A method for improving the efficiency of feature selection in genetic programming scheduling rules for job shop scheduling
By using a surrogate-assisted genetic programming method, the complexity of fitness evaluation is reduced by utilizing a surrogate model, which solves the problem of low feature selection efficiency in dynamic job shop scheduling and improves shop production efficiency.
Patent Information
- Application Number
- CN202210656998.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-06-10
- Publication Date
- 2024-11-19
- Estimated Expiration
- 2042-06-10
AI Technical Summary
Existing technologies, when faced with dynamic job shop scheduling problems, suffer from low efficiency in feature selection algorithms, which are unable to cope with complex dynamic shop environments and unpredictable events, resulting in computational difficulties and low scheduling efficiency.
We employ a surrogate-assisted genetic programming (SGP) method based on a simplified model. Scheduling rules are generated by evolution using the original training set and the surrogate training set, respectively. This surrogate model reduces the complexity of fitness evaluation and improves feature selection efficiency.
It significantly reduces the training time of genetic programming and improves the efficiency of feature selection algorithms, thereby improving workshop production efficiency and enterprise work efficiency.
Smart Images

Figure CN115034615B_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of workshop scheduling problems in manufacturing industry, and more specifically, relates to a method for improving the efficiency of feature selection in agent-assisted genetic programming scheduling rules based on a simplified model, so as to further improve workshop production efficiency. Background Art
[0002] Job Shop Scheduling (JSS) has always been a hot research topic in the field of intelligent manufacturing and artificial intelligence, and has many applications in different industries such as manufacturing and project scheduling. With the rapid development of information technology and the intensification of global market competition, corporate interests and the diversified needs of customers are closely related. Therefore, companies are increasingly concerned about how to formulate workshop production scheduling plans (such as assigning workpieces to machines and determining the processing order of assigned workpieces on each machine) to optimize criteria such as process time, on-time delivery or customer satisfaction, which will benefit the company and thus improve its profits or reputation.
[0003] At present, for solving the job shop scheduling problem, the patent (publication number "CN112598309A", name "Job Shop Scheduling Method Based on Keras") has been published to optimize the job shop scheduling problem by integrating neural network and Q learning methods. The patent (publication number "CN110458478A", name "Job Shop Scheduling Method Based on Discrete Invasive Weed Algorithm") proposes a scheduling method based on an advanced new evolutionary algorithm, and makes it more intelligent to solve the job shop scheduling problem in the manufacturing industry. Although these methods can solve the job shop scheduling problem to a certain extent, when the shop scheduling faces the challenges of the dynamics and uncertainty of the shop environment (such as machine failure, task cancellation), this brings computational difficulties to the traditional meta-heuristic optimization technology (such as the above method). On the contrary, the scheduling rules can cope with complex dynamic shop environment changes and respond to unpredictable events that are about to arrive at the shop. Therefore, scheduling rules are widely used in solving the dynamic job shop scheduling problem (DJSSP).
[0004] When the scheduling rules are generated by evolution through the hyper-heuristic genetic programming (GP) method, the selection of feature terminal sets is crucial. Existing results have shown that the use of effective feature subsets significantly improves the test performance compared to the use of all feature sets. However, the existing feature selection algorithms used to generate job shop scheduling rule models need to be improved and enhanced. The present invention is aimed at the dynamic shop scheduling problem with the scheduling goal of minimizing the average process time. A method for feature selection based on a simplified model of surrogate-assisted genetic programming (SGP) is proposed to intelligently design and generate scheduling rules. The surrogate model is used to reduce the complexity of fitness evaluation to reduce the time of the genetic programming training process and improve the efficiency of the feature selection algorithm. It is of great significance to the problem of workshop production scheduling efficiency. Summary of the invention
[0005] In order to improve the efficiency of feature selection in the job shop scheduling problem using genetic programming, the present invention uses the original training set and the proxy training set to evolve and generate scheduling rules respectively, and then performs feature selection on the shop terminal set. The proxy model can obtain a shop feature terminal set similar to the original model without losing the evaluation accuracy, and can greatly reduce the time of training rules.
[0006] In order to achieve the above object, the present invention adopts the following technical solutions.
[0007] A method for improving the efficiency of feature selection in genetic programming scheduling rules for job shop scheduling, characterized in that based on the definition: there are m machines in the shop and n workpieces to be processed; workpiece i (i = 1, 2, ... n) has j processes J i =(J i1 ,...J ij ), the jth process J of workpiece i ij According to a certain process sequence, the specified machine m ij When processing on, the processing time required is t ij ; The time when workpiece i arrives at the workshop dynamically and randomly in a certain distribution form is a i , the delivery time of workpiece i is d i ;
[0008] Select training instances and use hyper-heuristic genetic programming (GP) to run N times to generate N scheduling rules;
[0009] For each rule and each corresponding feature, the original model and the proxy model are used to calculate the contribution value of each feature to each rule;
[0010] Set to the median of all contributions.
[0011] In the above method, the original model fitness and the surrogate model fitness The average normalized target value is:
[0012]
[0013]
[0014] In the above method, It is the optimal reference target value, but the optimal value is usually unknown. The present invention sets it as the target value of the reference rule "2PT+WINQ+NPT", which is the optimal effective scheduling rule for the target value.
[0015] Secondly, the feature selection algorithm is described as follows: i To the priority function Contribution and
[0016]
[0017]
[0018] The two contributions are: Delete x from i The difference between the fitness values before and after.
[0019] In the above method, the original model and the proxy model are applied to calculate the contribution value of each feature to each rule according to Formula 3 and Formula 4 respectively.
[0020] In the above method, the median of all contribution values is based on the GP's contribution to the best rule.
[0021] In the above method, the GP generates scheduling rules including:
[0022] Step 1: Use the ramped-half-and-half generation method to randomly initialize the population. Given a GP tree depth of (2, 6), the GP tree has a variety of sizes and shapes to maintain population diversity. Randomly initialize the population from the function set {+, -, *, / , min, max} and the terminal set shown in the attached table. Figure 2 shown.
[0023] Step 2: Set the number of iterations, i.e. the maximum number of generations, to 51;
[0024] Step 3: Evaluate the fitness of each initialized rule;
[0025] Step 4: If the generated best rule has better training performance than the current best rule, then update the current best rule. If the stopping condition (i.e., the maximum number of genetic generations in SGP) is met, then SGP will stop;
[0026] Step 5: If the stopping condition is not met, the tournament selection method is used to perform genetic operations such as replication, crossover and mutation and the elite retention strategy to create an intermediate population. The size parameter of the tournament selection method is set to 7. Compared with the original population, a larger intermediate population size increases the diversity of the population and also provides an opportunity to obtain better rules. The fitness of all rules in the intermediate population is evaluated by the proxy model;
[0027] Step 6: Use the proxy model to evaluate the fitness of the generated intermediate population, and then select 1024 individuals from good to bad according to the fitness to enter the next generation, and return to step 3 until the stopping condition is met.
[0028] In the above method, step 5 specifically includes:
[0029] Step 501: replication operation, selecting from all individuals of the 0th generation through the tournament selection method, the individuals with smaller fitness have a greater probability of being inherited to the next generation population, and the replication rate is 5%;
[0030] Step 502: Crossover operation, as shown in the attached Figure 3 As shown, two parent GP trees are randomly selected, and their child nodes are randomly selected to form offspring through crossover. The crossover rate is 85%, and the maximum depth of the crossover subtree is 8.
[0031] Step 503: mutation operation, as shown in the attached Figure 4 As shown in the figure, a child node is randomly selected as a mutation point in the parent generation, and a subtree is randomly generated to replace the subtree with the mutation point as the root node in the parent generation. The mutation rate is 10%, and the maximum depth of the mutation subtree is 8. At the same time, the elite retention strategy can ensure that the best individuals in the population are inherited to the next generation.
[0032] Step: 504: The original population size is set to 1024, and the intermediate population size is set to 256*7=1792.
[0033] The advantages of the present invention are:
[0034] (1) In the experiment, the rules are analyzed by calculating the length of the best rule and the mean value of leaves (number of terminals). Fig.12 The length and mean of leaves of the best rule obtained from 30 independent runs of the original model and the surrogate model in a dynamic job shop scenario to minimize the average process time are given. As expected, the surrogate model obtains simpler rules than the original model, with lower values for the length of the best rule and the mean of leaves.
[0035] (2) Attachment Figure 5 It also shows that the average length of rules per generation obtained from 30 independent runs on the original and proxy models is smaller for the proxy model than for the original model.
[0036] (3) Since one of the main goals of this invention is to improve the efficiency of feature selection, the calculation time is analyzed. Figure 6 The computation time of 30 independent runs of GP on the original model training set and the proxy model training set is given, which greatly improves the efficiency of feature selection, thereby improving workshop production efficiency and further improving the work efficiency of the enterprise. BRIEF DESCRIPTION OF THE DRAWINGS
[0037] Figure 1 It is the flow chart of the SGP algorithm for surrogate model evaluation.
[0038] Figure 2 is an example of a randomly initialized population.
[0039] Figure 3 It is the crossover operation of the GP tree.
[0040] Figure 4 It is the mutation operation of GP tree.
[0041] Figure 5 is the average of the rule lengths in 30 independent runs across two training sets in a specific embodiment.
[0042] Figure 6 is the running time of 30 independent training rules in two training sets in a specific embodiment.
[0043] Figure 7 It is the meaning of the symbolic expression in the feature selection in the specific embodiment.
[0044] Figure 8 It is the symbolic representation of the workshop feature terminal set in a specific embodiment.
[0045] Fig. 9 It is the feature correlation of 30 independent runs of ET-GP on the original training set in a specific embodiment.
[0046] Fig.10 is the feature correlation of 30 independent runs of ET-SGP on the proxy training set in a specific embodiment.
[0047] Fig.11 It is the key terminal set for the workshop scenario in the two training sets in the specific embodiment.
[0048] Fig.12is the mean of the length and leaves of the best rule obtained in 30 independent runs of two training sets in a specific embodiment.
[0049] Fig.13 It is the test performance of different workshop scenarios in specific embodiments. DETAILED DESCRIPTION
[0050] In order to make the purpose and technical solution of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention, and are not used to limit the present invention.
[0051] 1. First, the method principle of the present invention is introduced.
[0052] First, a mathematical model is established. In DJSSP, there are m machines in the workshop and n workpieces to be processed; workpiece i (i = 1, 2, ... n) has j processes J i =(J i1 ,...J ij ), the jth process J of workpiece i ij According to a certain process sequence, the specified machine m ij When processing on, the processing time required is t ij ; The time when workpiece i arrives at the workshop dynamically and randomly in a certain distribution form is a i , the delivery time of workpiece i is d i .
[0053] At the same time, define the original model fitness and the surrogate model fitness The average normalized target value on instance set I:
[0054]
[0055]
[0056] Ideally, It is the optimal reference target value, but the optimal value is usually unknown. The present invention sets it as the target value of the reference rule "2PT+WINQ+NPT", which is the optimal effective scheduling rule for the target value.
[0057] Secondly, the feature selection algorithm is described as follows: i To the priority function Contribution and
[0058]
[0059]
[0060] The two contributions are: Delete x from i The difference between the fitness values before and after. A positive difference means that x has been deleted. i After that, the new scheduling rule will lead to worse performance; on the contrary, negative values indicate that x i right produces a negative contribution, and removing it leads to better performance; if and Then x i No contribution, no impact on performance. According to equations (1) and (2), define feature x i The correlation χ ω (x i ) to optimize the scheduling rules generated under different job shop scenarios ω. If a feature x i has better relevance, then it will contribute more in the workshop scenario. ω (x i ) is defined based on the contribution of GP to the optimal rule, and has the following calculation steps.
[0061] Step 1: Select a set of training instances
[0062] Step 2: Use all workshop features As the terminal set, GP is applied to evolve and generate scheduling rules;
[0063] Step 3: Perform 30 independent GP runs and obtain the 30 best rules
[0064] Step 4: For each rule and each feature The two models calculate x through equations (3) and (4) respectively. i arrive Contribution
[0065] Step 5: Set χ ω (x i ) is set to all The median of the values (k=1,...30).
[0066] The genetic programming for feature selection of the original model is denoted as FS-GP (Feature Selection-Genetic Programming), and the genetic programming for feature selection of the surrogate model is denoted as FS-SGP (Feature Selection-Surrogate Genetic Programming). The framework of FS-GP and FS-SGP includes three stages. In the first stage, GP of all feature terminal sets is run 30 times independently to obtain 30 best rules. In the second stage, the GP of each feature x is calculated based on the 30 best rules. i The correlation χ ω (x i ), and select features with correlation greater than 0 to form a new set of key terminals The feature selection correlation results are shown in Appendix Tables 3, 4, and 5. In the third stage, the new terminal set was used again. GP to obtain new rules.
[0067] Finally, the purpose of using a method to improve the efficiency of feature selection in agent-assisted genetic programming based on a simplified model is to reduce the complexity of the fitness evaluation of the original model and keep its accuracy level within an acceptable range. The design strategy of the agent model reduces the complexity of the original model by reducing the number of machines and the number of preheating jobs. The agent model sets the number of jobs and the number of preheating jobs to 500 and 100, respectively. Since the simplification of the model makes it statistically inaccurate (caused by a small amount of repetitions and deviations), the agent model cannot provide the absolute performance of the original workshop (for example, minimizing the average process time). Therefore, what we should prove is that rule a is better than rule b based on the absolute performance of the original model, then the agent model should also prove that a is better than b, that is, relative performance. The algorithm uses three types of fitness functions at different stages. The fitness of the original model and the surrogate model fitness Fitness is the true fitness (absolute performance) obtained using the original model, is the estimated fitness obtained using the surrogate model, and the best fitness It's a rule Performance in a particular generation. Because simulation is expensive, it is impractical to use a large number of copies to obtain fitness; therefore, only one copy is used per generation to evolve the rules to obtain Based on the best fitness After the intermediate population is generated, the proxy model is used to evaluate the fitness of the newly generated rules. Although these three fitness function evaluation algorithms are complex, they are more effective and can reduce the overall training time of rules, improve the efficiency of feature selection algorithms, and thus improve the production efficiency of the entire workshop. The algorithm flow chart is as follows Figure 1shown.
[0068] The present invention takes job shop scheduling as the background problem, generates a corresponding model according to the number of workpieces and the number of machines, and sets the relevant parameters of the dynamic job shop scheduling problem, as shown in the attached table. The entire algorithm solution steps are as follows:
[0069] Step 1: Use the ramped-half-and-half generation method to randomly initialize the population. Given a GP tree depth of (2, 6), the GP tree has a variety of sizes and shapes to maintain population diversity. Randomly initialize the population from the function set {+, -, *, / , min, max} and the terminal set shown in the attached table. Figure 2 shown.
[0070] Step 2: Set the number of iterations, i.e. the maximum number of generations, to 51;
[0071] Step 3: Evaluate the fitness of each initialized rule;
[0072] Step 4: If the generated best rule has better training performance than the current best rule, then update the current best rule. If the stopping condition (i.e., the maximum number of genetic generations in SGP) is met, then SGP will stop;
[0073] Step 5: If the stopping condition is not met, the tournament selection method is used to perform genetic operations such as replication, crossover and mutation and the elite retention strategy to create an intermediate population. The size parameter of the tournament selection method is set to 7. Compared with the original population, a larger intermediate population size increases the diversity of the population and also provides an opportunity to obtain better rules. The fitness of all rules in the intermediate population is evaluated by the proxy model;
[0074] Step 501: replication operation, selecting from all individuals of the 0th generation through the tournament selection method, the individuals with smaller fitness have a greater probability of being inherited to the next generation population, and the replication rate is 5%;
[0075] Step 502: Crossover operation, as shown in the attached Figure 3 As shown, two parent GP trees are randomly selected, and their child nodes are randomly selected to form offspring through crossover. The crossover rate is 85%, and the maximum depth of the crossover subtree is 8.
[0076] Step 503: mutation operation, as shown in the attached Figure 4 As shown in the figure, a child node is randomly selected as a mutation point in the parent generation, and a subtree is randomly generated to replace the subtree with the mutation point as the root node in the parent generation. The mutation rate is 10%, and the maximum depth of the mutation subtree is 8. At the same time, the elite retention strategy can ensure that the best individuals in the population are inherited to the next generation.
[0077] Step: 504: the original population size is set to 1024, and the intermediate population size is set to 256*7=1792;
[0078] Step 6: Use the proxy model to evaluate the fitness of the generated intermediate population, and then select 1024 individuals from good to bad according to the fitness to enter the next generation, and return to step 3 until the stopping condition is met.
[0079] 2. The following is a specific example of the principle of the above method.
[0080] The present invention sets the parameters of the workshop simulation experiment according to the established model. There are 10 machines in the workshop and the number of workpieces is 4000. Considering the stability of the workshop state and the effectiveness of the simulation experiment, 1000 workpieces will be used for preheating at the beginning of each simulation experiment. The process of the workpiece arriving at the workshop obeys the Poisson distribution, that is, the workpiece arrives at the workshop dynamically according to the exponential distribution interval time. The number of processes for each workpiece is randomly generated from the discrete uniform distribution DU[2,10], the processing time of each process obeys the discrete uniform distribution DU[1,49], and the candidate processing machine for each process is randomly generated from 10 machines by sampling without replacement. At the same time, the machine utilization rate is 0.8, 0.85, 0.9, and 0.95 for testing; missing means that the number of processes for each job is randomly selected between 2 and 10, obeying U(2,10); full means that the number of processes for all jobs is 10; the delivery margin coefficient of the workpiece is set to 4. At the same time, the dynamic job shop scheduling problem should meet the following assumptions:
[0081] Assumption 1: Starting from time zero, all machines are operational.
[0082] Assumption 2: When one process of a workpiece is completed, it is immediately sent to the next machine to process its next process (transmission time is negligible).
[0083] Assumption 3: Each machine can only process one process at a time.
[0084] Assumption 4: Scheduling is non-preemptive. Once a process starts processing on a machine, it cannot be interrupted by other operations.
[0085] Assumption 5: Different operations of a workpiece cannot be processed together at any given time.
[0086] Then, the scheduling rules generated by GP evolution are used to make decisions based on the workshop information. In the workshop decision queue, when a machine is idle, the scheduling rules are applied to multiple workpieces waiting to be processed on the machine to calculate their priorities and select the workpiece with the highest priority for processing. For example, in the shortest processing time (SPT) rule, the workpiece with the shortest processing time has the highest priority. After all workpieces in the workshop are processed, their target values are calculated and the fitness of the rules is evaluated.
[0087] After the first run, the best rule in 51 generations was selected, and 30 independent runs were performed at the same time, and the 30 best rules were selected for feature selection. The feature contribution of the two training sets was calculated according to equations (3) and (4), and the correlation of each feature was calculated respectively. The key terminal set with correlation greater than 0 was selected, and then GP was run again. The process is the same as that of running with all feature terminal sets, which will not be repeated here.
[0088] Finally, the performance of the 51st generation of rules after feature selection was tested on the test set, and the average performance of 30 independent runs was calculated. The Wilcoxon signed rank test with α = 0.05 was performed on the results before and after feature selection. The statistically significantly better rules are marked in bold. Fig.13 .
[0089] The specific embodiments described herein are merely examples of the spirit of the present invention. Those skilled in the art may make various modifications or additions to the specific embodiments described or replace them in similar ways, but they will not deviate from the spirit of the present invention or exceed the scope defined by the appended claims.
Claims
1. A method for improving the efficiency of feature selection in genetic programming scheduling rules for job shop scheduling, It is characterized in that Based on the definition: there are m machines in the workshop, n workpieces to be processed; workpiece i (i = 1, 2, ... n) has j processes J i =(J i1 ,...J ij ), the jth process J of workpiece i ij According to a certain process sequence, the specified machine m ij When processing on, the processing time required is t ij ; The time when workpiece i arrives at the workshop dynamically and randomly in a certain distribution form is a i , the delivery time of workpiece i is d i ; Select training instances and use hyper-heuristic genetic programming (GP) to run N times to generate N scheduling rules; For each rule and each corresponding feature, the original model and the proxy model are used to calculate the contribution value of each feature to each rule; Set to the median of all contributions; Original model fitness and the surrogate model fitness The average normalized target value is: is the best reference target value, which is set as the target value of the reference rule "2PT+WINQ+NPT", which is the best effective scheduling rule for the target value; is the priority function, Secondly, the feature selection algorithm is described as follows: i To the priority function Contribution and The two contributions are: Delete x from i The difference between the fitness values before and after; GP generates scheduling rules including: Step 1: Use the ramped-half-and-half generation method to randomly initialize the population. Given a GP tree depth of (2, 6), the GP tree has a variety of sizes and shapes to maintain population diversity. Randomly initialize the population from the function set {+, -, *, / , min, max} and the terminal set shown in the attached table. Step 2: Set the number of iterations, i.e. the maximum number of generations, to 51; Step 3: Evaluate the fitness of each initialized rule; Step 4: If the generated best rule has better training performance than the current best rule, the current best rule is updated; if the stopping condition, i.e., the maximum number of genetic generations in SGP, is met, SGP will stop; Step 5: If the stopping condition is not met, the tournament selection method is used to perform genetic operations such as replication, crossover and mutation, and the elite retention strategy to create an intermediate population. The size parameter of the tournament selection method is set to 7. Compared with the original population, the larger intermediate population size increases the diversity of the population and also provides the opportunity to obtain better rules. The fitness of all rules in the intermediate population is evaluated by the proxy model. Step 6: Use the proxy model to evaluate the fitness of the generated intermediate population, and then select 1024 individuals from good to bad according to the fitness to enter the next generation, and return to step 3 until the stopping condition is met.
2. According to claim 1, It is characterized in that The original model and the proxy model are used to calculate the contribution value of each feature to each rule according to Formula 3 and Formula 4 respectively.
3. A method for improving the efficiency of feature selection in genetic programming scheduling rules for job shop scheduling according to claim 1, It is characterized in that The median of all contribution values is based on the GP's contribution to the best rule.
4. A method for improving the efficiency of feature selection in genetic programming scheduling rules for job shop scheduling according to claim 1, It is characterized in that Step 5 specifically includes: Step 501: replication operation, selecting from all individuals of the 0th generation through the tournament selection method, the individuals with smaller fitness have a greater probability of being inherited to the next generation population, and the replication rate is 5%; Step 502: crossover operation, randomly select two parent GP trees, randomly select child nodes therein, crossover to form child generations, the crossover rate is 85%, and the maximum depth of the crossover subtree is 8; Step 503: mutation operation, randomly select a child node in the parent generation as a mutation point, randomly generate a subtree to replace the subtree with the mutation point as the root node in the parent generation, the mutation rate is 10%, and the maximum depth of the mutation subtree is 8; at the same time, the elite retention strategy can ensure that the best individuals in the population are inherited to the next generation; Step: 504: The original population size is set to 1024, and the intermediate population size is set to 256*7=1792.
