LLM and reinforcement learning method and system for AGV flexible workshop scheduling

Abstract

The application belongs to the technical field of intelligent manufacturing and production scheduling, and particularly relates to an LLM and reinforcement learning method and system for AGV flexible workshop scheduling, which comprises the following steps: (1) multi-objective cooperative initialization: initial solutions with advantages in maximum completion time and total energy consumption are generated to form an initial population; (2) multi-objective search guided by LLM: the distribution state of the population in the multi-objective space is analyzed by a large language model, and the crossover and mutation strategies are dynamically adjusted to guide the population to approach the Pareto optimal frontier; (3) multi-objective local search driven by Q-learning: based on the Pareto level and target value of the current solution, an energy-saving neighborhood operation is adaptively selected by Q-learning for local optimization; (4) multi-objective feedback and population cooperation: the multi-objective performance indicators of the population are monitored in real time, the LLM guidance strategy and the Q-learning exploration rate are dynamically adjusted according to the feedback, and the convergence and diversity of the population are maintained.

Description

Technical Field

[0001] This invention belongs to the field of intelligent manufacturing and production scheduling technology, specifically an LLM and reinforcement learning method and system for scheduling flexible workshops with AGVs, applicable to multi-objective optimization problems with material transportation constraints in flexible workshops. Background Technology

[0002] As the global manufacturing industry accelerates its transformation towards intelligent and green practices, the Flexible Job Shop Scheduling Problem (FJSP), a core combinatorial optimization problem in production systems, is becoming increasingly complex and important. In practical manufacturing systems, finite transportation resources such as Automated Guided Vehicles (AGVs) are widely used to transfer workpieces between machines to achieve collaborative production. However, most current research either simplifies transportation time as part of processing time or assumes unlimited transportation resources. This idealized approach severely weakens the practical application value of scheduling solutions. In fact, there is a strong coupling relationship between processing tasks and transportation tasks: the preceding processing determines the starting point and start time of transportation, while the completion of transportation tasks directly affects the initiation of subsequent processing. Therefore, comprehensively considering the Flexible Job Shop Scheduling Problem with finite AGVs (FJSP with automatic guided vehicles, FJSP-AGV), i.e., simultaneously solving the four sub-problems of machine allocation, process sequencing, AGV allocation, and AGV scheduling, is crucial for improving the overall efficiency of manufacturing systems.

[0003] Due to its NP-hard nature and complex inter-problem constraints, the solution space of FJSP-AGV is exceptionally large and complex. Existing research still has the following obvious limitations in addressing this challenge: (1) Insufficient model integration and limited research scale: Early research focused on small-scale problem instances, with insufficient exploration of medium and large-scale problems that are more relevant to reality. Although metaheuristic methods such as Iterative Local Search (ILS) and Hybrid Genetic Algorithm (IGA) have emerged in recent years, their optimization performance is still insufficient when facing large-scale problems, and it is difficult to obtain satisfactory solutions within an acceptable time. (2) Lack of algorithm adaptation and learning ability: Traditional evolutionary algorithms (EA) and their variants usually adopt fixed search strategies and parameter settings when solving FJSP-AGV. During the search process, they are difficult to adaptively balance "exploration" and "exploitation" according to the problem characteristics and evolutionary state, and are prone to getting trapped in local optima and have poor convergence speed. Although hyperheuristic algorithms improve adaptability by managing underlying heuristic rules through high-level frameworks, the design of their rule selection strategies still heavily relies on domain knowledge and lacks the ability to learn and improve from the search process. (3) The dilemma of multi-objective collaborative optimization: Minimizing the maximum completion time (Makespan) and minimizing the total energy consumption (TEC) are inherently conflicting. Shortening the manufacturing cycle may require AGVs to run at high speed and machines to operate at high power, thereby increasing energy consumption; conversely, slowing down for energy saving will slow down the production progress. Existing methods mostly use linear weighting to transform multi-objectives into single-objectives, but the weight setting is subjective and it is difficult to obtain a uniformly distributed Pareto optimal solution set, which cannot provide decision-makers with diverse choices. (4) Insufficient depth of intelligent technology integration: Although reinforcement learning (RL) has been introduced to adaptively select neighborhood search operators or adjust parameters, existing reinforcement learning methods often have limited effectiveness in the core link of crossover / mutation in evolutionary algorithms. These methods often fail to adequately consider the overall state of the population and the multi-objective attributes of individuals, making it difficult for the selected paired individuals to efficiently cooperate in producing high-quality offspring, thus limiting further improvements in the overall performance of the algorithm. Meanwhile, large language models (LLMs) possess excellent context awareness and reasoning capabilities, but how to deeply embed them into the decision loop of optimization algorithms, dynamically analyze the population state, and guide the evolutionary direction, rather than merely using them as peripheral tools, remains an unexplored frontier.

[0004] In recent years, the integration of machine learning and evolutionary algorithms has become an emerging research trend. However, effective integration models of RL and EA still need to be further developed, especially in terms of fully leveraging the learning mechanisms to mine the information generated during the search process, which has great potential.

[0005] Therefore, to address the above problems, there is an urgent need for a new solution that can deeply integrate the decision-making capabilities of advanced AI models with the global search framework of evolutionary algorithms. Summary of the Invention

[0006] To address the problems of fragmented processing and transportation scheduling, insufficient utilization of cross-resource information, and low stability of scheduling results in existing technologies, this invention provides an LLM and reinforcement learning method and system for flexible workshop scheduling with AGVs.

[0007] To achieve the above objectives, the technical solution of the present invention is as follows:

[0008] An LLM and reinforcement learning method for flexible shop floor scheduling with AGVs, characterized by the following steps:

[0009] (1) Multi-objective collaborative initialization: Generate initial solutions that have advantages in both maximum completion time and total energy consumption objectives, and form an initial population;

[0010] (2) LLM-guided multi-objective search: By analyzing the distribution of the population in the multi-objective space through a large language model, the crossover and mutation strategies are dynamically adjusted to guide the population to approach the Pareto optimal frontier;

[0011] (3) Q-learning-driven multi-objective local search: Based on the Pareto level of the current solution and the objective value, Q-learning is used to adaptively select energy-saving neighborhood operations for local optimization;

[0012] (4) Multi-objective feedback and population collaboration: Real-time monitoring of the population's multi-objective performance indicators, and dynamic adjustment of the LLM guidance strategy and Q-learning exploration rate based on feedback to maintain the population's convergence and diversity.

[0013] Preferably, the multi-objective collaborative initialization in step (1) includes: generating a solution that adopts the earliest completion rule and optimizes the maximum completion time; generating a solution that adopts the lowest energy consumption machine and AGV allocation rule and optimizes the total energy consumption; generating an insertion heuristic equilibrium solution based on weighted sum evaluation; and generating several random solutions to construct a well-distributed initial population.

[0014] Preferably, the LLM-guided multi-objective search in step (2) includes: using Pareto solution set distribution breadth, solution set improvement degree, and diversity index as input; the large language model infers and outputs high-level strategy suggestions based on the input; and dynamically adjusting the crossover and mutation strategies according to the suggestions to achieve intelligent search.

[0015] Preferably, the Q-learning-driven multi-objective local search in step (3) includes: the state is defined as the Pareto hierarchy of the solution and the normalized value of the objective; the action space includes efficiency-oriented and energy-efficiency-oriented neighborhood operations; the reward function is designed based on the Pareto dominance relationship; and the optimal operation is selected for local optimization through Q-learning.

[0016] Preferably, the multi-objective feedback and population collaboration in step (4) includes: periodically calculating the hypervolume index of the Pareto solution set; generating a feedback signal when optimization stagnation is detected; increasing the mutation probability and Q-learning exploration rate; and performing environmental selection based on non-dominated sorting and crowding distance to maintain population diversity.

[0017] An LLM and reinforcement learning system for flexible workshop scheduling with AGVs includes: a multi-objective collaborative initialization module for generating a well-distributed initial population; an LLM-guided multi-objective search module for dynamically adjusting the evolutionary strategy based on the population state; a Q-learning-driven multi-objective local search module for adaptively selecting energy-efficient neighborhood operations; and a multi-objective feedback and population collaboration module for monitoring the optimization process and dynamically adjusting system parameters.

[0018] Preferably, the LLM-guided multi-objective search module is based on a large language model, which takes the multi-objective state information of the population as input and outputs high-level search strategy suggestions.

[0019] Preferably, the Q-learning-driven multi-objective local search module adopts a state-action-reward mechanism, combining the Pareto characteristics of the solution to select the optimal local operation.

[0020] Preferably, the multi-objective feedback and population collaboration module monitors the optimization process through hypervolume indicators and generates feedback signals when stagnation is detected, dynamically adjusting the exploration strategy.

[0021] Preferably, the system adopts a three-segment coding scheme to represent the scheduling solution, including process coding, machine allocation coding and AGV allocation coding, which fully expresses the collaborative scheduling scheme of processing and transportation.

[0022] Compared with the prior art, the present invention has the following advantages:

[0023] (1) This invention precisely defines the multi-objective optimization problem of FJSP-AGV and constructs a hybrid integer programming model that simultaneously considers the maximum completion time and total energy consumption, providing clear optimization criteria for green manufacturing.

[0024] (2) This invention pioneered an LLM-guided multi-objective search framework, enabling the algorithm to understand complex multi-objective search states and make high-level strategic decisions, realizing intelligent and adaptive search of Pareto fronts, and breaking through the limitations of fixed parameters in traditional algorithms.

[0025] (3) The Q-learning multi-objective local search module designed in this invention can adaptively select the most effective energy-saving scheduling operation according to the Pareto characteristics of the solution, realize the synergistic optimization of efficiency and energy efficiency at the micro level, and significantly improve the accuracy of local search.

[0026] (4) By constructing a multi-objective feedback closed loop, the present invention enables the entire system to have the ability to self-perceive, self-evaluate and self-adjust during the solution process, and can dynamically balance the search efforts between different objectives, and finally stably produce a high-quality and highly diverse Pareto optimal solution set.

[0027] (5) The present invention has a clear logical hierarchy and the proposed intelligent optimization framework has strong versatility and scalability. It can be widely applied to other production scheduling and logistics optimization fields that require balancing multiple objectives such as efficiency, energy consumption, and cost. Attached Figure Description

[0028] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0029] Figure 1 This is a schematic diagram of the structure of the flexible workshop scheduling problem with a limited number of AGVs that this invention solves.

[0030] Figure 2 This is the three-segment coding scheme for this problem in the present invention.

[0031] Figure 3 This is a flowchart of the algorithm framework for the LLM and reinforcement learning co-evolutionary system for energy efficiency optimization in this invention.

[0032] Figure 4 This is an example Gantt chart of the problem-solving process in this invention. Detailed Implementation

[0033] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0034] The calculation of maximum completion time and total energy consumption is shown in formulas (1) to (5):

[0035]

[0036] Wherein, formula (1) represents the calculation of the maximum completion time, C i This is the completion time of workpiece i. Formula (2) indicates that the total energy consumption consists of processing energy consumption PE, idle energy consumption IE, and transportation energy consumption TE. PE and IE refer to the energy consumption of the machine in processing and idle states, respectively. TE represents the energy consumption generated by AGV transportation. The calculation method for PE is shown in formula (3), pt ijk Indicates processing time, pe ijk Indicates process O ij The processing power X on machine k ijk The variable is used to determine whether to process the product on machine k. The calculation method for IE is shown in formula (4). k This represents the idle time of machine k, i.e. k This represents the idle energy consumption of machine k per unit time. The calculation method for TE is shown in formula (5), tt ijv Z represents the transportation time, te represents the transportation energy consumption per unit time, and Z represents the transportation time. ijv This represents the decision variable, determining whether to use AGVv for transportation.

[0037] Example 1

[0038] This invention provides an LLM and reinforcement learning method and system for flexible workshop scheduling with AGVs, used to solve for FJSP-AGV, with the optimization objective being cooperative Makespan and TEC. This multi-objective optimization system, as an intelligent solver for this problem, includes the following modules:

[0039] (1) Multi-objective cooperative initialization module. In this module, a multi-objective cooperative initialization strategy is designed to consciously generate initial solutions that have advantages in both maximum completion time and total energy consumption objectives, so as to ensure the initial distribution of the population on the Pareto front.

[0040] (2) LLM-guided multi-objective search module. In this module, a large language model is introduced as a high-level strategic decision-maker. By analyzing the distribution, convergence and diversity of the population in the multi-objective space, it dynamically guides the crossover and mutation strategies of the evolutionary algorithm to efficiently approximate the Pareto optimal solution set.

[0041] (3) Q-learning-driven multi-objective local search module. In this module, an adaptive local searcher based on Q-learning is designed. Its state space is fused to understand the position information in the target space, and the action space is a series of energy-saving domain operations for machines, AGVs and processing speed. The optimal operator selection strategy is learned through multi-objective reward signals.

[0042] (4) Multi-objective feedback and population coordination module. This module constructs a multi-objective feedback evaluation system to monitor the hypervolume, spacing and other indicators of the Pareto front in real time, and dynamically adjusts the guidance strategy of LLM, the exploration-exploitation balance of Q-learning and the reorganization of the population according to the feedback signal to maintain the convergence and diversity of the solution set.

[0043] Preferably, in module (1), the multi-objective collaborative initialization strategy first generates a scheduling solution that adopts the earliest completion rule and is biased towards optimizing the maximum completion time; secondly, it generates a scheduling solution that adopts the lowest energy consumption machine and AGV allocation rule and is biased towards optimizing the total energy consumption; the third solution is generated by an insertion heuristic method based on weighted sum evaluation to seek a balance between the two objectives; the remaining solutions are randomly generated to jointly constitute a high-quality initial population that is distributed across multiple objectives.

[0044] Preferably, in module (2), the LLM-guided multi-objective search module uses the Pareto solution set distribution breadth of the current population, the degree of solution set improvement between generations, and population diversity indicators as input prompts. The large language model infers based on this and outputs a high-level decision: whether to focus on exploring the boundary solutions of the unknown target space or to strengthen the fine search within the known Pareto region to improve the convergence and distribution uniformity of the solution set, and dynamically adjust the crossover and mutation strategies accordingly to achieve intelligent search for the shape and density of the Pareto front.

[0045] Preferably, in module (3), the state of the Q-learning-driven multi-objective local search module is defined as the Pareto front level to which the current solution belongs and its standardized values ​​in terms of energy consumption and completion time. Its action space consists of two types of specially designed domain operations: one is efficiency-oriented operations, such as "inserting a critical process on the critical path" or "swapping the machine allocation of adjacent processes in the factory"; the other is energy-efficiency-oriented operations, such as "reassigning processes on high-energy-consuming machines to low-energy-consuming compatible machines" or "merging AGV transportation tasks to reduce idle travel". Its reward function is designed based on the Pareto dominance relationship between the new solution and the original solution and their respective objective function values, thereby guiding Q-learning to intelligently select local search operations that can effectively and synergistically improve the maximum completion time and total energy consumption.

[0046] Preferably, in module (4), the multi-objective feedback and population collaboration module periodically calculates the hypervolume index of the Pareto solution set to monitor the optimization process. When it is detected that the hypervolume has not significantly improved for a consecutive preset number of generations, it is determined that the search has stalled, and a feedback signal to strengthen the exploration is generated. This signal will trigger the adjustment of exploratory parameters such as mutation probability and guide the Q-learning module to increase its exploration rate in order to try more diverse domain operators. At the same time, this module integrates an environment selection mechanism based on fast non-dominated sorting and crowding calculation, which effectively removes redundant solutions and introduces new search directions for the population while ensuring that elite individuals are preserved, thereby systematically maintaining the search momentum.

[0047] Example 2

[0048] Figure 1 This invention illustrates a typical scenario of the FJSP-AGV problem addressed by this invention. Within a flexible workshop, multiple machines, some with identical or different functions, are used to process a batch of workpieces according to their respective process routes. A limited number of Automated Guided Vehicles (AGVs) are deployed within the workshop to transport workpieces between machines. The objective of this invention is to simultaneously allocate appropriate machines to processes, determine the processing sequence of processes, assign AGVs to transport tasks, and plan the transport paths and times for the AGVs, ultimately achieving synergistic optimization of both the maximum completion time (Makespan) and the total system energy consumption (TEC).

[0049] Figure 2 This invention presents a three-tiered coding scheme for this problem. The FJSP-AGV scheduling solution is represented by a three-layer coding scheme: the first layer is the process coding, which determines the global processing sequence of all processes based on the repetitive arrangement of workpieces; the second layer is the machine allocation coding, which sequentially selects a specific processing machine from its set of available machines for each process; the third layer is the AGV allocation coding, which assigns an AGV to each transportation task arising from differences in machines between processes. This coding structure fully expresses the collaborative scheduling scheme for processing and transportation.

[0050] Figure 3 The complete workflow of the system described in this invention is illustrated in flowchart form. The entire process is an iterative evolutionary closed loop, starting with population initialization and ending with the satisfaction of termination conditions, ultimately outputting a set of high-quality Pareto-optimal scheduling schemes. The core steps are as follows:

[0051] 1. Startup and parameter initialization: Set algorithm parameters, such as population size, maximum number of iterations, initial values ​​of crossover and mutation probabilities, LLM bootstrap frequency, Q-learning parameters (learning rate α, discount factor γ, exploration rate ε), etc.

[0052] 2. Multi-objective collaborative initialization (corresponding module (1)): Execute the initialization strategy to generate the initial population. This step specifically generates three types of guided solutions (time-optimal, energy-optimal, and equilibrium solutions) and several random solutions to ensure that the population has a good initial distribution in the target space.

[0053] 3. Evaluation and Pareto Classification: Calculate the objective function value (Makespan and TEC) for each individual in the population. Then, use the fast non-dominated sorting algorithm to classify all individuals in the population into strata (the first stratum is the current Pareto optimal frontier), and calculate the crowding distance between individuals within the same non-dominated stratum to measure the distribution density of the solution.

[0054] 4. LLM-guided macro-strategy adjustment (corresponding module (2)): In each generation or every few generations, the system aggregates the Pareto front distribution status, hypervolume change trend, population diversity indicators, etc. of the current population into a structured description and inputs it into the Large Language Model (LLM). Based on this context, the LLM performs analysis and reasoning, and outputs macro-suggestions for the current generation's search strategy, such as "exploration in low-energy regions should be strengthened" or "the uniformity of the current front solution set should be improved". The system transforms this suggestion into a dynamic adjustment of the probability and type of evolutionary operators (crossover, mutation).

[0055] 5. Evolutionary Operations (Crossing and Mutation): Based on the LLM-guided adjustment strategy in step 4, select parent individuals from the current population and apply the corresponding crossover and mutation operators to generate offspring individuals.

[0056] 6. Q-learning-driven adaptive local search (corresponding module (3)): For newly generated offspring individuals (or focusing on elite individuals among them), a Q-learning-based local optimizer is initiated. The process is as follows:

[0057] a. State awareness: Calculate the normalized position and Pareto level of the solution in the current target space based on the target value of the solution, and use it as the state s.

[0058] b. Action selection: The local searcher selects an action a from the predefined domain operation action library based on the current state s and the Q table using an ε-greedy strategy.

[0059] c. Execution and Evaluation: Execute action a to perturb the current solution, generate a new solution, and calculate its objective value.

[0060] d. Reward Calculation and Learning: Calculate the immediate reward r based on the Pareto dominance relationship between the new solution and the original solution, and use the formula Q(s) t , a t )←Q(s t , a t )+α(r(t+1) +γmaxQ(s (t+1) ,a (t+1) )-Q(s t ,a t Update the Q-value to complete one learning cycle. This process can be repeated multiple times for a single solution to fully explore its improvement potential.

[0061] 7. Environmental Selection and Population Renewal (an important component of module (4)): The parent population is merged with the offspring population after local search. The merged population is then subjected to rapid non-dominated sorting and crowding calculation again. Based on the principle of elite preservation, individuals with higher non-dominated levels are selected first; within the same level, individuals with larger crowding distances are selected first (to maintain diversity), thereby selecting a new generation of population.

[0062] 8. Multi-objective feedback and adaptive adjustment (corresponding to the core of module (4): The system periodically (e.g., every 10 generations) calculates the hypervolume (HV) index of the Pareto solution set of the new generation population. The search progress is judged by monitoring the growth of HV: if the progress is good, the current parameters of each module are maintained. If stagnation is detected (e.g., HV has not improved significantly for several generations): a feedback signal is generated. This signal will directly affect the Q-learning module, increase its exploration rate ε, and encourage the attempt of new operations; at the same time, this stagnation information will also serve as an important context input to guide the next round of LLM, prompting it to make more exploratory strategic decisions.

[0063] 9. Main Loop Check: Check if the preset termination condition has been met. If not, proceed to the next generation evolutionary optimization loop; if met, output the current Pareto optimal solution set and end the process.

[0064] Figure 4 This is an example Gantt chart for solving the problem in this invention. As shown, four workpieces are assigned to a factory with four production machines and two AGVs for processing. Based on the proposed encoding scheme, a Gantt chart is given with the encoding OS=[4,1,2,3,4,1,3,2,1], MA=[2,1,4,4,2,1,3,3,2], and AA=[2,1,1,1,2,2,2,1,1]. The first process for workpiece 4 begins with its first operation. The first operation for workpiece 4 is based on... Select the second machine in the optional machine set, namely machine 3, and have AGV1 transport the goods from the warehouse to machine 3. Continue in this manner until all workpieces have completed their processing steps.

[0065] The advantages of this system are as follows: (1) In the Q-learning-driven multi-objective local search module, its state space integrates the Pareto level and standardized objective value of the solution in the objective space. This design enables the algorithm to accurately identify the position characteristics of the current solution in the dimensions of energy consumption and completion time, thereby more accurately evaluating and selecting the most effective energy-saving scheduling operation, and guiding the search process to efficiently advance towards the real Pareto front; (2) The multi-objective collaborative initialization and feedback evaluation mechanism constructed by this system deeply integrates the characteristics of strong coupling between processing and transportation in the FJSP-AGV problem: during initialization, it constructs separate targets for completion time and energy consumption. The superior solution guides the population to cover the key area of ​​the Pareto front; the feedback module dynamically identifies the search bottleneck and adjusts the exploration direction by monitoring indicators such as hypervolume in real time, so that the system can always maintain a sensitive response to the core features of the problem, thereby fundamentally improving the overall quality and evolution efficiency of the solution; (3) The whole system realizes the dynamic self-evolution of the algorithm search strategy through macro-strategy adjustment and multi-objective feedback closed loop guided by LLM: it can not only actively enhance the exploration when the search stagnates, but also intelligently balance global exploration and local development according to the shape and density of the Pareto front, thereby systematically avoiding invalid search and continuously improving the quality and diversity of the solution set.

[0066] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. An LLM and reinforcement learning method for flexible shop floor scheduling with AGVs, characterized in that, Includes the following steps: (1) Multi-objective collaborative initialization: Generate initial solutions that have advantages in both maximum completion time and total energy consumption objectives, and form an initial population; (2) LLM-guided multi-objective search: By analyzing the distribution of the population in the multi-objective space through a large language model, the crossover and mutation strategies are dynamically adjusted to guide the population to approach the Pareto optimal frontier; (3) Q-learning-driven multi-objective local search: Based on the Pareto level of the current solution and the objective value, Q-learning is used to adaptively select energy-saving neighborhood operations for local optimization; (4) Multi-objective feedback and population collaboration: Real-time monitoring of the population's multi-objective performance indicators, and dynamic adjustment of the LLM guidance strategy and Q-learning exploration rate based on feedback to maintain the population's convergence and diversity.

2. The method according to claim 1, characterized in that, The multi-objective collaborative initialization in step (1) includes: generating a solution that adopts the earliest completion rule and optimizes the maximum completion time; generating a solution that adopts the lowest energy consumption machine and AGV allocation rule and optimizes the total energy consumption; generating an insertion heuristic equilibrium solution based on weighted sum evaluation; and generating several random solutions to construct a well-distributed initial population.

3. The method according to claim 1, characterized in that, The LLM-guided multi-objective search in step (2) includes: using Pareto solution set distribution breadth, solution set improvement degree, and diversity index as input; the large language model infers and outputs high-level strategy suggestions based on the input; and dynamically adjusting the crossover and mutation strategies according to the suggestions to achieve intelligent search.

4. The method according to claim 1, characterized in that, The Q-learning-driven multi-objective local search in step (3) includes: the state is defined as the Pareto hierarchy of the solution and the normalized value of the objective; the action space includes efficiency-oriented and energy-efficiency-oriented neighborhood operations; the reward function is designed based on the Pareto dominance relationship; and the optimal operation is selected for local optimization through Q-learning.

5. The method according to claim 1, characterized in that, The multi-objective feedback and population coordination mentioned in step (4) include: The hypervolume index of the Pareto solution set is calculated periodically; a feedback signal is generated when optimization stagnation is detected; the mutation probability and Q-learning exploration rate are increased; and environmental selection is performed based on non-dominated ranking and crowding distance to maintain population diversity.

6. An LLM and reinforcement learning system for flexible workshop scheduling with AGVs, characterized in that, include: A multi-objective collaborative initialization module is used to generate a well-distributed initial population; An LLM-guided multi-objective search module is used to dynamically adjust evolutionary strategies based on population status. A Q-learning-driven multi-objective local search module for adaptive selection of energy-efficient neighborhood operations; The multi-objective feedback and population collaboration module is used to monitor the optimization process and dynamically adjust system parameters.

7. The system according to claim 6, characterized in that, The LLM-guided multi-objective search module uses a large language model as its core, inputs the multi-objective state information of the population, and outputs high-level search strategy suggestions.

8. The system according to claim 6, characterized in that, The Q-learning-driven multi-objective local search module adopts a state-action-reward mechanism and selects the optimal local operation by combining the Pareto characteristics of the solution.

9. The system according to claim 6, characterized in that, The multi-objective feedback and population collaboration module monitors the optimization process through hypervolume indices and generates feedback signals when stagnation is detected, dynamically adjusting the exploration strategy.

10. The system according to claim 6, characterized in that, The system uses a three-segment coding scheme to represent the scheduling solution, including process coding, machine allocation coding and AGV allocation coding, which fully expresses the collaborative scheduling scheme of processing and transportation.

Images