A multi-strategy adaptive bat algorithm-based computing offloading method

Abstract

This invention discloses a computational offloading method based on a multi-strategy adaptive bat algorithm. It addresses the issues of task offloading and resource allocation in mobile IoT networks supporting multiple MECs, and proposes an edge server scheduling model to mitigate the shortcomings of edge execution models, which may lead to resource conflicts and waste. The model minimizes latency while comprehensively considering energy consumption factors to prevent excessive energy consumption due to excessive pursuit of latency. It also limits and optimizes latency and energy consumption for computationally intensive applications, making the solution more practical, more suitable for the needs of computationally intensive applications, and more efficient in resource allocation. Furthermore, a computational offloading method based on a multi-strategy adaptive bat algorithm is proposed to provide a computational offloading scheme for the model. This algorithm makes decisions on offloading selection and order to achieve rational resource allocation and better energy efficiency.

Description

Technical Field

[0001] This invention relates to methods for task offloading and resource allocation in networks, and more particularly to a computational offloading method based on a multi-strategy adaptive bat algorithm. Background Technology

[0002] With the rapid development of Internet of Things (IoT) technology and the widespread adoption of smart terminals, an increasing number of devices with computationally intensive applications will connect to the Internet©. According to predictions by International Data Corporation (IDC), it is estimated that 41.6 billion connected IoT devices will generate 79.4 ZB of data by 2025, and this figure will reach 2142 ZB by 2035. Simultaneously, the rapid development of intelligent applications such as Virtual Reality (VR), Augmented Reality (AR), and video streaming places higher demands on computing and storage resources. The tasks generated by these applications are time-sensitive and require sufficient computing resources, but the battery life of the devices generating these tasks is typically limited, resulting in lower capacity to handle resource-intensive services and impacting the Quality of Service (QoS) of these applications.

[0003] Mobile cloud computing (MCC) was once considered a solution to the aforementioned challenges and has been widely applied in related fields. MCC allows users to offload tasks to cloud computing infrastructure and then send the results back to mobile devices for further processing, thus addressing the computing power limitations of user devices. However, because MCC requires users to send large amounts of data to distant cloud servers, the MCC scheme places a significant load on the mobile internet during transmission and is also subject to bandwidth limitations from the mobile internet, resulting in high latency. Although channel bandwidth has been continuously increasing with the development of communication technology, the utilization rate of radio spectrum is approaching its theoretical optimal value. Today, the continuous influx of internet-connected terminal devices has led to a decreasing channel resource for cloud resource computing and a continuous decline in computing power, urgently requiring a new technology to break this deadlock.

[0004] To address the aforementioned issues, mobile edge computing (MEC) is used to compensate for the weaknesses of cloud computing. Unlike MCC, which uploads all traffic to the data center, MEC introduces a new concept—the edge. This refers to the network and computing resources that traverse from the data source to the data center, including servers with computing, storage, and network resources such as routers, gateways, access points (Base Stations, BSs), and base stations (Base Stations, BSs). By offloading all or part of the computing workload that would otherwise be transmitted and run on the cloud computing data center to MEC servers closer to the end user, MEC reduces the computing load and bandwidth consumption during transmission in the cloud data center, while also offering the following advantages: 1) Deploying computing resources near the user provides low-latency services; 2) Offloading tasks to the edge reduces the energy consumption of user devices; 3) The edge can obtain real-time information about users (such as location, pedestrian flow, and road conditions); 4) The edge is a distributed architecture, making it difficult to be targeted by attacks and providing privacy protection for users.

[0005] Furthermore, compared to MCC, the computing resources on edge nodes of MEC are far less than those of cloud data centers. Therefore, when MEC provides services to mobile users, how to efficiently offload the computing tasks of mobile devices to MEC nodes is a major research issue in academia, typically including complete offloading and partial offloading. Many existing studies focus on single MEC servers, lacking research on task offloading and resource allocation in multi-MEC server scenarios. In addition, current models differ somewhat from reality and cannot effectively approximate processing latency, task transmission, and energy consumption. On the other hand, some applications on mobile devices (such as multimedia streams) require real-time processing; if not processed within a specified time slot, the data becomes useless. For these applications, ensuring the long-term overall system performance is more important than improving the performance of individual tasks. Therefore, considering the overall system performance, joint optimization of computing offloading and resource allocation in multi-server, multi-user IoT networks is essential. Summary of the Invention

[0006] The technical problem to be solved by the embodiments of the present invention is to provide a computational unloading method based on a multi-strategy adaptive bat algorithm. This method can make decisions on unloading selection and unloading order to achieve rational resource allocation and better energy efficiency.

[0007] To address the aforementioned technical problems, this invention provides a computational offloading method based on a multi-strategy adaptive bat algorithm, comprising the following steps: S1: Model the computation offloading model and divide the edge computing model into three modules: local execution, computation offloading, and edge execution. Solve the latency and energy consumption of different devices in each module for processing different tasks to obtain the basic bundle function P1 of the multi-user multi-MEC server edge computing model. S2: When the energy consumption threshold is exceeded, the penalty function is used to penalize the server, increasing the cost for user devices to offload tasks to the server, thus obtaining the final constraint function P2 of the computational offloading model for multi-user multi-MEC servers. S3: The bat's position in the bat algorithm is updated using three update strategies: staying at the current edge server, moving closer to the best edge server, and moving once in any direction. Variables S and G, which vary with the number of iterations, control the probability of the bat choosing different strategies in the global and local searches. rand is a randomly generated number between 0 and 1. The global search update strategy is: ; The local search update strategy is as follows: S4: All individuals in the bat algorithm will be fully encoded. The elements of a bat individual can take values ​​from 0 to... N The dimension of the bat is consistent with the total number of tasks. Then, by encoding the constraint function P2, the fitness function of the problem is obtained, and the total system cost is calculated when all tasks are assigned to the corresponding servers. S5: Obtain the multi-user multi-MEC server computation offloading model through steps S1 and S2; obtain the multi-strategy adaptive bat algorithm that can solve the computation offloading strategy of the model through step S3; and encode the model and algorithm through step S4 to combine the model and algorithm to solve the computation offloading strategy of the multi-user multi-MEC server computation offloading model.

[0008] The fundamental bundle function P1 is defined as follows: in, v i This represents the computational unloading decision vector for task i. T i It is the total system delay. I loc (i） , I off,j (i） and I pen (i） All are indicator functions.I loc (i） Used to indicate whether a task is executed locally. I off,j (i） Indicates task i Is it in the first j One edge server n j The calculation and unloading are performed on top. I pen (i） This indicates that the task cannot be completed within the specified time under the current decision. T o This represents a delay penalty for not completing the task. 0≤E i ≤E max This represents the energy consumption constraint for processing task i; 0 ≤ P i ≤P max Indicates mobile device k i Transmission power constraint; 0≤f loc,i ≤f loc,max This represents the CPU frequency constraint for locally computed mobile devices; 0 ≤ f edg,i ≤f edg,max This indicates the CPU frequency constraint of the edge server after computation is unloaded.

[0009] The penalty function penalty is: in, I E (i) It is an indicator function that indicates whether the edge node has reached its maximum energy consumption limit. E i >E max ,if E i >E max This indicates that the energy consumption limit has been exceeded and a penalty order is required. I E (i)=1 Otherwise I E (i)=0 ; g represents the constant penalty coefficient that balances energy consumption and delay.

[0010] Wherein, the constraint function P2 is: in vi Indicates task The calculation of the unloading decision vector, T i It is the total system delay. 0≤E i Indicates task processing i Energy consumption constraints; 0≤P i ≤P max Indicates mobile device k i Transmission power constraint; 0≤f loc,i ≤f loc,max This represents the CPU frequency constraint for locally computed mobile devices; 0 ≤ f edg,i ≤f edg,max This indicates the CPU frequency constraint of the edge server after computation is unloaded.

[0011] The update strategy for the flight strategy towards the optimal solution in S3 is as follows: in v t ij This represents the bat's velocity at iteration t, where M is the inertial parameter. f ij The pulse frequency, x t ij The value represents the bat's position at iteration t, step represents the random number Lévy flew, and N represents the number of MEC servers.

[0012] The formula for calculating the total system cost when all tasks are assigned to corresponding servers in step S4 is as follows: .

[0013] Implementing the embodiments of the present invention has the following beneficial effects: 1. This invention is a computational offloading model that comprehensively measures computational latency and device energy consumption during computational offloading. Based on research into mobile edge computing, a model that is closer to reality than existing computational offloading models is proposed. The model comprehensively considers multiple data sources, including latency, energy consumption, load balancing, latency thresholds, and device data. It solves the problem of insufficient variable consideration and inconsistency with reality in existing models, thereby improving the model's similarity to real-world conditions.

[0014] 2. This invention proposes a computational offloading method based on the adaptive discrete bat algorithm. This algorithm discretizes the bat algorithm and employs unique modifications to the global and local searches through dynamic adjustment of the search update strategy and multi-strategy design, enhancing the algorithm's ability to search for optimal solutions in discrete space. Furthermore, it compensates for the premature convergence of the bat algorithm by incorporating a random flight strategy in the global search, and proposes a dimension-swapping strategy in the local search to improve local space retrieval. This algorithm is not only applicable to solving moving edge computation problems but can also be transferred to other NP-hard problems in discrete space, demonstrating broad application prospects.

[0015] 3. Experiments show that this invention can significantly reduce the total system cost of computational offloading compared to existing methods. Specifically, under baseline parameter conditions, this method achieves an improvement of over 15% compared to other methods. Figure 3 As shown. By changing the energy consumption constraint ( Figure 4 ), change the number of mobile devices ( Figure 5 ), change the number of edge servers ( Figure 6 ), observe the convergence status ( Figure 7 ) and observation running time ( Figure 8 This method was compared with computational offloading methods based on Artificial Fish Swarm Algorithm (AFSA), Bacterial Foraging Algorithm (BFA), Gene Algorithm (GA), Particle Swarm Optimization Algorithm (PSO), Local Computation (LC), and Random Offloading (RO) from multiple perspectives, and the experimental data concluded that this method has significant advantages in all of the above situations. Attached Figure Description

[0016] Figure 1 Calculate the unloading system model; Figure 2 Adaptive Discrete Bat Algorithm Flowchart; Figure 3 It compares the test results of 5 algorithms; Figure 4 As maximum energy consumption increases, the data from each algorithm increases; Figure 5 As the number of mobile devices increases, the data from each algorithm will grow. Figure 6 As the number of edge servers changes, the data for each algorithm changes; Figure 7 Convergence status of each algorithm; Figure 8 The running time of each algorithm. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings.

[0018] An embodiment of the present invention provides a computational offloading method based on a multi-strategy adaptive bat algorithm, which is implemented through the following steps.

[0019] Step 1: Model the computational unloading model. The system model is as follows: Figure 1 As shown. In the multi-user, multi-MEC server compute offloading model, each mobile device has a compute-intensive task and can offload its task to a nearby edge server via a wireless network. Meanwhile, this method considers that real-time tasks in practical applications may have time constraints; the task should be executed before the time limit threshold. Specifically, the edge computing model is divided into three modules: local execution, compute offloading, and edge execution. The latency and energy consumption of different devices processing different tasks in each module are calculated separately. The fundamental bundle function P1 of the multi-user, multi-MEC server edge computing model is obtained, as shown below: The ultimate goal of P1 is to minimize the delay of the entire problem, where v i Indicates task i The calculation of the unloading decision vector, i.e., the task i The specific edge servers are assigned. T i It is the total system delay, in the calculation T i In the formula I loc (i） , I off,j (i） and I pen (i） All of these are indicator functions. I loc (i） Used to indicate whether a task is executed locally. I off,j (i） Indicates task i Is it in the first j One edge server n j The calculation and unloading are performed on top. I pen (i）This indicates that the task cannot be completed within the specified time under the current decision. T o This represents a delay penalty for not completing the task. (After the task has been defined...) i After the decision vector, the task i Unable to meet the time limit threshold Previously executed, task i It will be considered discarded.

[0020] 0≤E i ≤E max Indicates task processing i Energy consumption constraints; 0≤P i ≤P max Indicates mobile device k i Transmission power constraint; 0≤f edg,i ≤f edg,max This represents the CPU frequency constraint for locally computed mobile devices; 0 ≤ f edg,i ≤f edg,max This indicates the CPU frequency constraint of the edge server after computation is unloaded.

[0021] Step Two: This method addresses the load balancing shortcomings in the model. It considers that without energy consumption limits on MEC servers, if a particular MEC server has higher processing efficiency than others, the algorithm will offload most computational tasks to that server. This does not meet the requirements for load balancing. Therefore, this method limits the energy consumption of each MEC server. When the energy consumption threshold is exceeded, a penalty is imposed on that server, increasing the cost for user devices to offload tasks to that server. Specifically, a penalty function `penalty` is added, representing the energy limit as a penalty for the total energy consumption after all tasks are offloaded, thus achieving load balancing. The `penalty` function is as follows: I E (i) It is an indicator function that indicates whether the edge node has reached its maximum energy consumption limit. E i >E max ,if E i > E max This indicates that the energy consumption limit has been exceeded and a penalty order is required. I E (i)=1 Otherwise IE (i)=0 ; g represents a constant penalty coefficient that balances energy consumption and latency. Adding penalty to P1 yields the final constraint function P2 for the computational offloading model of a multi-user, multi-MEC server: in v i Indicates task The calculation of the unloading decision vector, T i It is the total system delay. 0≤E i Indicates task processing i Energy consumption constraints (no longer have a maximum energy consumption limit); 0≤P i ≤P max Indicates mobile device k i Transmission power constraint; 0≤f edg,i ≤f edg,max This represents the CPU frequency constraint for locally computed mobile devices; 0 ≤ f edg,i ≤f edg,max This indicates the CPU frequency constraint of the edge server after computation is unloaded.

[0022] Step 3: For multi-MEC server computation offloading models, this paper proposes a computation offloading method based on the bat algorithm in mobile edge computing scenarios, using a multi-strategy adaptive bat algorithm to solve the computation offloading decision of the model. This method first discretizes the bat algorithm, where the values ​​of elements in each bat are integers between 0 and N. Simultaneously, the position and velocity of each bat are meaningful integers, and the values ​​of each dimension of the position will not exceed the total number of currently selectable nodes (both local and remote). One edge server is selected for unloading, for a total of N+1 nodes.

[0023] bat i In the j Each dimension in terms of frequency f ij In position x ij With speed v ij Bats fly randomly to find prey. They also automatically adjust their pulse frequency and pulse emission rate based on their proximity to prey. r ∈[0,1].

[0024] r t ijLet represent the loudness at the t-th iteration. The algorithm updates the bat's position in two ways to find the optimal strategy, as shown below.

[0025] In the global search, three update strategies were designed to update the position of individual bats: staying at the current edge server, moving closer to the best edge server, and moving once in any direction. A variable S, which varies with the number of iterations, was designed to control the probability of the bat choosing different strategies. The algorithm maximizes the search range while avoiding getting trapped in local optima when the number of iterations is low. As the number of iterations increases, the algorithm gradually shifts towards searching near the current optimal solution, thus solving the problems of premature convergence and lack of population diversity in the later stages of the bat algorithm.

[0026] S is a parameter that controls the program to execute three strategies with a certain probability based on the number of iterations, as shown below: This method considers that in a discrete environment, adjacent nodes in the same dimension are not logically similar. Therefore, while improving the original algorithm's update strategy, it adds dwell and random retrieval strategies, with different strategies having different objectives. For each virtual bat, its position in the t-th iteration is defined. x t ij The strategy for updating bat locations is as follows: .

[0027] x t ij Let represent the position of bat i in the j-th dimension at iteration t, and rand is a randomly generated number between 0 and 1.

[0028] The random search strategy aims to retrieve as many possibilities as possible during the global search, encouraging the algorithm to explore more solutions in the early stages of iteration while preventing premature convergence. The dwell strategy, which maintains the current position, aims to preserve the best possible position of the current node, allowing it to better retain its position in the early stages of iteration and preventing premature convergence. The fly-to-optimal-solution strategy is as follows.

[0029] The optimal bat position in the t-th iteration x t * and speed v t ij . β 0∈[0,1] is a uniformly distributed random number, and the pulse frequency is... f ij The definition of is: By incorporating Lévy flight random numbers and inertial parameters into the speed update strategy, the algorithm's global retrieval capability is enhanced. The bat speed update strategy is as follows: Here, M is the inertial parameter, which was inspired by the DPSO algorithm. Its purpose is to balance the bat's own speed with its speed of flight towards the optimal solution, preventing the bat from converging to a local optimum too early, and ultimately enhancing the bat's global search capability. M ∈ N [0,1] is a random vector that follows a normal distribution, and step is a Lévy flight random number.

[0030] Solution x t ij The definition of time step t is: N represents the number of MEC servers, and N+1 represents the total number of options available to the user. x t+1 ij , x t ij , v t +1 ij , N ∈ N In a discrete environment, the values ​​for each dimension of speed and solution are natural numbers, and should be less than the total number of solutions (a value of 0 for the current dimension indicates local processing).

[0031] When performing a local search, this method designs a local search scheme to address the problem that there is no logical similarity between two adjacent points in discrete space. During the local search phase, each bat randomly walks near the current global optimum to generate a new solution, i.e., it starts its local search from the global optimum. To address the problem of the discrete bat algorithm easily getting trapped in local optima, the local search strategy in the discrete case is changed from flying towards the current optimum to executing a strategy of dwelling and random flight with a certain probability, i.e., searching for surrounding strategies based on the current solution. A parameter G, similar to S, is designed as follows: G allows us to control the update scheme for particles under discrete conditions. In the case of local updates, there are three schemes: random generation, dimension swapping, and maintaining the optimal state. The update schemes are shown below: Random generation involves randomly generating values ​​for the current dimension to expand the search range based on the current optimal solution. Dimension swapping involves exchanging the current dimension value with another dimension value (only unprocessed dimensions can be swapped; if the current dimension is the last one, the best value is selected), aiming to perturb the optimal solution and expand the algorithm's search range. Maintaining the best value aims to keep the solution close to the current optimal solution.

[0032] As the iteration progresses, the pulse emission rate... It will be updated. The update process is described below: in γ m It is a constant, and γ>0 When t→∞, r t ij → r 0 ij Initial emission rate r 0 ∈[0,1].

[0033] The flowchart of the multi-strategy adaptive bat algorithm is as follows: Figure 2 As shown.

[0034] Step Four: The user-multi-MEC server edge computing model is encoded. This involves encoding parameters such as the total number of tasks generated by mobile devices, the total number of edge servers, the CPU frequency of mobile devices, the CPU frequency of edge servers, the transmission efficiency of tasks from mobile devices to edge servers, and the computational complexity of both mobile devices and edge servers. Then, all individuals in the bat algorithm are fully encoded (full encoding). The elements of each bat can take any integer from 0 to N. The dimension of the bat is the same as the total number of tasks.

[0035] Then, by encoding the constraint function P2, the fitness function of the problem is obtained. The fitness function represents the total system cost when all tasks are assigned to the corresponding servers according to the current method, and the calculation formula is: The meaning of the formula is shown in step two, constraint function P2. k Indicates the total number of tasks. T iThis represents the total system delay, and g represents the equilibrium energy consumption. E i For the current task i energy consumption E max Due to energy consumption constraints, the fitness function is the sole criterion for evaluating the quality of individual bats in the bat algorithm. The ultimate goal of the algorithm is to obtain the bat with the lowest fitness (compute the unloading strategy).

[0036] Step 5: The multi-user, multi-MEC server computational offloading model is obtained through steps one and two. Step three is used to obtain the multi-strategy adaptive bat algorithm that can solve the computational offloading strategy of the model. Step four is used to encode the model and the algorithm and combine the model and the algorithm to solve the computational offloading strategy of the multi-user, multi-MEC server computational offloading model.

[0037] The above description is merely a preferred embodiment of the present invention and should not be construed as limiting the scope of the invention. Therefore, any equivalent variations made in accordance with the claims of the present invention are still within the scope of the present invention.

Claims

1. A computational unloading method based on a multi-strategy adaptive bat algorithm, characterized in that, Includes the following steps: S1: Model the computation offloading model and divide the edge computing model into three modules: local execution, computation offloading, and edge execution. Solve the latency and energy consumption of different devices in each module for processing different tasks to obtain the basic bundle function P1 of the multi-user multi-MEC server edge computing model. S2: When the energy consumption threshold is exceeded, the penalty function is used to penalize the server, increasing the cost for user devices to offload tasks to the server, thus obtaining the final constraint function P2 of the computational offloading model for multi-user multi-MEC servers. S3: The bat position in the bat algorithm is updated using three update strategies: staying at the current edge server, moving closer to the best edge server, and moving once in any direction. Variables S and G, which vary with the number of iterations, control the probability of the bat choosing different strategies in the global and local searches. The global search update strategy is as follows: ； The local search update strategy is as follows: in, x t ij This represents the position of bat i in the j-th dimension at iteration t, and rand is a randomly generated random number between 0 and 1. The update strategy for the optimal flight strategy is as follows: in v t ij This represents the bat's velocity at iteration t, where M is the inertial parameter. f ij The pulse frequency, x t ij The value represents the bat's position at iteration t, step represents the random number Lévy flew by, and N represents the number of MEC servers; S4: All individuals in the bat algorithm will be fully encoded. The elements of a bat individual can take values ​​from 0 to... N The dimension of the bat is consistent with the total number of tasks. Then, by encoding the constraint function P2, the fitness function of the problem is obtained, and the total system cost is calculated when all tasks are assigned to the corresponding servers. S5: Obtain the multi-user multi-MEC server computation offloading model through steps S1 and S2; obtain the multi-strategy adaptive bat algorithm that can solve the computation offloading strategy of the model through step S3; and encode the model and algorithm through step S4 to combine the model and algorithm to solve the computation offloading strategy of the multi-user multi-MEC server computation offloading model.

2. The computational unloading method based on the multi-strategy adaptive bat algorithm according to claim 1, characterized in that, The fundamental bundle function P1 is defined as follows: in, v i This represents the computational unloading decision vector for task i. T i It is the total system delay. I loc (i） , I off,j (i） and I pen (i） All are indicator functions. I loc (i） Used to indicate whether a task is executed locally. I off,j (i） Indicates task i Is it in the first j One edge server n j The calculation and unloading are performed on top. I pen (i） This indicates that the task cannot be completed within the specified time under the current decision. T o This represents a delay penalty for not completing the task. 0≤E i ≤E max This represents the energy consumption constraint for processing task i; 0 ≤ P i ≤P max Indicates mobile device k i Transmission power constraint; 0≤f loc,i ≤f loc,max This represents the CPU frequency constraint for locally computed mobile devices; 0 ≤ f edg,i ≤f edg,max This indicates the CPU frequency constraint of the edge server after computation is unloaded.

3. The computational unloading method based on the multi-strategy adaptive bat algorithm according to claim 2, characterized in that, The penalty function penalty is: in, I E (i) It is an indicator function that indicates whether the edge node has reached its maximum energy consumption limit, i.e. E i >E max ,if E i > E max This indicates that the energy consumption limit has been exceeded and a penalty order is required. I E (i)=1 Otherwise I E (i)=0 ; g represents the constant penalty coefficient that balances energy consumption and delay.

4. The computational unloading method based on the multi-strategy adaptive bat algorithm according to claim 3, characterized in that, The constraint function P2 is: in, v i Indicates task The calculation of the unloading decision vector, T i It is the total system delay. 0≤E i Indicates task processing i Energy consumption constraints; 0≤P i ≤P max Indicates mobile device k i Transmission power constraint; 0≤f loc,i ≤f loc,max This represents the CPU frequency constraint for locally computed mobile devices; 0 ≤ f edg,i ≤f edg,max This indicates the CPU frequency constraint of the edge server after computation is unloaded.

5. The computational unloading method based on the multi-strategy adaptive bat algorithm according to claim 3, characterized in that, The formula for calculating the total system cost when all tasks are assigned to the corresponding servers in S4 is as follows: 。

Images