Material transfer optimal scheduling method and system based on large language model

Abstract

The invention relates to the technical field of optimal scheduling, and discloses a material transfer optimal scheduling method and system based on a large language model. The method comprises the steps of initializing a population, constructing cue words, calling a large language model to generate a filial generation, detecting validity, evaluating and selecting fitness, carrying out linkage adjustment based on a temperature parameter and a population structure, and outputting a current optimal solution. The system corresponds to the method. According to the invention, through deep fusion of the large language model and the material transfer optimization scheduling, the production workshop material transfer efficiency is significantly improved.

Description

Technical Field

[0001] This application relates to the field of optimization scheduling technology, specifically a material transfer optimization scheduling method and system based on a large language model. Background Technology

[0002] In production workshops or warehousing scenarios, Automated Guided Vehicles (AGVs) need to sequentially reach multiple material pickup / delivery points to complete material transfer tasks. The material transfer optimization scheduling problem for a single AGV can be abstracted as follows: Starting from its initial position, the AGV visits each point in the task point set in a certain order, with each point visited only once, in order to minimize the total travel distance. This problem is a typical optimization scheduling problem, characterized by being discrete, non-convex, and having a solution space that grows exponentially with the number of task points. Therefore, it is usually classified as an NP-hard problem in terms of computational complexity.

[0003] In existing technologies, evolutionary algorithms are often used to solve this type of single AGV material transfer optimization scheduling problem. Although evolutionary algorithms have powerful global optimization capabilities, their performance often depends on manually designed operators such as selection, crossover, and mutation, resulting in problems such as strong reliance on domain experience and insufficient transfer adaptability.

[0004] In recent years, with the continuous development of Large Language Models (LLMs), impressive results have been achieved in many fields. Through learning from large amounts of text data, these large language models absorb the accumulated human knowledge, including valuable experience and deep intuition in the field of optimization algorithm design, providing new ideas for solving the aforementioned challenges. Existing techniques have proposed LLM-driven evolutionary methods, which guide the LLM to select parents from the current population and perform crossover and mutation to generate offspring through prompt words, thereby reducing reliance on manual operators. However, in practical applications of single AGV material handling optimization scheduling problems, existing LLM-driven evolutionary methods are still prone to problems such as: generated sequences not meeting validity requirements, difficulty in continuously improving path structures, recurrence of inferior offspring, and shortcomings in temperature population adaptive adjustment, leading to limited convergence efficiency and solution quality.

[0005] Therefore, there is an urgent need for a new material transfer optimization scheduling technology based on a large language model, which can improve the quality of single AGV material transfer optimization scheduling by means of offspring effectiveness detection and repair, path structure optimization, avoidance of inferior offspring and temperature combined with population adaptive linkage adjustment. Summary of the Invention

[0006] The purpose of this application is to provide a material transfer optimization scheduling method and system based on a large language model, in order to solve the technical problems in the existing technology, such as the generation sequence not meeting the validity requirements, the path structure being difficult to continuously improve, the repeated occurrence of inferior offspring, and the inadequacy in temperature population adaptive adjustment, which lead to limited convergence efficiency and solution quality.

[0007] To achieve the above objectives, this application provides a material transfer optimization scheduling method based on a large language model, applied to a single automated guided vehicle (AGV) material transfer optimization scheduling scenario. The method includes: S1: Obtain the material transfer scheduling data to be optimized from the production site data source, including at least the transfer task order, workshop map, AGV initial position and speed limit; construct the cost matrix based on the material transfer scheduling data to construct the initial population; the population individuals are structured access sequence sequences, and the initialization process does not involve a large language model; S2: Construct a prompt word, which includes at least a problem description, population context information, and task instructions; the problem description includes at least a set of pick-up and delivery points for a single automated guided vehicle, corresponding distance cost information, and task node sequence requirements; the task instructions require outputting a structured access sequence and ensuring that each point is accessed only once. S3: Generate offspring individuals based on prompt words by calling a large language model. The offspring individuals are structured access sequence sequences. S4: Perform a validity check on the offspring individuals. If the check fails, trigger a validity repair to obtain the repaired offspring individuals. S5: Merge the offspring individuals that have passed the effectiveness test or repair with the current population individuals, and conduct fitness assessment and selection based on the total driving distance of the merged individuals to form the next generation population; S6: For the evolutionary process of optimizing material transfer scheduling for a single automated guided vehicle, temperature parameters and population structure are adjusted in conjunction to dynamically switch between exploration and convergence. S7: When the preset maximum number of generations is reached or the preset optimization target is met, the evolution is terminated and the current best individual is output. The best individual is converted into a navigation instruction sequence for controlling a single automated guided vehicle to transfer materials, so as to optimize the material transfer scheduling of a single automated guided vehicle.

[0008] Preferably, S2 also includes prompt word optimization based on a sub-path fine-tuning mechanism, which is based on the 2-opt operator.

[0009] As a preferred option, S2 also includes a dynamic sub-path obstacle avoidance mechanism and a failed sample adaptive guidance mechanism; The dynamic sub-path obstacle avoidance mechanism defines the frequently occurring path segments in the failed paths as the first negative signal, and guides the large language model to avoid generating new paths that include the path segments corresponding to the first negative signal based on the first negative signal. The failure sample adaptive guidance mechanism defines inferior offspring as the second negative signal, and guides the large language model based on the second negative signal to avoid generating new paths that include features corresponding to the second negative signal.

[0010] Preferably, the coordinated adjustment of temperature parameters and population structure is based on a dual-channel coordinated adjustment mechanism, which specifically includes: The temperature regulation is used to dynamically switch between exploration and convergence modes at different stages, and is coordinated with the population size for adjustment; and the temperature regulation is set with a minimum time interval.

[0011] Preferably, the temperature regulation is based on continuous unimproved algebraic triggering, including: Obtain consecutive unimproved algebras and their corresponding preset negative threshold groups, wherein the preset negative threshold groups include multiple preset negative thresholds; Based on the continuous unimproved algebra, the system iterates through the preset negative threshold group. When the continuous unimproved algebra is greater than the corresponding preset negative threshold, the temperature is increased based on multiple preset increase rates.

[0012] Preferably, the linkage adjustment based on population size specifically involves dynamically adjusting the population size within a preset number of rounds after the temperature adjustment is completed, based on the current temperature and the severity of failure; wherein, the severity of failure is determined based on the number of consecutive generations without improvement, and the adjustment of the population size involves adding new individuals to the population.

[0013] Preferably, the temperature regulation is further based on continuously improved algebraic triggering, including: Obtain the continuously improved algebra and its corresponding preset positive threshold; When the number of consecutive improvement generations is greater than the preset positive threshold, the temperature is rolled back based on a preset reduction rate.

[0014] To achieve the above objectives, this application also provides a material transfer optimization scheduling system based on a large language model, which applies the material transfer optimization scheduling method based on a large language model as described above, including: The initialization module is used to obtain the material transfer scheduling data to be optimized from the production site data source, including at least the transfer task order, workshop map, AGV initial position and speed limit; construct a cost matrix based on the material transfer scheduling data to build the initial population; the population individuals are structured access sequence sequences, and the initialization process does not involve a large language model; The prompt word construction module is used to construct prompt words, which include at least a problem description, population context information, and task instructions. The problem description includes at least a set of pick-up and delivery points for a single automated guided vehicle, corresponding distance cost information, and task node sequence requirements. The task instructions require the output of a structured access sequence and satisfy that each point is accessed only once. The offspring generation module is used to generate offspring individuals based on prompt words by calling the large language model. The offspring individuals are structured access sequence sequences. The validity detection module is used to perform validity checks on offspring individuals. If the check fails, validity repair is triggered to obtain the repaired offspring individuals. The evaluation and selection module is used to merge offspring individuals that have passed the effectiveness test or repair with current population individuals, and to evaluate and select the next generation population based on the total driving distance of the merged individuals. The linkage adjustment module is used to dynamically switch between exploration and convergence by adjusting temperature parameters and population structure in response to the evolutionary process of material transfer optimization scheduling for a single automated guided vehicle. The optimal individual output module is used to terminate evolution and output the current optimal individual when the preset maximum number of generations is reached or the preset optimization target is met. The optimal individual is converted into a navigation instruction sequence for controlling a single automated guided vehicle to transfer materials, so as to optimize the material transfer scheduling of a single automated guided vehicle.

[0015] To achieve the above objectives, this application also provides a computer device for material transfer optimization scheduling based on a large language model, including at least one processor, at least one memory, and a data bus; The processor and the memory communicate with each other via the data bus; The memory stores program instructions that can be executed by the processor, which calls the program instructions to execute the material transfer optimization scheduling method based on a large language model as described above.

[0016] To achieve the above objectives, this application also provides a storage medium storing a computer program thereon, which, when executed by a processor, implements the material transfer optimization scheduling method based on a large language model as described above.

[0017] Beneficial Effects: Compared with existing technologies, the material transfer optimization scheduling method and system based on a large language model proposed in this application constructs a complete technical chain for the single AGV material transfer optimization scheduling problem, from the task point set to the candidate access order and then to the AGV executable instructions: In the initialization stage, it does not rely on the large language model to build a basic population, avoiding interference from initial model bias on the search process; the prompt words integrate the task point set with its travel cost information, population context, and task instructions, enabling the large language model to output the candidate access order in a structured form; in the offspring generation stage, validity detection and feasibility repair that satisfy the constraints of the single AGV material transfer optimization scheduling problem are introduced, ensuring the generated results... It can be directly used for subsequent evaluation and AGV execution; by embedding the 2-opt sub-path fine-tuning mechanism into the prompt words, it achieves rapid local improvement of the path structure, improving the quality of offspring and the convergence speed; by forming a negative signal closed loop through dynamic sub-path obstacle avoidance and adaptive guidance of failed samples, it reduces the recurrence of inferior segments and inferior offspring, improving the effectiveness of the search; and by adjusting the temperature parameter and population size through dual-channel linkage, it enhances exploration when stagnant and accelerates convergence when continuously improving, thereby improving convergence efficiency while ensuring the quality of offspring. Finally, the optimal access order is converted into an AGV-executable navigation task instruction sequence to complete the material transfer scheduling of a single AGV, reducing the threshold for solving and deployment. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 The overall architecture diagram of the material transfer optimization scheduling method based on a large language model provided in the embodiments of this application; Figure 2 A flowchart illustrating the material transfer optimization scheduling method based on a large language model provided in this application embodiment; Figure 3 A flowchart illustrating the temperature adaptive adjustment process provided in this application embodiment; Figure 4 A flowchart illustrating the population adaptive adjustment process provided in this application embodiment; Figure 5 A flowchart illustrating the algorithm of the material transfer optimization scheduling method based on a large language model provided in this application embodiment; Figure 6 The structural block diagram of the material transfer optimization scheduling system based on a large language model provided in the embodiments of this application is shown.

[0020] The implementation, functional features, and advantages of this invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0021] The technical solutions in the embodiments of this application will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0022] In this document, the term "comprising" is intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0023] In practical applications, the basic form of single AGV task sequence planning can be described as follows: Within a manufacturing plant, warehouses and workstations correspond to multiple task points. A single AGV must sequentially reach these task points within a single scheduling cycle to complete material transfer. Task points and their identifiers are generated by the warehouse / manufacturing management system, and the travel costs between task points are provided by the plant map, forming a cost matrix. Determining the order in which AGVs access these task nodes, minimizing the total travel distance while satisfying all transfer requirements, is the core problem that single AGV task sequence planning needs to solve. From a technical modeling perspective, the above single AGV material transfer scheduling problem can be abstracted as follows: In a weighted graph consisting of warehouse nodes and multiple workstation nodes, determine a sequence of paths that access all specified task nodes without repetition, and minimize the path cost while satisfying the corresponding constraints.

[0024] In practical industrial applications, the aforementioned problems are not only affected by the number of nodes, but also strongly correlated with factors such as workshop topology and dynamic changes in tasks, making it difficult for fixed strategies or static rules to adapt to different production scenarios. Therefore, the single AGV task sequence planning problem places higher demands on the scheduling method's solution space exploration capabilities and problem adaptability.

[0025] Reference Figure 1 , Figure 1 This is the overall architecture diagram of the material transfer optimization scheduling method based on a large language model in this embodiment.

[0026] like Figure 1 As shown, in terms of overall architecture, the material transfer optimization scheduling method based on a large language model disclosed in this embodiment adopts the basic framework of traditional evolutionary algorithms and optimizes it, specifically as follows: For a given material transfer optimization scheduling problem T, the initialization process is performed first: based on the workshop topology, task set, path cost, and constraints, N feasible scheduling schemes are generated as an initial candidate set. Then, the iterative optimization phase begins.

[0027] In each iteration, a set of prompt information for the large language model is constructed and updated. The set of prompt information includes at least: (1) a structured description of the problem (including task points, start and end points, cost indicators, constraints, etc.), (2) standardized task instructions (including output format, feasibility requirements, etc.), and (3) guidance information based on historical samples. Based on the optimization of previous generations, prompt words containing sub-path obstacle avoidance and failure sample guidance mechanisms are constructed to form a prompt context from generation obstacle avoidance to quality improvement.

[0028] Driven by the prompt information, the large language model performs scheme generation and update on the current candidate set: it selects candidate schemes from the candidate set as reference inputs, generates a new set of candidate schemes, performs local fine-tuning on the sub-paths in the schemes, and performs validity checks on them, so that the output meets the workshop path constraints and output structure requirements.

[0029] Subsequently, the newly generated candidate solutions are merged with the current candidate set, and sorted and filtered using the fitness evaluation module, retaining the N solutions with the best quality as the candidate set for the next round. In addition, the system performs coordinated regulation of temperature and population. When the number of iterations reaches a preset upper limit or the termination condition is met, the iteration ends, and the scheduling scheme with the best fitness in the candidate set is output as the final result.

[0030] The main process of the material transfer optimization scheduling method based on the large language model in this embodiment will now be described in detail.

[0031] Reference Figure 2 , Figure 2 This is a flowchart of the material transfer optimization scheduling method based on a large language model in this embodiment.

[0032] like Figure 2 As shown in the figure, this embodiment discloses a material transfer optimization scheduling method based on a large language model, applied to the material transfer optimization scheduling problem. The method includes: S1: Obtain the material transfer scheduling data to be optimized from the production site data source, including at least the transfer task order, workshop map, AGV initial position and speed limit; construct the cost matrix based on the material transfer scheduling data to construct the initial population; the population individuals are a structured access sequence, and the initialization process does not involve a large language model.

[0033] In this specific application, basic data for the problem T to be solved is obtained from data sources on the production site and structured modeling is completed, then an initial candidate scheduling set is generated. Data sources include, but are not limited to: transfer task orders (including pick-up points, delivery points, etc.) issued by the Manufacturing Execution System / Warehouse System, workshop maps, and operating parameters such as the initial position and speed limit of the AGV fed back by the AGV controller. A cost matrix is ​​constructed based on the above data, and each transfer task is abstracted as a set of nodes or work stations with paired pick-up and delivery points. N initial individuals are generated as an initial candidate scheduling set P: each individual represents the task execution order of a single AGV as a structured sequence, satisfying feasibility constraints (such as task coverage, valid start and end points, valid format, and the order of task nodes). This stage uses a random method to generate candidate sequences and does not invoke LLM.

[0034] S2: Construct a prompt word, which includes at least a problem description, population context information, and task instructions. The problem description includes at least a set of pick-up and delivery points for a single automated guided vehicle, corresponding distance cost information, and task node sequence requirements. The task instructions require the output of a structured access sequence and satisfy that each point is accessed only once.

[0035] In the specific application of this embodiment, in each iteration, based on the structured data of the problem to be solved, the current population state, and the optimization objective, a prompt word matching the current solution process is constructed. The prompt word serves as the input carrier for the large language model, used to transmit the specific description of the optimization scheduling problem, material transfer optimization scheduling constraints, evaluation criteria, and candidate scheduling set states to the large language model through a unified linguistic interface. This enables the large language model to generate a child scheduling set that satisfies the workshop material transfer constraints under a controlled output format and to update the candidate scheduling set. To ensure that the prompt word can stably guide the large language model to complete the candidate solution update process, this embodiment divides the core components of the prompt word into the following three parts: Problem Description: The prompts should clearly and systematically describe the basic elements of the optimization problem to be solved, including the problem background, objective function, constraints, and structured output format. In this embodiment, the prompts should explicitly provide the location information of the warehouse and workstation, and specify the constraints of the single AGV scheduling solution, such as allowing each task point / workstation to be accessed only once, the path must form a closed loop or meet preset start-end point requirements, and the order of task nodes. Simultaneously, the prompts should clearly state that the optimization objective is to minimize the total travel distance as the cost indicator, and the output must be a structured sequence that can be parsed by the system to ensure that the offspring individuals generated by the large language model can directly enter the subsequent validity detection and fitness evaluation processes.

[0036] Population context information: The prompt should also provide information on several representative individuals in the current population and their fitness to enhance the large language model's ability to understand the context of the current search state. Representative individuals should include at least those with better and worse performance, along with their corresponding total distance, constraint violations, and other evaluation results. This allows the large language model to learn from the quality differences between different individuals the distinction between feasible and low-cost structural features and high-cost or infeasible structural features, thus favoring the inheritance of superior structures and avoiding low-quality structures when generating new solutions.

[0037] Task Instructions: The prompts must include clear instructions on the operational process to guide the large language model in updating the offspring scheduling set according to a preset sequence. The task instructions should at least include: selecting a parent generation based on the current population, performing crossover and mutation on the parent generation, and outputting the offspring scheduling set as required. During the crossover and mutation process, the prompts will incorporate knowledge from the field of Evolutionary Computation (EC), requiring the LLM to select the most suitable operator from multiple options based on the current logistics transfer status of a single AGV, and clearly indicating the type of operator used and the generated new solution. This approach allows the large model to not only generate results but also output its reasoning process for subsequent verification and analysis.

[0038] Although large language models can generate offspring individuals under the guidance of prompts, their ability to correct local optima is relatively limited when dealing with material transfer optimization scheduling problems. This can easily lead to situations where newly generated offspring contain obvious redundant edges, detour segments, or non-shortest sub-paths. To further enhance the material transfer optimization scheduling capability of offspring under constraints, this embodiment proposes a sub-path fine-tuning mechanism and incorporates the constraints corresponding to this mechanism into the system prompts to guide the output of the large language model.

[0039] Specifically, S2 also includes prompt word optimization based on a sub-path fine-tuning mechanism, which is based on the 2-opt operator.

[0040] In this specific application, the local commutation logic of 2-opt is embedded in the prompt words as an executable linguistic rule. This guides the large language model to perform a local commutativity check on the current path structure after completing the basic path generation, based on the given path cost information. Specifically, it checks whether there are commutable non-adjacent edge pairs in the path, so that the total path cost can be reduced through a local reversal operation while satisfying task constraints and effectiveness. When it is detected that the local reversal operation can reduce the cost, the large language model is required to output the improved individual after the reversal, and... <refined>The system returns labels to differentiate between "original generated individuals" and "fine-tuned individuals" and prioritizes them for subsequent evaluation. By incorporating the sub-path fine-tuning rule into the prompt, the system can improve the quality of local paths in candidate solutions and enhance the availability and stability of the sub-scheduling set without introducing additional external solution calculation modules.

[0041] While the aforementioned evolutionary process driven by a large language model demonstrates significant optimization capabilities, it may still experience prolonged stagnation when the actual workshop task structure is complex, meaning that the optimal solution remains unchanged for multiple generations. To improve the ability to escape and convergence efficiency in such stagnation scenarios, this embodiment proposes a dynamic sub-path obstacle avoidance mechanism and a failure sample adaptive guidance mechanism, and integrates these two mechanisms with the prompt word structure, applying directional constraints to the large language model generation process through negative signal injection.

[0042] Specifically, S2 also includes a dynamic sub-path obstacle avoidance mechanism and a failed sample adaptive guidance mechanism; The dynamic sub-path obstacle avoidance mechanism defines frequently occurring path segments in failed paths as the first negative signal, and guides the large language model to avoid generating new paths that include path segments corresponding to the first negative signal based on the first negative signal. The failure sample adaptive guidance mechanism defines inferior offspring as the second negative signal, and guides the large language model based on the second negative signal to avoid generating new paths that include features corresponding to the second negative signal.

[0043] In the specific application of this embodiment, the dynamic sub-path obstacle avoidance mechanism statistically models "potentially failed paths" based on the current population and historical evaluation results: when no better individuals are obtained for several consecutive generations, the scheduling set in the current population is automatically collected as a set of potential failed paths, and frequency statistics are performed on the path segments that meet the detection logic; when a certain path segment appears frequently in the set of potential failed paths, it is judged as a high-risk local structure related to stagnation, and the path segment is written as the first negative signal into the prompt word. In the form of natural language constraints, the large language model is explicitly required to avoid structures containing the segment when generating new paths, thereby reducing the probability of repeatedly generating the same type of failed structure and improving the search coverage.

[0044] In the specific application of this embodiment, both the failure sample adaptive guidance mechanism and the dynamic sub-path obstacle avoidance mechanism utilize negative feedback information to constrain the generation process, but they operate on different targets: the dynamic sub-path obstacle avoidance mechanism focuses on high-frequency local segments common to inferior offspring, while the failure sample adaptive guidance mechanism focuses on the overall structural features of inferior offspring. Therefore, the negative example information fed back to the prompt words by the failure sample adaptive guidance mechanism consists of representative complete inferior offspring and their failure reasons descriptions. This enables the large language model to form a tendency to identify and avoid inferior structural features during semantic reasoning, and actively avoid structural features corresponding to the second negative signal when generating new candidate solutions, thereby reducing the probability of getting stuck in local stagnation and improving global convergence ability.

[0045] S3: Generate offspring individuals based on prompt words by calling a large language model. The offspring individuals are structured access sequence sequences.

[0046] In the specific application of this embodiment, after the prompt words are generated, a pre-trained large language model, such as GPT-3.5 or GPT-4o, is called through an Application Programming Interface (API) to generate progeny words based on the input natural language content. These progeny words are then transformed into a structured form to ensure that they meet the constraints required by the specific combinatorial optimization problem.

[0047] S4: Perform a validity check on the offspring individuals. If the check fails, trigger validity repair to obtain a repaired offspring individual. In the specific application of this embodiment, the validity check on the offspring individuals is performed as follows: when duplicate points, missing points, or formats that do not meet the preset structured form or the specific requirements and constraints of the single AGV optimization scheduling problem listed above are detected, validity repair is triggered to obtain a repaired offspring individual.

[0048] S5: Merge the offspring individuals that have passed the validity test or repair with the current population individuals. Based on the total travel distance of the merged individuals, perform fitness evaluation and selection to form the next generation population. In the specific application of this embodiment, the newly generated offspring will be merged with the current population, and then all individuals will be scored according to the fitness evaluation function. The top N optimal individuals will be selected to enter the next generation.

[0049] S6: For the evolutionary process of optimizing material transfer scheduling for a single automated guided vehicle, temperature parameters and population structure are adjusted in conjunction to dynamically switch between exploration and convergence.

[0050] Text generation in large language models is controlled by a temperature parameter: higher temperatures result in more random generation, suitable for early exploration; lower temperatures lead to more stable output, which is beneficial for later convergence. Meanwhile, population size also affects the search process: during exploration, a larger population size helps increase population diversity; while appropriately reducing the population size during the convergence phase helps improve convergence performance.

[0051] In response to this issue, this embodiment proposes an innovative temperature-population linkage adjustment mechanism to address the switching between "exploration" and "convergence" modes in the logistics transfer optimization scheduling problem. The system performs dual-channel linkage adjustment of temperature parameters and population size based on the material transfer scheduling evaluation results.

[0052] Specifically, the coordinated adjustment of temperature parameters and population structure is based on a dual-channel coordinated adjustment mechanism, which is as follows: The system dynamically switches between exploration and convergence modes at different stages through temperature regulation, and adjusts these modes in conjunction with population size; the temperature regulation is set with a minimum time interval. In this embodiment, the minimum time interval for temperature adjustment is no less than two generations to prevent frequent fluctuations from increasing search instability.

[0053] Reference Figure 3 , Figure 3 This is a flowchart of the temperature adaptive adjustment process in this embodiment.

[0054] like Figure 3 As shown, specifically, the temperature regulation is based on continuous unimproved algebraic triggering, including: Obtain consecutive unimproved algebras and their corresponding preset negative threshold groups, wherein the preset negative threshold groups include multiple preset negative thresholds; Based on the continuous unimproved algebra, the system iterates through the preset negative threshold group. When the continuous unimproved algebra is greater than the corresponding preset negative threshold, the temperature is increased based on multiple preset increase rates.

[0055] like Figure 3 As shown, specifically, the temperature regulation is also based on continuously improved algebraic triggering, including: Obtain the continuously improved algebra and its corresponding preset positive threshold; When the number of consecutive improvement generations is greater than the preset positive threshold, the temperature is rolled back based on a preset reduction rate.

[0056] like Figure 3 As shown, in this specific application, temperature regulation is primarily triggered by the number of consecutive unimproved algebra iterations (worse_iter). When the number of consecutive failure rounds reaches a specific threshold, the temperature will increase according to the following tiered rules: When worse_iter exceeds 10, 20, and 30, the temperature is increased by increments of 0.005, 0.007, and 0.010 respectively, with a maximum increase of 1.3.

[0057] Meanwhile, when the consecutive improvement algebra better_iter > 5 occurs, the system will back down the temperature by an amount of 0.005 to maintain stable convergence behavior.

[0058] It should be noted that in this embodiment, the preset negative thresholds in the preset negative threshold group are 10, 20, and 30, and the preset increase values ​​are 0.005, 0.007, and 0.010. The preset positive threshold is 5, and the preset decrease value is 0.005.

[0059] Reference Figure 4 , Figure 4 This is a flowchart of the population adaptive adjustment process in this embodiment.

[0060] like Figure 4 As shown, specifically, the linkage adjustment based on population size involves dynamically adjusting the population size within a preset number of rounds after the temperature adjustment is completed, based on the current temperature and the severity of failure; wherein, the severity of failure is determined based on the number of consecutive generations without improvement, and the adjustment of the population size involves adding new individuals to the population.

[0061] like Figure 4 As shown, in the specific application of this embodiment, within 1 to 2 rounds after each temperature adjustment, the system dynamically adjusts the population size `pop_num` based on the current temperature and the severity of failure to enhance the coverage of the solution space. When the temperature reaches 1.05, 1.1, or 1.2, and is accompanied by `worse_iter` exceeding 20, 25, and 30 respectively, the system will add 1, 2, and 3 individuals to the population. To avoid excessive population growth leading to a surge in computational overhead, the system also sets a maximum population size limit.

[0062] It should be noted that the preset number of rounds in this embodiment is 1 to 2, the judgment value corresponding to the current temperature is 1.05, 1.1 or 1.2, and the severity of failure is determined based on the accompanying worse_iter exceeding 20, 25 or 30.

[0063] S7: When the preset maximum number of generations is reached or the preset optimization target is met, the evolution is terminated and the current best individual is output. The best individual is converted into a navigation instruction sequence for controlling a single automated guided vehicle to transfer materials, so as to optimize the material transfer scheduling of a single automated guided vehicle.

[0064] In the specific application of this embodiment, when the algorithm reaches the preset maximum number of algebras G or meets certain optimization objectives, such as reaching a specified accuracy or finding the optimal solution, the algorithm terminates and outputs the current optimal solution.

[0065] Reference Figure 5 , Figure 5 This is a flowchart of the material transfer optimization scheduling method based on a large language model in this embodiment.

[0066] like Figure 5 As shown, the algorithm corresponding to the material transfer optimization scheduling method based on the large language model in this embodiment corresponds to S1 to S7 above. Its input is the optimization problem T, the maximum number of iterations G, and the population size N. Its output is the optimal solution s. .

[0067] To systematically evaluate the optimization effect of the Large Language Model-based Material Transport Scheduling (LLM-MTS) method in this embodiment, a series of experimental tasks covering different scales were designed using the classic single AGV production workshop material transport problem as the test platform. The comprehensive capabilities of LLM-MTS in terms of optimization performance, convergence behavior, and algorithm mechanism effectiveness were verified from multiple perspectives.

[0068] For the test problem setup, this embodiment uses the material transfer problem in a production workshop as the test task. The test instances are generated by the warehouse / manufacturing management system and are called "tue". The instance sizes vary, including four numbers of nodes: n=10, 15, 20, 25, and each size contains four different instances, for a total of 16 instances. The optimal solution for all instances is obtained through an exact solver.

[0069] To ensure the universality of the algorithm framework and minimize reliance on prior domain knowledge, LLM-MTS employs a random initialization strategy in the initial stage of evolutionary search. At the start of the process, LLM-MTS transforms the test instances generated by the warehouse / manufacturing management system into specific mathematical expressions to be solved, completing structured modeling. Under the constraints, for each instance to be optimized, a set of N initial generations is directly generated to form the initial population.

[0070] Regarding the baseline algorithm settings, this embodiment uses four classic heuristic algorithms for comparison with LLM-MTS as a benchmark: Nearest Neighbor (NN), Farthest Insertion (FI), Nearest Insertion (NI), and Random Insertion (RI). These are all widely used heuristic algorithms. The Nearest Neighbor algorithm starts from a random node and selects the next node closest to the current node to add to the path each time, until all nodes are included. The insertion algorithm, on the other hand, selects one node to insert into the current path at each step to minimize the cost of adding nodes. Specifically, FI selects the node farthest from any other node in the path, NI selects the node closest to any other node in the path, and RI randomly selects the node to insert.

[0071] Regarding parameter settings and evaluation indicators: In terms of parameter settings, the population size N of LLM-MTS was set to 16, and the maximum number of iterations G was set to 250. These two figures were obtained by balancing the full utilization of algorithm performance with the saving of computational resources (such as time and API consumption). The large language model used by the algorithm is OpenAI's GPT-3.5-turbo model, which is accessed via API.

[0072] Regarding the evaluation metrics, specifically the performance evaluation metrics, this experiment uses the Optimality Gap to evaluate the algorithm's performance. The calculation formula is as follows: In addition, the number of instances in which the algorithm finds the optimal solution and the average number of iterations required for the algorithm to find the optimal solution on the first attempt will be recorded to measure the convergence speed.

[0073] The final data obtained from the experiment are summarized in the table below: The data in the table is presented as average optimality gap ± standard deviation, and the rightmost column represents the average number of iterations ± standard deviation required for successful convergence. The number of successful convergences on four instances is recorded in parentheses. If the algorithm fails to find the optimal solution on a test set, the result is recorded as N / A. From the above data, we can draw the following conclusions: (1) LLM-MTS demonstrates excellent optimization capabilities in small-scale single AGV logistics transfer optimization scheduling problems (with 10 and 15 task points). It can stably find the optimal solution in almost all instances, and outperforms all traditional heuristic algorithms in terms of result accuracy and convergence speed. Even when the solution space for instances with 15 nodes exceeds the order of magnitude, LLM-MTS can still accurately locate the optimal path in a finite number of evolution rounds. This result fully demonstrates the advantages of large language models in structural modeling and generalized reasoning. In particular, it can still achieve better performance than explicit constructive property algorithms in the absence of domain-specific heuristic information, which fully illustrates the potential of using prompt word design to drive the language model search to effectively approximate the solution of NP-hard problems.

[0074] (2) LLM-MTS performs slightly worse than traditional heuristic algorithms other than the nearest neighbor algorithm on medium-sized problems (20 task points), especially the nearest insertion and farthest insertion algorithms. On larger-scale problems (25 task points), the convergence difficulty of LLM-MTS increases significantly, the solution quality declines, and the optimality gap of some instances is significantly higher than the baseline method. The diversity of path structures generated by the language model in the search space decreases significantly with the increase in problem size, suggesting that the population information carried in the prompt words may still be insufficient to fully stimulate the language model's understanding and exploration of the global properties of the solution space. This phenomenon indicates that the current prompt construction strategy and interaction mechanism have a scalability bottleneck for larger-scale problems.

[0075] Reference Figure 6 , Figure 6 This is a block diagram of the material transfer optimization scheduling system based on a large language model in this embodiment.

[0076] like Figure 6 As shown, this embodiment also discloses a material transfer optimization scheduling system based on a large language model, which applies the material transfer optimization scheduling method based on a large language model as described above, including: The initialization module is used to obtain the material transfer scheduling data to be optimized from the production site data source, including at least the transfer task order, workshop map, AGV initial position and speed limit; construct a cost matrix based on the material transfer scheduling data to build the initial population; the population individuals are structured access sequence sequences, and the initialization process does not involve a large language model; The prompt word construction module is used to construct prompt words, which include at least a problem description, population context information, and task instructions. The problem description includes at least a set of pick-up and delivery points for a single automated guided vehicle, corresponding distance cost information, and task node sequence requirements. The task instructions require the output of a structured access sequence and satisfy that each point is accessed only once. The offspring generation module is used to generate offspring individuals based on prompt words by calling the large language model. The offspring individuals are structured access sequence sequences. The validity detection module is used to perform validity checks on offspring individuals. If the check fails, validity repair is triggered to obtain the repaired offspring individuals. The evaluation and selection module is used to merge offspring individuals that have passed the effectiveness test or repair with current population individuals, and to evaluate and select the next generation population based on the total driving distance of the merged individuals. The linkage adjustment module is used to dynamically switch between exploration and convergence by adjusting temperature parameters and population structure in response to the evolutionary process of material transfer optimization scheduling for a single automated guided vehicle. The optimal individual output module is used to terminate evolution and output the current optimal individual when the preset maximum number of generations is reached or the preset optimization target is met. The optimal individual is converted into a navigation instruction sequence for controlling a single automated guided vehicle to transfer materials, so as to optimize the material transfer scheduling of a single automated guided vehicle.

[0077] This embodiment also discloses a computer device for combinatorial optimization processing of large language models, including at least one processor, at least one memory and a data bus; The processor and the memory communicate with each other via the data bus; The memory stores program instructions that can be executed by the processor, which calls the program instructions to execute the material transfer optimization scheduling method based on a large language model as described above.

[0078] This embodiment also discloses a storage medium on which a computer program is stored. When the computer program is executed by a processor, it implements the material transfer optimization scheduling method based on a large language model as described above.

[0079] It should be noted that the material transfer optimization scheduling system, computer equipment, and storage medium based on the large language model in this embodiment correspond to the aforementioned material transfer optimization scheduling method based on the large language model. Therefore, any content not specifically described in the material transfer optimization scheduling system, computer equipment, and storage medium based on the large language model in this embodiment, including but not limited to functional definitions, working principles, and technical effects, can be referred to the description in the aforementioned material transfer optimization scheduling method based on the large language model, and will not be repeated here.

[0080] In the embodiments provided in this application, it should be understood that the embodiments described herein can be implemented in hardware, software, firmware, middleware, code, or any suitable combination thereof. For hardware implementation, the processor may be implemented in one or more of the following: application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field-programmable gate arrays (FPGAs), processors, controllers, microcontrollers, microprocessors, other electronic units designed to implement the functions described herein, or combinations thereof. For software implementation, some or all of the processes of the embodiments may be performed by a computer program instructing the associated hardware. During implementation, the program may be stored in a computer-readable storage medium or transmitted as one or more instructions or code on a computer-readable storage medium. Computer-readable storage media include computer storage media and communication media, wherein communication media include any medium that facilitates the transmission of a computer program from one place to another. Storage media may be any available medium accessible to a computer. Computer-readable storage media may include, but are not limited to, RAM, ROM, EEPROM, CD-ROM or other optical disk storage, magnetic disk storage media or other magnetic storage devices, or any other medium capable of carrying or storing desired program code having the form of instructions or data structures and accessible to a computer.

[0081] Finally, it should be noted that the above description is only a preferred embodiment of this application and is not intended to limit this application. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.< / refined>

Claims

1. A material transfer optimization scheduling method based on a large language model, applied to a single automated guided vehicle (AGV) material transfer optimization scheduling scenario, characterized in that, The method includes: S1: Obtain the material transfer scheduling data to be optimized from the production site data source, including at least the transfer task order, workshop map, AGV initial position and speed limit; construct the cost matrix based on the material transfer scheduling data to construct the initial population; the population individuals are structured access sequence sequences, and the initialization process does not involve a large language model; S2: Construct a prompt word, which includes at least a problem description, population context information, and task instructions; the problem description includes at least a set of pick-up and delivery points for a single automated guided vehicle, corresponding distance cost information, and task node sequence requirements; the task instructions require outputting a structured access sequence and ensuring that each point is accessed only once. S3: Generate offspring individuals based on prompt words by calling a large language model. The offspring individuals are structured access sequence sequences. S4: Perform a validity check on the offspring individuals. If the check fails, trigger a validity repair to obtain the repaired offspring individuals. S5: Merge the offspring individuals that have passed the effectiveness test or repair with the current population individuals, and conduct fitness assessment and selection based on the total driving distance of the merged individuals to form the next generation population; S6: For the evolutionary process of optimizing material transfer scheduling for a single automated guided vehicle, temperature parameters and population structure are adjusted in conjunction to dynamically switch between exploration and convergence. S7: When the preset maximum number of generations is reached or the preset optimization target is met, the evolution is terminated and the current best individual is output. The best individual is converted into a navigation instruction sequence for controlling a single automated guided vehicle to transfer materials, so as to optimize the material transfer scheduling of a single automated guided vehicle.

2. The material transfer optimization scheduling method based on a large language model according to claim 1, characterized in that, S2 also includes prompt word optimization based on a sub-path fine-tuning mechanism, which is based on the 2-opt operator.

3. The material transfer optimization scheduling method based on a large language model according to claim 1, characterized in that, S2 also includes a dynamic sub-path obstacle avoidance mechanism and a failed sample adaptive guidance mechanism; The dynamic sub-path obstacle avoidance mechanism defines the frequently occurring path segments in the failed paths as the first negative signal, and guides the large language model to avoid generating new paths that include the path segments corresponding to the first negative signal based on the first negative signal. The failure sample adaptive guidance mechanism defines inferior offspring as the second negative signal, and guides the large language model based on the second negative signal to avoid generating new paths that include features corresponding to the second negative signal.

4. The material transfer optimization scheduling method based on a large language model according to claim 1, characterized in that, The coordinated regulation of temperature parameters and population structure is based on a dual-channel coordinated regulation mechanism, which specifically includes: The temperature regulation is used to dynamically switch between exploration and convergence modes at different stages, and is coordinated with the population size for adjustment; and the temperature regulation is set with a minimum time interval.

5. The material transfer optimization scheduling method based on a large language model according to claim 4, characterized in that, The temperature regulation is based on continuous unimproved algebraic triggering, including: Obtain consecutive unimproved algebras and their corresponding preset negative threshold groups, wherein the preset negative threshold groups include multiple preset negative thresholds; Based on the continuous unimproved algebra, the system iterates through the preset negative threshold group. When the continuous unimproved algebra is greater than the corresponding preset negative threshold, the temperature is increased based on multiple preset increase rates.

6. The material transfer optimization scheduling method based on a large language model according to claim 5, characterized in that, The linkage adjustment based on population size specifically involves dynamically adjusting the population size within a preset number of rounds after the temperature adjustment is completed, based on the current temperature and the severity of failure; wherein, the severity of failure is determined based on the number of consecutive generations without improvement, and the adjustment of the population size involves adding new individuals to the population.

7. The material transfer optimization scheduling method based on a large language model according to claim 4, characterized in that, The temperature regulation is also based on continuously improved algebraic triggering, including: Obtain the continuously improved algebra and its corresponding preset positive threshold; When the number of consecutive improvement generations is greater than the preset positive threshold, the temperature is rolled back based on a preset reduction rate.

8. A material transfer optimization scheduling system based on a large language model, employing the material transfer optimization scheduling method based on a large language model as described in any one of claims 1 to 7, characterized in that, include: The initialization module is used to obtain the material transfer scheduling data to be optimized from the production site data source, including at least the transfer task order, workshop map, AGV initial position and speed limit; The initial population is constructed by building a cost matrix based on material transfer and scheduling data; the individuals in the population are structured access sequence sequences, and the initialization process does not involve a large language model. The prompt word construction module is used to construct prompt words, which include at least a problem description, population context information, and task instructions. The problem description includes at least a set of pick-up and delivery points for a single automated guided vehicle, corresponding distance cost information, and task node sequence requirements. The task instructions require the output of a structured access sequence and satisfy that each point is accessed only once. The offspring generation module is used to generate offspring individuals based on prompt words by calling the large language model. The offspring individuals are structured access sequence sequences. The validity detection module is used to perform validity checks on offspring individuals. If the check fails, validity repair is triggered to obtain the repaired offspring individuals. The evaluation and selection module is used to merge offspring individuals that have passed the effectiveness test or repair with current population individuals, and to evaluate and select the next generation population based on the total driving distance of the merged individuals. The linkage adjustment module is used to dynamically switch between exploration and convergence by adjusting temperature parameters and population structure in response to the evolutionary process of material transfer optimization scheduling for a single automated guided vehicle. The optimal individual output module is used to terminate evolution and output the current optimal individual when the preset maximum number of generations is reached or the preset optimization target is met. The optimal individual is converted into a navigation instruction sequence for controlling a single automated guided vehicle to transfer materials, so as to optimize the material transfer scheduling of a single automated guided vehicle.

9. A computer device for optimizing and scheduling material transfer based on a large language model, characterized in that, Includes at least one processor, at least one memory, and a data bus; The processor and the memory communicate with each other via the data bus; The memory stores program instructions that can be executed by the processor, which invokes the program instructions to execute the material transfer optimization scheduling method based on a large language model as described in any one of claims 1 to 7.

10. A storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the material transfer optimization scheduling method based on a large language model as described in any one of claims 1 to 7.

Images