Multi-strategy fusion algorithm suitable for edge computing task scheduling

Abstract

The invention discloses a multi-strategy fusion algorithm suitable for edge computing task scheduling, and belongs to the field of edge computing and intelligent scheduling. According to the algorithm, a three-stage co-evolution mechanism of dynamic parameter regulation and control, resource perception disturbance and feedback type guide updating is constructed aiming at the problems of parameter static property, low search efficiency, non-uniform load distribution and the like in the prior art. Initial distribution is guided through a dynamic adaptation degree function of task calculation density and node load, search diversity is enhanced by adopting a disturbance kernel function of local aggregation degree perception, and a task completion state and a feedback evaluation index of delay response are introduced to optimize a scheduling behavior. According to the algorithm, multi-strategy staged collaboration is realized by fusing a collaboration evolution mode, and the resource utilization rate, the task response timeliness and the system stability of edge nodes are effectively improved. According to the method, the problems that a traditional algorithm is prone to local optimum and unbalanced in load in a heterogeneous environment are solved, and an intelligent and self-adaptive task scheduling solution is provided for an edge computing scene.

Description

Technical Field

[0001] The present invention relates to the field of edge computing and intelligent scheduling, and in particular to a multi-strategy fusion algorithm suitable for edge computing task scheduling. Background Art

[0002] With the rapid development of the Internet of Things (IoT), smart devices, and 5G technology, edge computing has become a vital complement to traditional cloud computing due to its proximity to users, low latency, and minimal bandwidth usage. In edge computing environments, task scheduling is a key mechanism connecting computing resources with business needs, and its efficiency directly impacts system response latency, resource utilization, and overall service quality.

[0003] Existing task scheduling methods mainly rely on heuristic or intelligent optimization algorithms, such as genetic algorithms, particle swarm optimization, ant colony algorithms, and whale optimization algorithms. These methods can effectively solve scheduling problems in the early stages, but they still have the following key issues:

[0004] Parameter staticity leads to weak adaptability: Most scheduling algorithms use fixed parameter settings and cannot make adaptive adjustments based on changes in task scale and fluctuations in edge node status, affecting the stability of the algorithm in complex environments.

[0005] The search behavior is simple and prone to falling into local optimal solutions: Traditional optimization algorithms often use fixed search models and lack the coordination of mutation and perturbation mechanisms, resulting in a limited exploration range of the solution space and falling into local optimal solutions.

[0006] Unbalanced load distribution and lack of feedback mechanism: Most algorithms ignore the dynamic load characteristics of nodes and fail to establish a feedback loop between task scheduling and execution status, resulting in low resource utilization and a disconnect between scheduling results and system operation status.

[0007] Although some improvement strategies in existing research have attempted to introduce multiple search operators or mutation mechanisms for fusion optimization, most of them lack the systematic design of the division of the action stages and the regulation logic between strategies. The fusion method is mostly static superposition, which fails to give full play to the synergistic advantages between strategies.

[0008] Therefore, there is an urgent need for a scheduling optimization algorithm with adaptive parameter control, disturbance-enhanced search and state-driven feedback control capabilities, which can achieve intelligent scheduling, balanced system resources and stable operating performance while ensuring the real-time performance of tasks. Summary of the Invention

[0009] The present invention aims to solve the problems of extended task scheduling, low resource utilization and poor system stability caused by parameter staticity, low search efficiency and uneven load in edge computing environments. By designing a multi-strategy fusion algorithm, combined with dynamic parameter control, disturbance-enhanced search and state feedback mechanism, adaptive scheduling optimization is achieved to improve scheduling efficiency, resource balance and system stability.

[0010] In order to solve the above technical problems, the present invention adopts the following technical solutions:

[0011] The present invention provides a multi-strategy fusion algorithm suitable for edge computing task scheduling, comprising the following steps:

[0012] Step S1: Establish a scheduling model in a heterogeneous edge computing environment, including:

[0013] S1.1 defines the heterogeneous computing characteristics of the task set and the dynamic resource status of the node set;

[0014] S1.2 Combine task calculation density and node load status to design a heuristic allocation strategy to generate an initial scheduling solution population;

[0015] Step S2: Implement multi-strategy fusion optimization iteration, including:

[0016] S2.1 Dynamic Perception Guidance Phase: Generates scheduling adaptability indicators based on task computation characteristics and node resource status, and constructs guidance vectors to adjust task allocation directions;

[0017] S2.2 Perturbation Enhancement Adjustment Stage: Dynamically calculate the perturbation weight based on the local aggregation of individuals in the population, and combine the optimal guidance direction with the adaptive perturbation term to generate mutant individuals. The mutant individuals maintain high perturbation intensity in the search dense area, enhancing the ability to escape; while the perturbation is weakened in the search sparse area, which is conducive to local convergence;

[0018] S2.3 State feedback driven phase: feedback evaluation indicators are constructed through task completion rate and delay penalty, and the population evolution direction is dynamically corrected based on historical scheduling results;

[0019] Step S3: Execute phased fusion collaborative evolution control, including:

[0020] In the early stages of iteration, the dynamic perception guidance mechanism is activated first, the perturbation-enhanced regulation intensity is enhanced in the middle stages, and the state feedback mechanism is strengthened in the later stages;

[0021] Dynamically adjust the weight coefficients of each strategy according to the iterative process to form a collaborative optimization link of structure guidance, disturbance regulation and feedback correction;

[0022] Step S4: Output the final scheduling plan that meets the delay constraints and resource balancing requirements to complete the optimal allocation of edge computing tasks.

[0023] In the above solution, step s1 further includes the following steps:

[0024] S1.1 defines the heterogeneous computing characteristics of the task set and the dynamic resource status of the node set, which specifically includes the following steps:

[0025] S1.1.1 Define the task set T = {t1, t2, ..., t N Each task t in i The computational cost C i , data volume D i , define the edge node set E={e1,e2,...,e N Each node e in j The computing power of R j , bandwidth resources B j and real-time load status;

[0026] S1.1.2 Define the scheduling decision matrix X = {x ij}, where x ij =0 means task t i Assigned to node e j , otherwise 0;

[0027] S1.1.3 Computational density Φ of the calculation task i and the node load ratio ρ j ,in:

[0028]

[0029] S1.2 Combine the task computing density and node load status to design a heuristic allocation strategy to generate an initial scheduling solution population, which specifically includes the following steps:

[0030] Calculate density Φ based on tasks i and node load ratio ρ j , the following heuristic allocation strategy is used to generate the initial candidate solution set P = {X 1 ,X 2 ,...,X n}:

[0031] (a) Prioritize high-computation-density tasks to nodes with low load ratios;

[0032] (b) Prioritize high-data-volume tasks to high-bandwidth resource nodes;

[0033] (c) Generate an initial population containing n scheduling schemes based on the allocation results, and select the individual with the best objective function as the global guidance benchmark.

[0034] In the above scheme, in step s2,

[0035] S2.1 Dynamic Perception Guidance Phase: Task-Based Density Φ Calculation i Construct a scheduling adaptation function with the node resource status to generate the guidance vector G i To adjust the direction of individual evolution, the direction of individual update is determined by G i OK, where:

[0036]

[0037] Among them, λ j Represents node e j The current average number of waiting tasks, Represents the task t i Among all edge nodes j=1,2,…,M, find the best fit η ij The node number j that reaches the maximum value;

[0038] S2.2 Perturbation-enhanced regulation stage: Calculating individual X k The local density perception factor δ k , construct the disturbance weight ω k To generate adaptive disturbance terms and update individual positions:

[0039]

[0040]

[0041] ω k =exp(-α·δ k )

[0042] in represents the updated position of the kth individual at iteration t+1, represents the current position of the kth individual at iteration number t, N k Table is individual X k The neighborhood individual set, X l It's X k neighboring individuals, is the global optimal individual, N(0,δ 2 ) is the standard normal distribution disturbance term, β is the guide weight, and α is the disturbance sensitivity coefficient;

[0043] S2.3 State feedback driving stage: Mapping function f according to task completion state s (i, j) and delayed response function f τ (X k ) Construct feedback evaluation index Γ k Driver individual update:

[0044]

[0045] Among them, θ1, θ2 are the weight coefficients of completion rate and delay penalty respectively, γ is the feedback control gain coefficient, τ i For task t i Actual completion time, For its deadline.

[0046] In the above scheme, in step s3,

[0047] S3.1 Build a phased collaborative optimization closed-loop system and dynamically adjust the strategy weights according to the iterative process:

[0048] (a) Activate the dynamic perception guidance mechanism at the beginning of the iteration and set the strategy weight to α t =0.7,β t =0.2,γ t =0.1, giving priority to ensuring the structural rationality of the initial scheduling plan;

[0049] (b) Increase the perturbation enhancement adjustment weight to β in the middle of the evolution t =0.5,α t =0.3,γ t =0.2, strengthen population diversity search;

[0050] (c) Strengthen the weight of the state feedback mechanism to γ ​​in the population stabilization stage t =0.5,α t =0.3,β t =0.2, driving the scheduling scheme to converge in stability;

[0051] S3.2 achieves the superposition of multiple strategies by integrating the control function, which is expressed as:

[0052] F opdate =α t ·D struct +β t ·D perturb +γ t ·D feedback

[0053] Among them D struct is the structural guidance action, D perturb is the disturbance adjustment amplitude, D feedback is the feedback correction strength, weight coefficient α t ,β t ,γ t Linearly decrease or increase according to the iteration round t;

[0054] S3.3 introduces individual state-aware gating function control strategy participation:

[0055]

[0056] when When allowing individual X k Participate in the current round of fusion update, otherwise keep the original state, where ξ is the activation threshold, σ(·) is the sigmoid mapping function, Γ k It is an individual feedback evaluation indicator.

[0057] In summary, the fusion collaborative evolution model constructed by the present invention realizes the full process control of decoupling-collaboration-fusion of multiple strategies in edge computing scheduling scenarios, constructs an evolutionary link combining parameter adjustment and behavior optimization, and effectively improves the convergence efficiency, system robustness and resource adaptability of the task scheduling algorithm.

[0058] Because the present invention adopts the above technical solution, it has the following beneficial effects:

[0059] 1. This invention adopts a computational load-aware individual initialization strategy to solve the problem of poor population search directionality and low early exploration efficiency caused by blind initialization of individual structures in traditional algorithms, thereby achieving the effect of improving the effectiveness of the initial population and the search potential.

[0060] 2. The present invention adopts a local density-aware disturbance enhancement mechanism to solve the problem that the evolution process is prone to falling into local optimality and premature convergence of the population, thereby achieving the effect of improving the algorithm's ability to escape and maintain diversity.

[0061] 3. The present invention adopts a state-driven feedback regulation mechanism to solve the problems of lack of feedback and insufficient adaptive ability in the evolutionary control process, and achieves the effect of dynamically adjusting the search intensity and direction based on the scheduling state.

[0062] 4. The present invention solves the problems of tight policy coupling and poor global performance of traditional scheduling methods in heterogeneous resource environments through an adaptive evolution process of multi-strategy fusion, thereby achieving the effect of improving the robustness of scheduling solutions and the flexibility of resource adaptation. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 Structural diagram of the disturbance enhancement regulation mechanism.

[0064] Figure 2 Structure diagram of the state-driven evolutionary feedback mechanism.

[0065] Figure 3 Overall flow chart of the multi-strategy fusion scheduling algorithm. DETAILED DESCRIPTION

[0066] The purpose of the present invention is to propose a multi-strategy fusion algorithm suitable for edge computing task scheduling, aiming to solve the problems of extended scheduling time, low resource utilization and poor system stability caused by parameter staticity, low search efficiency and uneven load distribution in the edge computing environment. The present invention realizes environmental adaptation of algorithm parameters, directional enhancement of the search process and feedback optimization of scheduling behavior by constructing a three-stage collaborative evolution mechanism of dynamic parameter control, resource perception disturbance and feedback-guided update, comprehensively improving the response efficiency of edge scheduling, rationality of resource allocation and global search stability, and meeting the composite requirements of intelligent, flexible and controllable task scheduling in edge computing scenarios.

[0067] In order to achieve the above objectives, the present invention innovatively establishes a fusion optimization framework of "computational load perception-search strategy self-adjustment-state feedback enhancement", organically integrates task characteristics, node status and algorithm evolution process, and constructs a structurally complete, strategy-coupled and responsive intelligent scheduling algorithm, which significantly improves the adaptability of task scheduling and the system performance of scheduling results.

[0068] Furthermore, in order to verify the practical effect of the multi-strategy fusion scheduling algorithm proposed in the present invention in the edge computing scenario, the following uses the distributed mobile application automation testing platform as a typical application scenario to expand the specific application process of the algorithm in actual task scheduling problems.

[0069] The platform consists of multiple edge nodes, each connected to several mobile devices. These nodes must regularly execute batch tasks from the test task scheduling system. These tasks include automated UI interaction commands such as clicks, swipes, and inputs, accompanied by the upload of large amounts of log and behavioral data. These tasks exhibit significant heterogeneity, timeliness, and computational load variations. The system must ensure that these tasks are completed efficiently while meeting latency constraints, avoiding node congestion and performance bottlenecks.

[0070] The present invention provides a multi-strategy fusion algorithm suitable for edge computing task scheduling, comprising the following steps:

[0071] S1: Establish a scheduling model in a heterogeneous edge computing environment, including:

[0072] The present invention considers a typical edge computing task scheduling scenario, where the task set is denoted as T = {t1, t2, ..., t N}, each task t i Including calculation amount C i and data volume D i ; The edge node set is represented as E={e1,e2,...,e N}, each node e j Corresponding computing power R j , bandwidth resources B jAnd the current system load status. The scheduling scheme is based on the decision matrix X = {x ij} means, where x ij =0 means task t i Assigned to node e j , otherwise it is 0. The computational density of the task is defined as The load ratio of the node is expressed as The two together serve as initial reference indicators for resource allocation.

[0073] In the population initialization phase, a heuristic allocation strategy is designed based on task characteristics and node status:

[0074] High-density tasks are matched with low-load nodes first, and high-bandwidth nodes are given priority to undertake high-data-volume tasks. The initial candidate solution set P = {X 1 ,X 2 ,...,X n At the same time, the optimal individual is selected as the global guide throughout the subsequent evolution process. This stage completes the variable modeling and initial solution generation of the scheduling problem, providing the structural and numerical foundation for the multi-strategy fusion optimization mechanism.

[0075] As a specific implementation method, refer to the following:

[0076] Assume that the task set is T = {t1, t2, ..., t N}, including 40 application test tasks, each task t i Has the following properties: Calculate the load γ i Indicates the number of UI operation steps included in the task, with a value range of 200–1000 (unit instruction number); communication load θ i Indicates the log data upload amount, ranging from 0.5–5MB; maximum completion time Set to 3–10 seconds after task submission. The edge node set is N=

[0077] {n1,n2,...,n6}, the configuration information of each node is as follows:

[0078]

[0079] To quickly generate an initial feasible solution, the platform calculates the density Φ based on the task i Calculate the idleness ρ with the node j Construct a heuristic scheduling propensity score matrix. The initial task allocation is based on the principle of "high-density tasks → high-resource nodes" to form the initial population P (0) .

[0080] Application of S2 multi-strategy fusion mechanism in actual scheduling

[0081] To improve the algorithm's environmental adaptability and evolutionary efficiency in edge scheduling, this paper designs a three-stage fusion optimization mechanism, including dynamic perception guidance, disturbance enhancement regulation, and state feedback drive. These three mechanisms respectively improve task modeling, search strategy, and scheduling result evaluation, and work together during the evolutionary process to build a complete intelligent scheduling process.

[0082] S201: Dynamic Perception Guidance Mechanism

[0083] In traditional scheduling algorithms, the individual evolution direction often relies solely on the static objective function, failing to fully integrate the task computing attributes and node resource status, resulting in poor scheduling stability under resource heterogeneity and dynamic load conditions. To this end, the present invention introduces a dynamic guidance mechanism driven by the task computing factor, dynamically adjusting the individual update direction through computing density information, and improving the structural perception ability of the search. Specifically, based on Φ i Guide individuals to comprehensively consider computing density and current node load when allocating tasks, and define the scheduling fitness function:

[0084]

[0085] Among them, λ j Represents node e j The current average number of waiting tasks. The higher the fitness, the more suitable the node is for carrying the current task. During the individual position update process, the fitness function is used to construct the guidance vector:

[0086]

[0087] As the pre-allocation direction of the current task among individuals, it guides the individual's update behavior to move towards a better matching direction between tasks and nodes.

[0088] On the one hand, this mechanism dynamically models the combined relationship between task computing load and communication overhead. On the other hand, it explicitly intervenes in individual behavior control through fitness evaluation, making the evolving behavior resource-aware and scheduling-oriented, providing a structural basis for subsequent disturbance regulation strategies.

[0089] As a possible specific implementation, refer to the following:

[0090] Deployment of structural guidance mechanism in scenarios

[0091] At the beginning of each round of evolution, the system i The calculation density and current node load are combined to dynamically adjust the scheduling priority. In practice, the following scheduling preference function is used:

[0092]

[0093] Among them, Lj The total amount of computational resources currently assigned to node j. This function reflects the relative adaptability of the remaining resources of the node to the task load. The larger the value, the more inclined to allocate. The initial update direction of the individual is based on the maximum η ij The scheduling tendency function used in this embodiment is a simplified example implementation of the structure guidance mechanism in the invention, and its variable structure can be adjusted according to the actual task load model.

[0094] S202 Search disturbance enhancement adjustment strategy

[0095] After introducing the dynamic perception guidance mechanism (see S201), the initial evolution of individuals has a structural direction. However, in the subsequent search process, if the perturbation pattern is too simple, it may still fall into a local optimum or cause premature convergence of the population. Traditional mutation strategies (such as single Gaussian perturbation) lack the ability to respond to local convergence states, making it difficult to maintain population diversity and search activity in complex scheduling scenarios.

[0096] To this end, the present invention constructs a responsive disturbance enhancement regulation mechanism (see Figure 1 ), by introducing the perturbation kernel function of local aggregation perception, the perturbation intensity is adaptively adjusted according to the density of the search area where the individual is currently located, and the differentiated perturbation behavior control based on the density distribution is realized. First, define the individual X k Local density perception factor:

[0097]

[0098] Among them, N k Represents individual X k The neighborhood of the first q individuals closest to the current individual. This factor measures the degree of aggregation of the local area where the current individual is located. The perturbation weight function based on this factor is constructed as ω k =exp(-α·δ k ) where α is the disturbance sensitivity coefficient. The disturbance direction is to optimally guide the individual X * As the target, the perturbation kernel is superimposed:

[0099]

[0100] Among them, β is the guide weight, N(0,δ 2 ) is the standard normal disturbance term. k The perturbation amplitude is adaptively adjusted to maintain a high perturbation intensity in the dense search area, thereby enhancing the ability to escape; while the perturbation is weakened in the sparse search area, which is conducive to local convergence.

[0101] As a supplement to the S201 guidance mechanism, this mechanism realizes the dynamic enhancement of disturbance behavior based on structural perception, effectively alleviates the problems of search traps and individual premature convergence, and lays the population foundation for the final scheduling quality optimization.

[0102] As a possible specific implementation, refer to the following:

[0103] Local implementation of disturbance-enhanced regulation mechanism

[0104] In each mid-term iteration, the platform calculates the local density perception metric of the individual:

[0105]

[0106] Among them, N k is the distance between individuals X in the population k The 5 most recent individuals. Adaptively adjust the disturbance intensity ω based on the density value k , generating the disturbance term:

[0107] Δx k =ω k ·∈where∈~N(0,0.05 2 )

[0108] During the scheduling process, when the individual density δ in a certain area k <0.1, the perturbation amplitude is reduced by 50% for refined search; when δ k When it is >0.4, it is increased by 2 times to try to escape the local trap.

[0109] S203 state-driven evolutionary feedback mechanism

[0110] Under the influence of computational perception guidance (S201) and disturbance enhancement regulation (S202), individuals initially possessed structural rationality and search flexibility. However, in actual evolution, the current strategy still focuses on static expected optimality and lacks feedback on historical scheduling results. This makes the algorithm ineffective in self-correction, especially when faced with task failures or severe delays, limiting the continued improvement of scheduling performance.

[0111] To this end, the present invention designs a state-driven evolution feedback mechanism (see Figure 2 ), introduce the task completion state mapping function and the scheduling delay response function, build a feedback control path, and establish a dynamic closed loop between the scheduling behavior and the search strategy. First, define individual X k The task completion status mapping function:

[0112]

[0113] And construct the task completion rate matrix: S k =[fs (i,j)] N×M , and define the delay response function of the scheduling scheme as:

[0114]

[0115] Among them, τ i For task t i Actual completion time, Its deadline. This function measures the overall delay penalty of the scheduling scheme. Based on the above function, a fusion feedback evaluation index is constructed:

[0116]

[0117] Finally, the feedback-driven update rule is constructed:

[0118]

[0119] Here, θ1 and θ2 are the weight coefficients for completion rate and delay penalty, respectively, and γ is the feedback control gain coefficient. This rule dynamically adjusts the evolutionary guidance intensity based on the actual historical scheduling results, reinforcing successful allocation behaviors and penalizing scheduling failure areas, thereby improving the global coordination of the population and the stability of task completion.

[0120] Compared to traditional algorithms that rely solely on static objective optimization, this mechanism, centered on state feedback, dynamically perceives the system's operating status, forming a closed-loop evolutionary control chain of "perception-evaluation-feedback-guidance." This not only improves the scheduling algorithm's adaptability to both timeliness and completion rate, but also organically integrates the strategy logic and action phases with the structural guidance mechanism of S2.1 and the disturbance regulation strategy of S2.2, ensuring global consistency and local adaptability of the integrated optimization process.

[0121] As a possible specific implementation, refer to the following:

[0122] Practical deployment of the state feedback mechanism

[0123] After each round of scheduling is completed, the platform records the actual completion time of each task Calculate the latency metric:

[0124]

[0125] Based on the delay metric d o , calculate the delayed response function:

[0126]

[0127] And count the task completion rate Construct comprehensive feedback score:

[0128] Γ k =0.8·r-0.2·f τ (X k )

[0129] This score is used as the individual X k The guidance correction coefficient is used to participate in the update of the next round of scheduling behavior.

[0130] Finally, the feedback-driven update rule is constructed:

[0131]

[0132] S3 Fusion Collaboration Evolution Model

[0133] In order to effectively integrate the three heterogeneous strategies of task feature drive, disturbance enhancement regulation and state feedback guidance, the present invention adopts a fusion collaborative evolution model. This model takes the task scheduling process as the main line and embeds multiple strategies into different stages of evolution to form a multi-stage collaborative optimization closed-loop system of structure-disturbance-feedback. The overall algorithm process structure is as follows: Figure 3 shown.

[0134] First, in the initial stage (before population construction and the first round of iteration), the task computing amount perception guidance mechanism (S201) is activated to generate an individual initial solution space that meets structural rationality based on the task load distribution and node heterogeneity capabilities, ensuring that the evolution direction has initial scheduling perception.

[0135] Entering the mid-stage of evolution, as individuals continue to iterate and update, the scheduling plan tends to converge locally. At this point, the perturbation enhancement adjustment mechanism (S202) is introduced as the core of the dynamic regulation of the mutation operation. By guiding the perturbation intensity through the combined factors of task balance and mutation direction, the algorithm can break the local optimum and achieve cross-region search, effectively improving the breadth of solution space exploration.

[0136] Finally, in the late stage of evolution and the stage of population stabilization, the state-driven evolution feedback mechanism (S203) is started to introduce the execution feedback of the previous round of scheduling plans into the current iteration, guide high-quality scheduling plans to converge towards the target, suppress resource allocation deviations and individual degradation trends, and improve the practical implementation and long-term stability of the scheduling results.

[0137] In addition, the fusion collaborative evolution model also introduces the individual state perception gating function: It is used to control whether individuals participate in the current round of fusion updates, prevent low-quality individuals from disrupting the overall evolutionary direction, and realize feedback gating control and strategy selection adaptation at the individual level.

[0138] As a possible specific implementation, refer to the following:

[0139] Deployment of fusion collaboration mechanism in mission life cycle

[0140] The system dynamically controls the activation ratio of the three strategies throughout the entire algorithm evolution cycle, using the following control rules.

[0141] In the first 0–10 rounds of iteration, structural guidance is mainly used (weight distribution: guidance 0.7, perturbation 0.2, feedback 0.1); in the middle stage (rounds 11–30), the perturbation ratio is increased to enhance search diversity (perturbation 0.5, guidance 0.3, feedback 0.2); in the late stage (rounds 31–50), feedback guidance is strengthened (feedback 0.5, guidance 0.3, perturbation 0.2) to promote convergence of scheduling stability.

[0142] Fusion control function:

[0143] F opdate =α t ·D struct +β t ·D perturb +γ t ·D feedback

[0144] The three items are the scheduling adjustment range of each sub-strategy, α t ,β t ,γ t As the rounds change, the values ​​are dynamically assigned according to the above ratio.

[0145] S4: Output the final scheduling plan that meets the delay constraints and resource balancing requirements, and complete the optimized allocation of edge computing tasks.

Claims

1. A multi-strategy fusion algorithm suitable for edge computing task scheduling, characterized by: The following steps are involved: Step S1: Establish a scheduling model in a heterogeneous edge computing environment, including: S1.1 defines the heterogeneous computing characteristics of the task set and the dynamic resource status of the node set; S1.2 Combine task calculation density and node load status to design a heuristic allocation strategy to generate an initial scheduling solution population; Step S2: Implement multi-strategy fusion optimization iteration, including: S2.1 Dynamic Perception Guidance Phase: Generates scheduling adaptability indicators based on task computation characteristics and node resource status, and constructs guidance vectors to adjust task allocation directions; S2.2 Perturbation Enhancement Adjustment Stage: Dynamically calculate the perturbation weight based on the local aggregation of individuals in the population, and combine the optimal guidance direction with the adaptive perturbation term to generate mutant individuals. The mutant individuals maintain high perturbation intensity in the search dense area, enhancing the ability to escape; while the perturbation is weakened in the search sparse area, which is conducive to local convergence; S2.3 State feedback driven phase: feedback evaluation indicators are constructed through task completion rate and delay penalty, and the population evolution direction is dynamically corrected based on historical scheduling results; Step S3: Execute phased fusion collaborative evolution control, including: In the early stages of iteration, the dynamic perception guidance mechanism is activated first, the perturbation-enhanced regulation intensity is enhanced in the middle stages, and the state feedback mechanism is strengthened in the later stages; Dynamically adjust the weight coefficients of each strategy according to the iterative process to form a collaborative optimization link of structure guidance, disturbance regulation and feedback correction; Step S4: Output the final scheduling plan that meets the delay constraints and resource balancing requirements to complete the optimal allocation of edge computing tasks.

2. The method according to claim 1, characterized in that S1.1 defines the heterogeneous computing characteristics of the task set and the dynamic resource status of the node set, which specifically includes the following steps: S1.1.1 Define the task set T = {t1, t2, ..., t N Each task t in i The computational cost C i , data volume D i , define the edge node set E = {e1, e2, ..., e N Each node e in j The computing power of R j , bandwidth resources B j and real-time load status; S1.1.2 Define the scheduling decision matrix X = {x ij }, where x ij =0 means task t i Assigned to node e j , otherwise 0; S1.1.3 Computational density Φ of the calculation task i and the node load ratio ρ j ,in: S1.2 Combine the task computing density and node load status to design a heuristic allocation strategy to generate an initial scheduling solution population, which specifically includes the following steps: Calculate density Φ based on tasks i and node load ratio ρ j , the following heuristic allocation strategy is used to generate the initial candidate solution set P = {X 1 , X 2 ,...,X n }: (a) Prioritize high-computation-density tasks to nodes with low load ratios; (b) Prioritize high-data-volume tasks to high-bandwidth resource nodes; (c) Generate an initial population containing n scheduling schemes based on the allocation results, and select the individual with the best objective function as the global guidance benchmark.

3. The method according to claim 1, characterized in that In step s2, S2.1 Dynamic Perception Guidance Phase: Task-Based Density Φ Calculation i Construct a scheduling adaptation function with the node resource status to generate the guidance vector G i To adjust the direction of individual evolution, the direction of individual update is determined by G i OK, where: Among them, λ j Represents node e j The current average number of waiting tasks, Represents the task t i Among all edge nodes j=1, 2, ..., M, find the node that makes the fitness η ij The node number j that reaches the maximum value; S2.2 Perturbation-enhanced regulation stage: Calculating individual X k The local density perception factor δ k , construct the disturbance weight ω k To generate adaptive disturbance terms and update individual positions: oh k =exp(-a·d k ) in represents the updated position of the kth individual at iteration t+1, represents the current position of the kth individual at iteration number t, N k Table is individual X k The neighborhood individual set, X l It's X k neighboring individuals, is the global optimal individual, N(0,δ 2 ) is the standard normal distribution disturbance term, β is the guide weight, and α is the disturbance sensitivity coefficient; S2.3 State feedback driving stage: Mapping function f according to task completion state s (i, j) and delayed response function f τ (X k ) Construct feedback evaluation index Γ k Driver individual update: Among them, θ1 and θ2 are the weight coefficients of completion rate and delay penalty respectively, γ is the feedback control gain coefficient, τ i For task t i Actual completion time, For its deadline.

4. The method according to claim 1, wherein In step s3, S3.1 Build a phased collaborative optimization closed-loop system and dynamically adjust the strategy weights according to the iterative process: (a) Activate the dynamic perception guidance mechanism at the beginning of the iteration and set the strategy weight to α t =0.7,β t =0.2,γ t =0.1, giving priority to ensuring the structural rationality of the initial scheduling plan; (b) Increase the perturbation enhancement adjustment weight to β in the middle of the evolution t =0.5,α t =0.3,γ t =0.2, strengthen population diversity search; (c) Strengthen the weight of the state feedback mechanism to γ ​​in the population stabilization stage t =0.5,α t =0.3,β t =0.2, driving the scheduling scheme to converge in stability; S3.2 achieves the superposition of multiple strategies by integrating the control function, which is expressed as: F opdate =a t ·D struct +b t ·D perturb +g t ·D feedback Among them D struct is the structural guidance action, D perturb is the disturbance adjustment amplitude, D feedback is the feedback correction strength, weight coefficient α t , β t , γ t Linearly decrease or increase according to the iteration round t; S3.3 introduces individual state-aware gating function control strategy participation: when When allowing individual X k Participate in the current round of fusion update, otherwise keep the original state, where ξ is the activation threshold, σ(·) is the sigmoid mapping function, Γ k It is an individual feedback evaluation indicator.

Images