Node selection method and communication apparatus

Abstract

The present application provides a node selection method and a communication apparatus. The method comprises: sending first information to N terminal devices, the first information comprising information about a first data set of a network device; receiving second information from each terminal device among the N terminal devices, the second information comprising a statistical distance of each terminal device among the N terminal devices, the statistical distance being used for indicating the distance between a second data set of each terminal device among the N terminal devices and the first data set; determining M terminal devices on the basis of the second information; sending indication information, the indication information being used for indicating the M terminal devices; receiving the second data set from each terminal device among the M terminal devices; and performing model training on the basis of the second data set of each terminal device among the M terminal devices. On the basis of the described method, the network device can selectively sample training data sets, improving the convergence speed of model training, and reducing communication overhead.

Description

A node selection method and communication device

[0001] This application claims priority to the Chinese patent application filed with the China Patent Office on January 22, 2024, with application number 202410092807.6 and title “A Node Selection Method and Communication Device,” the entire contents of which are incorporated by reference into this application. Technical Field

[0002] The present application relates to the field of communication technology, and more specifically, to a node selection method and a communication device. Background Art

[0003] To improve the intelligence and automation of networks, inference models, such as artificial intelligence (AI) and machine learning (ML) models, are being applied to a growing number of technical fields. Models are typically acquired through training. In some application scenarios, model training and the collection of the required training datasets may be deployed on different devices. For example, multiple terminal devices can collect the training datasets required for model training and send them to the network device, which then performs model training to update the model. However, if all terminal devices send the training datasets to the network device, this would incur excessive communication losses and computational overhead.

[0004] In existing solutions, network devices randomly sample training datasets collected by terminal devices for training, but this results in slow model convergence. Therefore, a new approach is urgently needed that allows network devices to selectively sample training datasets, improve model convergence speed, and reduce communication overhead. Summary of the Invention

[0005] The present application provides a node selection method and a communication device that enable network devices to selectively sample training data sets, improve the convergence speed of model training, and reduce communication overhead.

[0006] In the first aspect, a node selection method is provided. The method can be executed by a network device, or can also be executed by a component of the network device (such as a chip or a chip system or a circuit). There is no limitation on this. For the sake of ease of description, the following is explained using the example of execution by a network device.

[0007] The method includes: sending first information to N terminal devices, the first information including information of a first data set of a network device; receiving second information from each of the N terminal devices, the second information including a statistical distance of each of the N terminal devices, the statistical distance of each of the N terminal devices being used to indicate a distance between a second data set of each of the N terminal devices and the first data set; determining M terminal devices based on the second information; sending indication information, the indication information being used to indicate the M terminal devices; receiving a second data set from each of the M terminal devices; and performing model training based on the second data set of each of the M terminal devices.

[0008] Wherein, N is an integer greater than 1; M is a positive integer, and M is less than or equal to N.

[0009] Based on the above technical solution, the terminal device is selected by considering the statistical distance between the training data set collected by each terminal device among multiple terminal devices and the data set of the network device, that is, the training data set collected by the terminal device is selected. The network device can selectively sample the training data set, effectively utilize the data in the distributed environment, improve the convergence speed of model training, and reduce communication overhead.

[0010] In combination with the first aspect, in certain implementations of the first aspect, sending indication information includes: sending a broadcast signal, the broadcast signal including first indication information, the first indication information being used to indicate a set of M terminal devices; or, sending second indication information to each of the M terminal devices, the second indication information being used to indicate each terminal device.

[0011] Based on the above technical solution, the network device notifies the selected M terminal devices through broadcast signals or separate indication information, so that they know that they have been selected and can participate in model training, avoiding unnecessary communication. Since only the selected terminal devices need to participate in the transmission of data in model training, the system communication overhead is reduced.

[0012] In combination with the first aspect, in certain implementations of the first aspect, M terminal devices are determined based on the second information, including: determining the M terminal devices based on the statistical distance ratio corresponding to each of the N terminal devices, the statistical distance ratio being the ratio of the statistical distance of each of the N terminal devices to the maximum statistical distance among the statistical distances of the N terminal devices.

[0013] In some possible implementations, the M terminal devices are the M terminal devices with the largest statistical distance ratio among the N terminal devices.

[0014] Based on the above technical solution, the network device can select the terminal device according to the statistical distance between the data sets. Selecting the terminal device with the largest statistical distance ratio, that is, the data set with a larger sampling error, can accelerate the convergence of the model.

[0015] In combination with the first aspect, in certain implementations of the first aspect, the second information also includes the communication capability of each of the N terminal devices, and the communication capability includes at least one of the following: the distance between each of the N terminal devices and the network device, or the energy consumption per unit data volume of each of the N terminal devices.

[0016] In some possible implementations, selecting M terminal devices based on the second information includes: determining the M terminal devices based on the statistical distance ratio and communication capability ratio corresponding to each of the N terminal devices; wherein the statistical distance ratio is the ratio of the statistical distance of each of the N terminal devices to the maximum statistical distance among the statistical distances of the N terminal devices, and the communication capability ratio is the ratio of the communication capability of each of the N terminal devices to the maximum communication capability among the communication capabilities of the N terminal devices.

[0017] In some possible implementations, the M terminal devices are the M terminal devices with the largest target value among the N terminal devices, and the target value is the sum of the statistical distance ratio and the communication capability ratio, or the weighted sum.

[0018] Based on the above technical solution, when selecting a terminal device, the network device comprehensively considers the two factors of statistical distance and communication capability, and flexibly adjusts the weights of the two factors in the selection of the terminal device according to a weighted method. This comprehensive selection method can balance the impact of statistical distance and communication capability, and can adapt to the different requirements for communication capability and data similarity in different scenarios.

[0019] In combination with the first aspect, in some implementations of the first aspect, the information of the first data set is the first data set, and the statistical distance is the distance between the second data set of each terminal device in the N terminal devices and the first data set.

[0020] In some possible implementations, the information of the first data set is statistical information of the first data set, and the statistical distance is the distance between the statistical information of the second data set and the statistical information of the first data set of each terminal device in the N terminal devices.

[0021] Optionally, the distance includes at least one of the following: Fréchet distance; maximum mean difference; Wasserstein distance; Kolmogorov-Smirnov distance.

[0022] Based on the above technical solution, the distance between data sets is obtained according to the information of the corresponding first data set, the computing power of multiple terminal devices is reasonably utilized, the overhead is reduced, and there are multiple options for distance calculation methods, which helps to consider the impact of different distance calculation methods on model training.

[0023] On the second aspect, a node selection method is provided. The method can be executed by a terminal device, or can also be executed by a component of the terminal device (such as a chip or a chip system or a circuit). There is no limitation on this. For the sake of ease of description, the following is explained using the example of execution by a terminal device.

[0024] The method includes: receiving first information from a network device, the first information including information of a first data set of the network device; sending second information, the second information including a statistical distance, the statistical distance being used to indicate the distance between the second data set of the terminal device and the first data set, the second information being used to determine M terminal devices; receiving indication information, the indication information being used to indicate the M terminal devices; and sending a second data set in response to the indication information, the second data set being used for model training.

[0025] Wherein, M is a positive integer.

[0026] Based on the above technical solution, statistical distance helps to obtain the difference information between the data set of the terminal device and the data set of the network device, providing a statistical distance measurement for network device selection. It can effectively utilize data in a distributed environment, improve the convergence speed of model training, and reduce communication overhead.

[0027] In combination with the second aspect, in certain implementations of the second aspect, receiving indication information includes: receiving a broadcast signal, the broadcast signal including first indication information, the first indication information being used to indicate a set of M terminal devices; or, receiving second indication information, the second indication information being used to indicate the terminal device.

[0028] Based on the above technical solution, by receiving broadcast signals or receiving separate indication information, it is learned that it has been selected by the network device to participate in model training, thereby transmitting the data set required for model training, avoiding unnecessary communication, and thus reducing overall communication overhead.

[0029] In combination with the second aspect, in some implementations of the second aspect, the second information also includes communication capability, and the communication capability includes at least one of the following: the distance between the terminal device and the network device, or the energy consumption per unit data volume of the terminal device.

[0030] Optionally, the terminal device obtains the distance from the network device, or obtains its own energy consumption per unit data volume.

[0031] Based on the above technical solution, the terminal device obtains its own communication capabilities, which helps network devices to comprehensively consider factors such as communication efficiency and energy consumption when making selections.

[0032] In combination with the second aspect, in certain implementations of the second aspect, the information of the first data set is the first data set, and the statistical distance is the distance between the second data set and the first data set.

[0033] In some possible implementations, the information of the first data set is statistical information of the first data set, and the statistical distance is the distance between the statistical information of the second data set and the statistical information of the first data set.

[0034] Optionally, the distance includes at least one of the following: Fréchet distance; maximum mean difference; Wasserstein distance; Kolmogorov-Smirnov distance.

[0035] Based on the above technical solution, the terminal device can calculate the distance between data sets based on the information of the corresponding first data set, reasonably utilize the computing power of multiple terminal devices, reduce overhead, and there are multiple options for distance calculation methods, which helps to consider the impact of different distance calculation methods on model training.

[0036] In a third aspect, the present application provides a communications device, which may include a module for performing the methods / operations / steps / actions described in the first or second aspects, wherein the module may be a hardware circuit, software, or a combination of hardware circuits and software. The communications device may be a first network element or a second network element, or a chip or circuit for the first or second network element.

[0037] In one implementation, the communication apparatus is a communication device. Exemplarily, the communication device may include a communication unit and / or a processing unit. The communication unit may be a transceiver or an input / output interface; the processing unit may be at least one processor. Alternatively, the transceiver may be a transceiver circuit. Alternatively, the input / output interface may be an input / output circuit.

[0038] In another implementation, the apparatus is a chip, chip system, or circuit used in a communication device. When the apparatus is a chip, chip system, or circuit used in a terminal device, the communication unit may be an input / output interface, interface circuit, output circuit, input circuit, pin, or related circuit on the chip, chip system, or circuit; and the processing unit may be at least one processor, processing circuit, or logic circuit.

[0039] In one example, the communication device is a network device, or the communication device is a chip, chip system, or circuit provided in the network device. In another example, the communication device is a terminal device, or the communication device is a chip, chip system, or circuit provided in the terminal device.

[0040] In a fourth aspect, the present application provides a communication device, comprising a processor configured to execute a computer program or instruction stored in a memory to perform the method provided in the first or second aspect, or any implementation thereof. Optionally, the communication device further comprises the memory. The communication device may be a network device or a terminal device, or a chip or circuit for a network device or a terminal device.

[0041] In a fifth aspect, the present application provides a communication device, comprising a processor and a communication interface, configured to execute the method provided in the first aspect or the second aspect, or any implementation thereof. Exemplarily, the communication interface may be a transceiver, a hardware circuit, a bus, a module, a pin, or other type of communication interface.

[0042] In a sixth aspect, a chip or chip system is provided, comprising a processor configured to execute a program or instruction to implement the method of the first aspect or any possible implementation of the first aspect, and to implement the method of the second aspect or any possible implementation of the second aspect. Optionally, the chip may further comprise a memory configured to store the program or instruction. Optionally, the chip may further comprise the transceiver.

[0043] In the seventh aspect, a computer-readable storage medium is provided, which includes instructions. When the instructions are executed by a processor, the method in the first aspect or any possible implementation of the first aspect is implemented, and the method in the second aspect or any possible implementation of the second aspect is implemented.

[0044] In an eighth aspect, a computer program is provided, which, when executed on a computer, enables the computer to execute the above-mentioned first aspect or second aspect, or the method provided in any implementation of the first aspect or second aspect.

[0045] In the ninth aspect, a computer program product is provided, which includes computer program code or instructions. When the computer program code or instructions are executed, the method in the first aspect or any possible implementation of the first aspect is implemented, and the method in the second aspect or any possible implementation of the second aspect is implemented.

[0046] In a tenth aspect, a communication system is provided, comprising one or more of the aforementioned network devices and terminal devices. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] FIG1 is a schematic diagram of a communication system provided in an embodiment of the present application.

[0048] FIG2 is a schematic diagram of a fully connected neural network provided in an embodiment of the present application.

[0049] FIG3 is a schematic diagram of distributed learning provided in an embodiment of the present application.

[0050] FIG4 is a schematic diagram of another distributed learning provided in an embodiment of the present application.

[0051] FIG5 is a node selection method provided in an embodiment of the present application.

[0052] FIG6 is another node selection method provided in an embodiment of the present application.

[0053] FIG7 is a schematic diagram of a communication device provided in an embodiment of the present application.

[0054] FIG8 is a schematic diagram of another communication device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0055] The technical solution in this application will be described below with reference to the accompanying drawings.

[0056] The technical solutions of the embodiments of the present application can be applied to various communication systems, such as: long term evolution (LTE) system, LTE frequency division duplex (FDD) system, LTE time division duplex (TDD) system, worldwide interoperability for microwave access (WiMAX) communication system, fifth generation (5G) system or new radio (NR) and future communication systems, vehicle-to-X (V2X), where V2X may include vehicle to network (V2N), vehicle to vehicle (V2V), vehicle to infrastructure (V2I), vehicle to pedestrian (V2P), etc., long term evolution technology for vehicle communication (LTE-V), Internet of Vehicles, machine type communication (MTC), Internet of Things (IoT), long term evolution technology for machine-to-machine communication (LTE-V), etc. evolution-machine (LTE-M), machine to machine (M2M), etc.

[0057] A device in a communication system can send signals to or receive signals from another device. Signals can include information, signaling, or data. The term "device" can also be replaced by an entity, network entity, node, communication device, communication module, network element, communication node, etc. This application uses devices as an example for description.

[0058] FIG1 shows a schematic diagram of a communication system provided by an embodiment of the present application. As shown in FIG1 , the communication system 100 may include one or more network devices, such as the network device 10 and the network device 20 shown in FIG1 . The communication system 100 may also include one or more terminal devices, such as the terminal device 30 and the terminal device 40 shown in FIG1 . The four terminal devices, namely, the network device 10, the network device 20, the terminal device 30, and the terminal device 40, may communicate with each other via a wireless link to exchange information.

[0059] In one implementation, the communication system may also include an AI entity. The network device may forward AI model-related data reported by the terminal device to the AI ​​entity. The AI ​​entity then performs AI-related operations such as training dataset construction and model training, and provides the output of AI-related operations, such as the trained AI model, model evaluation, and test results, to the network device. In another implementation, the AI ​​entity may be located within network device 10 or network device 20, that is, as a module of network device 10 or network device 20.

[0060] In practical applications, a network device can serve one or more terminal devices simultaneously. A terminal device can also access one or more network devices simultaneously. It is understood that network devices and terminal devices can also be referred to as communication devices, and the embodiments of this application do not limit the number of network devices and terminal devices included in the communication system.

[0061] Exemplarily, a network device may be a network-side device with wireless transceiver functions. A network device may be a device in a radio access network (RAN) that provides wireless communication functions for terminal devices, referred to as a RAN device. For example, the network device may be a base station, an evolved NodeB (eNodeB), a next generation NodeB (gNB) in a 5G mobile communication system, a base station that has been subsequently evolved by 3GPP, a transmission reception point (TRP), an access node in a WiFi system, a wireless relay node, a wireless backhaul node, etc. In communication systems using different radio access technologies (RAT), the names of devices with base station functions may be different. For example, in an LTE system, it may be called an eNB or eNodeB, and in a 5G system or NR system, it may be called a gNB. This application does not limit the specific name of the base station. The network device may include one or more co-located or non-co-located transmission and reception points. For another example, the network device may include at least one of the following items: one or more centralized units (CU), one or more distributed units (DU), and one or more radio units (RU). In different systems, CU (or CU-CP and CU-UP), DU or RU may also have different names, but those skilled in the art can understand their meanings. For example, the radio access network may also be an open radio access network (O-RAN) architecture. In the ORAN system, CU may also be called O-CU (open CU), DU may also be called O-DU, CU-CP may also be called O-CU-CP, CU-UP may also be called O-CU-UP, and RU may also be called O-RU. Any of the CU (or CU-CP, CU-UP), DU and RU in this application may be implemented by a software module, a hardware module, or a combination of a software module and a hardware module. Exemplarily, the functions of the CU may be implemented by one entity or different entities. For example, the functions of the CU are further divided, that is, the control plane and the user plane are separated and implemented through different entities, namely the control plane CU entity (i.e., CU-CP entity) and the user plane CU entity (i.e., CU-UP entity). The CU-CP entity and the CU-UP entity can be coupled with the DU to jointly complete the functions of the access network device.For example, the CU is responsible for processing non-real-time protocols and services, and implementing the functions of the radio resource control (RRC) and packet data convergence protocol (PDCP) layers. The DU is responsible for processing physical layer protocols and real-time services, and implementing the functions of the radio link control (RLC) layer, the media access control (MAC) layer, and the physical (PHY) layer. In this way, some functions of the wireless access network device can be implemented through multiple network function entities. These network function entities can be network elements in hardware devices, software functions running on dedicated hardware, or virtualized functions instantiated on a platform (for example, a cloud platform). The first device may also include an active antenna unit (AAU). The AAU implements some physical layer processing functions, radio frequency processing, and related functions of the active antenna. Since the information of the RRC layer will eventually become the information of the PHY layer, or be converted from the information of the PHY layer, under this architecture, high-layer signaling, such as RRC layer signaling, can also be considered to be sent by the DU, or by the DU+AAU. It can be understood that the network device can be a device including one or more of a CU node, a DU node, and an AAU node. In addition, the CU can be divided into a network device in the access network (radio access network, RAN), or the CU can be divided into a network device in the core network (core network, CN), and this application does not limit this. For example, in the vehicle to everything (V2X) technology, the access network device can be a road side unit (RSU). The multiple access network devices in the communication system can be base stations of the same type or different types. The base station can communicate with the terminal device or communicate with the terminal device through a relay station. In an embodiment of the present application, the device for realizing the function of the network device can be the network device itself, or it can be a device that can support the network device to realize the function, such as a chip system or a combination device or component that can realize the function of the access network device, and the device can be installed in the network device. In an embodiment of the present application, the chip system can be composed of chips, or it can include chips and other discrete devices.

[0062] Exemplarily, a terminal device may be a user-side device with wireless transceiver capabilities, and may be a fixed device, a mobile device, a handheld device (e.g., a mobile phone), a wearable device, an in-vehicle device, or a wireless device built into the above devices (e.g., a communication module, a modem, or a chip system, etc.). The terminal device is used to connect people, objects, machines, etc., and can be widely used in various scenarios, such as: cellular communications, device-to-device (D2D) communications, V2X communications, machine-to-machine / machine-type communications (M2M / MTC) communications, the Internet of Things, virtual reality (VR), augmented reality (AR), industrial control, self-driving, remote medical, smart grid, smart furniture, smart office, smart wearables, smart transportation, smart city, drones, robots, and other scenarios. Exemplarily, the terminal device can be a handheld terminal in cellular communication, a communication device in D2D, an IoT device in MTC, a surveillance camera in intelligent transportation and smart cities, or a communication device on a drone, etc. The terminal device may sometimes be referred to as user equipment (UE), user terminal, user device, user unit, user station, terminal, access terminal, access station, UE station, remote station, mobile device or wireless communication device, etc. The terminal device may also be a terminal device in an IoT system. IoT is an important part of the future development of information technology. Its main technical feature is to connect objects to the network through communication technology, thereby realizing an intelligent network of human-machine interconnection and object-to-object interconnection. In an embodiment of the present application, IoT technology can achieve massive connections, deep coverage, and terminal power saving through, for example, narrowband (NB) technology. In an embodiment of the present application, the device for realizing the function of the terminal device may be a terminal device, or a device that can support the terminal device to realize the function, such as a chip system or a combination device or component that can realize the function of the terminal device. The device can be installed in the terminal device.

[0063] The network equipment and terminal devices can be deployed on land, including indoors or outdoors, handheld or vehicle-mounted; they can also be deployed on water; they can also be deployed in the air on aircraft, balloons, and satellites. The embodiments of this application do not limit the scenarios in which the network equipment and terminal devices are located.

[0064] For example, the communication system 100 may further include an application function (AF) network element, which is a control plane network function provided by the operator network and is used to provide application layer information; the communication system 100 may further include a session management function (SMF) network element, which is a control plane network function provided by the operator network. In the embodiment of the present application, when the communication system 100 includes the AF network element and the SMF network element, the AF can send service-related information to the network device via the SMF.

[0065] It should be understood that the number and type of each device in the communication system shown in Figure 1 are for illustration only, and the present application is not limited to this. In actual applications, the communication system may also include more terminal devices, more network devices, more positioning devices, and other network elements, such as core network devices, and / or network elements for implementing artificial intelligence functions.

[0066] It should also be understood that the division of the modes, situations, categories and embodiments in the embodiments of the present application is only for the convenience of description and should not constitute a special limitation. The features of various modes, categories, situations and embodiments can be combined without contradiction.

[0067] It should also be understood that in the various embodiments of the present application, the size of the serial number of each process 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.

[0068] It should be noted that in the embodiments of the present application, "pre-setting", "pre-definition", "pre-configuration", etc. can be achieved by pre-saving corresponding codes, tables or other methods that can be used to indicate relevant information in the device (for example, including terminal devices and network devices). This application does not limit its specific implementation method, such as the pre-configured information in the embodiments of the present application.

[0069] For the sake of clarity, some terms in the embodiments of this application are explained below.

[0070] 1) Model: A function learned from data that can achieve a specific function / mapping. The model can be obtained based on artificial intelligence (AI) or machine learning (ML) technology, and therefore, it can also be called an artificial intelligence / AI model, a machine learning / ML model, etc. Commonly used algorithms for generating AI / ML models include: supervised learning, unsupervised learning, and reinforcement learning, and the corresponding models can be called supervised learning models, unsupervised learning models, and reinforcement learning models. For example, a supervised learning model can be a classification model, a prediction model, a regression model, etc., and an unsupervised learning model can be a clustering model. In addition, the model can also be obtained based on neural network (NN) technology, and this model can also be called a neural network model, a deep learning model, etc.

[0071] 2) Training Dataset: Data used for model training, validation, and testing in machine learning. The quantity and quality of this data will impact the effectiveness of machine learning. Training data can include the input to the AI ​​model, or the input and target output of the AI ​​model. The target output is the target value of the AI ​​model's output, also known as the true output value, output truth value, label, or label sample.

[0072] It should be understood that a training dataset is a collection of training samples, each of which is an input to a neural network, and the training dataset is used for model training. The training dataset is one of the most important parts of machine learning. The training process of machine learning is essentially to learn certain features of the training dataset so that the difference between the output of the neural network and the ideal target value (i.e., the label or output true value) is minimized under the training dataset. Normally, even if the same network structure is used, the weights and outputs of the neural networks trained using different training datasets are different. Therefore, the composition and selection of the training dataset determine, to some extent, the performance of the trained neural network.

[0073] 3) Supervised Learning: The goal of supervised learning is to learn the mapping between inputs (data) and outputs (labels) given a training dataset (consisting of multiple pairs of inputs and outputs). Furthermore, the goal is to apply this mapping to data outside the training dataset. The training dataset is a collection of correct input-output pairs.

[0074] Taking a fully connected neural network as an example, as shown in Figure 2, a schematic diagram of a fully connected neural network provided in an embodiment of the present application is shown. A fully connected neural network is also called a multilayer perceptron (MLP). An MLP consists of an input layer (left), an output layer (right), and multiple hidden layers (middle). Each layer contains several nodes, called neurons. Among them, the neurons in two adjacent layers are connected to each other.

[0075] Considering neurons in two adjacent layers, the output h of a neuron in the next layer is the weighted sum of all neurons x connected to it in the previous layer, after passing through the activation function. This can be expressed in a matrix as: h = f(wx + b) (Formula 1)

[0076] Among them, w is the weight matrix, b is the bias vector, and f is the activation function.

[0077] Then the output of the neural network can be recursively expressed as: y = f n (w n f n-1 (…)+b n ) (Formula 2)

[0078] Simply put, a neural network can be understood as a mapping relationship from an input data set to an output data set. Neural networks are usually randomly initialized, and the process of using existing data to obtain this mapping relationship from random w and b is called neural network training.

[0079] The specific training method is to use the loss function to evaluate the output of the neural network, and back-propagate the error. The gradient descent method is used to iteratively optimize w and b until the loss function reaches the minimum value. The gradient descent process can be expressed as:

[0080] Among them, θ is the parameter to be optimized (such as w and b above), L is the loss function, and η is the learning rate, which controls the step size of gradient descent.

[0081] The process of backpropagation can use the chain rule of partial derivatives, that is, the gradient of the previous layer parameters can be recursively calculated from the gradient of the next layer parameters. The formula can be expressed as:

[0082] With the advent of the big data era, every device (including terminal devices and network equipment) generates massive amounts of raw data in various forms every day. This data is created as "islands" and exists in every corner of the world. Traditional centralized learning requires each edge device to transmit local data to a central server. The central server then uses the collected data to train and learn models. However, this architecture has been increasingly limited by the following factors:

[0083] (1) Edge devices are widely distributed in various regions and corners of the world. These devices continuously generate and accumulate huge amounts of raw data at a rapid rate. If the center needs to collect raw data from all edge devices, it will bring huge communication losses and computing power requirements.

[0084] (2) As real-life scenarios become more complex, more and more learning tasks require edge devices to make timely and effective decisions and provide feedback. Traditional centralized learning involves uploading large amounts of data, which results in significant latency and makes it unable to meet the real-time requirements of actual task scenarios.

[0085] (3) Considering issues such as industry competition, user privacy and security, and complex administrative procedures, centralized data integration will face increasing resistance and constraints. Therefore, system deployment will increasingly tend to store data locally, while edge devices themselves perform local computing.

[0086] It should be understood that the above-mentioned central end can be the network device mentioned below, and the edge device can be the terminal device mentioned below, and this application does not limit this.

[0087] The technical solution provided by this application is introduced below.

[0088] FIG3 shows a schematic diagram of a distributed learning provided by an embodiment of the present application. As shown in FIG3 , in distributed learning, the network device can randomly select the terminal device to send the central data set, for example, the central data set D (including the data set D a 、D b 、D c ) is sent to the multiple terminal devices (such as terminal device #a, terminal device #b, and terminal device #c in Figure 3), and each of the multiple terminal devices uses local computing resources to train the model and sends the trained model W (including model W a 、W b 、W c ) is uploaded to the network device. In this learning architecture, the network device holds the entire dataset, eliminating the need for the end device to collect local datasets. This learning architecture leverages the computing power of the end device to help train the model, offloading the computational burden on the network device.

[0089] It should be understood that the central dataset is a dataset collected by network devices for model training.

[0090] While this distributed learning architecture reduces the computational burden on network devices to a certain extent, it also has some drawbacks. First, after each training round, the terminal device needs to upload its locally trained model to the network device. This involves transferring a large amount of model parameters, which can result in significant communication overhead, especially when the model is large or network conditions are poor. Second, because the network device only updates the model uploaded by the terminal device and does not use local data, the model convergence speed may be relatively slow. The network device may not be able to directly perceive the local data of each terminal device, which may affect the global performance of the model.

[0091] FIG4 shows a schematic diagram of another distributed learning method provided by an embodiment of the present application. As shown in FIG4 , the algorithm flow is roughly as follows:

[0092] (1) Network equipment initializes the model to be trained And broadcast it to all terminal devices, such as terminal device #1,...terminal device #M,...terminal device #N in the figure.

[0093] (2) Taking the Mth terminal device (terminal device #M) as an example, in the t∈[1,T]th round, terminal device #M is based on the local data set D M For the received global model Train to get local training results And report it to the network device.

[0094] (3) The network device aggregates and collects the local training results from all (or part) of the terminal devices. Assume that the terminal set that uploads the local model in round t is S t , the network device will use the number of samples of the corresponding terminal device as the weight to perform weighted averaging to obtain a new global model. The specific update rule is M∈[1,N], where D′ M Represents the dataset D M The network device then sends the latest version of the global model to The broadcast is sent to all terminal devices for a new round of training.

[0095] (4) Repeat steps (2) and (3) until the model finally converges or the number of training rounds reaches the upper limit.

[0096] For terminal device #M, in addition to reporting the local model to the network device In addition, the local gradient of training can be reported. The network device averages the local gradient and updates the global model according to the direction of the average gradient.

[0097] Therefore, in the above-mentioned distributed learning framework, the training data set exists at the terminal device, that is, the terminal device collects the local data set, performs local training, and reports the local results (model or gradient) obtained from the training to the network device. The network device itself does not have a data set and is only responsible for fusing the training results of the terminal device to obtain a global model and send it to the terminal device. However, due to the periodic fusion of the entire model through the federated averaging algorithm, the convergence speed is slow and the convergence performance has certain defects; secondly, this distributed learning usually requires the network device to send the model to multiple terminal devices. In scenarios with large model scales, the communication overhead will be too high, and the computing, storage and communication capabilities of the device will be high.

[0098] Therefore, how to select terminal devices to participate in training is a complex issue that may affect the effectiveness of model training. The embodiment of the present application provides a node selection method that can improve the convergence speed and the utilization efficiency of data and computing power during model training. At the same time, in the solution provided by the present application, the interaction between the network device and the terminal device is only the training data set, which can significantly reduce the communication overhead when the number of model parameters is large, and can also reduce the requirements for device computing, storage and communication capabilities.

[0099] FIG5 shows a node selection method 400 provided in an embodiment of the present application. As shown in FIG5 , the method 400 may include steps S410 to S470.

[0100] S410, the network device sends first information to the terminal device; correspondingly, the terminal device receives the first information.

[0101] Exemplarily, the RRC layer of the network device sends the first information to the terminal device.

[0102] The first information includes information of the central data set; the information of the central data set may be the central data set or statistical information of the central data set.

[0103] It should be understood that the central data set is the data set of the network device, and the data set of the network device is used for model training. For the sake of ease of description, this application uses the central data set to describe the data set of the network device, without limitation.

[0104] Optionally, the central data set is a data set obtained by the network device according to the updated model after the last round of model training is completed.

[0105] Optionally, the statistical information of the central data set includes at least one of the following: empirical distribution, statistical moment, compression, downsampling, etc.

[0106] For example, the empirical distribution is the probability distribution function of a given dataset obtained by normalizing the sample frequencies. It is the cumulative distribution function of the relative frequencies of each data point in the dataset and is used to describe the actual observed data distribution.

[0107] For example, a statistical moment is a statistic that describes the shape of a data distribution. Common moments include the mean, variance, skewness, and kurtosis. The mean describes the center of a data set, the variance measures the dispersion of the data, the skewness measures the symmetry of the distribution, and the kurtosis describes how peaked or flat the distribution is.

[0108] For example, compression can refer to the process of reducing the dimensionality of a dataset or extracting key features. This can be achieved through various compression algorithms, principal component analysis (PCA), feature selection, and other methods. Compression helps reduce data storage space and computational costs while preserving key information about the data's importance.

[0109] For example, downsampling involves extracting subsets from the original dataset, reducing the number of data points. This helps improve computational efficiency when processing large amounts of data, and in some cases, downsampling can occur without significantly losing important information about the dataset. In statistical analysis and machine learning, downsampling can be used to quickly explore data or reduce the size of the dataset during training.

[0110] It should be understood that the network device can send the first information to multiple terminal devices in the form of broadcast, and this application does not limit the number of the above-mentioned terminal devices.

[0111] Optionally, the first information further includes indication information for indicating a calculation method of the statistical distance.

[0112] It should be understood that the statistical distance is used to represent the distance between two data sets, and the calculation method of the statistical distance includes but is not limited to Fréchet distance, maximum mean difference, Wasserstein distance, and Kolmogorov-Smirnov distance.

[0113] To facilitate understanding of the solution of this application, the following briefly introduces the calculation method of the statistical distance.

[0114] (1) Fréchet distance

[0115] The Fréchet distance is a description of path space similarity. Specifically, given two curves or paths, each curve is considered as a function defined in a common parameter space. The Fréchet distance considers the movement of the two paths in the parameter space and, when comparing them, considers both the consistency of position and orientation.

[0116] Calculating the Fréchet distance involves finding the optimal match between two functions. It can be thought of as placing two points between two paths and then moving them along the paths so that they always remain on their corresponding paths. The Fréchet distance is the maximum distance between the two points during this entire movement.

[0117] The specific calculation method can be expressed by mathematical formula: F(P,Q)=inf αβ max t∈[0,1] d(P(α(t)),Q(β(t))) (Formula 5)

[0118] Among them, P and Q are two paths; α and β are mappings of parameterized paths P and Q; t is a variable in the parameter space; d is the distance metric defined between points in the parameter space; inf means taking the infimum of all possible α and β; max means taking the maximum value in the entire parameter space.

[0119] If we use Fréchet distance to measure the differences between datasets, we can regard each data point in the dataset as a point in the path, and the entire dataset as a path. That is, we can choose a metric to define the distance between data points, which usually depends on the type and nature of the dataset. For example, we can use Euclidean distance, Manhattan distance, or other distance metrics suitable for the data type; then we can introduce parameterization for the path of each dataset by corresponding data points to points in the parameter space, which means that each data point will have a corresponding position in the parameter space, thus forming a path.

[0120] For example, assuming there are two data sets X and Y, data sets X and Y can be regarded as paths X and Y, and the formula of Fréchet distance can be expressed as:

[0121] Therefore, the Fleche distance provides a quantitative way to measure the difference between two data sets. A smaller Fleche distance indicates that the data sets have higher shape similarity, while a larger Fleche distance indicates that the data sets have lower shape similarity.

[0122] (2) Maximum mean difference

[0123] Maximum Mean Discrepancy (MMD) is a method used to measure the distance between two different but related random variable distributions. MMD is widely used in machine learning, especially in transfer learning and generative model evaluation.

[0124] The core idea of ​​MMD is to measure the difference between two distributions by comparing their sample means. Its calculation formula is usually based on the kernel method, and one common definition is to calculate the MMD between two distributions P and Q by using the kernel function k:

[0125] Where f is the norm of the kernel function k in this space, and sup represents the supremum. The intuitive explanation of this formula is that MMD finds a maximizing function f over all possible kernel functions k, so that the mean difference between P and Q under this function is maximized.

[0126] In transfer learning, MMD is often used as a loss function to measure the distribution difference between the source and target domains. By minimizing MMD, the distributions of the two domains can be brought closer together, thereby improving the model's generalization performance on the target domain. Furthermore, MMD can be used to compare the distance between two datasets, especially when considering their distribution differences. Below is a simple example illustrating how to use MMD to compare two datasets.

[0127] Suppose there are two data sets, denoted as data set X and data set Y, each of which contains multiple samples. For example, the sample point of data set X is x, and the sample point of data set Y is y. We can compare the distribution differences between the two data sets by following the steps below:

[0128] First, an appropriate kernel function k is selected. This often depends on the nature of the data; for example, a Gaussian kernel is a good choice in many cases. Finally, the selected kernel function k is used to calculate the MMD, i.e., the MMD between the two datasets X and Y. The MMD can be calculated using the kernel method of the sample mean. The formula is as follows:

[0129] Therefore, the smaller the MMD value, the more similar the distributions of the two data sets are. Conversely, the larger the MMD value, the greater the difference in distributions between the two data sets.

[0130] (3) Wasserstein distance

[0131] The Wasserstein distance is used to compare two distributions and is particularly well-suited for distributions of different shapes and masses. Its core concept is to find the minimum cost required to transform one distribution into another, where the cost is weighted by the distance between the two distributions. One of the main advantages of the Wasserstein distance is that it is more robust to small variations between distributions and non-overlapping regions. The following example illustrates how to use the Wasserstein distance to compare two datasets.

[0132] Suppose there are two data sets, denoted as data set X and data set Y, each of which contains multiple samples. For example, the sample point of data set X is x, and the sample point of data set Y is y. We can compare the distribution differences between the two data sets by following the steps below:

[0133] First, each data set is considered as a probability distribution, where the probability of each sample occurring is equal to its frequency in the data set. Second, the distance between two data sets is calculated using the definition of the Wasserstein distance. The calculation of the Wasserstein distance involves finding an optimal coupling in the sample space so that the total cost of converting one distribution to another is minimized. The mathematical expression is as follows:

[0134] where X and Y are the distributions of two datasets, ∏(X,Y) is the set of all possible couplings between X and Y, and γ(x,y) is the cost function defined in the sample space.

[0135] Therefore, the value of the Wasserstein distance represents the minimum cost of converting from one distribution to another. A smaller Wasserstein distance generally indicates a higher similarity between the two data sets, while a larger distance indicates a greater difference between the two data sets.

[0136] (4) Kolmogorov-Smirnov (KS) distance

[0137] The KS distance is a statistic used to compare the dissimilarity between two probability distributions. It measures the maximum vertical distance between two cumulative distribution functions (CDFs). The KS distance is widely used to test whether two samples come from the same distribution, but it can also be used to compare the similarity of any two probability distributions. A KS distance of zero indicates that the two distributions are identical; a larger KS distance indicates that the two distributions are significantly different.

[0138] The KS distance is calculated as follows: First, for each sample set, sort it by numerical value and then calculate the cumulative distribution function for each value. This generates a CDF for each sample set. Finally, for each sample value, calculate the absolute difference between the two CDFs and find the largest difference. This largest difference is the KS distance. Mathematically, the KS distance can be calculated as:

[0139] Among them, F1(x) and F2(x) are the cumulative distribution functions of the two sample sets, and x is the sample point.

[0140] If we use KS distance to measure the difference between datasets, we can regard the datasets as probability distributions and thus indirectly use KS distance to compare the similarities between them. Suppose there are two datasets, represented as dataset X and dataset Y, each dataset contains multiple samples. We can compare the distribution differences between the two datasets by the following steps: obtain sample data from datasets X and Y; regard each dataset as a probability distribution and calculate their empirical distribution functions; use KS distance Computes the maximum difference between two empirical distribution functions.

[0141] Among them, F X (x) and F Y (x) are the empirical distribution functions of data sets X and Y respectively, and x is the sample point.

[0142] Therefore, when the KS distance is small, it means that the empirical distribution functions of the two data sets are relatively similar, that is, the difference between the two data sets is small; if the KS distance is large, it means that there is a large difference between the two data sets.

[0143] S420: The terminal device determines second information based on the first information.

[0144] The second information includes a statistical distance, and the terminal device determines the statistical distance between the local data set of the terminal device and the central data set according to the first information.

[0145] Optionally, the terminal device calculates the distance between the local data set and the central data set according to the calculation method of the statistical distance indicated by the network device and the central data set.

[0146] Optionally, the terminal device calculates the distance between the statistical information of the local data set and the statistical information of the central data set according to the calculation method of the statistical distance indicated by the network device and the statistical information of the central data set.

[0147] Optionally, the statistical information includes at least one of the following: empirical distribution, statistical moment, compression, downsampling, etc.

[0148] It should be understood that the local data set is the data set collected by the terminal device.

[0149] Optionally, the second information also includes the communication capability of the terminal device.

[0150] It should be understood that the communication capability of a terminal device refers to the various capabilities and performances that the terminal device possesses during communication, including the terminal device's ability to connect to network devices, supported communication protocols, transmission rate, power consumption efficiency, location services, and acquisition of sensor data.

[0151] For example, the communication capability of the terminal device can be measured by the distance from the network device.

[0152] It should be understood that the Signal-to-Interference plus Noise Ratio (SINR) is a metric used to measure signal quality in wireless communications. It is the ratio of the received signal power to the received interference and noise power, and is mathematically expressed as: Since the noise N and interference I are relatively stable, It can be seen that in order to ensure that the receiving end (e.g., network equipment) receives a similar SINR, the transmitting end (e.g., terminal equipment) transmit power P tx and the square of the distance x 2 should remain constant. Therefore, we can use x 2 To indicate the communication capability of the terminal device, that is, the distance between the terminal device and the network device can be used to indicate the communication capability of the terminal device.

[0153] Exemplarily, the communication capability of the terminal device may also be measured by the communication quality of the terminal device or the energy consumption per unit data volume of the terminal device.

[0154] It should be understood that communication capability can be measured by measuring indicators such as the quality, speed and power consumption of actual communication, that is, calculating the energy consumption of sending unit data volume based on historical communication conditions. This is because the historical communication conditions of the terminal device provide the performance of the terminal device under different conditions.

[0155] For example, a mobile application can use the mobile device's API (such as Android's Network Stats Manager) to monitor the application's data usage and data transmission speed. Such measurements can be used to calculate the energy consumption required to send a unit of data under specific network conditions.

[0156] S430, the terminal device sends second information to the network device, and correspondingly, the network device receives the second information.

[0157] The second information includes a statistical distance.

[0158] Optionally, the second information also includes communication capabilities.

[0159] S440: The network device determines the terminal device according to the received second information.

[0160] Optionally, the second information includes a statistical distance, and the network device determines the terminal device according to the statistical distance.

[0161] Exemplarily, the network device receives second information from each of N terminal devices, where the second information includes a statistical distance of each of the N terminal devices. The network device determines and selects M terminal devices with the largest statistical distance ratios based on the statistical distance ratios corresponding to each of the N terminal devices.

[0162] The statistical distance ratio is a ratio of the statistical distance of each terminal device among the N terminal devices to the maximum statistical distance among the statistical distances of the N terminal devices.

[0163] Based on this solution, network devices can select terminal devices based on the statistical distance between data sets. Terminal devices with the largest statistical distance ratio, i.e., data sets with larger sampling errors, are selected to accelerate model convergence.

[0164] Optionally, the second information further includes communication capability, and the network device determines the terminal device based on the statistical distance and communication capability.

[0165] Exemplarily, a network device receives second information from each of N terminal devices, where the second information includes a statistical distance of each of the N terminal devices and a communication capability of each of the N terminal devices. The network device determines M terminal devices having the largest target value based on the statistical distance ratio and the communication capability ratio corresponding to each of the N terminal devices. The target value is the sum, or a weighted sum, of the statistical distance ratio and the communication capability ratio.

[0166] Among them, the statistical distance ratio is the ratio of the statistical distance of each terminal device among the N terminal devices to the maximum statistical distance among the statistical distances of the N terminal devices, and the communication capability ratio is the ratio of the communication capability of each terminal device among the N terminal devices to the maximum communication capability among the communication capabilities of the N terminal devices.

[0167] Based on this solution, terminal devices are selected by comprehensively considering both communication capability and statistical distance, and the weights of the two factors in terminal device selection are flexibly adjusted according to a weighted approach. This comprehensive selection method can balance the impact of statistical distance and communication capability, and can adapt to the different requirements for communication capability and data similarity in different scenarios.

[0168] It should be noted that the specific calculation example of the above step S440 will be introduced in detail in method 500 and will not be repeated here.

[0169] S450, the network device sends instruction information to the terminal device; correspondingly, the terminal device receives the instruction information from the network device.

[0170] Optionally, the network device may broadcast a selected set of terminal devices (such as terminal device #1, terminal device #2, ... terminal device #M in FIG6 ).

[0171] Optionally, the broadcast includes an identifier of the selected terminal device.

[0172] Optionally, the network device may send a selected signaling to each of the M selected terminal devices, where the selected signaling is used to indicate that the terminal device has been selected by the network device.

[0173] S460: The terminal device sends a data set to the network device. Correspondingly, the network device receives the data set from the terminal device.

[0174] For example, the terminal device first needs to collect a local dataset. That is, the local dataset is a set of data generated, collected, or stored locally by the terminal device, including user-generated information, sensor records, data generated by applications, etc. This data may include various types, such as text, images, audio, sensor readings, etc. The specific content depends on the application scenario and function of the terminal device. This application does not limit this. Secondly, the terminal device establishes a communication connection with the network device, encapsulates and sends the local dataset to the network device, and realizes data transmission and sharing.

[0175] S470: The network device performs model training based on the received data set.

[0176] Exemplarily, the network device receives a data set from each of the selected M terminal devices, and the network device performs model training based on the data set of each of the M terminal devices.

[0177] It should be understood that the network device trains the model based on the received data set. The training can be performed for a fixed number of rounds or stopped when the model converges.

[0178] Optionally, the network device trains a new model or a model saved after a previous round of training based on the received data set to obtain an updated model.

[0179] Optionally, the central dataset is updated according to the updated model, which may include adding new data, deleting outdated data, or updating existing data.

[0180] It should be understood that the data set described above (eg, a central data set, or a data set of a terminal device) may be part or all of the data set itself, or may be indicative information of the data set, and this application does not limit this.

[0181] It should be noted that the updated central data set is used for the next round of terminal device selection.

[0182] The solution of this application should be applicable to the scenario of multiple rounds of sampling. Assume that K rounds of sampling are set, and each round of sampling in the K rounds of sampling includes at least two terminal devices and one network device. For example, the cth round of sampling in the K rounds of sampling includes N terminal devices and one network device, and the i-th terminal device in the N terminal devices collects the local data set D i , and calculate the local data set D i With the central dataset D c The network device selects M terminal devices according to the statistical distance and communication capability corresponding to each terminal device in the N terminal devices, and sends indication information according to the selection, where the indication information is used to indicate the selected M terminal devices. The M terminal devices respond to the indication information and send the collected local data sets to the network device. The network device performs T rounds of training on the model based on the data sets of each terminal device in the M terminal devices to obtain a trained and updated model W c .

[0183] For another example, the c+1th round of sampling in the K rounds of sampling includes H terminal devices and one network device. The i-th terminal device in the H terminal devices collects the local data set D i , and calculate the local data set D i With the central dataset D c+1 The network device selects Z terminal devices according to the statistical distance and communication capability corresponding to each of the H terminal devices, and sends indication information according to the selection, where the indication information is used to instruct the selected Z terminal devices. The Z terminal devices send the collected local data sets to the network device in response to the indication information. The network device performs T rounds of training on the model based on the data sets of each of the Z terminal devices to obtain a trained and updated model W c+1 .

[0184] It should be understood that H in the c+1th round of sampling is smaller than N in the cth round of sampling; the central data set D in the c+1th round of sampling c+1 is the network device according to the model W in the cth round of sampling c The dataset to be updated.

[0185] The following describes the solution of this application by taking the cth round of sampling in the K rounds of sampling as an example.

[0186] FIG6 shows a schematic diagram of a node selection method 500 provided in an embodiment of the present application. The method 500 may be regarded as a specific expansion of the method 400 and may include but is not limited to steps S510 to S580.

[0187] S510, the network device sends first information to N terminal devices (such as terminal device #1, terminal device #2, ..., terminal device #M, ..., terminal device #N in Figure 6).

[0188] Correspondingly, each of the N terminal devices receives the first information from the network device.

[0189] The first information includes information of the central data set; the information of the central data set may be the central data set or statistical information of the central data set.

[0190] It should be understood that the central data set is the data set of the network device, and the data set of the network device is used for model training. For the sake of ease of description, this application uses the central data set to describe the data set of the network device, without limitation.

[0191] Optionally, the central data set is a data set obtained by the network device according to the updated model after the last round of model training is completed.

[0192] For example, the last round model is the model W obtained after the network device is trained in the c-1 round of sampling. c-1 For the convenience of description, the central data set of this round of sampling is denoted by D c Indicates that the central data set of the previous round of sampling is D c-1 express.

[0193] It should be understood that during the c-1 round of sampling, the network device sends the first information to R terminal devices, and determines to select J terminal devices based on the statistical distance and communication capability sent by each of the R terminal devices. The J terminal devices send local data sets to the network device for training the model W in the c-2 round of sampling. c-2 , after the model is updated, the model W is obtained c-1 , according to the model W c-1 Central dataset D c-1 Update to get the central dataset D for round c sampling c Among them, R in the c-1th round of sampling is greater than N in the cth round of sampling.

[0194] Optionally, the statistical information of the central data set includes at least one of the following: empirical distribution, statistical moment, compression, downsampling, etc.

[0195] For the convenience of description, the statistical information of the central data set is denoted by T(D c ) indicates that the local data set collected by terminal device #i is represented by D i Indicates that i = 1, 2, ... M, ... N. For example, the local data set collected by terminal device #1 is D1, the local data set collected by terminal device #2 is D2, ..., the local data set collected by terminal device #M is DM ,…, the local data set collected by terminal device #N is D N , but the application is not limited thereto.

[0196] Optionally, the first information further includes indication information for indicating a calculation method of the statistical distance, where the calculation method of the statistical distance includes: Fréchet distance, maximum mean difference, Wasserstein distance, or Kolmogorov-Smirnov distance.

[0197] S520, each of the N terminal devices determines a statistical distance F1, F2, ..., F corresponding to the first information. M ,…F N .

[0198] It should be understood that the terminal device #i calculates the central data set D based on the first information c With the local dataset D i The statistical distance F i , i=1,2,…M,…N.

[0199] Optionally, the first information is a central data set D c When terminal device #i calculates the central data set D c With the local dataset D i The statistical distance F i .

[0200] For example, when the first information indicates that the calculation method of the statistical distance is the Fréchet distance, the statistical distance F(1) is:

[0201] For example, when the first information indicates that the calculation method of the statistical distance is the maximum mean difference, the statistical distance F(2) is:

[0202] For example, when the first information indicates that the calculation method of the statistical distance is Wasserstein distance, the statistical distance F(3) is:

[0203] Exemplarily, the first information indicates that the calculation method of the statistical distance is Kolmogorov-Smirnov distance, and the statistical distance F(4) is:

[0204] in, and They are the central dataset D c With the local dataset D i The empirical distribution of , x is the sample point.

[0205] Optionally, when the first information is the statistical information of the central data set, the terminal device #i calculates the statistical information T(D c ) and the statistical information T(D) of the local data set i ) statistical distance.

[0206] It should be understood that the central dataset D c Statistics of T(D c ) and the local dataset D i Statistics of T(D i ) can also reflect the statistical distance between the two data sets.

[0207] For example, when the first information indicates that the calculation method of the statistical distance is the Fréchet distance, the statistical distance F(5) is:

[0208] For example, when the first information indicates that the calculation method of the statistical distance is the maximum mean difference, the statistical distance F(6) is:

[0209] For example, when the first information indicates that the calculation method of the statistical distance is Wasserstein distance, the statistical distance F(7) is:

[0210] Exemplarily, the first information indicates that the calculation method of the statistical distance is Kolmogorov-Smirnov distance, and the statistical distance F(8) is:

[0211] in, and They are the central dataset D c With the local dataset D i The empirical distribution of , x is the sample point.

[0212] Optionally, the terminal device may further calculate a statistical distance between the central dataset and the local dataset according to a configuration method of statistical information.

[0213] Optionally, in S530, each of the N terminal devices obtains a corresponding communication capability C1, C2, ..., C M ,…C N .

[0214] For example, the terminal device #i obtains its own communication capability C e , i=1,2,…M,…N.

[0215] Optionally, the square of the distance between terminal device #i and the network device Can be used to represent communication capability C e.

[0216] It should be understood that SINR is an indicator used to measure the signal quality in wireless communications. It is the ratio of the received signal power to the received interference and noise power, and is mathematically expressed as: Since the noise N and interference I are relatively stable, It can be seen that in order to ensure that the receiving end receives a similar SINR, the transmitting end's transmission power P tx and the square of the distance x 2 should remain constant. Therefore, we can use x 2 To indicate the communication capability of the terminal device, that is, the distance between the terminal device and the network device can be used to indicate the communication capability of the terminal device.

[0217] Optionally, the communication capability of the terminal device may also be measured by the communication quality of the terminal device or the energy consumption per unit data volume of the terminal device.

[0218] It should be understood that the communication capability can be measured by measuring indicators such as the quality, speed and power consumption of actual communication, that is, calculating the energy consumption of sending unit data volume based on historical communication conditions. This is because the historical communication conditions of the terminal device provide the performance of the node under different conditions.

[0219] S540, each of the N terminal devices sends second information to the network device, where the second information includes a statistical distance F i Optionally, the second information also includes communication capability C i .

[0220] Correspondingly, the network device receives the second information from each of the N terminal devices.

[0221] S550: The network device determines M terminal devices according to the second information, where M≤N.

[0222] Optionally, the network device determines the M terminal devices according to the statistical distance of each of the N terminal devices.

[0223] Exemplarily, the network determines the maximum statistical distance max among the statistical distances of the N terminal devices according to the statistical distance of each terminal device of the N terminal devices. N F(D c ,D N ); Then the network device traverses N terminal devices and obtains the ratio of the statistical distance of each terminal device to the maximum statistical distance Select the statistical distance ratio The maximum number of terminal devices is M. Among them, D c is the central dataset, D i is the data set of terminal device #i, for example, DN It is the data set of terminal device #N.

[0224] Exemplarily, the network determines the maximum statistical distance max among the statistical distances of the N terminal devices according to the statistical distance of each terminal device of the N terminal devices. N F(T(D c )); Then the network device traverses N terminal devices and obtains the ratio of the statistical distance of each terminal device to the maximum statistical distance Select the statistical distance ratio The maximum number of M terminal devices. Among them, T(D c ) is the central dataset D c Statistics of T(D i ) is the data set D of terminal device #i i Statistics, for example, T(D N ) is the data set D of terminal device #N N Statistics.

[0225] Based on this solution, network devices can select terminal devices based on the statistical distance between data sets. Terminal devices with the largest statistical distance ratio, i.e., data sets with larger sampling errors, are selected to accelerate model convergence.

[0226] Optionally, the network device selects M terminal devices according to the statistical distance and communication capability of each of the N terminal devices.

[0227] For example, the network first determines the maximum statistical distance max among the statistical distances of the N terminal devices according to the statistical distance of each terminal device of the N terminal devices. N F(D c ,D N ) and the maximum communication capability max among the communication capabilities of N terminal devices N C N ; Then the network device traverses N terminal devices and obtains the ratio of the statistical distance of each terminal device to the maximum statistical distance And the ratio of the communication capability of each terminal device in N terminal devices to the maximum communication capability Select target value The maximum number of terminal devices is M. Among them, D c is the central dataset, D i is the data set of terminal device #i, for example, D N is the data set of terminal device #N; C i is the communication capability of terminal device #i, for example, C N It is the communication capability of terminal device #N.

[0228] For example, the network first determines the maximum statistical distance max among the statistical distances of the N terminal devices according to the statistical distance of each terminal device of the N terminal devices. N F(T(D c ),T(D N )) and the maximum communication capability max among the communication capabilities of N terminal devices N C N ; Then the network device traverses N terminal devices and obtains the ratio of the statistical distance of each terminal device to the maximum statistical distance And the ratio of the communication capability of each terminal device in N terminal devices to the maximum communication capability Select target value The maximum number of M terminal devices. Among them, T(D c ) is the central dataset D c Statistics of T(D i ) is the data set D of terminal device #i i Statistics, for example, T(D N ) is the data set D of terminal device #N N Statistics of C i is the communication capability of terminal device #i, for example, C N It is the communication capability of terminal device #N.

[0229] Based on this solution, the terminal equipment is selected by comprehensively considering the two factors of communication capability and statistical distance, which can balance the impact of statistical distance and communication capability and adapt to the different requirements for communication capability and data similarity in different scenarios.

[0230] Optionally, when a network device selects a terminal device, there is a clear preference for communication capability or statistical distance, and the weight between the two can be regulated by a hyperparameter α.

[0231] For example, the network device selects a target value A maximum of M terminal devices.

[0232] For example, the network device selects a target value A maximum of M terminal devices.

[0233] Based on this solution, by adjusting the hyperparameter α, the trade-off between communication capability and statistical distance can be flexibly balanced to meet the optimization goals in specific needs or scenarios. In practical applications, the optimal α value can be selected through methods such as cross-validation to achieve the best selection.

[0234] S560: The network device sends instruction information to the terminal device. Correspondingly, the terminal device receives the instruction information from the network device.

[0235] Optionally, the network device may broadcast a selected set of terminal devices (such as terminal device #1, terminal device #2, ... terminal device #M in FIG6 ).

[0236] Optionally, the broadcast includes an identifier of the selected terminal device.

[0237] Optionally, the network device may send a selection signaling to each of the M selected terminal devices, where the selection signaling is used to indicate that the terminal device has been selected by the network device.

[0238] Exemplarily, the network device sends a selected signaling to terminal device #1, terminal device #2, ... terminal device #M respectively.

[0239] S570: Each of the selected M terminal devices sends a data set to the network device.

[0240] Correspondingly, the network device receives a data set from each of the M terminal devices.

[0241] For example, each of the M terminal devices sends a data set (D1, D2, ..., D M ).

[0242] S580, the network device receives the data set (D1, D2, ..., D M ) to train the model.

[0243] Optionally, the network device can determine a fixed number of training rounds T, perform model training, and obtain an updated model W c ;

[0244] Optionally, according to model W c The central data set is updated, and the updated central data set can be used for the next round of sampling.

[0245] After obtaining the training data set required by the model, the network device can train the model.

[0246] Optionally, if the model has not yet converged and multiple operations need to be performed, the process shown in FIG6 can be repeated.

[0247] For example, the network device performs T model training in the cth round of sampling in the K rounds of sampling to obtain the model W c , repeat the process shown in Figure 6, the network device can perform T times of model training again in the c+1 round of sampling to obtain the model W c+1 , where the value of K can be preset, or sampling can be stopped until the model converges.

[0248] It should be understood that the central data set of the next round of sampling, that is, the c+1 round of sampling, is the central data set updated after the end of the c round of sampling.

[0249] Optionally, the c+1th round of sampling can reduce the number of terminal devices. For example, H terminal devices can be selected for the c+1th round of sampling, where H is less than N.

[0250] The solution provided in the embodiment of the present application is that after the network device performs K*T times of model training in K rounds of sampling, a model is obtained after completing K*T times of model training, which can improve the convergence speed during the model training process. In addition, the solution of the present application is to train the model in multiple rounds, which can reduce the requirements for device computing, storage and communication capabilities.

[0251] It can be understood that when the operation of the model, such as training and / or inference, is not performed on the network device, the above S470 and S580 steps can be replaced by the operation network element of the model. The method also includes the network device sending information or reporting information related to the operation of the model to the operation network element of the model.

[0252] It should be understood that the data set described above (eg, a central data set, or a data set of a terminal device) may be part or all of the data set itself, or may be indicative information of the data set, and this application does not limit this.

[0253] The above describes in detail the node selection method provided by this application. As mentioned above, the method provided by this application can be applied to AI scenarios. The above model is an AI model. The method of this application can also be applied to non-AI scenarios. For example, if the network device needs to conditionally select a data set, the method of this application can be used.

[0254] The node selection method embodiment provided by the present application is described above in conjunction with Figures 5 and 6. The device embodiment provided by the present application is described below in conjunction with Figure 7. It should be understood that the device embodiment and the method embodiment correspond to each other, and similar descriptions can refer to the method embodiment.

[0255] Figure 7 shows a schematic diagram of an apparatus 700 provided in an embodiment of the present application. The apparatus 700 may include a transceiver unit 710, a storage unit 720, and a processing unit 730. The transceiver unit 710 is configured to receive or send information and / or data and may also be referred to as a communication interface or communication unit. The storage unit 720 is configured to implement corresponding storage functions and store corresponding information and / or data. The processing unit 730 is configured to perform data processing, so that the apparatus 700 implements the aforementioned node selection method.

[0256] In a possible implementation, the apparatus 700 may only include a transceiver unit 710 and a processing unit 730 .

[0257] As a design, the apparatus 700 may execute the actions executed by the network device in the above method embodiment.

[0258] The apparatus 700 includes a transceiver unit 710 and a processing unit 730. The transceiver unit 710 is configured to: transmit first information, the first information including information about a first data set of a network device; and receive second information from each of N terminal devices, the second information including a statistical distance of each of the N terminal devices, the statistical distance of each of the N terminal devices indicating a distance between the second data set of each of the N terminal devices and the first data set. The processing unit 730 is configured to determine M terminal devices based on the second information.

[0259] In one possible implementation, the transceiver unit 710 is further configured to send instruction information for instructing M terminal devices to receive a second data set from each of the M terminal devices. The processing unit 730 is further configured to perform model training based on the second data set for each of the M terminal devices.

[0260] In one possible implementation, the transceiver unit 710 is specifically used to send a broadcast signal, which includes first indication information, and the first indication information is used to indicate a set of M terminal devices; or, to send second indication information to each of the M terminal devices, and the second indication information is used to indicate each terminal device.

[0261] In one possible implementation, the processing unit 730 is specifically used to determine M terminal devices based on the statistical distance ratio corresponding to each terminal device in the N terminal devices, where the statistical distance ratio is the ratio of the statistical distance of each terminal device in the N terminal devices to the maximum statistical distance among the statistical distances of the N terminal devices.

[0262] The M terminal devices are the M terminal devices with the largest statistical distance ratio among the N terminal devices.

[0263] In one possible implementation, the processing unit 730 is specifically used to determine M terminal devices based on the statistical distance ratio and communication capability ratio corresponding to each of the N terminal devices; wherein the statistical distance ratio is the ratio of the statistical distance of each of the N terminal devices to the maximum statistical distance among the statistical distances of the N terminal devices, and the communication capability ratio is the ratio of the communication capability of each of the N terminal devices to the maximum communication capability among the communication capabilities of the N terminal devices.

[0264] Among them, the M terminal devices are the M terminal devices with the largest target value among the N terminal devices, and the target value is the sum of the statistical distance ratio and the communication capability ratio, or the weighted sum.

[0265] As a design, the apparatus 700 may execute the actions executed by the terminal device in the above method embodiment.

[0266] The device 700 includes: a transceiver unit 710 and a processing unit 730, the transceiver unit 710 is used to: receive first information from a network device, the first information including information of a first data set of the network device; send second information, the second information including a statistical distance, the statistical distance being used to indicate the distance between the second data set of the terminal device and the first data set, the second information being used to determine M terminal devices; receive indication information, the indication information being used to indicate M terminal devices; and send a second data set in response to the indication information, the second data set being used for model training.

[0267] In one possible implementation, the transceiver unit 710 is specifically used to receive a broadcast signal, which includes first indication information, and the first indication information is used to indicate a set of M terminal devices; or, receive second indication information, and the second indication information is used to indicate the terminal device.

[0268] In a possible implementation, the processing unit 730 is configured to determine a statistical distance.

[0269] Optionally, the information of the first data set is the first data set, and the statistical distance is the distance between the second data set and the first data set.

[0270] Optionally, the information of the first data set is statistical information of the first data set, and the statistical distance is the distance between the statistical information of the second data set and the statistical information of the first data set.

[0271] FIG8 shows a schematic diagram of another apparatus 800 provided in an embodiment of the present application.

[0272] The device 800 may include a memory 810, a processor 820, and a communication interface 830. The memory 810, processor 820, and communication interface 830 are connected via an internal connection path. The memory 810 is used to store instructions, and the processor 820 is used to execute the instructions stored in the memory 810 to control the communication interface 830 to obtain information, or to enable the device 800 to implement the aforementioned model training method. Optionally, the memory 810 may be coupled to the processor 820 via an interface, or may be integrated with the processor 820.

[0273] In one possible implementation, the apparatus 800 may include only a processor 820 and a communication interface 830 .

[0274] It should be noted that the communication interface 830 may be a transceiver such as, but not limited to, a transceiver. The communication interface 830 may also include an input / output interface.

[0275] The processor 820 stores one or more computer programs, which include instructions. When the instructions are executed by the processor 820, the device 800 executes the model training method in each of the above embodiments.

[0276] During implementation, each step of the above method can be completed by an integrated logic circuit of the hardware in the processor 820 or by instructions in the form of software. The method disclosed in conjunction with the embodiments of the present application can be directly embodied as being executed by a hardware processor, or can be executed by a combination of hardware and software modules in the processor. The software module can be located in a storage medium mature in the art, such as a random access memory, a flash memory, a read-only memory, a programmable read-only memory or an electrically erasable programmable memory, a register, etc. The storage medium is located in the memory 810, and the processor 820 reads the information in the memory 1110 and completes the steps of the above method in combination with its hardware. To avoid repetition, it will not be described in detail here.

[0277] In one possible implementation, the apparatus 800 may include only the processor 820 and the communication interface 830 , but not the memory 810 .

[0278] Optionally, the communication interface 830 in FIG. 8 may implement the transceiver unit 710 in FIG. 7 , and the processor 820 in FIG. 8 may implement the processing unit 730 in FIG. 7 .

[0279] An embodiment of the present application further provides a computer-readable storage medium, wherein the computer-readable storage medium stores a program code. When the computer program code is executed on a computer, the computer executes any one of the methods in FIG. 5 or FIG. 6 .

[0280] An embodiment of the present application further provides a computer program product, which includes a computer program. When the computer program is executed, the computer executes any one of the methods in Figure 5 or Figure 6 above.

[0281] An embodiment of the present application further provides a chip or chip system, including: a circuit, which is used to execute any one of the methods in Figure 5 or Figure 6 above.

[0282] An embodiment of the present application also provides a communication system, including: a first device and a second device, the first device is used to execute the actions / steps executed by the network device in Figure 5 or Figure 6; the second device is used to execute the actions / steps executed by the terminal device in Figure 5 or Figure 6.

[0283] Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0284] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0285] In the several embodiments provided in this application, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are merely schematic. For example, the division of the units is merely a logical function division. In actual implementation, there may be other division methods, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0286] The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of these units may be selected to achieve the purpose of this embodiment according to actual needs.

[0287] In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.

[0288] If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.

[0289] The above description is merely a specific 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.

Claims

1. A node selection method, characterized in that, Including: Sending first information to N terminal devices, where the first information includes information on a first data set of a network device, and N is an integer greater than 1; Receiving second information from each of the N terminal devices, where the second information includes the statistical distance of each of the N terminal devices, and the statistical distance of each of the N terminal devices is used to indicate the distance between the second data set of each of the N terminal devices and the first data set; Determining M terminal devices according to the second information, where M is a positive integer and M is less than or equal to N; Sending indication information, where the indication information is used to indicate the M terminal devices; Receiving the second data set from each of the M terminal devices; Performing model training according to the second data set of each of the M terminal devices.

2. The method according to claim 1, wherein The sending of the indication information includes: Sending a broadcast signal, where the broadcast signal includes first indication information, and the first indication information is used to indicate the set of the M terminal devices; or, Sending second indication information to each of the M terminal devices, where the second indication information is used to indicate each of the terminal devices.

3. The method according to claim 1 or 2, characterized in that, Determining M terminal devices according to the second information includes: Determining the M terminal devices according to the statistical distance ratio corresponding to each of the N terminal devices, where the statistical distance ratio is the ratio of the statistical distance of each of the N terminal devices to the maximum statistical distance among the statistical distances of the N terminal devices.

4. The method according to claim 3, wherein: The M terminal devices are the M terminal devices with the largest statistical distance ratio among the N terminal devices.

5. The method according to any one of claims 1 to 4, characterized in that The second information further includes the communication capability of each of the N terminal devices, and the communication capability includes at least one of the following: the distance between each of the N terminal devices and the network device, and the energy consumption per unit data volume of each of the N terminal devices.

6. The method according to claim 5, characterized in that Determining M terminal devices according to the second information includes: Determining M terminal devices according to the statistical distance ratio and the communication capability ratio corresponding to each of the N terminal devices; wherein, the statistical distance ratio is the ratio of the statistical distance of each of the N terminal devices to the maximum statistical distance among the statistical distances of the N terminal devices, and the communication capability ratio is the ratio of the communication capability of each of the N terminal devices to the maximum communication capability among the communication capabilities of the N terminal devices.

7. The method according to claim 6, wherein: The M terminal devices are the M terminal devices with the largest target value among the N terminal devices, and the target value is the sum, or weighted sum, of the statistical distance ratio and the communication capability ratio.

8. The method according to any one of claims 1 to 7, wherein: The information of the first data set is the first data set, and the statistical distance is the distance between the second data set of each of the N terminal devices and the first data set.

9. The method according to any one of claims 1 to 7, characterized in that The information of the first data set is the statistical information of the first data set, and the statistical distance is the distance between the statistical information of the second data set of each of the N terminal devices and the statistical information of the first data set.

10. The method according to claim 8 or 9, characterized in that, The distance includes at least one of the following: Fréchet distance; maximum mean discrepancy; earth mover's (Wasserstein) distance; Kolmogorov-Smirnov distance.

11. A node selection method, characterized in that, including: Receiving first information from a network device, the first information including information of a first data set of the network device; Sending second information, the second information including a statistical distance for indicating the distance between a second data set of a terminal device and the first data set, the second information being used to determine M terminal devices, where M is a positive integer; Receiving indication information for indicating the M terminal devices; In response to the indication information, sending the second data set for model training.

12. The method according to claim 11, wherein The receiving the indication information includes: Receiving a broadcast signal including first indication information for indicating a set of the M terminal devices; or, Receiving second indication information for indicating the terminal device.

13. The method according to claim 11 or 12, characterized in that, The second information further includes communication capabilities, and the communication capabilities include at least one of the following: the distance between the terminal device and the network device, the energy consumption per unit data volume of the terminal device.

14. The method according to any one of claims 11 to 13, characterized in that The information of the first data set is the first data set, and the statistical distance is the distance between the second data set and the first data set.

15. The method according to any one of claims 11 to 13, characterized in that The information of the first data set is the statistical information of the first data set, and the statistical distance is the distance between the statistical information of the second data set and the statistical information of the first data set.

16. The method according to claim 14 or 15, characterized in that The distance includes at least one of the following: Fréchet distance; maximum mean discrepancy; earth mover's (Wasserstein) distance; Kolmogorov-Smirnov distance.

17. A communication device, characterized in that, including a module or unit for performing the method according to any one of claims 1 to 10.

18. A communication device, characterized in that, including: A processor and a memory, the processor being coupled to the memory for reading and executing instructions in the memory to perform the method according to any one of claims 1 to 10.

19. A communication device, characterized in that, including a module or unit for performing the method according to any one of claims 11 to 16.

20. A communication device, characterized in that, including: A processor and a communication interface, the communication interface being connected to the processor, the communication interface being configured to obtain a program or instructions, and the processor executing the method according to any one of claims 1 to 10 or executing the method according to any one of claims 11 to 16 by running the program or instructions.

21. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer program code, which, when run on a computer, causes the computer to execute the method according to any one of claims 1 to 10 or claims 11 to 16.

22. A chip or a chip system, characterized in that, Comprising: a circuit configured to execute the method according to any one of claims 1 to 10 or claims 11 to 16.

23. A communication system, characterized in that, Comprising: A first device configured to execute the method according to any one of claims 1 to 10; A second device configured to execute the method according to any one of claims 11 to 16.

Images