Topology link-aware task offloading method for v2v and v2i combined system

Abstract

The application discloses a topology link perception task cooperation unloading method for a V2V and V2I combined system. A delay minimization problem of sequential subtasks in a V2I and V2V combined system model is expressed as a path optimization problem and modeled as a Markov decision process to obtain an action of optimal unloading scheme selection in a time window, and then an execution task process is obtained. The execution task process is divided into time windows, the task calculation amount in each time window under the current pulse parameter is calculated, then Gaussian fitting is performed on the task calculation amount to obtain a Gaussian curve. Gaussian curve parameters are fitted to predict the task calculation amount in the time window under other pulse parameters. In combination with TLCO and STW, the application realizes efficient prediction of the calculation workload of the V2V and V2I systems in different time slices, can provide an optimization basis for unloading decision, and reduces the calculation delay and communication overhead.

Description

TECHNICAL FIELD

[0001] The application belongs to the application of reinforcement learning in the task offloading of the communication field, and particularly relates to a topology link-aware task cooperative offloading method for a V2V and V2I joint system. BACKGROUND

[0002] With the continuous progress of intelligent transportation systems, a series of innovative technologies have emerged in the field of Internet of Vehicles (IoV), among which Vehicle-to-Vehicle (V2V) and Vehicle-to-Infrastructure (V2I) communication technologies are particularly prominent. These technologies enable real-time exchange of information between vehicles and road infrastructure, providing strong support for intelligent traffic management, safe driving, and environmental perception.

[0003] In complex traffic environments such as urban congestion areas or emergency rescue scenarios, vehicles need to respond quickly and work cooperatively. To address these challenges, researchers have proposed cluster-based computing strategies, organizing vehicles into clusters and having cluster head vehicles handle computing tasks centrally, thereby improving overall computing efficiency and Quality of Service (QoS). In addition, V2V and V2I communication technologies play an important role in collision warning systems at unsignalized intersections, effectively reducing the risk of traffic accidents through accurate vehicle position and speed information.

[0004] However, the high-speed movement characteristics of vehicles result in instability of communication links, which poses a challenge to the stability and reliability of the Internet of Vehicles. To address this issue, researchers have proposed various strategies, such as dynamic resource allocation based on link duration and distance, and computing task offloading considering link reliability, computing resources, and relative speed. These strategies aim to optimize transmission efficiency in V2V and V2I communication, ensuring that critical information can be delivered in a timely and accurate manner.

[0005] Privacy protection is also an issue that cannot be ignored in the development of the Internet of Vehicles. When vehicles engage in V2V communication, sensitive information may be exposed, raising concerns about personal privacy. To balance communication efficiency and privacy protection, researchers are exploring how to build secure and reliable communication link topologies while ensuring communication quality. These structures need to consider multiple factors, such as link stability, data transmission security, and vehicle mobility.

[0006] In practical applications, simply relying on statistical information to allocate computing resources is not feasible, as demand is dynamically changing. For this reason, reinforcement learning has become an effective tool for optimizing resource allocation in vehicular networks. Through reinforcement learning, vehicles can adjust their communication and computing strategies based on real-time environmental feedback, enabling more flexible and efficient resource management. However, existing research often overlooks the importance of topology structure between vehicles, which is crucial in actual V2V communication. SUMMARY

[0007] To achieve efficient resource allocation and task offloading based on various topological link structures, a topology link-aware collaborative offloading method is proposed: Topological Link-Aware Task Co-Offloading Method for Joint V2V and V2I System (STW-TLCO). To find the optimal offloading strategy under different topological link structures, the topology link-aware part of the offloading process in the joint V2V and V2I system is modeled as a Markov Decision Process (MDP), and a Deep Double-Q-Network (DDQN) is used to solve the optimization problem. Based on the innovative prediction framework of TLCO and sliding time window module (STW-TLCO), the computing workload of the joint V2V and V2I system at different time slices is predicted. This prediction is achieved by analyzing the pulse parameters, which play a crucial role in determining the arrival time of tasks within the system.

[0008] In a first aspect, the present application provides a topology link-aware task co-offloading method for a joint V2V and V2I system, which is implemented using the following steps:

[0009] In a second aspect, the present application provides a topology link-aware task co-offloading system for implementing the method, which includes:

[0010] A data acquisition module is responsible for acquiring vehicular network data.

[0011] A data processing module is responsible for obtaining the optimal offloading scheme selection actions for each time window under the V2I and V2V joint system model.

[0012] A time window task computation prediction module is responsible for obtaining the execution task process based on the optimal offloading scheme selection actions for each time window, dividing the execution task process into time windows, calculating the task computation in each time window, and performing Gaussian fitting on the time window task computation under the current pulse parameter to predict the time window task computation under other pulse parameters.

[0013] In a third aspect, the present application provides an electronic device comprising a processor and a memory, the memory storing machine executable instructions capable of being executed by the processor, and the processor executes the machine executable instructions to implement the method.

[0014] In a fourth aspect, the present application provides a machine readable storage medium storing machine executable instructions, which when invoked and executed by a processor, cause the processor to implement the method.

[0015] The present application has the beneficial effects of:

[0016] The present application adopts the TLCO algorithm to provide accurate scheduling execution time for the tasks of the vehicle by considering the task arrival time and computing capacity, and ensures efficient completion of the tasks within a limited time window. The STW framework dynamically adjusts the time window for task execution through the sliding time window technology, adapts to the uncertainty of task arrival and the changes of system load, thereby optimizing resource allocation, and uses Gaussian curve fitting to predict the computing amount of each time window, so that the present application can more accurately predict the system workload. In combination with TLCO and STW, the present application realizes efficient prediction of the computing workload of the V2V and V2I system in different time slices, can provide optimization basis for offloading decision, and reduces computing delay and communication overhead. Therefore, the present application develops a fine-grained partial offloading model for sequential sub-tasks in a dynamic V2I and V2V system with a topology link, optimizes the collaborative offloading strategy, and realizes efficient utilization of computing resources, thereby reducing the waste of computing resources caused by improper task scheduling and improving the overall energy efficiency ratio of the system. BRIEF DESCRIPTION OF DRAWINGS

[0017] Figure 1 is the principle diagram of the method proposed by the present application.

[0018] Figure 2 is the overall architecture of the V2V and V2I combined system proposed by the present application.

[0019] Figure 3 is all possible structure patterns of the topology link when the number of vehicles M=4.

[0020] Figure 4 is the average delay of different numbers of sub-tasks, wherein the number of sub-tasks K=[5, 6, 7].

[0021] Figure 5 is the average delay of different numbers of vehicles, wherein the number of vehicles M=[5, 6, 7].

[0022] Figure 6is the average delay of assisting vehicles at different speeds, where v0 = [16, 18, 20] m / s.

[0023] Figure 7 is the impact of learning rate, seed, and speed on the reward of the training process.

[0024] Figure 8 is the computational load in different time windows when δ ∈ [1 / 9, 1 / 7, 1 / 5, 1 / 3, 1].

[0025] Figure 9 is the cumulative computational load in different time windows when δ ∈ [1 / 9, 1 / 7, 1 / 5, 1 / 3, 1].

[0026] Figure 10 is the difference between the predicted and actual values of the time window task computational load at different δ, where δ is 1 / 6, 1 / 8, and 1 / 10 in (a)-(c) figures, respectively. DETAILED DESCRIPTION

[0027] The present application is further described below with reference to the accompanying drawings and specific examples.

[0028] Vehicle-to-vehicle (V2V) and vehicle-to-infrastructure (V2I) joint offloading is an effective method to utilize the residual computing resources of neighboring devices, thus expanding the coverage of the Internet of Vehicles computing resource supply. However, many studies ignore the importance of topology communication caused by the rapid movement of vehicles, privacy, and communication intention. In order to achieve efficient task offloading in the face of various topology link structures, the present application proposes a novel topology link-aware task collaborative offloading method designed for partial offloading in V2V and V2I joint systems. The sequential sub-task offloading process is modeled as a Markov decision process, and the deep double Q network (DDQN) algorithm is used to optimize the total delay of the proposed system. Based on the innovative prediction framework of the TLCO and sliding time window module (STW-TLCO), the computational load in different time windows is accurately predicted using the pulse parameter. Compared with the state-of-the-art method, the proposed TLCO algorithm is more effective and superior. In addition, the STW-TLCO algorithm also shows high accuracy, with an R-square value of over 99%, confirming its prediction ability.

[0029] The present embodiment provides a topology link-aware task collaborative offloading method for V2V and V2I joint systems, as shown in Figure 1 The method comprises the following steps:

[0030] Step 1: Obtain Internet of Vehicles data, which includes the data transmission rate R, task computing capacity f and connection mode Mode of topological links of vehicles and intelligent roadside equipment RSU (such as Figure 3 ), as well as the vehicle's mission data volume D, mission computation workload ω, coordinates (x, y), speed v, and mission initial time T start ;

[0031] Step 2: Establish a V2I and V2V joint system model and formulate the delay minimization problem of sequential subtasks in the V2I and V2V joint system model as a path optimization problem.

[0032] The V2V and V2I joint system model is as follows Figure 2 At least a number of ordinary vehicles V1, V2, ..., V M , consisting of an auxiliary vehicle V0 and an intelligent roadside device RSU; ordinary vehicles selectively communicate with other ordinary vehicles to form a V2V link, the auxiliary vehicle can communicate with any vehicle to form a V2A link, and the intelligent roadside device RSU can communicate 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 device has the ability to execute tasks from the auxiliary vehicle itself and unload tasks from ordinary vehicles;

[0033] The specific task offloading relationship is: the vehicle V 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 based on the previous subtasks; the arrival time of the initial task assigned to each ordinary vehicle is random, Indicates vehicle V m The arrival time of a task is equal to the arrival time of its initial task;

[0034] 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 of 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 ;

[0035] The delay minimization problem of sequential subtasks consists of three parts: i) execution time, which depends on the computation amount and the computing power of the device; ii) transmission time, which is calculated according to the data size and the transmission rate (if there is a link); iii) waiting time, which is determined by the offloading strategy of the previous subtask and the limitation of available computing resources;

[0036] Assume that Dev0 corresponds to the auxiliary vehicle V0, Dev M+1 corresponds to the intelligent roadside device, Dev1, Dev2, …, Dev M corresponds to the ordinary vehicle V1, V2, …, V M respectively, and Dev l is denoted as the device Dev l , which has the ability to execute the subtask assigned to it.

[0037] 1) Execution time:

[0038] The execution time of the subtask task m,k is denoted as:

[0039]

[0040] wherein denotes the execution time of the kth subtask task m of the ordinary vehicle V m,k on the device Dev l ; exe represents the execution process; ω m,k denotes the computation amount of the subtask task m,k , m ∈ {0, 1, 2, …, M}, k ∈ {1, 2, …, K}; f l denotes the computing power of the device Dev l , with the unit of cycles per second.

[0041] The formula for determining whether the task task m,K is executed on the device Dev l is as follows:

[0042]

[0043] 2) Transmission time:

[0044] The intelligent roadside device RSU and the auxiliary vehicle have the ability to communicate with all ordinary vehicles; however, due to privacy, trust authorization, and communication intent, etc., the links between ordinary vehicles will be limited; these limitations are represented as Mask m = [mask m,1 , maskm,2 ,…,mask m,M ].

[0045]

[0046] where m'∈{1,2,…,M},m≠m';task m,k is not always from the vehicle V m , unless k=1; introduce CD m,k to represent the device currently responsible for executing the task task m,k , that is:

[0047] When k=1, the vehicle V m performs the subtask task m,k itself;

[0048] When k>1, the device performing the subtask task m,k is denoted as CD m,k ; therefore, the transmission time of the subtask task m,k is calculated as:

[0049]

[0050] where cur represents the CD m,k , D m,k represents the data size of the subtask task m,k , and tr represents the transmission process; is an indication value indicating whether the subtask task m,k is transmitted to the target device Dev l ; mask cur,l is an indication value indicating whether there is a link from the CD m,k to the Dev l ;

[0051] The calculation formula of the transmission rate R m,k of the current device CD l to the device Dev cur,l is:

[0052]

[0053] 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 represent the bandwidth and distance between the current device CD m,k and the device Dev l , respectively;

[0054] 3) Waiting time: ​

[0055] When a vehicle or an intelligent roadside device is occupied by a current subtask task m,k , other subtasks offloaded to the occupied device need to wait for the completion of the subtask task m,k ;

[0056] When a subtask task m,k is selected to be executed on a device Dev l , the waiting time is calculated as follows:

[0057]

[0058]

[0059] where, is the occupation time of the device Dev l , K m is the number of completed subtasks of the vehicle V m , T m,k is the arrival time of task l to the device Dev m,k,l , and ΔT l is the total delay of task m,k executed on the device Dev m ; m is the time when the vehicle V m starts to generate subtasks, K m is the number of completed subtasks of the vehicle V sum ; The delay minimization problem of a dynamic V2I and V2V joint system is formulated as a path optimization problem, and the possibility of offloading tasks according to the order of subtasks is considered; the problem is formulated as follows:

[0060]

[0061]

[0062] where T sum represents the total delay of executing all subtasks on all devices;

[0063] Step 3: Based on step 2, model the path optimization problem as a Markov decision process;

[0064] 3-1 Define the state space and the state of each time step; specifically:

[0065] The state of time step t p is defined as:

[0066]

[0067]

[0068] Among them, is the result of the last subtask that needs to be transmitted back to vehicle V m MO ∈ Mode M is the mode of the topological link; is V m at the current position in the p-th time step, and is V m speed;

[0069] 3-2 Define the action space, and the action selection at each time step is given by the reinforcement learning method; the reinforcement learning method uses the deep double Q network DDQN;

[0070] The action space is A = [a0, a1,..., a M+1 , where a(t p ) = a0 means that at time t p subtask is executed on the auxiliary vehicle, a(t p ) = [a1,..., a M means that at time t p subtask is executed on a certain ordinary vehicle, a(t p ) = a M+1 means that at time t p subtask is executed on the intelligent roadside unit RSU;

[0071] 3-3 Define the reward, define different rewards and punishments according to different situations, and the core is to minimize the total delay;

[0072]

[0073] Among them C1, C2, C3 are three different constants used to adjust the punishment and reward according to the preference of the optimization problem;

[0074] The definition of condition 1 is as follows: If a(t p ) is not in , and the task number k p at time t p = K, then the task result is preferentially transmitted back to the vehicle The definition of condition 2 is as follows: If a(t p ) is not in , and k p < K, then the task result is preferentially transmitted to the connected device in ;

[0075] 3-4 Deep Double Q Network (DDQN) learns and selects the action that obtains the highest reward according to the state at each time to realize the minimum total time delay of the system; the Deep Double Q Network (DDQN) comprises a main network and a target network;

[0076] The target network outputs Q p according to the action a T at the current time;

[0077] The Q function of the main network is updated as follows:

[0078]

[0079] where r p is the reward measured by the time delay in t p , h p+1 and a p+1 are the hidden representation and action at the next time, Q M (t p+1 ) is the Q value prediction of the main network at t p+1 , and Q T is the output of the target network; is the Q function of the main network at time t p ; represents the Q function of the updated main network, that is, the Q function obtained after the update

[0080] a p+1 output by that can obtain the maximum is the action selected by the optimal offloading scheme at time t p+1 ;

[0081] Step 4: According to the action selected by the optimal offloading scheme in the time window in step 3, the task execution process is obtained, and the task execution process is divided into time windows, each window has the same size and is fixed, and the task computation amount in each time window is calculated; the task computation amount in the time window under the current pulse parameter is Gaussian fitted to obtain a Gaussian curve; the Gaussian curve parameters are fitted again to predict the Gaussian curve parameters under other pulse parameters, and the task computation amount in the time window under other pulse parameters is predicted by reconstructing the Gaussian curve.

[0082] 4-1 The action obtained in step 3 is used to execute all tasks, the task execution process is divided into time windows, and the task computation amount in each time window is calculated; specifically:

[0083] The task computation amount of each time window is the sum of the task computation amounts of multiple sub-tasks in the time window, and the task computation amount of each sub-task is specifically:

[0084]

[0085] in Indicates task m,k exist On execution, The computing power of wd Indicates the start time of the time window, t u Indicates the size of the time window; represents the vehicle V under the pulse parameter δ m task m,k Arrival time, lie in within the scope; In the time slot of the impulse workflow, tasks from multiple vehicles arrive randomly at the time slot within a multiple of δ; a k ∈[a0,a M+1 ] represents the action of selecting the optimal unloading solution obtained in step 3; C is a constant; Indicates a k Execute subtask task m,k Total delay of

[0086] Therefore, the computational workload of the time window is:

[0087]

[0088] 4-2 According to the task calculation amount of each time window under the pulse parameter δ, the following Gaussian curve is fitted,

[0089]

[0090] Where y = Com(t wd ,δ), x represents the time window, b, c, d are the Gaussian curve parameters.

[0091] Then, we use the power function to fit b, c, and d to predict the corresponding Gaussian curve parameters for other pulse parameters δ'. Based on the predicted Gaussian curve parameters for other pulse parameters, we construct a Gaussian curve to obtain the task computational load for each time window under the pulse parameter δ'.

[0092] Figure 3 This is an example of all possible structural modes of topological links when the number of vehicles M = 4. There are 10 connection modes between the 4 devices, forming different masks.

[0093] Figure 4 is the average delay of different numbers of subtasks, where the number of subtasks K = [5, 6, 7]. Figure 4As shown, TLCO-DDQN (i.e., the method without step 4 in the present invention) performs significantly better than other methods. When M = 5 and K = 5, TLCO-DDQN improves by 5.19% over TLCO-DQN (i.e., replacing the DDQN in TLCO-DDQN with a deep Q network DQN), 3.23% over TLCO-DuelingDQN, 22.57% over Greedy, 43.34% over Greedy-Noseg (i.e., the Greedy algorithm that merges subtasks into a single task), 58.72% over All Edge, 136.12% over All Local, and 186.82% over Random. As K increases, the average execution time increases due to the increase in the total computational workload of the subtasks. At the same time, the total computing power of the device remains unchanged, resulting in a narrower optimization space. Therefore, the advantages of the DRL-based algorithm gradually weaken. Among the non-DRL based algorithms, Greedy achieves the best offloading decision, while the three DRL based algorithms, TLCO-DDQN, TLCO-DQN and TLCO-DuelingDQN, consistently outperform the non-DRL based algorithms under different K values.

[0094] Figure 5 is the average delay for different numbers of vehicles, where the number of vehicles M = [5, 6, 7]. Figure 5 As shown in the figure, TLCO-DDQN significantly outperforms other methods. For example, when the experiment is set to M = 6 and K = 5, TLCO-DDQN significantly improves 7.57% over TLCO-DQN, 1.52% over TLCO-DuelingDQN, 11.44% over Greedy, 30.35% over Greedy-Noseg, 66.47% over All Edge, 96.91% over All Local, and 162.31% over Random. As the value of M increases, the average execution time also increases. Compared with the increase in K value, the results based on the DRL algorithm show greater efficiency improvement. This is because as M increases, the available equipment of the vehicle also increases, thereby expanding the optimization space of the average time. Among non-DRL-based algorithms, the greedy algorithm achieved the best offloading decision, while the three DRL-based algorithms—TLCO-DDQN, TLCODQN, and TLCO-DuelingDQN—consistently outperformed the non-DRL-based algorithms for different values ​​of M. As the number of available devices increases, the optimization space also expands, and the superiority of DRL-based methods becomes apparent, demonstrating their ability to effectively handle a variety of offloading scenarios in V2I and V2V systems.

[0095] Figure 6is the average delay of the assisting car at different speeds, where v0= [16, 18, 20] m / s. To study the effect of the car speed on the experimental results, we set the car speed to 16, 18, and 20 m / s for comparison. As shown in Figure 6 , the change of the assisting car speed has little effect on the experimental results. This observation indicates that the proposed method can robustly determine the optimal unloading strategy regardless of the change of the car speed.

[0096] Figure 7 is the effect of learning rate, seed, and speed on the reward of the training process. Considering that different random seeds will lead to randomness in the experimental results, we set the seed of TLCO-DDQN to 0, 500, and 1000, respectively, to reduce the effect of randomness on the experimental results. In addition, we also compare the effect of different learning rates and assisting car speeds on the experimental results, with learning rates set to 0.005, 0.007, and 0.01, and speeds set to 16, 18, and 20 m / s, respectively. As shown in Figure 7 , the different colors of the range represent the average reward of different parameter sets of TLCO-DDQN. The smaller the range, the smaller the effect of different parameters. Different speeds and seeds have little effect on the convergence speed and results, while different learning rates only speed up the convergence, and the convergence results are consistent with different speeds and seeds.

[0097] Figure 8 is the computation amount in different time windows when δ∈[1 / 9, 1 / 7, 1 / 5, 1 / 3, 1]. We set the pulse parameter δ∈[1 / 9, 1 / 7, 1 / 5, 1 / 3, 1], which corresponds to the degree of task congestion. As δ increases, the arrival time of the task is delayed, resulting in a decrease in task density. The size of the time window is set to 0.1 seconds, and the interval of the time window is set to 0.02 seconds. As shown in Figure 8 , as δ decreases, the peak of the computation comes earlier and reaches a higher level, resulting in the completion of the task in advance. For example, when δ = 1, the computation amount is 0 near the 100th sliding time window; when δ = 1 / 7, the computation amount is 0 near the 80th sliding time window, and the peak is higher and earlier when δ = 1 / 7. When the task is sparse, the peak of the computation amount appears after a period of time. When the task is intensive, the peak of the computation amount appears earlier.

[0098] Figure 9 is the cumulative computation amount in different time windows when δ∈[1 / 9, 1 / 7, 1 / 5, 1 / 3, 1]. We set the pulse parameter δ∈[1 / 9, 1 / 7, 1 / 5, 1 / 3, 1]. As δ increases, the task completion time is advanced, i.e., the peak of the cumulative computation amount is advanced.

[0099] Figure 10is the difference between the predicted value and the actual value of the calculated amount for different δ. We predict the Gaussian curve parameters corresponding to the calculated amount within the sliding window according to different δ ∈ [1 / 8, 1 / 6, 1 / 4, 1 / 2] values. These predicted values are compared with the actual values calculated by TLCO. The fitting results for different δ values are shown in Figure 10 .

Claims

1. A method for topology link-aware task offloading in V2V and V2I combined system, characterized in that The method comprises the following steps: Step 1: acquiring Internet of Vehicles data, wherein the Internet of Vehicles data comprises data transmission rates, task computing capabilities and connection modes of topological links of vehicles and intelligent roadside devices RSUs, and task data volumes, task computing workloads, coordinate positions, speeds and task initial times of the vehicles; Step 2: establishing a V2I and V2V joint system model, and expressing a delay minimization problem of sequential subtasks in the V2I and V2V joint system model as a path optimization problem; Step 3: modeling the path optimization problem as a Markov decision process, and using a deep double Q network DDQN to solve the optimization problem, so as to obtain an action of optimal offloading scheme selection in a time window; Step 4: acquiring an execution task process according to the action of optimal offloading scheme selection in the time window in step 3, dividing the execution task process into time windows, calculating task computing amounts in the time windows, fitting a Gaussian curve of the task computing amounts in the time window under a current pulse parameter, predicting Gaussian curve parameters under other pulse parameters, and predicting the task computing amounts in the time windows under the other pulse parameters by reconstructing the Gaussian curve.

2. The method of claim 1, wherein The V2V and V2I joint system model is composed of at least a plurality of ordinary vehicles V1, V2,…,V M , an auxiliary vehicle V0 and an intelligent roadside device RSU; the ordinary vehicles are selectively connected with other ordinary vehicles to form a V2V link, the auxiliary vehicle can be connected with any vehicle to form a V2A link, and the intelligent roadside device RSU can be connected 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 connected 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 connected ordinary vehicles; and the intelligent roadside device has the ability to execute tasks from the auxiliary vehicle itself and unload tasks from ordinary vehicles.

3. The method of claim 1, wherein The delay minimization problem of sequential subtasks in the V2V and V2I joint system model comprises execution times, transmission times and waiting times; Dev0 corresponds to the auxiliary vehicle V0, Dev M+1 corresponds to the intelligent roadside device, Dev1, Dev2, …, Dev M corresponds to the ordinary vehicles V1, V2, …, V M , denoted as device Dev l , where l ∈ {0, 1, …, M+1}; the device Dev l all have the ability to perform the subtasks assigned to them; The delay minimization problem of the dynamic V2I and V2V joint system is expressed as a path optimization problem, and the expression is as follows: wherein respectively for the ordinary vehicle V m the kth subtask task m,k of the device Dev l the execution time, transmission time, waiting time, T sum denotes the total delay for executing all subtasks on all devices.

4. The method of claim 1, wherein Step 3 specifically comprises: 3-1 defining a state space and a state of each time step; 3-2 defining an action space, and an action selection of each time step is given by a reinforcement learning method; the reinforcement learning method uses a deep double Q network DDQN; 3-3 defining a reward, and defining different rewards and penalties according to different situations to minimize the total time delay; 3-4 the deep double Q network DDQN learns and selects an action with the highest reward according to a state at each time to minimize the total time delay of the system; the deep double Q network DDQN comprises a main network and a target network; Target network according to action a at current time p , output Q T ; The Q function of the main network is updated as follows: where r p It is through t p The reward is measured by the time delay in p+1 and a p+1 They are the hidden 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 Q function of the updated main network; by being able to obtain the maximum of output a p+1 is the optimal unloading scheme selection action at time t p+1 .

5. The method of claim 1 wherein Step 4 specifically comprises: 4-1 using the action obtained in step 3 to execute all tasks, dividing an execution task process into time windows, and calculating task computing amounts in the time windows; 4-2 fitting a Gaussian curve according to the task computing amounts in the time windows under a current pulse parameter; using a power function to fit Gaussian curve parameters to predict corresponding Gaussian curve parameters under other pulse parameters; constructing a Gaussian curve according to the predicted corresponding Gaussian curve parameters under the other pulse parameters, and then obtaining the task computing amounts in the time windows under the other pulse parameters.

6. The method of claim 5, wherein Step 4-1 specifically comprises: The task computing amount of each time window is a sum of task computing amounts of multiple subtasks in the time window, and the task computing amount of each subtask is specifically as follows: wherein denotes a task m,k In executes, the computing power of the task, t wd denotes the start time of the time window, t u denotes the size of the time window; denotes the arrival time of the vehicle V m task m,k ; a k ∈ [a0, a M+1 ] denotes the action of the optimal unloading scheme selection obtained in step 3, C is a constant; denotes a k total delay of the execution of the subtask taskm, k ; Thus, the task computation amount of the time window 7. The method of claim 5, wherein The Gaussian curve under the current pulse parameter in step 4-2 is specifically as follows: where y = Com(t wd , δ), x represents the time window, and b, c, d are Gaussian curve parameters.

8. A topology-aware task offload system implementing the method of any of claims 1-7, characterized in that The method comprises the following steps: A data acquisition module is responsible for acquiring Internet of Vehicles data; A data processing module is responsible for acquiring an action of optimal offloading scheme selection in each time window under a V2I and V2V joint system model. The time window task computation amount prediction module is responsible for obtaining an execution task process according to an action selected by the optimal unloading scheme under the time window, dividing the execution task process into time windows, and calculating the task computation amount in each time window; and performing Gaussian fitting on the time window task computation amount under the current pulse parameter to predict the time window task computation amount under other pulse parameters.

9. An electronic device, comprising: The processor and the memory, the memory stores machine executable instructions capable of being executed by the processor, the processor executes the machine executable instructions to realize the method in any one of claims 1-7.

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 the processor, the machine executable instructions cause the processor to realize the method in any one of claims 1-7.

Images