the embodiments of the present application is intended to include but is not limited to these and any other suitable types of memories.

[0285] The methods disclosed in the above embodiments of the present application can be applied to the first processor 901 or implemented by the first processor 901. The first processor 901 may be an integrated circuit chip with signal processing capabilities. During implementation, each step of the above method can be completed by hardware integrated logic circuits in the first processor 901 or by software instructions. The above-mentioned first processor 901 can be a general-purpose processor, a digital signal processor (DSP), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The first processor 901 can implement or execute the various methods, steps, and logic block diagrams disclosed in the embodiments of the present application. The general-purpose processor can be a microprocessor or any conventional processor, etc. The steps of the methods disclosed in the embodiments of the present application can be directly embodied as being executed by a hardware decoding processor, or can be executed by a combination of hardware and software modules in the decoding processor. The software module can be located in a storage medium, which is located in the first memory 903. The first processor 901 reads the information in the first memory 903 and completes the steps of the above method in combination with its hardware.

[0286] It should be noted that Figure 10 This is a schematic diagram of the structure of the terminal in the embodiment of the present application. Figure 10 As shown, the second network element 1000 includes: a second processor 1001 and a second memory 1003 . Optionally, the second network element 1000 may further include a second communication interface 1002 .

[0287] It is understood that the second memory 1003 can be a volatile memory or a non-volatile memory, or can include both volatile and non-volatile memories. Among them, the non-volatile memory can be a read-only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPROM), a magnetic random access memory (FRAM), a flash memory, a magnetic surface memory, an optical disc, or a compact disc read-only memory (CD-ROM); the magnetic surface memory can be a magnetic disk memory or a magnetic tape memory. The volatile memory can be a random access memory (RAM), which is used as an external cache. By way of example and not limitation, many forms of RAM are available, such as static random access memory (SRAM), synchronous static random access memory (SSRAM), dynamic random access memory (DRAM), synchronous dynamic random access memory (SDRAM), double data rate synchronous dynamic random access memory (DDRSDRAM), enhanced synchronous dynamic random access memory (ESDRAM), synchronous link dynamic random access memory (SLDRAM), and direct rambus random access memory (DRRAM).The second memory 1003 described in the embodiments of the present application is intended to include but is not limited to these and any other suitable types of memories.

[0288] The methods disclosed in the above embodiments of the present application can be applied to the second processor 1001 or implemented by the second processor 1001. The second processor 1001 may be an integrated circuit chip with signal processing capabilities. During implementation, each step of the above method can be completed by hardware integrated logic circuits in the second processor 1001 or by software instructions. The above second processor 1001 may be a general-purpose processor, a digital signal processor (DSP), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The second processor 1001 can implement or execute the various methods, steps, and logic block diagrams disclosed in the embodiments of the present application. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the methods disclosed in the embodiments of the present application can be directly implemented as being executed by a hardware decoding processor, or can be executed by a combination of hardware and software modules in the decoding processor. The software module can be located in a storage medium located in the second memory 1003. The second processor 1001 reads the information in the second memory 1003 and completes the steps of the above method in combination with its hardware.

[0289] In order to implement the method provided in the embodiment of the present application, the embodiment of the present application also provides a model generation system, such as Figure 11 As shown, Figure 11 This is a schematic diagram of the structure of the model generation system of the embodiment of the present application, which includes: a first network element 1101 and a second network element 1102.

[0290] Here, it should be noted that the specific processing procedures of the first network element 1101 and the second network element 1102 have been described in detail above and will not be repeated here.

[0291] In an exemplary embodiment, the device may be implemented by one or more application-specific integrated circuits (ASICs), DSPs, programmable logic devices (PLDs), complex programmable logic devices (CPLDs), field-programmable gate arrays (FPGAs), general-purpose processors, controllers, microcontrollers (MCUs), microprocessors, or other electronic components to perform the aforementioned method.

[0292] It should be understood that "one embodiment" or "an embodiment" mentioned throughout the specification means that the specific features, structures or characteristics related to the embodiment are included in at least one embodiment of the present application. Therefore, "in one embodiment" or "in an embodiment" appearing throughout the specification does not necessarily refer to the same embodiment. In addition, these specific features, structures or characteristics can be combined in one or more embodiments in any suitable manner. It should be understood that in the various embodiments of the present application, the size of the serial numbers of the above-mentioned processes does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present application. The above-mentioned serial numbers of the embodiments of the present application are for description only and do not represent the advantages and disadvantages of the embodiments.

[0293] It should be noted that, in this application, the terms "comprises," "includes," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a process, method, article, or apparatus comprising a series of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of other identical elements in the process, method, article, or apparatus comprising the element.

[0294] The methods disclosed in the several method embodiments provided in this application can be arbitrarily combined without conflict to obtain new method embodiments.

[0295] The features disclosed in the several product embodiments provided in this application can be arbitrarily combined without conflict to obtain new product embodiments.

[0296] The features disclosed in the several method or device embodiments provided in this application can be arbitrarily combined without conflict to obtain new method embodiments or device embodiments.

[0297] The above is merely an embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of this application. Therefore, the scope of protection of this application should be based on the scope of protection of the claims.

[0298] It should be noted that: "first", "second", etc. are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence.

[0299] In addition, the technical solutions described in the embodiments of the present application can be arbitrarily combined without conflict.

[0300] The above description is merely a preferred embodiment of the present application and is not intended to limit the scope of protection of the present application.

Claims

1. A model generation method, characterized in that: Applied to the first network element, including: receiving request information of the model to be required sent by the second network element; the request information at least including first information related to the application of the model to be required and second information related to performance requirements; Determining a first model in a preset registration model based on the first information and the second information; generating a second model according to the second information and the first model; The second model is trained to obtain a trained third model; the third model is used to send to the second network element.

2. The method according to claim 1, characterized in that The first information at least includes: Problem label information; the problem label information represents the task goal of the required model for solving the problem; Problem input information; the problem input information represents input variable information used by the demand model to solve the problem; Problem output information; the problem output information represents the output variable information of the demand model used to solve the problem.

3. The method according to claim 1, characterized in that The second information at least includes: Model calculation amount limit; the model calculation amount limit represents the calculation amount limit of the model to be required; Model parameter quantity limitation; the model parameter quantity limitation represents the parameter quantity limitation of the model to be required.

4. The method according to claim 1, wherein The first information further includes accuracy requirement information; and determining the first model in a preset registration model based on the first information and the second information includes: Performing a text semantic search on the first information in a preset registration model to obtain a plurality of first candidate models; Determining a plurality of second candidate models that meet the accuracy requirement information from the plurality of candidate models; Sorting the plurality of second candidate models according to the accuracy performance index of each second candidate model to obtain a ranking result of the plurality of second candidate models; The first model is determined based on the ranking result and the second information.

5. The method according to claim 3, characterized in that Generating a second model according to the second information and the first model includes: Calculating a first calculation amount and a first parameter amount of the first model; determining a first ratio of the model computation amount limit to the first computation amount and a second ratio of the model parameter amount limit to the first parameter amount; Determining whether both the first ratio and the second ratio are greater than a preset threshold; If both the first ratio and the second ratio are not greater than the preset threshold, adjusting the first model; When both the first ratio and the second ratio are greater than the preset threshold, adjusting the first model is stopped and a second model is generated.

6. The method according to claim 5, wherein adjusting the first model comprises at least one of the following: reducing the number of layers of the first model; Reduce the number of neurons in each layer of the first model.

7. The method according to claim 5, characterized in that The method further comprises: determining a ratio of the first ratio to the second ratio; The magnitude of the first model is adjusted based on the magnitude of the ratio.

8. The method according to claim 1, characterized in that The step of training the second model to obtain a trained third model includes: Obtaining a historical data set based on a training request for the second model; Training the second model using the historical data set to obtain a loss function of the second model; Parameters of the second model are optimized based on the loss function to obtain a trained third model.

9. The method according to claim 8, characterized in that The loss function includes at least one of the following: Mean squared error loss function; Negative log-likelihood loss function; KL divergence loss function.

10. The method according to claim 8, characterized in that The parameter optimization method includes one of the following: Stochastic Gradient Descent (SGD); Root mean square propagation RMSProp; Adaptive gradient AdaGrad; Adaptive Matrix Estimation Adam.

11. A model generation method, characterized in that: Applied to the second network element, including: Sending request information for a model to be required to a first network element; the request information includes at least first information related to the application of the model to be required and second information related to performance requirements; the first information and the second information are used by the first network element to determine a first model from a preset registered model based on the first information and the second information; generating a second model based on the second information and the first model; and training the second model to obtain a trained third model. Receive the third model sent by the first network element.

12. The method according to claim 11, characterized in that The first information at least includes: Problem label information; the problem label information represents the task goal of the required model for solving the problem; Problem input information; the problem input information represents input variable information used by the demand model to solve the problem; Problem output information; the problem output information represents the output variable information of the demand model used to solve the problem.

13. The method according to claim 11, characterized in that The second information at least includes: Model calculation amount limit; the model calculation amount limit represents the calculation amount limit of the model to be required; Model parameter quantity limitation; the model parameter quantity limitation represents the parameter quantity limitation of the model to be required.

14. A model generation device, characterized in that: Set in the first network element, including: A first receiving unit is configured to receive request information of the model to be required sent by the second network element; the request information at least includes first information related to the application of the model to be required and second information related to performance requirements; a determining unit, configured to determine a first model in a preset registered model based on the first information and the second information; a generating unit, configured to generate a second model according to the second information and the first model; A training unit is used to train the second model to obtain a trained third model; the third model is used to send to the second network element.

15. A lightweight model generation device, characterized in that: Set in the second network element, including: A sending unit, configured to send request information for a model to be required to a first network element; the request information including at least first information related to an application of the model to be required and second information related to performance requirements; the first information and the second information being used by the first network element to determine a first model from a preset registered model based on the first information and the second information; generating a second model based on the second information and the first model; and training the second model to obtain a trained third model. The second receiving unit is used to receive the third model sent by the first network element.

16. A first network element, characterized in that: include: a first communication interface and a first processor; wherein, The first communication interface is configured to receive request information of the model to be required sent by the second network element; the request information at least including first information related to the application of the model to be required and second information related to performance requirements; The first processor is used to determine a first model in a preset registration model based on the first information and the second information; generate a second model according to the second information and the first model; train the second model to obtain a trained third model; and the third model is used to send to the second network element.

17. A second network element, characterized in that: include: A second communication interface and a second processor; wherein, The second communication interface is used to send request information of the model to be required to the first network element; the request information includes at least first information related to the application of the model to be required and second information related to performance requirements; the first information and the second information are used by the first network element to determine the first model in the preset registration model based on the first information and the second information; generate a second model according to the second information and the first model; train the second model to obtain a trained third model; and receive the third model sent by the first network element.

18. A first network element, characterized in that: include: a first processor and a first memory for storing a computer program capable of being executed on the processor, Wherein, when the first processor is used to run the computer program, it executes the steps of the method according to any one of claims 1 to 10.

19. A second network element, characterized in that: include: a second processor and a second memory for storing a computer program capable of being executed on the processor, Wherein, when the second processor is used to run the computer program, it executes the steps of the method according to claims 11 to 13.

20. A computer program product comprising a computer program, characterized in that The computer program implements the steps of the method according to any one of claims 1 to 10 when executed by a processor; or, the computer program implements the steps of the method according to claims 11 to 13 when executed by a processor.

21. A storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 10 are implemented, or the steps of the method according to claims 11 to 13 are implemented.

Images