Vehicle-mounted service resource allocation and calculation unloading method based on active learning multi-objective optimization

Abstract

The invention belongs to the technical field of Internet of Vehicles, and particularly relates to a vehicle-mounted service unloading method based on active learning multi-objective optimization. The method comprises the following steps: firstly, acquiring an application scene of the Internet of Vehicles, and establishing a time delay model and an energy consumption model of a vehicle-mounted service; according to the established time delay and energy consumption model, time delay and energy consumption corresponding to the service are solved, and a multi-objective optimization problem is constructed; secondly, solving the multi-objective optimization problem by adopting a multi-objective optimization algorithm based on active learning to obtain a Pareto optimal solution set; and finally, according to the obtained Pareto optimal solution set, completing resource allocation and calculation unloading of the vehicle-mounted service. According to the method, the advantages of active learning are integrated, the number of times of needing to evaluate the target function is reduced while the Pareto optimal solution set close to the exhaustion search algorithm is obtained, and therefore the cost of obtaining the data set in the Internet of Vehicles is reduced, and meanwhile the adaptability of the algorithm is improved.

Description

Technical field:

[0001] The present invention belongs to the technical field of vehicle networking, and specifically relates to a vehicle-borne business resource allocation and calculation offloading method based on active learning multi-objective optimization. Background technology:

[0002] The Internet of Vehicles is a large-scale system network that conducts wireless communication and information exchange between vehicles and X (X: vehicles, roads, pedestrians and cloud platforms, etc.) based on the in-vehicle network, inter-vehicle network and vehicle-cloud network, in accordance with the agreed communication protocols and data interaction standards. It is an integrated network that can realize intelligent traffic management, intelligent dynamic information services and intelligent vehicle control, and is a typical application of Internet of Things technology in the field of transportation systems. Among them, a large amount of network resources will be consumed during the transmission of vehicle-borne communication services, and network resources are limited. Therefore, it is necessary to allocate network resources efficiently.

[0003] At present, a key issue in the development of vehicle-mounted services is how to deal with the contradiction between limited computing resources and transmission resources and the large number of transmission and computing requirements and strict latency requirements of vehicle-mounted applications. In response to this key issue, the document with application number "N202410860567.X" discloses "a method for dynamic adjustment of edge nodes in the Internet of Vehicles based on reinforcement learning", with the goal of maximizing reliability and minimizing energy consumption, using a multi-objective decision analysis method to convert multiple objectives into a single objective, and using an asynchronous advantage actor critic method for multi-threaded solution, which ultimately improves the reliability of the vehicle edge computing network and shortens latency. The problem with this document is that the non-convex multi-objective optimization problem is converted into a single-objective optimization problem for solution, which will result in poor solution stability. In addition, the document with application number "CN202410301427.9" discloses a "task offloading system and method based on deep reinforcement learning and vehicle fog computing framework", which uses a multi-agent deep reinforcement learning algorithm to determine the task offloading decision and the transmission power of each vehicle, which can effectively reduce the average delay and average energy consumption of the task, while minimizing the task offloading cost of the entire system. The problem with this file is that it requires a large dataset for training, and obtaining a large-scale dataset in a complex and highly dynamically changing Internet of Vehicles environment is a very costly undertaking.

[0004] The common problems of the above existing technologies are mainly the poor stability of the solution obtained after the multi-objective optimization problem is converted into a single objective and the high cost of obtaining data sets in the Internet of Vehicles. The above shortcomings make the above two methods unable to cope with the challenges of resource allocation in the highly dynamic environment of the future Internet of Vehicles. Summary of the invention:

[0005] The purpose of the present invention is to provide a method for in-vehicle business resource allocation and computation offloading based on active learning multi-objective optimization, so as to solve the problems of poor stability of the prior art, high cost of acquiring data sets in the Internet of Vehicles, and inability to cope with the difficulty of resource allocation in the highly dynamic environment of the future Internet of Vehicles.

[0006] To achieve the above purpose, the steps of the technical solution adopted by the present invention are as follows: A method for vehicle-borne business resource allocation and calculation offloading based on active learning multi-objective optimization, comprising the following steps:

[0007] Step 1: Obtain the application scenarios of the Internet of Vehicles and establish the delay model and energy consumption model under different unloading modes of the vehicle-mounted services;

[0008] Step 2: According to the established delay model and energy consumption model, the delay and energy consumption corresponding to the service are obtained, and a multi-objective optimization problem is constructed;

[0009] Step 3: Solve the multi-objective optimization problem by adopting a multi-objective optimization algorithm based on active learning to obtain a Pareto optimal solution set;

[0010] Step 4: According to the obtained Pareto optimal solution set, complete the resource allocation and computation offloading of the vehicle-mounted business.

[0011] Furthermore, in the above step 1, in the latency model and energy consumption model in the local computing mode:

[0012] When part of a task is processed locally by the vehicle, its total delay is the computation delay. The delay of task n in local computation mode is defined as

[0013]

[0014] Among them, S n represents the amount of data input to task n, C n represents the computational density of task n, Indicates the CPU frequency of the vehicle carrying task n;

[0015] The energy consumption of task n in local computing mode is

[0016]

[0017] Where β is the effective switched capacitance of the vehicle.

[0018] Furthermore, in the above step 1, in the latency model and energy consumption model in the data-based MEC computing mode:

[0019] The total delay time in this mode is:

[0020]

[0021] The total energy consumption in this mode is:

[0022]

[0023] Furthermore, in the above step 1, in the delay model and energy consumption model in the instruction-based MEC computing mode:

[0024] The total computational latency in instruction-based MEC computing mode is

[0025]

[0026] The energy consumption of the instruction-based MEC computing mode is the sum of the computing energy consumption and transmission energy consumption of the base station, that is,

[0027]

[0028] Furthermore, in step 1 above, the total delay to complete the task is the maximum of the local processing delay and the MEC server processing delay:

[0029] The total delay T n yes

[0030]

[0031] The total energy consumption of task n is

[0032]

[0033] Furthermore, in the above step 2, the average delay of one of the optimization objectives is:

[0034]

[0035] The average energy consumption of the second optimization goal is:

[0036]

[0037] Furthermore, in the above step 2, the various resource constraints in the system are:

[0038]

[0039]

[0040] Among them, f RSU Indicates the total computing resources on the RSU.

[0041] Furthermore, the multi-objective optimization algorithm based on active learning in the above step 3 has as input a set of decision variables and values ​​of some objective functions, and outputs a set of Pareto optimal solutions.

[0042] Furthermore, the computation offloading in the above step 4 refers to local computation, data-based MEC computation and instruction-based MEC computation.

[0043] Furthermore, the resource allocation in the above step 4 includes task division ratio, communication resource allocation and computing resource allocation.

[0044] Compared with the prior art, the present invention has the following beneficial effects:

[0045] 1. According to the non-convexity characteristic of the optimization problem formed in the present invention, compared with using the conversion method to convert it into a single-objective optimization problem to solve, it is directly constructed as a multi-objective optimization problem, and the method of finding the Pareto optimal solution is selected to solve it, thereby improving the stability of the solution.

[0046] 2. Aiming at the problem that the Internet of Vehicles environment is highly dynamic and the cost of obtaining accurate data sets is high, the present invention draws on the idea of ​​active learning when solving multi-objective optimization problems, and proposes a multi-objective optimization method based on active learning to solve the multi-objective optimization problem. This method can obtain a Pareto optimal solution set that is similar to the exhaustive search when a small number of accurate objective functions are required. While ensuring the accuracy of the desired Pareto optimal solution set, the number of times the objective function needs to be evaluated is reduced, thereby reducing the cost of obtaining data sets in the Internet of Vehicles, while improving the adaptability of the algorithm. The method of the present invention can effectively allocate resources in the highly dynamically changing environment of the future Internet of Vehicles. Description of the drawings:

[0047] Figure 1 A flow chart provided for the present invention;

[0048] Figure 2 A task offloading model diagram based on integrated synaesthesia and computing in the Internet of Vehicles scenario provided by the present invention;

[0049] Figure 3 A framework diagram of a multi-objective optimization algorithm based on active learning provided by the present invention;

[0050] Figure 4 A comparison chart of simulation results of the multi-objective optimization algorithm based on active learning and the exhaustive search algorithm provided by the present invention;

[0051] Figure 5 A comparison chart of the number of objective function evaluations required by the multi-objective optimization algorithm based on active learning and the exhaustive search algorithm provided by the present invention. Specific implementation method:

[0052] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0053] The framework of the present invention is as follows Figure 1 As shown, first, the application scenario of the Internet of Vehicles is obtained, and the delay model and energy consumption model of the in-vehicle service are established; then, based on the established delay and energy consumption models, the delay and energy consumption corresponding to the in-vehicle service are obtained, and a multi-objective optimization problem is constructed; the multi-objective optimization problem is solved by adopting a multi-objective optimization algorithm based on active learning to obtain a Pareto optimal solution set; finally, based on the obtained Pareto optimal solution set, the resource allocation and computational offloading of the in-vehicle service are completed.

[0054] Based on the above ideas, the present invention proposes a vehicle-borne business resource allocation and calculation offloading method based on active learning multi-objective optimization, which specifically includes the following steps:

[0055] Step 1: Obtain the application scenarios of the Internet of Vehicles and establish the latency model and energy consumption model under different unloading modes of the vehicle-mounted services.

[0056] There are three types of offloading modes, namely local computing, data-based mobile edge computing (MEC) and instruction-based MEC computing. Specifically, local computing means that the vehicle processes the vehicle services locally; data-based MEC computing means that the vehicle transmits the data required for processing the vehicle services to a nearby base station, and the base station uses the mobile edge computing server to process the vehicle services; instruction-based MEC computing means that the vehicle transmits the instructions required for processing the vehicle services to a nearby base station, and the base station uses the MEC server to process the vehicle services.

[0057] The integrated framework of synergy and computing is a system architecture that integrates communication, perception and computing capabilities. Based on this framework and combined with the fact that the MEC server on the vehicle and its nearby roadside unit in the vehicle network can perceive the environmental content of the same location from different perspectives, the content perceived on the MEC can be converted into the environmental content perceived by the vehicle through coordinate conversion operations. Therefore, only the instruction information including the vehicle location, vehicle service type, and vehicle service size needs to be uploaded, and the MEC server can complete the vehicle service, which greatly reduces the delay required to upload a large amount of video content.

[0058] See also Figure 2The vehicle networking vehicle service offloading scenario includes multiple vehicles and multiple base stations. Each base station is equipped with a MEC server with powerful computing power. Vehicles within the coverage of the base station can offload vehicle services to the base station for computing and processing. For each task, there are three possible computing modes in the considered inter-sensory computing integration framework, namely local computing, data-based MEC computing, and instruction-based MEC computing. In the inter-sensory computing integration framework, the vehicle task is defined as S n Represents the input data size of the task, C n It is expressed as the number of CPU cycles required for the task, I n The instruction size corresponding to the task, Represents the required number of CPU cycles for the coordinate transformation.

[0059] The delay model includes the data transmission delay and the vehicle-borne service calculation delay. The transmission delay can be calculated by the transmission rate and data size, while the calculation delay can be calculated by the computing power and task complexity.

[0060] The energy consumption model takes into account the energy consumption of data transmission and the computing energy consumption of processing vehicle-mounted services.

[0061] The following decision variables are included in the process of establishing the delay model and energy consumption model: data upload mode x n , when x n =1 indicates upload command, x n =0 means uploading data; task division ratio η n ∈[0,1], represents the proportion of task n offloaded to the MEC server; the bandwidth w allocated by RSU to the vehicle carrying task n n And the computing resources f allocated by the MEC server to the vehicle carrying task n n .

[0062] (I) Establish the delay model and energy consumption model of the local computing mode of the vehicle-mounted business in the Internet of Vehicles:

[0063] When part of the task is processed locally by the vehicle, its total delay is the computational delay. Let f n l represents the CPU frequency of the vehicle carrying task n. The delay of task n in local computing mode is defined as

[0064]

[0065] Among them, S n represents the amount of data input to task n, C n Represents the computational density of task n.

[0066] Similarly, the resource cost in local computing mode only involves the computational energy cost. Let β denote the effective switching capacitance of the vehicle, which depends on the chip structure, then the energy consumption of task n in local computing mode is

[0067]

[0068] (II) Establishing the latency model and energy consumption model of the vehicle-borne business in the data-based MEC computing mode in the Internet of Vehicles:

[0069] In the data-based MEC mode, the task input data perceived by the vehicle is transmitted to a nearby base station and calculated on the co-located MEC server. Therefore, the delay of this mode includes task upload delay, calculation delay, and downlink feedback delay.

[0070] The delay of transmitting data from the vehicle to the base station is

[0071]

[0072] Among them B total is the total bandwidth resource of the system, w n is the bandwidth ratio allocated by the base station to the vehicle user carrying task n, p n is the transmission power of the vehicle carrying task n, is the noise power during transmission. Then the calculation delay of the MEC server is

[0073]

[0074] where f n is the CPU frequency assigned to task n in the MEC server. Since the task result is too small, the downlink delay can be ignored. Therefore, the total delay time in this mode is

[0075]

[0076] In this process, the cost of completing the task is more complicated than the local computing mode. It consists of two parts, one is the transmission energy consumption and the other is the MEC computing energy consumption. The transmission energy consumption is:

[0077]

[0078] Calculate the energy consumption as

[0079]

[0080] Where γ is the effective switching capacitance of the RSU.

[0081] So the total energy consumption in this mode is

[0082]

[0083] (III) Establishing the delay model and energy consumption model of the vehicle-borne business in the instruction-based MEC computing mode in the Internet of Vehicles:

[0084] When the subtask is executed on the top, when the vehicle transmits the calculation instruction to the base station, the delay includes the calculation instruction transmission delay, instruction conversion delay, calculation delay and downlink transmission delay. The transmission delay when uploading the calculation instruction to the base station is

[0085]

[0086] Among them I n is the transfer instruction size of task n.

[0087] In order to make the data point of view acquired by the roadside unit consistent with the data point of view that the vehicle wants to process at this time, the environmental data perceived by the roadside unit needs to be preprocessed. This can be achieved by performing matrix operations using the vehicle coordinates contained in the calculation instructions. The conversion time is related to the size of the perception data and is expressed as

[0088]

[0089] After the calculation instruction reaches the base station, the calculation delay of the base station is

[0090]

[0091] Since the transmission time of the calculation instruction is small compared to the coordinate conversion time, the transmission time is ignored. Therefore, the total calculation delay in the instruction-based MEC calculation mode is

[0092]

[0093] The energy consumption of the instruction-based MEC computing mode is the sum of the computing energy consumption and transmission energy consumption of the base station, that is,

[0094]

[0095] Step 2: Based on the established delay and energy consumption model, obtain the delay and energy consumption corresponding to the business and construct a multi-objective optimization problem.

[0096] The present invention analyzes the delay and cost of the three modes from the perspective of the vehicle user. When the vehicle needs to process a task, the user hopes to complete the task within the required time and at the lowest cost. The delay in task completion depends on the longest delay of the subtasks in the three modes. n With the help of

[0097]

[0098] The total delay to complete the task is the maximum of the local processing delay and the MEC server processing delay. n yes

[0099]

[0100] The system energy consumption to complete task n is the sum of the energy consumption of each of the three modes. The total energy consumption of task n is

[0101]

[0102] Finally, the average delay and average energy consumption are selected as the optimization targets of the present invention, and the multi-objective optimization problem finally formed is as follows:

[0103]

[0104] Among them, f RSU Indicates the total computing resources on the RSU.

[0105] Step three: Solve the multi-objective optimization problem by adopting a multi-objective optimization algorithm based on active learning to obtain a Pareto optimal solution set.

[0106] In the multi-objective optimization algorithm: the input is a set of decision variables and the values ​​of some objective functions, and the output is a set of Pareto optimal solutions.

[0107] like Figure 3 As shown in Figure 1, the multi-objective optimization algorithm based on active learning is mainly divided into the following four steps when obtaining the Pareto optimal solution set: modeling, discarding, covering and sampling. In the tth iteration, the algorithm keeps the undecided point set U t And predict the set of points that belong to the Pareto set. In this iterative process, from U t or S t Select a sampling point from the above. The algorithm discards U t The point with a higher probability of being dominated by another point will be determined to belong to the Pareto set from U t Move to S t When all undecided points are sampled, discarded, or determined to belong to the Pareto set with high probability, the algorithm stops at a certain iteration point T. At this time, the obtained set S t This is the desired Pareto set, which is used as the decision output.

[0108] Step 4: Based on the Pareto optimal solution set obtained in step 3, complete the resource allocation and computation offloading of the vehicle-mounted service:

[0109] The resource allocation includes task split ratio, communication resource allocation and computing resource allocation. The task split ratio refers to splitting the task into two parts, one part is calculated by the vehicle locally, and the other part is calculated by the MEC server. The computing resource allocation refers to how much computing resources the MEC server allocates to the task vehicle, and the communication resource allocation refers to how much bandwidth is allocated to each task vehicle.

[0110] The computing offloading mentioned here refers to local computing, data-based MEC computing, and instruction-based MEC computing.

[0111] In order to test the method of the present invention, the experimental parameters of the present invention are set as shown in Table 2.

[0112] Table 2 Experimental parameters

[0113] parameter value parameter value Number of users 3 Bandwidth of RSU 40MHz Transmitting power of the vehicle 300mW Computing power of RSU <![CDATA[10 12 cycle / s]]> Vehicle CPU frequency <![CDATA[4x10 9 Hz]]> Task size 40-50M Uplink maximum transmission power 23dBm Instruction data size 2.5-5M Gaussian white noise power -173dBm MEC maximum calculation frequency <![CDATA[2x10 10 Hz]]>

[0114] The basic configuration of the experimental equipment is NVIDIA GeForce RTX 3090.

[0115] Figure 4 This is a simulation comparison result diagram of the multi-objective optimization algorithm based on active learning and the exhaustive search algorithm proposed in the present invention, where the Pareto frontier obtained by the exhaustive search is the theoretical optimal solution. Figure 4 As shown in Figure 2, the multi-objective optimization based on active learning has a close Pareto frontier to the exhaustive search. This shows that the multi-objective optimization algorithm based on active learning has good performance and high accuracy.

[0116] Figure 5 This is a simulation comparison of the number of objective function evaluations required by the multi-objective optimization algorithm based on active learning and the exhaustive search. Figure 5 As shown in the figure, the multi-objective optimization algorithm based on active learning has a similar Pareto frontier as the exhaustive search algorithm, but the multi-objective optimization algorithm based on active learning requires a much smaller number of objective function evaluations. This greatly reduces the difficulty of obtaining accurate data sets in a complex and dynamic Internet of Vehicles environment. Figure 4 and Figure 5 It can be concluded that the multi-objective optimization algorithm based on active learning proposed in the present invention greatly reduces the cost of obtaining data sets in a complex Internet of Vehicles environment while ensuring that the obtained Pareto optimal solution set has a high accuracy rate.

[0117] The above description is only a preferred embodiment of the present invention and is not intended to limit the protection scope of the present invention. Any equivalent structural changes made using the contents of the description and drawings of the present invention should be included in the patent protection scope of the invention.

Claims

1. A method for vehicle-borne service resource allocation and computation offloading based on active learning multi-objective optimization, characterized in that: The following steps are involved: Step 1: Obtain the application scenarios of the Internet of Vehicles and establish the delay model and energy consumption model under different unloading modes of the vehicle-mounted services; Step 2: According to the established delay model and energy consumption model, the delay and energy consumption corresponding to the service are obtained, and a multi-objective optimization problem is constructed; Step 3: Solve the multi-objective optimization problem by adopting a multi-objective optimization algorithm based on active learning to obtain a Pareto optimal solution set; Step 4: According to the obtained Pareto optimal solution set, complete the resource allocation and computation offloading of the vehicle-mounted business.

2. The method for vehicle-borne service resource allocation and computation offloading based on active learning multi-objective optimization according to claim 1, characterized in that: In the step 1, in the latency model and energy consumption model in the local computing mode: When part of a task is processed locally by the vehicle, its total delay is the computation delay. The delay of task n in local computation mode is defined as Among them, S n represents the amount of data input to task n, C n represents the computational density of task n, Indicates the CPU frequency of the vehicle carrying task n; The energy consumption of task n in local computing mode is Where β is the effective switched capacitance of the vehicle.

3. The method for vehicle-borne service resource allocation and computation offloading based on active learning multi-objective optimization according to claim 2, characterized in that: In step 1, in the latency model and energy consumption model under the data-based MEC computing mode: The total delay time in this mode is: The total energy consumption in this mode is:

4. The method for vehicle-borne service resource allocation and computation offloading based on active learning multi-objective optimization according to claim 3, characterized in that: In step 1, in the delay model and energy consumption model in the instruction-based MEC computing mode: The total computational latency in instruction-based MEC computing mode is The energy consumption of the instruction-based MEC computing mode is the sum of the computing energy consumption and transmission energy consumption of the base station, that is, 5. A vehicle-borne service unloading method based on active learning multi-objective optimization according to any one of claims 2 to 4, characterized in that: In step 2, the total delay for completing the task is the maximum of the local processing delay and the MEC server processing delay: The total delay T n yes The total energy consumption of task n is 6. The method for vehicle-borne service resource allocation and computation offloading based on active learning multi-objective optimization according to claim 5, characterized in that: In step 2, the average delay of one of the optimization objectives is: The average energy consumption of the second optimization goal is:

7. The method for vehicle-borne service resource allocation and computation offloading based on active learning multi-objective optimization according to claim 6, characterized in that: In step 2, the various resource constraints are: Among them, f RSU Indicates the total computing resources on the RSU.

8. The method for vehicle-borne service resource allocation and computation offloading based on active learning multi-objective optimization according to claim 7, characterized in that: The multi-objective optimization algorithm based on active learning in step 3 has as input a set of decision variables and values ​​of some objective functions, and outputs a set of Pareto optimal solutions.

9. The method for vehicle-borne service resource allocation and computation offloading based on active learning multi-objective optimization according to claim 8, characterized in that: The computation offloading in step 4 refers to local computation, data-based MEC computation, and instruction-based MEC computation.

10. The method for vehicle-borne service resource allocation and computation offloading based on active learning multi-objective optimization according to claim 9, characterized in that: The resource allocation in step 4 includes task division ratio, communication resource allocation and computing resource allocation.

Images