Large language model driven self-evolution operator flexible job shop scheduling method
Patent Information
- Application Number
- CN202610961416.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-06-30
- Publication Date
- 2026-10-13
AI Technical Summary
[0088]本申请的有益效果是，通过构建大语言模型驱动的优化算子生成机制，将优化算子定义为作用于调度解空间的动态搜索变换函数，使调度算法在初始化阶段无需预设固定优化算子，从而突破传统启发式算法依赖人工经验设计算子的局限，提升方法对不同柔性作业车间调度问题的适应能力；通过构建多算子并行优化机制，使多个优化算子能够同时作用于当前调度解集，并通过解质量改进值、单位调用优化效率以及稳定性指标等性能评价参数对各优化算子进行量化评估，从而实现对高性能算子的保留以及低性能算子的淘汰，提升调度搜索过程的有效性与稳定性；通过将算子综合评分、调度问题统计特征以及历史算子性能信息反馈至大语言模型，实现优化算子的动态修改、参数调整与复合生成，使优化算子能够在搜索过程中持续进化，从而形成更适应当前调度问题特征的搜索策略，避免传统固定算子启发式算法搜索能力受限的问题；通过构建“算子生成—并行优化—性能评估—算子更新—解集更新”的闭环迭代机制，使优化算子与调度解集在搜索过程中协同演化，从而有效提升柔性作业车间调度问题的全局寻优能力、收敛效率以及求解稳定性。
Smart Images
Figure CN122905264A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of intelligent manufacturing and optimized scheduling technology, specifically to a flexible job shop scheduling method driven by a large language model and featuring self-evolving operators. Background Technology
[0002] The Flexible Job Shop Scheduling (FJSP) problem is a core aspect of manufacturing production management. Its core task is to select the appropriate processing machine for each operation of each job, given a set of jobs and a set of machines, and to determine the processing sequence of all operations to optimize production goals. In scenarios considering only machines and jobs, minimizing the maximum completion time (Makespan) is the most critical optimization objective, directly impacting production cycle time and equipment utilization.
[0003] Most existing methods for solving the FJSP are based on heuristic or metaheuristic algorithms, such as genetic algorithms and particle swarm optimization. These methods typically rely on pre-designed optimization operators (such as crossover operators, mutation operators, or neighborhood search operators) to search the solution space and gradually obtain a better solution through iterative optimization. However, these methods still have the following limitations in practical applications:
[0004] First, the design of optimization operators is highly dependent on human experience and usually requires specialized design for specific problems. The applicability of operators is poor in different scheduling scenarios, making it difficult to achieve general application. Second, the operator structure usually remains fixed during the algorithm's operation and lacks the ability to dynamically adjust according to problem characteristics and search status, which limits the algorithm's exploration ability in complex search spaces. Third, the combination of different operators and scheduling strategies often rely on manual settings, making it difficult to fully explore the synergistic effect between operators and easily leading to the search process getting stuck in local optima.
[0005] Therefore, a scheduling method is needed that can automatically generate and continuously optimize operators based on the characteristics of the scheduling problem, so as to achieve adaptive evolution at the operator level and thus improve the solution quality and efficiency of the flexible job shop scheduling problem.
[0006] Application content
[0007] The purpose of this application is to provide a flexible job shop scheduling method driven by a large language model and based on self-evolving operators, in order to solve the problems of existing heuristic algorithms in solving the flexible job shop scheduling problem (FJSP), such as the reliance on manual design of optimization operators, difficulty in dynamically adapting to different scheduling scenarios, and limited optimization capabilities, thereby effectively improving the quality of scheduling results and the efficiency of solution.
[0008] To achieve the above objectives, the technical solution adopted by this application to solve its technical problem is as follows:
[0009] A flexible job shop scheduling method driven by a large language model and featuring self-evolving operators, the method comprising the following steps:
[0010] S1: Construct a flexible workshop scheduling problem model, design relevant indicators for the model, use the maximum completion time as the scheduling evaluation indicator, and generate an initial scheduling solution set;
[0011] S2: Collect data on relevant indicators of workshop scheduling problems, and generate a set of optimization operators based on the large language model;
[0012] S3: Perform multi-operator parallel optimization operation, apply the set of optimization operators to the current scheduling solution set to obtain a new scheduling solution, and obtain the maximum change in completion time before and after the action of each optimization operator based on the new scheduling solution;
[0013] S4: Evaluate the performance of each optimization operator and obtain the comprehensive operator score;
[0014] S5: Based on the comprehensive score of the operators, the optimized operators are updated to obtain a new set of optimized operators;
[0015] S6: Update the scheduling solution set based on the new set of optimization operators and determine the termination condition;
[0016] S7: Output the optimal scheduling scheme, that is, output the scheduling scheme with the minimum maximum completion time in the current scheduling solution set as the final result.
[0017] Furthermore, S1 specifically includes:
[0018] S1.1: Construct the initial scheduling solution set:
[0019] ;
[0020] in, This represents the initial scheduling solution set. The number of initial solutions. Indicates the first One initial scheduling solution;
[0021] S1.2: Initialize the optimization operator set container:
[0022] ;
[0023] in, This represents the initial set of optimization operators, where no specific optimization operators are preset in the initial stage.
[0024] S1.3: Based on the maximum completion time The scheduling evaluation index is calculated using the following formula:
[0025]
[0026] in, Represents a scheduling solution. Indicates workpiece The completion time of the final process. The total number of workpieces. For workpiece The number of processes.
[0027] Furthermore, S2 specifically includes:
[0028] S2.1: Constructing Input Information for a Large Language Model
[0029]
[0030] in, Indicates the first The information set input into the large language model, Information indicating the scale of the scheduling problem includes at least the number of jobs, the number of machines, the number of processes, the set of available machines, and the processing time. It represents the statistical characteristics of the current scheduling solution set, including at least the current best maximum completion time, the average maximum completion time, and the degree of fluctuation in the solution set; It represents historical operator performance information, including historical improvement values, invocation efficiency, and overall score for each operator;
[0031] S2.2: The large language model determines the input information. generate A set of candidate optimization operators is formed.
[0032]
[0033] in, This represents the operator generation function driven by the large language model. Indicates the first The set of optimization operators Indicates the first The first generation An optimization operator;
[0034] S2.3: Define each optimization operator as a transformation function acting on the scheduling solution space:
[0035]
[0036] in, Indicates the first The first generation A scheduling solution Indicates after the first The candidate new scheduling solutions generated after the application of the optimization operators.
[0037] Furthermore, S3 specifically includes:
[0038] S3.1: Optimize the set of operators Each optimization operator in the algorithm acts on the current scheduling solution set. This yields a new subset of scheduling solutions corresponding to each operator:
[0039]
[0040] in, Indicates the first Dispatch set of the scheduler Indicates the first A new subset of scheduling solutions generated by an optimization operator;
[0041] S3.2: For any scheduling solution in the scheduling solution set , No. The process of an optimization operator is defined as follows:
[0042]
[0043] in, Indicates the first The operator parameters of an optimization operator include at least the target, the scope of action, and the adjustment intensity; the operation mode of the optimization operator includes process exchange, process insertion, machine reallocation, bottleneck process migration, or a combination of the above methods;
[0044] S3.3: Record the maximum change in completion time before and after each optimization operator is applied. The calculation formula is as follows:
[0045]
[0046] in, Indicates the first The optimization operator for the first... The improvement margin of each scheduling solution; when When the operator improves the quality of the scheduling solution, it indicates that the operator improves the quality of the scheduling solution.
[0047] Furthermore, S4 specifically includes:
[0048] S4.1: Computation Operator Solution quality improvement value:
[0049]
[0050] in, Indicates the first The average effective improvement of each optimization operator, Indicates the first The number of times each optimization operator is applied;
[0051] S4.2: Calculate the efficiency of operator unit call optimization:
[0052]
[0053] in, Indicates the first The optimization efficiency per unit time of each optimization operator. Indicates the first The cumulative running time of each optimization operator;
[0054] S4.3: Calculate the operator stability index, the formula of which is:
[0055]
[0056] in, Indicates the first Stability metrics for optimization operators Indicates the first The average improvement of each optimization operator;
[0057] S4.4: Normalize the above indicators to obtain... , and And construct the operator comprehensive scoring function:
[0058]
[0059] in, Indicates the first The overall score of the optimization operators , , These are the weighting coefficients.
[0060] Furthermore, S5 specifically includes:
[0061] S5.1: Retain the preferred operator based on the comprehensive operator score:
[0062]
[0063] in, Denotes the set of preferred operators to be retained. Indicates the operator retention threshold;
[0064] S5.2: Eliminate low-performance operators based on comprehensive operator scores:
[0065] $
[0066] in, This represents the set of low-performance operators that are being phased out. This represents the operator elimination threshold, and ;
[0067] S5.3: Input the structural information, operator parameters, and comprehensive score of the optimized operator into the large language model to generate the improved operator:
[0068]
[0069] in, Indicates by the first The improved operator obtained by modifying the preferred operator;
[0070] S5.4: Select two or more operators from the preferred set of operators, and generate composite operators from the large language model:
[0071]
[0072] in, and Represent two optimization operators, This represents operator combination operations. This represents the newly generated composite operator;
[0073] S5.5: Forming a new set of optimization operators:
[0074]
[0075] in, This represents the set of improved operators generated by modifying the preferred operator. This represents the set of new composite operators generated by the combination of preferred operators.
[0076] Furthermore, S6 specifically includes:
[0077] S6.1: Merge the new scheduling solutions generated by all optimization operators to obtain a set of candidate scheduling solutions:
[0078] ;
[0079] S6.2: Select the solution with the optimal objective function from the candidate scheduling solution set. Each solution forms a new set of scheduling solutions:
[0080] ;
[0081] in, This indicates that the first [time] is selected after sorting by the longest completion time in ascending order. One scheduling solution;
[0082] S6.3: Determine if the iteration termination condition is met. The iteration terminates when any of the following conditions are met:
[0083] or ;
[0084] in, This represents the threshold for the maximum number of iterations. Indicates the first The optimal objective function value of the generation, This represents the interval of historical iterations used to determine convergence. This represents the convergence accuracy threshold; if the termination condition is not met, return to S3 to continue execution.
[0085] Output the scheduling scheme with the minimum maximum completion time in the current scheduling solution set as the final scheduling result:
[0086]
[0087] in, This represents the optimal scheduling scheme in the final output.
[0088] The beneficial effects of this application are as follows: By constructing a large language model-driven optimization operator generation mechanism, the optimization operator is defined as a dynamic search transformation function acting on the scheduling solution space. This eliminates the need for pre-setting fixed optimization operators during the initialization phase of the scheduling algorithm, thus overcoming the limitation of traditional heuristic algorithms that rely on manual experience to design operators and improving the method's adaptability to different flexible job shop scheduling problems. Furthermore, by constructing a multi-operator parallel optimization mechanism, multiple optimization operators can be applied simultaneously to the current scheduling solution set. Performance evaluation parameters such as solution quality improvement value, unit call optimization efficiency, and stability index are used to quantitatively evaluate each optimization operator, thereby achieving the retention of high-performance operators and the elimination of low-performance operators, improving the efficiency of scheduling. This method enhances the effectiveness and stability of the scheduling search process. By feeding back operator comprehensive scoring, scheduling problem statistical characteristics, and historical operator performance information to a large language model, it enables dynamic modification, parameter adjustment, and composite generation of optimization operators. This allows optimization operators to continuously evolve during the search process, forming search strategies that are more adapted to the characteristics of the current scheduling problem, thus avoiding the limitation of traditional fixed-operator heuristic algorithms in terms of search capability. Furthermore, by constructing a closed-loop iterative mechanism of "operator generation—parallel optimization—performance evaluation—operator update—solution set update," optimization operators and scheduling solutions co-evolve during the search process, thereby effectively improving the global optimization capability, convergence efficiency, and solution stability of the flexible job shop scheduling problem. Attached Figure Description
[0089] The present application will be further described below with reference to the accompanying drawings and embodiments:
[0090] Figure 1This application includes a flowchart of the method steps.
[0091] Figure 2 System log diagram for generating initial optimization operators;
[0092] Figure 3 Optimize the log diagram of the operator's self-evolution process. Detailed Implementation
[0093] The purpose of this application is to provide a large language model-driven self-evolving operator scheduling optimization method for solving flexible job shop scheduling. This method can overcome the problems of traditional heuristic algorithms in solving flexible job shop scheduling problems, such as the reliance on human experience in designing optimization operators, fixed optimization strategies, and insufficient global optimization capabilities, thereby improving the quality and efficiency of the algorithm in solving flexible job shop scheduling tasks.
[0094] This application proposes a flexible job shop scheduling method driven by a large language model and featuring self-evolving operators. First, a flexible job shop scheduling problem model is constructed, and then solved within a swarm heuristic optimization framework. The specific parameters of the problem model are as follows:
[0095] (1) Core parameters:
[0096] Workpiece set ,in This represents the total number of workpieces.
[0097] machine set ,in This represents the total number of machines.
[0098] Each workpiece Include One process;
[0099] Process The optional machine set is In the machine The processing time is ;
[0100] variable Indicates the start time of the process. Indicates the completion time.
[0101] (2) Constraints:
[0102] Process sequence constraint: Processes on the same workpiece must be processed in a predetermined order;
[0103] Machine exclusivity constraint: Each machine can only process one operation at any given time;
[0104] Machine applicability constraints: Each process can only be performed on the set of machines available to it.
[0105] The optimization objective is to minimize the maximum completion time Makespan, i.e.:
[0106] .
[0107] The technical solution adopted by this application to solve its technical problem is:
[0108] The flowchart of this application is as follows: Figure 1 As shown, this application provides a flexible job shop scheduling method driven by a large language model and featuring self-evolving operators, comprising the following steps:
[0109] S1: Construct a flexible job shop scheduling problem model, design relevant model indicators, use the maximum completion time as the scheduling evaluation indicator, and generate an initial scheduling solution set; after generation, execute S2; specifically including:
[0110] S1.1: Construct the initial scheduling solution set:
[0111] ;
[0112] in, This represents the initial scheduling solution set. The number of initial solutions. Indicates the first One initial scheduling solution;
[0113] S1.2: Initialize the optimization operator set container:
[0114] ;
[0115] in, This represents the initial set of optimization operators, where no specific optimization operators are preset in the initial stage.
[0116] S1.3: Based on the maximum completion time The scheduling evaluation index is calculated using the following formula:
[0117] ;
[0118] in, Represents a scheduling solution. Indicates workpiece The completion time of the final process. The total number of workpieces. For workpiece The number of processes.
[0119] S2: Collect data on relevant indicators of job shop scheduling, and generate a set of optimization operators based on a large language model; after generation, execute S3; specifically including:
[0120] S2.1: Constructing input information for a large language model:
[0121] ;
[0122] in, Indicates the first The information set input into the large language model, Information indicating the scale of the scheduling problem includes at least the number of jobs, the number of machines, the number of processes, the set of available machines, and the processing time. It represents the statistical characteristics of the current scheduling solution set, including at least the current best maximum completion time, the average maximum completion time, and the degree of fluctuation in the solution set; It represents historical operator performance information, including historical improvement values, invocation efficiency, and overall score for each operator;
[0123] S2.2: The large language model determines the input information. generate A set of candidate optimization operators is formed.
[0124] ;
[0125] in, This represents the operator generation function driven by the large language model. Indicates the first The set of optimization operators Indicates the first The first generation An optimization operator;
[0126] S2.3: Define each optimization operator as a transformation function acting on the scheduling solution space:
[0127] ;
[0128] in, Indicates the first The first generation A scheduling solution Indicates after the first The candidate new scheduling solutions generated after the application of the optimization operators.
[0129] S3: Perform multi-operator parallel optimization operations, applying the set of optimization operators to the current scheduling solution set to obtain a new scheduling solution, and obtaining the maximum change in completion time before and after the application of each optimization operator based on the new scheduling solution; after execution, execute S4; specifically including:
[0130] S3.1: Optimize the set of operators Each optimization operator in the algorithm acts on the current scheduling solution set. This yields a new subset of scheduling solutions corresponding to each operator:
[0131] ;
[0132] in, Indicates the first Dispatch set of the scheduler Indicates the first A new subset of scheduling solutions generated by an optimization operator;
[0133] S3.2: For any scheduling solution in the scheduling solution set , No. The process of an optimization operator is defined as follows:
[0134] ;
[0135] in, Indicates the first The operator parameters of an optimization operator include at least the target, the scope of action, and the adjustment intensity; the operation mode of the optimization operator includes process exchange, process insertion, machine reallocation, bottleneck process migration, or a combination of the above methods;
[0136] S3.3: Record the maximum change in completion time before and after each optimization operator is applied. The calculation formula is as follows:
[0137] ;
[0138] in, Indicates the first The optimization operator for the first... The improvement margin of each scheduling solution. When When the operator improves the quality of the scheduling solution, it indicates that the operator improves the quality of the scheduling solution.
[0139] S4: Evaluate the performance of each optimization operator to obtain a comprehensive operator score. After the evaluation is completed, execute S5; specifically including:
[0140] S4.1: Computation Operator Solution quality improvement value:
[0141] ;
[0142] in, Indicates the first The average effective improvement of each optimization operator, Indicates the first The number of times each optimization operator is applied;
[0143] S4.2: Calculate the efficiency of operator unit call optimization:
[0144] ;
[0145] in, Indicates the first The optimization efficiency per unit time of each optimization operator. Indicates the first The cumulative running time of each optimization operator;
[0146] S4.3: Calculate the operator stability index, the formula of which is:
[0147] ;
[0148] in, Indicates the first Stability metrics for optimization operators Indicates the first The average improvement of each optimization operator;
[0149] S4.4: Normalize the above indicators to obtain... , and And construct the operator comprehensive scoring function:
[0150] ;
[0151] in, Indicates the first The overall score of the optimization operators , , Let be the weight coefficient, and satisfy...
[0152] S5: Based on the comprehensive score of the operators, optimize and update the operators to obtain a new set of optimized operators; after the update is completed, execute S6; specifically including:
[0153] S5.1: Retain the preferred operator based on the comprehensive operator score:
[0154] ;
[0155] in, Denotes the set of preferred operators to be retained. Indicates the operator retention threshold;
[0156] S5.2: Eliminate low-performance operators based on comprehensive operator scores:
[0157] ;
[0158] in, This represents the set of low-performance operators that are being phased out. This represents the operator elimination threshold, and ;
[0159] S5.3: Input the structural information, operator parameters, and comprehensive score of the optimized operator into the large language model to generate the improved operator:
[0160] ;
[0161] in, Indicates by the first The improved operator obtained by modifying the preferred operator;
[0162] S5.4: Select two or more operators from the preferred set of operators, and generate composite operators from the large language model:
[0163] ;
[0164] in, and Represent two optimization operators, This represents operator combination operations. This represents the newly generated composite operator;
[0165] S5.5: Forming a new set of optimization operators:
[0166] ;
[0167] in, This represents the set of improved operators generated by modifying the preferred operator. This represents the set of new composite operators generated by the combination of preferred operators.
[0168] S6: Update the scheduling solution set based on the new set of optimization operators and determine the termination condition; specifically including:
[0169] S6.1: Merge the new scheduling solutions generated by all optimization operators to obtain a set of candidate scheduling solutions:
[0170] ;
[0171] S6.2: Select the solution with the optimal objective function from the candidate scheduling solution set. Each solution forms a new set of scheduling solutions:
[0172] ;
[0173] in, This indicates that the first [time] is selected after sorting by the longest completion time in ascending order. One scheduling solution;
[0174] S6.3: Determine if the iteration termination condition is met. The iteration terminates when any of the following conditions are met:
[0175] or ;
[0176] in, This represents the threshold for the maximum number of iterations. Indicates the first The optimal objective function value of the generation, This represents the interval of historical iterations used to determine convergence. This represents the convergence accuracy threshold; if the termination condition is not met, return to S3 to continue execution.
[0177] S7: Output the optimal scheduling scheme, that is, output the scheduling scheme with the minimum maximum completion time in the current scheduling solution set as the final result.
[0178] Output the scheduling scheme with the minimum maximum completion time in the current scheduling solution set as the final scheduling result:
[0179] ;
[0180] in, This represents the optimal scheduling scheme in the final output.
[0181] The beneficial effects of this application are as follows: By constructing a large language model-driven optimization operator generation mechanism, the optimization operator is defined as a dynamic search transformation function acting on the scheduling solution space. This eliminates the need for pre-setting fixed optimization operators during the initialization phase of the scheduling algorithm, thus overcoming the limitation of traditional heuristic algorithms that rely on manual experience to design operators and improving the method's adaptability to different flexible job shop scheduling problems. Furthermore, by constructing a multi-operator parallel optimization mechanism, multiple optimization operators can be applied simultaneously to the current scheduling solution set. Performance evaluation parameters such as solution quality improvement value, unit call optimization efficiency, and stability index are used to quantitatively evaluate each optimization operator, thereby achieving the retention of high-performance operators and the elimination of low-performance operators, improving the efficiency of scheduling. This method enhances the effectiveness and stability of the scheduling search process. By feeding back operator comprehensive scoring, scheduling problem statistical characteristics, and historical operator performance information to a large language model, it enables dynamic modification, parameter adjustment, and composite generation of optimization operators. This allows optimization operators to continuously evolve during the search process, forming search strategies that are more adapted to the characteristics of the current scheduling problem, thus avoiding the limitation of traditional fixed-operator heuristic algorithms in terms of search capability. Furthermore, by constructing a closed-loop iterative mechanism of "operator generation—parallel optimization—performance evaluation—operator update—solution set update," optimization operators and scheduling solutions co-evolve during the search process, thereby effectively improving the global optimization capability, convergence efficiency, and solution stability of the flexible job shop scheduling problem.
[0182] The effectiveness of this application can be further illustrated by the following simulation experiments.
[0183] To verify the effectiveness of the proposed method in the flexible job shop scheduling problem, the proposed large language model-driven self-evolutionary operator flexible job shop scheduling method was implemented in Python, and comparative experiments were conducted with several typical heuristic algorithms. The comparative algorithms included Genetic Algorithm (GA), Particle Swarm Optimization (PSO), Ant Colony Optimization (ACO), and Tabu Search (TS). A classic flexible job shop scheduling test case was selected as the research object in the experiment, and its set of processing machines and processing time data are shown in Tables 1 and 2, respectively. The main experimental parameters were set as follows: scheduling solution set size... Maximum number of iterations Initial number of optimization operators Operator retention threshold Operator elimination threshold The weight coefficients of the operator evaluation function are respectively ; ; Operator stability statistics count Convergence criterion algebra Convergence threshold The experimental results are shown in Table 3.
[0184] Table 1 Processing Machines
[0185]
[0186] Table 2 Processing Time Table
[0187]
[0188] Table 3 Comparison of Simulation Results
[0189]
[0190] As can be seen from the data in Table 3, different traditional heuristic algorithms exhibit significantly different optimization characteristics when solving this scheduling instance. For example, the particle swarm optimization algorithm can quickly reduce the scheduling target value in the early stages of the search, but due to the lack of an effective local search mechanism, its improvement in the later stages is relatively small; the ant colony optimization algorithm performs well in terms of search stability, but due to the slow pheromone update speed, its overall convergence speed is relatively low; the tabu search algorithm has strong local search capabilities, but it is quite sensitive to the quality of the initial solution and is prone to getting trapped in local optima when the search space is large.
[0191] In comparison, the method proposed in this application demonstrates both superior search efficiency and optimization capability during the solution process. Specifically, the method converges to a stable solution in the 14th generation, while the traditional genetic algorithm requires 27 generations to reach convergence. Regarding the quality of the final solution, the method achieved a maximum completion time of 16 seconds, which is superior to the 18 seconds of the genetic algorithm, 19 seconds of the particle swarm optimization algorithm, 20 seconds of the ant colony optimization algorithm, and 19 seconds of the tabu search algorithm. Furthermore, the average initial solution time of the method proposed in this application is 21.47 seconds, significantly lower than the other comparative algorithms, indicating that the initial operator generated by the large language model effectively guides the search process.
[0192] To further illustrate the actual operation of the operator generation and evolution mechanism in the method of this application, Figure 2 and Figure 3 The system's runtime logs for the initialization phase and the operator evolution phase are presented separately.
[0193] in, Figure 2 The system's operation during the initialization phase is demonstrated. First, the system generates an initial set of scheduling solutions. Based on this, it constructs an operator generation request by combining the scale information of the scheduling problem and the statistical characteristics of the current set of solutions. Then, it calls a large language model to generate multiple candidate optimization operators. Subsequently, the system performs interface verification and filtering on the generated optimization operators, forming an initial set of optimization operators and adding it to the operator pool for subsequent scheduling optimization processes.
[0194] Figure 3 The process of operator evolution is illustrated. After the system performs parallel optimization on the scheduling solution set using the current set of optimized operators, it further evaluates the scheduling effect of each optimized operator and calculates the corresponding performance evaluation index and comprehensive score. Based on this, the system updates the set of optimized operators according to the operator performance evaluation results, including retaining high-performance operators, eliminating low-performance operators, and calling the large language model to generate improved operators or new composite operators, thus forming an updated set of optimized operators.
[0195] Through the cyclical mechanism of initial solution generation—operator generation—parallel optimization—performance evaluation—operator update—solution set update, the method in this application can continuously improve the search strategy during the optimization process, enabling the optimization operator to gradually adapt to the characteristics of the current scheduling problem and continuously improve the quality of the scheduling solution. This application realizes the automatic generation and dynamic evolution of the optimization operator, avoiding dependence on fixed optimization strategies, and can effectively improve the global optimization capability, convergence efficiency, and solution stability of the flexible job shop scheduling problem. It is suitable for scheduling optimization problems in complex dynamic manufacturing environments. The experimental results in Table 3 show that the method in this application outperforms the traditional fixed-operator heuristic algorithm in terms of final solution quality and convergence efficiency, thus verifying the effectiveness of the method.
[0196] The above description is merely a preferred embodiment of this application and does not constitute any limitation on this application. Any person skilled in the art can make many possible variations and modifications to the technical solution of this application, or modify it into equivalent embodiments, without departing from the scope of the technical solution of this application, using the methods and techniques disclosed above. Therefore, any simple modifications, equivalent substitutions, equivalent changes, and modifications made to the above embodiments based on the technical essence of this application, without departing from the content of the technical solution of this application, shall still fall within the protection scope of the technical solution of this application.
Claims
1. A flexible job shop scheduling method driven by a large language model and featuring self-evolving operators, characterized in that... The method includes the following steps: S1: Construct a flexible workshop scheduling problem model, design relevant indicators for the model, use the maximum completion time as the scheduling evaluation indicator, and generate an initial scheduling solution set; S2: Collect data on relevant indicators of workshop scheduling problems, and generate a set of optimization operators based on the large language model; S3: Perform multi-operator parallel optimization operation, apply the set of optimization operators to the current scheduling solution set to obtain a new scheduling solution, and obtain the maximum change in completion time before and after the action of each optimization operator based on the new scheduling solution; S4: Evaluate the performance of each optimization operator and obtain the comprehensive operator score; S5: Based on the comprehensive score of the operators, the optimized operators are updated to obtain a new set of optimized operators; S6: Update the scheduling solution set based on the new set of optimization operators and determine the termination condition; S7: Output the optimal scheduling scheme, that is, output the scheduling scheme with the minimum maximum completion time in the current scheduling solution set as the final result.
2. The method according to claim 1, characterized in that, S1 specifically includes: S1.1: Construct the initial scheduling solution set: ； in, This represents the initial scheduling solution set. The number of initial solutions. Indicates the first One initial scheduling solution; S1.2: Initialize the optimization operator set container: ； in, This represents the initial set of optimization operators, where no specific optimization operators are preset in the initial stage. S1.3: Based on the maximum completion time The scheduling evaluation index is calculated using the following formula: ； in, Represents a scheduling solution. Indicates workpiece The completion time of the final process. The total number of workpieces. For workpiece The number of processes.
3. The method according to claim 2, characterized in that, S2 specifically includes: S2.1: Constructing Input Information for a Large Language Model ； in, Indicates the first The information set input into the large language model, Information indicating the scale of the scheduling problem includes at least the number of jobs, the number of machines, the number of processes, the set of available machines, and the processing time. It represents the statistical characteristics of the current scheduling solution set, including at least the current best maximum completion time, the average maximum completion time, and the degree of fluctuation in the solution set; It represents historical operator performance information, including historical improvement values, invocation efficiency, and overall score for each operator; S2.2: The large language model determines the input information. generate A set of candidate optimization operators is formed. ； in, This represents the operator generation function driven by the large language model. Indicates the first The set of optimization operators Indicates the first The first generation An optimization operator; S2.3: Define each optimization operator as a transformation function acting on the scheduling solution space: ； in, Indicates the first The first generation A scheduling solution Indicates after the first The candidate new scheduling solutions generated after the application of the optimization operators.
4. The method according to claim 3, characterized in that, S3 specifically includes: S3.1: Optimize the set of operators Each optimization operator in the algorithm acts on the current scheduling solution set. This yields a new subset of scheduling solutions corresponding to each operator: ； in, Indicates the first Dispatch set of the scheduler Indicates the first A new subset of scheduling solutions generated by an optimization operator; S3.2: For any scheduling solution in the scheduling solution set , No. The process of an optimization operator is defined as follows: ； in, Indicates the first The operator parameters of an optimization operator include at least the target, the scope of action, and the adjustment intensity; the operation mode of the optimization operator includes process exchange, process insertion, machine reallocation, bottleneck process migration, or a combination of the above methods; S3.3: Record the maximum change in completion time before and after each optimization operator is applied. The calculation formula is as follows: ； in, Indicates the first The optimization operator for the first... The improvement margin of each scheduling solution; when When the operator improves the quality of the scheduling solution, it indicates that the operator improves the quality of the scheduling solution.
5. The method according to claim 4, characterized in that, S4 specifically includes: S4.1: Computation Operator Solution quality improvement value: ； in, Indicates the first The average effective improvement of each optimization operator, Indicates the first The number of times each optimization operator is applied; S4.2: Calculate the efficiency of operator unit call optimization: ； in, Indicates the first The optimization efficiency per unit time of each optimization operator. Indicates the first The cumulative running time of each optimization operator; S4.3: Calculate the operator stability index, the formula of which is: ； in, Indicates the first Stability metrics for optimization operators Indicates the first The average improvement of each optimization operator; S4.4: Normalize the above indicators to obtain... , and And construct the operator comprehensive scoring function: ； in, Indicates the first The overall score of the optimization operators , , These are the weighting coefficients.
6. The method according to claim 5, characterized in that, S5 specifically includes: S5.1: Retain the preferred operator based on the comprehensive operator score: ； in, Denotes the set of preferred operators to be retained. Indicates the operator retention threshold; S5.2: Eliminate low-performance operators based on comprehensive operator scores: ； in, This represents the set of low-performance operators that are being phased out. This represents the operator elimination threshold, and ; S5.3: Input the structural information, operator parameters, and comprehensive score of the optimized operator into the large language model to generate the improved operator: ； in, Indicates by the first The improved operator obtained by modifying the preferred operator; S5.4: Select two or more operators from the preferred set of operators, and generate composite operators from the large language model: ； in, and Represent two optimization operators, This represents operator combination operations. This represents the newly generated composite operator; S5.5: Forming a new set of optimization operators: ； in, This represents the set of improved operators generated by modifying the preferred operator. This represents the set of new composite operators generated by the combination of preferred operators.
7. The method according to claim 6, characterized in that, S6 specifically includes: S6.1: Merge the new scheduling solutions generated by all optimization operators to obtain a set of candidate scheduling solutions: ； S6.2: Select the solution with the optimal objective function from the candidate scheduling solution set. Each solution forms a new set of scheduling solutions: ； in, This indicates that the first [time] is selected after sorting by the longest completion time in ascending order. One scheduling solution; S6.3: Determine if the iteration termination condition is met. The iteration terminates when any of the following conditions are met: or ; in, This represents the threshold for the maximum number of iterations. Indicates the first The optimal objective function value of the generation, This represents the interval of historical iterations used to determine convergence. This represents the convergence accuracy threshold; if the termination condition is not met, return to S3 to continue execution.
8. The method according to claim 7, characterized in that, S7 specifically includes: outputting the scheduling scheme with the minimum maximum completion time in the current scheduling solution set as the final scheduling result: ； in, This represents the optimal scheduling scheme in the final output.