An ultra-reliable and low-latency adaptive task offloading method for vehicle networks

Abstract

The present invention relates to an ultra-reliable and low-latency in-vehicle network adaptive task unloading method, which mainly includes the following steps: S1: In an in-vehicle network based on edge computing, a requesting vehicle generates a computing task, and the initial matching node of the requesting vehicle is determined based on the vehicle position and the upper limit of the RSU (Roadside Units) load; S2: When a requesting vehicle has multiple candidate matching nodes, a two-way reward and penalty selection algorithm based on node normalized reputation and interruption probability is used to determine the matching node. S3: High-reliability and low-latency adaptive task unloading of the requesting vehicle is achieved by minimizing the probability cost of the requesting vehicle. S4: Design an adaptive unloading algorithm based on probability cost minimization of time-varying channels. The beneficial effects of the present invention are: the ultra-reliable and low-latency in-vehicle network adaptive task unloading method, a method for high-reliability and low-latency adaptive task unloading of the requesting vehicle is achieved by minimizing the probability cost of the requesting vehicle.

Description

Technical Field

[0001] The present invention relates to the technical field of vehicle network adaptive task offloading, and in particular to an ultra-reliable and low-latency vehicle network adaptive task offloading method. Background Art

[0002] The Internet of Vehicles is an emerging concept in intelligent transportation systems that helps solve various traffic and driving problems, and thus plays a vital role in a safe and efficient transportation system. The development of the Internet of Things and wireless technologies has also further promoted the realization of emerging in-vehicle applications with advanced functions. However, the emergence of these computationally intensive and latency-sensitive in-vehicle applications may put a lot of pressure on vehicles with limited resources, making it difficult for vehicles to guarantee the quality of service required by the applications.

[0003] The dynamic changes in the task offloading environment caused by the high-speed movement of vehicle terminals in the vehicle network will lead to delay jitter. In addition, the existing vehicle network adaptive task offloading methods have different degrees of delays, which can easily cause the reliability of vehicle network communication to decrease, thus posing a threat to traffic safety.

[0004] Therefore, it is necessary to design an ultra-reliable and low-latency vehicle network adaptive task offloading method to address the above problems. Summary of the invention

[0005] The purpose of the present invention is to provide an ultra-reliable and low-latency vehicle network adaptive task offloading method to solve the problems raised in the above background technology.

[0006] To achieve the above object, the present invention provides the following technical solution: an ultra-reliable and low-latency vehicle network adaptive task offloading method, the method mainly comprising the following steps:

[0007] S1: In the edge computing-based Internet of Vehicles, the requesting vehicle generates a computing task, and the initial matching node of the requesting vehicle is determined based on the vehicle location and the upper limit of the RSU (Roadside Units);

[0008] S2: All service vehicles within the V2V communication range of the requesting vehicle can be used as candidates for V2V matching of the requesting vehicle. When a requesting vehicle has multiple candidate matching nodes, a two-way reward and penalty selection algorithm based on node normalized reputation and interruption probability is used to determine the matching node.

[0009] S3: In high-speed mobile Internet of Vehicles, the time-varying nature of the channel between the requesting vehicle and the service node is taken into account, and the matching decision and sub-channel allocation decision of the requesting vehicle are combined into the probabilistic cost model of the task requesting vehicle. By minimizing the probabilistic cost of the requesting vehicle, highly reliable and low-latency adaptive task offloading of the requesting vehicle is achieved.

[0010] S4: Design an adaptive offloading algorithm based on probability cost minimization over time-varying channels.

[0011] Furthermore, in step S1, there are I requesting vehicles, an RSU equipped with a MEC server and J service vehicles. The service vehicles and the MEC server are collectively referred to as service nodes. The requesting vehicle and the service node set are represented by A v ={a1,a2,…,a i} and S v ={s0,s1,s2,…,s j}, where node s0 represents the MEC server. I request vehicle tasks are matched with J+1 service nodes, and the request vehicle i unloads the task to the service node j that matches it. The matching of the request vehicle with the service node s0 is called V2I matching, and the matching with other service nodes except s0 is called V2V matching, and other service nodes except s0 can match at most one vehicle. The matching variable is defined as z ij ∈{0,1},z ij =1 indicates that the requesting vehicle i matches the service node j, otherwise, z ij = 0. In addition, In order to effectively utilize spectrum resources, the requesting vehicles for V2I matching and V2V matching work in the same frequency band, and the spectrum is divided into N sub-channels, represented as N = {1, 2, ..., N}, and the bandwidth of each sub-channel is B. Define the variable δ i,n ∈{0,1} is the subchannel allocation variable, if δ i,n =1, it means that subchannel n is assigned to the requesting vehicle i, otherwise, δ i,n =0.

[0012] Furthermore, the process of determining the initial matching node of the requesting vehicle includes:

[0013] Request the position of vehicle i with (x i ,y i ),i∈A v The coverage area of ​​RSU is represented by m r , requesting vehicle i is within the RSU coverage area, then (x i ,y i )∈m r ;otherwise The upper limit of the RSU load is in, Indicates the maximum computing capacity of RSU, C r Indicates the number of CPU cycles required for RSU to calculate a unit bit of data. Indicates the maximum tolerable delay of vehicles within the RSU coverage area. And (xi ,y i )∈m r , then z i0 =1,z ij =0,j∈S v \s0; otherwise, z i0 =0,z ij =1,j∈S v \s0, where Q r (t) represents the load of RSU at the beginning of time t.

[0014] Furthermore, in step S2, the two-way reward and penalty selection algorithm based on node normalized reputation and interruption probability includes:

[0015] 1) Determine the service capacity of candidate matching nodes: The set of candidate matching nodes for requesting vehicle i is represented as and The normalized reputation value of candidate matching node k is defined as:

[0016]

[0017] Among them, N k represents the number of times the candidate matching node k has been selected historically, f k represents the size of computing resources currently available at candidate matching node k. The interruption probability between requesting vehicle i and candidate matching node k is represented as P out (γ ik ≤γ th ), where γ th is the signal-to-interference-noise ratio threshold, and the service capacity of the candidate matching node k for the requesting vehicle i is defined as

[0018] SC ik =cre k ·(1-P out (γ ik ≤γ th ))

[0019] 2) The requesting vehicle sends a service request: select the V2V matching requesting vehicle i to send a request to the candidate matching node set N i alt The candidate matching node k with the strongest service capability s Send a service request.

[0020] 3) Alternative matching node response: The maximum tolerable delay of the requesting vehicle i for selecting V2V matching is expressed as T i max , alternative matching node k s The time required to execute the task of requesting vehicle i is expressed as like V2V requests vehicle i to pay a price To node k s Provide rewards; if Matching node k s By price Request vehicle i to pay compensation to V2V, where Represents V2V matching of vehicle i to node k s The reward bid per unit of latency saved, Represents V2V matching of vehicle i to node k s The penalty price for exceeding the delay, and The V2V matching vehicle i's demand willingness price is expressed as: like Alternative matching node k s Accept the service request of V2V matching request vehicle i, otherwise, the alternative matching node k s Reject the service request for vehicle i, where Indicates the candidate matching node k s expected price.

[0021] 4) V2V matching request vehicle determines the matching node: For the request vehicle i that selects V2V matching, if the requested candidate node k s Denial of service, then select the candidate node k with the next best service capability s' ,like Then request vehicle i to increase the demand willingness price and re-request alternative node k s service, where ε is the interruption probability threshold.

[0022] Furthermore, in step S3, the time-varying channel gain between the requesting vehicle i and the service node j is represented as h ij (t), the probability cost of requesting vehicle i is defined as

[0023]

[0024] in, and They represent the task transmission time for requesting vehicle i to perform V2I matching and V2V matching with the service node respectively. Definition The target optimization problem is modeled as:

[0025] P1:min U

[0026] stz ij ∈{0,1},i∈A v ,j∈S v

[0027]

[0028]

[0029] δ i,n ∈{0,1},i∈A v ,n∈N.

[0030] Furthermore, in step S4, the probability cost minimization adaptive offloading algorithm based on time-varying channels includes:

[0031] 1) Matching decision optimization based on deep neural network: At time slot t, the channel gain h(t) is used as the input of the deep neural network with parameter θ, based on the current matching strategy Get Slack Matching Action According to the matching action Based on the transformation probability Perform action exploration, generate L binary matching actions, substitute them into problem P1 for calculation, and select the minimum calculation result U * (h(t),z l ) corresponding to the matching action. At the same time, the channel gain and the currently obtained best matching action (h(t), z l ) is saved as a set of samples in the experience replay pool, and some samples are randomly selected from the experience replay pool regularly for deep neural network training, that is, to update the parameters θ t , get the new network parameters θ t+1 According to the new matching strategy Generate new matching actions and samples.

[0032] 2) Optimization of subchannel allocation strategy based on graph coloring model: Construct weight interference graph G = (A v ,ξ), where Represents the interference weight of each edge. Establish the channel connection matrix C and interference matrix O, and set the uncolored vertex set to Φ = A v The number of orthogonal subchannels is N, indicating that there are N available colors. First, from the vehicle set A v N vehicles are selected as vertices and N colors are used to color the selected N vertices. Specifically, the Hungarian algorithm is used to allocate orthogonal subchannels to maximize the transmission rate of N requesting vehicles. The completion of N vertex coloring indicates that N orthogonal subchannels have been allocated and N vehicles have been selected. Define the set of colored vertices as N a , then the set of uncolored vertices is updated to Φ = Φ\N a , the number of uncolored vertices is updated to |Φ| = |A v |-N. In the remaining |A v|-When coloring N vertices, the color that currently has the least interference with the requesting vehicle i is assigned to vehicle i to reduce the interference with vehicle i∈Φ. Therefore, the color with the least interference to vehicle i is obtained by searching for the highest transmission rate. According to the previous vertex coloring, update the channel connection matrix and the interference matrix. Set the corresponding variables of the assigned colors in the channel connection matrix to 1, calculate the interference to the vehicle after the channel allocation is updated, and update the corresponding variables in the interference matrix. Remove the colored vertex i from the uncolored vertex set Φ and update Φ. Check whether the uncolored vertex set Φ is empty. If so, the sub-channel allocation of all vehicles is completed and the algorithm ends. Otherwise, repeat the above process.

[0033] Compared with the prior art, the beneficial effects of the present invention are: combining the matching decision and sub-channel allocation decision of the requesting vehicle into the probabilistic cost model of the task requesting vehicle, realizing high-reliability and low-latency adaptive task offloading of the requesting vehicle by minimizing the probabilistic cost of the requesting vehicle, and at the same time, all service vehicles within the V2V communication range of the requesting vehicle can be used as alternative options for V2V matching of the requesting vehicle. When a requesting vehicle has multiple alternative matching nodes, a two-way reward and penalty selection algorithm based on node normalized reputation and interruption probability is used to determine the matching node to improve the reliability of the method. DETAILED DESCRIPTION

[0034] The present invention provides a technical solution: an ultra-reliable and low-latency vehicle network adaptive task offloading method, which mainly includes the following steps:

[0035] S1: In the edge computing-based Internet of Vehicles, the requesting vehicle generates a computing task, and the initial matching node of the requesting vehicle is determined based on the vehicle location and the upper limit of the RSU (Roadside Units);

[0036] S2: All service vehicles within the V2V communication range of the requesting vehicle can be used as candidates for V2V matching of the requesting vehicle. When a requesting vehicle has multiple candidate matching nodes, a two-way reward and penalty selection algorithm based on node normalized reputation and interruption probability is used to determine the matching node.

[0037] S3: In high-speed mobile Internet of Vehicles, the time-varying nature of the channel between the requesting vehicle and the service node is taken into account, and the matching decision and sub-channel allocation decision of the requesting vehicle are combined into the probabilistic cost model of the task requesting vehicle. By minimizing the probabilistic cost of the requesting vehicle, highly reliable and low-latency adaptive task offloading of the requesting vehicle is achieved.

[0038] S4: Design an adaptive offloading algorithm based on probability cost minimization of time-varying channels;

[0039] The specific operation is as follows: the matching decision and sub-channel allocation decision of the requesting vehicle are combined into the probabilistic cost model of the task requesting vehicle. The high-reliability and low-latency adaptive task offloading of the requesting vehicle is achieved by minimizing the probabilistic cost of the requesting vehicle. At the same time, all service vehicles within the V2V communication range of the requesting vehicle can be used as alternative options for V2V matching of the requesting vehicle. When a requesting vehicle has multiple alternative matching nodes, a two-way reward and penalty selection algorithm based on node normalized reputation and interruption probability is used to determine the matching node to improve the reliability of this method.

[0040] In step S1, there are I requesting vehicles, an RSU equipped with a MEC server and J service vehicles. The service vehicles and MEC servers are collectively referred to as service nodes. The requesting vehicles and the service node sets are represented by A and v ={a1,a2,…,a i} and S v ={s0,s1,s2,…,s j}, where node s0 represents the MEC server. I request vehicle tasks are matched with J+1 service nodes, and the request vehicle i unloads the task to the service node j that matches it. The matching of the request vehicle with the service node s0 is called V2I matching, and the matching with other service nodes except s0 is called V2V matching, and other service nodes except s0 can match at most one vehicle. The matching variable is defined as z ij ∈{0,1},z ij =1 indicates that the requesting vehicle i matches the service node j, otherwise, z ij = 0. In addition, In order to effectively utilize spectrum resources, the requesting vehicles for V2I matching and V2V matching work in the same frequency band, and the spectrum is divided into N sub-channels, represented as N = {1, 2, ..., N}, and the bandwidth of each sub-channel is B. Define the variable δ i,n ∈{0,1} is the subchannel allocation variable, if δ i,n =1, it means that subchannel n is assigned to the requesting vehicle i, otherwise, δ i,n = 0, the process of determining the initial matching node of the requesting vehicle includes: the position of the requesting vehicle i is expressed as (x i ,y i ),i∈A v The coverage area of ​​RSU is represented by m r , requesting vehicle i is within the RSU coverage area, then (x i ,y i )∈m r ;otherwise The upper limit of the RSU load is in, Indicates the maximum computing capacity of RSU, C rIndicates the number of CPU cycles required for RSU to calculate a unit bit of data. Indicates the maximum tolerable delay of vehicles within the RSU coverage area. And (x i ,y i )∈m r , then z i0 =1,z ij =0,j∈S v \s0; otherwise, z i0 =0,z ij =1,j∈S v \s0, where Q r (t) represents the load of RSU at the beginning of time t;

[0041] In step S2, the two-way reward and penalty selection algorithm based on node normalized reputation and interruption probability includes:

[0042] 1) Determine the service capacity of candidate matching nodes: The set of candidate matching nodes for requesting vehicle i is represented as and The normalized reputation value of candidate matching node k is defined as:

[0043]

[0044] Among them, N k represents the number of times the candidate matching node k has been selected historically, f k represents the size of computing resources currently available at candidate matching node k. The interruption probability between requesting vehicle i and candidate matching node k is represented as P out (γ ik ≤γ th ), where γ th is the signal-to-interference-noise ratio threshold, and the service capacity of the candidate matching node k for the requesting vehicle i is defined as

[0045] SC ik =cre k ·(1-P out (γ ik ≤γ th ))

[0046] 2) The requesting vehicle sends a service request: select the V2V matching requesting vehicle i to send a request to the candidate matching node set N i alt The candidate matching node k with the strongest service capability s Send a service request.

[0047] 3) Alternative matching node response: The maximum tolerable delay of the requesting vehicle i for selecting V2V matching is expressed as T imax , alternative matching node k s The time required to execute the task of requesting vehicle i is expressed as like V2V requests vehicle i to pay a price To node k s Provide rewards; if Matching node k s By price Request vehicle i to pay compensation to V2V, where Represents V2V matching of vehicle i to node k s The reward bid per unit of latency saved, Represents V2V matching of vehicle i to node k s The penalty price for exceeding the delay, and The V2V matching vehicle i's demand willingness price is expressed as: like Alternative matching node k s Accept the service request of V2V matching request vehicle i, otherwise, the alternative matching node k s Reject the service request of vehicle i, where Indicates the candidate matching node k s expected price.

[0048] 4) V2V matching request vehicle determines the matching node: For the request vehicle i that selects V2V matching, if the requested candidate node k s Denial of service, then select the candidate node k with the next best service capability s' ,like Then request vehicle i to increase the demand willingness price and re-request alternative node k s service, where ε is the interruption probability threshold;

[0049] In step S3, the time-varying channel gain between the requesting vehicle i and the service node j is represented as h ij (t), the probability cost of requesting vehicle i is defined as

[0050]

[0051] in, and They represent the task transmission time for requesting vehicle i to perform V2I matching and V2V matching with the service node respectively. Definition The objective optimization problem is modeled as:

[0052] P1:min U

[0053] stz ij ∈{0,1},i∈A v,j∈S v

[0054]

[0055]

[0056] δ i,n ∈{0,1},i∈A v ,n∈N;

[0057] In step S4, the probability cost minimization adaptive offloading algorithm based on time-varying channels includes:

[0058] 1) Matching decision optimization based on deep neural network: At time slot t, the channel gain h(t) is used as the input of the deep neural network with parameter θ, based on the current matching strategy Get Slack Matching Action According to the matching action Based on the transformation probability Perform action exploration, generate L binary matching actions, substitute them into problem P1 for calculation, and select the minimum calculation result U * (h(t),z l ) corresponding to the matching action. At the same time, the channel gain and the currently obtained best matching action (h(t), z l ) is saved as a set of samples in the experience replay pool, and some samples are randomly selected from the experience replay pool regularly for deep neural network training, that is, to update the parameters θ t , get the new network parameters θ t+1 According to the new matching strategy Generate new matching actions and samples.

[0059] 2) Optimization of subchannel allocation strategy based on graph coloring model: Construct weight interference graph G = (A v ,ξ), where Represents the interference weight of each edge. Establish the channel connection matrix C and interference matrix O, and set the uncolored vertex set to Φ = A v The number of orthogonal subchannels is N, indicating that there are N available colors. First, from the vehicle set A v N vehicles are selected as vertices and N colors are used to color the selected N vertices. Specifically, the Hungarian algorithm is used to allocate orthogonal subchannels to maximize the transmission rate of N requesting vehicles. The completion of N vertex coloring indicates that N orthogonal subchannels have been allocated and N vehicles have been selected. Define the set of colored vertices as N a , then the set of uncolored vertices is updated to Φ = Φ\N a , the number of uncolored vertices is updated to |Φ| = |A v |-N. In the remaining |Av |-When coloring N vertices, the color that currently has the least interference with the requesting vehicle i is assigned to vehicle i to reduce the interference with vehicle i∈Φ. Therefore, the color with the least interference to vehicle i is obtained by searching for the highest transmission rate. According to the previous vertex coloring, update the channel connection matrix and the interference matrix. Set the corresponding variables of the assigned colors in the channel connection matrix to 1, calculate the interference to the vehicle after the channel allocation is updated, and update the corresponding variables in the interference matrix. Remove the colored vertex i from the uncolored vertex set Φ and update Φ. Check whether the uncolored vertex set Φ is empty. If so, the sub-channel allocation of all vehicles is completed and the algorithm ends. Otherwise, repeat the above process.

Claims

1. An ultra-reliable and low-latency vehicle network adaptive task offloading method, characterized in that: The following steps are involved: S1: In the Internet of Vehicles based on edge computing, the requesting vehicle generates a computing task. Based on the relationship between the requesting vehicle location and the RSU coverage, and whether the current load of the RSU exceeds the computing capacity threshold, the matching strategy represented by the matching variable is used to determine the initial matching node; S2: All service vehicles within the V2V communication range of the requesting vehicle can be used as candidates for V2V matching of the requesting vehicle. When there are multiple candidate matching nodes for the requesting vehicle, the service capability is calculated based on the normalized reputation value and interruption probability of the candidate nodes, and the matching node is selected in combination with the reward and penalty mechanism; S3: In high-speed mobile vehicle networks, the time-varying channel between the requesting vehicle and the service node is considered, and the matching decision and sub-channel allocation decision of the requesting vehicle are combined into the probabilistic cost model of the task requesting vehicle. By minimizing the optimization problem of the probabilistic cost and combining the time-varying channel gain, adaptive task offloading of the requesting vehicle is achieved; S4: Design an adaptive offloading algorithm based on deep neural network optimization matching decision and graph coloring model optimization sub-channel allocation, and minimize the probability cost of time-varying channels by iteratively updating network parameters and dynamically allocating sub-channels; in, In step S1, there are I requesting vehicles, an RSU equipped with a MEC server and J service vehicles. The service vehicles and MEC servers are collectively referred to as service nodes. The requesting vehicles and the service node sets are represented by A and v ={a1,a2,…,a i } and S v ={s0,s1,s2,…,s j }, where node s0 represents the MEC server; I request vehicle tasks are matched with J+1 service nodes, and the request vehicle i unloads the task to the service node j that matches it. The matching of the request vehicle with the service node s0 is called V2I matching, and the matching with other service nodes except s0 is called V2V matching, and other service nodes except s0 can match at most one vehicle; the matching variable is defined as z ij ∈{0,1},z ij =1 indicates that the requesting vehicle i matches the service node j, otherwise, z ij =0; In addition, In order to effectively utilize spectrum resources, the requesting vehicles for V2I matching and V2V matching work in the same frequency band, and the spectrum is divided into N sub-channels, represented by N = {1, 2, ..., N}, and the bandwidth of each sub-channel is B; define the variable δ i,n ∈{0,1} is the subchannel allocation variable, if δ i,n =1, it means that subchannel n is assigned to the requesting vehicle i, otherwise, δ i,n =0; The process of determining the initial matching node for the requesting vehicle includes: Request the position of vehicle i with (x i ,y i ),i∈A v The coverage area of ​​RSU is represented by m r , requesting vehicle i is within the RSU coverage area, then (x i ,y i )∈m r ;otherwise The upper limit of the RSU load is in, Indicates the maximum computing capacity of RSU, C r Indicates the number of CPU cycles required for RSU to calculate a unit bit of data. represents the maximum tolerable delay of vehicles within the coverage area of ​​the RSU; if And (x i ,y i )∈m r , then z i0 =1,z ij =0,j∈S v \s0; otherwise, z i0 =0,z ij =1,j∈S v \s0, where Q r (t) represents the load of RSU at the beginning of time t, D i Indicates the size of the computational task requesting vehicle i to input; In step S2, the two-way reward and penalty selection algorithm based on node normalized reputation and interruption probability includes: 1) Determine the service capacity of candidate matching nodes: The set of candidate matching nodes for requesting vehicle i is represented as and The normalized reputation value of candidate matching node k is defined as: Among them, N k represents the number of times the candidate matching node k has been selected historically, f k represents the size of computing resources currently available at candidate matching node k, and the interruption probability between requesting vehicle i and candidate matching node k is expressed as P out (γ ik ≤γ th ), where γ th is the signal-to-interference-noise ratio threshold, and the service capacity of the candidate matching node k for the requesting vehicle i is defined as SC ik =cre k ·(1-P out (γ ik ≤γ th )) 2) The requesting vehicle sends a service request: select the V2V matching requesting vehicle i to send a request to the candidate matching node set N i alt The candidate matching node k with the strongest service capability s Send a service request; 3) Alternative matching node response: The maximum tolerable delay of the requesting vehicle i for selecting V2V matching is expressed as T i max , alternative matching node k s The time required to execute the task of requesting vehicle i is expressed as like V2V requests vehicle i to pay a price To node k s Provide rewards; if Matching node k s By price Request vehicle i to pay compensation to V2V, where Represents V2V matching of vehicle i to node k s The reward bid per unit of latency saved, Represents V2V matching of vehicle i to node k s The penalty price for exceeding the delay, and The V2V matching vehicle i's demand willingness price is expressed as: like Alternative matching node k s Accept the service request of V2V matching request vehicle i, otherwise, the alternative matching node k s Reject the service request of vehicle i, where Indicates the candidate matching node k s expected price; 4) V2V matching request vehicle determines the matching node: For the request vehicle i that selects V2V matching, if the requested candidate node k s Denial of service, then select the candidate node k with the next best service capability s' ,like Then request vehicle i to increase the demand willingness price and re-request alternative node k s service, where ε is the interruption probability threshold; Indicates the candidate matching node k s The number of CPU cycles required to calculate a unit bit of data; In step S3, the time-varying channel gain between the requesting vehicle i and the service node j is represented as h ij (t), the probability cost of requesting vehicle i is defined as in, and They represent the task transmission time for requesting vehicle i to perform V2I matching and V2V matching with the service node respectively; The target optimization problem is modeled as: P1:min U s.t.z ij ∈{0,1},i∈A v ,j∈S v d i,n ∈{0,1},i∈A v ,n∈N.

2. According to claim 1, an ultra-reliable and low-latency in-vehicle network adaptive task offloading method is characterized in that: In step S4, the probability cost minimization adaptive offloading algorithm based on time-varying channels includes: 1) Matching decision optimization based on deep neural network: At time slot t, the channel gain h(t) is used as the input of the deep neural network with parameter θ, based on the current matching strategy Get Slack matching action According to the matching action Based on the transformation probability Perform action exploration, generate L binary matching actions, substitute them into problem P1 for calculation, and select the minimum calculation result U * (h(t),z l ) corresponding to the matching action; at the same time, the channel gain and the currently obtained best matching action (h(t), z l ) is saved as a set of samples in the experience replay pool, and some samples are randomly selected from the experience replay pool regularly for deep neural network training, that is, to update the parameters θ t , get the new network parameters θ t+1 ; According to the new matching strategy Generate new matching actions and samples; 2) Optimization of subchannel allocation strategy based on graph coloring model: Construct weight interference graph G = (A v ,ξ), where Represents the interference weight of each edge; establishes the channel connection matrix C and the interference matrix O, and sets the uncolored vertex set to Φ = A v ; The number of orthogonal subchannels is N, indicating that there are N available colors. First, from the vehicle set A v Select N vehicles as vertices, use N colors to color the selected N vertices; specifically use the Hungarian algorithm to allocate orthogonal subchannels to maximize the transmission rate of N requesting vehicles; the completion of N vertex coloring indicates that N orthogonal subchannels have been allocated and N vehicles have been selected; define the set of colored vertices as N a , then the set of uncolored vertices is updated to Φ = Φ\N a , the number of uncolored vertices is updated to |Φ| = |A v |-N; for the remaining |A v |-N vertices are colored, the color that currently has the least interference with the requesting vehicle i is assigned to vehicle i to reduce the interference with vehicle i∈Φ, so the color with the least interference to vehicle i is obtained by searching for the highest transmission rate; according to the previous vertex coloring, the channel connection matrix and the interference matrix are updated; the corresponding variables of the assigned colors in the channel connection matrix are set to 1, the interference to the vehicle after the channel allocation is updated is calculated and the corresponding variables in the interference matrix are updated; the colored vertex i is removed from the uncolored vertex set Φ and Φ is updated; check whether the uncolored vertex set Φ is empty. If so, the sub-channel allocation of all vehicles is completed and the algorithm ends. Otherwise, the above process is repeated.

Images