Joint task unloading and resource optimization design method in communication, inductance and calculation integrated Internet of Vehicles

Abstract

The invention relates to a joint task unloading and resource optimization design method in a communication-sensing-calculation integrated internet of vehicles, and belongs to the technical field of wireless communication. In the scheme, a task vehicle carries a communication-sensing integrated system, and needs to cooperatively complete a large task and independently complete respective sensing tasks; large-scale tasks can be split into a plurality of sequence dependent medium-scale tasks, and task vehicles can select local processing or unloading to service nodes. According to the method, an efficient task unloading and resource allocation algorithm is jointly designed, and an unloading strategy X, a power allocation strategy P and a bandwidth allocation strategy B are optimized, so that the maximum sub-task splitting number constraint, the maximum bandwidth constraint, the maximum transmitting power constraint, the vehicle communication range constraint and the vehicle radar sensing threshold constraint are satisfied; and the time delay of a large task is minimized. Simulation results show that the sensing performance is considered, the processing time delay of the large chain task in the communication, sensing and calculation integrated Internet of Vehicles is effectively reduced, and the method has very high application value and practical significance.

Description

Technical Field

[0001] The present invention belongs to the technical field of wireless communication networks, and in particular relates to an optimization method for joint task offloading and resource allocation in a telematics-computing integrated vehicle network. Background Art

[0002] With the paradigm shift of the fifth-generation mobile communication network (5G), the sixth-generation mobile communication network (6G) is no longer just focused on the expansion of communication services, but is centered on the construction of the integration of sensing, communication and computing (ISCC) architecture. The future frequency band will expand to millimeter waves and even terahertz, which has the characteristics of large bandwidth and high penetration. At the same time, the deployment of larger-scale antennas and denser networks in the future will greatly promote the development of perception systems. On the other hand, the hardware design and configuration of communication systems and perception systems are becoming more and more similar, which promotes the integration of communication systems and perception systems, allowing the two to share limited resources, namely spectrum, energy and hardware platforms, thereby improving the efficiency of perception and communication and achieving an increase in integration gain. In the future, through in-depth research on 6G Integrated Sensing and Communication (ISAC) technology, it is expected to provide diversified perception services, which provides important technical support for the further development of the field of Vehicular Networks (VNs). However, in this case, it is difficult to complete the diverse sensing services with the low computing power of the vehicle itself, and the data of such applications are usually large in volume, highly complex, and extremely sensitive to latency, which also poses serious challenges to traditional cloud-based networks. Therefore, Vehicular Edge Computing (VEC) technology is studied to solve the challenge of long-distance offloading. In VEC, task offloading enables vehicles to offload computing tasks to nearby RSUs with computing capabilities. However, the traditional vehicle-to-roadside unit (V2R) offloading method may bring high latency and energy consumption. For example, a large number of tasks occupy the limited computing resources of RSUs, resulting in an excessive computing burden on RSUs, which requires additional waiting time. Fortunately, future vehicles can be considered as a distributed computing system because they can carry devices with certain computing power. Therefore, task vehicles with limited resources can transfer computing tasks to vehicles with available resources for execution through vehicle-to-vehicle (V2V) offloading. Therefore, by flexibly combining VEC and V2V methods, VNs can achieve efficient processing of a large number of complex sensing services.

[0003] Common perception services in future VNs are usually more complex, such as assisted navigation, mobile augmented reality and virtual reality applications. In addition to completing such tasks, vehicles also have tasks to maintain the performance of the vehicle's own perception system, and these two types of tasks will compete for limited transmission power and bandwidth resources. In addition, these large tasks are described in many studies as consisting of multiple interdependent medium-sized tasks, and such applications are usually modeled as directed acyclic graphs (DAGs), in which medium-sized tasks need to be processed in a specific order. Medium-sized tasks can often be split into multiple small tasks, and when processing these subtasks, problems such as computational offloading, resource allocation, and balanced communication perception performance are involved. The present invention designs a method for joint task offloading and resource optimization to achieve balanced perception performance while quickly completing tasks.

[0004] After searching, the application publication number CN111918311 B, a method for task offloading and resource allocation of Internet of Vehicles based on 5G mobile edge computing, belongs to the field of mobile communication technology. By adding a backup MEC server to make up for the problem of insufficient computing resources in the network, resources are shared between MEC servers to solve the problem of load imbalance, and the task offloading strategy and computing resource allocation are combined to minimize network overhead. Since this optimization problem is difficult to solve directly, the present invention proposes a heuristic algorithm to convert the original problem into local computing resource allocation and task offloading sub-problems and solve them separately. Among them, the local computing resource allocation sub-problem adopts the method of finding the extreme value of the derivative to obtain the optimal solution, and the task offloading sub-problem combines task offloading and MEC resource allocation, and adopts the Lagrangian optimization algorithm to solve it, thereby minimizing network delay and energy consumption overhead.

[0005] Traditional technologies focus more on task scheduling and low-dimensional resource allocation when solving chain tasks, and the proposed algorithms are relatively simple, which is ineffective in dealing with high-dimensional resource allocation problems in chain tasks. This patent gradually splits and solves complex problems by decoupling task offloading and resource allocation, and uses an improved deep reinforcement learning algorithm to solve high-dimensional resource allocation problems. Summary of the invention

[0006] The present invention aims to solve the above problems of the prior art. A joint task offloading and resource optimization design method in a telematics integrated vehicle network is proposed. The technical solution of the present invention is as follows:

[0007] A method for designing joint task offloading and resource optimization in a telematics-computing integrated vehicle network includes the following steps:

[0008] The task vehicles are equipped with a communication and perception integrated system. The task vehicles complete a large task together and complete their own perception tasks independently. The large task is split into multiple sequence-dependent medium-sized tasks. The task vehicles choose to process locally or offload to the service node. Under the premise of satisfying the maximum number of subtask splitting constraints, maximum bandwidth constraints, maximum transmission power constraints, vehicle communication range constraints and vehicle radar perception threshold constraints, the target problem is established and the task offloading strategy X, fine-grained bandwidth allocation strategy B(t) and power allocation strategy P(t) are solved to achieve the optimization of the total task delay.

[0009] Furthermore, the target problem is established under the premise of satisfying the maximum number of subtask splitting constraints, the maximum bandwidth constraint, the maximum transmission power constraint, the vehicle communication range constraint and the vehicle radar perception threshold constraint, specifically including:

[0010] In order to ensure that each medium-sized task can be split into a maximum number of x j , each small task x j,n,k The indivisibility of max Constraint, maximum available transmit power P max Constraints, the maximum communication range constraint between TV and SN, the radar rate threshold constraint of TV, the joint task offloading strategy X, the fine-grained bandwidth allocation strategy B(t) and power allocation strategy P(t), minimize the completion delay of chained large tasks. The target problem can be established as the following expression:

[0011] P1:

[0012]

[0013] The first constraint represents the splitting constraint for the medium-sized task, where the parameter represents the completion delay of a medium-sized task, where represents the delay of local processing, where x j,0 Indicates the number of subtasks left for local computation, x j Indicates the total number of tasks, d j Indicates the amount of data for the total task, c j Indicates the number of CPU revolutions required to complete 1 bit of data. Indicates the amount of computing resources of the vehicle; Indicates the processing delay from offloading to SN, including uplink delay Processing delay and downlink delay

[0014] The second and third constraints represent the power p j,n (t) and bandwidth b j,n(t) The fourth constraint represents the communication range constraint, where represents the communication time between a pair of TV and SN, represents the maximum tolerance of the communication time; the fifth represents the radar estimation rate constraint, where represents the radar estimation rate of TV at time t, η represents the minimum threshold of radar estimation rate; the sixth represents the non-negative constraint of resource allocation; the seventh represents the small task x j,n,k The indivisibility constraint is enforced to ensure that each small task is processed by one SN.

[0015] By solving the target problem, we obtain the task offloading strategy X, the fine-grained bandwidth allocation strategy B(t) and the power allocation strategy P(t), thus optimizing the total task delay.

[0016] Furthermore, pass get, Indicates the uplink transmission rate from TV to SN, h j,n (t) represents the channel condition, σ 2 , b j,n (t), B, and The definition of is the same as that in requirement 3, Indicates the processing delay at the SN end. Can be get.

[0017] Furthermore, the target problem is divided into a task offloading subproblem and a resource allocation subproblem, wherein the task offloading subproblem is solved using a heuristic algorithm; after obtaining the offloading decision, deep reinforcement learning is used to allocate resources on a small time scale, and finally the offloading and resource allocation strategies are obtained.

[0018] Furthermore, the task offloading sub-problem is solved by using a heuristic algorithm, which specifically includes:

[0019] Design a two-step algorithm with two time scales. The optimal offloading strategy is obtained by the designed heuristic algorithm on the large time scale. This problem can be reduced to P2. Then, resources are allocated according to the obtained offloading strategy on the small time scale. This problem can be reduced to P3. For the resource allocation sub-problem, the deep reinforcement learning DRL algorithm is used to solve it.

[0020] P2:

[0021] stC1,C4,C7

[0022] P3:

[0023] stC2~C6

[0024] First, the vehicle information and task information are taken as input, and then the parameters are initialized, where the task allocation matrix is ​​a zero matrix and the resource allocation matrix is ​​calculated by the formula defined as follows:

[0025]

[0026] It is assumed that the power and bandwidth The allocation ratio is the same. Assuming that the radar occupies the same proportion of bandwidth and power, the minimum resource ratio that meets the estimation threshold is res R ,◇is the trade-off index between computing resources and channel conditions, It is the normalized representation of the computing resources of all service nodes within the communication range of the mission vehicle. It is the normalized representation of the channel conditions between the task vehicle and these service nodes. The algorithm consists of three loops. The outer loop traverses all task vehicles, the middle loop traverses all subtasks, and the inner loop traverses all service nodes and local nodes. The subtask will be assigned to a node that minimizes the delay increase of this task.

[0027] Furthermore, for the resource allocation sub-problem, the deep reinforcement learning DRL algorithm is used to solve it, including:

[0028] DRL combines reinforcement learning (RL) and neural networks, combining RL decisions with neural networks in mapping state space and action space. The classic RL system consists of an agent and an environment. The agent receives feedback from the environment, takes appropriate actions in different states, and continuously learns and adjusts strategies, trying to learn the best actions in each round of interaction with the environment, so that the final cumulative reward reaches the maximum.

[0029] DRL is built on MDP. In DRL, MDP is used to model the observable environment. Several key elements of MDP include: state space, action space and reward function. After the constructed optimization problem is converted into an MDP problem, the corresponding state space, action space and reward function are obtained.

[0030] Furthermore, the state space, action space and reward function specifically include:

[0031] S(t)={id TV (t),z TV (t)} represents the state space, where id TV (t) is the number of the current mission vehicle, z TV(t)=[z1(t),z2(t),...,z N (t)] represents the completion progress of subtasks on all service nodes of the current task vehicle;

[0032] A(t)={B(t),P(t)} represents the action space, where Β(t)=[b1(t),b2(t),...,b N+1 (t)] represents the bandwidth allocation matrix, P(t) = [p1(t), p2(t), ..., p N+1 (t)] represents the transmission power allocation matrix;

[0033] represents the reward of the agent for each round of interaction with the environment, R rad (t) represents the reward corresponding to the radar estimation rate. slot (t) represents the amount of data processed by the agent in each round of interaction with the environment, which can be expressed as:

[0034]

[0035] When the remaining tasks in the current time slot are insufficient to be processed, the amount of tasks to be processed is Otherwise (1-z n (t))×x j,n / x j ×d j , z n (t) indicates the progress of subtask completion. slot (t) = t up (t)+t sn (t) represents the total time consumed by the agent to process the amount of data in each round of interaction with the environment, t up (t) represents the time consumed by the agent to upload data in each round of interaction with the environment, which can be expressed as:

[0036]

[0037] t sn (t) represents the time consumed by the SN for computing and processing the uploaded data in each round of interaction between the agent and the environment. The radar estimation rate reward of the vehicle in this time slot can be expressed as:

[0038]

[0039] Where r0 and r1 are set constants; if the threshold is met, then r0 is used, otherwise The intelligent agent pursues the maximum reward in each interaction with the environment, thereby achieving efficient use of resources at each small time scale while meeting the radar estimation rate threshold, and ultimately achieving the goal of minimizing the total task delay.

[0040] An electronic device comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the method for designing joint task offloading and resource optimization in a telematics-computing integrated vehicle network as described in any one of the items is implemented.

[0041] A non-transitory computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the joint task offloading and resource optimization design method in a telematics-computing integrated vehicle network as described in any one of the items.

[0042] The advantages and beneficial effects of the present invention are as follows:

[0043] The present invention proposes a solution for joint task offloading and resource allocation in the integrated telematics and computing vehicle network. The offloading tasks, power resources and bandwidth resources are allocated in consideration of large and small time scales and dynamic VNs, so that the total completion delay of the task is minimized. At the same time, the radar estimation rate is guaranteed to be met. Simulation verification shows that the present invention performs very well under different task computing densities, TV density, perception thresholds and SN computing power, and realizes efficient complex task processing, providing important technical support for the development of future VNs, and has strong application value and development potential.

[0044] The present invention decouples the sub-problems of joint task offloading and high-dimensional resource allocation into sub-problems and solves them separately, thereby reducing the complexity of the problems. It adopts a simple heuristic algorithm to solve the task allocation sub-problem, and then adopts an improved deep reinforcement learning algorithm to solve the resource allocation sub-problem, thereby cleverly balancing the resource competition between offloading tasks and perception tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 It is a schematic diagram of the structure of a system model of a preferred embodiment provided by the present invention;

[0046] Figure 2 : Convergence comparison diagram of learning curve;

[0047] Figure 3 : Relationship diagram between SN computing resources and total task latency;

[0048] Figure 4 : Relationship between task computing density and total task latency;

[0049] Figure 5 : Relationship diagram between radar estimation rate threshold and total task delay;

[0050] Figure 6 : Relationship between the number of TVs and total task latency. DETAILED DESCRIPTION

[0051] The following will describe the technical solutions in the embodiments of the present invention in detail in conjunction with the accompanying drawings in the embodiments of the present invention. The described embodiments are only part of the embodiments of the present invention.

[0052] The technical solution of the present invention to solve the above technical problems is:

[0053] The present invention proposes a joint task offloading and resource optimization method in the integrated telematics and computing vehicle network. The VNs system includes 4 lanes, J task vehicles (TV) and N service nodes (SN), where the TV is equipped with a single antenna, several cameras and radars. All TVs in the VNs cooperate to complete a large task. Each TV is responsible for completing a medium-sized task that can be disassembled into X small tasks, and selectively offloading the small tasks to the SN. By formulating a detailed offloading plan and a fine-grained resource allocation plan (including power and bandwidth), the task completion delay can be effectively reduced while ensuring radar performance.

[0054] In the proposed VNs system, TV and SN are respectively denoted as TV = {1, ..., j, ..., J}, j∈N and SN = {1, ..., n, ..., N}, n∈N. The present invention represents the physical entity virtual mapping of TV as:

[0055] IR TV ={F TV ,P TV ,V TV ,L TV ,C TV ,D TV ,X TV ,Ra TV}

[0056] in represents the local computing resources of TV, P TV ={p1,...,p j ,...,p J} represents TV transmission power, Indicates the speed of TV, represents the TV two-dimensional coordinate, C TV ={c1,...,c j ,...,c J} represents the amount of computing resources required to complete the task. TV ={d1,...,d j,...,d J} represents the amount of task input data, X TV ={x1,...,x j ,...,x J}. Assuming the tasks can be evenly split, x j Indicates the maximum number of subtasks that a task can be split into. Indicates the maximum communication coverage of the TV.

[0057] Similarly, the physical entity virtual mapping of SN is expressed as:

[0058] IR SN ={F SN ,V SN ,L SN ,Ra SN}

[0059] in represents the local computing resources that SN estimates to have, represents the speed of SN, represents the two-dimensional coordinates of SN, Indicates the maximum coverage of the SN.

[0060] In order to measure the impact of vehicle movement on the decision of computing offloading tasks, the present invention first models the vehicle movement process. At the initial time t0, the coordinates of TV and SN can be expressed as: and Then the Euclidean distance between the two can be calculated as:

[0061]

[0062] in represents the initial relative distance between TV and SN on the X-axis, represents the initial relative distance between TV and SN on the y-axis, which can be expressed as:

[0063]

[0064] Due to the different directions of vehicles, the speed of the vehicle can be positive or negative. This topic stipulates that the speed of the vehicle traveling to the right is a positive value. This topic assumes that the vehicle does not change lanes and the speed of the vehicle in each lane is the same. Since the vehicle only moves in the X-axis direction, the relative distance between TV and SN on the X-axis from t0 to t1 can be expressed as:

[0065]

[0066] in Considering the coverage of SN and TV, the communication range between TV and SN is recorded as:

[0067]

[0068] Therefore, the maximum communication time between TV and SN can be expressed as:

[0069]

[0070] In rainy and foggy weather or at night, the environmental information collected by the on-board camera is often not accurate enough, so radar cooperation is needed. For radar perception, this project uses G j,tar (t) TV represents the detection response to the target and is expressed as:

[0071]

[0072] where ξ0 is the reflection coefficient determined by the radar cross section, is the path loss, which can be expressed as:

[0073]

[0074] in is the channel gain at the reference distance (i.e. 1m), is the path loss factor, dis j,tar (t) represents the distance between TV and the target, so the radar estimation rate of TV for the target can be expressed as:

[0075]

[0076] Where δ represents the radar duty cycle, μ represents the radar pulse period, η is a constant determined by the radar waveform, and p j,0 represents the power used for radar sensing, b j,0 The bandwidth ratio used for radar perception, B represents the total uplink bandwidth, represents the variance of the predicted radar return value, σ 2 Represents the environmental noise variance. Considering that the object needs to be detected by the vehicle when it enters the field of view, the radar estimation rate threshold η is required to be:

[0077]

[0078] where θ j,radar Indicates the detection distance requirement for sensing the target.

[0079] The present invention divides the task offloading types into two categories: local execution and V2X offloading.

[0080] (1) Local computing:

[0081] The task completion delay of local computing can be expressed as:

[0082]

[0083] where x j,0 Indicates the number of subtasks left for local computation, x j Indicates the total number of tasks, d j Indicates the amount of data for the total task, c j Indicates the number of CPU revolutions required to complete 1 bit of data. Indicates the computing resources of the vehicle.

[0084] (2) V2X offloading:

[0085] Assuming that there is large-scale fading in the channel between TV and SN, the channel gain between them is expressed as:

[0086] h j,n (t) = G t G r |dis j,n (t) -2

[0087] G t and G r They represent the gains of the transceiver respectively. Assuming the OFDMA access method is adopted, the uplink transmission rate from TV to SN can be expressed as:

[0088]

[0089] where b j,n Indicates the bandwidth ratio allocated by TV to SN, p j,n represents the power resource occupied by the communication link. Assume that the time required for the TV upload task is Then the data upload process satisfies:

[0090]

[0091] Using the above formula, we can calculate Therefore, the processing time of the task on the SN side can be described as:

[0092]

[0093] After the calculation is completed, SN transmits the data to the next TV, and its return time satisfies:

[0094]

[0095] The present invention regards the transmission power and bandwidth resource consumed by the downlink as the same as the uplink, so the downlink rate can be expressed as:

[0096]

[0097] Then the offloading task delay of a pair of TV and SN can be described as:

[0098]

[0099] Then the total time taken to complete a large task can be expressed as:

[0100]

[0101] According to the order of TV task completion, after all TV tasks are completed, RSU summarizes the calculation results of all tasks and then sends them to all vehicles. Considering that the summarized data is relatively small, the transmission delay can be ignored. Then the total task completion delay of the system can be expressed as

[0102] The goal of this invention is to jointly optimize task offloading decisions, transmission power and bandwidth resource allocation to minimize the completion delay of large tasks. Then the optimization problem can be expressed as:

[0103] P1:

[0104]

[0105] The first constraint represents the number of splits of the medium-sized task, and the second and third constraints represent the power p j,n (t) and bandwidth b j,n (t) The fourth constraint represents the communication range constraint, where represents the communication time between a pair of TV and SN, represents the maximum tolerance of the communication time; the fifth represents the radar estimation rate constraint, where represents the radar estimation rate of TV at time t, η represents the minimum threshold of radar estimation rate; the sixth represents the non-negative constraint of resource allocation; the seventh represents the small task x j,n,k The indivisibility constraint is set to ensure that each small task is processed by one SN.

[0106] It can be observed that delay optimization is a mixed integer nonlinear programming problem, and task offloading and resource allocation are coupled with each other. It is challenging to directly solve the above problems with traditional optimization methods, and the classic dual alternating optimization method provides an effective idea for solving such problems. However, when facing tasks with huge amounts of data, the unchanging resource allocation strategy during the long task execution cycle will cause a lot of unnecessary resource waste, thereby reducing the delay of task completion. In addition, the high-dimensional decision variables involved in this problem are relatively complex, which makes the decision dimension of this problem large in scale, and it is impractical to directly use the model-free deep reinforcement learning method in such a large decision space. Therefore, the present invention designs a two-step algorithm with dual time scales, and obtains the optimal offloading strategy on a large time scale through the designed heuristic algorithm. This problem can be attributed to P2; then, on a small time scale, resources are allocated according to the obtained offloading strategy, which effectively avoids the disadvantages of the fixed resource allocation strategy in a large time scale. This problem can be attributed to P3. As for the resource allocation sub-problem, due to the mutual coupling of the allocation of bandwidth resources and power resources, the traditional resource allocation algorithm is difficult to solve, so it is considered to use the deep reinforcement learning (DRL) algorithm to solve it.

[0107] P2:

[0108] stC1,C4,C7

[0109] P3:

[0110] stC2~C6

[0111] First, the vehicle information (position, speed, computing power, etc.) and task information (CPU, data volume, etc.) are taken as input, and then some parameters are initialized, such as task allocation and resource allocation, where the task allocation matrix is ​​a zero matrix and the resource allocation matrix is ​​calculated by the formula defined as follows:

[0112]

[0113] It is assumed that the power and bandwidth The allocation ratio is the same. Assuming that the radar occupies the same proportion of bandwidth and power, the minimum resource ratio that meets the estimation threshold is res R ,◇is the trade-off index between computing resources and channel conditions, It is the normalized representation of the computing resources of all service nodes within the communication range of the mission vehicle. It is the normalized representation of the channel conditions between the task vehicle and these service nodes. The algorithm consists of three layers of loops, and the main idea can be summarized as follows: the outer loop traverses all task vehicles, the middle loop traverses all subtasks, and the inner loop traverses all service nodes and local nodes. The subtask will be assigned to a node that minimizes the delay increase of this task.

[0114] DRL effectively combines reinforcement learning (RL) and neural networks, combining the advantages of RL in decision-making with the advantages of neural networks in mapping state space and action space. The classic RL system consists of an agent and an environment. The agent receives feedback from the environment, takes appropriate actions in different states, and continuously learns and adjusts its strategy, trying to learn the optimal action in each round of interaction with the environment, so that the final cumulative reward reaches the maximum.

[0115] The theoretical basis of DRL is built on MDP. In DRL, MDP is used to model the observable environment. Several key elements of MDP include: state space, action space and reward function. Next, we will briefly introduce the corresponding state space, action space and reward function after the optimization problem constructed in this chapter is converted into an MDP problem.

[0116] S(t)={id TV (t),z TV (t)} represents the state space, where id TV (t) is the number of the current mission vehicle, z TV (t)=[z1(t),z2(t),...,z N (t)] represents the completion progress of subtasks on all service nodes of the current task vehicle.

[0117] A(t)={B(t),P(t)} represents the action space, where Β(t)=[b1(t),b2(t),...,b N+1 (t)] represents the bandwidth allocation matrix, P(t) = [p1(t), p2(t), ..., p N+1 (t)] represents the transmission power allocation matrix.

[0118] Represents the reward of the agent for each round of interaction with the environment. slot (t) represents the amount of data processed by the agent in each round of interaction with the environment, which can be expressed as:

[0119]

[0120] When the remaining tasks in the current time slot are insufficient to be processed, the amount of tasks to be processed is Otherwise (1-z n (t))×x j,n / x j ×d j . T slot (t) = t up (t)+t sn (t) represents the total time consumed by the agent to process the amount of data in each round of interaction with the environment, t up (t) represents the time consumed by the agent to upload data in each round of interaction with the environment, which can be expressed as:

[0121]

[0122] t sn (t) represents the time consumed by the SN to calculate and process the uploaded data in each round of interaction between the agent and the environment. The specific calculation method is shown in Formula 14. The radar estimation rate reward of the vehicle in this time slot can be expressed as:

[0123]

[0124] Where r0 and r1 are set constants. If the threshold is met, then r0 is used; otherwise, The intelligent agent pursues the maximum reward in each interaction with the environment, thereby achieving efficient use of resources at each small time scale while meeting the radar estimation rate threshold, and ultimately achieving the goal of minimizing the total task delay.

[0125] In the invention, in order to verify the advantages of the proposed joint power and bandwidth resource allocation scheme, the scheme was simulated, and the convergence performance of the proposed algorithm and the performance of the delay minimization scheme were analyzed. The simulation parameters are set as follows: the number of task vehicles J = 2, the number of service nodes N = 3 (including 1 RSU, 2 vehicles), where the task vehicles and service vehicles are evenly distributed on 4 lanes with a length of 100m (RSU is on the roadside). In addition, other system parameter information is shown in the following table:

[0126] Table 1 System environment parameter settings

[0127]

[0128]

[0129] This method is compared with the binary method (considering medium-sized tasks as an indivisible whole), the coarse resource allocation method (optimizing task offloading and resource allocation on a large time scale) and the one-dimensional resource allocation method (optimizing only bandwidth resources), and the following simulation results are obtained:

[0130] Figure 2It means that as the number of learning rounds increases, the learning curve tends to be stable, proving that the algorithm has good convergence.

[0131] Figure 3 It represents the changing relationship between SN computing resources and the completion delay of large tasks. First, it can be seen that when SN computing resources gradually increase, the completion delay of the task is decreasing; but as the computing resources increase to a certain extent, the rate of delay reduction slows down. This is because the proportion of SN computing delay in the total delay is small enough. When computing resources are abundant, the impact on the total task delay will be small. Obviously, the proposed method is significantly better than other methods in all stages. This is because the binary offloading method does not fully utilize the computing resources of SN; the rough resource method allocation will result in the situation where resources are still allocated for completed tasks, resulting in resource waste; the monotonous resource allocation does not consider the impact of power.

[0132] Figure 4 It represents the changing relationship between the computing density of tasks and the completion delay of large tasks. First, it can be seen that when the computing density gradually increases, the completion delay of tasks is increasing; considering that the growth of total delay is composed of transmission delay and computing delay, when the high computing density further increases, the growth of task delay will slow down, but the rough resource allocation method is more advantageous than monotonic resource allocation at high computing density. This is because at high computing density, the computing delay accounts for a high proportion, and the two-layer alternating optimization algorithm used in rough resource allocation can reallocate small tasks, and the proposed method is significantly better than other methods in all stages.

[0133] Figure 5 It represents the changing relationship between the radar estimation rate and the completion delay of large tasks. First, it can be seen that when the radar estimation rate gradually increases, the completion delay of the task is increasing; when the radar estimation rate rises to a certain extent, the growth of the delay begins to accelerate. This is because the amount of resources required for the high radar estimation rate to increase further does not increase linearly. At this time, maintaining the radar estimation rate will consume a lot of resources, so the resources left for task offloading will be greatly reduced; when the radar estimation rate is low, the monotonic resource allocation method is slightly better than the rough resource allocation method. This is because the power required to meet the radar threshold is low at this time, and the fixed power resources will not have too much impact on the total delay. In addition, it can be seen that the proposed method is significantly better than other methods in all stages.

[0134] Figure 6 It shows the changing relationship between the number of TVs and the completion delay of large tasks. It can be seen that when the number of TVs gradually increases, the completion delay of the task shows an approximately linear upward trend, and this method is better than the monotonic resource allocation method, which is better than the rough resource allocation method, while the binary offloading method performs the worst.

[0135] The systems, devices, modules or units described in the above embodiments may be implemented by computer chips or entities, or by products with certain functions.

[0136] Computer readable media include permanent and non-permanent, removable and non-removable media that can be implemented by any method or technology to store information. Information can be computer readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, compact disk read-only memory (CD-ROM), digital versatile disk (DVD) or other optical storage, magnetic cassettes, magnetic tape magnetic disk storage or other magnetic storage devices or any other non-transmission media that can be used to store information that can be accessed by a computing device. As defined herein, computer readable media does not include temporary computer readable media (transitory media), such as modulated data signals and carrier waves.

[0137] It should also be noted that the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, commodity or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, commodity or device. In the absence of more restrictions, the elements defined by the sentence "comprises a ..." do not exclude the existence of other identical elements in the process, method, commodity or device including the elements.

[0138] The above embodiments should be understood to be only used to illustrate the present invention and not to limit the protection scope of the present invention. After reading the contents of the present invention, technicians can make various changes or modifications to the present invention, and these equivalent changes and modifications also fall within the scope defined by the claims of the present invention.

Claims

1. A joint task offloading and resource optimization design method in a telematics-computing integrated vehicle network, characterized in that: The following steps are involved: The task vehicles are equipped with a communication and perception integrated system. The task vehicles complete a large task together and complete their own perception tasks independently. The large task is split into multiple sequence-dependent medium-sized tasks. The task vehicles choose to process locally or offload to the service node. Under the premise of satisfying the maximum number of subtask splitting constraints, maximum bandwidth constraints, maximum transmission power constraints, vehicle communication range constraints and vehicle radar perception threshold constraints, the target problem is established and the task offloading strategy X, fine-grained bandwidth allocation strategy B(t) and power allocation strategy P(t) are solved to achieve the optimization of the total task delay.

2. The method for designing joint task offloading and resource optimization in the integrated telematics and computing vehicle network according to claim 1 is characterized in that: The target problem is established under the premise of satisfying the maximum number of subtask splitting constraints, the maximum bandwidth constraint, the maximum transmission power constraint, the vehicle communication range constraint and the vehicle radar perception threshold constraint, specifically including: In order to ensure that each medium-sized task can be split into a maximum number of x j , each small task x j,n,k The indivisibility of the maximum available bandwidth B max Constraint, maximum available transmit power P max Constraints, the maximum communication range constraint between TV and SN, the radar rate threshold constraint of TV, the joint task offloading strategy X, the fine-grained bandwidth allocation strategy B(t) and power allocation strategy P(t), minimize the completion delay of chained large tasks. The target problem can be established as the following expression: The first constraint represents the splitting constraint for the medium-sized task, where the parameter represents the completion delay of a medium-sized task, where represents the delay of local processing, where x j,0 Indicates the number of subtasks left for local computation, x j Indicates the total number of tasks, d j Indicates the amount of data for the total task, c j Indicates the number of CPU revolutions required to complete 1 bit of data. Indicates the amount of computing resources of the vehicle; Indicates the processing delay from offloading to SN, including uplink delay Processing delay and downlink delay The second and third constraints represent the power p j,n (t) and bandwidth b j,n (t) The fourth constraint represents the communication range constraint, where represents the communication time between a pair of TV and SN, represents the maximum tolerance of the communication time; the fifth represents the radar estimation rate constraint, where represents the radar estimation rate of TV at time t, η represents the minimum threshold of radar estimation rate; the sixth represents the non-negative constraint of resource allocation; the seventh represents the small task x j,n,k The indivisibility constraint is enforced to ensure that each small task is processed by one SN; By solving the target problem, we obtain the task offloading strategy X, the fine-grained bandwidth allocation strategy B(t) and the power allocation strategy P(t), thus optimizing the total task delay.

3. The method for designing joint task offloading and resource optimization in the integrated telematics and computing vehicle network according to claim 2 is characterized in that: pass get, Indicates the uplink transmission rate from TV to SN, h j,n (t) represents the channel condition, σ 2 represents the environmental noise variance, and They represent the maximum number of splittable tasks for each medium-sized task and the number of small tasks assigned to different SNs. represents the amount of data for each medium-sized task, represents the uplink transmission delay from TV to SN, b j,n (t) represents the bandwidth share of each TV-SN channel, B represents the total bandwidth, Indicates the uplink transmission power from TV to SN, Indicates the processing delay at the SN end. Indicates the number of CPU revolutions required to calculate bit data. Indicates the total computing power of SN, Can be get.

4. The method for designing joint task offloading and resource optimization in the integrated telematics and computing vehicle network according to claim 2 is characterized in that: The target problem is divided into a task offloading subproblem and a resource allocation subproblem, wherein the task offloading subproblem is solved by a heuristic algorithm; after obtaining the offloading decision, deep reinforcement learning is used to allocate resources on a small time scale, and finally the offloading and resource allocation strategies are obtained.

5. The method for designing joint task offloading and resource optimization in the integrated telematics and computing vehicle network according to claim 4 is characterized in that: The task offloading sub-problem is solved by using a heuristic algorithm, which specifically includes: Design a two-step algorithm with two time scales. The optimal offloading strategy is obtained by the designed heuristic algorithm on the large time scale. This problem can be reduced to P2. Then, resources are allocated according to the obtained offloading strategy on the small time scale. This problem can be reduced to P3. For the resource allocation sub-problem, the deep reinforcement learning DRL algorithm is used to solve it. stC1,C4,C7 stC2~C6 First, the vehicle information and task information are taken as input, and then the parameters are initialized, where the task allocation matrix is ​​a zero matrix and the resource allocation matrix is ​​calculated by the formula defined as follows: It is assumed that the power and bandwidth The allocation ratio is the same. Assuming that the radar occupies the same proportion of bandwidth and power, the minimum resource ratio that meets the estimation threshold is res R ,◇is the trade-off index between computing resources and channel conditions, It is the normalized representation of the computing resources of all service nodes within the communication range of the mission vehicle. It is the normalized representation of the channel conditions between the task vehicle and these service nodes. The algorithm consists of three loops. The outer loop traverses all task vehicles, the middle loop traverses all subtasks, and the inner loop traverses all service nodes and local nodes. The subtask will be assigned to a node that minimizes the delay increase of this task.

6. The method for designing joint task offloading and resource optimization in the integrated telematics and computing vehicle network according to claim 5 is characterized in that: For the resource allocation sub-problem, the deep reinforcement learning DRL algorithm is used to solve it, including: DRL combines reinforcement learning (RL) and neural networks, combining RL decisions with neural networks in mapping state space and action space. The classic RL system consists of an agent and an environment. The agent receives feedback from the environment, takes appropriate actions in different states, and continuously learns and adjusts strategies, trying to learn the best actions in each round of interaction with the environment, so that the final cumulative reward reaches the maximum. DRL is built on MDP. In DRL, MDP is used to model the observable environment. Several key elements of MDP include: state space, action space and reward function. After the constructed optimization problem is converted into an MDP problem, the corresponding state space, action space and reward function are obtained.

7. The method for designing joint task offloading and resource optimization in the integrated telematics and computing vehicle network according to claim 6 is characterized in that: The state space, action space and reward function specifically include: S(t)={id TV (t),z TV (t)} represents the state space, where id TV (t) is the number of the current mission vehicle, z TV (t)=[z1(t),z2(t),...,z N (t)] represents the completion progress of subtasks on all service nodes of the current task vehicle; A(t)={B(t),P(t)} represents the action space, where Β(t)=[b1(t),b2(t),...,b N+1 (t)] represents the bandwidth allocation matrix, P(t) = [p1(t), p2(t), ..., p N+1 (t)] represents the transmission power allocation matrix; represents the reward of the agent for each round of interaction with the environment, R rad (t) represents the reward corresponding to the radar estimation rate, D slot (t) represents the amount of data processed by the agent in each round of interaction with the environment, which can be expressed as: When the remaining tasks in the current time slot are insufficient to be processed, the amount of tasks to be processed is Otherwise (1-z n (t))×x j,n / x j ×d j , z n (t) represents the progress of subtask completion, T slot (t) = t up (t)+t sn (t) represents the total time consumed by the agent to process the amount of data in each round of interaction with the environment, t up (t) represents the time consumed by the agent to upload data during each round of interaction with the environment, which can be expressed as: t sn (t) represents the time consumed by the SN for computing and processing the uploaded data in each round of interaction between the agent and the environment. The radar estimation rate reward of the vehicle in this time slot can be expressed as: Where r0 and r1 are set constants; if the threshold is met, then r0 is used, otherwise The intelligent agent pursues the maximum reward in each interaction with the environment, thereby achieving efficient use of resources at each small time scale while meeting the radar estimation rate threshold, and ultimately achieving the goal of minimizing the total task delay.

8. An electronic device, characterized in that: It includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the joint task offloading and resource optimization design method in the integrated telematics and computing network as described in any one of claims 1 to 7.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, the method for designing joint task offloading and resource optimization in the integrated telematics and computing network of vehicles as described in any one of claims 1 to 7 is implemented.

Images