Link topology adaptive unloading method for edge calculation of Internet of Vehicles

Abstract

The invention discloses a link topology self-adaptive unloading method for edge calculation of the Internet of Vehicles. Establishing a dynamic V2I and V2V combined system, and expressing a delay minimization problem of sequential subtasks in the combined system as a path optimization problem; defining a state space of a Markov decision process, a state of each time step, an action space and an award, and modeling a path optimization problem into the Markov decision process; in the Markov decision process, hidden information in the state of time step length is adaptively learned according to different link topologies by using the GCN, and the hidden information is provided for the DDQN to select an action for obtaining the highest reward so as to realize the minimum total time delay. According to the method, a fine-grained partial unloading model is developed for sequential subtasks in a dynamic V2I and V2V combined system with topological links so as to optimize a collaborative unloading strategy, and based on a GCN and DDQN self-adaptive unloading method, different link topologies are adapted, and the total time delay is minimized.

Description

Technical Field

[0001] The present invention belongs to the technical field of reinforcement learning and graph machine learning, and specifically is a link topology adaptive unloading method for edge computing of Internet of Vehicles. Background Art

[0002] In recent years, with the development of communication technology, some innovative modes have emerged, such as vehicle-to-vehicle (V2V) communication, which establishes direct and decentralized links between nearby vehicles, and vehicle-to-infrastructure (V2I) communication, through which vehicles connect with infrastructure elements and exchange information. These modes enable vehicles to effectively exchange real-time information, laying the foundation for many applications such as collision avoidance, cooperative driving, and traffic management.

[0003] In high-demand scenarios such as traffic congestion or emergency situations, vehicles can pool computing power and coordinate work to handle the increased load, and nearby vehicles and infrastructure can share computing resources. Cluster-based algorithms treat buses as cluster heads, overcome resource limitations, and meet the Quality of Service (QoS) requirements of emergency information. Unsignaled intersection collision warning systems based on V2V and V2I can improve driver safety. Some algorithms model the connectivity of vehicles with roadside units (RSUs) and vehicles, and estimate the connectivity and unloading capacity between vehicles through traffic classification, overload control, and admission control. Then some algorithms optimize the transmission mode and power level in V2V and V2I environments to maximize the total capacity of the V2I link and ensure QoS requirements. It is obvious that V2V and V2I collaborative computing paradigms can effectively utilize resources, improve response time, and improve overall system performance.

[0004] However, the rapid movement of vehicles will lead to unstable communication links and dynamic changes in network topology, which will seriously affect data transmission performance and hinder the development of the Internet of Vehicles (IoV). Some methods select service vehicles by considering the link duration and distance between nodes, and use idle resources on vehicle nodes to offload computing tasks. Other methods select nearby vehicles with available idle computing resources to achieve concurrent task processing by considering factors such as link reliability, distance, available computing resources, and relative speed. In addition, V2V communication has also raised concerns about potential privacy violations and unauthorized exposure of personal data to the outside, causing some car owners to choose not to participate in V2V communication, making it impossible for some vehicles to establish connections with other vehicles. Therefore, in order to achieve reliable communication, the links between vehicles must form a certain topology to ensure communication quality and privacy security. The reliability of the link is closely related to a variety of factors. The link topology formed under the influence of indicators constructed according to these different factors will also be different.

[0005] In practice, it is impractical to decide which vehicles provide computing resources based on statistical information about resource requirements, because choices based solely on statistics cannot adapt to changing needs over time. Therefore, many studies have used reinforcement learning to optimize offloading strategies in IoT environments, because reinforcement learning allows vehicles to learn and optimize their communication strategies based on the penalties and rewards of the environment. By adopting reinforcement learning techniques, vehicles can learn how to dynamically optimize important factors such as transmission power, channel allocation, and resource allocation, thereby achieving more efficient and reliable V2V communication. Unfortunately, existing studies have ignored the high dynamics of topological connections between vehicles, which is crucial in real-world V2V communication scenarios.

[0006] Terminology explanation:

[0007] V2V: Vehicle-to-Vehicle, vehicle-to-vehicle communication;

[0008] V2A: Vehicle-to-Assist vehicle, communication between vehicle and assist vehicle;

[0009] V2I: Vehicle-to-Infrastructure, communication between vehicles and infrastructure. Summary of the invention

[0010] The purpose of the present invention is to deal with the unloading decision problem of the joint system of vehicle-to-vehicle (V2V) and vehicle-to-infrastructure (V2I) under different topological links, and proposes a link topology adaptive offloading method for edge computing of the Internet of Vehicles, namely Link Topology Adaptive Offloading (LTAO), which is used to adaptively offload tasks according to different link topologies in the joint system of V2V and V2I. It minimizes the total delay of the joint V2V and V2I system composed of multiple mobile vehicles. In order to find the optimal offloading strategy under different topological link structures, the topological link perception partial offloading process in the joint system of V2V and V2I is modeled as a Markov decision process (Markov Decision Process, MDP), and the graph convolutional neural network (Graph Convolutional Networks, GCN) and the deep double Q network (Double Deep-Q-Network, DDQN) are used to solve the optimization problem.

[0011] In a first aspect, the present invention provides a link topology adaptive offloading method for edge computing of an Internet of Vehicles, comprising the following steps:

[0012] Step 1: Obtain data sources, including topological links between vehicles and intelligent roadside devices, data transmission rate, and task computing capability, as well as the vehicle’s task data volume, task computing workload, coordinates, speed, and task initial time;

[0013] Step 2: Build a joint vehicle-to-vehicle and vehicle-to-infrastructure system, and then formulate the delay minimization problem of sequential subtasks in the joint system as a path optimization problem;

[0014] The vehicle-to-vehicle and vehicle-to-infrastructure joint system includes a plurality of ordinary vehicles V1, V2, ..., V M , an auxiliary vehicle V0 and an intelligent roadside device, M is an integer greater than 1;

[0015] Ordinary vehicles selectively communicate with other ordinary vehicles to form a V2V link, auxiliary vehicles communicate with any ordinary vehicle to form a V2A link, and intelligent roadside equipment communicates with any vehicle to form a V2I link; wherein, the auxiliary vehicle is responsible for generating tasks, and has the ability to execute its own designated tasks and unload tasks from other networked ordinary vehicles; the ordinary vehicle is responsible for generating tasks, and has the ability to execute its own designated tasks and unload tasks from other networked ordinary vehicles; the intelligent roadside equipment has the ability to execute tasks from the auxiliary vehicle itself and unload tasks from ordinary vehicles;

[0016] Step 3: Define the state space NF of the Markov decision processp , the state of each time step, the action space A, the reward, and the path optimization problem is modeled as a Markov decision process;

[0017] The Markov decision process uses a graph convolutional neural network (GCN) to adaptively learn the state S of the time step according to different link topologies. p The hidden information in the network is provided to the deep dual Q network DDQN to select the action with the highest reward to minimize the total delay.

[0018] In a second aspect, the present invention provides an electronic device, characterized in that it includes a processor and a memory, the memory stores machine executable instructions that can be executed by the processor, and the processor executes the machine executable instructions to implement the method.

[0019] In a third aspect, the present invention provides a machine-readable storage medium, characterized in that the machine-readable storage medium stores machine-executable instructions, and when the machine-executable instructions are called and executed by a processor, the machine-executable instructions prompt the processor to implement the method.

[0020] The present invention has the beneficial effects:

[0021] This paper develops a fine-grained partial offloading model for sequential subtasks in a dynamic V2I and V2V joint system with topological links to optimize the collaborative offloading strategy and thus achieve efficient utilization of computing resources. In addition, this paper also proposes a link topology adaptive offloading method based on GCN and DDQN to adapt to different link topologies and minimize the total delay. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 This is the overall architecture of the V2V and V2I joint system proposed in the present invention.

[0023] Figure 2 These are some possible structural patterns of topological links when the number of vehicles M = 4.

[0024] Figure 3 This is a schematic diagram of the link topology adaptive offloading method TLAO-DDQN for edge computing of Internet of Vehicles proposed in the present invention.

[0025] Figure 4 is the average delay of different methods under different numbers of vehicles, where the number of vehicles M = [5, 6, 7, 10, 15].

[0026] Figure 5 This is the comparison between TLAO-DDQN and the greedy algorithm and the greedy algorithm without subtask division, where the number of vehicles M = [5, 6, 7, 10, 15]. DETAILED DESCRIPTION

[0027] The present invention is further described below in conjunction with the accompanying drawings and specific embodiments.

[0028] Offloading methods play a key role in optimizing computation, minimizing latency, and enhancing the overall performance of vehicular edge computing (VEC) systems by transferring tasks between vehicles and edge servers or other computing resources. However, many works have ignored the topological links arising from differences in privacy requests, communication preferences, and transmission quality, which greatly affect the search space of optimal offloading decisions. Herein, a graph reinforcement learning (GRL) method is proposed to adaptively optimize the sum of latencies of tasks in joint vehicle-to-vehicle (V2V) and vehicle-to-infrastructure (V2I) communications, called linktopology-adaptive offloading (LTAO) method. The LTAO method consists of two modules in series: a graph convolutional network (GCN) module and a deep reinforcement learning (DRL) module, where the former module extracts features from the current state and the latter module outputs an offloading decision for the state. Extensive experimental results demonstrate the effectiveness and superiority of the proposed LTAO method.

[0029] Therefore, this embodiment provides a link topology-adaptive offloading method (LTAO) for edge computing of Internet of Vehicles. Figure 3 As shown in the figure, it includes the modeling of the V2V and V2I joint system, the input and output of the DDQN model, and the training process. First, the task is divided into K sequential subtasks, and the subtask information and topological link information are input from the environment as node features and adjacency matrices. Secondly, the node features and adjacency matrix are used as state information to learn new hidden representations through GCN, and the output of GCN is used as the input of DDQN for the next step of action selection learning to make the best choice to update the environment at the next moment.

[0030] The method comprises the following steps:

[0031] Step 1: Obtain data sources, including topological links between vehicles and intelligent roadside devices, data transmission rate R and task computing capacity f, as well as the vehicle's task data volume D, task computing workload, coordinates (x, y), speed v and task initial time T stsrt ;

[0032] The intelligent roadside equipment refers to a roadside unit RSU (Road Side Unit);

[0033] Step 2: Build a joint vehicle-to-vehicle and vehicle-to-infrastructure system, then formulate the delay minimization problem of sequential subtasks in the joint system as a path optimization problem and model it as a Markov decision process;

[0034] like Figure 1 As shown, the vehicle-to-vehicle and vehicle-to-infrastructure joint system includes a plurality of ordinary vehicles V1, V2, ..., V M , an auxiliary vehicle V0 and an intelligent roadside device, M is an integer greater than 1;

[0035] Ordinary vehicles selectively communicate with other ordinary vehicles to form a V2V link, auxiliary vehicles communicate with any ordinary vehicle to form a V2A link, and intelligent roadside equipment communicates with any vehicle (referring to ordinary vehicles and auxiliary vehicles) to form a V2I link; wherein, the auxiliary vehicle is responsible for generating tasks, and has the ability to execute its own designated tasks and unload tasks from other networked ordinary vehicles; the ordinary vehicle is responsible for generating tasks, and has the ability to execute its own designated tasks and unload tasks from other networked ordinary vehicles; the intelligent roadside equipment has the ability to execute tasks from the auxiliary vehicle itself and unload tasks from ordinary vehicles;

[0036] The specific task offloading relationship is: m The task is divided into K subtasks, where m∈{0,1,2,…,M} and the subtasks related to the same vehicle are arranged in order, and the subsequent subtasks are generated according to the previous subtasks; the initial task arrival time assigned to each ordinary vehicle is random, Indicates vehicle V m The arrival time of the task;

[0037] For each vehicle V m , whose subtask set is represented as Task m ={task m,1 ,task m,2 ,…,task m,K}; Specifically, vehicle V m The first subtask task m,1It can be executed within the vehicle itself or offloaded to intelligent roadside equipment, auxiliary vehicle V0 or other ordinary communication vehicles; the last subtask task m,K After successful completion, the results need to be transmitted back to the vehicle V m ;

[0038] The delay minimization problem of sequential subtasks in a joint V2I and V2V system consists of three parts: i) execution time, which depends on the amount of computation and the computing power of the device; ii) transmission time, which is calculated based on the data size and transmission rate (if a link exists); iii) waiting time, which is determined by the offloading strategy of the previous subtask and the limitation of available computing resources.

[0039] 1) Execution time:

[0040] Assume that Dev0 corresponds to auxiliary vehicle V0, Dev M+1 Corresponding to intelligent roadside equipment, Dev1, Dev2, ..., Dev M Corresponding to ordinary vehicles V1, V2, ..., V M , so it is recorded as device Dev l (where l∈{0,1,…,M+1}; device Dev l All have the ability to perform the subtasks assigned to them;

[0041] Subtask m,k The execution time is expressed as:

[0042]

[0043] in For ordinary vehicles V m The kth subtask task m,k ,Device l Execution time; exe represents the execution process; ω m,k Represents a subtask task m,k The computational workload, m∈{0,1,2,…,M},k∈{1,2,…,K}; fl represents the device Dev l The computing power of , in cycles per second;

[0044] Used to determine the task task m,K Is it on the device Dev l The calculation formula is as follows:

[0045]

[0046] 2) Transmission time:

[0047] Intelligent roadside equipment and auxiliary vehicles have the ability to communicate with all ordinary vehicles; however, due to privacy, trust authorization, and communication intent, the links between ordinary vehicles are limited; these restrictions are manifested as Mask m =[mask m,1 ,mask m,2 ,…,mask m,M ];

[0048]

[0049] Where m'∈{1,2,…,M}, m≠m';

[0050] task m,k Not always from the vehicle V m Transmit, unless k = 1. Introducing CD m,k Indicates the current task task m,k equipment, namely:

[0051] When k = 1, the vehicle V m Execute subtasks task by itself m,k ;

[0052] When k>1, execute subtask task m,k The device is recorded as CD m,k ; Therefore, the subtask task m,k Transmission time The calculation formula is:

[0053]

[0054] Where cur represents CD m,k ,D m,k Represents a subtask task m,k The data size, tr represents the transmission process; Represents a subtask task m,k Whether to transfer to the target device Dev l The indication value of mask cur,l Indicates from CD m,k To Dev l An indication value of whether a link exists;

[0055] Current Device CD m,k To deviceDev l The transmission rate R cur,l The calculation formula is:

[0056]

[0057] Where h is the channel fading coefficient, P is the transmission power, θ is the path loss index, and ω0 is the white Gaussian noise power; B cur,l and d cur,l Represents the current device CD m,k With the deviceDev l Bandwidth and distance between

[0058] 3) Waiting time:

[0059] When a vehicle or intelligent roadside equipment is m,k When the device is occupied, other subtasks offloaded to the occupied device need to wait for the subtask task m,k Finish;

[0060] When the subtask m,k Select Device l When executing, wait time The calculation is as follows:

[0061]

[0062] in, Device l The occupation time, K m is the vehicle V m The number of completed subtasks, It is a task m,k Arrival DeviceDev l The arrival time, ΔT m,k,l Is on the device Dev l Tasks executed on m,k The total delay of K m is the vehicle V m The number of completed subtasks, Indicates vehicle V m The time when the subtask starts to be generated;

[0063] The delay minimization problem of the joint V2I and V2V system is formulated as a path optimization problem, taking into account the possibility of offloading tasks according to their sequential subtask dependencies; the problem is formulated as follows:

[0064]

[0065] Where T sum Represents the total latency of executing all subtasks on all devices;

[0066] Finally, the path optimization problem is modeled as a Markov decision process;

[0067] Step 3: Based on step 2, define the state space NF of the Markov decision processp and the state at each time step;

[0068] In the V2I and V2V joint system, the time step t p Status S p Defined as:

[0069] S p =[NF p ,A p ] Formula (11)

[0070]

[0071] Among them, A p is the time t p Adjacency matrix of topological links; Yes V m At time t p The current position in Yes V m speed;

[0072] Step 4: Based on steps 2 and 3, define the action space A of the Markov decision process, and the action selection at each time step is given by the reinforcement learning method; the reinforcement learning method uses the deep dual Q network DDQN in step 6;

[0073] The action space is A = [a0, a1, ..., a M+1 ], where a(t p )=a0 represents the time t p Subtasks Executed on the auxiliary vehicle, a(t p )=[a1,…,a M ] represents the time t p Subtasks Executed on a normal vehicle, a(t p )=a M+1 Indicates time t p Subtasks Executed on intelligent roadside equipment;

[0074] Step 5: Based on steps 2, 3, and 4, define the reward of the Markov decision process. Define different rewards and penalties according to different situations. The core goal is to minimize the total delay.

[0075]

[0076] Condition 1 is defined as follows: If a(t p ) is not and at time t p The task number k p =K, then the task The result is preferentially transmitted back to the vehicle m p ∈ {0, 1, 2, …, M}, k p ∈ {1, 2, …, K}; represents the current device CD p at time t p and the connection status of other devices;

[0077] Condition 2 is defined as follows: If a(t p ) is not in and k p < K, then the task is preferentially transmitted to the connected device in ;

[0078] where C1, C2, C3 are three different constants used to adjust penalties and rewards according to the preferences of the optimization problem; represents the execution time of p the k -th sub-task p of a general vehicle a(t p ) at time t, represents the transmission time of p the k -th sub-task p of a general vehicle a(t p ) at time t, represents the waiting time when p the k -th sub-task p of a general vehicle is selected to be executed on a(t p );

[0079] Step 6: Based on Steps 2, 3, 4, and 5, use the graph convolutional neural network GCN to adaptively learn the state S of the time step according to different link topologies p in the hidden information and provide it to the deep double Q-network DDQN to select the action that obtains the highest reward to minimize the total delay of the system; specifically:

[0080] 6-1 Input the state information S p = [NF p , A p into the two-layer graph convolutional neural network GCN to learn the hidden information:

[0081]

[0082] where hp t p The hidden information at the moment is represented by W0 and W1, which are weight matrices. is the symmetric normalized adjacency matrix;

[0083] h p Input two layers of Multi-Layer Perceptron (MLP) to obtain the Q value:

[0084] Q p =MLP(h p )=W3ReLU(W2h p ) Formula (17)

[0085] Where Q p Indicates time t p Q value; W2, W3 are the weight matrices of MLP;

[0086] 6-2 The deep dual Q network DDQN includes a main network and a target network;

[0087] The target network is based on the current action a p , output Q T ;

[0088] The Q function of the main network is updated as:

[0089]

[0090] where r p It is through t p The reward measured by the time delay in h p+1 and a p+1 They are the hidden information representation and action at the next moment, Q M (t p+1 ) is the main network at t p+1 Q value prediction at the moment, Q T is the output of the target network; is time t p The Q function of the main network; represents the updated Q function of the main network;

[0091] By being able to obtain the maximum of Output a p+1 That is, at time t p+1 The action selected by the optimal unloading solution. According to the action selected by the optimal unloading solution, the task execution process can be known.

[0092] Figure 2These are some possible connection modes of topological links when the number of vehicles M=4, forming different adjacency matrices and Masks.

[0093] Figure 4 is the average delay of different methods under different vehicle numbers, where the number of vehicles M = [5, 6, 7, 10, 15]. Figure 4 As shown in the figure, the results of TLAO-DDQN of the present invention are significantly better than other methods. For example, when the experiment is set to M=5 and K=5, TLAO-DDQN is 3.71% better than TLAO-DQN, 16.9% better than Greedy, 16.9% better than Greedy-noseg, 58.49% better than All Edge, 66.26% better than All Local, and 71.76% better than Random. This is because as M increases, the available equipment of the vehicle also increases, thereby expanding the optimization space of the average time. Among the non-DRL-based algorithms, the greedy algorithm achieves the best unloading decision, while the DRL-based algorithms TLAO-DDQN and TLAO-DQN always perform better than the non-DRL-based algorithms at different M values. As the available equipment increases, the optimization space also expands, and the superiority of the DRL-based methods becomes obvious, indicating that they can effectively handle various unloading situations in V2I and V2V systems.

[0094] Figure 5 This is the comparison between TLAO-DDQN and the greedy algorithm and the greedy algorithm without subtask division, where the number of vehicles M = [5, 6, 7, 10, 15]. Figure 5 As shown in the figure, the results of TLAO-DDQN are significantly better than the two different greedy algorithms. For example, when the experiment is set to K = 5, corresponding to different M = [5, 6, 7, 10, 15], TLAO-DDQN improves by 29.81%, 16.90%, 10.05%, 11.62% and 10.60% compared with Greedy; and improves by 36.47%, 38.63%, 32.90%, 18.99% and 18.96% compared with Greedy-noseg. Since the computing power of RSU remains unchanged, as M increases, the computing power of V2V and V2I becomes insufficient, and the performance improvement of LTAO decreases as M increases, which means that the optimization space decreases.

Claims

1. A link topology adaptive offloading method for edge computing in Internet of Vehicles, characterized by: The method comprises the following steps: Step 1: Obtain data sources, including topological links between vehicles and intelligent roadside devices, data transmission rate, and task computing capability, as well as the vehicle’s task data volume, task computing workload, coordinates, speed, and task initial time; Step 2: Build a joint vehicle-to-vehicle and vehicle-to-infrastructure system, and then formulate the delay minimization problem of sequential subtasks in the joint system as a path optimization problem; The vehicle-to-vehicle and vehicle-to-infrastructure joint system includes a plurality of ordinary vehicles V1, V2, ..., V M , an auxiliary vehicle V0 and an intelligent roadside device, M is an integer greater than 1; Ordinary vehicles selectively communicate with other ordinary vehicles to form a V2V link, auxiliary vehicles communicate with any ordinary vehicle to form a V2A link, and intelligent roadside equipment communicates with any vehicle to form a V2I link; wherein, the auxiliary vehicle is responsible for generating tasks, and has the ability to execute its own designated tasks and unload tasks from other networked ordinary vehicles; the ordinary vehicle is responsible for generating tasks, and has the ability to execute its own designated tasks and unload tasks from other networked ordinary vehicles; the intelligent roadside equipment has the ability to execute tasks from the auxiliary vehicle itself and unload tasks from ordinary vehicles; Step 3: Define the state space NF of the Markov decision process p , the state of each time step, the action space A, the reward, and the path optimization problem is modeled as a Markov decision process; The Markov decision process uses a graph convolutional neural network (GCN) to adaptively learn the state S of the time step according to different link topologies. p The hidden information in the deep double Q network DDQN is provided to select the action with the highest reward to minimize the total delay.

2. The method according to claim 1, characterized in that The intelligent roadside equipment refers to a roadside unit RSU.

3. The method according to claim 1, characterized in that The task offloading relationship in the joint system of vehicle-to-vehicle and vehicle-to-infrastructure is as follows: m The task is divided into K subtasks, where m∈{0,1,2,…,M} and the subtasks related to the same vehicle are arranged in order, and the subsequent subtasks are generated according to the previous subtasks; the initial task arrival time assigned to each ordinary vehicle is random, Indicates vehicle V m The arrival time of the task.

4. The method according to claim 3, characterized in that In the combined vehicle-to-vehicle and vehicle-to-infrastructure system, for each vehicle V m , whose subtask set is represented as Task m ={task m,1 ,task m,2 ,…,task m,K Specifically, vehicle V m The first subtask task m,1 It can be executed within the vehicle itself or offloaded to intelligent roadside equipment, auxiliary vehicle V0 or other ordinary communication vehicles; the last subtask task m,K After successful completion, the results need to be transmitted back to the vehicle V m .

5. The method according to claim 3, characterized in that In vehicle-to-vehicle and vehicle-to-infrastructure joint systems, the delay minimization problem of sequential subtasks consists of the following three parts: execution time, transmission time, and waiting time. The delay minimization problem of the joint system is formulated as a path optimization problem, and the possibility of offloading tasks according to their sequential subtask dependencies is considered.

6. The method according to claim 5, characterized in that The specific process of obtaining the execution time is: Assume that Dev0 corresponds to auxiliary vehicle V0, Dev M+1 Corresponding to intelligent roadside equipment, Dev1, Dev2, ..., Dev M Corresponding to ordinary vehicles V1, V2, ..., V M , so it is recorded as device Dev l (where l∈{0,1,…,M+1}; device Dev l All have the ability to perform the subtasks assigned to them; Subtask m,k The execution time is expressed as: in For ordinary vehicles V m The kth subtask task m,k ,Device l Execution time; exe represents the execution process; ω m,k Represents a subtask task m,k The computational workload of f, m∈{0,1,2,…,M},k∈{1,2,…,K}; l Indicates the device Dev l The computing power of , in cycles per second; Used to determine the task task m,K Is it on the device Dev l The calculation formula is as follows: The specific process of obtaining the transmission time is: Intelligent roadside equipment and auxiliary vehicles have the ability to communicate with all ordinary vehicles; however, due to privacy, trust authorization, and communication intent, the links between ordinary vehicles are limited; these restrictions are manifested as Mask m =[mask m,1 ,mask m,2 ,…,mask m,M ]; Where m'∈{1,2,…,M}, m≠m'; When k = 1, the vehicle V m Execute subtasks task by itself m,k ; When k>1, execute subtask task m,k The device is recorded as CD m,k ; Therefore, the subtask task m,k Transmission time The calculation formula is: Where cur represents CD m,k ,D m,k Represents a subtask task m,k The data size, tr represents the transmission process; It indicates the subtask task m,k Whether to transfer to the target device Dev l The indication value of mask cur,l Indicates from CD m,k To Dev l An indication value of whether a link exists; Current Device CD m,k To deviceDev l The transmission rate R cur,l The calculation formula is: Where h is the channel fading coefficient, P is the transmission power, θ is the path loss index, and ω0 is the white Gaussian noise power; B cur,l and d cur,l Represents the current device CD m,k With the deviceDev l Bandwidth and distance between The specific process of obtaining the waiting time is: When a vehicle or intelligent roadside equipment is m,k When the device is occupied, other subtasks that are offloaded to the occupied device need to wait for the subtask task m,k Finish; When the subtask m,k Select Device l When executing, wait time The calculation is as follows: in, Device l The occupation time, K m is the vehicle V m The number of completed subtasks, It is a task m,k Arrival DeviceDev l The arrival time, ΔT m,k,l Is on the device Dev l Tasks executed on m,k The total delay, K m is the vehicle V m The number of completed subtasks, Indicates vehicle V m The time when the subtask starts to be generated; The delay minimization problem of the joint system is formulated as a path optimization problem as follows: Where T sum Represents the total latency of executing all subtasks on all devices.

7. The method according to claim 1, characterized in that The state space NF of the Markov decision process in step 3 p And the state of each time step is specifically: In the combined vehicle-to-vehicle and vehicle-to-infrastructure systems, the time step t p Status S p Defined as: S p = [NF p , A p Equation (11) Among them, A p is the time t p Adjacency matrix of topological links; Yes V m At time t p The current position in Yes V m speed; The action space of the Markov decision process A=[a0,a1,…,a M+1 ], where a(t p )=a0 represents the time t p Subtasks Executed on the auxiliary vehicle, a(t p )=[a1,…,a M ] represents the time t p Subtasks Executed on a normal vehicle, a(t p )=a M+1 Indicates time t p Subtasks Executed on intelligent roadside equipment; The rewards of the Markov decision process are specifically: in C1, C2, and C3 are three different constants used to adjust penalties and rewards according to the preferences of the optimization problem; Indicates time t p For ordinary vehicles kth p Subtasks a(t p ) execution time, Indicates time t p For ordinary vehicles kth p Subtasks a(t p ) transmission time, Indicates time t p For ordinary vehicles kth p Subtasks Choose in a(t p ) when executing on the 8. The method according to claim 7, characterized in that The Markov decision process in step 3 is specifically: (1) The status information S p =[NF p ,A p ] Input a two-layer graph convolutional neural network GCN to learn hidden information: where h p t p The hidden information at the moment is represented by W0 and W1, which are weight matrices. is the symmetric normalized adjacency matrix; h p Input two layers of Multi-Layer Perceptron (MLP) to obtain the Q value: Q p =MLP(h p )=W3ReLU(W2h p ) Formula (17) Where Q p Indicates time t p Q value; W2, W3 are the weight matrices of MLP; (2) The deep dual Q network DDQN includes a main network and a target network; The target network is based on the current action a p , output Q T ; The Q function of the main network is updated as: where r p It is through t p The reward measured by the time delay in h p+1 and a p+1 They are the hidden information representation and action at the next moment, Q M (t p+1 ) is the main network at t p+1 Q value prediction at the moment, Q T is the output of the target network; is time t p The Q function of the main network; represents the updated Q function of the main network; By being able to obtain the maximum of Output a p+1 That is, at time t p+1 The action selected by the optimal unloading solution.

9. An electronic device, characterized in that: The invention comprises a processor and a memory, wherein the memory stores machine executable instructions that can be executed by the processor, and the processor executes the machine executable instructions to implement the method according to any one of claims 1 to 8.

10. A machine-readable storage medium, characterized in that: The machine-readable storage medium stores machine-executable instructions. When the machine-executable instructions are called and executed by a processor, the machine-executable instructions prompt the processor to implement the method described in any one of claims 1 to 8.

Images