CN111918245B - Multi-agent-based vehicle speed perception calculation task unloading and resource allocation method - Google Patents

Abstract

The invention discloses a computing task unloading and resource allocation method based on multi-agent vehicle speed perception, comprising the following steps: collecting vehicle-side computing tasks, and dividing vehicle-side computing tasks into key tasks, high-level Priority tasks and low-priority tasks; under different computing resources and wireless resource allocation, respectively, offload to the VEC server, execute locally, and continue to wait for the corresponding delay and energy consumption, and then under the constraints of the delay threshold of each task, form The objective function is to reduce the energy consumption of the vehicle-side processing task; convert the objective function into a Markov decision process; train the multi-agent reinforcement learning network; input the state of the vehicle and edge server to be assigned to the trained In the multi-agent reinforcement learning network, the task offloading and resource allocation results are obtained, and this method can effectively improve the overall performance of the vehicle in the VEC server scenario.

Description

Computing task offloading and resource allocation method based on multi-agent vehicle speed perception

technical field

The invention belongs to the technical field of wireless communication, and relates to a computing task offloading and resource allocation method based on multi-agent vehicle speed perception.

Background technique

With the rapid development of the Internet of Things, intelligent in-vehicle applications (including autonomous driving, image-assisted navigation, and multimedia entertainment) have been widely used in intelligent vehicles, which can provide drivers and passengers with a more comfortable and safe environment. However, these in-vehicle applications consume a lot of computing resources and require extremely low processing time. Cloud servers with powerful computing and storage capabilities can be used to handle offload computing tasks from vehicles, but may lead to high latency due to long-distance transmission. , In order to deal with the drawbacks of cloud servers, Vehicle Edge Computing (VEC) came into being.

The VEC server is closer to the vehicle terminal and has powerful computing power, and is densely deployed next to the roadside unit (RSU). By offloading computing-consuming tasks to the VEC server, the processing delay and energy consumption of in-vehicle applications can be significantly reduced.

The development of VEC also faces many challenges. For example, the impact of vehicle speed on the delay threshold of computing tasks in the VEC network, the interaction between task processing delay and energy consumption and task offloading and resource allocation strategies. Therefore, the research on the task delay threshold of vehicle speed perception and the interaction of task processing delay and energy consumption with task offloading and resource allocation strategies are crucial to the overall performance of VEC.

SUMMARY OF THE INVENTION

The purpose of the present invention is to overcome the above-mentioned shortcomings of the prior art, and to provide a computing task offloading and resource allocation method based on multi-agent vehicle speed perception, which can effectively improve the overall performance of the vehicle in the VEC server scenario.

In order to achieve the above object, the multi-agent-based vehicle speed sensing method for offloading computing tasks and resource allocation according to the present invention includes the following steps:

1) Collect vehicle-side computing tasks, divide vehicle-side computing tasks into key tasks, high-priority tasks, and low-priority tasks according to the types of vehicle-side computing tasks, and then determine key tasks, high-priority tasks, and low-priority tasks according to vehicle speed. task delay threshold;

2) For key tasks, high-priority tasks and low-priority tasks, calculate different computing resources and wireless resource allocations, offload to the VEC server, execute locally, and continue to wait for the corresponding delay and energy consumption, and then calculate the corresponding delay and energy consumption of each task. Under the constraint of the delay threshold, an objective function is formed to reduce the energy consumption of the vehicle-side processing task;

3) Convert the objective function obtained in step 2) into a Markov decision process, and initialize the state space, action space and reward of the Markov decision process;

4) Obtain new states, actions and rewards according to the multi-agent reinforcement learning network, and store the new states, actions and rewards in the experience playback pool;

5) When the data in the experience playback pool reaches the threshold, train the multi-agent reinforcement learning network until the multi-agent reinforcement learning network converges;

6) Input the state of the vehicle terminal and edge server to be allocated into the multi-agent reinforcement learning network after training, get the task offloading and resource allocation results, and complete the computing task offloading and resource allocation based on multi-agent reinforcement learning for vehicle speed perception distribute.

的时延门限

为：According to the bandwidth and delay requirements of the vehicle-side computing tasks, the vehicle-side computing tasks are divided into key tasks φ ₁ , high-priority tasks φ ₂ and low-priority tasks φ ₃ , key tasks φ ₁ , high-priority tasks φ ₂ and The delay thresholds corresponding to the low-priority task φ ₃ are Thr ₁ , Thr ₂ and Thr ₃ respectively. The delay threshold of the key task φ ₁ is 10ms, the delay threshold of the low-priority task φ ₃ is 100ms, and the high-priority task φ 3 has a delay threshold of 100ms. _The level task φ2 is related to the current speed of the vehicle, where the task

delay threshold

for:

时，

当

时，

为车辆k的速度，Thr₂为道路限速v_max对应的时延门限。Among them, when

hour,

when

hour,

is the speed of vehicle k, and Thr ₂ is the delay threshold corresponding to the road speed limit v _max .

为：Transmission rate of uplink channel n assigned to vehicle k by the VEC server

for:

为上行信道干扰，信道带宽

为VEC服务器的上行总带宽，

为VEC服务器的上行信道个数，

为上行信道集合；where σ ² is the noise power, P is the transmission power,

is the uplink channel interference, the channel bandwidth

is the total upstream bandwidth of the VEC server,

is the number of upstream channels of the VEC server,

is the set of uplink channels;

表示车辆k与VEC服务器之间的上行信道n是否分配给车辆k，若分配，则

为1，否则，则

为0，得车辆k与VEC服务器之间的上行传输速率

为：Assume

Indicates whether the uplink channel n between vehicle k and the VEC server is allocated to vehicle k, if so, then

is 1, otherwise, then

is 0, get the uplink transmission rate between vehicle k and VEC server

for:

为：Transmission rate of downlink channel n assigned to vehicle k by the VEC server

for:

为下行信道干扰，信道带宽

为VEC服务器的下行总带宽，

为VEC服务器的下行信道个数，

为下行信道集合；where σ ² is the noise power, P is the transmission power,

is the downlink channel interference, the channel bandwidth

is the total downlink bandwidth of the VEC server,

is the number of downlink channels of the VEC server,

is a set of downlink channels;

表示车辆k与VEC服务器之间的下行信道n是否分配给车辆k，若分配，则

为1，否则，则

为0，得车辆k与VEC服务器之间的下行传输速率

为：Assume

Indicates whether the downlink channel n between vehicle k and the VEC server is allocated to vehicle k, if so, then

is 1, otherwise, then

is 0, the downlink transmission rate between vehicle k and VEC server

for:

卸载到VEC服务器执行消耗的总时延

为：task of vehicle k

The total latency consumed by offloading to the VEC server for execution

for:

为向上取整函数，

为车辆k的任务

的文件大小，

为处理车辆k的任务

需要的计算密度，

为车辆k下载的任务

的文件大小相对于原上传任务缩小的比例，

为VEC服务器给车辆k的任务

分配的计算资源比例，f^VEC为本地VEC服务器的CPU频率，

为车辆与VEC服务器之间的上行传输速率，

为下行传输速率。in,

is the round-up function,

task for vehicle k

file size,

for the task of handling vehicle k

required computational density,

Tasks downloaded for vehicle k

The size of the file is reduced relative to the original upload task,

Task given to vehicle k for the VEC server

The proportion of computing resources allocated, f ^VEC is the CPU frequency of the local VEC server,

is the uplink transmission rate between the vehicle and the VEC server,

is the downlink transmission rate.

在本地执行的时间消耗

为：Compute the task of vehicle k based on the CPU frequency f ^k assigned by vehicle k

Time spent executing locally

for:

能够选择继续等待、卸载到本地VEC服务器以及本地执行，设

表示车辆k的任务

是否继续等待，当继续等待，则

为1，否则，则

为0，设继续等待时间为T_h，

表示车辆k的任务

是否卸载到本地VEC服务器，当

为1时，则表示车辆k的任务

卸载到本地VEC服务器，对于车辆k的任务

从产生任务

到执行动作完成需要花费的总时延

为：At time t, the task of vehicle k

Can choose to continue to wait, offload to the local VEC server and execute locally, set

represents the task of vehicle k

Whether to continue waiting, when continuing to wait, then

is 1, otherwise, then

is 0, set the continuation waiting time as _Th ,

represents the task of vehicle k

Whether to offload to the local VEC server, when

When it is 1, it means the task of vehicle k

Offload to the local VEC server, for the task of vehicle k

Generate tasks from

The total delay it takes until the execution of the action is completed

for:

为根据车辆k的任务

产生的时间。in,

for the task according to vehicle k

generated time.

卸载到VEC服务器时，卸载任务的能量消耗包括上传计算任务消耗的能量及下载任务消耗的能量，卸载到VEC服务器的能量消耗

为：When the computing task of vehicle k

When offloading to the VEC server, the energy consumption of the offloading task includes the energy consumed by uploading computing tasks and the energy consumed by downloading tasks, and the energy consumption of offloading to the VEC server.

for:

在本地处理时，根据处理车辆k的任务

所需要的能量密度

得任务在本地处理的能量消耗

为：When the computing task of vehicle k

When processed locally, according to the task of processing vehicle k

required energy density

The energy consumption of the task to be processed locally

for:

At time t, after the vehicle executes the unloading strategy, the energy E(t) consumed by all vehicles within the service range of the local VEC server is:

Under the condition of limited task delay threshold, computing resources and wireless resources, the objective function formed to reduce the energy consumption of vehicle-side processing tasks is:

in,

The present invention has the following beneficial effects:

The method for unloading computing tasks and allocating resources based on multi-agent vehicle speed perception according to the present invention first collects computing tasks on the vehicle side during specific operations, and divides the tasks into key tasks and high-priority tasks according to the types of computing tasks on the vehicle side. and low-priority tasks, combined with the vehicle speed to obtain the delay threshold of different computing tasks, for different computing tasks, calculate different computing resources and wireless resource allocation, offload to the VEC server, execute locally, and continue to wait for the corresponding delay and energy. Under the constraints of the task delay threshold, an objective function is formed to reduce the energy consumption of the vehicle side, so as to comprehensively consider factors such as the position, speed, task queue, computing resources, wireless resources of the vehicle side, and computing resources and wireless resources of the VEC server side. , within the task processing delay threshold, the energy consumption of the vehicle can be effectively reduced, and then the objective function is transformed into a Markov decision process, and then the multi-agent reinforcement learning network is trained, and finally the trained multi-agent is used. The reinforcement learning network offloads computing tasks and allocates resources to improve the overall performance of the vehicle in the VEC server scenario.

Description of drawings

Fig. 1 is the schematic flow chart of the present invention;

Figure 2 is a distribution diagram of the vehicle average task completion delay (the number of vehicles is 7) corresponding to the five algorithms;

Fig. 3 is the distribution diagram of the vehicle average task completion delay (the number of vehicles is 9) corresponding to the five algorithms;

Figure 4 is a distribution diagram of the vehicle average task completion delay (the number of vehicles is 11) corresponding to the five algorithms;

Figure 5 is a distribution diagram of the vehicle average task completion delay (the number of vehicles is 13) corresponding to the five algorithms;

Fig. 6 is the distribution diagram of the vehicle average task energy consumption (the number of vehicles is 7) corresponding to the five algorithms;

Fig. 7 is the distribution diagram of vehicle average task energy consumption (the number of vehicles is 9) corresponding to five algorithms;

Figure 8 is a distribution diagram of the average task energy consumption of vehicles (the number of vehicles is 11) corresponding to the five algorithms;

Figure 9 is a distribution diagram of the average task energy consumption of vehicles (the number of vehicles is 13) corresponding to five algorithms;

Figure 10 is a distribution diagram of the vehicle average task completion delay (vehicle speed range is 30-50Km/h) corresponding to the five algorithms;

Figure 11 is a distribution diagram of the average task energy consumption of the vehicle (vehicle speed range is 30-50Km/h) corresponding to the five algorithms;

Figure 12 is a distribution diagram of the vehicle average task completion delay (vehicle speed range is 50-80Km/h) corresponding to the five algorithms;

Figure 13 is a distribution diagram of the average vehicle task energy consumption (vehicle speed range is 50-80Km/h) corresponding to the five algorithms.

Detailed ways

Below in conjunction with accompanying drawing, the present invention is described in further detail:

车辆的个数为K，在时刻t，车辆k需要处理的计算任务为

任务产生的时间为

任务在VEC服务器处理消耗的时延为

任务在本地执行消耗的时延为

则对于任务

从产生到执行动作完成需要花费的总时延

为：Assume that the set of vehicles within the service range of the local VEC server is set to

The number of vehicles is K. At time t, the computing task that vehicle k needs to process is:

The time when the task is generated is

The delay consumed by the task processing in the VEC server is

The latency consumed by the local execution of the task is

then for the task

The total delay from generation to completion of the execution of the action

for:

表示车辆k的任务

是否继续等待，如果继续等待，则

为1，否则，则

为0，继续等待时间设为T_h，

表示车辆k的任务

是否卸载到本地VEC服务器，

为1，则车辆k的任务

卸载到本地VEC服务器。in,

represents the task of vehicle k

Whether to continue to wait, if continue to wait, then

is 1, otherwise, then

is 0, the continuation waiting time is set to _Th ,

represents the task of vehicle k

Whether to offload to the local VEC server,

is 1, then the task of vehicle k

Offload to the local VEC server.

任务在本地执行消耗的能量为

当车辆执行卸载策略后，本地VEC服务器服务范围内所有车辆消耗的能量E(t)为：At time t, the energy consumed by the task processing on the VEC server is

The energy consumed by the task to execute locally is

When the vehicle executes the unloading strategy, the energy E(t) consumed by all vehicles within the service range of the local VEC server is:

The present invention takes minimizing the energy consumption of the vehicle end as the optimization goal, and under the condition that the task delay threshold, computing resources and wireless resources are limited, the corresponding optimization problem is:

s.t.

The computing task offloading and resource allocation method based on multi-agent vehicle speed perception according to the present invention comprises the following steps:

1) Collect vehicle-side computing tasks, divide vehicle-side computing tasks into key tasks, high-priority tasks, and low-priority tasks according to the types of vehicle-side computing tasks, and then determine key tasks, high-priority tasks, and low-priority tasks according to vehicle speed. task delay threshold;

2) For key tasks, high-priority tasks and low-priority tasks, calculate different computing resources and wireless resource allocations, offload to the VEC server, execute locally, and continue to wait for the corresponding delay and energy consumption, and then calculate the corresponding delay and energy consumption of each task. Under the constraint of the delay threshold, an objective function is formed to reduce the energy consumption of the vehicle-side processing task;

3) Convert the objective function obtained in step 2) into a Markov decision process, and initialize the state space, action space and reward of the Markov decision process;

4) Obtain new states, actions and rewards according to the multi-agent reinforcement learning network, and store the new states, actions and rewards in the experience playback pool;

5) When the data in the experience playback pool reaches the threshold, train the multi-agent reinforcement learning network until the multi-agent reinforcement learning network converges;

6) Input the state of the vehicle terminal and edge server to be allocated into the converged multi-agent reinforcement learning network, get the task offloading and resource allocation results, and complete the computing task offloading and resource allocation based on multi-agent reinforcement learning for vehicle speed perception .

The following is a detailed description with reference to Figure 1:

Step 11) According to the type of the task, the task is divided into key tasks, high-priority tasks and low-priority tasks, and the delay thresholds of different computing tasks are obtained in combination with the vehicle speed, and the specific process is as follows;

的时延门限

为：According to the bandwidth and delay requirements of the vehicle-side computing tasks, the vehicle-side computing tasks are divided into key tasks φ ₁ , high-priority tasks φ ₂ and low-priority tasks φ ₃ , key tasks φ ₁ , high-priority tasks φ ₂ and The delay thresholds corresponding to the low-priority task φ ₃ are Thr ₁ , Thr ₂ and Thr ₃ respectively. The delay threshold of the key task φ ₁ is 10ms, the delay threshold of the low-priority task φ ₃ is 100ms, and the high-priority task φ 3 has a delay threshold of 100ms. The level task φ ₂ is related to the current speed of the vehicle, and the task

delay threshold

for:

时，

当

时，

为车辆k的速度，Thr₂为道路限速v_max对应的时延门限。Among them, when

hour,

when

hour,

is the speed of vehicle k, and Thr ₂ is the delay threshold corresponding to the road speed limit v _max .

为：Step 12) For different computing tasks, when the task is offloaded to the VEC server for processing, the consumption delay

for:

为向上取整函数，

为车辆k的任务

的文件大小，

为处理车辆k的任务

需要的计算密度，

为车辆k下载的任务

的文件大小相对于原上传任务缩小的比例，

为VEC服务器给车辆k的任务

分配的计算资源比例，f^VEC为本地VEC服务器的CPU频率，

为车辆与VEC服务器之间的上行传输速率，

为下行传输速率；in,

is the round-up function,

task for vehicle k

file size,

for the task of handling vehicle k

required computational density,

Tasks downloaded for vehicle k

The size of the file is reduced relative to the original upload task,

Task given to vehicle k for the VEC server

The proportion of computing resources allocated, f ^VEC is the CPU frequency of the local VEC server,

is the uplink transmission rate between the vehicle and the VEC server,

is the downlink transmission rate;

为：At this time, the corresponding vehicle-side energy consumption

for:

Among them, P is the signal transmission power of the vehicle;

在本地执行的时间消耗

为：When the task is executed locally, the task of vehicle k is obtained according to the CPU frequency f ^k allocated by vehicle k

Time spent executing locally

for:

为：At this time, the corresponding vehicle-side energy consumption

for:

所需要的能量密度

Among them, the task of processing vehicle k

required energy density

Taking minimizing the energy consumption of the vehicle as the optimization goal, under the condition of limited task delay threshold, computing resources and wireless resources, the corresponding optimization problem is:

s.t.

Calculate the delay and energy consumption of different unloading locations and resource allocation, and form an optimization goal to reduce vehicle-side energy consumption under the delay constraint;

Step 13) Initialize the states, actions and rewards of the Markov decision process, at time t, define the state space of vehicle k as s _k (t), the state space of vehicle k includes the state information of other vehicles and the state of the VEC server information, _sk (t) is:

表示在时刻t车辆k是否选择卸载位置(·)，如果选择，则

为1，否则，则

为0，

表示在时刻t，VEC服务器给车辆k分配的计算资源比例，

表示在时刻t，VEC服务器的上行信道资源是否空闲，

表示在时刻t，VEC服务器的下行信道资源是否空闲，因此，系统的状态空间定义为：S_t＝(s₁(t),...s_k(t)...,s_K(t))；Among them, v _k (t), d _k (t), _ck (t) represent the speed, position and file size of the vehicle k at time t, respectively, and rb _VEC (t) represents the current remaining VEC server at time t. Calculate ability,

Indicates whether the vehicle k selects the unloading position (·) at time t, if so, then

is 1, otherwise, then

is 0,

represents the proportion of computing resources allocated by the VEC server to vehicle k at time t,

Indicates whether the uplink channel resources of the VEC server are idle at time t,

Indicates whether the downlink channel resources of the VEC server are idle at time t. Therefore, the state space of the system is defined as: S _t =(s ₁ (t),...s _k (t)...,s _K (t) );

For vehicle k, its action space is whether to continue to wait, whether to unload to the VEC server, the computing power allocated by the VEC, and the uplink and downlink sub-channels allocated by the VEC server, namely:

Therefore, at time t, the action space of the vehicle is: A _t ={a ₁ (t),... _ak (t)...,a _K (t)};

When the state after the action a _k (t) taken by the vehicle k does not satisfy the conditions (c1)-(c7), the reward function is:

Among them, when the condition of ( ) is not satisfied, Λ _{( )} is -1, otherwise, Λ _{( )} is 0, and l1, Γ1, Γ2, Γ3, Γ4 are experimental parameters.

When the state of vehicle k after taking action satisfies all conditions (c1)-(c4), the reward function is defined as:

r _k (t)=l ₂ +exp(Thr _k (t)-D _k (t))

Among them, l ₂ is the experimental parameter, exp( ) is the exponential function, when the state of the vehicle k after taking the action satisfies all the conditions (c1)-(c5), the reward function is:

r(t)=l ₃ +Γ ₅ ·exp(E _k (t))

Among them, l ₃ , Γ ₅ are experimental parameters.

Step 14) obtain new states, actions and rewards according to the multi-agent reinforcement learning network, and store them in the experience playback pool;

Step 15) judging whether the data in the experience playback pool reaches the threshold, when the data in the experience playback pool reaches the threshold, the multi-agent enhanced learning network is trained until the multi-agent enhanced learning network converges;

表示所有agent的策略集合，则对于agent k的确定性策略μ_k，其梯度表示为：For the centralized training process, which consists of K agents, the parameters of the multi-agent reinforcement learning network are θ={θ ₁ ,...,θ _K }, let

represents the policy set of all agents, then for the deterministic policy μ _k of agent k, its gradient is expressed as:

为经验回放区，由一系列的状态、动作以及奖励组成，即：(S,A,S',R)，

为集中式的动作-价值函数，输入为所有agents的动作和一些状态信息，输出为agent k的Q值，对于评价网络

根据Loss函数进行更新，即：in,

is the experience playback area, which consists of a series of states, actions and rewards, namely: (S, A, S', R),

is a centralized action-value function, the input is the actions of all agents and some state information, and the output is the Q value of agent k. For the evaluation network

Update according to the Loss function, ie:

where γ is the discount factor, and the action network is updated by minimizing the agent's policy gradient, namely:

Among them, X is the size of the mini-batch, and j is the index of the sample.

Step 17) Input the state of the vehicle terminal and the edge server to be allocated into the converged multi-agent reinforcement learning network, obtain the task offloading and resource allocation results, and complete the computing task offloading and resource allocation based on the multi-agent reinforcement learning vehicle speed perception distribute.

Simulation

The simulation platform is implemented in the Python3.7 environment, and the tensorflow version is 1.15.0. The detailed simulation parameter settings are shown in Table 1 and Table 2. In the experimental results, the existing computing offloading and resource allocation algorithms are AL, AV, RD and The EDG algorithm, the corresponding algorithm of the present invention is the JDEE-MADDPG algorithm.

Table 1

Table 2

parameter set value parameter set value layers 3 layer type Fully Connected number of hidden units 512 Judging the network learning rate 0.001 optimizer Adam Action network learning rate 0.0001 cycle 100000 activation function Relu Sampling pool 128 buffer size 20000

The average task completion delay of the five algorithms is compared. This experiment mainly evaluates the distribution of the average task completion delay corresponding to each algorithm in different vehicle numbers. The experimental results are shown in Figure 2, Figure 3, Figure 4 and Figure 5. As can be seen from Fig. 2, Fig. 3, Fig. 4 and Fig. 5, compared with the AL, AV and RD algorithms, the JDEE-MADDPG algorithm proposed by the present invention can always maintain a lower task completion delay for each vehicle. It is because the present invention can allocate computing resources and wireless resources to vehicles more accurately according to task priority, task size, vehicle speed and vehicle channel state. In addition, the task completion delay of some vehicles of the EDG algorithm is smaller than that proposed by the present invention. JDEE-MADDPG algorithm, because the algorithm proposed by the present invention sacrifices a little task completion delay to reduce the energy consumption of the vehicle terminal under the premise of not exceeding the task delay threshold.

Compared with other algorithms, the present invention can always maintain a lower energy consumption level, because the The invention formulates optimal offloading and resource allocation strategies according to task priority, task size, vehicle speed and vehicle channel state, and reduces task energy consumption of all vehicles as much as possible.

When the vehicle speed ranges are [30,50]Km/h and [50,80]Km/h respectively, the average task completion delay of the five algorithms is compared with the average task energy consumption. The experimental results are shown in Figure 10, Figure 11, As shown in Fig. 12 and Fig. 13, compared with the AL, AV and RD algorithms, the present invention performs better in terms of task delay and energy consumption, because the present invention can utilize the state information of the vehicle terminal and the VEC server, namely Vehicle location, vehicle speed, task queue, channel status and remaining computing resources to make optimal offloading and resource allocation strategies. In addition, the reason why the task completion delay of some vehicles in the present invention is higher than that of the EDG algorithm is that the JDEE-MADDPG algorithm proposed in the present invention can allocate more VEC servers to high-speed vehicles under the premise of not exceeding the task delay threshold. Wireless and computing resources are used to reduce the overall energy consumption on the vehicle side, thus causing the task completion delay of some vehicles to be higher than that of the EDG algorithm.

The above embodiments are only used to illustrate the technical solutions of the present invention, but not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that the foregoing embodiments can still be used for The recorded technical solutions are modified, or some or all of the technical features thereof are equivalently replaced; and these modifications or replacements do not make the essence of the corresponding technical solutions depart from the scope of the technical solutions of the embodiments of the present invention.

Claims (1)

1. a computing task unloading and resource allocation method based on multi-agent vehicle speed perception, is characterized in that, comprises the following steps: 1) Collect vehicle-side computing tasks, divide vehicle-side computing tasks into key tasks, high-priority tasks, and low-priority tasks according to the types of vehicle-side computing tasks, and then determine key tasks, high-priority tasks, and low-priority tasks according to vehicle speed. task delay threshold; 2) For key tasks, high-priority tasks and low-priority tasks, calculate different computing resources and wireless resource allocations, offload to the VEC server, execute locally, and continue to wait for the corresponding delay and energy consumption, and then calculate the corresponding delay and energy consumption of each task. Under the constraint of the delay threshold, an objective function is formed to reduce the energy consumption of the vehicle-side processing task; 3) Convert the objective function obtained in step 2) into a Markov decision process, and initialize the state space, action space and reward of the Markov decision process; 4) Obtain new states, actions and rewards according to the multi-agent reinforcement learning network, and store the new states, actions and rewards in the experience playback pool; 5) When the data in the experience playback pool reaches the threshold, train the multi-agent reinforcement learning network until the multi-agent reinforcement learning network converges; 6) Input the state of the vehicle terminal and edge server to be allocated into the multi-agent reinforcement learning network after training, get the task offloading and resource allocation results, and complete the computing task offloading and resource allocation based on multi-agent reinforcement learning for vehicle speed perception distribute; According to the bandwidth and delay requirements of the vehicle-side computing tasks, the vehicle-side computing tasks are divided into key tasks φ ₁ , high-priority tasks φ ₂ and low-priority tasks φ ₃ , key tasks φ ₁ , high-priority tasks φ ₂ and The delay thresholds corresponding to the low-priority task φ ₃ are Thr ₁ , Thr ₂ and Thr ₃ respectively. The delay threshold of the key task φ ₁ is 10ms, the delay threshold of the low-priority task φ ₃ is 100ms, and the high-priority task φ 3 has a delay threshold of 100ms. _The level task φ2 is related to the current speed of the vehicle, let

represents the task of vehicle k at time t, then at this time, the task

delay threshold

for:

Among them, v _max is the maximum speed limit of the road, and α ² is the variance of the normal function. In order to ensure that the probability of the current vehicle speed within the maximum speed limit v _max of the road exceeds 95%, then α=v _max /1.96, exp( ) is Exponential function to base e, when

hour,

when

hour,

is the speed of vehicle k at time t, and Thr ₂ is the delay threshold corresponding to the maximum speed limit v _max of the road; Transmission rate of uplink channel n assigned to vehicle k by the VEC server

for:

where σ ² is the noise power, P is the transmission power,

is the channel interference of the uplink channel n between the vehicle k and the VEC server,

is the channel gain of the uplink channel n between the vehicle k and the VEC server, the channel bandwidth

equal

is the total upstream bandwidth of the VEC server,

is the number of upstream channels of the VEC server,

is the set of uplink channels between the vehicle and the VEC, log ₂ ( ) is a logarithmic function with base 2; Assume

Indicates whether the uplink channel n between vehicle k and the VEC server is allocated to vehicle k, if so, then

is 1, otherwise, then

is 0, get the uplink transmission rate between vehicle k and VEC server

for:

Transmission rate of downlink channel n assigned to vehicle k by the VEC server

for:

where σ ² is the noise power, P is the transmission power,

is the channel interference of downlink channel n between vehicle k and VEC server,

is the channel gain of the downlink channel n between the vehicle k and the VEC server, the channel bandwidth

equal

is the total downlink bandwidth of the VEC server,

is the number of downlink channels of the VEC server,

is the set of downlink channels between the vehicle and the VEC; Assume

Indicates whether the downlink channel n between vehicle k and the VEC server is allocated to vehicle k, if so, then

is 1, otherwise, then

is 0, the downlink transmission rate between vehicle k and VEC server

for:

task of vehicle k

The total latency consumed by offloading to the VEC server for execution

for:

in,

is the round-up function,

task for vehicle k

file size,

for the task of handling vehicle k

required computational density,

Tasks downloaded for vehicle k

The size of the file is reduced relative to the original upload task,

Task given to vehicle k for the VEC server

The proportion of computing resources allocated, f ^VEC is the CPU frequency of the local VEC server,

is the uplink transmission rate between vehicle k and the VEC server,

is the downlink transmission rate between vehicle k and the VEC server; Calculate the task of vehicle k according to the CPU frequency f ^k allocated by vehicle k itself

Time spent executing locally

for:

At time t, the task of vehicle k

Can choose to continue to wait, offload to the local VEC server and execute locally, set

represents the task of vehicle k

Whether to continue waiting, when continuing to wait, then

is 1, otherwise, then

is 0, set the continuation waiting time as _Th ,

represents the task of vehicle k

Whether to offload to the local VEC server, when

When it is 1, it means the task of vehicle k

Offload to the local VEC server, for the task of vehicle k

Generate tasks from

The total delay it takes until the execution of the action is completed

for:

in,

for the task according to vehicle k

the time of generation; When the computing task of vehicle k

When offloading to the VEC server, the energy consumption of the offloading task includes the energy consumed by uploading computing tasks and the energy consumed by downloading tasks, and the energy consumption of offloading to the VEC server.

for:

When the computing task of vehicle k

When processed locally, according to the task of processing vehicle k

required energy density

The energy consumption of the task to be processed locally

for:

At time t, when the vehicle executes the unloading strategy, the energy E(t) consumed by all vehicles within the service range of the local VEC server is:

Under the condition that the task delay threshold, computing resources and wireless resources are limited, the objective function formed to reduce the energy consumption of the vehicle-side processing task is:

in,

is the vehicle set, and T is the total time of vehicle driving simulation.

Images