A joint optimization method of task offloading and service cache based on mobile edge computing

Abstract

The present invention belongs to the field of mobile communication technology, and specifically relates to a task offloading and service cache joint optimization method based on mobile edge computing; the method comprises: constructing a mobile edge computing system network model; constructing a communication model, a computing offloading model and a service cache model based on the mobile edge computing system network model; establishing a task offloading and service cache joint optimization problem according to the communication model, the computing offloading model and the service cache model; using a distributed edge collaborative offloading and service cache method based on a DPFL algorithm to solve the task offloading and service cache joint optimization problem, and obtaining a task offloading strategy and a service cache strategy; taking into account the differences in user preferences for applications, using a personalized federated learning algorithm to predict popular services based on historical request information, so that a MEC server can make better decisions, thereby achieving low task processing latency and device energy consumption, as well as a high cache hit rate.

Description

Technical Field

[0001] The present invention belongs to the technical field of mobile communications, and particularly relates to a joint optimization method for task offloading and service caching based on mobile edge computing. Background Art

[0002] With the development and integration of the global Internet and Internet of Things (IoT) technologies, the exponential growth of terminal devices and intelligent applications has been promoted. Various applications such as autonomous driving, intelligent healthcare, and virtual reality all require low latency and low energy consumption to ensure their high reliability and provide users with a high-quality experience. However, with the use of Beyond 5G networks, many small base stations are densely deployed in the cell, which may lead to a decline in network performance and backhaul congestion. At the same time, due to the long propagation delay between the user equipment and the remote cloud, the application frequently obtains services related to specific computing tasks through the cloud, so the application's demand for latency cannot be met. Mobile edge computing (MEC) is a shared application mode with communication, computing, and storage functions. Compared with cloud computing, MEC pushes the frontier of data and services from the centralized cloud infrastructure to the logical edge of the Internet, which can effectively solve problems such as latency and network congestion.

[0003] Using a cache in the MEC system to dynamically store services required by specific applications has become an effective method, which can save the bandwidth resources of the backhaul link and reduce the task processing latency. There are many works considering the research on related issues of computing offloading and edge caching under the MEC framework. For example, a study on the service caching and computing offloading problem in an MEC system uses mixed integer non-linear programming to jointly optimize the computing offloading decision and service cache placement, and designs a complexity-reducing alternating minimization technique to alternately update the offloading decision and cache placement strategy. A study on the multi-user computing offloading and data caching problem proposes a joint data caching and computing offloading strategy for a hybrid mobile cloud, and uses the alternating direction method of multipliers to obtain the optimal performance. A study on the joint optimization problem of service caching, computing migration, and resource allocation proposes a method based on semidefinite relaxation and an effective approximation algorithm for alternating optimization to solve the computing offloading and service caching problems in a multi-user multi-task scenario. However, in the Beyond 5G distributed network, due to the high autonomy of a large number of diverse edge nodes, for example, small base stations, Wi-Fi APs, there are different communication characteristics, caching capabilities, and computing capabilities among edge nodes, which may cause the problem that the cached content of edge nodes does not fully match the task requirements of users. Therefore, in different edge node service areas, edge nodes should pre-cache services that meet the task requirements of users according to the characteristics of heterogeneous tasks and user preferences. Summary of the Invention

[0004] Aiming at the deficiencies of the existing technology, the present invention proposes a joint optimization method for task offloading and service caching based on mobile edge computing, and the method includes:

[0005] S1: Construct a network model of the mobile edge computing system;

[0006] S2: Based on the network model of the mobile edge computing system, construct a communication model, a computing offloading model and a service caching model;

[0007] S3: Establish a joint optimization problem of task offloading and service caching according to the communication model, the computing offloading model and the service caching model;

[0008] S4: Use a distributed edge collaborative offloading and service caching method based on the DPFL algorithm to solve the joint optimization problem of task offloading and service caching, and obtain a task offloading strategy and a service caching strategy.

[0009] Preferably, the network model of the mobile edge computing system includes: N edge nodes, and the edge node set is defined as An MEC server is deployed on each edge node; there are U user devices in the service area of each edge node, and the user set is defined as There are Q service types in the network, and the service type set in the network is defined as In edge node n, the task of the i-th user device u i,n is μ i,n (t) = (g i,n (t), q i,n (t), ρ i,n (t), τ i,n (t)), where g i,n (t) is the size of the computing task, q i,n (t) is the service required by the task, ρ i,n (t) is the number of CPU cycles required per byte to process the task, and τ i,n (t) represents the maximum delay that can be tolerated to complete the task; the tasks of all users under the edge node are represented as μ n (t) = {μ 1,n (t), μ 2,n (t), …, μ U,n (t)}.

[0010] Preferably, the communication model includes: The task is sent from the user device to the edge node through the wireless uplink channel and is processed by the MEC server deployed by the edge node.

[0011] Preferably, the computing offloading model includes: any edge node has four computing task processing methods, and different task processing methods have different task processing delays and user equipment energy consumption; the four task processing methods are: local computing, offloading to an associated node for processing, forwarding the offloaded task to an adjacent node through the associated node for processing, and offloading to the cloud for processing.

[0012] Preferably, the service caching model includes: using binary variables to represent the service caching decision of edge node n at time slot t; if edge node n caches service q i,n , then otherwise, Within time slot t, the service caching decision of edge node n is expressed as

[0013] Preferably, the process of establishing the joint optimization problem of task offloading and service caching includes: constructing a cost-minimization optimization objective function based on the task processing delay and user equipment energy consumption under different task processing methods; constructing optimization constraint conditions, including task processing delay constraints, service storage capacity constraints, and task offloading decision constraints; establishing the joint optimization problem of task offloading and service caching according to the cost-minimization optimization objective function and the joint optimization constraint conditions.

[0014] Furthermore, the joint optimization problem of task offloading and service caching is expressed as:

[0015]

[0016]

[0017]

[0018]

[0019]

[0020] where, Φ n (t) represents the total cost of completing all user computing tasks under edge node n at time slot t, represents the service caching decision of edge node n at time slot t, represents the offloading decision of edge node n at time slot t, T i,n (t) represents the task processing delay of edge node n at time slot t, τ i,n (t) represents the maximum tolerable task processing delay of edge node n at time slot t, represents the set of edge nodes, represents the set of user equipment, represents whether edge node n caches service q at time slot t, cq Denote the storage space of service q, C n Denote the service storage space of edge node n Denote the set of service types Denote the offloading decision of user equipment i under edge node n at time slot t, and N denotes the number of edge nodes.

[0021] Preferably, the process of solving the joint optimization problem of task offloading and service caching includes:

[0022] Describe the joint optimization problem of task offloading and service caching as a Markov decision problem and train a DQN model to obtain a trained DQN model; use the DQN model as the global model and train the global model using a personalized federated learning training model to obtain a trained global model; edge nodes obtain task offloading policies and service caching policies from the trained global model.

[0023] Furthermore, the process of training the DQN model includes:

[0024] Regard the edge node as a DQN agent, initialize the state, action, and reward of the agent, and initialize the estimation network and the target network;

[0025] Obtain task offloading policies and service caching policies according to the current state in the estimation network, execute actions according to the task offloading policies and service caching policies and update the reward, and enter the next state;

[0026] Generate experiences according to the current state, next state, action, and reward, sample multiple experiences to train the estimation network and the target network, and obtain trained estimation network and target network;

[0027] The formula for updating the reward is:

[0028]

[0029] where, R n (t) represents the reward value, χ n (t) represents whether the service stored by edge node n exceeds the total capacity, U represents the number of user devices, β i,n (t) represents whether the task processing delay meets the terminal requirements, Φ n (t) represents the total cost of completing all user computing tasks under edge node n within time slot t.

[0030] Furthermore, the process of training the global model using a personalized federated learning training model includes:

[0031] Take the edge nodes as clients and the cloud server as the central controller; within a decision-making cycle, multiple clients download the global model and service caching policy from the central controller and use the downloaded global model as the local model.

[0032] Each client trains a personalized model according to the current service demand; updates the local model based on the personalized model and uploads the updated local model to the central controller.

[0033] The central controller updates the global model based on multiple local models to obtain a trained global model.

[0034] The beneficial effects of the present invention are as follows: Considering factors such as users' differentiated needs and dynamic time-varying network environments, aiming at minimizing task processing latency and device energy consumption, the present invention proposes a joint optimization method for task offloading and service caching based on mobile edge computing, and proposes a distributed edge collaborative offloading and service caching method based on deep reinforcement learning and personalized federated learning. Personalized federated learning can solve the problems of FL in non-independent and identically distributed data and shared models, and through training the DQN model, perform additional training on the data of local edge nodes. Considering the differences in users' application preferences, use the personalized federated learning algorithm to predict popular services based on historical request information, enable the MEC server to make better decisions, achieve the goal of regional model personalization, provide adaptive service caching policies for users in heterogeneous regions, and protect users' data privacy at the same time. A large number of simulation results show that the proposed scheme of the present invention can achieve lower latency and device energy consumption, as well as higher cache hit rates, verifying the effectiveness of the proposed scheme. Description of the Drawings

[0035] Figure 1 It is a flowchart of the joint optimization method for task offloading and service caching based on mobile edge computing in the present invention;

[0036] Figure 2 It is a schematic diagram of the network model of the mobile edge computing system in the present invention;

[0037] Figure 3 It is a block diagram of the personalized federated learning model training in the present invention;

[0038] Figure 4 It is a performance comparison diagram between the present invention and the comparative method. Detailed Embodiments

[0039] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0040] The present invention proposes a joint optimization method for task offloading and service caching based on mobile edge computing, as Figure 1 shown, the method includes the following contents:

[0041] S1: Construct a network model of a mobile edge computing system.

[0042] As Figure 2 shown, the network model of the mobile edge computing system includes N edge nodes, and the edge node set is defined as Each edge node is deployed with an MEC server, enabling the edge node to have certain storage and computing capabilities. The maximum CPU frequency of the MEC server on the edge node is f n , and the storage capacity is C n , which is used to store services associated with specific computing tasks; there are U user devices in the service area of each edge node, and the user set is defined as There are Q service types in the network, and the service type set in the network is defined as In edge node n, the task of the i-th user device u i,n is μ i,n (t)=(g i,n (t),q i,n (t),ρ i,n (t),τ i,n (t)), where g i,n (t) is the size of the computing task, q i,n (t) is the service required by the task, ρ i,n (t) is the number of CPU cycles required per byte for processing the task, and τ i,n (t) represents the maximum latency that can be tolerated to complete the task; the tasks of all users under the edge node are represented as μ n (t)={μ 1,n (t),μ 2,n (t),…,μ U,n (t)}.

[0043] S2: Construct a communication model, a computing offloading model, and a service caching model based on the network model of the mobile edge computing system.

[0044] The communication model includes: The task is sent from the user equipment to the edge node through the wireless uplink channel and processed by the MEC server deployed by the edge node. The present invention mainly studies the transmission delay of the uplink and the task calculation delay. Due to the high downlink transmission rate between the MEC server and the user equipment, the present invention ignores the download delay of the calculation result.

[0045] Device u i,n The data transmission rate between it and the edge node n can be calculated as:

[0046]

[0047] where B n is the channel bandwidth of the edge node n, N s is the maximum number of service devices of the channel, P i,n (t) is the transmission power of the user equipment u i,n H i,n represents the channel gain between the user equipment u i,n and the edge node n, and σ 2 is the Gaussian channel noise.

[0048] The computing offloading model includes: Any edge node has four computing task processing methods, and different task processing methods have different task processing delays and user equipment energy consumption; the four task processing methods are: local computing, offloading to an associated node for processing, forwarding the offloaded task to an adjacent node for processing through an associated node, and offloading to the cloud for processing.

[0049] Define the task offloading decision of the edge node n at time slot t as where represents local computing in the device u i,n , represents that the edge node n processes the task request of the device u i,n , represents that the computing task of the device u i,n is offloaded to the cloud server for processing.

[0050] 1) Local computing:

[0051] A user equipment with sufficient power can execute the computing task by itself. Assuming that the device has all the services required to process its task, the computing frequency of the device u i,n is f i,n . Therefore, the processing delay i,n of the task μ for local computing can be expressed as:

[0052]

[0053] Device u i,n Local computing energy consumption is as follows:

[0054]

[0055] Among them, e is the energy consumption coefficient depending on the chip architecture. Preferably, take e = 10 -8 .

[0056] 2) Offload to the associated node for processing

[0057] The local associated node n executes the computing task of device u i,n . If the associated node n caches the service q i,n required for task μ i,n (t), the processing delay i,n of task μ can be expressed as:

[0058]

[0059] 3) Forward the offloaded task to the adjacent node for processing through the associated node

[0060] Task μ i,n (t) can also be offloaded from the associated node n to the adjacent node m for execution by wire. If the adjacent node m caches the service q i,n , using to represent the average transmission rate from the associated node n to the adjacent node m. The transmission rates between different nodes are different and are related to the distance between them. Therefore, the processing delay i,n of task μ can be expressed as:

[0061]

[0062] 4) Offload to the cloud for processing

[0063] When the edge node cannot provide the service required for the user to execute the task, the service needs to be obtained from the cloud server. Since the cloud server has sufficient cloud computing resources and powerful computing capabilities, therefore, the computing delay of the cloud server is ignored in the present invention. Using the offloading decision to represent offloading the computing task of device u i,n to the cloud server for processing, the processing delay i,n of task μ can be expressed as

[0064]

[0065] Among them, ψ represents the end-to-end delay between the local edge node n and the cloud server.

[0066] For the task processing methods of categories 2-4, device u i,n The energy consumption only occurs during the process of sending the computing task to the local edge node n. Therefore, the energy consumed by the user device is all:

[0067]

[0068] The service caching model includes: the edge node can reduce the task processing delay by placing specific services required for computing tasks. It is assumed in the present invention that all edge nodes in the network can exchange service caching information, and the cloud server has sufficient computing and storage resources to store all services. Define the storage space of each service as c q . The binary variable is used to represent the service caching decision of edge node n at time slot t; if edge node n caches service q i,n , then otherwise, During time slot t, the service caching decision of edge node n is expressed as Since the user's preference for tasks varies in different time slots, each edge node needs to provide applicable caching decisions for the users in its respective area

[0069] S3: Establish a joint optimization problem of task offloading and service caching according to the communication model, computing offloading model and service caching model.

[0070] Construct a cost-minimizing optimization objective function according to the task processing delay and user device energy consumption under different task processing methods.

[0071] Through the above analysis, it can be seen that in the MEC system, the user's service experience is determined by the task completion delay and energy consumption. The task processing delay within time slot t can be expressed as:

[0072]

[0073] The energy consumption of the device is expressed as:

[0074]

[0075] Define λ i,n and (1 - λ i,n ) to represent the preference parameters of the delay and energy consumption of task μ i,n (t) respectively. Considering the preferences of different users, define the task execution cost as the weighted sum of the delay and energy consumption. Then the cost consumed for processing task μ i,n is:

[0076] Φ i,n (t) = λ i,n T i,n (t) + (1 - λ i,n )E i,n (t)

[0077] Therefore, within the service range of the edge node n, the total cost of completing all user computing tasks is:

[0078]

[0079] Construct optimization constraints, including task processing delay constraints, service storage capacity constraints, and task offloading decision constraints. Establish a joint optimization problem of task offloading and service caching based on the cost minimization optimization objective function and the joint optimization constraints. Specifically:

[0080] According to different offloading and caching strategies, the processing delay T i,n and energy consumption E i,n of the task may be different. The objective of the present invention is to make full use of distributed edge computing resources, design a collaborative task offloading and service caching scheme, minimize the task processing delay and device energy consumption while meeting the maximum tolerable delay of the computing task. The joint optimization problem of task offloading and service caching established by the present invention is described as:

[0081]

[0082]

[0083]

[0084]

[0085]

[0086] Among them, Φ n (t) represents the total cost of completing all user computing tasks under the edge node n in time slot t, represents the service caching decision of the edge node n in time slot t, represents the offloading decision of the edge node n in time slot t, T i,n (t) represents the task processing delay of the edge node n in time slot t, τ i,n (t) represents the maximum tolerable delay of the task processing of the edge node n in time slot t, represents the offloading decision of the user device i under the edge node n in time slot t; The C1 constraint ensures that the task μ i,n(t) does not exceed the maximum tolerable time delay, and the C2 constraint ensures that the memory occupied by the services placed at the edge node n does not exceed the total storage capacity C n . The C3 constraint is used to determine the caching decision of the edge node n for the service q. The C4 constraint states that only one offloading decision can be selected for each task.

[0087] S4: Use the distributed edge collaborative offloading and service caching method based on the DPFL algorithm to solve the joint optimization problem of task offloading and service caching, and obtain the task offloading strategy and service caching strategy.

[0088] In the state where future information cannot be obtained, such as the arrival of random tasks and dynamic and complex network conditions, it is difficult for edge nodes to obtain the optimal offloading decision and service caching placement strategy. In addition, the caching strategy depends on the frequency of requests for various applications by local users in the current area, and there are certain differences in the caching strategies of different areas. Edge nodes can predict services by collecting historical request information of local users, but there will be a serious risk of privacy leakage in the process of collecting user information. Therefore, the present invention proposes a distributed edge collaborative offloading and service caching method based on deep reinforcement learning and personalized federated learning (DPFL algorithm), and the edge node is used as a DQN agent to learn the task offloading strategy and service caching strategy.

[0089] Deep reinforcement learning algorithm (DQN):

[0090] Describe the joint optimization problem of task offloading and service caching as a Markov decision problem and train the DQN model to obtain a trained DQN model.

[0091] Due to the heterogeneity of the service areas of edge nodes, edge nodes need to make computing offloading and service caching decisions according to the continuously generated task requests of users. Its state space and action space are huge, and it will face the problem of high-dimensional space; the present invention uses the DQN algorithm to solve the above problems, specifically:

[0092] Take the edge node as a DQN agent, initialize the state, action and reward of the agent, and initialize the estimation network and target network.

[0093] 1) State space: In time slot t, the environmental state that the system can observe includes the task information in all edge nodes, the computing resources of each edge node, the service caching strategy and storage capacity, as well as the channel bandwidth. Therefore, the state space of the system environment can be expressed as:

[0094]

[0095] Among them, f N (t) represents the computing frequency of the edge node N, C NDenotes the service cache storage capacity of the edge node N, B N Denotes the channel bandwidth of the edge node N.

[0096] 2) Observation space: The observation space of the agent is defined as the current information observed by the agent in the edge network. At the initial moment of time slot t, the agent receives task information (including task size, requested service of the task, number of CPU cycles required by the task, and maximum tolerable delay), computing frequency from the devices within its coverage area, and takes its own computing resources and service cache as the observation state. Then the agent determines the task offloading decision and service cache placement strategy. Therefore, the partial observation space of agent n can be described as:

[0097]

[0098] 3) Action space: The tasks requested by users are variable. To adapt to the continuous changes in the dynamic environment, the agent decides where to process the task requests (including user devices, edge nodes, or cloud servers), and selects the services to be cached. Define A n (t) as the system action of the state space, then the action space of agent n at time slot t can be described as

[0099] 4) Reward function: According to the joint optimization problem of task offloading and service caching, the goal of each agent is to minimize the task delay and energy consumption while satisfying the limited storage resource C n and the maximum tolerable delay. When the agent takes an action at time slot t - 1, the corresponding reward is fed back to the agent at time slot t. Based on the obtained reward, the agent updates its policy to obtain the optimal result. Since the reward causes the agent to reach the optimal policy, and the policy directly determines the computing offloading decision and service caching policy of the MEC server; the reward return of agent n can be calculated by the following formula:

[0100]

[0101] where, R n (t) represents the reward value, represents whether the service stored by edge node n exceeds the total capacity, β i,n (t) = η 2 H(τ i,n (t) - T i,n (t)) represents whether the task processing delay meets the terminal requirements; H(·) is the step function, η 1 and η 2 are weight coefficients. Therefore, if the constraints of storage capacity and delay requirements are met, then there are χ n (t) = η 1 and β i,n (t) = η2 。

[0102] Obtain the task offloading policy and service caching policy according to the current state in the estimation network, execute actions according to the task offloading policy and service caching policy, and update the reward, then enter the next state;

[0103] Generate experiences based on the current state, next state, actions and rewards, sample multiple experiences to train the estimation network and the target network, and obtain the trained estimation network and target network.

[0104] Personalized federated learning model:

[0105] Personalized federated learning allows devices or edge nodes to jointly train a global model under the coordination of a cloud server in a cloud-edge paradigm. After learning and training the global model, personalized learning methods can be adopted on the client side to deploy personalized models according to the user needs in different regions. Therefore, the present invention uses the pFedMe algorithm (personalized federated learning) to predict popular services, provides a personalized service caching scheme for the differentiated needs of users, and protects the data privacy of users at the same time. The pFedMe algorithm formulates personalized federated learning as the following two-layer problem:

[0106]

[0107] where θ n represents the personalized model of edge node n, ζ represents the global model, and α represents the regularization parameter that controls the strength ζ of the personalized model. A larger α (ensuring that θ n and ζ are as similar as possible) can benefit from unreliable data in rich data aggregation, and a smaller α can help clients with sufficient valid data to perform personalization first. The pFedMe method allows edge nodes to update the local model in different directions while not deviating from ζ.

[0108] Use the DQN model as the global model, and train the global model using the personalized federated learning training model to obtain the trained global model; specifically:

[0109] As Figure 3 shown, the present invention takes edge nodes deployed in multiple regions as clients and a cloud server as a central controller; at the beginning of each decision period t, N clients first receive the current service caching policy of the global model W t (weights of DQN);

[0110] Each client trains the personalized model θ i,n according to the current service demand q n ; expressed as:

[0111]

[0112] Among them, represents the d-dimensional real number set, and y n (θ n ) represents the expectation on the data distribution of client n.

[0113] The client uses the parameter δ to approximate and updates the local model The updated local model is uploaded to the central controller.

[0114] The central controller updates the global model according to the received multiple local models to obtain a trained global model, and updates the global model with the parameter σ, which is expressed as:

[0115]

[0116] The client downloads the updated global model W t+1 , obtains the task offloading policy and service caching policy; updates the local cached service by obtaining services from the cloud server according to the received service caching policy, executes the corresponding task offloading policy and service caching policy, and then the client enters the next decision cycle.

[0117] Evaluate the present invention: To prove the performance advantages of the DPFL algorithm proposed by the present invention, the present invention studies and compares the performance effects of the distributed edge collaborative offloading and service caching (DECM) scheme under multiple algorithms, including the DQN algorithm, the DFL algorithm (the fusion algorithm of DQN and federated learning), and the DPFL algorithm. As Figure 4As shown in the figure, it can be seen that the DPFL algorithm has the best performance in terms of cache hit rate, latency, energy consumption, and revenue. Since the MEC system continuously learns the service preferences of users in the current area, the DPFL algorithm predicts the probability of future services being requested by training a neural network, thus providing a good basis for the personalized caching strategy of each edge node. The service prediction model trained by the DFL algorithm is shared among edge nodes, and edge nodes cannot solve the problem of differential requirements and incomplete matching of cached content according to the user preferences in the current area. The DQN algorithm cannot enable the system to learn service popularity, so it cannot capture the changes in service popularity brought about by user differential requirements, resulting in the lowest cache hit rate. After the system training is stable, most tasks are calculated on local edge nodes, and the average latency and energy consumption of tasks gradually decrease and converge with the number of iterations. Since the DPFL algorithm provides a service caching strategy for each edge node to adapt to local users, local edge nodes can handle most user tasks, so the task latency and energy consumption are the lowest and converge first. In summary, compared with the existing comparison methods, the DPFL algorithm proposed in the present invention has lower latency and device energy consumption and higher cache hit rate.

[0118] The above embodiments further illustrate the purpose, technical solutions, and advantages of the present invention. It should be understood that the above embodiments are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made to the present invention within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A task offloading and service cache joint optimization method based on mobile edge computing, characterized in that: include: S1: Build a network model for mobile edge computing system; The network model of the mobile edge computing system includes: N edge nodes, defining the edge node set Each edge node is deployed with a MEC server; each edge node has U user devices in its service area, defining the user set There are Q types of services in the network, and the service type set in the network is defined as In edge node n, the i-th user device u i,n The task is μ i,n (t)=(g i,n (t),q i,n (t),ρ i,n (t),τ i,n (t)), where g i,n (t) is the size of the computational task, q i,n (t) is the service required by the task, ρ i,n (t) is the number of CPU cycles required to process each byte of the task, τ i,n (t) represents the maximum tolerable delay for completing the task; the tasks of all users under the edge node are represented by μ n (t) = {μ 1,n (t),μ 2,n (t),…,μ U,n (t)}; S2: Based on the network model of the mobile edge computing system, a communication model, a computing offloading model and a service cache model are constructed; the communication model includes: tasks are sent from the user equipment to the edge node through the wireless uplink channel and processed by the MEC server deployed on the edge node; the computing offloading model includes: any edge node has four computing task processing modes, and different task processing modes have different task processing delays and user equipment energy consumption; the four task processing modes are: local computing, offloading to associated nodes for processing, forwarding the offloaded tasks to adjacent nodes for processing through associated nodes, and offloading to the cloud for processing; the service cache model includes: using binary variables represents the service caching decision of edge node n in time slot t; if edge node n caches service q i,n ,but otherwise, In time slot t, the service cache decision of edge node n is expressed as S3: Establish the joint optimization problem of task offloading and service caching according to the communication model, computation offloading model and service caching model; the process of establishing the joint optimization problem of task offloading and service caching includes: constructing the cost minimum optimization objective function according to the task processing delay and user equipment energy consumption under different task processing modes; constructing optimization constraints, including task processing delay constraints, service storage capacity constraints and task offloading decision constraints; establishing the joint optimization problem of task offloading and service caching according to the cost minimum optimization objective function and the joint optimization constraints; the joint optimization problem of task offloading and service caching is expressed as: Among them, Φ n (t) represents the total cost of completing all user computing tasks under edge node n in time slot t, represents the service cache decision of edge node n at time slot t, represents the offloading decision of edge node n at time slot t, T i,n (t) represents the task processing delay of edge node n in time slot t, τ i,n (t) represents the maximum tolerable delay of task processing by edge node n in time slot t, represents the set of edge nodes, Represents a collection of user devices. Indicates whether edge node n caches service q in time slot t, c q represents the storage space of service q, C n represents the service storage space of edge node n, Represents a service type set, represents the offloading decision of user equipment i under edge node n in time slot t, and N represents the number of edge nodes; S4: The distributed edge collaborative offloading and service caching method based on the DPFL algorithm is used to solve the joint optimization problem of task offloading and service caching, and the task offloading strategy and service caching strategy are obtained. The process of solving the joint optimization problem of task offloading and service caching includes: The task offloading and service cache joint optimization problem is described as a Markov decision problem and the DQN model is trained to obtain a trained DQN model; the DQN model is used as a global model, and a personalized federated learning training model is used to train the global model to obtain a trained global model; the edge node obtains the task offloading strategy and the service cache strategy from the trained global model; the process of training the DQN model includes: Use the edge node as a DQN agent, initialize the agent's state, action, and reward, and initialize the estimation network and target network; In the estimation network, the task offloading strategy and service cache strategy are obtained according to the current state, and actions are executed and rewards are updated according to the task offloading strategy and service cache strategy, and the next state is entered; Generate experience based on the current state, next state, action, and reward, sample multiple experiences to train the estimation network and target network, and obtain the trained estimation network and target network; The formula for updating the reward is: Among them, R n (t) represents the reward value, χ n (t) indicates whether the services stored in edge node n exceed the total capacity, U indicates the number of user devices, β i,n (t) indicates whether the task processing delay meets the terminal requirements, Φ n (t) represents the total cost of completing all user computing tasks under edge node n in time slot t; The process of training a global model through personalized federated learning training includes: The edge node is used as the client and the cloud server is used as the central controller. In one decision cycle, multiple clients download the global model and service cache policy from the central controller and use the downloaded global model as the local model. Each client trains a personalized model θ based on the current service requirements n , expressed as: in, represents a d-dimensional real number set, y n (θ n ) represents the expectation of the data distribution of client n, ζ represents the global model, and α represents the regularization parameter that controls the strength of the personalized model ζ; the client uses parameter δ to approximate And update the local model The updated local model Upload to the central controller; The central controller updates the global model based on multiple local models to obtain a trained global model, which is expressed as: Among them, the client downloads the updated global model W t+1 , σ represents the parameters used to update the global model.

Images