An edge computing-oriented task offloading and resource allocation method

Abstract

This invention discloses a task offloading and resource allocation method for edge computing. The method comprises three aspects: relay selection, offloading decision, and resource allocation, and is used to address the task latency and energy consumption issues of mobile users in the Internet of Things (IoT) environment. First, a virtual backbone network architecture is introduced, providing efficient multi-hop offloading services to users through a connectivity control set. Then, a deep reinforcement learning algorithm based on the connectivity control set is proposed to find the shortest path from the mobile user to the mobile edge computing server. Furthermore, based on the high coupling relationship between offloading decision and resource allocation, this invention designs a DLIO algorithm to solve the joint optimization problem of offloading and resource allocation, with the goal of minimizing system cost. This invention can effectively solve the problem of task offloading and resource allocation for mobile users in ubiquitous edge IoT systems.

Description

Technical Field

[0001] This invention relates to a method for task offloading and resource allocation for edge computing, belonging to the fields of Internet of Things and artificial intelligence technology. Background Technology

[0002] With the widespread adoption of general-purpose smart terminals and ubiquitous network technologies, a series of new communication concepts, such as the Internet of Things and edge computing, have been proposed and are gradually being realized in the 5G and the upcoming 6G era. Various new communication services are emerging in people's daily lives, including virtual reality, natural language processing, and autonomous driving. However, the contradiction between limited user computing resources and the need for real-time, rapid processing of the large amounts of data generated by applications is an unavoidable and pressing problem when faced with numerous computationally intensive and latency-sensitive tasks.

[0003] Mobile edge computing (MEC), as an innovative network computing paradigm, is widely considered a promising technology. It can address the dilemma between unprecedented growth in computing demands and the need for high-quality-of-service (QoS) computing. In typical MEC systems, servers with abundant computing and storage resources are deployed at the network edge, bringing them closer to mobile users and enabling them to provide timely computing services. Specifically, mobile users with limited computing resources can offload computing tasks to nearby MEC servers via wireless communication, reducing task latency and user energy consumption, thereby improving QoS. However, even when mobile users face computationally intensive tasks with strict latency constraints, the computing resources of MEC servers are not unlimited, and they may not be able to serve all mobile users. Furthermore, excessive offloading in an MEC network can lead to communication interference and computational congestion between servers and users, reducing system benefits. The offloading strategy of mobile users determines whether they perform computation offloading and the specific location of the offloading, which is closely related to the user's energy consumption and task latency. Meanwhile, the energy efficiency of the computation offloading strategy largely depends on the wireless resource allocation scheme, which is affected by interference management and power control schemes. Therefore, it is unwise to consider only a single performance metric of the system. A joint optimization strategy is needed for offloading and resource allocation in multi-user ubiquitous MEC IoT systems.

[0004] The joint optimization problem is a mixed-integer nonlinear programming (MINP) problem, which is NP-hard and difficult to solve directly using existing mathematical methods. Traditional numerical methods mainly include dynamic programming and branch calibration, but the computational complexity of these two methods increases exponentially with the scale of ubiquitous MEC (Multi-access Edge Computing) systems. Although much research has focused on the joint optimization problem in ubiquitous MEC environments, there is currently no widely applicable solution, nor is there a multi-hop forwarding scheme designed specifically for ubiquitous MEC systems. Summary of the Invention

[0005] Purpose of the invention: In view of the problems and shortcomings of the existing technology, the present invention proposes a task offloading and resource allocation method for edge computing, which solves the problem of intelligent computing offloading and resource allocation of mobile devices in ubiquitous edge Internet of Things.

[0006] Technical solution: The present invention provides a method for task offloading and resource allocation for edge computing, comprising the following steps:

[0007] Step 1: Introduce a virtual backbone network architecture and build a connectivity control set. When a mobile user has a computing task that needs to be processed but does not have sufficient computing power, it can choose to offload all or part of the task to the MEC server for processing via CDS;

[0008] Step 2: Use a deep reinforcement learning algorithm based on connected control sets to find the shortest unloading path;

[0009] Step 3: Solve the joint optimization problem of unloading and resource allocation using the DLIO algorithm.

[0010] Step 4: Process the task according to the overall optimal intelligent computing scheme.

[0011] Step 1 involves introducing a virtual backbone network architecture and constructing a connectivity control set, as detailed below:

[0012] (1) In the ubiquitous edge Internet of Things system, the location of each mobile user is randomly deployed. For ease of analysis, this invention assumes that a connected graph can be formed between the mobile edge computing (MEC) server and the mobile user. That is, all mobile users can directly or indirectly offload tasks to the MEC server. A virtual backbone network consists of all core users and MEC servers, where all core users are regarded as nodes with radius r′ and MEC servers are regarded as nodes with radius R′.

[0013] (2) Construct a connectivity control set (CDS) in the virtual backbone network to ensure that all core users can forward tasks to the MEC server through the nodes in the CDS. The construction of the CDS is mainly divided into three stages. First, a control set is found using the greedy coloring method. Then, a CDS is formed by constructing a Steiner tree. Finally, each leaf node in the Steiner tree is judged in turn. If the remaining CDS still has connectivity after deleting the leaf node, then the node is deleted from the current CDS. If deleting the leaf node affects the connectivity of the remaining CDS, then the leaf node is retained.

[0014] Step 2: Use a deep reinforcement learning algorithm based on connected control sets to find the shortest unloading path, as follows:

[0015] (1) Deep reinforcement learning is used in the relay node selection phase to search for the shortest forwarding path from each user to the server in the CDS network. The process of users searching for the shortest path is described as a Markov decision process (MDP).

[0016] (2) In the interaction with the CDS network environment, the next node with the lowest transmission cost is selected according to the reward function, and finally the shortest path from the user to the server is obtained.

[0017] Step 3: Solve the joint optimization problem of unloading and resource allocation using the DLIO algorithm, as detailed below.

[0018] (1) The solution space of the first layer combinatorial optimization depends on the number of users in the CDS network. As the number of users increases, the number of decisions also increases exponentially.

[0019] (2) The second layer of optimization of the DLIO algorithm is the optimization of the objective function, which aims to optimize the total delay and energy consumption by using an improved PSO algorithm to optimize the resource allocation strategy.

[0020] Step 4: Process the task according to the overall optimal intelligent computing scheme, that is, the task offloading scheme that minimizes the total system overhead by rationally allocating limited computing resources. Based on the virtual backbone network constructed in Step 1, construct CDS, apply Step 2 to find the shortest offloading path for mobile users, and then use Step 3 to solve the joint optimization problem of offloading and resource allocation. After obtaining the optimal intelligent computing scheme in Step 4, the task is completed.

[0021] Compared with the prior art, the present invention has the following advantages:

[0022] (1) This invention studies a multi-hop offload management scheme in a multi-user MEC environment, which solves the problem that mobile users cannot directly offload tasks to the MEC server under certain special circumstances.

[0023] (2) This invention is the first to apply a CDS-based relay forwarding mechanism to the task offloading process in a ubiquitous MEC Internet of Things system. This solution can better manage core relay users and reduce unnecessary energy consumption.

[0024] (3) This invention proposes an intelligent three-layer joint optimization strategy for the offloading and resource allocation problem of ubiquitous MEC Internet of Things system. By limiting CPU, bandwidth and power, the joint optimization strategy for computation offloading and resource allocation is formulated as a mixed integer nonlinear programming (MINP) problem.

[0025] (4) Based on the high coupling between unloading decision and resource allocation, this invention proposes a two-layer intelligent optimization algorithm (DLIO), which combines the ant colony optimization algorithm (ACS) and the particle swarm optimization algorithm (PSO) to obtain the optimal solution to the above MINP problem. Attached Figure Description

[0026] Figure 1 This is a structural example diagram of mobile user task offloading in a mobile edge IoT system. Detailed Implementation

[0027] To enhance understanding of the present invention, the invention will be further explained below with reference to specific embodiments.

[0028] Example 1: See Figure 1 A method for task offloading and resource allocation for edge computing includes the following steps:

[0029] Step 1: Introduce the virtual backbone network architecture and construct the connectivity control set. This example uses a pseudo-unit disk graph (PUDG) G=(V,E) to represent the virtual backbone network, where V={v0,v1,v2,…,v…} i ,…,v c} represents all nodes in the graph, v0 represents the MEC server, and the transmission range is R′. i Representing core users E is the set of edges between nodes in G. Mobile users are randomly distributed around the MEC server, and the set of mobile users is U = {U1, U2, ..., U...}. k ,…,U N} represents the transmission range of each mobile user as r′. Mobile users are divided into core users and edge users based on whether they are within the transmission range of the MEC server. If a mobile user is within the transmission range of the MEC server, they are called a core user; otherwise, they are called an edge user. A core user is represented as... Edge users are represented as U C ∪UE =U. Each edge user has at least one neighbor that is a core user, to ensure that edge users can forward tasks to the MEC server through their neighboring core users.

[0030] Unlike traditional UDG, PUDG is characterized by having only one special node representing the MEC server. Its transmission range is significantly larger than that of other nodes, whose transmission range is r′. The disk centered on the node represents the transmission range of the user (MEC server). When v a and v b The Euclidean distance between them satisfies D(v a ,v b When )≤r′,a,b∈[1,c],v a and v b They can send messages to each other, i.e., v a and v b They are adjacent and have bidirectional edges. When D(v) i When v0)≤r′, we can say that from v i There is a one-way edge to v0, and v0 is v i The neighbor, which means v i Messages can be sent to v0. Due to the specific nature of the system model, messages can be sent from node v0 to any node v. i The distances all satisfy D(v0,v) i )≤R′, so from v0 to v i There is always a one-way edge. Therefore, it can be ensured that the computation results from the MEC server can be sent to any core user quickly, with negligible latency.

[0031] In this example, each mobile user, where U k For each ∈U, there is a computationally intensive and latency-sensitive task. To complete, M k Indicates the size of the input data, C k This represents the total number of CPU cycles required to complete the computational task. It's a task delay constraint. When a mobile user U... kWhen a mobile user has a computational task to process but lacks sufficient computing power, it can choose to offload all or part of the task to the MEC server via CDS. Therefore, during task offloading, transmission occurs only within the core user nodes of the CDS, not across all core user nodes in the network. It's important to note that although all mobile users in offload computing mode eventually offload their tasks to the MEC server, the transmission paths for core and edge users differ. Edge users first offload their computational tasks to core users, who then forward the tasks to the MEC server via CDS using the shortest route. Core users, on the other hand, directly forward their tasks to the MEC server via CDS using the shortest path.

[0032] In this example, the construction of the CDS is mainly divided into three stages. First, a control set is found using a greedy coloring method. Then, a Steiner tree is constructed to form a CDS. Finally, the size of the CDS is reduced by excluding nodes that do not affect connectivity without losing connectivity. The specific construction process of the CDS is as follows:

[0033] Phase 1: Constructing the Data Structure (DS). In this phase of the algorithm proposed in this example, a Mindset (MIS) is greedily constructed as the DS using a labeling method. First, this example finds the node v with the largest number of neighbors, excluding the MEC server node. i Dye it black and include it in DS, where D = D∪{v i Then, color all of that node's neighbors gray. Next, find the node with the most white neighbors among the remaining white nodes, color it black, and then color all its neighbors gray. Repeat this process until... Figure 1 The set of all black nodes, D, is formed until there are no more white nodes. Figure 1 The MIS is also a DS.

[0034] Phase 2: Constructing the Steiner Tree. In this phase, a Steiner tree with the fewest Steiner nodes needs to be found to connect the nodes in DS. First, the MEC server node is selected as the leader, C = C∪{v0}. Nodes in D are called terminals. A path from a node in D to the leader is checked based on the non-increasing number of neighbors, where these neighbors consist only of black nodes. If so, the non-increasing order of adjacent nodes is checked; otherwise, a path with the fewest gray nodes that cannot form a cycle is selected, the gray nodes in the path are colored black, and added to C. Finally, the set C∪D of all black nodes constitutes the set CDSS, also known as the Steiner tree.

[0035] Phase 3: Pruning the CDS. The goal of this example is to minimize the size of the CDS, thereby reducing network costs. The control node is redundant if the resulting CDS still connects and covers all nodes in the network by removing the control node. First, check all leaf nodes v in the Steiner tree. i To determine whether there exists a single leaf node v i The gray node under control, i.e., the leaf node v being inspected. i Check if all gray nodes in a leaf node have only one black neighbor. If no such gray node exists, move the leaf node v to a black neighbor. i Remove nodes from the CDS and color them blue; if gray nodes exist, continue checking until all leaf nodes have been checked. The final collection of all black nodes constitutes an approximate MCDS for the entire graph.

[0036] When an approximate MCDS is found, the CDSRL algorithm described above is used to search for the set of shortest paths from each mobile user to the MEC server, and the number of relay nodes m on the transmission path is calculated.

[0037] Step 2: Use a deep reinforcement learning algorithm based on connected control sets to find the shortest unloading path.

[0038] Sub-step 2-1: Establish a system model for task unloading and resource allocation.

[0039] There are two different task offloading schemes for different types of mobile users. For edge users... It first divides the computation task into two subtasks, computes one subtask, and then sends the other subtask to its neighboring core users via a single-hop relay channel. The core users then forward the subtask to the MEC server via CDS based on the shortest path. For core users... It first divides the task into two subtasks, then sends the subtasks that need to be forwarded to the MEC server via the shortest path in the CDS, while the remaining subtasks are computed locally. Let θ k ∈[0,1] represents task A k The unloading ratio, k∈[1,N]. According to θ k The value indicates the following three task uninstallation modes:

[0040] (1) If θ k =0, means MU U k Choose to perform the task entirely locally.

[0041] (2) If 0 < θ k <1, meaning θ k M kTasks of a certain size will be offloaded to the MEC server for computation, while the remaining (1-θ) tasks will be performed on the MEC server. k M k The bits will be calculated locally.

[0042] (3) If θ k =1, which means mobile user U k Choose to completely offload the task to the MEC server for computation.

[0043] In addition, this invention mainly considers the joint optimization of task offloading and resource allocation in multi-hop ubiquitous edge IoT systems, and mainly introduces resource allocation strategies from three aspects: power, bandwidth and CPU.

[0044] (1) Power: P = {P1, P2, ..., P} k ,…,P N}, where P k Let P represent the transmission power allocated to the k-th user. max It is the maximum transmission power of a MU.

[0045] (2) CPU: C = {C1, C2, ..., C} k ,…,C N}, where C k C represents the number of CPU cycles allocated to user k. max It is the maximum CPU cycle.

[0046] (3) Bandwidth: in Indicates relay user The bandwidth allocated to the l-th task, where i∈[1,c].

[0047] Power and CPU allocation is a real-time, dynamically adjustable resource allocation method to achieve the optimal solution. However, bandwidth allocation differs from the other two parameters; it is a policy-based allocation.

[0048] Marginal users The achievable data transmission rate is shown in formula (1):

[0049]

[0050] Where B represents user Offloading channel spectral bandwidth, P k Indicates transmission power. σ represents the channel gain. 2 Defined as noise power.

[0051] Core users The achievable data transmission rate is shown in formula (2):

[0052]

[0053] in, Indicated by user The spectral bandwidth of the offload channel allocated to subtask l, P k Indicates transmission power. σ represents the channel gain. 2 Defined as noise power.

[0054] It is important to note that edge users do not need to allocate bandwidth, while core users do need to allocate bandwidth when acting as relay nodes in order to improve task forwarding speed and thus reduce task forwarding time.

[0055] This paper presents two computing modes: a fully local computing mode and a hybrid offload computing mode.

[0056] (1) Fully local computation

[0057] In fully local computing mode, the user's computing tasks are handled directly by the user. The time and energy consumption required to execute local computing tasks can be expressed by formulas (3) and (4):

[0058]

[0059]

[0060] Where, γ k It is the effective capacitance coefficient, which depends on the mobile user U. k The chip architecture, M k Indicates the size of the input data, C k f represents the total number of CPU cycles required to complete the computational task. k It's U k CPU frequency.

[0061] (2) Hybrid unloading calculation

[0062] In the hybrid offload computing mode, users first need to divide the task into two parts. One part is computed locally, and the other part is offloaded to the MEC server for computation. Since the server's computing power is quite large, this paper does not consider the MEC server's energy consumption and computation time.

[0063] There are two types of mobile users in this article, therefore u k Set user metrics to specify user categories: u k =1 indicates mobile user U k The type is edge user, u k =0 indicates mobile user U k The type is core users.

[0064] U k The local computation time of the task can be calculated using formula (5):

[0065]

[0066] Where θ k M represents the task unloading ratio. k Indicates the size of the input data, C k f represents the total number of CPU cycles required to complete the computational task. k It's U k CPU frequency.

[0067] The total time for task unloading includes the unloading time for edge users and the forwarding time for core users, as shown in formula (6):

[0068]

[0069] Where m represents the number of users on the shortest path from the core user to the MEC server in CDS, θ k M represents the task unloading ratio. k Indicates the size of the input data, C k This indicates the total number of CPU cycles required to complete the computational task.

[0070] The energy consumption of the local computing task and the unloading task are shown in Equations (7) and (8), respectively:

[0071]

[0072]

[0073] Where γ k It is the effective capacitance coefficient, θ k M represents the task unloading ratio. k Indicates the size of the input data, C k f represents the total number of CPU cycles required to complete the computational task. k It's U k CPU frequency, P k Indicates transmission power.

[0074] In summary, according to formulas (5) and (6), the total time consumed in the hybrid unloading mode is shown in formula (9):

[0075]

[0076] According to formulas (7) and (8), the total energy consumption under the hybrid unloading mode is shown in formula (10):

[0077]

[0078] Considering that latency and energy consumption are both important indicators for measuring system performance, this invention uses their weighted sum as the total cost of completing a task.

[0079] The total cost of the system completing all tasks is shown in formula (11):

[0080]

[0081] Where β k ∈[0,1] are the weighting coefficients for delay and energy consumption when calculating system cost, T k This is the total time consumed in the hybrid unloading mode, E k This represents the total energy consumption under hybrid offloading mode.

[0082] Based on the transmission characteristics of MEC (Multi-access Edge Computing), higher energy consumption of mobile devices results in greater transmission power, higher data transmission rates, and lower latency, and vice versa. However, with increased energy consumption, the lifespan of mobile devices will be shortened. Therefore, a trade-off needs to be struck between device usage time (related to energy consumption) and service latency. This invention employs a multi-objective optimization method to model the trade-off between latency and energy consumption, aiming to minimize energy consumption and task processing latency across all nodes in the system. The optimization objective is to minimize the total cost of the system completing the task.

[0083] Sub-step 2-2: Use a deep reinforcement learning algorithm to find the shortest forwarding path.

[0084] In the relay node selection phase, this invention uses deep reinforcement learning to search for the shortest forwarding path from each edge user to the server in the CDS network. This invention describes the process of edge users searching for the shortest path as a Markov decision process. During interaction with the CDS network environment, the next node with the lowest transmission cost is selected based on the reward, ultimately obtaining the shortest path from the edge user to the server. The detailed MDP definition is as follows:

[0085] (1) State s: Each mobile user on the CDS network is represented as a state. The current network location of an edge user is its current state, and the next possible state of an edge user is any connected user in the CDS.

[0086] (2) Action a: The actions of edge users consist of directed edges from the current core user to the next core user in the CDS network. Unlike the traditional reinforcement learning state-action set, the user positions in the CDS network are randomized, so different users have different action sets.

[0087] (3) Reward R(s,a): The immediate reward obtained by a user for performing an operation in the current state. In this invention, the reward is set as the communication transmission cost between users in the CDS network.

[0088] (4) Discount factor γ: γ∈[0,1] is a trade-off between current and future reward realization. If γ=0, the edge user only cares about the transmission cost of the next user. If γ=1, the edge user cares about the transmission cost of all future users.

[0089] To search for the nearest transmission path to the server in the CDS network, the smart device initializes state s0 (i.e., the edge user node), and then selects the next CDS user node based on the reward, until the data reaches the server. Let Q(s,a) represent the state action value function, and its update method is shown in formula (12):

[0090]

[0091] Where α t Q is the learning rate in time slot t, γ is the discount factor, and Q is the learning rate in time slot t. t (s,a) represents the current state action value function in time slot t, Q t (s,a′) represents the state action value function of the next action a′ in time slot t, R t (s,a) represents the immediate reward a user receives for performing action a in the current state s.

[0092] In CDS, the design of the reward function is crucial for mobile users to find the transmission path with the lowest communication cost. Different reward function models lead to different learning strategies. To find the route with the lowest transmission cost in the CDS backbone network, this invention defines the reward function in the algorithm as the negative value of the weight (communication cost) of the CDS backbone network connection edge. Furthermore, due to the instability of the backbone network, directly using negative numbers as the reward for the mobile user's search process will lead to instability in the algorithm update. Therefore, an average reward strategy is introduced to obtain a stable reward value and accelerate convergence.

[0093] Step 3: Solve the joint optimization problem of unloading and resource allocation using the DLIO algorithm.

[0094] The optimization problem of this invention involves dual optimization: optimization of computational patterns and continuous optimization of computational resources. Optimization of computational patterns is essentially an NP-hard combinatorial optimization problem. Solving combinatorial optimization problems is extremely challenging. While deterministic algorithms (such as branch and bound algorithms) can obtain optimal solutions within a sufficiently large space, their time cost is unacceptable. Furthermore, offloading decisions are chaotic, thus necessitating the use of intelligent algorithms. Clearly, computational pattern decisions and resource allocation in the optimization problem are coupled. Regarding resource allocation strategies, different offloading decisions directly affect the allocation ratio of resources such as power and cache for mobile users. For offloading decisions (i.e., computational pattern decisions), the resource allocation ratio directly affects the estimation of mobile device performance. It can be said that only under the optimal offloading decision can the optimal resource allocation strategy be obtained.

[0095] To address the aforementioned coupled optimization problem, this invention proposes a two-level intelligent optimization algorithm, DLIO. The first-level optimization is a combinatorial optimization of computational patterns, aiming to minimize the latency and energy consumption of all mobile users. Since deterministic algorithms (such as branch and bound algorithms) struggle with high-complexity problems, this invention proposes a novel ant colony algorithm for offloading decision optimization.

[0096] Sub-step 3-1: Optimize the task unloading decision.

[0097] The solution space for the first-layer combinatorial optimization depends on the number of users in the CDS network. As the number of users increases, the number of decisions also grows exponentially. It is important to note that many tasks cannot be unloaded due to time constraints. Therefore, during the task initialization phase, tasks that cannot be unloaded are removed to reduce the computational complexity of unloading decisions. Next, this invention specifies the conditions under which tasks cannot be unloaded.

[0098] According to resource allocation rules, given limited server computing resources, if the CPU computing resources and bandwidth allocated to edge users are insufficient to complete the offloading computation, then edge users cannot select the offloading mode. Let... The delay constraint for user i is shown in equation (13); the conditions for edge users to offload tasks are as follows:

[0099]

[0100] in This is part of the time calculated locally. It is the CDS forwarding time, C i It represents the computing resources allocated to user i, where B represents the transmission bandwidth, and P represents the computing resources allocated to user i. i Indicates transmission power. σ represents the channel gain. 2Defined as noise power. In other words, edge users must complete computational tasks on time within the constraint of the maximum transmission rate.

[0101] When a mobile user selects a computing mode, the local pheromone is updated using formula (14):

[0102]

[0103] Where ω is the attenuation coefficient and v0 is the initial pheromone.

[0104] At the end of each iteration of the DLIO algorithm, based on the allocated resources... Update global variable v ij .

[0105] Sub-step 3-2: Continuous optimization of computing resources.

[0106] Specifically, a modified PSO algorithm is used to optimize resource allocation strategies, thereby optimizing total latency and energy consumption. The PSO is initialized with a set of random particles (i.e., random solutions in iteration n). The particle tracks two "extremes". Perform self-update. After completing the search... and Then, the particle will update its velocity and position using formula (15):

[0107]

[0108] Where v i (n) is the particle velocity relative to the nth iteration, x i (n) is the particle position. η,κ1,κ2,r1,r2 are usually subject to the following constraints: η∈[0,1.2],κ1∈[0,2],κ2∈[0,2],r1∈[0,1],r2∈[0,1].

[0109] ηv i The function of (n) is to make the particle move in the same direction as the original direction, where η is the acceleration factor. This can cause particles to tend to return to positions in the search space where they have very high fitness values. It can move particles to the positions that the swarm moved to in the previous iteration.

[0110] For nonconvex problems P, this invention proposes an improved PSO algorithm that integrates the Hessian matrix and Lagrange-Kuhn-Tucker conditions into the fitness function of PSO.

[0111] Step 4: Process tasks according to the overall optimal intelligent computing scheme. The optimization objective of this invention is to minimize system latency and energy consumption. Based on the virtual backbone network constructed in Step 1, a connectivity control set is built. Step 2 is used to find the shortest offload path for mobile users. Then, Step 3 is used to solve the joint optimization problem of offloading and resource allocation. Finally, based on the optimal task offloading and resource allocation scheme obtained in Steps 1, 2, and 3, each mobile user selects the optimal offloading scheme, rationally allocating limited computing resources to minimize the total system overhead.

[0112] Those skilled in the art will recognize that the embodiments described herein are intended to help readers understand the principles of the invention. It should be understood that the embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. After reading this invention, any modifications of the invention in various equivalent forms by those skilled in the art fall within the scope defined by the appended claims.

Claims

1. An edge computing oriented task offloading and resource allocation method, characterized in that, Comprising the following steps: Step 1: introduce a virtual backbone network architecture, build a connected control set; Step 2: find the shortest offloading path by using a deep reinforcement learning algorithm based on the connected control set; Step 3: solve the joint optimization problem of offloading and resource allocation by DLIO algorithm; Step 4: process tasks according to the overall optimal intelligent computing scheme; Wherein, step 1: introduce a virtual backbone network architecture, build a connected control set, as follows: (1) In the ubiquitous edge Internet of Things system, the location of each mobile user is randomly deployed, assuming that a connected graph is formed between the mobile edge computing (MEC) server and the mobile user, that is, all mobile users directly or indirectly offload tasks to the MEC server, and a virtual backbone network is composed of all core users and MEC servers, wherein all core users are regarded as nodes with a radius of , and the MEC server is regarded as a node with a radius of ; (2) A connected control set CDS is constructed in the virtual backbone network to ensure that all core users forward tasks to the MEC server through the nodes in the CDS, and the construction of CDS mainly includes three stages: first, find a control set using greedy coloring method, then form a CDS by constructing a Steiner tree, finally, judge each leaf node in the Steiner tree in turn, if the remaining CDS still has connectivity after deleting the leaf node, delete the node from the current CDS; if deleting the leaf node affects the connectivity of the remaining CDS, keep the leaf node; Step 2: find the shortest offloading path by using a deep reinforcement learning algorithm based on the connected control set, as follows: (1) In the relay node selection stage, deep reinforcement learning is used to search for the shortest forwarding path from each user to the server in the CDS network, and the process of searching for the shortest path by the user is described as a Markov decision process MDP; (2) In the interaction with the CDS network environment, the next node with the lowest transmission cost is selected according to the reward function, and the shortest path from the user to the server is finally obtained; Step 3: solve the joint optimization problem of offloading and resource allocation by DLIO algorithm, as follows, (1) The solution space of the first layer of combined optimization depends on the number of users in the CDS network, and as the number of users increases, the number of decisions also increases exponentially; (2) The second layer optimization of DLIO algorithm is the optimization of objective function, the goal is to optimize the total delay and energy consumption, and an improved particle swarm optimization algorithm PSO is used to optimize the resource allocation strategy. 2.The method of claim 1, wherein, Step 4: process tasks according to the overall optimal intelligent computing scheme, that is, by reasonably allocating limited computing resources, the total cost of the system can be minimized in the task offloading scheme, according to step 1 to build a virtual backbone network and build a CDS, apply step 2 to find the shortest offloading path of mobile users, and then solve the joint optimization problem of offloading and resource allocation by step 3, after obtaining the optimal intelligent computing scheme in step 4, complete task 2.

Images