Edge computing unloading method

Abstract

An embodiment of the invention discloses an edge computing unloading method, which introduces a personalized federated learning mechanism, considers unique characteristics and requirements of each edge device, realizes personalized unloading decisions, improves resource utilization and computing efficiency, meets requirements under different application scenes, adopts personalized edge computing unloading, and improves unloading efficiency. Only model parameters or gradient information is transmitted instead of original data, data privacy is effectively protected, the risks of data leakage and privacy invasion are reduced, the security and credibility of a system are enhanced, local knowledge of multiple edge devices is fused into a global model through personalized edge computing unloading, the generalization ability and accuracy of the model are improved, and the system performance is improved. The global model can better adapt to various complex scenes and change requirements, through personalized unloading decision and iterative optimization, calculation tasks are reasonably allocated to edge devices or edge servers, optimal allocation of resources is achieved, calculation efficiency and service quality are improved, and requirements and expectations of users are met.

Description

Technical Field

[0001] The present invention relates to the field of edge computing technology, and in particular to a method for edge computing offloading. Background Art

[0002] With the rapid development of the Internet of Things (IoT) and big data technologies, the amount of data generated by edge devices (such as smartphones, sensors, etc.) has exploded. However, due to the limited computing resources and storage capacity of edge devices, how to efficiently process this data has become a challenge. Edge computing, as an emerging computing paradigm, can effectively alleviate this problem by offloading computing tasks to edge servers. However, existing edge computing offloading models often ignore the differences in computing power, data characteristics, etc. between different edge devices, and cannot achieve personalized offloading decisions.

[0003] There are some shortcomings in the currently proposed offloading models.

[0004] 1) Existing edge computing offloading models often adopt a unified offloading strategy, ignoring the differences in computing power, data characteristics, application scenarios, etc. between different edge devices, and cannot achieve personalized offloading decisions. This leads to inefficient resource utilization and poor computing efficiency.

[0005] 2) In the existing computing offloading model, edge devices are usually required to transmit raw data or computing results to the central server for processing. This approach is prone to the risk of data leakage and privacy violation, especially when processing sensitive data.

[0006] 3) Due to the limited computing resources and data of edge devices, their local models may not generalize well to all scenarios. Existing offloading models often fail to effectively integrate the local knowledge of multiple edge devices into the global model, limiting the accuracy and generalization ability of the model.

[0007] 4) Existing offloading models may not be reasonable in resource allocation and cannot be dynamically adjusted according to the actual situation and needs of edge devices. This may lead to waste or shortage of computing resources, affecting computing efficiency and service quality. Summary of the invention

[0008] In order to solve the above technical problems, the present invention proposes a method for edge computing offloading, which can improve resource utilization efficiency and computing efficiency, meet the needs of different application scenarios, reduce the risks of data leakage and privacy infringement, enhance the security and credibility of the system, enable the global model to better adapt to various complex scenarios and changing needs, and improve computing efficiency and service quality to meet user needs and expectations.

[0009] An embodiment of the present invention provides a method for offloading edge computing, comprising the following steps:

[0010] The cloud server initializes the parameters of the local local models of all edge servers and sends them to each edge server for training;

[0011] The user device sends a computing task request to the edge server;

[0012] The edge server receives the computing task request and trains the local model using a deep learning optimization algorithm based on the DQN model to obtain a personalized local model.

[0013] The edge server sends the local model update increment to the cloud server, and the cloud server uses the local model update increment to update the hypernetwork model;

[0014] The cloud server weights and aggregates the final super network through three layers of personalized super network models and one layer of global super network model, and then aggregates it with the local local models corresponding to each edge server to output a personalized local local model for the edge server;

[0015] The cloud server sends the personalized local partial model to the corresponding edge server again;

[0016] Repeat the above steps until the specified number of iterations R or the loss function L converges, and the average loss function of all edge servers reaches the minimum convergence;

[0017] The cloud server notifies all edge servers to stop training.

[0018] The computing task request further includes:

[0019] Task parameters: Indicates that any i belongs to M, from 1 to M, where R(i) represents the task request of the i-th terminal device, The definition of R(i) is (I i , C i , T i ), in the task request parameter information, I i Indicates the size of the input data, C i Indicates the number of CPU cycles required to complete the task, T i represents time, i represents the i-th terminal device;

[0020] Network configuration parameters: Each terminal device consumes a certain amount of transmission energy when unloading, and the task unloading time and energy consumption is the key indicator considered, R(i) represents the computing task offloaded to the edge server, and the transmission power P of the terminal device in the current time block i , the channel gain h from i to the edge server i , channel bandwidth W i , Gaussian white noise power σ 2 ;

[0021] Terminal configuration parameters: the number of calculation cycles required for the terminal to process each bit of data C_loc, the terminal's calculation frequency F_loc, the terminal's calculation energy efficiency ratio k_loc, the terminal's data transmission energy efficiency ratio k_t, the terminal's data transmission limit T, and the distance d from the terminal to its associated base station;

[0022] Base station configuration parameters: the number of calculation cycles C_e required for the base station to process each bit of data, the calculation frequency F_e of the base station, the calculation energy efficiency ratio k_e of the base station, the noise power spectrum density No_e of the base station, and the bandwidth B_e of each base station.

[0023] Furthermore, the DQN model is a three-layer neural network, including an input layer, several hidden layers and an output layer, which is used to approximate the Q function. The parameters of the DQN model include weights and biases.

[0024] The cloud server updates the hypernetwork model using the increment of the local partial model update, further comprising:

[0025] Updated increment Δθ j Expressed as: Δθ t =θ′ t -θ t , where θ′ t is the model parameter obtained by the edge server through local training, and θ t are the model parameters received from the cloud server,

[0026] For the super network corresponding to edge server j The following is v j , ν and Update method:

[0027]

[0028] where Δν j represents the change of the local hypernetwork embedding vector corresponding to the current edge server j, Δν represents the change of the global embedding vector, represents the change of partial network parameters corresponding to the current edge server j, represents the change of global hypernetwork parameters, The local model parameters corresponding to the current edge server j are represented by Δθj represents the increment of the local model parameters corresponding to the current edge server j, {θ l1 ,θ l2 ,…,θ ln} represents the parameters of each layer of the local model corresponding to edge server j, They represent their respective gradients.

[0029] It further includes: the objective function of the cloud server, and the update rules of the embedding vector and hypernetwork parameters corresponding to the edge server j,

[0030] The local model parameters of each edge server j are obtained by weighted aggregation according to ε j :

[0031]

[0032] in It is expressed as:

[0033]

[0034] Through each layer of model parameters and corresponding ε j Aggregate to get the final model parameters of each layer.

[0035] The objective function of the algorithm is expressed as follows:

[0036]

[0037] Get the average minimum value of the loss function of the final local model parameters under N edge servers,

[0038] Where V = {v1, v2, ..., V N} is the embedding vector corresponding to each edge server j,

[0039] is the hypernetwork parameter corresponding to each edge server j, ν is the global embedding vector, is the global hypernetwork parameter, and the algorithm converts the local model parameters The optimization problem is converted into an embedding vector ν j , local embedding vector ν, hypernetwork parameters and the global hypernetwork parameters

[0040] Here are the update rules:

[0041]

[0042]

[0043] The above formula is used to calculate the local embedding vector v i The loss function for edge server j Gradient The loss function of the global embedding vector ν with respect to edge server j is Gradient Global hypernetwork parameters The loss function for edge server j Gradient And calculate the hypernetwork parameters The loss function for edge server j Gradient represents the local model parameters corresponding to the current edge server j, Δθ j represents the increment of the local model parameters corresponding to the current edge server j, {θ l1 ,θ l2 ,…,θ ln} represents the parameters of each layer of the local model corresponding to the current edge server, represent their respective gradients, express About loss function gradient.

[0044] The cloud server weights and aggregates the final super network through three layers of personalized super network models and one layer of global super network model, and then aggregates with the local local models corresponding to the edge servers, and outputs personalized local local models for the edge servers, further comprising:

[0045] The final value of the hypernetwork is ε j , ε j Defined as

[0046]

[0047] The algorithm's aggregation weight matrix ε j The update process:

[0048]

[0049] Suppose there are 4 dictionaries {β 1 , β 2 , β 3 , β 4}, each dictionary contains the same set of keys K, and for each key k∈K, its corresponding value is the vector

[0050] Corresponding to each dictionary β x and key k, calculate the vector The L2 norm of

[0051]

[0052] in, Representation vector The xth component of ;

[0053] Compute the sum of the L2 norms of each key k in all dictionaries

[0054]

[0055] Perform a weighted summation on each key k according to the ratio of the L2 norm to obtain a weighted average vector

[0056]

[0057] Finally, the weighted aggregate dictionary is

[0058]

[0059] get

[0060] Then ε j and the local model parameters of the respective edge servers Generate different personalized models by weighted aggregation

[0061]

[0062] Finally, the personalized model parameters will be obtained Again to the respective edge servers.

[0063] In the above technical solution, by introducing a personalized federated learning mechanism, the unique characteristics and needs of each edge device can be fully considered to achieve personalized offloading decisions, which can improve resource utilization efficiency and computing efficiency and meet the needs of different application scenarios.

[0064] By adopting a personalized edge computing offloading method, only model parameters or gradient information are transmitted instead of original data, which effectively protects data privacy. This reduces the risk of data leakage and privacy infringement and enhances the security and credibility of the system.

[0065] By integrating the local knowledge of multiple edge devices into the global model through personalized edge computing offloading, the generalization and accuracy of the model are improved, which enables the global model to better adapt to various complex scenarios and changing needs.

[0066] Through personalized offloading decisions and iterative optimization, computing tasks can be reasonably allocated to edge devices or edge servers to achieve optimal resource allocation. This can improve computing efficiency and service quality and meet user needs and expectations.

[0067] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] The above and other purposes, features and advantages of the present invention will become more apparent by describing the embodiments of the present invention in more detail in conjunction with the accompanying drawings. The accompanying drawings are used to provide a further understanding of the embodiments of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings, the same reference numerals generally represent the same components or steps.

[0069] Figure 1 It is a flow chart of edge computing offloading in an embodiment of the present invention. DETAILED DESCRIPTION

[0070] Below, the exemplary embodiments according to the present invention will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments of the present invention, and it should be understood that the present invention is not limited to the exemplary embodiments described here.

[0071] It should be noted that the relative arrangement of components and steps, the numerical expressions and numerical values ​​set forth in these embodiments do not limit the scope of the present invention unless specifically stated otherwise.

[0072] Those skilled in the art can understand that the terms "first", "second", etc. in the embodiments of the present invention are only used to distinguish different steps, devices or modules, etc., and neither represent any specific technical meaning nor indicate the necessary logical order between them.

[0073] It should also be understood that, in the embodiments of the present invention, “plurality” may refer to two or more than two, and “at least one” may refer to one, two or more than two.

[0074] It should also be understood that any component, data or structure mentioned in the embodiments of the present invention can generally be understood as one or more, unless explicitly limited or otherwise indicated in the context.

[0075] The technical solution of the present invention is to solve the following problems:

[0076] 1) Personalized offloading requirements: Different edge devices have different requirements for computing offloading due to differences in hardware configuration, application scenarios, data characteristics, etc. By introducing a personalized federated learning mechanism, the unique characteristics and requirements of each edge device can be fully considered to achieve personalized offloading decisions.

[0077] 2) Data privacy protection: In edge computing environments, edge devices usually hold a large amount of sensitive data. How to protect this data from being leaked while achieving computation offload is an important issue. By using personalized edge computing offload, only model parameters or gradient information are transmitted instead of original data, effectively protecting data privacy.

[0078] 3) Fusion of global information and information knowledge: Although edge devices have certain computing capabilities, their local models may not be well generalized to all scenarios due to limited resources. Through personalized edge computing offloading, the local knowledge of multiple edge devices is integrated into the global model, which improves the generalization ability and accuracy of the model.

[0079] 4) Optimize resource allocation: In edge computing environments, computing resources and network resources are limited. Through personalized offloading decisions, computing tasks can be reasonably allocated to edge devices or edge servers, thereby optimizing resource allocation and improving computing efficiency.

[0080] The task offloading model is personalized for multiple edge computing scenarios with multiple terminal devices and a single edge server. Each edge server scenario includes N = {1, 2, 3, ..., N} terminal devices MDs and a base station integrated with an edge server. The edge server can provide computing offloading services for mobile devices with limited resources. N MDs can use the wireless network to send task requests (task data parameters, network parameters, base station configuration parameters, etc.) to the edge server to ensure that the requested task can be completed while meeting the constraints.

[0081] Assume that each MD has an indivisible task within T time blocks. Represents a task. i Indicates the size of the input data, C i represents the number of CPU cycles required to complete the task. Our goal is to perform edge offloading under the constraints of computing power, latency, and task offloading to minimize the total energy consumption of the system.

[0082] The model constructed below is the variables contained in the request task sent by MDs to the edge server.

[0083] 1) Task offloading model

[0084] The tasks of MDs can be completed locally or transferred to edge servers. The task offloading decision vector α, α i , i∈N is defined as the task offloading strategy, α i =1 means offloading the computation to the edge server, α i =0 means to perform calculation locally. The task offloading strategy is as follows:

[0085] α i ∈{0, 1}, i∈N

[0086] 2) Transmission model

[0087] MDs can offload computing tasks to edge servers for execution. Each MD consumes a certain amount of transmission energy when offloading, and the task offloading time is and energy consumption is the key indicator to be considered. R(i) represents the offloading of computing tasks to edge servers, P i represents the transmission power of MD in the current time block, h i represents the channel gain from i to the edge server, W i represents the channel bandwidth, σ 2 is the Gaussian white noise power. The following is the Shannon formula obtained:

[0088]

[0089] The energy consumption required to offload task R(i) to the server i∈N

[0090] 3) Computational model

[0091] Task R(i) can be computed locally or offloaded to an edge server to ensure that the task is completed within time T. Assume Indicates the time completed locally. Indicates the time it takes to complete on the edge server, with the following time limits

[0092] 4) Local computing model

[0093] Each MD has an R(i) in T and cannot be divided into subtasks. i The CPU frequency F is assigned to task R(i) i loc , F i loc .

[0094] Next, consider local computing energy consumption use Indicates MD i Chip energy consumption coefficient.

[0095] For MD i Not affected by excessive energy consumption and heat, while ensuring that it can be completed within the specified time, the computing frequency of the computing task R(i) when executed locally will not exceed its maximum computing frequency F i max .

[0096] 5) Edge Server Computing Model

[0097] F i e represents the CPU frequency assigned by the server to task R(i), F e max represents the maximum computing frequency of the edge server. Assume that k is the chip energy coefficient of the edge server, and the CPU frequency assigned to R(i) by the edge server is F i e .

[0098] Considering the resource constraints of the edge computing server for computing tasks, it is necessary to consider the overall resource constraints of the server. The resource constraints are as follows:

[0099] In energy consumption In terms of performance, energy consumption will depend directly on the overall server load and the computing resources used.

[0100] 6) Problem description

[0101] Through appropriate task offloading strategies, the total energy consumption of the edge server system can be minimized under the constraints of the above model. The above model is used to solve the optimal offloading strategy. Therefore, this problem can be modeled as a combinatorial optimization problem. It contains discrete variables α and continuous variables t up , t e , and is an NP-hard problem:

[0102]

[0103] stα i ∈{0, 1}, i∈N

[0104]

[0105] To solve the optimization problem involving discrete variables, we adopt a deep reinforcement learning algorithm based on DQN to obtain the task offloading strategy of the optimization problem, and the edge server can train an optimal task offloading model.

[0106] For multiple different edge computing environments, a personalized task offloading model is trained for each edge computing environment. Considering N servers with non-IID tasks, assuming is the dataset of the ith server, where x j is the jth request task (including task parameters, network configuration parameters, terminal configuration parameters, base station configuration parameter information, etc.), y j is the corresponding offloading strategy. The size of the data set on the i-th server is represented by m i Let θ i is the model parameter of server i, and the goal of this problem is finally expressed as:

[0107]

[0108] where θ = {θ1, ..., θ N} is a personalized parameter set for all servers. is the i-th dataset D i The loss function of the associated server. is the cross entropy loss function.

[0109] In order to solve the personalized offloading model in the multi-edge server scenario, based on the above problems, this paper proposes a personalized task offloading algorithm based on a federated learning algorithm with a multi-layer hypernetwork. Through this algorithm, a personalized offloading model can be trained for each edge server.

[0110] Figure 1 Flowchart of edge computing offloading in an embodiment of the present invention. Figure 1 As shown, the edge computing offloading process includes the following steps:

[0111] Step 101: The cloud server initializes the parameters of the local local models of all edge servers and sends them to each edge server for training.

[0112] The global model parameters refer to the initialization parameters of the local models of all edge servers. A three-layer neural network is used to approximate the Q function (i.e., action value function).

[0113] The DQN model is a three-layer neural network, including an input layer, several hidden layers, and an output layer, which is used to approximate the Q function. The parameters of the DQN model include weights and biases. The parameters of the DQN model mainly include:

[0114] Weights: Each connection between a neuron and the neurons in the next layer has a weight. These weights are updated during training using the back-propagation algorithm.

[0115] Biases: In addition to weights, each neuron has a bias term that affects the activation level of the neuron.

[0116] Step 102: The user device sends a computing task request to the edge server.

[0117] The computing task request further includes:

[0118] Task parameters: Indicates that any i belongs to M, from 1 to M, where R(i) represents the task request of the i-th terminal device, The definition of R(i) is (I i , C i , T i ), in the task request parameter information, I i Indicates the size of the input data, C i Indicates the number of CPU cycles required to complete the task, T i represents time, i represents the i-th terminal device;

[0119] Network configuration parameters: Each terminal device consumes a certain amount of transmission energy when unloading, and the task unloading time and energy consumption is the key indicator considered, R(i) represents the computing task offloaded to the edge server, and the transmission power P of the terminal device in the current time block i , the channel gain h from i to the edge server i , channel bandwidth W i , Gaussian white noise power σ 2 ;

[0120] Terminal configuration parameters: the number of calculation cycles required for the terminal to process each bit of data C_loc, the terminal's calculation frequency F_loc, the terminal's calculation energy efficiency ratio k_loc, the terminal's data transmission energy efficiency ratio k_t, the terminal's data transmission limit T, and the distance d from the terminal to its associated base station;

[0121] Base station configuration parameters: the number of calculation cycles C_e required for the base station to process each bit of data, the calculation frequency F_e of the base station, the calculation energy efficiency ratio k_e of the base station, the noise power spectrum density No_e of the base station, and the bandwidth B_e of each base station.

[0122] By sending these data to the edge server, the offloading model on the edge server can be trained accordingly based on the data and output a corresponding offloading strategy.

[0123] Step 103: The edge server receives the computing task request and uses a deep learning optimization algorithm based on the DQN model to train the local model to obtain a personalized local model.

[0124] After each edge server receives a task request, it will use a deep learning optimization algorithm based on the DQN model to train the local model. Through training, a personalized local model can be obtained for each edge server. The DQN model is usually a multi-layer neural network used to approximate the Q function (i.e., the action-value function). A typical DQN structure may include an input layer, several hidden layers, and an output layer.

[0125] We use a three-layer DQN network model. By receiving task requests sent by user devices, the DQN network on the edge server performs local iterative training based on these data, involving forward propagation to calculate the loss, and then calculating the gradient through the backpropagation algorithm. After a certain number of iterations, the edge server accumulates these gradients and updates the local model parameters according to the optimization algorithm (such as SGD, Adam, etc.).

[0126] Step 104: The edge server sends the increment of the local partial model update to the cloud server, and the cloud server updates the super network model using the increment of the local partial model update.

[0127] After each edge server performs local training iterations, it updates the local model parameters by the increment Δθ i The increment is actually the difference between the model parameters received from the cloud server last time and the current local model parameters. Specifically, the model parameters received by the edge server from the cloud at time t are θt. After local training, the model parameters are updated to θt′. Then, the updated increment Δθ i It can be expressed as: Δθ t =θ′ t -θ t Here θ′ t is the model parameter obtained by the edge server through local training, and θ t are the model parameters received from the remote end.

[0128] The cloud server updates the hypernetwork model using the increment of the local model update, further including:

[0129] Updated increment Δθ j Expressed as: Δθ t =θ′ t -θ t , where θ′ t is the model parameter obtained by the edge server through local training, and θ t are the model parameters received from the cloud server,

[0130] For the super network corresponding to edge server j The following is vj , ν and Update method:

[0131]

[0132] where Δv j represents the change of the local hypernetwork embedding vector corresponding to the current edge server j, Δv represents the change of the global embedding vector, represents the change of partial network parameters corresponding to the current edge server j, represents the change of global hypernetwork parameters, The local model parameters corresponding to the current edge server j are represented by Δθ j represents the increment of the local model parameters corresponding to the current edge server j, {θ l1 ,θ l2 ,…,θ ln} represents the parameters of each layer of the local model corresponding to edge server j, They represent their respective gradients.

[0133] Step 105: The cloud server weights and aggregates the final super network through three layers of personalized super network models and one layer of global super network model, and then aggregates them with the local local models corresponding to the edge servers, and outputs personalized local local models for the edge servers.

[0134] The objective function of the cloud server, and the update rules of the embedding vector and hypernetwork parameters corresponding to edge server j,

[0135] The local model parameters of each edge server j are obtained by weighted aggregation according to ε i :

[0136]

[0137] in It is expressed as:

[0138]

[0139] Through each layer of model parameters and corresponding ε j Aggregate to get the final model parameters of each layer.

[0140] The objective function of the algorithm is expressed as follows:

[0141]

[0142] Get the average minimum value of the loss function of the final local model parameters under N edge servers,

[0143] Where V = {v1, v2, ..., V N} is the embedding vector corresponding to each edge server j,

[0144] is the hypernetwork parameter corresponding to each edge server j, ν is the global embedding vector, is the global hypernetwork parameter, and the algorithm converts the local model parameter θ j The optimization problem is converted into an embedding vector v j , local embedding vector V, hypernetwork parameters and the global hypernetwork parameters

[0145] Here are the update rules:

[0146]

[0147] The above formula is used to calculate the local embedding vector v i The loss function for edge server j Gradient The loss function of the global embedding vector V with respect to edge server j is Gradient Global hypernetwork parameters The loss function for edge server j Gradient And calculate the hypernetwork parameters The loss function for edge server j Gradient represents the local model parameters corresponding to the current edge server j, Δθ j represents the increment of the local model parameters corresponding to the current edge server j, {θ l1 ,θ l2 ,…,θ ln} represents the parameters of each layer of the local model corresponding to the current edge server, represent their respective gradients, express About loss function gradient.

[0148] The cloud server weights and aggregates the final super network through three layers of personalized super network models and one layer of global super network model, and then aggregates with the local local models corresponding to the edge servers, and outputs personalized local local models for the edge servers, further comprising:

[0149] The final value of the hypernetwork is ε j , ε j Defined as

[0150]

[0151] The algorithm's aggregate weight matrix ε j The update process:

[0152]

[0153] Suppose there are 4 dictionaries {β 1 , β 2 , β 3 , β 4}, each dictionary contains the same set of keys K, and for each key k∈K, its corresponding value is the vector

[0154] Corresponding to each dictionary β x and key k, calculate the vector The L2 norm of

[0155]

[0156] in, Representation vector The xth component of ;

[0157] Compute the sum of the L2 norms of each key k in all dictionaries

[0158]

[0159] Perform a weighted summation on each key k according to the ratio of the L2 norm to obtain a weighted average vector

[0160]

[0161] Finally, the weighted aggregate dictionary is

[0162]

[0163] get

[0164] Then ε j and the local model parameters of the respective edge servers Generate different personalized models by weighted aggregation

[0165]

[0166] Finally, the personalized model parameters will be obtained Again to the respective edge servers.

[0167] Step 106: The cloud server sends the personalized local partial model to the corresponding edge server again.

[0168] Step 107: Determine whether the specified number of iterations R is reached or the loss function L converges. The goal is The average value of the loss function of all edge servers reaches minimum convergence.

[0169] If not, repeat the above steps 101 to 106 until the specified number of iterations R or the loss function L converges, and the average value of the loss function of all edge servers reaches the minimum convergence, and then go to step 108.

[0170] Step 108: The cloud server notifies all edge servers to stop training.

[0171] In the above embodiments, by introducing a personalized edge computing offloading mechanism, a personalized offloading decision can be generated according to the characteristics of each edge device (such as computing power, data characteristics, application scenarios, etc.). This greatly improves resource utilization efficiency, so that each edge device can perform the optimal offloading operation according to its actual situation.

[0172] By adopting a personalized edge computing approach, only model parameters or gradient information are transmitted instead of original data, effectively protecting data privacy. This approach reduces the risk of data leakage and enhances the security of the system.

[0173] Through the personalized edge computing offloading algorithm, the local information of multiple edge devices can be integrated into the global model, improving the generalization and accuracy of the model. This fusion method enables the global model to better adapt to various complex scenarios and changing needs, improving the practicality of the model.

[0174] Due to personalized offloading decisions and dynamic resource allocation, the system can use resources more efficiently, reduce computing delays, and improve the overall performance of the system. This is especially important for application scenarios that require real-time response.

[0175] The basic principle of the technical solution is described above in conjunction with specific embodiments. However, it should be pointed out that the advantages, strengths, effects, etc. mentioned in the technical solution are only examples and not limitations, and it cannot be considered that these advantages, strengths, effects, etc. must be possessed by each embodiment of the technical solution. In addition, the specific details of the above technical solution are only for the purpose of illustration and facilitation of understanding, rather than limitation, and the above details do not limit the technical solution to the need to adopt the above specific details to be implemented.

[0176] Each embodiment in this specification is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the system embodiment, since it basically corresponds to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment.

[0177] The block diagrams of the devices, apparatuses, equipment, and systems involved in the present technical solution are only illustrative examples and are not intended to require or imply that they must be connected, arranged, and configured in the manner shown in the block diagrams. As will be appreciated by those skilled in the art, these devices, apparatuses, equipment, and systems may be connected, arranged, and configured in any manner. Words such as "including", "comprising", "having", and the like are open words, referring to "including but not limited to", and may be used interchangeably therewith. The words "or", "and", and "used herein refer to the words "and / or", and may be used interchangeably therewith, unless the context clearly indicates otherwise. The words "such as", used herein, refer to the phrase "such as but not limited to", and may be used interchangeably therewith.

[0178] The method and apparatus of the present technical solution may be implemented in many ways. For example, the method and apparatus of the present technical solution may be implemented by software, hardware, firmware, or any combination of software, hardware, and firmware. The above order of steps for the method is for illustration only, and the steps of the method of the present technical solution are not limited to the order specifically described above, unless otherwise specifically stated. In addition, in some embodiments, the present technical solution may also be implemented as a program recorded in a recording medium, which includes machine-readable instructions for implementing the method according to the present technical solution. Therefore, the present technical solution also covers a recording medium storing a program for executing the method according to the present technical solution.

[0179] It should also be noted that, in the apparatus, equipment and method of the present technical solution, each component or each step can be decomposed and / or recombined. These decompositions and / or recombinations should be regarded as equivalent solutions of the present technical solution. The above description of the aspects of the technical solution is provided to enable any technician in the field to make or use the present technical solution. Various modifications to these aspects are very obvious to those skilled in the art, and the general principles defined herein can be applied to other aspects without departing from the scope of the present technical solution. Therefore, the present technical solution is not intended to be limited to the aspects shown here, but in accordance with the widest range consistent with the principles and novel features of this technical solution.

[0180] The above description has been given for the purpose of illustration and description. In addition, this description is not intended to limit the embodiments of the present technical solution to the form of the technical solution herein. Although multiple example aspects and embodiments have been discussed above, those skilled in the art will recognize certain variations, modifications, changes, additions and sub-combinations thereof.

Claims

1. A method for offloading edge computing, It is characterized in that The following steps are involved: The cloud server initializes the parameters of the local local models of all edge servers and sends them to each edge server for training; The user device sends a computing task request to the edge server; The edge server receives the computing task request and trains the local model using a deep learning optimization algorithm based on the DQN model to obtain a personalized local model. The edge server sends the local model update increment to the cloud server, and the cloud server uses the local model update increment to update the hypernetwork model; The cloud server weights and aggregates the final super network through three layers of personalized super network models and one layer of global super network model, and then aggregates it with the local local models corresponding to each edge server to output a personalized local local model for the edge server; The cloud server sends the personalized local partial model to the corresponding edge server again; Repeat the above steps until the specified number of iterations R or the loss function L converges, and the average loss function of all edge servers reaches the minimum convergence; The cloud server notifies all edge servers to stop training.

2. The method for edge computing offloading according to claim 1, It is characterized in that The computing task request further includes: Task parameters: Indicates that any i belongs to M, from 1 to M, where R(i) represents the task request of the i-th terminal device, The definition of R(i) is (I i , C i , T i ), in the task request parameter information, I i Indicates the size of the input data, C i Indicates the number of CPU cycles required to complete the task, T i represents time, i represents the i-th terminal device; Network configuration parameters: Each terminal device consumes a certain amount of transmission energy when unloading, and the task unloading time and energy consumption is the key indicator considered, R(i) represents the computing task offloaded to the edge server, and the transmission power P of the terminal device in the current time block i , the channel gain h from i to the edge server i , channel bandwidth W i , Gaussian white noise power σ 2 ; Terminal configuration parameters: the number of calculation cycles required for the terminal to process each bit of data C_loc, the terminal's calculation frequency F_loc, the terminal's calculation energy efficiency ratio k_loc, the terminal's data transmission energy efficiency ratio k_t, the terminal's data transmission limit T, and the distance d from the terminal to its associated base station; Base station configuration parameters: the number of calculation cycles C_e required for the base station to process each bit of data, the calculation frequency F_e of the base station, the calculation energy efficiency ratio k_e of the base station, the noise power spectrum density No_e of the base station, and the bandwidth B_e of each base station.

3. The method for edge computing offloading according to claim 1, It is characterized in that The DQN model is a three-layer neural network, including an input layer, several hidden layers and an output layer, and is used to approximate the Q function. The parameters of the DQN model include weights and biases.

4. The method for offloading edge computing according to claim 1, It is characterized in that The cloud server updates the hypernetwork model using the increment of the local partial model update, further comprising: Updated increment Δθ j Expressed as: Δθ t =θ′ t -θ t , where θ′ t is the model parameter obtained by the edge server through local training, and θ t are the model parameters received from the cloud server, For the super network corresponding to edge server j The following is v j , ν and Update method: Where Δv j represents the change of the local hypernetwork embedding vector corresponding to the current edge server j, Δv represents the change of the global embedding vector, represents the change of partial network parameters corresponding to the current edge server j, represents the change of global hypernetwork parameters, The local model parameters corresponding to the current edge server j are represented by Δθ j represents the increment of the local model parameters corresponding to the current edge server j, {θ l1 ,θ l2 , ..., θ ln } represents the parameters of each layer of the local model corresponding to edge server j, They represent their respective gradients.

5. The edge computing offloading method according to claim 1, It is characterized in that Further including: The objective function of the cloud server, and the update rules of the embedding vector and hypernetwork parameters corresponding to edge server j, The local model parameters of each edge server j are obtained by weighted aggregation according to ε j : in It is expressed as: Through each layer of model parameters and corresponding ε j Aggregate to get the final model parameters of each layer. The objective function of the algorithm is expressed as follows: Get the average minimum value of the loss function of the final local model parameters under N edge servers, Where V = {v 1 , v 2 , ..., v N } is the embedding vector corresponding to each edge server j, is the hypernetwork parameter corresponding to each edge server j, ν is the global embedding vector, is the global hypernetwork parameter, and the algorithm converts the local model parameter θ j The optimization problem is converted into an embedding vector v j , local embedding vector V, hypernetwork parameters and the global hypernetwork parameters Here are the update rules: The above formula is used to calculate the local embedding vector v i The loss function for edge server j Gradient The loss function of the global embedding vector V with respect to edge server j is Gradient Global hypernetwork parameters The loss function for edge server j Gradient And calculate the hypernetwork parameters The loss function for edge server j Tax rate represents the local model parameters corresponding to the current edge server j, Δθ j represents the increment of the local model parameters corresponding to the current edge server j, {θ l1 ,θ l2 , ..., θ ln } represents the parameters of each layer of the local model corresponding to the current edge server, represent their respective gradients, express About loss function gradient.

6. The method for offloading edge computing according to claim 1, It is characterized in that The cloud server weights and aggregates the final super network through three layers of personalized super network models and one layer of global super network model, and then aggregates with the local local models corresponding to the edge servers, and outputs personalized local local models for the edge servers, further comprising: The final value of the hypernetwork is ε j , ε j Defined as The algorithm's aggregation weight matrix ε j The update process: Suppose there are 4 dictionaries {β 1 , β 2 , β 3 , β 4 }, each dictionary contains the same set of keys K, and for each key k∈K, its corresponding value is the vector Corresponding to each dictionary β x and key k, calculate the vector The L2 norm of in, Representation vector The xth component of ; Compute the sum of the L2 norms of each key k in all dictionaries Perform a weighted summation on each key k according to the ratio of the L2 norm to obtain a weighted average vector Finally, the weighted aggregate dictionary is get Then ε j and the local model parameters of the respective edge servers Generate different personalized models by weighted aggregation Finally, the personalized model parameters will be obtained Again to the respective edge servers.

Images