A high expansibility federated learning method for power mobile edge computing

Abstract

The application discloses a high expansibility federated learning method for power mobile edge computing. In view of the expansibility problem of mobile edge computing in the power industry, the application improves the expansibility of the system by deploying multiple mutually cooperative edge servers. The application has two new features: 1) it can adaptively determine the local model update times of each working device; 2) it can dynamically adjust the model synchronization transmission between the working device and the edge server. Through the above two characteristics, the application designs a control mechanism. It consists of two key modules: an adaptive iteration number determination module based on reinforcement learning technology on the edge server side; and a dynamic selection module of the working device and the edge server based on the incentive mechanism on the working device side. Compared with the previous method, the application can effectively improve the expansibility of the power mobile edge computing field, has higher model precision and lower energy consumption.

Description

Technical Field

[0001] In response to the scalability problem of federated learning, the present invention proposes a highly scalable federated learning method for power mobile edge computing. Background Art

[0002] Federated learning technology is widely used in related fields such as mobile edge computing for electric power. However, the complex and real-world network environments present challenges for deploying federated learning tasks. Furthermore, with the increasing number of sensors embedded within large-scale equipment in the power industry, ensuring the scalability of the entire system is crucial. Therefore, developing a highly scalable federated learning method for mobile edge computing is of great significance to the power industry.

[0003] Currently, federated learning tasks deployed in the field of electric mobile edge computing typically rely on a single edge server. However, this single-edge server approach cannot meet the requirements for training a large number of worker nodes simultaneously performing federated learning. Furthermore, due to network resource limitations in the field, if multiple workers simultaneously upload model parameters to the edge server, network congestion can occur, making federated learning training inefficient. In recent years, new implementations based on multi-edge server technology have been used to improve the effectiveness of federated learning training, providing new insights for this study on highly scalable federated learning methods for electric mobile edge computing scenarios. This paper aims to implement a highly scalable federated learning method for the electric mobile edge computing field. Summary of the Invention

[0004] The purpose of this section is to summarize some aspects of the embodiments of the present invention and briefly introduce some preferred embodiments. Some simplifications or omissions may be made in this section and the abstract and title of this application to avoid obscuring the purpose of this section, the abstract and the title of the invention, and such simplifications or omissions should not be used to limit the scope of the present invention.

[0005] In view of the above-mentioned and / or existing problems in the field of existing power mobile edge computing, the present invention is proposed.

[0006] Therefore, the problem to be solved by the present invention is how to achieve high scalability of power edge computing.

[0007] To solve the above technical problems, the present invention provides the following technical solutions: a highly scalable federated learning method for power mobile edge computing, which includes:

[0008] Deploy the adaptive algorithm module on the edge server side based on multi-edge server technology;

[0009] Deploy the dynamic selection module on the working node side;

[0010] In any training round, each worker device adapts the number of local iterations and dynamically selects the edge server for model synchronization.

[0011] As an optimal solution of the highly scalable federated learning method for electric mobile edge computing described in the present invention, a control mechanism for a highly scalable federated learning method for electric mobile edge computing is designed, and the control mechanism includes an algorithm module based on deep reinforcement learning and a dynamic selection module based on an incentive mechanism.

[0012] As a preferred solution of the highly scalable federated learning method for electric mobile edge computing described in the present invention, an algorithm module based on deep reinforcement learning is deployed on the edge server side. The module uses curiosity exploration to help the algorithm module converge quickly. At the same time, in order to achieve highly scalable federated learning for electric mobile edge computing, the module designs the corresponding state space, action space, and external reward function based on deep reinforcement learning. The module uses a spatiotemporal error method to update the training model within the module. Based on known information training, the locally expected number of iterations for all working devices of each edge server and the corresponding model synchronization binary variables are obtained;

[0013] By deploying a dynamic selection module based on an incentive mechanism on the working device side, the working device is helped to select the optimal edge server for model synchronization, thereby improving the scalability of the federated learning method under electric mobile edge computing, while improving the accuracy of the training model and reducing the resource overhead of the federated learning method under electric mobile edge computing.

[0014] As a preferred solution of the highly scalable federated learning method for electric mobile edge computing described in the present invention, the control mechanism includes an algorithm module based on deep reinforcement learning, which is a DDQN algorithm module with curiosity exploration, using To represent the state, reward and external action of the client in edge server k in the tth training round, the reward here is It consists of two parts: internal rewards and external rewards Right now:

[0015]

[0016] Since the curiosity exploration in the deep reinforcement learning algorithm module has two identical neural networks, namely the curiosity target network and curiosity prediction network in, and Corresponding to the model parameters of these two neural networks, so far, the internal reward The mean square error of both can be defined as:

[0017]

[0018] Where the curiosity target network Can be minimized by gradient descent To update the parameters, while the curiosity prediction network Is updated by soft update.

[0019] As a preferred scheme of the high expansibility federated learning method for power mobile edge computing, the core technology of the deep reinforcement learning-based algorithm module is DDQN, including two neural network parameter architectures consistent deep learning model Q value approximation function And Q value function Where θ' k The model parameter information of Q value approximation function and Q value function respectively, the neural network model parameters of DDQN are updated by space-time difference, and the space-time difference target is represented by

[0020]

[0021] γ∈[0,1) is a discount coefficient used to reduce the importance of future rewards, and the space-time difference error is defined as

[0022]

[0023] Update by gradient descent method That is

[0024]

[0025] Where, α represents the expansion step.

[0026] As a preferred scheme of the high expansibility federated learning method for power mobile edge computing, the control mechanism includes a deep reinforcement learning-based algorithm module, and the state space, action space and external reward function of reinforcement learning in the module are also designed, wherein the state space is defined as Indicates the global model parameters of the edge server k in t training rounds; Composed of vector Indicates the remaining amount of all resources of the working device i, assuming that there are U types of resources, then the initial total amount of any type of resource u in the working device i is r i,u ​​In the t-th training round, assuming that each execution of the local iteration training of the working device i consumes units of u-type resources; and the model synchronization between the working device i and the edge server k consumes units of u-type resources, then:

[0027]

[0028]

[0029]

[0030] W (t) denotes the network topology between the edge servers in the t-th training round, where den(·) is the density function of the network topology graph.

[0031] As a preferred scheme of the high expansibility federated learning method for power mobile edge computing, the action space is defined as After receiving the state space , the client in the edge server k will generate an action matrix This matrix is composed of M decision vectors (i = 1, 2, … M), which means that each edge server k will make action decisions for all M working devices participating in model training. In the t-th training round, the action decision of a certain working device i generated by the client in any edge server k is composed of two decision variables, which are the expected number of local iteration times of the working device i and a 0 / 1 binary variable whether the working device i expects the edge server k to perform model synchronization

[0032] For the client outside any edge server k, the external reward function is defined as:

[0033]

[0034] wherein are positive expansion factors, β is a regularization coefficient, ∈ represents the expected accuracy of the model set in different federated learning training tasks, the function calculates the actual accuracy of the model corresponding to x;

[0035]

[0036] Indicates the number of devices that actually synchronize the model with the edge server in the tth training round.

[0037]

[0038] As a preferred solution of the highly scalable federated learning method for electric mobile edge computing described in the present invention, the control mechanism further includes a dynamic selection module based on an incentive mechanism;

[0039] definition To express the probability that working device i selects edge server k for model synchronization, it can be expressed as a piecewise function:

[0040]

[0041] Design a dynamic incentive for each worker device, which is obtained based on the edge server k selected by worker device i in training round t:

[0042]

[0043] in, is the pulse variable obtained based on the model performance on edge server k in the tth training round; σ i is the contribution ratio related to the distribution of working equipment data; and R i It is a compensation constant obtained by the working equipment participating in the training. Model:

[0044]

[0045] Where ω is a parameter related to the degree of impulse, and is the maximum model accuracy that can be achieved theoretically between servers. In the tth round of training, after the working device i chooses to join the model synchronization device set of the edge server k, the resource overhead is

[0046]

[0047] Therefore, the utilization rate of working device i choosing to join edge server k for model synchronization can be calculated as:

[0048]

[0049] in, Represents a linear function. In summary, among the many edge servers, the average utilization rate of working device i that wants to synchronize the model is

[0050]

[0051] In order to pursue higher utilization, each device will determine the most appropriate server for model synchronization.

[0052] A computer device includes a memory and a processor, wherein the memory stores a computer program, and is characterized in that the processor implements the steps of the above method when executing the computer program.

[0053] A computer-readable storage medium stores a computer program thereon, wherein the computer program implements the steps of the above method when executed by a processor.

[0054] The beneficial effects of the present invention are: compared with previous methods, the present invention can effectively improve the scalability of the field of electric mobile edge computing, and has higher model accuracy and lower energy consumption. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for describing the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. Those skilled in the art can also derive other drawings based on these drawings without inventive effort. Among them:

[0056] Figure 1 This is a schematic diagram of the overall architecture of a highly scalable federated learning method for power mobile edge computing in Example 1.

[0057] Figure 2 This is a basic flow chart of the control mechanism of a highly scalable federated learning method for power mobile edge computing in Example 1.

[0058] Figure 3 This is a schematic diagram of the deep reinforcement learning-based modules involved in the control mechanism of a highly scalable federated learning method for electric mobile edge computing in Example 1.

[0059] Figure 4 This is a graph showing the convergence effect of the test set loss model for training an SVM model using the MNIST dataset using a highly scalable federated learning method for electric mobile edge computing in Example 2 and other comparative methods.

[0060] Figure 5 This is a graph showing the convergence effect of the test set loss model for a highly scalable federated learning method for electric mobile edge computing in Example 2 and other comparative methods using the MNIST dataset to train the LeNet model.

[0061] Figure 6This is a graph showing the convergence effect of the test set loss model for training the LSTM model using the AG NEWS dataset using a highly scalable federated learning method for electric mobile edge computing in Example 2 and other comparative methods. DETAILED DESCRIPTION

[0062] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0063] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0064] Secondly, the term "one embodiment" or "embodiment" herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in various places throughout this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive of other embodiments.

[0065] Example 1

[0066] Reference Figure 1 , which is the first embodiment of the present invention. This embodiment provides a highly scalable federated learning method for electric mobile edge computing.

[0067] Reference Figure 2 In this embodiment, a control mechanism is implemented internally. The mechanism includes an algorithm module based on deep reinforcement learning and a dynamic selection module based on an incentive mechanism.

[0068] Reference Figure 3 The control mechanism in this embodiment has an algorithm module based on deep reinforcement learning. This module is a DDQN algorithm module with curiosity exploration. To represent the state, reward and external action of the client in edge server k in the tth training round. Here the reward It consists of two parts: internal rewards and external rewards Right now:

[0069]

[0070] Since the curiosity exploration in the deep reinforcement learning algorithm module has two identical neural networks, namely the curiosity target network and curiosity prediction network in, and Corresponding to the model parameters of these two neural networks respectively. In this embodiment, three layers of 128×128 convolutional layers with ReLU activation function are selected as the curiosity target network and curiosity prediction network models. It can be defined as the mean square error between the two:

[0071]

[0072] The core technology of the algorithm module based on deep reinforcement learning is DDQN. It includes two deep learning models with the same neural network parameter architecture and Q value approximation function. and Q-value function in θ' k They are the model parameter information of the Q-value approximation function and the Q-value function respectively. In this embodiment, both the Q-value approximation function and the Q-value function use a three-layer fully connected neural network as a deep learning model. The first two fully connected layers of the model have 400 neurons, and the third fully connected layer has 300 neurons. ReLU is used as the activation function between every two layers. The neural network model parameters of DDQN are updated through the time-space difference. To express the spatiotemporal difference target:

[0073]

[0074] γ∈[0,1) is a discount factor used to reduce the importance of future rewards. Define the temporal and spatial difference error for:

[0075]

[0076] Minimize by gradient descent To update,

[0077]

[0078] Where α represents the expansion step size.

[0079] This embodiment also designs the corresponding state space, action space and external reward function in the deep reinforcement learning algorithm module within the control mechanism. The state space is defined as represents the global model parameters of edge server k in t training rounds; By vector It represents the remaining amount of all resources of working device i. Assuming that there are U types of resources, the initial total amount of any type of resource u in working device i is r i,uIn the tth round of training, it is assumed that each time the worker device i performs local iterative training, it consumes Units of u-type resources; while the model synchronization between working device i and edge server k consumes Units of u-type resources, then:

[0080]

[0081]

[0082]

[0083] W (t) represents the network topology between edge servers in the tth round of training. Here, den(·) is the density function of the network topology graph. The action space is defined as In the received state space After that, the client in edge server k will generate the action matrix This matrix has M decision vectors (i=1,2,…M). This means that each edge server k will make action decisions for all M working devices participating in model training. In the tth training round, the action decision of a working device i generated by the client in any edge server k is It consists of two decision variables, namely the number of local model iterations expected by the working device i and a 0 / 1 binary variable indicating whether you want to synchronize the model between the worker device i and the edge server k. For any client external reward function within edge server k Defined as:

[0084]

[0085] in, are all positive expansion factors, and β is the regularization coefficient. ∈ refers to the expected accuracy of the model set in different federated learning training tasks. The function calculates the actual accuracy of the model corresponding to the current model x.

[0086]

[0087] Indicates the number of devices that actually synchronize the model with the edge server in the tth training round.

[0088]

[0089] The control mechanism in this embodiment also includes a dynamic selection module based on an incentive mechanism. To express the probability that working device i selects edge server k for model synchronization. Here it can be expressed as a piecewise function:

[0090]

[0091] Design a dynamic incentive for each worker device, which is obtained based on the edge server k selected by worker device i in training round t:

[0092]

[0093] in, is the pulse variable obtained based on the model performance on edge server k in the tth training round; σ i is the contribution ratio related to the distribution of working equipment data. This implementation uses the Fisher-Pearson coefficient as the definition of σ i The key component of R i is a compensation constant obtained by the working equipment participating in the training. Model:

[0094]

[0095] Where ω is a parameter related to the degree of impulse, and is the maximum model accuracy that can be achieved theoretically between servers. In the tth round of training, after the working device i chooses to join the model synchronization device set of the edge server k, the resource overhead is

[0096]

[0097] Therefore, the utilization rate of working device i choosing to join edge server k for model synchronization can be calculated as:

[0098]

[0099] in, Represents a linear function. The linear function selected in this embodiment is a linear function of y=x. In summary, among the many edge servers, the average utilization rate of the working device i that wants to synchronize the model is

[0100]

[0101] In order to pursue higher utilization, each device will determine the most appropriate server for model synchronization.

[0102] Example 2

[0103] To demonstrate the technical effectiveness of this method, this example uses a traditional FedAvg method, a hierarchical FedCH method, and a P-FedAvg method based on the FedAvg algorithm for comparative testing. This comparison of experimental results, using scientific evidence, verifies the effectiveness of this method.

[0104] Using the model test accuracy and energy consumption values, the present method and the present embodiment are respectively compared in three different benchmark experiments (data set: MNIST, training model SVM; data set MNIST, training model LeNet; data set: AG NEWS, training model LSTM) using the traditional FedAvg method, the FedCH method based on the hierarchical architecture, and the P-FedAvg method based on the FedAvg algorithm.

[0105] During training, the batch size is set to 128, and the grid search is performed from the set {10 -2 ,10 -1}. Set the number of edge servers B to 5 and the number of working devices M to 500. The communication topology between edge servers is generated using the Metropolis-Hastings rule. The topology can be changed from an undirected complete graph to a simple ring graph. In the embodiment, let U = 1 and only consider energy consumption as the only resource type in the experiment. Gaussian distribution is used to simulate the energy consumption corresponding to each device i. In which, let the energy consumption be calculated The mean and standard deviation of -3 and 10 -3 / 3

[67]

[98] ; communication energy consumption The mean and standard deviation are set to 2×10 -2 and 2×10 -2 / 3. Finally, in the experiment, the external reward function ∈ is set to 100, β is set to 0.1, They are set to 100, 32 and 1 respectively. In addition, in order to avoid the randomness of the experimental results, the average results of 5 independent runs are taken for the above indicators. The experimental results are shown in Table 1 and Table 2. The relevant model convergence effect diagram is shown in Figure 4 、 Figure 5 、 Figure 6 shown.

[0106] Table 1: Comparative experimental results of model test accuracy

[0107]

[0108] Table 2: Comparative experimental results of energy consumption values

[0109]

[0110] As shown in Table 1 and Table 2, in the power mobile edge computing scene, 500 working devices are deployed on a large scale, and the method has excellent high expansibility. In terms of model test accuracy, the method has better model progress. In terms of energy consumption, the energy consumption of the method is lower.

[0111] As shown in Table 1 and Table 2, in the power mobile edge computing scene, 500 working devices are deployed on a large scale, and the method has excellent high expansibility. In terms of model test accuracy, the method has better model progress. In terms of energy consumption, the energy consumption of the method is lower. Figure 4 、 Figure 5 and Figure 6 As shown in Table 1 and Table 2, in the power mobile edge computing scene, 500 working devices are deployed on a large scale, and the method has excellent high expansibility. In terms of model test accuracy, the method has better model progress. In terms of energy consumption, the energy consumption of the method is lower.

[0112] Embodiment 3

[0113] An embodiment of the present application, which is different from the first two embodiments, is that the function, if realized in the form of a software function unit and sold or used as an independent product, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application or the part that essentially contributes to the prior art or part of the technical solution can be embodied in the form of a software product stored in a storage medium, including a number of instructions for causing 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: a U disk, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a magnetic disk or an optical disk, and various program code storage media.

[0114] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered a list of executable instructions for implementing logical functions, and can be specifically embodied in any computer-readable medium for use by an instruction execution system, device or apparatus, such as a computer-based system, a system including a processor, or other system that can fetch and execute instructions from the instruction execution system, device or apparatus, or in conjunction with these instructions. For the purpose of this specification, "computer-readable medium" can be any device that can contain, store, communicate, propagate or transport programs for use by an instruction execution system, device or apparatus, or in conjunction with these instructions.

[0115] More specific examples (a non-exhaustive list) of computer-readable media include the following: an electrical connection with one or more wires (electronic devices), a portable computer disk cartridge (magnetic devices), a random access memory (RAM), a read-only memory (ROM), an erasable and programmable read-only memory (EPROM or flash memory), a fiber optic device, and a portable compact disc read-only memory (CDROM). In addition, the computer-readable medium may even be paper or other suitable medium on which the program is printed, since the program may be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, deciphering, or processing in another suitable manner as necessary, and then stored in a computer memory.

[0116] It should be understood that various parts of the present invention can be implemented using hardware, software, firmware, or a combination thereof. In the above-described embodiments, multiple steps or methods can be implemented using software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented using hardware, as in another embodiment, any one of the following technologies known in the art or a combination thereof can be used: a discrete logic circuit having a logic gate circuit for implementing a logic function on a data signal, an application-specific integrated circuit having a suitable combination of logic gate circuits, a programmable gate array (PGA), a field programmable gate array (FPGA), etc.

[0117] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A highly scalable federated learning method for electric mobile edge computing, characterized by: include, Deploy the adaptive algorithm module on the edge server side based on multi-edge server technology; Deploy the dynamic selection module on the working node side; In any training round, each worker device adapts the number of local iterations and dynamically selects the edge server for model synchronization; Design a highly scalable federated learning control mechanism for electric mobile edge computing. The control mechanism includes an algorithm module based on deep reinforcement learning and a dynamic selection module based on an incentive mechanism. A deep reinforcement learning-based algorithm module is deployed on the edge server side. This module uses curiosity exploration to help the algorithm module converge quickly. The corresponding state space, action space, and external reward function are designed based on deep reinforcement learning. The spatiotemporal error method is used to update the training model within the module. Based on known information, each edge server obtains the local expected number of iterations for all working devices and the corresponding model synchronization binary variable. By deploying a dynamic selection module based on an incentive mechanism on the working device side, it helps the working device select the optimal edge server for model synchronization, thereby improving the scalability of the federated learning method in electric mobile edge computing; The algorithm module based on deep reinforcement learning is a dual-depth Q network DDQN algorithm module with curiosity exploration, which uses To represent the state, reward and external action of the client in edge server k in the tth training round, the reward here is It consists of two parts: internal rewards and external rewards Right now: Since the curiosity exploration in the deep reinforcement learning algorithm module has two identical neural networks, namely the curiosity target network and curiosity prediction network in, and Corresponding to the model parameters of these two neural networks, so far, the internal reward It is defined as the mean square error between the two: Among them, the curiosity target network Minimize by gradient descent Perform parameter updates, and the curiosity prediction network The model is updated through soft update; The core technology of the algorithm module based on deep reinforcement learning is DDQN, which includes two deep learning models with consistent neural network parameter architecture and Q value approximation function. and Q-value function in θ ' k They are the model parameter information of the Q-value approximation function and the Q-value function, and the neural network model parameters of DDQN are updated through the time-space difference. To express the spatiotemporal difference target: γ∈[0,1) is a discount factor used to reduce the importance of future rewards, defining the spatiotemporal error for: Minimize by gradient descent To update, Among them, α represents the expansion step size; The corresponding state space, action space and external reward function are designed based on deep reinforcement learning. The state space is defined as represents the global model parameters of edge server k in t training rounds; By vector It represents the remaining amount of all resources of working device i. Assuming that there are U types of resources, the initial total amount of any type of resource u in working device i is r i,u , in the tth round of training, assuming that each time the worker device i performs local iterative training, it consumes Units of u-type resources; while the model synchronization between working device i and edge server k consumes Units of u-type resources, then: W (t) represents the network topology between edge servers in the tth training round, where den(·) is the density function of the network topology graph; The action space is defined as In the received state space After that, the client in edge server k will generate the action matrix This matrix has M decision vectors (i=1,2,…M), which means that each edge server k will make action decisions for all M working devices participating in model training. In the tth training round, the action decision of a working device i generated by the client in any edge server k It consists of two decision variables, namely the number of local model iterations expected by the working device i and a 0 / 1 binary variable indicating whether you want to synchronize the model between the worker device i and the edge server k. For any client external reward function within edge server k Defined as: in, are all positive expansion factors, β is the regularization coefficient, ∈ refers to the expected accuracy of the model set in different federated learning training tasks, The function calculates the actual accuracy of the model corresponding to the current model x; Indicates the number of devices that actually synchronize the model with the edge server in the tth training round; The dynamic selection module based on the incentive mechanism includes defining To represent the probability that working device i selects edge server k for model synchronization, Expressed as a piecewise function: Design a dynamic incentive for each worker device, which is obtained based on the edge server k selected by worker device i in training round t: in, is the pulse variable obtained based on the model performance on edge server k in the tth training round; σ i is the contribution ratio related to the distribution of working equipment data; and R i It is a compensation constant obtained by the working equipment participating in the training. Model: Among them, ω is a parameter about the degree of impulse, and is the maximum model accuracy that can be achieved theoretically between servers. In the tth round of training, after the working device i chooses to join the model synchronization device set of the edge server k, the resource overhead is Therefore, the utilization rate of working device i choosing to join edge server k for model synchronization is calculated as: in, Represents a linear function. In summary, among many edge servers, the average utilization rate of working device i that wants to synchronize models is expressed as, In order to pursue higher utilization, each device will determine the most appropriate server for model synchronization.

2. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to claim 1 are implemented.

3. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to claim 1 are implemented.

Images