Edge computing-based collaborative path planning and resource allocation methods for autonomous vehicles

Abstract

This invention relates to a collaborative path planning and resource allocation method for autonomous vehicles (RVs) based on edge computing. A reinforcement learning method is designed for the path planning problem, and a multi-RV collaborative search method based on particle swarm optimization is designed for the resource allocation problem. The method considers the coupling relationship between the path planning and resource allocation variables, as well as limitations such as RV mobility, edge server distribution, time-sensitive onboard tasks, and edge server computing capabilities. By utilizing collaborative path planning among RVs to proactively balance edge server computing resources, the invention effectively resolves the contradiction between the changing task requirements of RVs and edge server resources, achieving cross-domain load balancing between the traffic flow domain and the edge computing resource domain. Compared to path planning and resource allocation algorithms that solve separately, this algorithm has significant advantages in terms of service latency and travel time for RV tasks.

Description

Technical Field

[0001] This invention relates to the field of wireless communication technology, and in particular to a method for cooperative path planning and resource allocation for unmanned vehicles based on edge computing. Background Technology

[0002] Intelligent autonomous vehicles, leveraging onboard sensors and computing power to understand their surroundings and achieve autonomous driving, are key to safer and more efficient transportation. However, due to limitations in perception range and computing power, most vehicles may not be able to handle some data-intensive tasks locally with their limited computing capabilities, hindering the optimization of traffic safety and efficiency. Multi-access edge computing (MEC) offloads computation, enabling autonomous vehicles to handle these time-sensitive and data-intensive computing tasks for emerging Internet of Vehicles (IoV) applications. However, the changing paths of autonomous vehicles force real-time changes in traffic flow, causing constantly shifting resource requirements for task execution. If the autonomous vehicle terminal cannot acquire resources in a timely manner, it can lead to system decision delays and traffic accidents. Therefore, given the growing demand for edge intelligence, real-time resource optimization in edge computing and path planning for autonomous vehicles become crucial.

[0003] To address the increasingly complex computational tasks of unloading unmanned vehicles, research has focused on edge server resource allocation strategies. However, the impact of traffic conditions on edge resource allocation has been largely neglected. Vehicles typically utilize edge computing resources passively, without leveraging their dynamic path planning to dynamically adjust and optimize these resources. Therefore, to ensure smooth traffic flow and reduce interference from unmanned vehicles competing for the same edge computing resources, an effective edge resource allocation mechanism is needed. This mechanism should dynamically adjust the allocated computing resources in real-time to meet the changing demands of unmanned vehicle path planning services. Summary of the Invention

[0004] This invention provides a method for collaborative path planning and resource allocation for autonomous vehicles based on edge computing. It involves simultaneously finding effective edge server resource allocation and rationally planning the path for the autonomous vehicle, proactively balancing the computational resource load on the vehicle's edge computing using the vehicle's path planning technology. This addresses cross-domain load balancing between the computational resource domain and the traffic flow domain, aiming to balance the service time delay and travel time delay of the autonomous vehicle, and achieving joint optimization of edge computing resource allocation and autonomous vehicle path planning.

[0005] An edge computing-based method for cooperative path planning and resource allocation for autonomous vehicles includes the following steps:

[0006] Step 1: Design constraints and establish a problem model;

[0007] Step 2: Based on the problem model established in Step 1, establish a path planning model when the resource allocation strategy is determined, and establish a resource allocation model when the autonomous vehicle selects a determined driving path.

[0008] Step 3: For the path planning model established in Step 2, use reinforcement learning to optimize the path planning of the autonomous vehicle. Based on the autonomous vehicle's own path probability vector, select the path and detect the travel time of each path to obtain the reward function. Then update the path selection probability vector in the path selection probability set space to obtain the optimal travel path.

[0009] Step 4: For the resource allocation model established in Step 2, use multi-vehicle cooperative search based on particle swarm optimization to obtain the optimal resource allocation strategy, and obtain the resource allocation strategy with the minimum service latency through online interaction and iteration.

[0010] Step 5: Combining Step 3 and Step 4, the path planning and resource allocation are jointly optimized using an alternating iterative method. This is achieved by dividing each time slot into two parts, which are then optimized separately for the path planning problem and the resource allocation problem.

[0011] Preferably, the problem model established in step 1 of this invention includes:

[0012] Communication model:

[0013] In an intelligent transportation road network deployed with edge computing infrastructure, an ensemble of driverless vehicles Roadside Unit (RSU) Collection An edge server (ES) is deployed on the RSU; the task of the autonomous vehicle i is represented by a 4-tuple. in, γ represents the input data size of driverless car i. i ξ represents the size of the onboard task of autonomous vehicle i (i.e., the number of CPU cycles required to complete the onboard computing task generated by autonomous vehicle i). i Indicates the deadline for autonomous vehicle i to complete its task, ο i This represents the ratio of task input to output for autonomous vehicle i, where the task output size of autonomous vehicle i is... Time is discretized into non-overlapping, equal-length time slots, with time intervals of... The maximum task completion deadline for the autonomous vehicle determines the number of time slots.

[0014] The number of RSUj used to provide computational offloading services for autonomous vehicles is represented as follows:

[0015]

[0016] Where, αi,j It is an indicator function that, when driverless vehicle i selects RSUj as its computational unloading target, α i,j =1, otherwise α i,j =0;

[0017] For autonomous vehicles i and RSUj, the uplink and downlink data are given by the Shannon-Harry theorem, that is, in the g-th time slot, the maximum transmission rates of the uplink and downlink are expressed as follows: and Where m = u is used for upward movement and m = d is used for downward movement, therefore...

[0018]

[0019] Where W represents bandwidth, l j σ represents the number of offload services running simultaneously on RSU j. 2 P represents noise power. t m Indicates transmission power. Let θ represent the small-scale fading channel power gain of unmanned vehicle i and RSU j in the g-th time slot, θ represent the path loss exponent, and k0 is a constant coefficient.

[0020] The uplink wireless transmission delay of the unmanned vehicle i in the g-th time slot is ,

[0021]

[0022] Since the autonomous vehicle's mission has a completion deadline, variables are introduced. Let represent the size of the downlink output data from ESj to the autonomous vehicle i in the g-th time slot, and let be the downlink wireless transmission delay in the g-th time slot.

[0023]

[0024] in The wireless transmission delay for the unmanned vehicle i.

[0025] Preferably, step 1 of the present invention, which establishes the problem model, further includes:

[0026] Computational model:

[0027] When autonomous vehicle i offloads the computational task to RSUj, the computational latency is,

[0028]

[0029] Where, γ i C represents the number of CPU cycles required to complete the task generated by driverless car i. j Indicates the computational power of RSUj, lj This represents the number of uninstall services running simultaneously on RSUj.

[0030] Preferably, step 1 of the present invention, which establishes the problem model, further includes:

[0031] Traffic model:

[0032] The set of road segments in a road network is represented by E = {e1,...,e...} l ,…,e p The set of intersections is represented by I = {I1, I2, ..., I}. q Each autonomous vehicle has its own starting point O and destination point D. Before the autonomous vehicle starts driving, the RSU first recommends an optimal driving path for each autonomous vehicle based on real-time traffic conditions and the shortest path algorithm. The driving path of autonomous vehicle i consists of an ordered continuous road segment, denoted as r. i ={e o ,e l ,…,e d}, e l Let segment e represent one of the road segments in which the autonomous vehicle i continuously travels; the speed of the autonomous vehicle on the road segment reflects the traffic conditions of the segment. l The number of driverless cars driving on the road is in, Indicates road segment e l The number of driverless cars that were backlogged in the previous time slot It is an indicator function that indicates when driverless car i selects road segment e. l At that time, indicator function otherwise, Section e l The average driving speed is:

[0033]

[0034] in, For road segment e l The free flow velocity on the surface, ψ l For road segment e l The number of driverless cars on the platform, cap l For road segment e l Maximum capacity;

[0035] Traffic speed reflects the level of traffic congestion. If the current speed exceeds a given threshold, vehicles can continue to be allocated to that road segment; segment e l weight ω l Calculated by the following formula:

[0036]

[0037] Among them, s l,maxIndicates road segment e l The maximum permissible speed; the more autonomous vehicles there are, the lower the average speed will be, and the higher the weight, the more congested the traffic segment will be; the weights of the road segments are normalized, that is,

[0038]

[0039] Based on the logit path selection model, recommended road segments are provided for autonomous vehicles.

[0040]

[0041] The recommended path set for driverless car i is r i ={e o ,e l ,…,e d}, where o and d represent the starting point and target point of autonomous vehicle i, respectively; autonomous vehicle i is on road segment e l The travel time is,

[0042]

[0043]

[0044] For driverless car i, follow path r i Travel time from the starting point to the destination.

[0045] Preferably, step 1 of the present invention, which establishes the problem model, further includes:

[0046] Optimization model:

[0047] The total travel time is the service delay time and travel time of driverless vehicle i from the starting point to the destination point.

[0048]

[0049] in, Let i be the travel time of the autonomous vehicle i from the starting point to the destination point. For the wireless transmission latency of the autonomous vehicle i, To minimize the average system cost, the computational latency of the computational task of autonomous vehicle i on RSUj is calculated.

[0050]

[0051] The optimization objective is expressed as follows:

[0052]

[0053] Preferably, the specific process of establishing the path planning and resource allocation model in step 2 of the present invention is as follows:

[0054] Given a fixed ES resource allocation strategy, optimize the path planning of the autonomous vehicle; problem P1 simplifies to:

[0055]

[0056] When the autonomous vehicle chooses a given driving path, problem P1 simplifies to:

[0057]

[0058] Preferably, the specific process of obtaining the optimal driving path in step 3 of the present invention is as follows:

[0059] 31) Initialize path selection probability

[0060] The set of probability vectors for an autonomous vehicle to choose a path is represented by: in, For driverless cars The probability vector for selecting a road segment. This indicates that the driverless car i selects road segment e. l The probability,

[0061]

[0062] 32) Update the reward function

[0063] For any autonomous vehicle i, they each detect their own travel time. The autonomous vehicle i selects a road segment e from the recommended path set. l ∈r i The reward function is denoted as Its update formula is expressed as follows:

[0064]

[0065] Among them, e i (k) is the road segment actually selected by driver i in k iterations; when e l When selected, the corresponding reward function value will be updated; otherwise, it will remain unchanged.

[0066] 33) Update the probability vector

[0067] The update formula for the probability vector is as follows:

[0068]

[0069] Where η1 and η2 are learning efficiencies, satisfying 0 < η2 < η1 < 1; e i,max This is the optimal road segment currently explored by driverless car i, denoted as .

[0070]

[0071] It is the normalized reward value of driverless car i in the kth iteration.

[0072] Preferably, the specific process of obtaining the resource allocation strategy with the minimum service latency in step 4 of the present invention is as follows:

[0073] 41) Particle swarm coding

[0074] Based on the particle swarm optimization algorithm, each autonomous vehicle acts as a particle, and the position vector of the particle is encoded into a 1·N one-dimensional vector X = (X1, X2, ..., XN). N The velocity is denoted as V = (V1, V2, ..., V). N For particle i, the current optimal position is denoted as X. O,i For the entire particle population, the current global optimal position is denoted as X. G ;

[0075] 42) Update particle velocity and position

[0076] By adding an inertial weight before the velocity to limit velocity spikes, the update formula for the PSO algorithm at any time step k is:

[0077]

[0078] Where κ1 and κ2 are acceleration factors, which are non-negative real numbers; ω is the inertia weight, whose value ranges from 0 to ω to 1; r1 and r2 are random numbers between 0 and 1;

[0079] 43) Set conditions for violating constraints

[0080] If the current Elasticsearch engine (ES) has reached its workload limit and can no longer allocate resources to provide computing services for the autonomous vehicle, a constraint violation (viol) is introduced. i If ESj is restricted from providing services to autonomous vehicle i when its workload exceeds its capacity, then viol i It is obtained through the following formula,

[0081]

[0082] Among them, C j This indicates the computational power of RSUj;

[0083] 44) Update resource allocation decisions

[0084] The latency detected by the autonomous vehicle is The corresponding ES resource allocation decision is denoted as The minimum delay time detected by driverless car i in the k-th iteration is denoted as . The corresponding optimal resource allocation decision for ES is denoted as The updated formula is as follows:

[0085]

[0086] Once all autonomous vehicles have completed their exploration, they exchange information about their current minimum latency and corresponding optimal resource allocation decisions for Elastic Compute Service (ES). After this information exchange, each autonomous vehicle will know its current global minimum latency. and corresponding optimal resource allocation decisions for Elasticsearch. They are respectively represented as,

[0087]

[0088] 45) Stopping conditions

[0089] When viol i When the value is greater than σ(k), the iteration stops, where σ(k) is a slack variable that ensures the position vector of each particle in the PSO algorithm is within the defined domain. As the number of feasible solutions increases, σ will adaptively decrease.

[0090]

[0091] Where, N fp (k) represents the total number of temporarily feasible particles in the population after the k-th iteration, and σ(0) is the average value of the total constraint violation degree of all particles at the beginning.

[0092] Preferably, step 5 of the present invention uses an alternating iterative method to jointly optimize path planning and resource allocation, and the specific process is as follows:

[0093] Each time slot is divided into two parts, and the path planning problem and resource allocation problem are optimized separately. The specific steps are as follows:

[0094] 51) When the ES resource allocation decision is determined, use step 3 to optimize the path planning of the autonomous vehicle;

[0095] 52) When the autonomous vehicle selects a certain driving path, use step 4 to optimize the ES resource allocation decision;

[0096] 53) When the time gap K > K max Stop iterating when the time is right.

[0097] This invention proposes a collaborative path planning and resource allocation method for autonomous vehicles (RVs) based on edge computing. This method combines path planning and resource allocation algorithms, considering the coupling relationship between the two variables, as well as limitations such as RV mobility, edge server distribution, time-sensitive onboard tasks, and edge server computing capabilities. By proactively balancing edge server computing resources through collaborative path planning among RVs, it effectively resolves the conflict between the changing task requirements of RVs and edge server resources, achieving cross-domain load balancing between the traffic flow domain and the edge computing resource domain. Compared to path planning and resource allocation algorithms that consider solutions separately, this algorithm demonstrates superior performance in terms of service latency and travel time for RV tasks. Attached Figure Description

[0098] Figure 1 This is a flowchart illustrating the path planning and resource allocation method of the present invention.

[0099] Figure 2 This is a flowchart illustrating the path planning method of the present invention.

[0100] Figure 3 This is a flowchart illustrating the resource allocation method of the present invention.

[0101] Figure 4 This is a flowchart illustrating the alternating iteration method of the present invention.

[0102] Figure 5 This is a schematic diagram of the time slot division model of the present invention.

[0103] Figure 6 This is a schematic diagram of the experimental simulation of the effect of different learning parameters on the path planning algorithm of the present invention.

[0104] Figure 7 This is a schematic diagram of the experimental simulation of the service delay time impact analysis of the present invention.

[0105] Figure 8 This is a schematic diagram of the experimental simulation of the impact analysis of travel time in this invention.

[0106] Figure 9 This is an experimental simulation diagram illustrating the performance analysis of task completion rate under different numbers of unmanned vehicles according to the present invention. Detailed Implementation

[0107] The technical solution of the present invention will now be described in detail with reference to the accompanying drawings:

[0108] like Figure 1 As shown, an edge computing-based method for cooperative path planning and resource allocation for autonomous vehicles includes the following steps:

[0109] Step 1: Design constraints and establish a problem model;

[0110] Step 2: Based on the problem model established in Step 1, establish a path planning model when the resource allocation strategy is determined, and establish a resource allocation model when the autonomous vehicle selects a determined driving path.

[0111] Step 3: For the path planning model established in Step 2, use reinforcement learning to optimize the path planning of the autonomous vehicle. Based on the autonomous vehicle's own path probability vector, select the path and detect the travel time of each path to obtain the reward function. Then update the path selection probability vector in the path selection probability set space to obtain the optimal travel path.

[0112] Step 4: For the resource allocation model established in Step 2, use multi-vehicle cooperative search based on particle swarm optimization to obtain the optimal resource allocation strategy, and obtain the resource allocation strategy with the minimum service latency through online interaction and iteration.

[0113] Step 5: Combining Step 3 and Step 4, the path planning and resource allocation are jointly optimized using an alternating iterative method. This is achieved by dividing each time slot into two parts, which are then optimized separately for the path planning problem and the resource allocation problem.

[0114] Step 1 of this invention, establishing the problem model, includes:

[0115] Communication model:

[0116] In an intelligent transportation road network deployed with edge computing infrastructure, an ensemble of driverless vehicles RSU collection An edge server (ES) is deployed on the RSU; the task of the autonomous vehicle i is represented by a 4-tuple. in, γ represents the input data size of driverless car i. i ξ represents the size of the onboard task of autonomous vehicle i (i.e., the number of CPU cycles required to complete the onboard computing task generated by autonomous vehicle i). i Indicates the deadline for autonomous vehicle i to complete its task, ο i This represents the ratio of task input to output for autonomous vehicle i, where the task output size of autonomous vehicle i is... Time is discretized into non-overlapping, equal-length time slots, with time intervals of... The maximum task completion deadline for the autonomous vehicle determines the number of time slots.

[0117] The number of RSUj used to provide computational offloading services for autonomous vehicles is represented as follows:

[0118]

[0119] Where, α i,j α is an indicator function that, when driverless vehicle i selects RSUj as its computational unloading target, determines... i,j =1, otherwise α i,j =0;

[0120] For autonomous vehicles i and RSUj, the uplink and downlink data are given by the Shannon-Harry theorem, that is, in the g-th time slot, the maximum transmission rates of the uplink and downlink are expressed as follows: and Where m = u is used for upward movement and m = d is used for downward movement, therefore...

[0121]

[0122] Where W represents bandwidth, l j σ represents the number of offload services running simultaneously on RSU j. 2 P represents noise power. t m Indicates transmission power. Let θ represent the small-scale fading channel power gain of autonomous vehicle i and RSU j in the g-th time slot, θ represent the path loss exponent, and k0 be a constant coefficient. p represents the position of the autonomous vehicle i in the g-th time slot. j Indicates the location of edge server j;

[0123] The uplink wireless transmission delay of the unmanned vehicle i in the g-th time slot is ,

[0124]

[0125] Since the autonomous vehicle's mission has a completion deadline, variables are introduced. Let represent the size of the downlink output data from ESj to the autonomous vehicle i in the g-th time slot, and let be the downlink wireless transmission delay in the g-th time slot.

[0126]

[0127] in The wireless transmission delay for the unmanned vehicle i.

[0128] Step 1 of this invention, which establishes the problem model, also includes:

[0129] Computational model:

[0130] When autonomous vehicle i offloads the computational task to RSUj, the computational latency is,

[0131]

[0132] Where, γi C represents the number of CPU cycles required to complete the task generated by driverless car i. j Indicates the computational power of RSUj, l j This represents the number of uninstall services running simultaneously on RSUj.

[0133] Step 1 of this invention, which establishes the problem model, also includes:

[0134] Traffic model:

[0135] The set of road segments in a road network is represented by E = {e1, ..., e}. l ,…,e p The set of intersections is represented by I = {I1, I2, ..., I}. q Each autonomous vehicle has its own starting point O and destination point D. Before the autonomous vehicle starts driving, the RSU first recommends an optimal driving path for each autonomous vehicle based on real-time traffic conditions and the shortest path algorithm. The driving path of autonomous vehicle i consists of an ordered continuous road segment, denoted as r. i ={e o ,e l ,…,e d}, e l Let segment e represent one of the road segments in which the autonomous vehicle i continuously travels; the speed of the autonomous vehicle on the road segment reflects the traffic conditions of the segment. l The number of driverless cars driving on the road is in, Indicates road segment e l The number of driverless cars that were backlogged in the previous time slot It is an indicator function that indicates when driverless car i selects road segment e. l At that time, indicator function otherwise, Section e l The average driving speed is:

[0136]

[0137] in, For road segment e l The free flow velocity on the surface, ψ l For road segment e l The number of driverless cars on the platform, cap l For road segment e l Maximum capacity;

[0138] Traffic speed reflects the level of traffic congestion. If the current speed exceeds a given threshold, vehicles can continue to be allocated to that road segment; segment e l weight ω l Calculated by the following formula:

[0139]

[0140] Among them, s l,max Indicates road segment e l The maximum permissible speed; the more autonomous vehicles there are, the lower the average speed will be, and the higher the weight, the more congested the traffic segment will be; the weights of the road segments are normalized, that is,

[0141]

[0142] Based on the logit path selection model, recommended road segments are provided for autonomous vehicles.

[0143]

[0144] The recommended path set for driverless car i is r i ={e o ,e l ,…,e d}, where o and d represent the starting point and target point of autonomous vehicle i, respectively; autonomous vehicle i is on road segment e l The travel time is,

[0145]

[0146] For driverless car i, follow path r i Travel time from the starting point to the destination.

[0147] Step 1 of this invention, which establishes the problem model, also includes:

[0148] Optimization model:

[0149] The total travel time is the service delay time and travel time of driverless vehicle i from the starting point to the destination point.

[0150]

[0151] Among them, based on formula (12) we get Let be the travel time of the unmanned vehicle i from the starting point to the target point; based on formula (5), we obtain Let be the wireless transmission delay of the unmanned vehicle i; based on formula (6), we obtain The computational latency of the computational task of autonomous vehicle i on RSUj is minimized; the average system cost is minimized.

[0152]

[0153] The optimization objective is expressed as follows:

[0154]

[0155] The specific process of establishing the path planning and resource allocation model in step 2 of this invention is as follows:

[0156] Given a fixed ES resource allocation strategy, optimize the path planning of the autonomous vehicle; problem P1 simplifies to:

[0157]

[0158] When the autonomous vehicle chooses a given driving path, problem P1 simplifies to:

[0159]

[0160] like Figure 2 As shown, the specific process of obtaining the optimal driving path in step 3 of this invention is as follows:

[0161] 31) Initialize path selection probability

[0162] The set of probability vectors for an autonomous vehicle to choose a path is represented by: in, For driverless cars The probability vector for selecting a road segment. This indicates that the driverless car i selects road segment e. l The probability,

[0163]

[0164] 32) Update the reward function

[0165] For any autonomous vehicle i, they each detect their own travel time. The autonomous vehicle i selects a road segment e from the recommended path set. l ∈r i The reward function is denoted as Its update formula is expressed as follows:

[0166]

[0167] Among them, e i (k) is the road segment actually selected by driver i in k iterations; when e l When selected, the corresponding reward function value will be updated; otherwise, it will remain unchanged.

[0168] 33) Update the probability vector

[0169] The update formula for the probability vector is as follows:

[0170]

[0171] Where η1 and η2 are learning efficiencies, satisfying 0 < η2 < η1 < 1; e i,maxThis is the optimal road segment currently explored by driverless car i, denoted as .

[0172]

[0173] It is the normalized reward value of driverless car i in the kth iteration.

[0174] like Figure 3 As shown, the specific process of obtaining the resource allocation strategy with the minimum service latency in step 4 of this invention is as follows:

[0175] 41) Particle swarm coding

[0176] Based on the particle swarm optimization algorithm, each autonomous vehicle acts as a particle, and the position vector of the particle is encoded into a 1·N one-dimensional vector X = (X1, X2, ..., XN). N The velocity is denoted as V = (V1, V2, ..., V). N For particle i, the current optimal position is denoted as X. O,i For the entire particle population, the current global optimal position is denoted as X. G ;

[0177] 42) Update particle velocity and position

[0178] By adding an inertial weight before the velocity to limit velocity spikes, the update formula for the PSO algorithm at any time step k is:

[0179]

[0180] Where κ1 and κ2 are acceleration factors, which are non-negative real numbers; ω is the inertia weight, whose value ranges from 0 to ω to 1; r1 and r2 are random numbers between 0 and 1;

[0181] 43) Set conditions for violating constraints

[0182] If the current Elasticsearch engine (ES) has reached its workload limit and can no longer allocate resources to provide computing services for the autonomous vehicle, a constraint violation (viol) is introduced. i If ESj is restricted from providing services to autonomous vehicle i when its workload exceeds its capacity, then viol i It is obtained through the following formula,

[0183]

[0184] Among them, C j This indicates the computational power of RSUj;

[0185] 44) Update resource allocation decisions

[0186] The latency detected by the autonomous vehicle is The corresponding ES resource allocation decision is denoted as The minimum delay time detected by driverless car i in the k-th iteration is denoted as . The corresponding optimal resource allocation decision for ES is denoted as The updated formula is as follows:

[0187]

[0188] Once all autonomous vehicles have completed their exploration, they exchange information about their current minimum latency and corresponding optimal resource allocation decisions for Elastic Compute Service (ES). After this information exchange, each autonomous vehicle will know its current global minimum latency. and corresponding optimal resource allocation decisions for Elasticsearch. They are respectively represented as,

[0189]

[0190] 45) Stopping conditions

[0191] When viol i When the value is greater than σ(k), the iteration stops, where σ(k) is a slack variable that ensures the position vector of each particle in the PSO algorithm is within the defined domain. As the number of feasible solutions increases, σ will adaptively decrease.

[0192]

[0193] Where, N fp (k) represents the total number of temporarily feasible particles in the population after the k-th iteration, and σ(0) is the average value of the total constraint violation degree of all particles at the beginning.

[0194] like Figure 4 As shown, step 5 of this invention utilizes an alternating iterative method to jointly optimize path planning and resource allocation. The specific process is as follows:

[0195] Each time slot is divided into two parts, and the path planning problem and resource allocation problem are optimized separately. The specific steps are as follows:

[0196] 51) When the ES resource allocation decision is determined, use step 3 to optimize the path planning of the autonomous vehicle;

[0197] 52) When the autonomous vehicle selects a certain driving path, use step 4 to optimize the ES resource allocation decision;

[0198] 53) When the time gap K > K max Stop iterating when the time is right.

[0199] Combination Figure 5 The time slot division model of the present invention will be described as follows:

[0200] Considering the operability and feasibility in actual traffic networks, the following mainly designs the time slot division in the iterative process. The time slot number is denoted as K, and one iteration of the alternating iterative algorithm is performed within each time slot. Each time slot is further divided into two parts, denoted as K1 and K2. In K1, a path planning algorithm is used to optimize the autonomous vehicle's path planning, and in K2, a resource allocation algorithm is used to optimize the resource allocation decision of the edge server. During the path planning and resource allocation processes, K1 and K2 are further divided into multiple hourly time slots, and one iteration of the path planning algorithm and the resource allocation algorithm is performed within each hourly time slot. Each hourly time slot in K1 is further divided into four parts: the first part is used to recommend a driving path for the autonomous vehicle, the second part is used to recommend a driving path for autonomous vehicle i based on its probability vector... The third part is used for the autonomous vehicles to detect their respective minimum delay times based on the selected path information, and the fourth part is used for learning, that is, updating the probability vector according to formulas (19) and (20). Similarly, each hourly slot in K2 is divided into four parts: the first part is used for population particle encoding and starting detection; the second part is used to detect the current minimum delay time of each autonomous vehicle and update the optimal position according to formula (27). The third part is used for the interaction between autonomous vehicles, updating the current global optimal position according to formula (29). Part Four calculates the next detection position of the unmanned vehicle according to formula (24).

[0201] To verify the effectiveness of the proposed algorithm, simulation experiments were conducted under different experimental scenarios. First, the impact of different learning parameter combinations on the path planning algorithm was analyzed. Then, on benchmark datasets with varying numbers of vehicles and task sizes, the proposed Alternating Iterative (AO) algorithm was compared with the Real-Time Path Replanning (RTRR) algorithm + Shortest Distance First (SDF) algorithm and the Social Vehicle Path Selection (SVRS) algorithm + Best Response Assignment (BA) algorithm in terms of travel time performance. Second, under the same experimental conditions, the proposed algorithm was compared with the Real-Time Path Replanning (RTRR) algorithm + Shortest Distance First (SDF) algorithm and the Social Vehicle Path Selection (SVRS) algorithm + Best Response Assignment algorithm in terms of service latency performance. Finally, by adjusting the task size and deadline, the task completion rate performance of the proposed algorithm under different numbers of autonomous vehicles was compared with the Real-Time Path Replanning (RTRR) algorithm + Shortest Distance First (SDF) algorithm and the Social Vehicle Path Selection (SVRS) algorithm + Best Response Assignment algorithm. The experimental platform used MATLAB running on a Windows 10 computer with an i7 processor and 16GB of RAM.

[0202] (1) The impact of different learning parameters on path planning algorithms

[0203] In path planning algorithms, the learned parameters η1 and η2 are crucial parameters affecting the algorithm's convergence. The total delay time varies with iteration under different settings of η1 and η2, as shown below. Figure 6 As shown, when η1 = 0.03 and η2 = 0.015, the algorithm converges fastest, but takes the longest time to converge. As the values ​​of η1 and η2 decrease, the convergence time decreases, but the convergence speed also decreases. Therefore, a trade-off between convergence speed and convergence accuracy should be considered when setting the values ​​of η1 and η2.

[0204] (2) Analysis of the impact of service delay time

[0205] Autonomous vehicles require high-quality service during operation; therefore, low communication and computing service latency is needed on the edge server side to ensure the quality of service for autonomous vehicles. Figure 7 The results show a comparison of average service latency under different methods. It can be seen that the alternating iterative optimization algorithm minimizes service latency for traffic scenarios with varying traffic volumes. This is because the alternating iterative optimization algorithm not only plans the optimal route for the autonomous vehicle but also rationally allocates resources to the autonomous vehicle based on the load of the edge servers. In other words, it actively balances the resource allocation of the edge servers using path planning technology, preventing resource overload. Furthermore, it can be seen that the alternating iterative optimization algorithm exhibits the slowest rate of change as traffic density and task size increase.

[0206] (3) Analysis of the impact of travel time

[0207] Driving time is a key performance indicator for evaluating the quality of autonomous vehicles (Vehicles) on real-world traffic networks. This invention primarily calculates the average driving time for all Vehicles to complete their driving tasks, such as... Figure 8 As shown, our proposed alternating iterative optimization algorithm significantly outperforms the other two combined algorithms. Under low traffic density, the results of the algorithms are similar because traffic flow is relatively evenly distributed across the road network, preventing potential congestion. However, as traffic volume increases, the alternating iterative optimization algorithm clearly outperforms the others. This is because it effectively guides traffic flow, evenly distributing autonomous vehicles across different road segments. Furthermore, under varying task sizes, the alternating iterative optimization algorithm also achieves high traffic efficiency, effectively allocating edge servers to provide computational and communication services to the autonomous vehicles.

[0208] (4) Performance analysis of task completion rate under different numbers of unmanned vehicles

[0209] To quantitatively analyze the performance of edge server resource allocation strategies, this invention compares... Figure 9 The completion rate of each algorithm is calculated under different numbers of autonomous vehicles. The task completion rate is defined as the effective returned data from the edge server divided by the total amount of output data required for the autonomous vehicle task. Among them, variables This represents the actual downlink output data size received by vehicle i in the g-th time slot. Specifically, to test the advantages of the proposed alternating iterative optimization algorithm, we... Figure 9 The distribution of task size and deadline is adjusted in the three subgraphs. In the same group of experiments, the more autonomous vehicles in the traffic network, the greater the traffic density. However, due to the limited computing resources of the edge server and the same task deadline, the edge server cannot compute more tasks compared to the case with fewer autonomous vehicles in the traffic network. Even so, under different traffic scenarios, our proposed alternating iterative optimization algorithm still outperforms the other two combined algorithms. From the results, we can conclude the importance of deadline to algorithm performance. Furthermore, it can be seen that the difference between the three comparative algorithms decreases with increasing traffic density. It can be envisioned that when the computing power of the edge server is large enough, Figure 9 The lines in the diagram will overlap, but for real-world traffic scenarios, the alternating iterative optimization algorithm can provide stable service latency and computation offloading services.

Claims

1. A method for cooperative path planning and resource allocation for unmanned vehicles based on edge computing, characterized in that... Includes the following steps: Step 1: Design constraints and establish a problem model; including: In an intelligent transportation road network with edge computing infrastructure, the wireless transmission delay of the unmanned vehicle is determined and a communication model is established based on the maximum uplink transmission rate and input data size of the unmanned vehicle, the maximum downlink transmission rate of the roadside unit, and the downlink data output size from the edge server deployed on the roadside unit to the unmanned vehicle. Establish a computational model to determine the computational delay when the autonomous vehicle offloads the computational task to the roadside unit; Based on the set of road segments and intersections in the road network, the travel time of autonomous vehicles is determined and a traffic model is established. An optimization model is established based on the total travel time of the unmanned vehicle, which consists of wireless transmission delay, computation delay, and travel time from the starting point to the target point. Step 2: Based on the problem model established in Step 1, establish a path planning model when the resource allocation strategy is determined, and establish a resource allocation model when the autonomous vehicle selects a determined driving path. Given a fixed edge server resource allocation strategy, optimize the autonomous vehicle's path planning; when the autonomous vehicle selects a fixed driving path, optimize the edge server resource allocation strategy. Step 3: For the path planning model established in Step 2, use reinforcement learning to optimize the path planning of the autonomous vehicle. Based on the autonomous vehicle's own path probability vector, select the path and detect the travel time of each path to obtain the reward function. Then update the path selection probability vector in the path selection probability set space to obtain the optimal travel path. Step 4: For the resource allocation model established in Step 2, use multi-vehicle cooperative search based on particle swarm optimization to obtain the optimal resource allocation strategy, and obtain the resource allocation strategy with the minimum service latency through online interaction and iteration. Step 5: Combining Step 3 and Step 4, the path planning and resource allocation are jointly optimized using an alternating iterative method. This is achieved by dividing each time slot into two parts, which are then optimized separately for the path planning problem and the resource allocation problem.

2. The method for cooperative path planning and resource allocation of unmanned vehicles based on edge computing according to claim 1, characterized in that... Step 1 above involves establishing the problem model, which includes: Communication model: In an intelligent transportation road network deployed with edge computing infrastructure, an ensemble of driverless vehicles Roadside Unit (RSU) Collection An edge server ES is deployed on the RSU; the task of the autonomous vehicle i is represented by a 4-tuple Φ. i ={θ i ,γ i ,ξ i ,ο i }; where θ i γ represents the input data size of driverless car i. i ξ represents the computational intensity of driverless car i. i Indicates the deadline for autonomous vehicle i to complete its task, ο i This represents the ratio of the task input to the output of autonomous vehicle i, where the task output of autonomous vehicle i is θ. i ·ο i Discretize time into non-overlapping, equal-length time slots, with time intervals of... The maximum task completion deadline for the autonomous vehicle determines the number of time slots. The number of RSUs j used to provide computational offloading services for autonomous vehicles is represented as follows: Where, α i,j It is an indicator function that, when driverless vehicle i selects RSU j as its computational unloading target, α i,j =1, otherwise α i,j =0; For driverless vehicle i and RSU j, the uplink and downlink data are given by the Shannon-Harry theorem, that is, in the g-th time slot, the maximum transmission rates of the uplink and downlink are expressed as: and Where m = u is used for upward movement and m = d is used for downward movement, therefore... Where W represents bandwidth, σ represents the number of offload services running simultaneously on RSU j. 2 P represents noise power. t m Indicates transmission power. Let θ represent the small-scale fading channel power gain of unmanned vehicle i and RSU j in the g-th time slot, θ represent the path loss exponent, and k0 is a constant coefficient. The uplink wireless transmission delay of the unmanned vehicle i in the g-th time slot is , Since the autonomous vehicle's mission has a completion deadline, variables are introduced. Let represent the size of the downlink output data from ESj to autonomous vehicle i in the g-th time slot, and let be the downlink wireless transmission delay in the g-th time slot. in The wireless transmission delay for the unmanned vehicle i.

3. The method for cooperative path planning and resource allocation of unmanned vehicles based on edge computing according to claim 2, characterized in that... Step 1 above involves establishing the problem model, which includes: Computational model: When autonomous vehicle i offloads the computational task to RSU j, the computational latency is, Where, γ i C represents the computational intensity of driverless car i. j This indicates the computational power of RSU j. This represents the number of uninstall services running simultaneously on RSUj.

4. The method for cooperative path planning and resource allocation of unmanned vehicles based on edge computing according to claim 3, characterized in that... Step 1 above involves establishing the problem model, which includes: Traffic model: The set of road segments in a road network is represented by E = {e1, ..., e}. l ,…,e p The set of intersections is represented by I = {I1, I2, ..., I}. q } represents the starting point o and the destination point d of each autonomous vehicle. Before the autonomous vehicle starts driving, the RSU first recommends an optimal driving path for each autonomous vehicle based on real-time traffic conditions and the shortest path algorithm. The driving path of autonomous vehicle i consists of an ordered continuous road segment, denoted as r. i ={e o ,e l ,…,e d The speed of an autonomous vehicle on a road segment reflects the traffic conditions of that segment; road segment e is defined as follows. l The number of driverless cars driving on the road is in, Indicates road segment e l The number of driverless cars that were backlogged in the previous time slot It is an indicator function that indicates when driverless car i selects road segment e. l At that time, indicator function otherwise, Section e l The average driving speed is: in, For road segment e l The free flow velocity on the surface, ψ l For road segment e l The number of driverless cars on the platform, cap l For road segment e l Maximum capacity; Traffic speed reflects the level of traffic congestion. If the current speed exceeds a given threshold, vehicles can continue to be allocated to that road segment; segment e l weight ω l Calculated by the following formula: Among them, s l,max Indicates road segment e l The maximum permissible speed; the more autonomous vehicles there are, the lower the average speed will be, and the higher the weight, the more congested the traffic segment will be; the weights of the road segments are normalized, that is, Based on the logit path selection model, recommended road segments are provided for autonomous vehicles. The recommended path set for driverless car i is r i ={e o ,e l ,…,e d }, where o and d represent the starting point and target point of autonomous vehicle i, respectively; autonomous vehicle i is on road segment e l The travel time is, Let be the travel time of driverless vehicle i from the starting point to the destination point.

5. The method for cooperative path planning and resource allocation of unmanned vehicles based on edge computing according to claim 4, characterized in that... Step 1 above involves establishing the problem model, which includes: Optimization model: The total travel time is the service delay time and travel time of driverless vehicle i from the starting point to the destination point. in, Let i be the travel time of the autonomous vehicle i from the starting point to the destination point. For the wireless transmission latency of the autonomous vehicle i, To minimize the average system cost, the computational latency of the computational task of autonomous vehicle i on RSU j is calculated. The optimization objective is expressed as follows:

6. The method for cooperative path planning and resource allocation of unmanned vehicles based on edge computing according to claim 5, characterized in that... The specific process of establishing the path planning and resource allocation model in step 2 above is as follows: Given a fixed ES resource allocation strategy, optimize the path planning of the autonomous vehicle; problem P1 simplifies to: When the autonomous vehicle chooses a given driving path, problem P1 simplifies to:

7. The method for cooperative path planning and resource allocation of unmanned vehicles based on edge computing according to claim 6, characterized in that... The specific process of obtaining the optimal driving path in step 3 above is as follows: 31) Initialize path selection probability The set of probability vectors for an autonomous vehicle to choose a path is represented by: in, For driverless cars The probability vector for selecting a road segment. This indicates that the driverless car i selects road segment e. l The probability, 32) Update the reward function For any autonomous vehicle i, they each detect their own travel time. The autonomous vehicle i selects a road segment e from the recommended path set. l ∈r i The reward function is denoted as Its update formula is expressed as follows: Among them, e i (k) is the road segment actually selected by driver i in the kth iteration; when e l When selected, the corresponding reward function value will be updated; otherwise, it will remain unchanged. 33) Update the probability vector The update formula for the probability vector is as follows: Where η1 and η2 are learning efficiencies, satisfying 0 < η2 < η1 < 1; e i,max This is the optimal road segment currently explored by driverless car i, denoted as . It is the normalized reward value of driverless car i in the kth iteration.

8. The method for cooperative path planning and resource allocation of unmanned vehicles based on edge computing according to claim 7, characterized in that... The specific process of obtaining the resource allocation strategy with the minimum service latency in step 4 above is as follows: 41) Particle swarm coding Based on the particle swarm optimization algorithm, each autonomous vehicle acts as a particle, and the position vector of the particle is encoded into a 1·N one-dimensional vector X = (X1, X2, ..., XN). N The velocity is denoted as V = (V1, V2, ..., V). N For particle i, the current optimal position is denoted as X. O,i For the entire particle population, the current global optimal position is denoted as X. G ; 42) Update particle velocity and position By adding an inertial weight before the velocity to limit velocity spikes, the update formula for the PSO algorithm at any time step k is: Where κ1 and κ2 are acceleration factors, which are non-negative real numbers; ω is the inertia weight, whose value ranges from 0 to ω to 1; r1 and r2 are random numbers between 0 and 1; 43) Set conditions for violating constraints If the current Elasticsearch engine (ES) has reached its workload limit and can no longer allocate resources to provide computing services for the autonomous vehicle, a constraint violation (viol) is introduced. i If ES j is restricted from providing services to autonomous vehicle i when its workload exceeds its capacity, then viol i It is obtained through the following formula, Among them, C j This indicates the computational power of RSU j; 44) Update resource allocation strategy The latency detected by the autonomous vehicle is The corresponding ES resource allocation strategy is denoted as The minimum delay time detected by driverless car i in the k-th iteration is denoted as . The corresponding optimal resource allocation strategy for Elasticsearch is denoted as . The updated formula is as follows: Once all autonomous vehicles have completed their exploration, they exchange information about their current minimum latency and corresponding optimal resource allocation strategies. After this information exchange, each autonomous vehicle will know its current global minimum latency. and corresponding optimal resource allocation strategies for Elasticsearch They are respectively represented as, 45) Stopping conditions When viol i When the value is greater than σ(k), the iteration stops, where σ(k) is a slack variable that ensures the position vector of each particle in the PSO algorithm is within the defined domain. As the number of feasible solutions increases, σ will adaptively decrease. Where, N fp (k) represents the total number of temporarily feasible particles in the population after the k-th iteration, and σ(0) is the average value of the total constraint violation degree of all particles at the beginning.

9. The method for cooperative path planning and resource allocation of unmanned vehicles based on edge computing according to claim 8, characterized in that... Step 5 above uses an alternating iterative method to jointly optimize path planning and resource allocation. The specific process is as follows: Each time slot is divided into two parts, and the path planning problem and resource allocation problem are optimized separately. The specific steps are as follows: 51) When the ES resource allocation strategy is determined, use step 3 to optimize the path planning of the autonomous vehicle; 52) When the autonomous vehicle selects a certain driving path, use step 4 to optimize the ES resource allocation strategy; 53) When the time gap K > K max Stop iterating when the time is right.

Images