A method for task migration based on multi-point cooperation in mobile edge computing

Abstract

The application belongs to the technical field of mobile communication, and particularly relates to a task migration method based on multi-point cooperation in mobile edge computing; the method comprises the following steps: constructing a network system model under a MEC scenario; constructing a cooperative communication model, a task computing model, a user mobility model and a load balancing model based on the network system model under the MEC scenario; constructing an offloading decision and a cooperative set association decision joint optimization problem according to the cooperative communication model, the task computing model, the user mobility model and the load balancing model; on a slow time scale, facing future load changes, performing node cooperation to form a cooperative set; solving the offloading decision and the cooperative set association decision joint optimization problem on a fast time scale according to the cooperative set to obtain an optimal offloading decision and an optimal cooperative set association decision; and the system performs task migration according to the optimal offloading decision and the optimal cooperative set association decision; the application can effectively reduce a task migration rate and a task execution time delay.

Description

Technical Field

[0001] This invention belongs to the field of mobile communication technology, specifically relating to a task migration method based on multi-point collaboration in mobile edge computing. Background Technology

[0002] With the increasing prevalence of smart devices and the Internet of Things (IoT), the demand for mobile data rates and server computing power from various emerging services is growing exponentially. Mobile Edge Computing (MEC) offloads computing and storage resources from cloud computing to edge devices closer to users, providing them with low-latency and highly reliable services. To meet the explosive high-traffic demands of users in hotspot areas, ultra-dense edge computing networks deploy a large number of base stations and servers in these areas to provide services. However, due to the dense deployment of base stations and servers, the coverage area of ​​nodes is reduced. In scenarios where a single node completes the task transmission and computing, high-speed mobile users can easily leave the coverage area of ​​the original offloading base station, leading to frequent handovers and task migrations, resulting in task interruptions or failures.

[0003] In a MEC environment, user mobility is a crucial factor affecting the quality of service. During high-speed user movement, issues such as task offloading failures, frequent switching, and task interruptions may occur. By considering cooperation among multiple nodes and utilizing multi-point cooperative transmission technology, wireless access points within a multi-point cooperative set can simultaneously provide reliable communication services to users, thereby expanding the service range. Numerous studies have been conducted on node-based collaborative task migration in MEC environments. For example, to address the frequent migration issues caused by high-speed user movement, a machine learning-based mobility awareness algorithm has been proposed to predict future service migrations and handover nodes, enabling advance task migration and the establishment of multi-point collaborative sets to reduce task interruptions caused by user handovers. Regarding resource allocation in MEC environments with CoMP reception capabilities, this problem is modeled as a mixed-integer nonlinear programming problem. A three-stage resource allocation algorithm based on the interference graph concept is proposed, including: computational resource allocation, subcarrier allocation and cell clustering, subcarrier reuse, and cell clustering. For the deployment of Service Function Chains (SFCs) in 6G wireless edge networks, a deployment algorithm based on the Actor-Critic framework is proposed to solve the long-term cost minimization SFC deployment problem by utilizing multi-point collaborative techniques to reduce interference between service chains.

[0004] The above-described user-centric collaborative set formation mechanism requires continuous prediction of user movement trajectories. However, in multi-user scenarios, the controller needs to collect and store a large amount of user trajectory information, which leads to significant prediction costs. Furthermore, the number of collaborative sets established based on users increases with the number of users, resulting in excessively high computational decision-making costs. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention proposes a task migration method based on multi-point collaboration in mobile edge computing, which includes:

[0006] S1: Construct a network system model in the MEC scenario;

[0007] S2: Based on the network system model in the MEC scenario, construct a collaborative communication model, a task computing model, a user mobility model, and a load balancing model;

[0008] S3: Construct a joint optimization problem of offloading decision and collaborative set association decision based on the collaborative communication model, task computing model, user mobility model and load balancing model;

[0009] S4: A two-time-scale decision model is used to solve the joint optimization problem of unloading decision and collaboration set association decision, and the optimal unloading decision and collaboration set association decision are obtained; the system performs task migration based on the optimal unloading decision and collaboration set association decision.

[0010] Preferably, the network system model in the MEC scenario specifically includes: U base stations and M users, with each base station equipped with an MEC server; the set of base stations is represented as follows. The user set is represented as The user's task is represented as in, Indicates the size of the task data. This indicates the number of CPU cycles required per bit of task. This indicates the maximum tolerable latency for the task. This represents the movement speed of user u in time slot t; the user selects a cooperative set. Unload task, define Represents the associated variable of the collaborative set. This indicates that the user has chosen to collaborate in the set. Perform the task. This indicates that the user did not select a collaboration set. definition This represents the decision variable for task unloading. This indicates that the task will be offloaded to the edge for execution. This indicates that the task will be executed locally.

[0011] Preferably, the user mobility model includes: using user dwell time as an evaluation indicator for whether the task can be successfully transmitted during cell handover; if the user's dwell time in the cooperative set... Less than the task transmission time Task transmission failed; the probability of task transmission failure is predicted based on the dwell time.

[0012] Preferred load balancing models include:

[0013] The predicted computing and storage resource usage of the server is calculated using an LSTM network. The expected load of the server is calculated based on the predicted computing and storage resource usage. An expected load threshold is set. When the expected load of the server is greater than the expected load threshold, the server is designated as a high-load node; otherwise, the server is designated as a low-load node.

[0014] High-load nodes are used to offload excess computing tasks to other neighboring nodes and do not accept computing tasks from other nodes. Low-load nodes are used to receive computing load from other nodes and do not need to offload computing tasks to other nodes. High-load and low-load nodes share computing resources and form a cooperative set, defining the expected load deviation of the cooperative set.

[0015] Preferably, the joint optimization problem of unloading decision and collaboration set association decision is expressed as:

[0016]

[0017]

[0018]

[0019]

[0020]

[0021]

[0022] in, This represents the user latency cost for successful execution of a task by user u in time slot t. Let T represent the set of time slots, and T represent the system time period. Represents a set of users. Represents a collaborative set. This represents the maximum tolerable delay for user u's task in time slot t. This represents the task offloading decision variable for user u in time slot t. This represents the collaborative set association variable for user u in time slot t. This indicates that user u in time slot t receives data from the cooperative set. The signal-to-interference-plus-noise ratio (SIR / NOT) of the signal, where p0 represents the SIR threshold. This represents the energy consumption of user equipment in time slot t. This represents the user equipment energy consumption budget.

[0023] Preferably, the process of solving the joint optimization problem of unloading decision and collaboration set association decision includes: a two-timescale decision model including a slow-timescale model and a fast-timescale model; the slow-timescale model updates the collaboration set according to the load information using a collaboration mechanism based on FCM algorithm and alliance game; the fast-timescale model solves the optimal unloading decision and collaboration set association decision using an Actor-Critic-based migration decision algorithm based on the collaboration set and user task information.

[0024] Furthermore, the process of updating the cooperation set using a cooperation mechanism based on the FCM algorithm and coalition game theory includes:

[0025] The FCM algorithm is used to cluster all server nodes to obtain K collaboration sets;

[0026] Repeatedly perform the merge and split process on K collaboration sets until the final collaboration set no longer changes;

[0027] Merging: Set the coalition utility function based on the expected load deviation of the cooperative set. Compare the Pareto advantages of the merged set and multiple cooperative sets based on the coalition utility function. If the Pareto advantage of the merged set is greater than the Pareto advantage of multiple cooperative sets, the merged set is taken as the new cooperative set. Otherwise, multiple cooperative sets are retained. The merged set is the union of multiple cooperative sets.

[0028] Splitting: Compare the Pareto advantage of the split subset with that of the original cooperative set. If the Pareto advantage of the split subset is greater than that of the original cooperative set, the split subset is adopted as the new cooperative set; otherwise, the original cooperative set is retained. The split subset is a subset of the original cooperative set.

[0029] Furthermore, the process of solving the optimal unloading decision and cooperative set association decision using the Actor-Critic-based transfer decision algorithm includes: abstracting the joint optimization problem of unloading decision and cooperative set association decision into a Markov decision process, with the base station acting as the agent, and constructing corresponding state space, action space, and reward function; each agent has an Actor network and a Critic network; both the Actor network and the Critic network consist of two neural networks with identical structures; the Actor network generates corresponding actions based on the current local observation state of a single agent and updates the reward function based on the actions, then enters the next state; the Critic network takes the actions output by the Actor network as input and adjusts the actions of the Actor network through the output policy gradient; empirical information is generated based on the current state, the next state, the actions, and the reward value; multiple empirical information samples are used to train the Actor network and the Critic network, and the network parameters are updated to obtain the trained Actor network and Critic network; the optimal unloading decision and cooperative set association decision are obtained based on the Actor network training results.

[0030] Furthermore, the reward function is:

[0031]

[0032] Among them, R t This represents the reward value obtained by the user performing an action in time slot t. This represents the user latency cost for successful execution of a task by user u in time slot t. This represents a set of users.

[0033] The beneficial effects of this invention are as follows: This invention proposes a task migration method based on multi-point collaboration in mobile edge computing for high-speed mobile scenarios. By constructing a collaborative communication model, a task computing model, a user mobility model, and a load balancing model, it constructs a joint optimization problem of offloading decision and collaboration set association decision. A two-time-scale decision model is designed to solve the node collaboration problem and the user task migration problem at two different time scales. Specifically, at the slow time scale, node collaboration is carried out in response to future load changes to form a collaboration set. Based on the collaboration set, the joint optimization problem of offloading decision and collaboration set association decision is solved at the fast time scale. Compared with the prior art, this invention establishes a collaboration set in response to future load changes, avoiding the "ping-pong effect" generated by previous online collaboration mechanisms. In addition, task migration based on multi-point collaboration can expand the service scope and aggregate computing capabilities, thereby effectively reducing the task migration rate and task execution latency. Attached Figure Description

[0034] Figure 1 This is a flowchart of the task migration method based on multi-point collaboration in mobile edge computing in this invention;

[0035] Figure 2 This is a schematic diagram of the network system model in the MEC scenario of this invention;

[0036] Figure 3 This is a structural diagram of the two-time-scale decision-making model in this invention;

[0037] Figure 4 The graph shows the average execution delay of tasks under different node distribution densities for the present invention and the comparative method. Detailed Implementation

[0038] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0039] This invention proposes a task migration method based on multi-point collaboration in mobile edge computing, such as... Figure 1 As shown, the method includes the following:

[0040] S1: Construct a network system model in the MEC scenario.

[0041] This invention can be applied to high-speed mobile scenarios. In the network system model of the MEC scenario constructed in this invention, there are U base stations and M users. Each base station is equipped with an MEC server, referred to as an edge service node. The computing capabilities of different servers are heterogeneous; for example... Figure 2 As shown, when a user is at location 1, they transmit tasks to a collaborative set of BS1, BS2, and BS3 using multi-point cooperative transmission technology. Together, these BS3 and BS2 provide communication services to the user. Servers deployed on the base stations also collaborate to provide computing services to the user. Servers within the collaborative set can share computing resources to ensure load balancing and prevent server overload. The set of base stations in the network is represented as follows: The user set is represented as The time period of the entire system is discretized as follows: The length of each time slot is τ.

[0042] Considering the random arrival of user tasks, the user task is represented as follows: in, Indicates the size of the task data (bits). This represents the number of CPU cycles required per bit of task (cycles / bit). This indicates the maximum tolerable latency for the task. Let represent the movement speed of user u in time slot t; considering a scenario where tasks arrive randomly and their statistical characteristics are unpredictable, when no task arrives in time slot t,

[0043] Assuming time slot t, a total of K cooperative sets can be formed. The k-th cooperative set can be represented as... Users can choose to unload tasks from collaborative collections and define... Represents the associated variable of the collaborative set, when This indicates that the user has chosen to collaborate in the set. Perform the task. This indicates that the user did not select a collaboration set. definition This represents the decision variable for task unloading. This indicates that the task will be offloaded to the edge for execution. This indicates that the task will be executed locally.

[0044] S2: Based on the network system model in the MEC scenario, construct a collaborative communication model, a task computing model, a user mobility model, and a load balancing model.

[0045] Collaborative communication model:

[0046] In this system, users within the same cell use Orthogonal Frequency Division Multiple Access (OFDMA). Interference between users within the same cell is ignored, but potential interference between users in different cells is considered. Assume the user equipment transmit power is fixed at P. u User u in time slot t receives data from the cooperative set. The signal-to-interference-plus-noise ratio (SIR) of a signal can be expressed as:

[0047]

[0048] Among them, P j D represents the power of the interfering user's transmission. u,m,t This represents the distance between user u and the service access point m, where α is the path loss constant. It is the fast fading coefficient of the channel allocated to user u by time slot node m in time slot t, and it follows a standard Gaussian distribution. Including collaborative sets Other cooperative sets besides N0, where N0 represents noise power.

[0049] t-slot user u will compute tasks Send to collaboration collection wireless transmission rate for:

[0050]

[0051] In the formula, B is the user's channel bandwidth.

[0052] The transmission time of a task is related to the channel transmission rate and the size of the task data, and the time slot user u-cooperative set is t. Communication task transmission delay It can be represented as:

[0053]

[0054] When a user sends a task to a node, it consumes its own energy. Due to the limited energy of user devices, this energy consumption cannot be ignored. At this time, the task... Transmission energy consumption It can be represented as:

[0055]

[0056] Task computation model:

[0057] When the task is executed locally, the local computation latency is... It can be represented as:

[0058]

[0059] in, The computing power of user equipment u.

[0060] User equipment power consumption only includes transmission power consumption and local computing power consumption, task power consumption. Energy consumption calculated locally It can be represented as:

[0061]

[0062] Here, v0 is the effective energy cost coefficient related to the chip architecture.

[0063] Each base station is equipped with a CPU with a maximum frequency of F. m A server provides computing services to users. When multiple users simultaneously offload their computations to a single server, the server provides services to multiple users concurrently through processor sharing. For simplicity, it is assumed that the server allocates the same amount of computing resources to each user; the computing resources allocated by server m to each user can be represented as follows: Because the time slot length is relatively short, assuming It remains unchanged within each time slot, but varies between different time slots. Things may change. Indicates in the collaborative set The number of servers in the collaboration set, and the average computing performance of the entire collaboration set. It can be represented as:

[0064]

[0065] When the task is unloaded to the collaboration set During computation, servers in the collaborative set share computing resources to provide services to users, and task computation latency... It can be represented as:

[0066]

[0067] The entire edge execution phase includes three parts: the user sends the task to the collaboration set, the collaboration set completes the task computation, and the computation result is sent back to the user. Since the task output is often small and the downlink transmission rate is relatively fast, the latency of the third part can be ignored. Considering the task transmission latency and the execution latency on the collaboration set, the total latency of task execution at the edge is... It can be represented as:

[0068]

[0069] The total latency of user task execution is expressed as:

[0070]

[0071] User mobility model:

[0072] To measure the speed of user mobility within the system, this invention uses user dwell time as a mobility evaluation index during cell handover. The dwell time distribution follows an exponential distribution, which can be specifically expressed as:

[0073]

[0074] In the formula, This represents the average dwell time, used to measure mobility intensity. The smaller the value, the stronger the user's mobility. The specific value can be obtained by analyzing historical data using machine learning methods.

[0075] If a user stays in a certain set for a certain amount of time Less than the task transmission time The task transmission will then fail. The probability of task transmission failure, predicted based on the dwell time, can be expressed as:

[0076]

[0077] The probability of successful task transmission is then:

[0078] Load balancing model:

[0079] This paper employs an LSTM network to predict the computational and storage resource usage of a computing server. LSTM is a time-series-based prediction algorithm. Because node load changes are periodic, LSTM can be used to predict future node load changes based on historical load information. The historical CPU and storage usage information of nodes in the network is input into the LSTM network, and the predicted CPU and storage resource usage values ​​are output.

[0080] To measure the server's load status, considering the utilization of server computing and storage resources, we define the expected load of server m. for:

[0081]

[0082] in, and These represent the predicted computational resource usage and storage resource usage of server m in time slot t, respectively. This represents the maximum computing power of server m. ω1 represents the maximum storage capacity of server m; ω2 represents the weight of computing resources and storage resources. The values ​​of computing resources and storage resources are determined according to their importance. Preferably, computing resources and storage resources can be equally important, so ω1 = ω2 = 0.5 can be taken.

[0083] Set the expected load threshold ΓL m The set of high-load nodes is defined as follows: Expected load value When a node is considered high-load, it needs to distribute excess computing tasks to other neighboring nodes and will not accept computing tasks from other nodes. A low-load node is defined as... Expected load value At that time, the node is considered a low-load node, which can accept the computing load of other nodes and does not need to unload computing tasks to other nodes.

[0084] To measure the load skew among nodes in a collaborative set and encourage high-load and low-load nodes to share computing resources, a collaborative set is defined. Expected load deviation for:

[0085]

[0086] Where n represents the set of collaborations The number of servers in the middle, Represents a collaborative set The average load across all nodes. When nodes form a collaborative network, the smaller the expected load deviation of the collaborative set, the better, to avoid uneven load distribution.

[0087] S3: Construct a joint optimization problem of offloading decision and collaborative set association decision based on the collaborative communication model, task computation model, user mobility model and load balancing model.

[0088] This invention considers the impact of mobility on task success. Since the impact of user mobility on task success primarily lies in the task transmission process, this invention addresses the user latency cost of successful task execution for user u in time slot t. The model is as follows:

[0089]

[0090] Because users move at high speeds within the network, offloading and collaboration set association decisions need to be made in real time. This invention uses minimizing task completion latency cost as the objective function, and the joint optimization problem of offloading and collaboration set association decisions can be modeled as follows:

[0091]

[0092]

[0093]

[0094]

[0095]

[0096]

[0097] in, Let T represent the set of time slots, and T represent the system time period. Let p0 represent the set of collaborations, and p0 represent the preset signal-to-interference-plus-noise ratio (SIR) threshold. Indicates the power consumption of user equipment. This represents the user equipment energy consumption budget; constraint C1 ensures that the total task execution delay does not exceed the maximum tolerable delay; C2 and C3 respectively ensure the values ​​of the offloading decision variable and the collaboration set association variable; C4 indicates that the signal-to-interference-plus-noise ratio between the user and the collaboration set should be greater than the preset minimum threshold p0 to ensure that the task can be transmitted successfully; and C5 is the user equipment energy consumption constraint.

[0098] S4: A two-time-scale decision model is used to solve the joint optimization problem of unloading decision and collaboration set association decision, and the optimal unloading decision and collaboration set association decision are obtained; the system performs task migration based on the optimal unloading decision and collaboration set association decision.

[0099] Considering that in high-speed mobile scenarios, users continuously generate computational tasks during their movement, requiring real-time offloading and migration decisions, a fast timescale decision-making approach is adopted. For example, in road safety scenarios, task generation timescales are often on the order of hundreds of milliseconds. Therefore, each time slot selects a suitable set of collaborators for offloading based on its own location. However, given the sampling timescale of real-world datasets, node load often changes significantly on the order of seconds or minutes, which is slower than the user task generation speed. Furthermore, this invention takes into account the changing trends of node load during collaboration; therefore, node collaboration decisions are considered to be made on a slow timescale.

[0100] like Figure 3 As shown, this invention proposes a two-timescale decision model to solve the joint optimization problem of unloading decision and cooperative set association decision; specifically:

[0101] Let T consecutive time slots be defined as a time scale, and let T be the set of time slots at the k-th time scale. k = {kT, kT+1, ..., (k+1)T*1}. Two time scales, fast and slow, are defined, and decisions are made on each time scale respectively, as detailed below:

[0102] 1) Slow time scale: At the beginning of each time scale, the load information is updated according to the load prediction algorithm, and a new cooperative set is formed by using a cooperative mechanism based on FCM and alliance game.

[0103] 2) Fast time scale: In each time slot, based on user task information and collaboration set, the best unloading decision and collaboration set association decision are found for the user using an Actor-Critic-based migration decision algorithm.

[0104] The process of updating the cooperation set using a cooperation mechanism based on the FCM algorithm and coalition game theory includes:

[0105] If only the current load of nodes is considered when forming a cooperation set, a "ping-pong effect" may occur, leading to instability in the resulting cooperation set. The load state of nodes will change over time, requiring continuous development of a game theory model to solve the problem. Therefore, a stable cooperation set is formed by comprehensively considering the long-term load changes of nodes. The future load changes of nodes are predicted using a load prediction algorithm, and then a stable cooperation set is formed based on the proposed cooperation algorithm.

[0106] To expand the service scope of collaborative sets, in addition to considering future node load changes, the relative distance between nodes also needs to be considered when forming collaborative sets. However, traditional distance-based data clustering algorithms, such as K-means, often produce relatively fixed clusters when calculating centroids, failing to form effective dynamic collaboration. This invention employs Fuzzy C-Means (FCM) clustering to pre-segment nodes, and then combines this with expected node load states to form stable collaborative sets through federated game theory; specifically:

[0107] The FCM algorithm, as a data clustering algorithm, assigns a weight to each data point and cluster to indicate the degree to which the current object belongs to that cluster. It obtains cluster centers by minimizing the clustering loss function. FCM divides several data points into multiple fuzzy clusters. In this invention, x... i This represents the node position. The clustering loss function is expressed as:

[0108]

[0109] Where K is the number of collaboration sets, n represents the number of collaboration points in the collaboration sets, and c j μ is the center of the cooperative set. ij τ represents the membership degree; τ is the membership degree index, used to control the fuzziness of the partitioned cooperative set.

[0110] To minimize the clustering loss function, the collaboration center c j The update rules are as follows:

[0111]

[0112] Membership degree μ ij The update rules are as follows:

[0113]

[0114] Collaboration Center c j and membership degree μ ij The above two formulas are used to iteratively update the algorithm until the FCM algorithm converges, resulting in K sets of collaborations.

[0115] The FCM algorithm only considers the positional relationships between nodes and does not consider their future load states. Based on the pre-segmentation of the FCM algorithm, the cooperation set is further adjusted according to the expected load of the nodes. Coalition game theory focuses on which coalitions can be formed and the collective and individual benefits generated during coalition formation. For the cooperation set adjustment problem in this invention, it can be modeled as a coalition game, specifically:

[0116] This invention defines a coalition game. The set of participants, i.e., all nodes in the network, is represented by Q, which represents the utility function value that the alliance can obtain. The goal of coalition game is to maximize the payoff value within each cooperative set, which represents the load fairness of the cooperative set. Each node, as a participant in the game, decides whether to join or leave a cooperative set. The alliance formed by multiple nodes should aim to maximize the alliance's utility function. The expected load deviation value of the cooperative set is taken. Standard deviation The reciprocal of the function is used as the utility function of the alliance. Represented as:

[0117]

[0118] Since nodes may form multiple different alliances, to describe which alliance a node is more inclined to choose, we define the concept of a stable alliance. A stable alliance is formed when no node is willing to leave its current alliance to form a new one. In this case, all nodes in the alliance can obtain higher utility than if they left their current alliance. To compare the strengths and weaknesses of the formed alliances, we introduce the concept of Pareto advantage: for two sets... and They are all alliances composed of the same subset of nodes if and only if For certain nodes, when the condition is strictly true, there exists Pareto is superior Represented as Pareto advantage indicates that compared to set Some nodes are more inclined to join. To obtain a greater utility value; specifically:

[0119] Repeatedly perform the merge and split process on K collaboration sets until the final collaboration set no longer changes;

[0120] Merging: Compare the Pareto advantage of the merged set with that of the multiple cooperative sets. If the Pareto advantage of the merged set is greater than that of the multiple cooperative sets, the merged set is taken as the final cooperative set; otherwise, the multiple cooperative sets are retained. The merged set is the union of the multiple cooperative sets.

[0121] Splitting: When the Pareto advantage of the split subset is greater than that of the original set, the original set is split to obtain a higher utility value, and the split subset is used as a new cooperative set; otherwise, the original cooperative set is retained.

[0122] The two actions of merging and splitting can be represented as follows:

[0123] Merge: When Pareto priority over subsets When merging subsets

[0124] Splitting: When a subset Pareto is superior At that time, split For a subset

[0125] Where l∈[1,K].

[0126] A group of nodes can operate and form a larger coalition if and only if such a combination increases the utility of at least one node without decreasing the utility of any other participant. Therefore, the merger decision ensures that the interests of all relevant nodes are not harmed. Similarly, a coalition can decide to split, and if the benefits of splitting are greater, it can split into smaller coalitions; stable cooperative sets can be formed in this way.

[0127] Optimization problem This is a mixed-integer nonlinear programming problem, difficult to solve directly. Considering the time-varying nature of channel state and location during user movement, this invention proposes an Actor-Critic-based migration decision algorithm. The process of solving the optimal offloading decision and cooperative set association decision using the Actor-Critic-based migration decision algorithm includes:

[0128] Without loss of generality, assuming that changes in the wireless environment and available computing resources exhibit Markov properties, the joint optimization problem of offloading decision and cooperative set association decision is abstracted as a Markov decision process, with the base station acting as the agent. Each agent has an Actor network and a Critic network; the Actor network and the Critic network consist of two neural networks with identical structures. The state space, action space, and reward function are constructed, specifically:

[0129] State space:

[0130] The state space is the environmental information observed by the agent regarding the users and nodes around them, including collaborative ensemble computing power, channel state, and user tasks, and is defined as follows:

[0131]

[0132] in, H(t) represents the average computational performance of all cooperative sets in the network, and H(t) represents the channel gain matrix. This represents a set of user tasks.

[0133] Action space:

[0134] Action space reflects the set of actions an agent can take, including unloading decisions and server association strategies, and is defined as follows:

[0135]

[0136] Reward Function: In each time slot, the controller responsible for offloading decisions and server association strategies receives a corresponding reward based on the action state. The optimal decision is found by maximizing the reward function value. To minimize the average task execution latency, this invention optimizes the problem... The reciprocal of the objective function is set as the reward value, expressed as:

[0137]

[0138] The Actor network generates corresponding actions based on the current local observation state of a single agent and updates the reward function according to the actions before entering the next state. The Critic network takes the actions output by the Actor network as input and adjusts the Actor network's actions based on the policy gradient of the output. Experience information is generated based on the current state, the next state, the actions, and the reward value. Multiple pieces of experience information are sampled to train the Actor and Critic networks. By alternating the execution of the Actor and Critic networks, a fully trained Actor network will generate the optimal offloading decision and the cooperative set association decision. The system performs task transfer based on the optimal offloading decision and the cooperative set association decision.

[0139] Evaluation of the present invention:

[0140] The load prediction algorithm is trained using CPU and storage usage data from servers in the Google Cluster Load Dataset, with the first 80% of the dataset used for training and the last 20% for testing. When using the FCM algorithm for node pre-clustering, the EUA dataset, which includes the geographical locations of all base stations (BS) in the Melbourne CBD area of ​​Australia, is used as the edge server location dataset. Figure 4 This demonstrates the variation in average task completion latency for users under different node distribution densities. Node distribution density is expressed as the ratio of the number of nodes to the number of users within a region. Figure 4As can be seen, with the increase of node distribution density, more nodes can provide services to users, and the average task completion latency gradually decreases. Moreover, compared with the recent CoMP cooperation scheme (SUN W, LIU J. 2-to-M coordinated multipoint-based uplink transmission in ultra-dense cellular networks[J]. IEEE Transactions on Wireless Communications, 2018, 17(12): 8342-8356.) and the CoMP-free scheme, the scheme proposed in this invention has a lower average task completion latency. This is because the scheme proposed in this invention takes into account the expected load changes of nodes when forming the cooperation set, which can better guarantee the computational performance of the cooperation set.

[0141] The above-described embodiments further illustrate the purpose, technical solution, and advantages of the present invention. It should be understood that the above-described embodiments are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made to the present invention within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for task migration based on coordinated multi-point in mobile edge computing, characterized in that, The method comprises the following steps: S1: constructing a network system model under a MEC scenario; S2: constructing a cooperative communication model, a task computing model, a user mobility model and a load balancing model based on the network system model under the MEC scenario; the load balancing model comprises: adopting an LSTM network to calculate a predicted computing resource occupation and a predicted storage resource occupation of a server, calculating an expected load of the server according to the predicted computing resource occupation and the predicted storage resource occupation, setting an expected load threshold, and regarding the server as a high-load node when the expected load of the server is greater than the expected load threshold, or regarding the server as a low-load node otherwise; the high-load node is used for dispersing redundant computing tasks to other adjacent nodes and not accepting computing tasks of other nodes, and the low-load node is used for receiving computing loads of other nodes and not unloading computing tasks to other nodes; the high-load node and the low-load node share computing resources and form a cooperative set, and an expected load deviation of the cooperative set is defined; S3: constructing a joint optimization problem of an offloading decision and a cooperative set association decision according to the cooperative communication model, the task computing model, the user mobility model and the load balancing model; the joint optimization problem of the offloading decision and the cooperative set association decision is represented as: in, express Time slot users User latency cost for successful task execution Represents a set of time slots. Indicates the system time period. Represents a set of users. Represents a collaborative set. express Time slot users The maximum tolerable latency for the task. express Time slot users Task unloading decision variables, express Time slot users Collaborative set related variables, express Time slot users Receive from collaborative collection Signal-to-interference-plus-noise ratio (SIR / N). This represents the signal-to-interference-plus-noise ratio (SINR) threshold. express Time-slot user equipment power consumption This represents the user equipment energy consumption budget; S4: solving the joint optimization problem of the offloading decision and the cooperative set association decision by adopting a two-time-scale decision model to obtain an optimal offloading decision and an optimal cooperative set association decision; the system performs task migration according to the optimal offloading decision and the optimal cooperative set association decision; the process of solving the joint optimization problem of the offloading decision and the cooperative set association decision comprises: the two-time-scale decision model comprises a slow-time-scale model and a fast-time-scale model; the slow-time-scale model updates the cooperative set by adopting a cooperative mechanism based on an FCM algorithm and a coalition game according to load information, and the fast-time-scale model solves the optimal offloading decision and the optimal cooperative set association decision by adopting an Actor-Critic-based migration decision algorithm according to the cooperative set and user task information; the process of updating the cooperative set by adopting the cooperative mechanism based on the FCM algorithm and the coalition game comprises: adopting the FCM algorithm to cluster all server nodes to obtain K cooperative sets; repeatedly performing a merging process and a splitting process on the K cooperative sets until the final cooperative set no longer changes; the merging process: setting a coalition utility function according to the expected load deviation of the cooperative set, comparing the Pareto advantage of a merged set and the Pareto advantage of multiple cooperative sets according to the coalition utility function, regarding the merged set as a new cooperative set when the Pareto advantage of the merged set is greater than the Pareto advantage of the multiple cooperative sets, or otherwise, retaining the multiple cooperative sets; wherein the merged set is the union set of the multiple cooperative sets; the splitting process: comparing the Pareto advantage of a split sub-set and the Pareto advantage of an original cooperative set, regarding the split sub-set as a new cooperative set when the Pareto advantage of the split sub-set is greater than the Pareto advantage of the original cooperative set, or otherwise, retaining the original cooperative set; wherein the split sub-set is the subset of the original cooperative set. 2.The method of claim 1, wherein, The network system model in the MEC scenario specifically includes: U base stations and M users, with each base station equipped with an MEC server; the set of base stations is represented as follows. The user set is represented as The user's task is represented as ,in, Indicates the size of the task data. This indicates the number of CPU cycles required per bit of task. This indicates the maximum tolerable latency for the task. express Time slot users Movement speed; user-selected collaboration set Unload task, define Represents the associated variable of the collaborative set. This indicates that the user has chosen to collaborate in the set. Perform the task. This indicates that the user did not select a collaboration set. ;definition This represents the decision variable for task unloading. This indicates that the task will be offloaded to the edge for execution. This indicates that the task will be executed locally. 3.The method of claim 1, wherein, The user mobility model includes: using user dwell time as an evaluation indicator for whether a task can be successfully transmitted during cell handover; if the user's dwell time in the cooperative set... Less than the task transmission time The task transmission failed; the probability of task transmission failure is predicted based on the dwell time. 4.The method of claim 1, wherein, The process of solving the optimal offloading decision and the cooperative set association decision by using the transfer decision algorithm based on Actor-Critic includes: abstracting the joint optimization problem of the offloading decision and the cooperative set association decision into a Markov decision process, taking the base station as an intelligent agent, and constructing a corresponding state space, action space and reward function; each intelligent agent has an Actor network and a Critic network; the Actor network and the Critic network are both composed of two neural networks with the same structure; the Actor network generates a corresponding action according to the current local observation state of a single intelligent agent and updates the reward function according to the action to enter the next state; the Critic network takes the action output by the Actor network as input and adjusts the action of the Actor network through the output policy gradient; experience information is generated according to the current state, the next state, the action and the reward value; a plurality of experience information is sampled to train the Actor network and the Critic network, update the network parameters, and obtain the trained Actor network and Critic network; and the optimal offloading decision and the cooperative set association decision are obtained according to the training result of the Actor network.

5. The method of claim 4, wherein, The reward function is: ； wherein, denotes a reward value obtained by a slot user performing an action, denotes a slot user a user latency cost for successful execution of a task by a slot user, denotes a set of users.

Images