A multi-objective hybrid flow shop scheduling method and system

Abstract

The application provides a multi-target mixed flow shop scheduling method and system, and belongs to the technical field of artificial intelligence. The application adopts a configurable MOEA framework, uses a CART enhanced I / F-Race method to automatically configure an optimal algorithm, and is used for solving a multi-target mixed flow shop scheduling problem. In subsequent F-Race iterations, before a new configuration is generated by using an F-Race learning model, a previously evaluated configuration in the F-Race is first used as training data to construct a CART model, and the CART model is subsequently used to predict the performance of the new configuration. Only the configuration predicted to be potential can enter the next F-Race iteration. The batch segmentation of each batch is considered in the application, and the maximum completion time and the total number of sub-batches are simultaneously optimized.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a multi-objective hybrid flow workshop scheduling method and system. Background Technology

[0002] The Hybrid Flowshop Scheduling Problem (HFSP), a branch of flowshop scheduling problems, has attracted widespread attention from researchers due to its significant theoretical and practical value. HFSP is commonly found in various flexible manufacturing workshops, including electronics, furniture, textiles, petrochemicals, and pharmaceuticals. HFSP requires determining the sequence of operations and machine allocation for each stage based on production constraints, and even in very small-scale instances, it has been proven to be NP-hard (Non-deterministic Polynomial).

[0003] In most current research on hybrid flow shop scheduling problems, jobs are indivisible; that is, each job cannot be moved to the downstream stage until it is completed at a specific stage. However, in many real-world scenarios, this negatively impacts production efficiency, where a job consists of a series of identical processing units. Batch flow technology can shorten production cycles, but it also has implementation costs, such as the transportation and management of subbatches. In other words, the larger the number of subbatches, the smaller the maximum completion time may be, but the corresponding transportation costs will increase. Therefore, considering the transportation costs of subbatches in practice, the total number of subbatches is limited. In summary, there is usually a trade-off between the maximum completion time and the total number of subbatches, which is the main optimization objective of flow shop scheduling.

[0004] Given the significant challenges posed by the multi-objective flow shop scheduling optimization problem, numerous solutions have been proposed. Among them, Multi-Objective Evolutionary Algorithms (MOEAs) have attracted considerable attention. MOEAs are capable of obtaining a set of approximate Pareto optimal solutions in a single run, and literature studies have demonstrated their superior performance. However, no universal algorithm exists applicable to all optimization problems. Therefore, it is necessary to configure the algorithm for the specific problem to improve its performance. It is well known that the performance of MOEAs is highly dependent on the configuration of algorithm parameters, including numerical and class parameters. On the one hand, traditional configuration methods (manual or simulation tuning) are often limited by past experience, making it difficult to obtain the optimal algorithm configuration and study how algorithm components interact. On the other hand, the entire process can be very time-consuming and tedious. Therefore, finding an efficient automatic algorithm design method to eliminate practitioners' reliance on prior knowledge and achieve automatic configuration of MOEAs has significant practical implications. Automated Algorithm Design (AAD) methods are gradually becoming a research area to alleviate these limitations.

[0005] The F-Race method, proposed by Birtari et al., is an AAD (Adaptive Optimization) approach that can simultaneously optimize both numerical and categorical parameters. Inspired by the Hoeffding Race, it was initially proposed for model selection in machine learning. F-Race continuously evaluates a given set of algorithm configurations across a series of test instances, eliminating inferior configurations once sufficient statistical evidence is obtained. In F-Race, initial candidate configurations are obtained through a full factorial design. However, when the number of parameters is large and the range of values ​​for each parameter is wide, the full factorial design may become impractical and computationally limited. To alleviate this problem, Balaprakash et al. proposed an iterative application of F-Race, called Iterated F-Race (I / F-Race). By learning useful information and competition mechanisms during the iteration process, I / F-Race can identify potential regions in the configuration space and find the optimal algorithm configuration.

[0006] Since its inception, I / F-Race has become a highly competitive AAD method. High-performance continuous optimization solvers have been successfully configured based on different functional test sets. Some researchers have also used I / F-Race to configure ant colony algorithms to solve different traveling salesman problems.

[0007] I / F-Race also has limitations. In a single F-Race iteration, given a set of test instances, configurations that statistically perform worse than other configurations are discarded. Test instances that have already been used cannot be reused in subsequent F-Race iterations. Furthermore, we know that a configuration that performs well on one instance may perform poorly on others. Therefore, historical evaluation information of configurations should be retained to comprehensively evaluate algorithm configurations.

[0008] The existing technology has at least the following shortcomings:

[0009] 1. Existing methods do not fully consider the variable batching characteristic. When there are start-up operations between different batches, frequent switching is required, which seriously affects processing efficiency.

[0010] 2. In a single F-Race iteration, given a set of test instances, the test instances that have already been used cannot be used in subsequent F-Race iterations. Therefore, it is impossible to find an algorithm configuration that fully matches the characteristics of the hybrid flow shop scheduling problem that takes into account the variable batching. Summary of the Invention

[0011] This invention aims to solve the existing multi-objective hybrid flow shop scheduling problem considering variable batching, while simultaneously optimizing maximum completion time and the total number of sub-batches. This invention provides a multi-objective hybrid flow shop scheduling method. Based on variable batching and the production characteristics of actual hybrid flow shops, a multi-objective hybrid flow shop scheduling model considering variable batching is established, including two objectives: maximum completion time and the total number of sub-batches. A configurable MOEA (CMOEA) framework is adopted. In subsequent F-Race iterations, before generating new configurations using the F-Race learning model, the configurations evaluated in previous F-Race iterations are first used as training data to construct a CART (Class and Regression Tree) model. To utilize historical information to improve sampling quality, this model is then used to predict the performance of new configurations. Only configurations predicted to have potential are included in the next F-Race iteration. Both the F-Race learning model and the CART model are based on real-time data and jointly undertake the core task of finding the optimal algorithm configuration, achieving the scheduling objectives of optimizing maximum completion time and the total number of sub-batches. This invention utilizes real-time data, which can intuitively reflect the impact of each parameter on the performance of the algorithm configuration.

[0012] This invention provides a multi-objective hybrid flow shop scheduling method, comprising the following steps:

[0013] Based on the actual production characteristics of the hybrid flow workshop, a multi-objective hybrid flow workshop scheduling model is established, including two objectives: maximum completion time and total number of sub-batches.

[0014] Build a configurable MOEA framework and collect multiple test cases of hybrid flow shop scheduling;

[0015] The optimal algorithm configuration is obtained by iterating through the following process:

[0016] Step 01: Based on test cases, select excellent algorithm configurations; determine the excellence of the algorithm according to the preset threshold of the optimization index;

[0017] Step 02: Construct the F-Race learning model based on a superior algorithm configuration;

[0018] Step 03: Build a CART model based on the tested algorithm configuration;

[0019] Step 04: Build a completely new algorithm configuration using the F-Race learning model;

[0020] Step 05: Use the CART model to predict the performance of the new algorithm configuration;

[0021] Step 06: Determine if the iteration termination condition is met. If it is met, output the optimal algorithm configuration. If it is not met, proceed to the next iteration based on the predicted high-performance algorithm configuration and execute step 01.

[0022] The F-Race learning model is used to generate new algorithm configurations, while the CART model is used to predict the performance of these new configurations. Both models work together to optimize the algorithm configurations.

[0023] The configurable MOEA framework selects a population evolution method based on the currently configured fitness evaluation algorithm, and adopts cooperative search, local search and population restart strategies to output the Pareto optimal solution for the current instance. The fitness evaluation algorithms include: the decomposition-based method DBM, the non-dominated sorting-based method NSBM and the index-based method IBM.

[0024] Preferably, constructing a CART model based on real-time data instances generated during the optimization process includes: using the evaluated algorithm configuration as training data, using parameter values ​​as features, using evaluation results as labels, and determining the label type based on a preset threshold for the optimization index. The label types include "excellent" and "poor".

[0025] Preferably, for the new configuration generated by the F-Race learning model, the CART model is used to predict the configuration performance, specifically including the following steps:

[0026] Step S001: Construct the CART model based on the evaluated algorithm configuration;

[0027] Step S002: Sample the F-Race learning model to generate a series of new algorithm configurations;

[0028] Step S003: Use the CART model to predict the new algorithm configurations. Algorithm configurations predicted as "good" are retained, and algorithm configurations predicted as "poor" are discarded.

[0029] Step S004: Mine the algorithm configuration information predicted as "good" and produce a new F-Race learning model.

[0030] Preferably, the configurable MOEA framework algorithm includes the following steps:

[0031] Initialize the population P according to the configurable initialization method;

[0032] The iterative process includes:

[0033] Depending on the configurable fitness value evaluation method, mating individuals are selected in different neighborhoods to generate offspring populations (OPs).

[0034] The target values ​​of individuals in population P and offspring population OP are obtained through cooperative search using a configurable decoding method, and the population is evaluated accordingly based on the different fitness value evaluation methods configured.

[0035] Depending on the different fitness value evaluation methods used for the configuration, appropriate environment selection should be made;

[0036] A variable neighborhood descent strategy is used for local search. If the number of consecutive failed updates using the neighborhood structure reaches the preset maximum number of consecutive failed updates, the neighborhood structure will switch to the next neighborhood.

[0037] If the number of consecutive failed updates using cooperative search and neighborhood structure is less than the maximum number of consecutive failed updates, execute the population restart method for each individual in population P.

[0038] If the iteration termination condition is met, the iteration ends and the Pareto nondominated solution found during the evolution process is output; otherwise, the iteration process returns to normal.

[0039] Preferably, during population initialization, if the fitness value evaluation method is a decomposition-based method (DBM), a set of uniformly distributed weight vectors is generated, and each individual is assigned a unique weight vector.

[0040] Preferably, depending on the configured fitness value evaluation method, mating individuals are selected in different neighborhoods to generate offspring populations, specifically including:

[0041] Depending on the configured fitness value evaluation method, a configurable cooperative method is used to generate a progeny population of a preset population size ps by applying it to the mating individuals selected from the population. Specifically:

[0042] If the fitness value evaluation method is the decomposition-based method DBM, then each individual selects a mating target in a neighborhood of a preset neighborhood size ns, and executes a configurable cooperative method to produce offspring until a subpopulation of a preset population size ps is produced.

[0043] If the fitness value evaluation method is the non-dominated ranking method NSBM and the index-based method IBM, then two mating individuals are selected in the tournament rules with a preset tournament size ts, and the cooperative method is used to produce offspring. This process is repeated with a preset population size ps times to produce a progeny population of the preset population size ps.

[0044] Preferably, the target values ​​of individuals in the population P and the offspring population OP are obtained through cooperative search using a decoding method, and corresponding population evaluations are performed based on different configured fitness value evaluation methods, specifically including:

[0045] The target values ​​of individuals in the population P and the offspring population OP are obtained using a configurable decoding method;

[0046] Depending on the configured fitness evaluation method, assign a fitness value using the corresponding fitness evaluation method:

[0047] If the fitness value evaluation method is the decomposition-based method DBM, then the decomposition-based method DBM shall be used for fitness value evaluation.

[0048] If the fitness value evaluation method is the non-dominated ranking method NSBM, then the fitness value evaluation method NSBM based on non-dominated ranking is used.

[0049] If the fitness value evaluation method is the indicator-based IBM method, then the indicator-based IBM method will be used for fitness value evaluation.

[0050] Preferably, depending on the configured fitness value evaluation method, corresponding environment selection is performed. Specifically, based on the different configured fitness value evaluation methods, the population P is updated using the offspring population OP through different corresponding environment selection methods, including:

[0051] If the fitness value evaluation method is the decomposition-based method DBM, then the decomposition-based method DBM is used to update the population P using the offspring population OP. Each individual in the offspring population OP is compared with individuals in the population P that have the same weight vector and their neighborhood. If an individual in the offspring population OP has a better fitness value, then the individual in the population P is updated to be an individual in the offspring population OP.

[0052] If the fitness value evaluation method is based on the non-dominated sorting method NSBM, then the offspring population OP is merged into the population P, and the individuals in the population are sorted according to the fast non-dominated sorting method, retaining the first preset population size ps individuals in the population P.

[0053] If the fitness value evaluation method is the index-based method IBM, then the index-based method IBM is used, the offspring population OP is merged into the population P, and the individuals with the lowest fitness values ​​are removed one by one until the population P has a preset population size of ps individuals.

[0054] Preferably, if the number of consecutive failed updates using cooperative search and neighborhood structure is less than the maximum number of consecutive failed updates, a population restart method is executed for each individual in population P. Specifically, a configurable population restart method is used to generate a replacement for each individual in population P. Otherwise, the individuals in population P remain unchanged.

[0055] This invention provides a multi-objective hybrid flow workshop scheduling system, using any of the above-mentioned multi-objective hybrid flow workshop scheduling methods, including multiple flow workshop devices to be scheduled and a control scheduling module;

[0056] The control and scheduling module establishes a multi-objective hybrid flow workshop scheduling model based on the actual production characteristics of the hybrid flow workshop, including two objectives: maximum completion time and total number of sub-batches. It constructs a configurable MOEA framework and collects multiple test cases for hybrid flow workshop scheduling. Based on these test cases, it selects superior algorithms and builds an F-Race learning model based on the configurations of these superior algorithms. It then builds a CART model based on the tested algorithm configurations. Finally, it uses the F-Race learning model to construct entirely new algorithm configurations and uses the CART model to predict the performance of these new configurations. Based on the predicted superior algorithms, it iterates repeatedly to obtain the optimal algorithm configuration. Finally, it selects the equipment to be scheduled in the flow workshop according to the optimal algorithm configuration and issues scheduling commands.

[0057] After receiving the scheduling instructions from the control and scheduling module, the equipment in the assembly line workshop that is to be scheduled will perform the corresponding operations.

[0058] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0059] (1) This invention introduces variable batching into HFSP, which takes into account not only batch order and machine allocation, but also batch segmentation (i.e., number of sub-batches and sub-batch size) of each batch, thereby achieving simultaneous optimization and maximization of completion time and total number of sub-batches.

[0060] (2) Based on the configurable algorithm framework, this invention introduces a CART-enhanced I / F-Race model to automatically construct a high-performance MOEA. Compared with other methods, the automatically generated MOEA performs very well and can achieve a better balance between maximum completion time and total number of sub-batches.

[0061] (3) The CART model developed in this invention is used to predict the performance of newly generated configurations. Only when the performance of a new configuration is predicted to be "good" will it have a chance to be selected. Otherwise, if its performance is predicted to be "poor", it will be ignored. The developed method has two advantages: first, it makes full use of the real-time data generated during the optimization process; second, it intuitively reflects the impact of parameter settings on configuration performance and can explore the characteristics of mixed flow shop scheduling problems that consider variable batching.

[0062] (4) In the local search phase, the present invention adopts a variable neighborhood descent strategy. If the number of consecutive update failures of an individual reaches the preset maximum number of consecutive update failures, it means that the current neighborhood structure may no longer have the potential to be utilized. In this case, it will switch to the next structure in the neighborhood structure set in the next generation. The present invention can systematically switch between various neighborhood structures, thereby making full use of each neighborhood structure and avoiding wasting computing resources. It can perform fast, comprehensive and efficient search of the solution space for the mixed flow shop scheduling problem considering variable batching.

[0063] (5) In the population restart phase, when multiple consecutive updates using the cooperative method and the local search method fail to update individuals in the population, the present invention uses the restart method to generate a replacement for the individual. Otherwise, the individual remains unchanged, which can help the individual escape the local optimal deadlock and escape the local optimal solution of the mixed flow shop scheduling problem considering variable batching. Attached Figure Description

[0064] Figure 1 This is a flowchart of a multi-objective hybrid flow workshop scheduling method according to an embodiment of the present invention. Detailed Implementation

[0065] The following is in conjunction with the appendix Figure 1 The specific embodiments of the present invention will be described in detail below.

[0066] This invention provides a multi-objective hybrid flow shop scheduling method, comprising the following steps:

[0067] Based on the actual production characteristics of the hybrid flow workshop, a multi-objective hybrid flow workshop scheduling model is established, including two objectives: maximum completion time and total number of sub-batches.

[0068] Build a configurable MOEA framework and collect multiple test cases of hybrid flow shop scheduling;

[0069] The optimal algorithm configuration is obtained by iterating through the following process:

[0070] Step 01: Based on test cases, select excellent algorithm configurations; determine the excellence of the algorithm according to the preset threshold of the optimization index;

[0071] Step 02: Construct the F-Race learning model based on a superior algorithm configuration;

[0072] Step 03: Build a CART model based on the tested algorithm configuration;

[0073] Step 04: Build a completely new algorithm configuration using the F-Race learning model;

[0074] Step 05: Use the CART model to predict the performance of the new algorithm configuration;

[0075] Step 06: Determine if the iteration termination condition is met. If it is met, output the optimal algorithm configuration. If it is not met, proceed to the next iteration based on the predicted high-performance algorithm configuration and execute step 01.

[0076] The F-Race learning model is used to generate new algorithm configurations, while the CART model is used to predict the performance of these new configurations. Both models work together to optimize the algorithm configurations.

[0077] The configurable MOEA framework selects a population evolution method based on the currently configured fitness evaluation algorithm, and adopts cooperative search, local search and population restart strategies to output the Pareto optimal solution for the current instance. The fitness evaluation algorithms include: the decomposition-based method DBM, the non-dominated sorting-based method NSBM and the index-based method IBM.

[0078] According to a specific embodiment of the present invention, constructing a CART model based on real-time data instances generated during the optimization process includes: using the evaluated algorithm configuration as training data, using parameter values ​​as features, using evaluation results as labels, and determining the label type according to a preset threshold of the optimization index. The label types include "excellent" and "poor".

[0079] According to a specific embodiment of the present invention, for a new configuration generated by the F-Race learning model, the CART model is used to predict the configuration performance, specifically including the following steps:

[0080] Step S001: Construct the CART model based on the evaluated algorithm configuration;

[0081] Step S002: Sample the F-Race learning model to generate a series of new algorithm configurations;

[0082] Step S003: Use the CART model to predict the new algorithm configurations. Algorithm configurations predicted as "good" are retained, and algorithm configurations predicted as "poor" are discarded.

[0083] Step S004: Mine the algorithm configuration information predicted as "good" and produce a new F-Race learning model.

[0084] According to a specific embodiment of the present invention, the configurable MOEA framework algorithm includes the following steps:

[0085] Initialize the population P according to the configurable initialization method;

[0086] The iterative process includes:

[0087] Depending on the configurable fitness value evaluation method, mating individuals are selected in different neighborhoods to generate offspring populations (OPs).

[0088] The target values ​​of individuals in population P and offspring population OP are obtained through cooperative search using a configurable decoding method, and the population is evaluated accordingly based on the different fitness value evaluation methods configured.

[0089] Depending on the different fitness value evaluation methods used for the configuration, appropriate environment selection should be made;

[0090] A variable neighborhood descent strategy is used for local search. If the number of consecutive failed updates using the neighborhood structure reaches the preset maximum number of consecutive failed updates, the neighborhood structure will switch to the next neighborhood.

[0091] If the number of consecutive failed updates using cooperative search and neighborhood structure is less than the maximum number of consecutive failed updates, execute the population restart method for each individual in population P.

[0092] If the iteration termination condition is met, the iteration ends and the Pareto nondominated solution found during the evolution process is output; otherwise, the iteration process returns to normal.

[0093] According to a specific embodiment of the present invention, during the initialization of the population, if the fitness value evaluation method is a decomposition-based method (DBM), a set of uniformly distributed weight vectors is generated, and each individual is assigned a unique weight vector.

[0094] According to a specific embodiment of the present invention, depending on the configured fitness value evaluation method, mating individuals are selected in different neighborhoods to generate a progeny population, specifically including:

[0095] Depending on the configured fitness value evaluation method, a configurable cooperative method is used to generate a progeny population of a preset population size ps by applying it to the mating individuals selected from the population. Specifically:

[0096] If the fitness value evaluation method is the decomposition-based method DBM, then each individual selects a mating target in a neighborhood of a preset neighborhood size ns, and executes a configurable cooperative method to produce offspring until a subpopulation of a preset population size ps is produced.

[0097] If the fitness value evaluation method is the non-dominated ranking method NSBM and the index-based method IBM, then two mating individuals are selected in the tournament rules with a preset tournament size ts, and the cooperative method is used to produce offspring. This process is repeated with a preset population size ps times to produce a progeny population of the preset population size ps.

[0098] According to a specific embodiment of the present invention, a decoding method is used to obtain the target values ​​of individuals in population P and offspring population OP through cooperative search, and corresponding population evaluation is performed according to different configured fitness value evaluation methods, specifically including:

[0099] The target values ​​of individuals in the population P and the offspring population OP are obtained using a configurable decoding method;

[0100] Depending on the configured fitness evaluation method, assign a fitness value using the corresponding fitness evaluation method:

[0101] If the fitness value evaluation method is the decomposition-based method DBM, then the decomposition-based method DBM shall be used for fitness value evaluation.

[0102] If the fitness value evaluation method is the non-dominated ranking method NSBM, then the fitness value evaluation method NSBM based on non-dominated ranking is used.

[0103] If the fitness value evaluation method is the indicator-based IBM method, then the indicator-based IBM method will be used for fitness value evaluation.

[0104] According to a specific embodiment of the present invention, corresponding environment selection is performed based on different configured fitness value evaluation methods. Specifically, the population P is updated using the offspring population OP through different corresponding environment selection methods based on the different configured fitness value evaluation methods, specifically including:

[0105] If the fitness value evaluation method is the decomposition-based method DBM, then the decomposition-based method DBM is used to update the population P using the offspring population OP. Each individual in the offspring population OP is compared with individuals in the population P that have the same weight vector and their neighborhood. If an individual in the offspring population OP has a better fitness value, then the individual in the population P is updated to be an individual in the offspring population OP.

[0106] If the fitness value evaluation method is based on the non-dominated sorting method NSBM, then the offspring population OP is merged into the population P, and the individuals in the population are sorted according to the fast non-dominated sorting method, retaining the first preset population size ps individuals in the population P.

[0107] If the fitness value evaluation method is the index-based method IBM, then the index-based method IBM is used, the offspring population OP is merged into the population P, and the individuals with the lowest fitness values ​​are removed one by one until the population P has a preset population size of ps individuals.

[0108] According to a specific embodiment of the present invention, if the number of consecutive failed updates using cooperative search and neighborhood structure is less than the maximum number of consecutive failed updates, a population restart method is performed on each individual in population P. Specifically, a configurable population restart method is used to generate a substitute to replace each individual in population P; otherwise, the individuals in population P remain unchanged.

[0109] This invention provides a multi-objective hybrid flow workshop scheduling system, using any of the above-mentioned multi-objective hybrid flow workshop scheduling methods, including multiple flow workshop devices to be scheduled and a control scheduling module;

[0110] The control and scheduling module establishes a multi-objective hybrid flow workshop scheduling model based on the actual production characteristics of the hybrid flow workshop, including two objectives: maximum completion time and total number of sub-batches. It constructs a configurable MOEA framework and collects multiple test cases for hybrid flow workshop scheduling. Based on these test cases, it selects superior algorithms and builds an F-Race learning model based on the configurations of these superior algorithms. It then builds a CART model based on the tested algorithm configurations. Finally, it uses the F-Race learning model to construct entirely new algorithm configurations and uses the CART model to predict the performance of these new configurations. Based on the predicted superior algorithms, it iterates repeatedly to obtain the optimal algorithm configuration. Finally, it selects the equipment to be scheduled in the flow workshop according to the optimal algorithm configuration and issues scheduling commands.

[0111] After receiving the scheduling instructions from the control and scheduling module, the equipment in the assembly line workshop that is to be scheduled will perform the corresponding operations.

[0112] Example 1

[0113] According to a specific embodiment of the present invention, the multi-objective hybrid flow workshop scheduling method of the present invention will be described in detail with reference to the accompanying drawings.

[0114] This invention provides a multi-objective hybrid flow shop scheduling method, comprising the following steps:

[0115] Based on the actual production characteristics of the hybrid flow workshop, a multi-objective hybrid flow workshop scheduling model is established, including two objectives: maximum completion time and total number of sub-batches.

[0116] Build a configurable MOEA framework and collect multiple test cases of hybrid flow shop scheduling;

[0117] The optimal algorithm configuration is obtained by iterating through the following process:

[0118] Step 01: Based on test cases, select excellent algorithm configurations; determine the excellence of the algorithm according to the preset threshold of the optimization index;

[0119] Step 02: Construct the F-Race learning model based on a superior algorithm configuration;

[0120] Step 03: Build a CART model based on the tested algorithm configuration;

[0121] Step 04: Build a completely new algorithm configuration using the F-Race learning model;

[0122] Step 05: Use the CART model to predict the performance of the new algorithm configuration;

[0123] Step 06: Determine if the iteration termination condition is met. If it is met, output the optimal algorithm configuration. If it is not met, proceed to the next iteration based on the predicted high-performance algorithm configuration and execute step 01.

[0124] The F-Race learning model is used to generate new algorithm configurations, while the CART model is used to predict the performance of these new configurations. Both models work together to optimize the algorithm configurations.

[0125] The configurable MOEA framework selects a population evolution method based on the currently configured fitness evaluation algorithm, and adopts cooperative search, local search and population restart strategies to output the Pareto optimal solution for the current instance. The fitness evaluation algorithms include: the decomposition-based method DBM, the non-dominated sorting-based method NSBM and the index-based method IBM.

[0126] Example 2

[0127] According to a specific embodiment of the present invention, the multi-objective hybrid flow workshop scheduling method of the present invention will be described in detail with reference to the accompanying drawings.

[0128] This invention provides a multi-objective hybrid flow shop scheduling method, comprising the following steps:

[0129] Based on the actual production characteristics of the hybrid flow workshop, a multi-objective hybrid flow workshop scheduling model is established, including two objectives: maximum completion time and total number of sub-batches.

[0130] Build a configurable MOEA framework and collect multiple test cases of hybrid flow shop scheduling;

[0131] The optimal algorithm configuration is obtained by iterating through the following process:

[0132] Step 01: Based on test cases, select excellent algorithm configurations; determine the excellence of the algorithm according to the preset threshold of the optimization index;

[0133] Step 02: Construct the F-Race learning model based on a superior algorithm configuration;

[0134] Step 03: Build a CART model based on the tested algorithm configuration;

[0135] Step 04: Build a completely new algorithm configuration using the F-Race learning model;

[0136] Step 05: Use the CART model to predict the performance of the new algorithm configuration;

[0137] Step 06: Determine if the iteration termination condition is met. If it is met, output the optimal algorithm configuration. If it is not met, proceed to the next iteration based on the predicted high-performance algorithm configuration and execute step 01.

[0138] The F-Race learning model is used to generate new algorithm configurations, while the CART model is used to predict the performance of these new configurations. Both models work together to optimize the algorithm configurations.

[0139] The configurable MOEA framework selects a population evolution method based on the currently configured fitness evaluation algorithm, and adopts cooperative search, local search and population restart strategies to output the Pareto optimal solution for the current instance. The fitness evaluation algorithms include: the decomposition-based method DBM, the non-dominated sorting-based method NSBM and the index-based method IBM.

[0140] The process of building a CART model based on real-time data instances generated during optimization includes: using the evaluated algorithm configuration as training data, using parameter values ​​as features, using evaluation results as labels, and determining the label type based on the preset threshold of the optimization index. The label types include "excellent" and "poor".

[0141] The configurable MOEA framework algorithm includes the following steps:

[0142] Initialize the population P according to the configurable initialization method;

[0143] The iterative process includes:

[0144] Depending on the configurable fitness value evaluation method, mating individuals are selected in different neighborhoods to generate offspring populations (OPs).

[0145] The target values ​​of individuals in population P and offspring population OP are obtained through cooperative search using configurable decoding methods, and corresponding population evaluations are performed for different configurable fitness value evaluation methods.

[0146] Depending on the different fitness value evaluation methods used for the configuration, appropriate environment selection should be made;

[0147] A variable neighborhood descent strategy is used for local search. If the number of consecutive failed updates using the neighborhood structure reaches the preset maximum number of consecutive failed updates, the neighborhood structure will switch to the next neighborhood.

[0148] If the number of consecutive failed updates using cooperative search and neighborhood structure is less than the maximum number of consecutive failed updates, execute the population restart method for each individual in population P.

[0149] If the iteration termination condition is met, the iteration ends and the Pareto nondominated solution found during the evolution process is output; otherwise, the iteration process returns to normal.

[0150] Example 3

[0151] According to a specific embodiment of the present invention, the multi-objective hybrid flow workshop scheduling method of the present invention will be described in detail with reference to the accompanying drawings.

[0152] This invention provides a multi-objective hybrid flow shop scheduling method, comprising the following steps:

[0153] Based on the actual production characteristics of the hybrid flow workshop, a multi-objective hybrid flow workshop scheduling model is established, including two objectives: maximum completion time and total number of sub-batches.

[0154] Build a configurable MOEA framework and collect multiple test cases of hybrid flow shop scheduling;

[0155] The optimal algorithm configuration is obtained by iterating through the following process:

[0156] Step 01: Based on test cases, select excellent algorithm configurations; determine the excellence of the algorithm according to the preset threshold of the optimization index;

[0157] Step 02: Construct the F-Race learning model based on a superior algorithm configuration;

[0158] Step 03: Build a CART model based on the tested algorithm configuration;

[0159] Step 04: Build a completely new algorithm configuration using the F-Race learning model;

[0160] Step 05: Use the CART model to predict the performance of the new algorithm configuration;

[0161] Step 06: Determine if the iteration termination condition is met. If it is met, output the optimal algorithm configuration. If it is not met, proceed to the next iteration based on the predicted high-performance algorithm configuration and execute step 01.

[0162] The F-Race learning model is used to generate new algorithm configurations, while the CART model is used to predict the performance of these new configurations. Both models work together to optimize the algorithm configurations.

[0163] The configurable MOEA framework selects a population evolution method based on the currently configured fitness evaluation algorithm, and adopts cooperative search, local search and population restart strategies to output the Pareto optimal solution for the current instance. The fitness evaluation algorithms include: the decomposition-based method DBM, the non-dominated sorting-based method NSBM and the index-based method IBM.

[0164] The process of building a CART model based on real-time data instances generated during optimization includes: using the evaluated algorithm configuration as training data, using parameter values ​​as features, using evaluation results as labels, and determining the label type based on the preset threshold of the optimization index. The label types include "excellent" and "poor".

[0165] The configurable MOEA framework algorithm includes the following steps:

[0166] Initialize the population P according to the configurable initialization method;

[0167] The iterative process includes:

[0168] Depending on the configurable fitness value evaluation method, mating individuals are selected in different neighborhoods to generate offspring populations (OPs).

[0169] The target values ​​of individuals in population P and offspring population OP are obtained through cooperative search using configurable decoding methods, and corresponding population evaluations are performed for different configurable fitness value evaluation methods.

[0170] Depending on the different fitness value evaluation methods used for the configuration, appropriate environment selection should be made;

[0171] A variable neighborhood descent strategy is used for local search. If the number of consecutive failed updates using the neighborhood structure reaches the preset maximum number of consecutive failed updates, the neighborhood structure will switch to the next neighborhood.

[0172] If the number of consecutive failed updates using cooperative search and neighborhood structure is less than the maximum number of consecutive failed updates, execute the population restart method for each individual in population P.

[0173] If the iteration termination condition is met, the iteration ends and the Pareto nondominated solution found during the evolution process is output; otherwise, the iteration process returns to normal.

[0174] Specifically, for the new configurations generated by the F-Race learning model, the CART model is used to predict the configuration performance, which includes the following steps:

[0175] Step S001: Construct the CART model based on the evaluated algorithm configuration;

[0176] Step S002: Sample the F-Race learning model to generate a series of new algorithm configurations;

[0177] Step S003: Use the CART model to predict the new algorithm configurations. Algorithm configurations predicted as "good" are retained, and algorithm configurations predicted as "poor" are discarded.

[0178] Step S004: Mine the algorithm configuration information predicted as "good" and produce a new F-Race learning model.

[0179] In the process of initializing the population, if the fitness value evaluation method is the decomposition-based method DBM, a set of uniformly distributed weight vectors are generated, and each individual is assigned a unique weight vector.

[0180] If the number of consecutive failed updates using cooperative search and neighborhood structure is less than the maximum number of consecutive failed updates, a population restart method is executed for each individual in population P. Specifically, a configurable population restart method is used to generate a replacement for each individual in population P. Otherwise, the individuals in population P remain unchanged.

[0181] Example 4

[0182] According to a specific embodiment of the present invention, the multi-objective hybrid flow workshop scheduling method of the present invention will be described in detail with reference to the accompanying drawings.

[0183] This invention provides a multi-objective hybrid flow shop scheduling method, comprising the following steps:

[0184] Based on the actual production characteristics of the hybrid flow workshop, a multi-objective hybrid flow workshop scheduling model is established, including two objectives: maximum completion time and total number of sub-batches.

[0185] Build a configurable MOEA framework and collect multiple test cases of hybrid flow shop scheduling;

[0186] The optimal algorithm configuration is obtained by iterating through the following process:

[0187] Step 01: Based on test cases, select excellent algorithm configurations; determine the excellence of the algorithm according to the preset threshold of the optimization index;

[0188] Step 02: Construct the F-Race learning model based on a superior algorithm configuration;

[0189] Step 03: Build a CART model based on the tested algorithm configuration;

[0190] Step 04: Build a completely new algorithm configuration using the F-Race learning model;

[0191] Step 05: Use the CART model to predict the performance of the new algorithm configuration;

[0192] Step 06: Determine if the iteration termination condition is met. If it is met, output the optimal algorithm configuration. If it is not met, proceed to the next iteration based on the predicted high-performance algorithm configuration and execute step 01.

[0193] The F-Race learning model is used to generate new algorithm configurations, while the CART model is used to predict the performance of these new configurations. Both models work together to optimize the algorithm configurations.

[0194] The configurable MOEA framework selects a population evolution method based on the currently configured fitness evaluation algorithm, and adopts cooperative search, local search and population restart strategies to output the Pareto optimal solution for the current instance. The fitness evaluation algorithms include: the decomposition-based method DBM, the non-dominated sorting-based method NSBM and the index-based method IBM.

[0195] The process of building a CART model based on real-time data instances generated during optimization includes: using the evaluated algorithm configuration as training data, using parameter values ​​as features, using evaluation results as labels, and determining the label type based on the preset threshold of the optimization index. The label types include "excellent" and "poor".

[0196] The configurable MOEA framework algorithm includes the following steps:

[0197] Initialize the population P according to the configurable initialization method;

[0198] The iterative process includes:

[0199] Depending on the configurable fitness value evaluation method, mating individuals are selected in different neighborhoods to generate offspring populations (OPs).

[0200] The target values ​​of individuals in population P and offspring population OP are obtained through cooperative search using configurable decoding methods, and corresponding population evaluations are performed for different configurable fitness value evaluation methods.

[0201] Depending on the different fitness value evaluation methods used for the configuration, appropriate environment selection should be made;

[0202] A variable neighborhood descent strategy is used for local search. If the number of consecutive failed updates using the neighborhood structure reaches the preset maximum number of consecutive failed updates, the neighborhood structure will switch to the next neighborhood.

[0203] If the number of consecutive failed updates using cooperative search and neighborhood structure is less than the maximum number of consecutive failed updates, execute the population restart method for each individual in population P.

[0204] If the iteration termination condition is met, the iteration ends and the Pareto nondominated solution found during the evolution process is output; otherwise, the iteration process returns to normal.

[0205] Specifically, for the new configurations generated by the F-Race learning model, the CART model is used to predict the configuration performance, which includes the following steps:

[0206] Step S001: Construct the CART model based on the evaluated algorithm configuration;

[0207] Step S002: Sample the F-Race learning model to generate a series of new algorithm configurations;

[0208] Step S003: Use the CART model to predict the new algorithm configurations. Algorithm configurations predicted as "good" are retained, and algorithm configurations predicted as "poor" are discarded.

[0209] Step S004: Mine the algorithm configuration information predicted as "good" and produce a new F-Race learning model.

[0210] In the process of initializing the population, if the fitness value evaluation method is the decomposition-based method DBM, a set of uniformly distributed weight vectors are generated, and each individual is assigned a unique weight vector.

[0211] If the number of consecutive failed updates using cooperative search and neighborhood structure is less than the maximum number of consecutive failed updates, a population restart method is executed for each individual in population P. Specifically, a configurable population restart method is used to generate a replacement for each individual in population P. Otherwise, the individuals in population P remain unchanged.

[0212] Depending on the configured fitness value evaluation method, mating individuals are selected within different neighborhoods to generate offspring populations, specifically including:

[0213] Depending on the configured fitness value evaluation method, a configurable cooperative method is used to generate a progeny population of a preset population size ps by applying it to the mating individuals selected from the population. Specifically:

[0214] If the fitness value evaluation method is the decomposition-based method DBM, then each individual selects a mating target in a neighborhood of a preset neighborhood size ns, and executes a configurable cooperative method to produce offspring until a subpopulation of a preset population size ps is produced.

[0215] If the fitness value evaluation method is the non-dominated ranking method NSBM and the index-based method IBM, then two mating individuals are selected in the tournament rules with a preset tournament size ts, and the cooperative method is used to produce offspring. This process is repeated with a preset population size ps times to produce a progeny population of the preset population size ps.

[0216] Example 5

[0217] According to a specific embodiment of the present invention, the multi-objective hybrid flow workshop scheduling method of the present invention will be described in detail with reference to the accompanying drawings.

[0218] This invention provides a multi-objective hybrid flow shop scheduling method, comprising the following steps:

[0219] Based on the actual production characteristics of the hybrid flow workshop, a multi-objective hybrid flow workshop scheduling model is established, including two objectives: maximum completion time and total number of sub-batches.

[0220] Build a configurable MOEA framework and collect multiple test cases of hybrid flow shop scheduling;

[0221] The optimal algorithm configuration is obtained by iterating through the following process:

[0222] Step 01: Based on test cases, select excellent algorithm configurations; determine the excellence of the algorithm according to the preset threshold of the optimization index;

[0223] Step 02: Construct the F-Race learning model based on a superior algorithm configuration;

[0224] Step 03: Build a CART model based on the tested algorithm configuration;

[0225] Step 04: Build a completely new algorithm configuration using the F-Race learning model;

[0226] Step 05: Use the CART model to predict the performance of the new algorithm configuration;

[0227] Step 06: Determine if the iteration termination condition is met. If it is met, output the optimal algorithm configuration. If it is not met, proceed to the next iteration based on the predicted high-performance algorithm configuration and execute step 01.

[0228] The F-Race learning model is used to generate new algorithm configurations, while the CART model is used to predict the performance of these new configurations. Both models work together to optimize the algorithm configurations.

[0229] The configurable MOEA framework selects a population evolution method based on the currently configured fitness evaluation algorithm, and adopts cooperative search, local search and population restart strategies to output the Pareto optimal solution for the current instance. The fitness evaluation algorithms include: the decomposition-based method DBM, the non-dominated sorting-based method NSBM and the index-based method IBM.

[0230] The process of building a CART model based on real-time data instances generated during optimization includes: using the evaluated algorithm configuration as training data, using parameter values ​​as features, using evaluation results as labels, and determining the label type based on the preset threshold of the optimization index. The label types include "excellent" and "poor".

[0231] The configurable MOEA framework algorithm includes the following steps:

[0232] Initialize the population P according to the configurable initialization method;

[0233] The iterative process includes:

[0234] Depending on the configurable fitness value evaluation method, mating individuals are selected in different neighborhoods to generate offspring populations (OPs).

[0235] The target values ​​of individuals in population P and offspring population OP are obtained through cooperative search using configurable decoding methods, and corresponding population evaluations are performed for different configurable fitness value evaluation methods.

[0236] Depending on the different fitness value evaluation methods used for the configuration, appropriate environment selection should be made;

[0237] A variable neighborhood descent strategy is used for local search. If the number of consecutive failed updates using the neighborhood structure reaches the preset maximum number of consecutive failed updates, the neighborhood structure will switch to the next neighborhood.

[0238] If the number of consecutive failed updates using cooperative search and neighborhood structure is less than the maximum number of consecutive failed updates, execute the population restart method for each individual in population P.

[0239] If the iteration termination condition is met, the iteration ends and the Pareto nondominated solution found during the evolution process is output; otherwise, the iteration process returns to normal.

[0240] Specifically, for the new configurations generated by the F-Race learning model, the CART model is used to predict the configuration performance, which includes the following steps:

[0241] Step S001: Construct the CART model based on the evaluated algorithm configuration;

[0242] Step S002: Sample the F-Race learning model to generate a series of new algorithm configurations;

[0243] Step S003: Use the CART model to predict the new algorithm configurations. Algorithm configurations predicted as "good" are retained, and algorithm configurations predicted as "poor" are discarded.

[0244] Step S004: Mine the algorithm configuration information predicted as "good" and produce a new F-Race learning model.

[0245] In the process of initializing the population, if the fitness value evaluation method is the decomposition-based method DBM, a set of uniformly distributed weight vectors are generated, and each individual is assigned a unique weight vector.

[0246] If the number of consecutive failed updates using cooperative search and neighborhood structure is less than the maximum number of consecutive failed updates, a population restart method is executed for each individual in population P. Specifically, a configurable population restart method is used to generate a replacement for each individual in population P. Otherwise, the individuals in population P remain unchanged.

[0247] Specifically, the target values ​​of individuals in the population P and the offspring population OP are obtained through cooperative search using a decoding method, and corresponding population evaluations are performed based on different configured fitness value evaluation methods, including:

[0248] The target values ​​of individuals in the population P and the offspring population OP are obtained using a configurable decoding method;

[0249] Depending on the configured fitness evaluation method, assign a fitness value using the corresponding fitness evaluation method:

[0250] If the fitness value evaluation method is the decomposition-based method DBM, then the decomposition-based method DBM shall be used for fitness value evaluation.

[0251] If the fitness value evaluation method is the non-dominated ranking method NSBM, then the fitness value evaluation method NSBM based on non-dominated ranking is used.

[0252] If the fitness value evaluation method is the indicator-based IBM method, then the indicator-based IBM method will be used for fitness value evaluation.

[0253] Example 6

[0254] According to a specific embodiment of the present invention, the multi-objective hybrid flow workshop scheduling method of the present invention will be described in detail with reference to the accompanying drawings.

[0255] This invention provides a multi-objective hybrid flow shop scheduling method, comprising the following steps:

[0256] Based on the actual production characteristics of the hybrid flow workshop, a multi-objective hybrid flow workshop scheduling model is established, including two objectives: maximum completion time and total number of sub-batches.

[0257] Build a configurable MOEA framework and collect multiple test cases of hybrid flow shop scheduling;

[0258] The optimal algorithm configuration is obtained by iterating through the following process:

[0259] Step 01: Based on test cases, select excellent algorithm configurations; determine the excellence of the algorithm according to the preset threshold of the optimization index;

[0260] Step 02: Construct the F-Race learning model based on a superior algorithm configuration;

[0261] Step 03: Build a CART model based on the tested algorithm configuration;

[0262] Step 04: Build a completely new algorithm configuration using the F-Race learning model;

[0263] Step 05: Use the CART model to predict the performance of the new algorithm configuration;

[0264] Step 06: Determine if the iteration termination condition is met. If it is met, output the optimal algorithm configuration. If it is not met, proceed to the next iteration based on the predicted high-performance algorithm configuration and execute step 01.

[0265] The F-Race learning model is used to generate new algorithm configurations, while the CART model is used to predict the performance of these new configurations. Both models work together to optimize the algorithm configurations.

[0266] The configurable MOEA framework selects a population evolution method based on the currently configured fitness evaluation algorithm, and adopts cooperative search, local search and population restart strategies to output the Pareto optimal solution for the current instance. The fitness evaluation algorithms include: the decomposition-based method DBM, the non-dominated sorting-based method NSBM and the index-based method IBM.

[0267] The process of building a CART model based on real-time data instances generated during optimization includes: using the evaluated algorithm configuration as training data, using parameter values ​​as features, using evaluation results as labels, and determining the label type based on the preset threshold of the optimization index. The label types include "excellent" and "poor".

[0268] The configurable MOEA framework algorithm includes the following steps:

[0269] Initialize the population P according to the configurable initialization method;

[0270] The iterative process includes:

[0271] Depending on the configurable fitness value evaluation method, mating individuals are selected in different neighborhoods to generate offspring populations (OPs).

[0272] The target values ​​of individuals in population P and offspring population OP are obtained through cooperative search using configurable decoding methods, and corresponding population evaluations are performed for different configurable fitness value evaluation methods.

[0273] Depending on the different fitness value evaluation methods used for the configuration, appropriate environment selection should be made;

[0274] A variable neighborhood descent strategy is used for local search. If the number of consecutive failed updates using the neighborhood structure reaches the preset maximum number of consecutive failed updates, the neighborhood structure will switch to the next neighborhood.

[0275] If the number of consecutive failed updates using cooperative search and neighborhood structure is less than the maximum number of consecutive failed updates, execute the population restart method for each individual in population P.

[0276] If the iteration termination condition is met, the iteration ends and the Pareto nondominated solution found during the evolution process is output; otherwise, the iteration process returns to normal.

[0277] Specifically, for the new configurations generated by the F-Race learning model, the CART model is used to predict the configuration performance, which includes the following steps:

[0278] Step S001: Construct the CART model based on the evaluated algorithm configuration;

[0279] Step S002: Sample the F-Race learning model to generate a series of new algorithm configurations;

[0280] Step S003: Use the CART model to predict the new algorithm configurations. Algorithm configurations predicted as "good" are retained, and algorithm configurations predicted as "poor" are discarded.

[0281] Step S004: Mine the algorithm configuration information predicted as "good" and produce a new F-Race learning model.

[0282] In the process of initializing the population, if the fitness value evaluation method is the decomposition-based method DBM, a set of uniformly distributed weight vectors are generated, and each individual is assigned a unique weight vector.

[0283] If the number of consecutive failed updates using cooperative search and neighborhood structure is less than the maximum number of consecutive failed updates, a population restart method is executed for each individual in population P. Specifically, a configurable population restart method is used to generate a replacement for each individual in population P. Otherwise, the individuals in population P remain unchanged.

[0284] Specifically, depending on the configured fitness value evaluation method, corresponding environment selection is performed. This involves updating the population P using the offspring population OP through different corresponding environment selection methods based on the configured fitness value evaluation method. This includes:

[0285] If the fitness value evaluation method is the decomposition-based method DBM, then the decomposition-based method DBM is used to update the population P using the offspring population OP. Each individual in the offspring population OP is compared with individuals in the population P that have the same weight vector and their neighborhood. If an individual in the offspring population OP has a better fitness value, then the individual in the population P is updated to be an individual in the offspring population OP.

[0286] If the fitness value evaluation method is based on the non-dominated sorting method NSBM, then the offspring population OP is merged into the population P, and the individuals in the population are sorted according to the fast non-dominated sorting method, retaining the first preset population size ps individuals in the population P.

[0287] If the fitness value evaluation method is the index-based method IBM, then the index-based method IBM is used, the offspring population OP is merged into the population P, and the individuals with the lowest fitness values ​​are removed one by one until the population P has a preset population size of ps individuals.

[0288] Example 7

[0289] According to a specific embodiment of the present invention, the multi-objective hybrid flow workshop scheduling method of the present invention will be described in detail with reference to the accompanying drawings.

[0290] This invention provides a multi-objective hybrid flow shop scheduling method, comprising the following steps:

[0291] Based on the actual production characteristics of the hybrid flow workshop, a multi-objective hybrid flow workshop scheduling model is established, including two objectives: maximum completion time and total number of sub-batches.

[0292] Build a configurable MOEA framework and collect multiple test cases of hybrid flow shop scheduling;

[0293] The optimal algorithm configuration is obtained by iterating through the following process:

[0294] Step 01: Based on test cases, select excellent algorithm configurations; determine the excellence of the algorithm according to the preset threshold of the optimization index;

[0295] Step 02: Construct the F-Race learning model based on a superior algorithm configuration;

[0296] Step 03: Build a CART model based on the tested algorithm configuration;

[0297] Step 04: Build a completely new algorithm configuration using the F-Race learning model;

[0298] Step 05: Use the CART model to predict the performance of the new algorithm configuration;

[0299] Step 06: Determine if the iteration termination condition is met. If it is met, output the optimal algorithm configuration. If it is not met, proceed to the next iteration based on the predicted high-performance algorithm configuration and execute step 01.

[0300] The F-Race learning model is used to generate new algorithm configurations, while the CART model is used to predict the performance of these new configurations. Both models work together to optimize the algorithm configurations.

[0301] The configurable MOEA framework selects a population evolution method based on the currently configured fitness evaluation algorithm, and adopts cooperative search, local search and population restart strategies to output the Pareto optimal solution for the current instance. The fitness evaluation algorithms include: the decomposition-based method DBM, the non-dominated sorting-based method NSBM and the index-based method IBM.

[0302] The process of building a CART model based on real-time data instances generated during optimization includes: using the evaluated algorithm configuration as training data, using parameter values ​​as features, using evaluation results as labels, and determining the label type based on the preset threshold of the optimization index. The label types include "excellent" and "poor".

[0303] The configurable MOEA framework algorithm includes the following steps:

[0304] Initialize the population P according to the configurable initialization method;

[0305] The iterative process includes:

[0306] Depending on the configurable fitness value evaluation method, mating individuals are selected in different neighborhoods to generate offspring populations (OPs).

[0307] The target values ​​of individuals in population P and offspring population OP are obtained through cooperative search using configurable decoding methods, and corresponding population evaluations are performed for different configurable fitness value evaluation methods.

[0308] Depending on the different fitness value evaluation methods used for the configuration, appropriate environment selection should be made;

[0309] A variable neighborhood descent strategy is used for local search. If the number of consecutive failed updates using the neighborhood structure reaches the preset maximum number of consecutive failed updates, the neighborhood structure will switch to the next neighborhood.

[0310] If the number of consecutive failed updates using cooperative search and neighborhood structure is less than the maximum number of consecutive failed updates, execute the population restart method for each individual in population P.

[0311] If the iteration termination condition is met, the iteration ends and the Pareto nondominated solution found during the evolution process is output; otherwise, the iteration process returns to normal.

[0312] Specifically, for the new configurations generated by the F-Race learning model, the CART model is used to predict the configuration performance, which includes the following steps:

[0313] Step S001: Construct the CART model based on the evaluated algorithm configuration;

[0314] Step S002: Sample the F-Race learning model to generate a series of new algorithm configurations;

[0315] Step S003: Use the CART model to predict the new algorithm configurations. Algorithm configurations predicted as "good" are retained, and algorithm configurations predicted as "poor" are discarded.

[0316] Step S004: Mine the algorithm configuration information predicted as "good" and produce a new F-Race learning model.

[0317] In the process of initializing the population, if the fitness value evaluation method is the decomposition-based method DBM, a set of uniformly distributed weight vectors are generated, and each individual is assigned a unique weight vector.

[0318] If the number of consecutive failed updates using cooperative search and neighborhood structure is less than the maximum number of consecutive failed updates, a population restart method is executed for each individual in population P. Specifically, a configurable population restart method is used to generate a replacement for each individual in population P. Otherwise, the individuals in population P remain unchanged.

[0319] Depending on the configured fitness value evaluation method, mating individuals are selected within different neighborhoods to generate offspring populations, specifically including:

[0320] Depending on the configured fitness value evaluation method, a configurable cooperative method is used to generate a progeny population of a preset population size ps by applying it to the mating individuals selected from the population. Specifically:

[0321] If the fitness value evaluation method is the decomposition-based method DBM, then each individual selects a mating target in a neighborhood of a preset neighborhood size ns, and executes a configurable cooperative method to produce offspring until a subpopulation of a preset population size ps is produced.

[0322] If the fitness value evaluation method is the non-dominated ranking method NSBM and the index-based method IBM, then two mating individuals are selected in the tournament rules with a preset tournament size ts, and the cooperative method is used to produce offspring. This process is repeated with a preset population size ps times to produce a progeny population of the preset population size ps.

[0323] Specifically, the target values ​​of individuals in the population P and the offspring population OP are obtained through cooperative search using a decoding method, and corresponding population evaluations are performed based on different configured fitness value evaluation methods, including:

[0324] The target values ​​of individuals in the population P and the offspring population OP are obtained using a configurable decoding method;

[0325] Depending on the configured fitness evaluation method, assign a fitness value using the corresponding fitness evaluation method:

[0326] If the fitness value evaluation method is the decomposition-based method DBM, then the decomposition-based method DBM shall be used for fitness value evaluation.

[0327] If the fitness value evaluation method is the non-dominated ranking method NSBM, then the fitness value evaluation method NSBM based on non-dominated ranking is used.

[0328] If the fitness value evaluation method is the indicator-based IBM method, then the indicator-based IBM method will be used for fitness value evaluation.

[0329] Specifically, depending on the configured fitness value evaluation method, corresponding environment selection is performed. This involves updating the population P using the offspring population OP through different corresponding environment selection methods based on the configured fitness value evaluation method. This includes:

[0330] If the fitness value evaluation method is the decomposition-based method DBM, then the decomposition-based method DBM is used to update the population P using the offspring population OP. Each individual in the offspring population OP is compared with individuals in the population P that have the same weight vector and their neighborhood. If an individual in the offspring population OP has a better fitness value, then the individual in the population P is updated to be an individual in the offspring population OP.

[0331] If the fitness value evaluation method is based on the non-dominated sorting method NSBM, then the offspring population OP is merged into the population P, and the individuals in the population are sorted according to the fast non-dominated sorting method, retaining the first preset population size ps individuals in the population P.

[0332] If the fitness value evaluation method is the index-based method IBM, then the index-based method IBM is used, the offspring population OP is merged into the population P, and the individuals with the lowest fitness values ​​are removed one by one until the population P has a preset population size of ps individuals.

[0333] Example 8

[0334] According to a specific embodiment of the present invention, the multi-objective hybrid flow workshop scheduling system of the present invention will be described in detail with reference to the accompanying drawings.

[0335] This invention provides a multi-objective hybrid flow workshop scheduling system, using any of the above-mentioned multi-objective hybrid flow workshop scheduling methods, including multiple flow workshop devices to be scheduled and a control scheduling module;

[0336] The control and scheduling module establishes a multi-objective hybrid flow workshop scheduling model based on the actual production characteristics of the hybrid flow workshop, including two objectives: maximum completion time and total number of sub-batches. It constructs a configurable MOEA framework and collects multiple test cases for hybrid flow workshop scheduling. Based on these test cases, it selects superior algorithms and builds an F-Race learning model based on the configurations of these superior algorithms. It then builds a CART model based on the tested algorithm configurations. Finally, it uses the F-Race learning model to construct entirely new algorithm configurations and uses the CART model to predict the performance of these new configurations. Based on the predicted superior algorithms, it iterates repeatedly to obtain the optimal algorithm configuration. Finally, it selects the equipment to be scheduled in the flow workshop according to the optimal algorithm configuration and issues scheduling commands.

[0337] After receiving the scheduling instructions from the control and scheduling module, the equipment in the assembly line workshop that is to be scheduled will perform the corresponding operations.

[0338] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention are included within the scope of protection of the present invention.

Claims

1. A multi-objective hybrid flow shop scheduling method, characterized in that, Includes the following steps: Based on the actual production characteristics of the hybrid flow workshop, a multi-objective hybrid flow workshop scheduling model is established, including two objectives: maximum completion time and total number of sub-batches. Build a configurable MOEA framework and collect multiple test cases of hybrid flow shop scheduling; The optimal algorithm configuration is obtained by iterating through the following process: Step 01: Based on test cases, select excellent algorithm configurations; determine the excellence of the algorithm according to the preset threshold of the optimization index; Step 02: Construct the F-Race learning model based on a superior algorithm configuration; Step 03: Build a CART model based on the tested algorithm configuration; Step 04: Build a completely new algorithm configuration using the F-Race learning model; Step 05: Use the CART model to predict the performance of the new algorithm configuration; Step 06: Determine if the iteration termination condition is met. If it is met, output the optimal algorithm configuration. If it is not met, proceed to the next iteration based on the predicted high-performance algorithm configuration and execute step 01. The F-Race learning model is used to generate new algorithm configurations, and the CART model is used to predict the performance of the new algorithm configurations. The configurable MOEA framework selects a population evolution method based on the currently configured fitness evaluation algorithm, and employs cooperative search, local search, and population restart strategies to output the Pareto optimal solution for the current instance. The fitness evaluation algorithm includes: a decomposition-based method. DBM Methods based on non-dominated sorting NSBM and indicator-based methods IBM; The configurable MOEA framework algorithm includes the following steps: Initialize the population P according to the configurable initialization method; The iterative process includes: Depending on the configurable fitness value evaluation method, mating individuals are selected in different neighborhoods to generate offspring populations (OPs). The target values ​​of individuals in population P and offspring population OP are obtained through cooperative search using configurable decoding methods, and corresponding population evaluations are performed for different configurable fitness value evaluation methods. Depending on the different fitness value evaluation methods used for the configuration, appropriate environment selection should be made; A variable neighborhood descent strategy is used for local search. If the number of consecutive failed updates using the neighborhood structure reaches the preset maximum number of consecutive failed updates, the neighborhood structure will switch to the next neighborhood. If the number of consecutive failed updates using cooperative search and neighborhood structure is less than the maximum number of consecutive failed updates, execute the population restart method for each individual in population P. If the iteration termination condition is met, the iteration ends and the Pareto nondominated solution found during the evolution process is output; otherwise, the iteration process returns to normal.

2. The multi-objective hybrid flow workshop scheduling method according to claim 1, characterized in that, The CART model is constructed based on real-time data instances generated during the optimization process, including: using the evaluated algorithm configuration as training data, using parameter values ​​as features, using evaluation results as labels, and determining the label type based on the preset threshold of the optimization index. The label types include "excellent" and "poor".

3. The multi-objective hybrid flow workshop scheduling method according to claim 2, characterized in that, For the new configurations generated by the F-Race learning model, the CART model is used to predict the configuration performance, which includes the following steps: Step S001: Construct the CART model based on the evaluated algorithm configuration; Step S002: Sample the F-Race learning model to generate a series of new algorithm configurations; Step S003: Use the CART model to predict the new algorithm configurations. Algorithm configurations predicted as "good" are retained, and algorithm configurations predicted as "poor" are discarded. Step S004: Mine the algorithm configuration information predicted as "good" and produce a new F-Race learning model.

4. The multi-objective hybrid flow workshop scheduling method according to claim 3, characterized in that, During population initialization, if the fitness value evaluation method is a decomposition-based method... DBM This generates a set of uniformly distributed weight vectors, and assigns a unique weight vector to each individual.

5. The multi-objective hybrid flow workshop scheduling method according to claim 4, characterized in that, Depending on the fitness value evaluation method configured, mating individuals are selected within different neighborhoods to generate offspring populations, specifically including: Depending on the configured fitness value evaluation method, a configurable cooperative method is used to generate a progeny population of a preset population size ps by applying it to the mating individuals selected from the population. Specifically: If the fitness value evaluation method is a decomposition-based method DBM Each individual selects a mating partner in a neighborhood of size ns and executes a configurable cooperative method to produce offspring until a subpopulation of size ps is produced. If the fitness value evaluation method is based on non-dominated ranking... NSBM and indicator-based methods IBM Then, in the tournament rules of the preset tournament size ts, two mating individuals are selected, and the cooperative method is executed to produce offspring. This process is repeated 10 times to produce a population of offspring with a size of 10 times.

6. The multi-objective hybrid flow workshop scheduling method according to claim 4, characterized in that, The target values ​​of individuals in the population P and the offspring population OP are obtained through cooperative search using a decoding method. Furthermore, population evaluation is performed based on different configured fitness value evaluation methods, specifically including: The target values ​​of individuals in the population P and the offspring population OP are obtained using a configurable decoding method; Depending on the configured fitness evaluation method, assign a fitness value using the corresponding fitness evaluation method: If the fitness value evaluation method is a decomposition-based method DBM Then a decomposition-based method is adopted. DBM Perform fitness value evaluation; If the fitness value evaluation method is based on non-dominated ranking... NSBM Then, a non-dominated sorting method is adopted. NSBM Perform fitness value evaluation; If the fitness value evaluation method is an index-based method IBM Then, an indicator-based approach is adopted. IBM Perform fitness value evaluation.

7. The multi-objective hybrid flow workshop scheduling method according to claim 4, characterized in that, Depending on the configured fitness value evaluation method, corresponding environment selection is performed. Specifically, based on the different configured fitness value evaluation methods, the population P is updated using the offspring population OP through different corresponding environment selection methods. This includes: If the fitness value evaluation method is a decomposition-based method DBM Then a decomposition-based method is adopted. DBM The offspring population OP is used to update the population P. Each individual in the offspring population OP is compared with individuals in population P that have the same weight vector and their neighborhood. If an individual in the offspring population OP has a better fitness value, then the individual in population P is updated to be an individual in the offspring population OP. If the fitness value evaluation method is based on non-dominated ranking... NSBM If the offspring population OP is merged into population P, the individuals in the population are sorted according to the fast non-dominated sorting method, and the first preset population size ps individuals are retained in population P. If the fitness value evaluation method is an index-based method IBM Then, an indicator-based approach is adopted. IBM The offspring population OP is merged into population P, and the individuals with the lowest fitness values ​​are removed sequentially until population P has a preset population size of ps individuals remaining.

8. The multi-objective hybrid flow workshop scheduling method according to claim 3, characterized in that, If the number of consecutive failed updates using cooperative search and neighborhood structure is less than the maximum number of consecutive failed updates, the population restart method is executed for each individual in population P. Specifically, a configurable population restart method is used to generate a replacement for each individual in population P. Otherwise, the individuals in population P remain unchanged.

9. A multi-objective hybrid flow workshop scheduling system, characterized in that, The multi-objective hybrid flow workshop scheduling method according to any one of claims 1-8 includes multiple flow workshop devices to be scheduled and a control and scheduling module; The control and scheduling module establishes a multi-objective hybrid flow workshop scheduling model based on the actual production characteristics of the hybrid flow workshop, including two objectives: maximum completion time and total number of sub-batches; it also constructs a configurable MOEA framework and collects multiple test cases for hybrid flow workshop scheduling. Based on test cases, select superior algorithms, and build an F-Race learning model based on the superior algorithm configurations; build a CART model based on the tested algorithm configurations; use the F-Race learning model to build a brand new algorithm configuration; and use the CART model to predict the performance of the brand new algorithm configuration. The algorithm with the best prediction is iterated repeatedly to obtain the optimal algorithm configuration; the equipment to be scheduled in the production line is selected according to the optimal algorithm configuration, and the scheduling instructions are issued. After receiving the scheduling instructions from the control and scheduling module, the equipment in the assembly line workshop that is to be scheduled will perform the corresponding operations.

Images