Rapid adaptive task unloading system and method based on information sharing

Abstract

The invention relates to the technical field of mobile edge computing, in particular to a fast adaptive task unloading system and method based on information sharing, a mobile layer is composed of vehicles running different computing tasks, a vehicle-mounted device has communication capacity and limited computing capacity, and when the vehicle-mounted device cannot process the computing tasks in time, the vehicle-mounted device can deal with the computing tasks in time. The vehicle-mounted equipment unloads the calculation task to an edge layer or a cloud layer for processing; the edge layer is composed of a wireless infrastructure and an edge server, is deployed at a network edge end close to a vehicle-mounted user, and provides computing power resources; the cloud layer is composed of a cloud server and located at the remote end of the network. According to a multi-arm bandit model based on context information, information sharing between vehicles is considered, information sharing between vehicles is depicted based on a kernel method, and a Gaussian process and UCB are combined to make a task unloading decision. Under the conditions of limited computing resources and small data scale, a computing task unloading decision is completed more quickly with less computing power consumption.

Description

Technical Field

[0001] The present invention relates to the technical field of mobile edge computing, and particularly relates to a fast-adaptive task offloading system and method based on information sharing. Background Art

[0002] Vehicular networking is one of the key technologies for realizing a smart city. With the improvement of communication technologies, the data traffic in vehicular networking has increased explosively, posing challenges to task processing. Cloud computing technology solves the problem of insufficient computing power of in-vehicle devices to a certain extent by providing extremely powerful computing power at a remote network end. However, due to the large amount of time consumed by data transmission, cloud computing technology cannot well solve the real-time task processing problem in vehicular networking. Mobile edge computing (MEC) reduces the task transmission time and improves the task computing efficiency by equipping edge devices with computing, storage, and communication resources at the network edge and offloading computing tasks to the edge rather than the cloud.

[0003] In MEC, the task offloading scheme is a key research issue. Due to the high mobility of vehicles, the channel resources and computing resources change rapidly, and the training method of collaborative learning is greatly affected by environmental changes, resulting in problems such as too long training time. Therefore, an algorithm that can quickly make offloading decisions in a time-varying environment needs to be designed.

[0004] One of the existing technologies is an algorithm that does not rely on learning, that is, a traditional heuristic algorithm [Z. Sun, G. Sun, Y. Liu, J. Wang, and D. Cao, "Bargain-match: A game theoretical approach for resource allocation and task offloading in vehicular edge computing networks," IEEE Transactions on Mobile Computing, vol. 23, no. 2, pp. 1655–1673, 2024.].

[0005] However, in the vehicular networking scenario, the rapid movement of vehicles causes the channel conditions and computing resources to change in real time, and the task requests of vehicles have extremely high requirements for latency. Its main technical disadvantages are as follows:

[0006] 1) Difficulty in handling dynamic environments: This type of algorithm that does not rely on learning depends on complete prior knowledge or assumption conditions to solve problems. It usually assumes that the environment and conditions of the problem are known and unchanged, and is not suitable for dealing with dynamic environment optimization problems.

[0007] 2) Large computational overhead: It is necessary to traverse the solutions in the entire search space, and in high-dimensional complex problems, the computational overhead is large.

[0008] 3) Lack of adaptability: It solves problems with the initially set knowledge, unable to adjust strategies according to the environment and lacking adaptability.

[0009] 4) Unable to discover potential laws from existing data: Its algorithm is based on existing models and formulas, and it cannot train and optimize strategies through existing data.

[0010] The second prior art is based on learning methods such as reinforcement learning and multi-agent learning [H. Peng and X. Shen, “Multi-agent reinforcement learning based resource management in mec and uav-assisted vehicular networks,” IEEE Journal on Selected Areas in Communications, vol. 39, no. 1, pp. 131–141, 2021.].

[0011] The time-varying communication environment affects the algorithm convergence speed. At the same time, due to the high algorithm complexity, a large amount of computing resources are required to support the deep learning algorithm, and a large amount of time is consumed for data inference. Specifically, its main disadvantages are as follows:

[0012] 1) High demand for computing resources: Deep learning requires a large amount of computing resources during the training phase and has extremely high requirements for computing power;

[0013] 2) Large data demand: It needs to rely on a large amount of labeled data for effective training;

[0014] 3) High computational complexity: The model contains a large number of parameters, with extremely high computational complexity. In practical applications, its computational delay is high and the response time is long;

[0015] 4) Poor environmental adaptability: In a time-varying environment, the change of environmental parameters causes the model to need to be frequently retrained, and data drift affects the performance of the model. Summary of the Invention

[0016] Taking into account the shortcomings of existing algorithms in terms of information completeness dependence and computational energy consumption, the present invention mainly solves the problem of online unloading scheduling in real-time scenarios with limited observation data and computing power. Unlike the multi-agent unloading problem, the present invention emphasizes promoting learning through information sharing between vehicles, rather than finding the optimal solution through collaborative learning. An algorithm based on the kernel method kernel is proposed, which combines the Gaussian process with the upper confidence bound (UCB) to describe information sharing between vehicles. Compared with existing methods, this method can quickly estimate real-time rewards in the case of imperfect information in real-time scenarios with limited observations, and can provide faster training time, occupy less computing power and memory, and is more suitable for real-time tasks with limited processing resources and data.

[0017] Compared with traditional task offloading methods, this kernel-based algorithm pays more attention to resource usage and computational latency issues. It mainly relies on vehicle information sharing and historical context information to obtain environmental data, so it has lower computational costs. However, this also makes this algorithm only suitable for medium and small-scale vehicle network offloading and computing problems, and cannot be applied to large-scale vehicle network environments like deep learning-based task offloading algorithms.

[0018] The complete technical solution provided by the present invention is:

[0019] The present invention discloses a fast adaptive task offloading system and method based on information sharing, a multi-armed bandit (CMAB) model based on context information, taking into account the information sharing between vehicles, and characterizing the information sharing between vehicles based on the kernel method, combining Gaussian process with UCB to make task offloading decisions. In the case of limited computing resources and small data scale, this method can complete the computing task offloading decision faster with less computing power consumption.

[0020] The fast adaptive task offloading system based on information sharing includes a mobile layer, an edge layer and a cloud layer. The mobile layer is composed of vehicles running different computing tasks. The on-board equipment has communication capabilities and limited computing capabilities. When the on-board equipment cannot process the computing tasks in time, the on-board equipment will offload the computing tasks to the edge layer or cloud layer for processing; the edge layer is composed of wireless infrastructure and edge servers, which are deployed at the edge of the network close to the on-board users and can provide stronger computing resources; the cloud layer is composed of cloud servers with powerful computing power, which have super computing power but are located at the far end of the network.

[0021] The fast adaptive task offloading method based on information sharing of the present invention takes into account the fluctuation of network environment and focuses on the problem of small and medium-scale real-time task processing under the conditions of imperfect information and computing power constraints. In this method, the vehicle user can quickly select the task offloading scheme based on the historical context information and the environmental information of other vehicles, which specifically includes the following steps:

[0022] S1. Capture the characteristics of the computing task according to the size and resource requirements of the task data, and define the computing model;

[0023] S2. Use the CMAB model to construct an efficient online parameter update planning mechanism by considering the correlation between historical context and action pairs;

[0024] S3. Use a kernel to describe the information sharing between vehicles, and design a collaborative server selection strategy based on the kernel method.

[0025] The specific solutions for the above steps are as follows:

[0026] The computing model in S1 is specifically:

[0027] For the edge server m, its computing and communication capabilities are represented by the CPU cycle f m,t and the wireless transmission rate b m,v,t The communication and computing capabilities vary with the environment and cannot be predicted in advance. The mobile layer consists of a group of vehicles, denoted as V, and the edge layer devices are denoted as M. The set of vehicles within the service range of the edge device M is V m .

[0028] Each vehicle v ∈ V periodically generates a computing task S v,t , and the characteristics of this task are characterized by the data size d v,t and the required CPU cycle C v,t . The data size is proportional to the required CPU cycle and can be expressed as (1)

[0029] C v,t = β v × d v,t (1)

[0030] β v is the proportionality factor, and different types of tasks correspond to different proportionality factors. When dealing with complex calculations, since more resources are required, the proportionality factor is high, and when dealing with model processing, the proportionality factor is low. The d of the same type of task v,t will fluctuate according to the capabilities of the sensors equipped on the vehicle.

[0031] In the description of task offloading, a v,t represents the task, and the server to which the computing task is offloaded is denoted as S v,t . Specifically, the total time for task calculation includes data transmission, task calculation, and result return. The amount of result data returned is very small, so the return time is ignored. For each task S v,t the transmission time is calculated according to the following situation, as shown in formula (2).

[0032]

[0033] In formula (2), b m,v,t and b m,c,t respectively represent the wireless transmission rate between vehicle v and edge server m, and the wired connection rate between edge server m and cloud server c; b m,v,t is expressed as formula (3):

[0034]

[0035] In the above formula, W represents the wireless bandwidth, P v is the transmission power of vehicle v, g m,v,t is the channel gain between edge server m and vehicle v, which changes dynamically with the mobility of the vehicle; for each task S v,t , the task calculation time is calculated in two cases, as shown in formula (4):

[0036]

[0037] Since the computing power of the cloud server is higher than that of the edge server, and the computing power of local devices is the lowest, with CPU cycles representing computing power, there is f c,t > f m,t > f v,t . Combining formula (2) and formula (4), the total task offloading time can be expressed as (5).

[0038]

[0039] S2 The specific method for constructing an efficient online parameter update planning mechanism is as follows:

[0040] The method designed by the present invention minimizes the average total task offloading time of each vehicle within the time length T. For each computing task a v,t , the optimization objective can be expressed as (6).

[0041]

[0042]

[0043] The constraint conditions in formula (6) ensure that each vehicle selects only one computing node to execute tasks in each time period; through the optimization objective model in (6), the solution space is determined by (2 + Q m ) T , where Q m represents the number of edge servers, and T represents the number of time periods. As the number of time periods increases, the solution space grows exponentially. At the same time, due to the dynamic environment, T v,tThe calculation is affected by the fluctuations in the resource requirements of real-time tasks and the changes in edge resources, and cannot be predicted in advance. Therefore, an online scheduling strategy is set up in this solution to achieve real-time decision-making quickly.

[0044] In the traditional multi-armed bandit problem, each action is regarded as an arm selected by the scheduler, and learning is carried out iteratively based on exploration-exploitation or UCB methods. To adapt to the dynamic vehicle environment, this method transforms the collaborative task offloading problem into a CMAB model, and establishes the scheduling strategy of each agent through the relationship between historical context information and actions. Since local vehicles can only observe task characteristics in advance, it is assumed that the context of vehicle v at time t is The reward function is used to measure the server selection gain, defined as the maximum delay caused by choosing a poor solution, minus the delay of executing the current task after selecting the offloading server. However, the maximum delay cannot be known in advance, so a constant is used instead. In the CMAB framework, the reward function F v is assumed to be a certain function of actions and contexts a parametric function of, for a given context select any action a v,t ∈ A, and the reward formula is as in (7)

[0045]

[0046] Since the environment where the vehicle is located fluctuates in real time, the edge server and cloud parameters change continuously and cannot be known by local vehicles. Therefore, G v is unknown. So, the vehicle needs to estimate the reward through the action and context pair , and the random reward y of each vehicle v,t can be expressed as (8).

[0047] y v,t = F(Z v,t ) + ε v,t (8)

[0048] where ε v,t is used to measure the deviation between the observed value and the true value caused by reasons such as hardware performance fluctuations and network delays. Thus, the objective function of the collaborative task offloading optimization problem based on CMAB can be rewritten as minimizing the expected value of the population regret, and its calculation formula is as shown in (9).

[0049]

[0050] where,

[0051] The specific method of S3 is:

[0052] The problem of edge-cloud collaborative task processing based on vehicle-to-vehicle information sharing and historical context information to be solved by the present invention is solved by using the context information shared among multiple vehicles on the basis of CMAB to promote the real-time estimation of the reward function. Specifically, the Gaussian model bG v,t (μ, k) is used to represent the distribution of the reward function G v,t of each vehicle v at time t, where μ and k represent the mean and covariance respectively.

[0053] The context information captures both the task characteristics k C and the temporal correlation k t , including four parameters: the time scale t, the data size d v,t , the required CPU cycles c v,t and the local computing power f v,t . At this time, the context information can be expressed as (10).

[0054]

[0055] For any two action-context pairs, its kernel function is designed as (11).

[0056]

[0057] When a vehicle receives a newly observed action pair, it encapsulates the data into a message and transmits it to nearby vehicles through single-hop communication. Each vehicle maintains a buffer to receive newly observed or received context-action pairs to update the model parameters. Given a historical data set containing context-action pairs, denoted as z v,t = [z v,t T , the rewards collected from the first moment to the t-th moment are: y v,t = [y v,t T , then the reward distribution can be expressed as (12).

[0058]

[0059] In the above formula, μ(z v,t ) is the mean vector, K v,t is the kernel matrix, and I is the identity matrix. At t + 1, the posterior distribution y v,t+1 can be expressed as (13).

[0060] y v,t+1 |z v,t ~ N(μ(z v,t+1 |z v,t ), σ(z v,t+1 |z v,t ​​)) (13)

[0061] Given the action-context pair z v,t+1 , update y v,t+1 's mean, covariance, and variance as follows (14),

[0062]

[0063] To simplify the notation, use (K v,t ) -1 to denote To ensure real-time computation, limit the number of elements in the historical dataset z v,t to the threshold θ v , so that |z v,t | ≤ θ v . Therefore, the kernel matrix at time t+1 can be calculated as:

[0064]

[0065] K v,t is the kernel matrix at the known time scale t, Based on this, the update method for each component of the inverse matrix (K v,t+1 ) -1 is shown in (16).

[0066]

[0067] K 11 , K 12 , K 21 , K 22 have dimensions n×n, 1×n, n×1, 1×1 respectively. Thus (k v,t+1 ) T is updated as:

[0068]

[0069] If |z v,t | > θ v , then remove the element z with the smallest time scale, i.e., z v,t ←z v,t \{z}, and update the distribution parameters according to the updated dataset z v,t using (14) as the rule.

[0070] After parameter update, the UCB method is needed to balance exploration and exploitation in CMAB. When selecting a server, given the context and the action a v,t is determined by equation (18).

[0071]

[0072] Represents exploitation, reflecting reward estimation based on cumulative knowledge; Represents exploration, capturing the uncertainty of potential rewards. In the initial stage of the algorithm, exploration dominates. As the process progresses, the confidence interval of each action gradually shrinks, and the algorithm increasingly tends to select actions that have provided higher rewards. β t Is used to balance the ratio of exploration and exploitation.

[0073] The technical effects of the present invention are:

[0074] 1) Learning and reasoning are carried out based on historical context information and vehicle information sharing. The change of environmental parameters has little impact on the algorithm convergence speed, and the algorithm is applicable to the actual application scenarios of the vehicle networking.

[0075] 2) Capturing the task characteristics and time correlation in the observed data, and designing an efficient reward function update rule based on the new context-action pairs received from other vehicles, overcoming the dependence of the algorithm on environmental information.

[0076] 3) The algorithm performs fast calculations with less computing power resources, and is particularly suitable for real-time task processing scenarios of medium and small-scale data. Specific implementation manner

[0077] The application scenario of this embodiment consists of a mobile layer, an edge layer, and a cloud layer. The edge layer consists of multiple intelligent connected vehicles. The vehicles move at high speed and collect environmental data through on-vehicle sensors. Since the local computing resources of intelligent connected vehicles are limited, an offloading decision for computing tasks, that is, action selection, is made: whether the task is processed locally or offloaded to the edge server or cloud server for processing. The intelligent connected vehicle can communicate with the edge server or cloud server. At the same time, the intelligent connected vehicles can perform single-hop communication. Since the main innovation point of this technical solution lies in the information sharing between vehicles, for better clarifying the applicability of the solution, only a single edge server and cloud server are considered in both the edge layer and the cloud layer.

[0078] The specific method is as follows:

[0079] Step 1: After the vehicle collects the surrounding environmental information through on-vehicle sensing devices, a random action selection a v,t is made, that is, the server (the vehicle itself, the edge server, or the cloud server) for processing the task is selected.

[0080] Step 2: After this action is applied to the current environment, a reward information feedback by the environment will be obtained. The reward is set as the completion time T of the task v,t。The currently observable environmental state and the selected action together form a context-action pair, which, together with the reward information, constitutes a set of sampled data. The vehicle will send this set of sampled data to surrounding neighboring vehicles via V2V communication and listen to V2V. The communication channel collects the sampled data transmitted by other vehicles.

[0081] Step 3: Based on the sampled data collected by itself and through V2V communication, an observation data set is obtained. According to this data set, the vehicle updates the relevant parameters of the Gaussian model, including the inverse matrix of the kernel matrix (K vt ) -1 , the mean μ, and the variance σ, etc.

[0082] Step 4: Given the currently observed context information, the vehicle calculates the expected reward function for each potential action according to the updated model parameters, and selects the action with the largest reward value as the action to be executed. After execution, if there are no subsequent tasks, it ends; otherwise, it returns to Step 2.

Claims

1. A fast-adaptive task offloading system based on information sharing, characterized in that, Includes mobile layer, edge layer and cloud layer; The mobile layer is composed of vehicles running different computing tasks. The on-board equipment has communication capabilities and limited computing capabilities. When the on-board equipment cannot process the computing tasks in time, the on-board equipment can offload the computing tasks to the edge layer or cloud layer for processing; The edge layer consists of wireless infrastructure and edge servers, which are deployed at the edge of the network close to the vehicle users to provide computing resources; The cloud layer is composed of cloud servers and is located at the remote end of the network.

2. A fast adaptation task offloading method based on information sharing, characterized in that The fast adaptive task offloading system based on information sharing as described in claim 1 is adopted; the multi-armed bandit CMAB model based on context information takes into account the information sharing between vehicles, and characterizes the information sharing between vehicles based on the kernel method, and combines the Gaussian process with UCB to make task offloading decisions.

3. The fast adaptation task offloading method based on information sharing according to claim 1, wherein The specific steps include: S1, captures the computing task characteristics and defines the computing model according to the size of the task data and resource requirements; S2, using the CMAB model, considering the correlation between historical context and action pairs, and constructing an efficient online parameter update planning mechanism; S3, describes the information sharing between vehicles with the kernel, and designs a collaborative server selection strategy based on the kernel method.

4. The fast adaptation task offloading method based on information sharing according to claim 3, characterized in that The calculation model in S1 is as follows: For the edge server m, its computing and communication capabilities are represented by the CPU cycle f m,t and the wireless transmission rate b m,v,t The mobile layer consists of a group of vehicles, denoted as V, the edge layer devices are denoted as M, and the set of vehicles within the service range of the edge device M is V m ; Each vehicle \(v\in V\) periodically generates a computing task \(S\). v,t The characteristics of this task are represented by the data size \(d\). v,t and the required CPU cycles \(C\). v,t The data size is proportional to the required CPU cycles, expressed as Equation (1): C v,t = β v × d v,t (1) β v is a scale factor; d for the same type of task v,t fluctuates according to the capabilities of the sensors equipped on the vehicle; Regarding the description of task offloading, a v,t represents the task, and the server to which the computing task is offloaded is denoted by S v,t Specifically, the total time for task computing includes data transmission, task computing, and result return. The amount of result data returned is very small, so the return time is negligible; for each task S v,t The transmission time is calculated according to the following conditions, as shown in formula (2); In formula (2), b m,v,t and b m,c,t respectively represent the wireless transmission rate between vehicle v and edge server m, and the wired connection rate between edge server m and cloud server c; b m,v,t is expressed as formula (3): In the above formula, W represents the wireless bandwidth, and P v is the transmission power of vehicle v, and g m,v,t is the channel gain between the edge server m and vehicle v, which changes dynamically with the mobility of the vehicle; for each task S v,t , the task computing time is calculated in two cases, as shown in formula (4): Since the computing power of the cloud server is higher than that of the edge server, and the computing power of the local device is the lowest, with the computing power represented by CPU cycles, there is f c,t > f m,t > f v,t ; Combining formula (2) and formula (4), the total task offloading time is expressed as (5):

5. The method for fast adaptation task offloading based on information sharing according to claim 4, characterized in that S2 builds an efficient online parameter update planning mechanism. The specific method is: to minimize the total average task offloading time of each vehicle within the time length T; for each computing task a v,t the optimization objective is expressed as (6): The constraint conditions in Equation (6) ensure that each vehicle selects only one computing node to execute tasks in each time period; through the optimization objective model in (6), the solution space is determined by (2 + Q m ) T , where Q m represents the number of edge servers, and T represents the number of time periods. As the number of time periods increases, the solution space grows exponentially; at the same time, due to the dynamic environment, the calculation of T v,t is affected by the fluctuations in the resource requirements of real-time tasks and the changes in edge resources, and cannot be predicted in advance; an online scheduling strategy is adopted to achieve real-time decision-making; The collaborative task offloading problem is transformed into a CMAB model, and the scheduling strategy of each agent is established through the relationship between historical context information and action pairs; Set the context of vehicle v at time t as The reward function is used to measure the server selection gain and is replaced by a constant; in the CMAB framework, the reward function F v The setting is a certain function of actions and contexts The parameter function of, for a given context Select any action a v,t The reward formula for ∈ A is as in (7): Since the environment in which the vehicle is located fluctuates in real time and the edge server and cloud parameters are constantly changing and cannot be known to the local vehicle, G v is unknown; Therefore, the vehicle needs to estimate the reward based on actions and context for The random reward y for each vehicle v,t is expressed as (8): y v,t = F(Z v,t ) + ε v,t (8) where ε v,t is used to measure the deviation between the observed value and the true value; thus, the objective function of the collaborative task offloading optimization problem based on CMAB is rewritten as minimizing the expected value of the population regret, and its calculation formula is shown in (9): Among them, 6. The method for fast adaptive task offloading based on information sharing according to claim 5, wherein The specific method of S3 is: Based on CMAB, by leveraging the context information shared among multiple vehicles, to facilitate the real-time estimation of the reward function; specifically, using the Gaussian model bG v,t (μ,k) to represent the distribution of the reward function G v,t for each vehicle v at time t, where μ and k represent the mean and covariance respectively; The context information captures the task characteristics simultaneously Time correlation k t , including four parameters: time scale t, data size d v,t , required CPU cycles c v,t and local computing power f v,t ; At this time, the context information is expressed as (10): For any two action-context pairs, its kernel function is designed as in (11); When a vehicle receives a newly observed action pair, it encapsulates the data into a message and transmits it to nearby vehicles via single-hop communication. Each vehicle maintains a buffer to receive newly observed or received context-action pairs to update the model parameters; Given a historical dataset containing context-action pairs, denote z v,t =[z v,t T , the rewards collected from the 1st moment to the t-th moment are: y v,t =[y v,t T , then the reward distribution is expressed as (12):​​ In the above formula, μ(z v,t ) is the mean vector, K v,t is the kernel matrix, and i is the identity matrix; at t + 1, the posterior distribution y v,t+1 is expressed as (13): y v,t+1 |z v,t ~N(μ(z v,t+1 |z v,t ),σ(z v,t+1 |z v,t )) (13) Given the action-context pair z v,t+1 , update y v,t+1 The mean, covariance, and variance are as follows (14): For simplicity of notation, use (K v,t ) -1 to denote To ensure real-time calculation, limit the threshold of the number of elements in the historical data set z v,t to θ v , such that |z v,t | ≤ θ v ; thus, the kernel matrix calculation at time t + 1 is as follows: K v,t is the kernel matrix of the known time scale t, Based on this, the update method for each component of the inverse matrix (K v,t+1 ) -1 is shown in (16): K 11 , K 12 , K 21 , K 22 have dimensions of n×n, 1×n, n×1, 1×1 respectively; thus (k v,t+1 ) T is updated to: If |z v,t | > θ v , then remove the element z with the smallest time scale, i.e., z v,t ← z v,t \{z}, and update the distribution parameters according to the updated data set z v,t in accordance with rule (14); After the parameter update, the UCB method is needed to balance exploration and exploitation of CMAB. When selecting a server, given the context and action a v,t is determined by Equation (18): Represents exploitation, reflecting reward estimation based on cumulative knowledge; Represents exploration, capturing the uncertainty of potential rewards; In the initial stage of the algorithm, exploration dominates. As the process progresses, the confidence interval of each action gradually shrinks, and the algorithm increasingly tends to choose actions that have provided higher rewards; β t Is used to balance the proportion of exploration and exploitation.

Images