Multi-edge computing task unloading and scheduling layering automatic optimization method based on large language model driving

Abstract

The invention relates to the technical field of edge computing and automatic optimization algorithms, and discloses a multi-edge computing task unloading and scheduling hierarchical automatic optimization method based on large language model driving, which comprises the following steps: S1, constructing a computing system, acquiring task data and edge server data to form a data set, dividing the data set into a training set, an optimization set and a verification set; and S2, constructing a joint task unloading and scheduling optimization model taking the total completion time delay of the system as an optimization target, taking minimization of the total completion time delay of the system as an optimization target function, and taking the total completion time delay of the system obtained by the joint task unloading and scheduling optimization model as a fitness value. According to the method, efficient and automatic optimization of task unloading and scheduling problems in the multi-edge computing environment is achieved, the intelligent level and generalization ability of a computing system are improved through an algorithm evolution mechanism driven by a large language model, and the method has important theoretical significance and wide engineering application prospects.

Description

Technical Field

[0001] This invention relates to the field of edge computing and automatic optimization algorithm technology, and in particular to a hierarchical automatic optimization method for multi-edge computing task offloading and scheduling based on a large language model. Background Technology

[0002] With the rapid development of augmented reality, industrial IoT, connected vehicles, and low-latency intelligent applications, traditional architectures relying on centralized cloud computing are gradually revealing problems such as severe network congestion, high response latency, and high bandwidth consumption. Multi-access edge computing (MEC), by pushing computing power down to the network edge to provide low-latency, highly reliable computing services to terminal devices, has become an important technical approach to solve these problems.

[0003] In computing systems, task offloading and task scheduling are two core issues that determine the overall performance of the system. Task offloading determines the execution location of each computing task, i.e., which edge server should handle it; task scheduling determines the execution order of multiple tasks within a server. Both affect the end-to-end latency, resource utilization, and stability of the computing system. Especially in complex environments with multiple tasks, multiple servers, and multiple resource constraints, the task offloading and scheduling problem often manifests as a tightly coupled optimization problem.

[0004] Existing technologies mainly employ the following methods to address this problem: one category is heuristic rules designed based on human experience, such as minimum delay priority allocation and shortest job priority scheduling; another category is search methods based on swarm intelligence optimization algorithms (such as particle swarm optimization and genetic algorithms); in recent years, deep reinforcement learning methods have also been used to construct state-action mapping models to achieve adaptive decision-making. However, the above methods still have the following defects and shortcomings in practical applications:

[0005] 1. Limited generalization ability due to reliance on manually designed rules: Traditional heuristic algorithms heavily rely on manually designed scoring functions or priority rules, which are often optimized for specific scales or load scenarios. When the scale of the computing system, task characteristics, or resource distribution changes, the original rules are difficult to maintain stability and lack cross-instance adaptive capabilities.

[0006] 2. Lack of explicit feasibility checking and repair mechanisms: Under multiple resource hard constraints, some optimization algorithms handle resource overrun problems through penalty functions, but these functions are prone to generating infeasible solutions in high-dimensional discrete spaces, especially in large-scale scenarios where the feasibility rate drops significantly. The existence of infeasible solutions not only affects the stability of the computational system but also reduces the practical engineering usability of the algorithm.

[0007] 3. Separation of decision-making between upper and lower layers, lack of computational system-level collaborative optimization: Existing methods often handle task offloading and internal server scheduling separately, using fixed scheduling strategies or simple combinations for joint application, lacking unified modeling and collaborative evolution mechanisms, making it difficult to achieve true end-to-end optimization.

[0008] 4. There is a contradiction between training cost and adaptability: Deep reinforcement learning-based methods usually require a large amount of training data and long-term iterative training. When environmental parameters change, the model often needs to be retrained, resulting in high training costs and long adaptation cycles, which is not conducive to rapid deployment and online adjustment in dynamic edge environments.

[0009] Therefore, there is an urgent need for a novel optimization method that can automatically generate high-quality offloading and scheduling strategies under strict resource constraints, has feasibility assurance capabilities, supports cross-instance generalization, and does not require large-scale pre-training, in order to improve the overall performance and engineering practical value of computing systems. Summary of the Invention

[0010] To address the technical problems existing in the prior art, this invention provides an automatic optimization method for multi-edge computing task offloading and scheduling based on a large language model.

[0011] This invention employs the following technical solution: a hierarchical automatic optimization method for multi-edge computing task offloading and scheduling based on a large language model, comprising the following steps:

[0012] S1. Construct a computing system, acquire task data and edge server data to form a dataset, and divide the dataset into a training set, an optimization set and a validation set;

[0013] S2. Construct a joint task offloading and scheduling optimization model with the total system completion delay as the optimization objective. Minimize the total system completion delay as the optimization objective function. Use the total system completion delay obtained from the joint task offloading and scheduling optimization model as the fitness value and set constraints.

[0014] S3. Construct a hierarchical optimization framework, which includes an offloading decision layer and a scheduling and sorting layer.

[0015] S4. Iterative improvement of heuristic rules: Heuristic rules are generated based on a large language model. The training set is input into a hierarchical optimization framework, and a performance feedback mechanism is used to iteratively improve the heuristic rules to generate the final heuristic rules.

[0016] S5. Iterative optimization of the hierarchical optimization framework: The optimization set is input into the hierarchical optimization framework, and the hierarchical optimization framework is iteratively optimized based on the hierarchical evolutionary search mechanism and the feasibility check and repair mechanism.

[0017] S6. Validate the hierarchical optimization framework of iterative optimization using a validation set;

[0018] S7. Input the computational task into the trained hierarchical optimization framework to generate optimized task unloading and scheduling strategies.

[0019] As a further improvement to the above scheme, in step S1, the computing system includes a task set and an edge server set.

[0020] Each task includes parameters for computing resource requirements, memory requirements, and data scale, while each edge server includes parameters for computing power, memory capacity, and communication bandwidth.

[0021] The dataset was divided into training, optimization, and validation sets in a 4:4:2 ratio.

[0022] As a further improvement to the above scheme, in step S2, the objective function for optimization is:

[0023] ,in Includes: task transmission latency Queue waiting time Actual calculation execution delay ;

[0024] The constraints include edge server computing resource constraints, communication bandwidth constraints, task allocation constraints, and task execution order constraints.

[0025] As a further improvement to the above scheme, in step S3, the unloading decision layer is used to generate candidate allocation schemes, and the scheduling and sorting layer is used to optimize the sorting of the task set inside each edge server based on the candidate allocation schemes generated by the unloading decision layer, and to calculate the queuing waiting time and completion time inside the edge server.

[0026] As a further improvement to the above scheme, in step S4, a large language model is constructed. The large language model automatically generates new heuristic rules by receiving optimization problem description information, current optimal solution information and historical function performance information. The generated new heuristic rules are embedded into the hierarchical optimization framework in the form of executable functions.

[0027] As a further improvement to the above scheme, in step S4, the performance feedback mechanism adopts the following steps:

[0028] The newly generated heuristic rules are evaluated for performance. The total completion time of the joint task unloading and scheduling optimization model system is used as the performance evaluation index of the heuristic rules, which is composed of the average fitness value and the proportion of feasible solutions. The evaluation results are fed back to the large language model as "historical performance" information in the next round of iterative optimization, so as to realize the iterative optimization of the heuristic rules.

[0029] As a further improvement to the above scheme, the hierarchical evolutionary search mechanism in step S5 is as follows: in the unloading layer, a population search is performed on the task allocation scheme, and candidate content is generated through selection, crossover and mutation operations; in the scheduling and sorting layer, the task sequences of each edge server are sorted and optimized based on heuristic rules.

[0030] The feasibility check and repair mechanism involves performing constraint checks and repairs during each iteration to ensure that resource constraints are met.

[0031] As a further improvement to the above scheme, the steps of the hierarchical evolutionary search mechanism in step S5 are as follows:

[0032] S11, Population Initialization:

[0033] In the unloading decision layer, a task allocation matrix is ​​randomly generated to form an initial population, and each individual task satisfies the unique task allocation constraint.

[0034] S12, Select Operation:

[0035] Individual task participants are sorted according to their fitness values, and a tournament selection or roulette wheel selection mechanism is used to retain the best individual task participants.

[0036] S13. Crossover and mutation operations:

[0037] The offloading layer employs a crossover operator based on task subset exchange and a mutation operator based on edge server reallocation; the mutation probability and crossover probability can be dynamically adjusted according to the iteration stage.

[0038] At the scheduling and sorting layer: the task execution sequence is reconstructed based on heuristic rules generated by the large language model.

[0039] As a further improvement to the above scheme, the feasibility check and repair mechanism in step S5 adopts the feasibility priority principle. Based on the computing resources of the edge server, the tasks are redistributed to the server with the lowest load according to the individual task load. Based on the communication bandwidth and communication volume, the tasks with the largest data volume are adjusted first. For tasks that are still unrepairable, a penalty is imposed to reduce their fitness.

[0040] As a further improvement to the above scheme, in step S7, the optimal task unloading and scheduling strategy is output. After reaching the preset number of iterations or the convergence condition, the optimal task allocation matrix and the corresponding execution order inside the edge server are output, and the system performance index is calculated.

[0041] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0042] 1. This invention introduces a large language model-driven automatic algorithm design mechanism in the multi-edge computing task offloading and scheduling problem; through semantic understanding and rule generation capabilities, it breaks through the limitations of traditional manual experience rule design, realizes the automatic generation and continuous optimization of heuristic strategies, and significantly improves the algorithm's adaptability and cross-scenario generalization ability.

[0043] 2. This invention achieves the co-evolution of unloading decision and scheduling order through a hierarchical optimization framework. Compared with existing methods that handle the two separately, it can uniformly evaluate and adjust under the computing system-level objectives, improve the overall optimization effect, and reduce the total latency of the computing system.

[0044] 3. This invention introduces a hierarchical evolutionary search and feasibility check and repair strategy, which effectively increases the proportion of feasible solutions in the discrete solution space, avoids resource overrun problems, and improves the engineering usability and stability of the algorithm.

[0045] 4. This invention can achieve policy evolution through prompt-driven and online iteration without the need for large-scale pre-training or multi-model integration, achieving a good balance between training cost and running efficiency; once high-quality heuristic rules are obtained, they can be quickly migrated to different scales or different load scenarios, significantly reducing retraining costs.

[0046] 5. The invention not only achieves efficient and automatic optimization of task offloading and scheduling problems in multi-edge computing environments, but also improves the intelligence level and generalization ability of computing systems through the algorithm evolution mechanism driven by a large language model, which has important theoretical significance and broad engineering application prospects. Attached Figure Description

[0047] Figure 1 A flowchart of an automatic optimization method for multi-edge computing task offloading and scheduling based on a large language model provided by the present invention;

[0048] Figure 2 The figure shows the experimental results of the hierarchical automatic optimization method for multi-edge computing task offloading and scheduling based on a large language model driven by the present invention under multiple scenario conditions. Detailed Implementation

[0049] The present invention will now be further described in conjunction with the accompanying drawings and specific embodiments. It should be noted that, without conflict, the various embodiments or technical features described below can be arbitrarily combined to form new embodiments.

[0050] Example 1:

[0051] Please combine Figure 1 This embodiment of a hierarchical automatic optimization method for multi-edge computing task offloading and scheduling based on a large language model includes the following steps:

[0052] S1. Construct a computing system, acquire task data and edge server data to form a dataset, and divide the dataset into a training set, an optimization set and a validation set;

[0053] The computing system includes a set of tasks and a set of edge servers.

[0054] Each task includes parameters for computing resource requirements, memory requirements, and data scale, while each edge server includes parameters for computing power, memory capacity, and communication bandwidth.

[0055] The dataset was divided into training, optimization, and validation sets in a 4:4:2 ratio.

[0056] S2. Construct a joint task offloading and scheduling optimization model with the total system completion delay as the optimization objective. Minimize the total system completion delay as the optimization objective function. Use the total system completion delay obtained from the joint task offloading and scheduling optimization model as the fitness value and set constraints.

[0057] The objective function to be optimized is:

[0058] ,in Includes: task transmission latency Queue waiting time Actual calculation execution delay ;

[0059] The constraints include edge server computing resource constraints, communication bandwidth constraints, task allocation constraints, and task execution order constraints.

[0060] S3. Construct a hierarchical optimization framework, which includes an offloading decision layer and a scheduling and sorting layer.

[0061] The unloading decision layer is used to generate candidate allocation schemes, and the scheduling and sorting layer is used to optimize the sorting of the task set inside each edge server based on the candidate allocation schemes generated by the unloading decision layer, and to calculate the queuing waiting time and completion time inside the edge server.

[0062] S4. Iterative improvement of heuristic rules: Heuristic rules are generated based on a large language model. The training set is input into a hierarchical optimization framework, and a performance feedback mechanism is used to iteratively improve the heuristic rules to generate the final heuristic rules.

[0063] A large language model is pre-built. The large language model automatically generates new heuristic rules by receiving optimization problem description information, current optimal solution information, and historical function performance information. The generated new heuristic rules are embedded into the hierarchical optimization framework in the form of executable functions.

[0064] The performance feedback mechanism follows these steps:

[0065] The newly generated heuristic rules are evaluated for performance. The total completion time of the joint task unloading and scheduling optimization model system is used as the performance evaluation index of the heuristic rules, which is composed of the average fitness value and the proportion of feasible solutions. The evaluation results are fed back to the large language model as "historical performance" information in the next round of iterative optimization, so as to realize the iterative optimization of the heuristic rules.

[0066] S5. Iterative optimization of the hierarchical optimization framework: The optimization set is input into the hierarchical optimization framework, and the hierarchical optimization framework is iteratively optimized based on the hierarchical evolutionary search mechanism and the feasibility check and repair mechanism.

[0067] The hierarchical evolutionary search mechanism consists of the following steps: in the unloading layer, a population search is performed on the task allocation scheme, and candidate content is generated through selection, crossover, and mutation operations; in the scheduling and sorting layer, the task sequences of each edge server are sorted and optimized based on heuristic rules.

[0068] The steps of the hierarchical evolutionary search mechanism are as follows:

[0069] S11, Population Initialization:

[0070] In the unloading decision layer, a task allocation matrix is ​​randomly generated to form an initial population, and each individual task satisfies the unique task allocation constraint.

[0071] S12, Select Operation:

[0072] Individual task participants are sorted according to their fitness values, and a tournament selection or roulette wheel selection mechanism is used to retain the best individual task participants.

[0073] S13. Crossover and mutation operations:

[0074] The offloading layer employs a crossover operator based on task subset exchange and a mutation operator based on edge server reallocation; the mutation probability and crossover probability can be dynamically adjusted according to the iteration stage.

[0075] At the scheduling and sorting layer: the task execution sequence is reconstructed based on heuristic rules generated by the large language model;

[0076] The feasibility check and repair mechanism consists of: performing constraint checks and repairs during each iteration to ensure that resource constraints are met;

[0077] The feasibility check and repair mechanism adopts the feasibility priority principle, that is, based on the computing resources of the edge server, the tasks are redistributed to the server with the lowest load according to the individual task load, and based on the communication bandwidth and communication volume, the tasks with the largest data volume are adjusted first; and the task individuals that are still unrepairable are penalized to reduce their fitness.

[0078] S6. Validate the hierarchical optimization framework of iterative optimization using a validation set;

[0079] S7. Input the computational task into the trained hierarchical optimization framework to generate the optimized task unloading and scheduling strategy content.

[0080] Output the optimal task unloading and scheduling strategy. After reaching the preset number of iterations or convergence conditions, output the optimal task allocation matrix and the corresponding execution order within the edge server, and calculate the system performance indicators.

[0081] Example 2:

[0082] Combination Figure 2 A hierarchical automatic optimization method for multi-edge computing task offloading and scheduling based on a large language model includes the following steps:

[0083] S1. Construct a computing system, acquire task data and edge server data to form a dataset, and divide the dataset into a training set, an optimization set and a validation set;

[0084] Build includes One edge server and The computing system performs a computing task and acquires task data and edge server data to form a dataset. The dataset is then divided into a training set, an optimization set, and a validation set in a 4:4:2 ratio.

[0085] Each task It has the following attribute parameter: Task input data volume Calculation requirements for the task ;

[0086] Each edge server It has the following resource parameters: available computing resources Communication bandwidth Current load status ;

[0087] Define binary task allocation variables :

[0088] when =1 indicates a task Assigned to server implement; =0 indicates otherwise; and a unique assignment constraint is set: .

[0089] S2. Construct a joint task offloading and scheduling optimization model with the total system completion time as the optimization objective. Use the minimization of the total system completion time as the optimization objective function. Use the total system completion time obtained from the joint task offloading and scheduling optimization model as the fitness value, thereby providing a unified performance evaluation standard for the subsequent hierarchical optimization framework and setting constraints.

[0090] The objective function is: ,in Includes: task transmission latency Queue waiting time Actual calculation execution delay .

[0091] The constraints include edge server computing resource constraints, communication bandwidth constraints, unique task allocation constraints, and unique task allocation constraints.

[0092] Among them, edge server computing resource constraints: the computing requirements of all assigned tasks on the edge server cannot exceed the edge server's computing capacity;

[0093] Communication bandwidth constraint: The amount of data transmitted by the task must not exceed the corresponding link bandwidth capacity;

[0094] Unique task assignment constraint: Each task must be assigned to one and only one edge server for execution;

[0095] Scheduling order constraint: Tasks on the same server are executed sequentially in a defined order.

[0096] S3. Construct a hierarchical optimization framework, which includes an offloading decision layer and a scheduling and sorting layer.

[0097] The unloading layer is responsible for generating candidate task allocation schemes, and each candidate solution is represented as a... The task allocation matrix is ​​used, and the unloading layer uses a population approach to represent multiple candidate solutions.

[0098] Given a task allocation matrix, the scheduling and sorting layer optimizes the sorting of the task set within each edge server.

[0099] Let the set of tasks assigned to edge server j be . Then the scheduling and sorting layer needs to determine a task execution sequence. , It is used to calculate the queuing time and completion time inside the edge server.

[0100] Meanwhile, the unloading layer and the scheduling and sorting layer achieve joint operation through a hierarchical collaboration mechanism. In each iteration, the unloading layer generates candidate allocation schemes, and the scheduling layer sorts the tasks and calculates the total system completion delay according to the current heuristic rules. The calculated fitness values ​​are fed back to the unloading layer for the next round of selection and mutation. Through the hierarchical collaboration mechanism, the original high-dimensional joint optimization problem is effectively decomposed and solved.

[0101] S4. Iterative improvement of heuristic rules: Heuristic rules are generated based on a large language model, and a performance feedback mechanism is used to iteratively improve the heuristic rules to generate the final heuristic rules.

[0102] Building a large language model: A large language model is a pre-trained language model built on the Transformer architecture. The large language model includes an encoding module, a context semantic understanding module, and a code generation module.

[0103] The encoding module is used to vectorize the input optimization problem description information, current optimal solution information, and historical function performance information;

[0104] The context semantic understanding module is used to extract the relationship features between the optimization objective and the constraints;

[0105] The code generation module automatically generates heuristic rules based on the prompts.

[0106] Heuristic rules include heuristic computation functions and task priority computation functions;

[0107] The optimization problem description information includes the objective function and constraints used for the current optimization.

[0108] The current optimal solution information consists of the optimal task allocation scheme and the corresponding optimal task ranking result generated by the hierarchical optimization framework under the current heuristic rule;

[0109] Historical function performance information includes the average fitness value and the proportion of feasible solutions corresponding to the current heuristic rule in multiple iterations;

[0110] The optimization objective is to optimize the overall performance indicators of the computational system, including minimizing the total system completion delay and the proportion of feasible solutions;

[0111] The prompts include optimization problem descriptions, current optimal solutions, and historical function performance information, which are used to guide the large language model to generate heuristic rules that adapt to the current optimization environment.

[0112] Meanwhile, the input information of the large language model is organized in the form of structured prompts, enabling the large language model to understand the current optimization state and improvement goals; the generated heuristic rules are embedded in the hierarchical optimization framework in the form of executable functions.

[0113] The performance feedback mechanism follows these steps:

[0114] The newly generated heuristic rules are evaluated for performance. The total completion time of the joint task unloading and scheduling optimization model system is used as the performance evaluation index of the heuristic rules, which is composed of the average fitness value and the proportion of feasible solutions. The evaluation results are fed back to the large language model as "historical performance" information in the next round of iterative optimization, so as to realize the iterative optimization of the heuristic rules.

[0115] Specifically:

[0116] First, the newly generated heuristic rules are embedded into the unloading decision layer and the scheduling and sorting layer;

[0117] Candidate task allocation schemes and task execution sequences are generated based on the newly generated heuristic rules;

[0118] The overall system completion delay is optimized using a joint task offloading and scheduling model.

[0119] The average fitness value and proportion of feasible solutions of the newly generated heuristic rules were statistically analyzed across multiple iterations.

[0120] The average fitness value and the proportion of feasible solutions are used together to form the performance evaluation index of the heuristic rule, and serve as "historical performance" information in the next round of iterative optimization.

[0121] S5. Iterative optimization of the hierarchical optimization framework: The training set is input into the hierarchical optimization framework, and the hierarchical optimization framework is iteratively optimized based on the hierarchical evolutionary search mechanism and the feasibility check and repair mechanism.

[0122] Hierarchical evolutionary search mechanism:

[0123] Population initialization:

[0124] Randomly generated at the unloading decision layer A task assignment matrix is ​​formed to constitute the initial population, and each individual task satisfies the task unique assignment constraint.

[0125] Select operation:

[0126] Individuals are ranked according to their fitness values, and tournament selection or roulette wheel selection mechanisms are used to retain superior individuals.

[0127] The selection mechanism involves randomly selecting a subset of individual tasks from the population, comparing the individual tasks in this subset according to their fitness values, and selecting the individual task with the best fitness value to enter the next generation of the population; this process is repeated until a specified number of individuals are generated.

[0128] Crossover and mutation operations:

[0129] At the offloading layer: a crossover operator based on task subset exchange is used; a mutation operator based on server reallocation is used; the mutation probability and crossover probability can be dynamically adjusted according to the iteration stage;

[0130] The crossover operator based on task subset exchange employs the following steps:

[0131] First, two task allocation matrices are randomly selected from the candidate task allocation schemes generated by the unloading layer as parent task allocation matrices;

[0132] A subset of tasks is formed by randomly selecting several task indices. The task indices are generated using a uniform random sampling method. A subset of tasks with different numbers is randomly selected from the entire task set according to a preset crossover ratio. The crossover ratio ranges from 20% to 50%.

[0133] Exchange the server allocation information of the two parent matrices at the corresponding positions in this task subset;

[0134] Then, two new child task assignment matrices are generated.

[0135] The mutation operator based on server reallocation includes the following steps:

[0136] Select a task randomly;

[0137] A new server is randomly selected from the set of servers that meet the resource constraints.

[0138] Reset the execution server for this task to the selected server;

[0139] Update the task assignment matrix.

[0140] In the early stages of evolution, a higher mutation probability is used to enhance the search space exploration capability; as the number of iterations increases, the mutation probability is gradually reduced to enhance the convergence stability of the algorithm; the mutation probability can be updated in the following way:

[0141]

[0142] Where t is the current iteration number and T is the maximum iteration number. The maximum mutation probability, This represents the minimum mutation probability.

[0143] At the scheduling layer: the task execution sequence is reconstructed based on the sorting rules generated by the large language model.

[0144] The feasibility check and repair mechanism adopts the feasibility priority principle. When a candidate task allocation scheme violates resource constraints, the feasibility is restored by readjusting the task allocation.

[0145] Specifically, when server computing resources are exceeded, the task with the greatest computing demand is prioritized and reallocated to the server with the most remaining resources; when communication bandwidth is exceeded, the task with the largest data volume is prioritized and transferred to the server with sufficient bandwidth resources.

[0146] For individuals that are still beyond repair, a penalty is imposed to reduce their fitness:

[0147] When a candidate solution still fails to meet resource constraints after repair, a penalty term is introduced into the fitness function:

[0148]

[0149] in λ represents the degree of constraint violation, and λ is the penalty coefficient.

[0150] Fitness calculation:

[0151] Total system completion time for each task: ,in It is calculated from transmission time, queuing time, and execution time.

[0152] Termination conditions:

[0153] Iteration stops when any of the following conditions are met:

[0154] Reach the maximum preset number of iterations;

[0155] The optimal solution for K consecutive generations showed no significant improvement, where the value of K ranged from 5 to 20.

[0156] The decrease in total system latency is less than the preset threshold.

[0157] S6. Validate the hierarchical optimization framework of the iterative optimization using a validation set. Input the validation set into the optimization framework and the total system completion time to verify the stability and adaptability under different task scales and resource configurations.

[0158] S7. Input the computational task into the trained hierarchical optimization framework to generate the optimized task unloading and scheduling strategy content.

[0159] Optimal strategy extraction:

[0160] After the termination condition is met, the individual with the best fitness value is selected from the population as the final output strategy. This strategy includes:

[0161] Task Unloading Matrix ,in for A binary matrix is ​​used to represent the execution server for each task;

[0162] Server internal scheduling sequence set , ,in Indicates server The optimal internal task execution order;

[0163] Performance index calculation:

[0164] Based on the optimal strategy, the overall performance of the computing system is quantitatively calculated, including:

[0165] Total system completion time ;

[0166] The proportion of feasible solutions is the percentage of individuals that satisfy the constraints throughout the entire evolution process, used to measure the stability of the algorithm.

[0167] Results visualization:

[0168] like Figure 2 As shown: The optimization results are visualized graphically;

[0169] in, Figure 2 (a) in the figure is a statistical chart of the computing system's operating results under different task scales and resource configurations:

[0170] Figure 2 In (a) of the above, the small-scale task scenario is: number of tasks N=30~50, number of edge servers M=3~5; the medium-scale task scenario is: number of tasks N=50~200, number of edge servers M=5~20; the large-scale task scenario is: number of tasks N=200~400, number of edge servers M=20~35.

[0171] The combination of random algorithm and FIFO algorithm means that the upper-level task unloading adopts a random allocation strategy, while the lower-level server internal scheduling adopts a first-in-first-out (FIFO) strategy.

[0172] The DRL algorithm + FIFO algorithm means that the upper-layer task offloading uses the Deep Reinforcement Learning (DRL) algorithm for decision-making, while the lower-layer server internal task scheduling uses the First-In-First-Out (FIFO) strategy.

[0173] H-EoH stands for Hierarchical Automatic Optimization. The upper-level task unloading adopts a heuristic evolutionary optimization strategy driven by a large language model, while the lower-level scheduling and sorting adopts a heuristic rule optimization strategy.

[0174] The hierarchical automatic optimization method (H-EoH) achieves the best overall performance in terms of total system completion time.

[0175] In small-scale task scenarios, the total system completion latency of the hierarchical automatic optimization method is 1.2015, which is about 24.6% lower than the 1.5927 of the random algorithm + FIFO algorithm and about 33.9% lower than the 1.8173 of the DRL + FIFO algorithm.

[0176] In medium-scale task scenarios, the total system completion latency of the hierarchical automatic optimization method is 1.3153, which is about 11.5% lower than the 1.4865 of the random algorithm + FIFO algorithm and about 23.9% lower than the 1.7287 of the DRL + FIFO algorithm.

[0177] In large-scale task scenarios, the total system completion latency of the hierarchical automatic optimization method is 1.4421, which is significantly better than the 1.6108 of the random algorithm + FIFO algorithm and the 2.0621 of the DRL + FIFO algorithm, reducing latency by approximately 10.5% and 30.1%, respectively.

[0178] The results of experiments at various scales show that the hierarchical automatic optimization method achieves optimal or near-optimal performance under most task scale conditions, especially in small and medium scale scenarios where its advantages are most significant. This indicates that the hierarchical automatic optimization method can effectively reduce the total system completion latency and maintain stable optimization capabilities under different load intensities.

[0179] Figure 2 Figure (b) shows a performance comparison between the proposed upper-layer task offloading strategy and various baseline algorithms. By comparing the changing trend of the average completion latency of the computing system under different task sizes, the optimization effect of the proposed offloading strategy in the task allocation stage can be intuitively demonstrated.

[0180] Figure 2 In (b) of the above, the small-scale task scenario is: number of tasks N=30~50, number of edge servers M=3~5; the medium-scale task scenario is: number of tasks N=50~200, number of edge servers M=5~20; the large-scale task scenario is: number of tasks N=200~400, number of edge servers M=20~35.

[0181] Greedy states that it employs a greedy strategy for task unloading, selecting the server with the lowest current latency for task allocation each time.

[0182] PSO means: Particle swarm optimization algorithm is used to make task unloading decisions, and a better task allocation scheme is obtained through particle swarm search.

[0183] DRL stands for: a task offloading policy network constructed using deep reinforcement learning methods, which achieves an adaptive task allocation scheme through state-action mapping.

[0184] H-EoH indicates that a hierarchical automatic optimization method is adopted, with the upper-level task unloading adopting a heuristic evolutionary optimization strategy driven by a large language model, and the lower-level scheduling and sorting adopting a heuristic rule optimization strategy.

[0185] The computational system performance of various task offloading decision algorithms was compared and analyzed under different task scales:

[0186] In small-scale task scenarios, the computational system performance index of the hierarchical automatic optimization method (H-EoH) is 6.6426, which is better than the particle swarm optimization algorithm (PSO) of 6.9107, the greedy algorithm (Greedy) of 6.9786 and the deep reinforcement learning algorithm (DRL) of 9.4958.

[0187] In medium-scale task scenarios, the performance index of the hierarchical automatic optimization method (H-EoH) is 6.5467, which is better than Greedy's algorithm (7.3315), PSO's algorithm (7.4323) and DRL's algorithm (9.6195).

[0188] In large-scale task scenarios, the performance index of the hierarchical automatic optimization method (H-EoH) is 6.3249, which is the best among all the compared algorithms; in comparison, the Greedy algorithm is 6.9927, the PSO algorithm is 7.6822, and the DRL algorithm is 9.6573.

[0189] The results above show that the hierarchical automatic optimization method exhibits superior global optimization capabilities across different scale scenarios, with its advantages being particularly pronounced in large-scale complex task environments. It can fully integrate the resource status of the computing system and historical optimization information to make global decisions, effectively avoiding the problem of traditional heuristic algorithms easily getting trapped in local optima. This significantly reduces the overall completion latency of the computing system and improves the load balancing capability among multiple edge nodes.

[0190] Figure 2 (c) is a performance comparison chart between the internal scheduling strategy of the lower-level server proposed in this invention and the baseline scheduling algorithm. By showing the changes in the completion latency of the computing system after optimizing the execution order of tasks inside the server, the performance advantages of the scheduling strategy of this invention in reducing queuing latency can be reflected.

[0191] Figure 2In (c) of the above, the small-scale task scenario is: number of tasks N=30~50, number of edge servers M=3~5; the medium-scale task scenario is: number of tasks N=50~200, number of edge servers M=5~20; the large-scale task scenario is: number of tasks N=200~400, number of edge servers M=20~35.

[0192] FIFO stands for First In First Out scheduling strategy, which is a strategy that executes tasks sequentially according to the time they arrive at the server.

[0193] SJF stands for Shortest Job First scheduling strategy, which prioritizes the execution of tasks with the least computational load.

[0194] LJF stands for Longest Job First (LJF) scheduling strategy, which prioritizes the execution of the task with the greatest computational cost.

[0195] H-EoH means: a hierarchical automatic optimization method is adopted, with the upper-level task unloading adopting a heuristic evolutionary optimization strategy driven by a large language model, and the lower-level scheduling and sorting adopting a heuristic rule optimization strategy.

[0196] A comparative analysis was conducted on the average completion latency of various server internal scheduling algorithms under different task scales:

[0197] In small-scale task scenarios, the average completion latency of the computation system of the hierarchical automatic optimization method (H-EoH) is 1.2015, which is significantly lower than 1.7228 for the first-in-first-out scheduling algorithm (FIFO) and 2.0489 for the longest job first algorithm (LJF), and also better than 1.4077 for the shortest job first algorithm (SJF).

[0198] In medium-scale task scenarios, the average completion latency of the computational system of the hierarchical automatic optimization method (H-EoH) is 1.3153, which is on par with the SJF algorithm, but still significantly better than the FIFO algorithm's 1.5676 and the LJF algorithm's 1.8612.

[0199] In large-scale task scenarios, the average completion latency of the computation system of the hierarchical automatic optimization method (H-EoH) is 1.4421, which is comparable to the performance of the SJF algorithm, but significantly better than the FIFO algorithm's 1.6968 and the LJF algorithm's 1.9562.

[0200] In summary, the hierarchical automatic optimization method (H-EoH) can maintain excellent latency control under different load conditions, especially in environments with smaller task sizes and more complex resource contention, where its performance advantage is more significant. It can dynamically adjust the task execution order according to the task load, avoiding the queuing and congestion problems caused by local optimal selection in traditional scheduling strategies, thereby effectively reducing the internal waiting latency of the server and improving resource utilization efficiency.

[0201] By comparison Figure 2 (b) and Figure 2 The experimental results in (c) show that the hierarchical automatic optimization method (H-EoH) achieves lower total system completion latency under different task sizes and computing system resource configurations. Compared with traditional heuristic algorithms, genetic algorithms, and random assignment strategies, the hierarchical automatic optimization method (H-EoH) maintains superior performance even with increased task numbers or enhanced resource constraints, indicating that the proposed hierarchical optimization framework can effectively improve the quality of task offloading and scheduling decisions.

[0202] Analyzing the performance changes of the policy functions generated by the large language model during algorithm iteration reveals that, as the evolutionary process continues, the newly generated policy functions (i.e., heuristic rules) continuously improve the overall fitness of the population. This indicates that the constructed closed-loop mechanism of "policy generation—performance evaluation—feedback optimization" effectively guides the algorithm's search direction, enabling heuristic rules to gradually evolve into better task offloading and scheduling strategies, rather than relying on random perturbations for improvement.

[0203] The performance fluctuations of different algorithms were statistically analyzed under repeated experimental conditions. The hierarchical automatic optimization method showed relatively small fluctuations in the results of multiple experiments, and the total system completion delay remained within a stable range. This indicates that the proposed hierarchical evolutionary search mechanism has good stability in complex search spaces and can continuously generate high-quality feasible solutions.

[0204] Multiple experiments were conducted under varying computational system parameters, including task size, number of servers, and resource distribution. The results show that the hierarchical automatic optimization method maintains good optimization performance across different experimental scenarios, adapting to new computational system environments without requiring redesigned heuristic rules or retrained models. This demonstrates that the hierarchical automatic optimization method possesses strong cross-scenario generalization ability and is applicable to dynamically changing computational system task offloading and scheduling optimization problems.

[0205] The hierarchical automatic optimization method proposed in this invention not only avoids the limitations of traditional manual rule design, but also achieves efficient solutions to high-dimensional combinatorial optimization problems through a hierarchical evolutionary search mechanism.

[0206] More importantly, this invention constructs a complete closed loop of "problem modeling - rule generation - evolutionary search - feedback optimization", which enables the algorithm to continuously improve itself and significantly enhances the performance and engineering practicality of the computing system.

[0207] This invention constructs a complete closed-loop optimization system consisting of "computational system modeling—hierarchical optimization—automatic rule generation—evolutionary search—result feedback";

[0208] Compared to traditional manually designed heuristic algorithms, this invention no longer relies on fixed rules or human experience for parameter tuning. Instead, it leverages a large language model's ability to understand the semantic information of the optimization problem to automatically generate and evolve heuristic strategies, achieving dynamic optimization at the algorithm's structural level. This mechanism enables the algorithm to adaptively adjust to different problem sizes and resource allocation conditions, significantly improving its cross-scenario generalization ability.

[0209] Compared to traditional swarm intelligence optimization algorithms, this invention decomposes the high-dimensional coupled optimization problem into two sub-problems—unloading decision and scheduling / sorting—through a hierarchical structure, effectively reducing search complexity while ensuring global goal consistency. Furthermore, by introducing rule generation and performance feedback mechanisms, the search direction is clearly guided, avoiding the slow convergence problem caused by purely random searches.

[0210] Compared to deep reinforcement learning-based methods, this invention does not require extensive pre-training with large amounts of historical data, nor does it require retraining the model for different scenarios. Through prompt-driven and online evolution mechanisms, continuous policy optimization can be achieved, striking a good balance between training costs and deployment efficiency, making it particularly suitable for dynamically changing edge computing environments.

[0211] Furthermore, this invention introduces a feasibility check and repair mechanism, which dynamically corrects or penalizes individuals that violate resource constraints during the evolutionary search process, significantly increasing the proportion of feasible solutions and enhancing the stability and engineering usability of the algorithm in resource-constrained environments.

[0212] Multi-scenario experiments have verified that this invention maintains low total system latency and high resource utilization under different task scales, server numbers, and resource distributions, while also exhibiting fast convergence speed and good load balancing. This demonstrates that this invention is not only theoretically innovative but also has high application value in practical engineering deployments.

[0213] In summary, this invention achieves intelligent and automated optimization design for task offloading and scheduling problems in multi-edge computing environments, breaking through the technical bottleneck of high costs associated with traditional manual rules that rely on model retraining. It has the following significant advantages:

[0214] It has the ability to automatically generate and evolve heuristic rules;

[0215] It possesses the ability for hierarchical collaborative optimization and high-dimensional problem decomposition;

[0216] It possesses a feasibility check and remediation mechanism under strict resource constraints;

[0217] It possesses cross-scenario generalization capabilities and rapid deployment capabilities;

[0218] Balancing performance optimization with computational efficiency.

[0219] Therefore, this invention has broad engineering application prospects in fields such as intelligent edge computing resource scheduling, industrial internet computing system optimization, vehicle network task allocation, and real-time computing service management.

[0220] The above embodiments are merely preferred embodiments of the present invention and should not be construed as limiting the scope of protection of the present invention. Any non-substantial changes and substitutions made by those skilled in the art based on the present invention shall fall within the scope of protection claimed by the present invention.

Claims

1. A hierarchical automatic optimization method for multi-edge computing task offloading and scheduling based on a large language model, characterized in that, Includes the following steps: S1. Construct a computing system, acquire task data and edge server data to form a dataset, and divide the dataset into a training set, an optimization set and a validation set; S2. Construct a joint task offloading and scheduling optimization model with the total system completion delay as the optimization objective. Minimize the total system completion delay as the optimization objective function. Use the total system completion delay obtained from the joint task offloading and scheduling optimization model as the fitness value and set constraints. S3. Construct a hierarchical optimization framework, which includes an offloading decision layer and a scheduling and sorting layer. S4. Iterative improvement of heuristic rules: Heuristic rules are generated based on a large language model. The training set is input into a hierarchical optimization framework, and a performance feedback mechanism is used to iteratively improve the heuristic rules to generate the final heuristic rules. S5. Iterative optimization of the hierarchical optimization framework: The optimization set is input into the hierarchical optimization framework, and the hierarchical optimization framework is iteratively optimized based on the hierarchical evolutionary search mechanism and the feasibility check and repair mechanism. S6. Validate the hierarchical optimization framework of iterative optimization using a validation set; S7. Input the computational task into the trained hierarchical optimization framework to generate optimized task unloading and scheduling strategies.

2. The hierarchical automatic optimization method for multi-edge computing task offloading and scheduling based on a large language model as described in claim 1, characterized in that, In step S1, the computing system includes a task set and an edge server set. Each task includes parameters for computing resource requirements, memory requirements, and data scale, while each edge server includes parameters for computing power, memory capacity, and communication bandwidth. The dataset was divided into training, optimization, and validation sets in a 4:4:2 ratio.

3. The hierarchical automatic optimization method for multi-edge computing task offloading and scheduling based on a large language model as described in claim 1, characterized in that, In step S2, the objective function to be optimized is: ,in Includes: task transmission latency Queue waiting time Actual calculation execution delay ; The constraints include edge server computing resource constraints, communication bandwidth constraints, task allocation constraints, and task execution order constraints.

4. The hierarchical automatic optimization method for multi-edge computing task offloading and scheduling based on a large language model as described in claim 1, characterized in that, In step S3, the unloading decision layer is used to generate candidate allocation schemes, and the scheduling and sorting layer is used to optimize the sorting of the task set inside each edge server based on the candidate allocation schemes generated by the unloading decision layer, and to calculate the queuing waiting time and completion time inside the edge server.

5. The hierarchical automatic optimization method for multi-edge computing task offloading and scheduling based on a large language model as described in claim 1, characterized in that, In step S4, a large language model is constructed. The large language model automatically generates new heuristic rules by receiving optimization problem description information, current optimal solution information, and historical function performance information. The generated new heuristic rules are embedded into the hierarchical optimization framework in the form of executable functions.

6. The hierarchical automatic optimization method for multi-edge computing task offloading and scheduling based on a large language model as described in claim 1, characterized in that, In step S4, the performance feedback mechanism is implemented using the following steps: The newly generated heuristic rules are evaluated for performance. The total completion time of the joint task unloading and scheduling optimization model system is used as the performance evaluation index of the heuristic rules, which is composed of the average fitness value and the proportion of feasible solutions. The evaluation results are fed back to the large language model as "historical performance" information in the next round of iterative optimization, so as to realize the iterative optimization of the heuristic rules.

7. The hierarchical automatic optimization method for multi-edge computing task offloading and scheduling based on a large language model as described in claim 1, characterized in that, In step S5, the hierarchical evolutionary search mechanism includes: performing a population search on the task allocation scheme at the unloading layer, generating candidate content through selection, crossover, and mutation operations; and optimizing the sorting of task sequences on each edge server based on heuristic rules at the scheduling and sorting layer. The feasibility check and repair mechanism involves performing constraint checks and repairs during each iteration to ensure that resource constraints are met.

8. The hierarchical automatic optimization method for multi-edge computing task offloading and scheduling based on a large language model as described in claim 1, characterized in that, In step S5, the hierarchical evolutionary search mechanism consists of the following steps: S11, Population Initialization: In the unloading decision layer, a task allocation matrix is ​​randomly generated to form an initial population, and each individual task satisfies the unique task allocation constraint. S12, Select Operation: Individual task participants are sorted according to their fitness values, and a tournament selection or roulette wheel selection mechanism is used to retain the best individual task participants. S13. Crossover and mutation operations: The offloading layer employs a crossover operator based on task subset exchange and a mutation operator based on edge server reallocation; the mutation probability and crossover probability can be dynamically adjusted according to the iteration stage. At the scheduling and sorting layer: the task execution sequence is reconstructed based on heuristic rules generated by the large language model.

9. The hierarchical automatic optimization method for multi-edge computing task offloading and scheduling based on a large language model as described in claim 1, characterized in that, In step S5, the feasibility check and repair mechanism adopts the feasibility priority principle. Based on the edge server's computing resources, tasks are redistributed to the server with the lowest load according to their individual load size. Based on communication bandwidth and communication volume, tasks with the largest data volume are prioritized for adjustment. For individuals whose tasks are still beyond repair, impose penalties to reduce their fitness.

10. The hierarchical automatic optimization method for multi-edge computing task offloading and scheduling based on a large language model as described in claim 1, characterized in that, In step S7, the optimal task unloading and scheduling strategy is output. After reaching the preset number of iterations or the convergence condition, the optimal task allocation matrix and the corresponding execution order inside the edge server are output, and the system performance index is calculated.

Images