Parameter adjustment device, parameter adjustment method, and parameter adjustment program

Abstract

In order to reduce the load of work for adjusting a parameter to be a value matching the preference of a worker, a parameter adjustment device (1) comprises: an evaluation value calculation unit (20) that simulates the operation of a machining device and calculates evaluation values of the machining results corresponding to the machining device parameter set to be used in the simulation; an optimal solution retrieval unit (10) that optimizes each of the evaluation values using a multi-objective optimization procedure, and retrieves a parameter set candidate matching the preference of the user; and an optimization preference information setting unit (40) that acquires, from the user, optimization preference information including user preference information pertaining to the parameter set and user preference information pertaining to the machining results, and reflects the optimization preference information in the optimization of the evaluation values.

Description

Parameter adjustment device, parameter adjustment method, and parameter adjustment program

[0001] This disclosure relates to a parameter adjustment device, a parameter adjustment method, and a parameter adjustment program used for adjusting parameters related to machine tools and the like.

[0002] In machine tools such as machining centers that process workpieces, it is necessary to adjust the parameters related to the machine tool to appropriate values ​​depending on the application and the workpiece. For example, the optimal values ​​for each parameter will differ depending on whether the operator prioritizes cycle time (processing time) over processing accuracy or, conversely, processing accuracy over cycle time. Patent Document 1 describes a technology that allows for the quantitative setting of the importance of processing time and processing accuracy, enabling more effective parameter adjustment in accordance with processing conditions.

[0003] Patent No. 5956619

[0004] The technology described in Patent Document 1 had a problem in that the accuracy of parameter adjustment deteriorated when the test program used for parameter adjustment differed from the program used for actual machining. Furthermore, in order to obtain adjustment results that suited the operator's preferences and maintain the accuracy of parameter adjustment, it was necessary to repeatedly perform a series of operations, such as changing the test program or modifying the adjustment range of each parameter and restarting the parameter adjustment work from the beginning, until the results converged to the operator's preferences. Conventionally, when parameters were to converge to the operator's preferences, it was necessary to perform the parameter adjustment work manually, which burdened the operator with additional effort and time. For this reason, there is a need for a technology that can converge the parameter adjustment values ​​to the operator's preferences while suppressing an increase in the operator's workload, compared to conventional technologies such as the technology described in Patent Document 1.

[0005] This disclosure is made in view of the above, and aims to provide a parameter adjustment device that can reduce the workload of adjusting parameters to values ​​that suit the operator's preferences.

[0006] To solve the above-mentioned problems and achieve the objectives, the parameter adjustment device according to this disclosure is characterized by comprising: an evaluation value calculation unit that simulates the operation of a processing machine and calculates evaluation values ​​of processing results corresponding to the parameter set of the processing machine used in the simulation; an optimal solution search unit that optimizes each evaluation value using a multi-objective optimization method and searches for candidate parameter sets that match the user's preferences; and an optimization preference information setting unit that obtains optimization preference information from the user, including user preference information regarding the parameter set and user preference information regarding the processing results, and reflects this in the optimization of the evaluation values.

[0007] According to this disclosure, a parameter adjustment device is obtained that can reduce the workload of adjusting parameters to values ​​that suit the operator's preferences.

[0008] Figures showing an example configuration of the parameter adjustment device according to Embodiment 1. Figures showing an example configuration of the optimal solution search unit included in the parameter adjustment device according to Embodiment 1. Figures showing an example configuration of the environment selection unit included in the optimal solution search unit of the parameter adjustment device according to Embodiment 1. Figures showing an overview of uncontrolled sorting. Figures explaining how to set the reference direction. Figures showing an example of a method for calculating relative scores using envelope analysis. Figures showing an example of a method for calculating relative scores using projection vectors. Flowchart showing an example of operation of the parameter adjustment device according to Embodiment 1. Flowchart showing an example of operation of the optimal solution search unit included in the parameter adjustment device according to Embodiment 1. Flowchart showing an example of operation of the environment selection unit of the optimal solution search unit included in the parameter adjustment device according to Embodiment 1. Figures showing an example of a method for correcting evaluation values ​​in the parameter adjustment device according to Modification 2 of Embodiment 1. Figures showing an example of another method for correcting evaluation values ​​in the parameter adjustment device according to Modification 2 of Embodiment 1. Figures showing an example of the configuration of the optimal solution search unit included in the parameter adjustment device according to Embodiment 2. Figures showing an example of a method for adjusting the tournament size by the tournament size adjustment unit of the optimal solution search unit according to Embodiment 2. Figures showing an example of a method for adjusting the tournament size by the tournament size adjustment unit of the optimal solution search unit according to Modification 5 of Embodiment 2. Figures showing an example of hardware that realizes the parameter adjustment device.

[0009] The parameter adjustment device, parameter adjustment method, and parameter adjustment program according to embodiments of this disclosure will be described in detail below with reference to the drawings.

[0010] Embodiment 1. Figure 1 shows an example of the configuration of the parameter adjustment device 1 according to Embodiment 1. The parameter adjustment device 1 comprises an optimal solution search unit 10, an evaluation value calculation unit 20, an evaluation value storage unit 30, an optimization preference information setting unit 40, a parameter storage unit 50, and a display unit 60. The parameter adjustment device 1 has the function of adjusting the parameters set on a processing machine (not shown) to appropriate values ​​according to the preferences of the user, such as an operator.

[0011] The optimal solution search unit 10 uses optimization algorithms such as grid search, random search, Newton's method, Bayesian optimization, and evolutionary computation to find one or more optimal solutions. In this embodiment, the optimal solution is a value obtained by adjusting the parameters set on the processing machine (hereinafter referred to as "processing machine parameters" or simply "parameters") according to the user's preferences.

[0012] In this embodiment, as an example, we will describe a case where the optimal solution search unit 10 performs multi-objective optimization using a genetic algorithm, which is a type of evolutionary computation. Therefore, we will first explain the outline of the genetic algorithm.

[0013] A genetic algorithm is an optimization method that incorporates the mechanism by which organisms reproduce and pass on superior individuals. In a genetic algorithm, the object of optimization is called an "individual." In this embodiment, the set of parameters of the processing machine and their corresponding evaluation values ​​are collectively referred to as an individual. Genetic algorithms are flexible and have various variations, but generally, they seek a solution by repeatedly performing "environmental selection," which leaves superior individuals for the next generation based on fitness; "crossover selection," which selects individuals to be crossed over based on fitness; "crossover," which generates two new individuals by converting the parameter values ​​of two individuals into binary and swapping the values; and "mutation," which probabilistically changes a part of an individual. Fitness in multi-objective optimization is an index that evaluates how well an individual is suited to multiple objective functions. The number of times operations such as environment selection and crossover are repeated is called the "number of generations," and the number of individuals in each generation is called the "population." In this embodiment, the number of times the optimal solution search unit 10 processes is defined as the number of generations. In other words, the first time the optimization process is performed in the optimal solution search unit 10 is called a generation. Also, let N be the population (where N is an integer greater than or equal to 2).

[0014] Next, we will explain the overview of multi-objective optimization. Generally, in multi-objective optimization, since the multiple objective functions being dealt with are in a trade-off relationship, there are multiple optimal solutions. Such solutions are called Pareto solutions, and the set of Pareto solutions is called the Pareto front. Multi-objective optimization can be broadly divided into two steps. In the first step, multiple solutions that approximate the Pareto front are presented to the user, and in the next step, the user selects a solution that suits their preferences from the presented multiple solutions. There are two perspectives for evaluating the presented multiple solutions (hereinafter referred to as search solutions): convergence and diversity. Convergence means how close the multiple search solutions are to the Pareto front, and diversity means how uniformly and broadly the multiple search solutions are distributed with respect to the Pareto front.

[0015] The parameter storage unit 50 stores the parameters output from the optimal solution search unit 10.

[0016] The evaluation value calculation unit 20 obtains new parameters generated in each generation of the multi-objective optimization using the genetic algorithm performed by the optimal solution search unit 10 from the parameter storage unit 50, simulates the operation of the processing machine when using the obtained parameters, and calculates one or more evaluation values ​​to evaluate the processing results by the processing machine.

[0017] The evaluation value storage unit 30 stores the evaluation value calculated by the evaluation value calculation unit 20.

[0018] The Optimization Preference Information Setting Unit 40 acquires preference information from the user regarding the setting range and constraints of each parameter of the processing machine, the target value and constraints of the evaluation value, and the genetic algorithm, and sets this information in the Optimal Solution Search Unit 10. The setting range and constraints of each parameter of the processing machine are examples of the user's preference information regarding the parameters of the processing machine. The target value and constraints of the evaluation value are examples of the user's preference information regarding the processing result. In this embodiment, the user can set parameters related to the reference direction, parameters related to crossover selection, parameters related to crossover, parameters related to mutation, and search termination conditions as preference information regarding the genetic algorithm. The search termination conditions may be the number of generations, the search time which indicates the length of time the search is performed, or a flag which indicates whether the evaluation value has reached the target value. Details of the information that can be set in the Optimization Preference Information Setting Unit 40 will be explained together when describing the components that utilize that information. In the following description, the above information acquired from the user by the Optimization Preference Information Setting Unit 40 will be collectively referred to as optimization preference information.

[0019] The display unit 60 acquires optimization preference information, parameters, evaluation values, and generation numbers from the optimal solution search unit 10 from the second generation onward, and displays them in association with each other. For example, if there are two types of evaluation values, the first being evaluation value #1 and the second evaluation value #2, the display may be divided into categories such as a priority category for evaluation value #1, a priority category for evaluation value #2, and a balanced category where evaluation value #1 and evaluation value #2 are balanced. Alternatively, all past and current search solutions may be displayed. Furthermore, if a target value for the evaluation value is set and there are multiple parameters that satisfy the target value, only solutions corresponding to the parameters that satisfy the target value may be displayed. If there are no parameters that satisfy the target value, solutions corresponding to the parameters may be displayed in order of increasing difference from the target value. If the user has met the search termination conditions and wishes to continue searching from the multiple parameters presented on the display unit 60, the parameter adjustment device 1 has a function that allows the user to select any parameter and further search around the selected parameter.

[0020] Figure 2 shows an example of the configuration of the optimal solution search unit 10 included in the parameter adjustment device 1 according to Embodiment 1. As shown in Figure 2, the optimal solution search unit 10 includes an optimization preference information acquisition unit 11, a genetic algorithm information processing unit 12, an initial parameter generation unit 13, an environment selection unit 14, a crossover selection unit 15, a crossover processing unit 16, and a mutation processing unit 17.

[0021] The optimization preference information acquisition unit 11 acquires optimization preference information from the optimization preference information setting unit 40 and outputs the acquired optimization preference information to the genetic algorithm information processing unit 12.

[0022] The Genetic Algorithm Information Processing Unit 12 acquires optimization preference information output from the Optimization Preference Information Setting Unit 40 via the Optimization Preference Information Acquisition Unit 11, acquires processing machine parameters from the Parameter Storage Unit 50, and acquires evaluation values ​​from the Evaluation Value Storage Unit 30. It then outputs information about the processing related to the Genetic Algorithm to the Environment Selection Unit 14, the Crossover Selection Unit 15, the Crossover Processing Unit 16, and the Mutation Processing Unit 17. The processing related to the Genetic Algorithm includes setting the Genetic Algorithm and managing the number of generations. The Genetic Algorithm Setting includes setting the initial parameter sampling method, crossover, mutation, tournament selection, and reference direction. Tournament selection is the method used in the Crossover Selection Unit 15. The reference direction is the vector that determines the direction of the search. The Genetic Algorithm Information Processing Unit 12 manages the number of generations, outputting sampling information to the initial parameter generation unit 13 for the first generation, and outputting relevant information to the environment selection unit 14, the crossover selection unit 15, the crossover processing unit 16, and the mutation processing unit 17 for the second generation and beyond.

[0023] The initial parameter generation unit 13 obtains information regarding the sampling of initial parameters from the genetic algorithm information processing unit 12 in the first generation, and generates initial parameters for the population size N based on that information. Examples of initial parameter sampling methods include random sampling, Latin hypersquare sampling, and manual sampling based on domain knowledge. Random sampling is a method of generating parameters randomly within a defined range. Latin hypersquare sampling is a method of randomly sampling parameters from the entire space within a defined range, minimizing overlap as much as possible. The initial parameter generation unit 13 outputs the generated initial parameters to the parameter storage unit 50.

[0024] The environment selection unit 14 selects N excellent solutions from N+M solutions (where N and M are integers greater than or equal to 2) based on their goodness of fit. Details of how the environment selection unit 14 selects N solutions will be described later.

[0025] The crossover selection unit 15 obtains the tournament size as parameter information related to crossover selection from the genetic algorithm information processing unit 12, and also obtains N individuals from the environment selection unit 14, which are the N solutions selected by the environment selection unit 14. The individuals obtained by the crossover selection unit 15 from the environment selection unit 14 are sets of parameters and evaluation values. After obtaining the parameter information and individuals, the crossover selection unit 15 selects L individuals (L is an integer of 2 or more) from the obtained N individuals to be crossed over by the crossover processing unit 16 through tournament selection. The L individuals that are crossed over by the crossover processing unit 16 may be selected from the N individuals multiple times. Tournament selection is carried out as follows: The crossover selection unit 15 randomly selects K individuals from the N individuals selected by the environment selection unit 14, and selects one individual from the K individuals based on some criterion. The crossover selection unit 15 selects one or more individuals to be crossed over by that individual using a similar operation, and creates a crossover pair. Note that the pairs to be crossed are not identical individuals, but pairs of different individuals. Different pairs may contain identical individuals. The crossover selection unit 15 repeats this tournament L times to select L individuals from N individuals. The number K of individuals randomly selected from N individuals is called the tournament size. The tournament size K is a parameter in tournament selection. In Embodiment 1, when selecting one individual from K individuals, the selection may be random, a known method such as NSGA2 (Non-dominated Sorting Genetic Algorithm II) may be used, the selection may be based solely on the relative score described later, or the selection may be based on the relative score only when selection is not possible by non-dominated sorting. The L selected individuals are output to the crossover processing unit 16.

[0026] The crossover processing unit 16 uses the L individuals selected by the crossover selection unit 15 based on the crossover parameters output from the genetic algorithm information processing unit 12 to generate M new individuals (parameters only) from the crossover pairs created by the crossover selection unit 15. While individuals are usually sets of parameters and evaluation values, the crossover processing unit 16 only generates new parameters, and the evaluation values ​​corresponding to the new parameters are calculated by the evaluation value calculation unit 20. Any method can be used for crossover; known methods such as SBX (Simulated Binary Crossover) may be used.

[0027] The mutation processing unit 17 performs a process to change the parameters of the new processing machine generated by the crossover processing unit 16 with a certain probability, based on the mutation parameters output from the genetic algorithm information processing unit 12. Any method can be used for mutation. For example, known methods such as Polynomial Mutation may be used. The M individuals (parameters only) generated by the mutation processing unit 17 are output to the parameter storage unit 50.

[0028] Next, the details of the environment selection unit 14 will be described. Figure 3 is a diagram showing an example of the configuration of the environment selection unit 14 included in the optimal solution search unit 10 of the parameter adjustment device 1 according to Embodiment 1.

[0029] The environment selection unit 14 includes an environment selection setting information acquisition unit 141, a non-controlling sort processing unit 142, a reference direction setting unit 143, an individual assignment unit 144, a relative score calculation unit 145, a number of affiliated individuals count unit 146, a parent individual selection unit 147, and a parent individual information storage unit 148.

[0030] The environment selection setting information acquisition unit 141 acquires environment selection setting information such as parameters, evaluation values, generation number, and reference direction from the genetic algorithm information processing unit 12 from the second generation onward, and outputs the acquired parameters and their corresponding evaluation values, along with the reference direction information, to the non-controlling sort processing unit 142. The parameters acquired by the environment selection setting information acquisition unit 141 are the parameters of the processing machine. In the genetic algorithm, the N individuals selected by the environment selection unit 14 and the N individuals generated by the initial parameter generation unit 13 are called parent individuals, and the M individuals generated by the crossover processing unit 16 and the mutation processing unit 17 are called offspring individuals. The number of offspring individuals M can be any number other than the population number N, but generally the number of offspring individuals M is often set to the population number N. In Figure 3, the output from the environment selection setting information acquisition unit 141 to the non-controlling sort processing unit 142 is offspring individuals, which is the case from the third generation onward. The individuals acquired by the environment selection unit 14 in the third generation and subsequent generations (L generation) are considered offspring individuals because they were generated by the crossover processing unit 16 and the mutation processing unit 17 in the L-1 generation. The individuals acquired by the environment selection unit 14 in the second generation are the individuals generated by the initial parameter generation unit 13 in the first generation, i.e., the parent individuals.

[0031] The non-dominated sorting unit 142 retrieves N parent individuals from the previous generation from the parent individual information storage unit 148, and also retrieves M child individuals output from the environment selection setting information acquisition unit 141. The non-dominated sorting unit 142 performs non-dominated sorting on the N+M parent and child individuals (hereinafter referred to as parent-child individuals) obtained from the previous generation, and ranks the individuals. Non-dominated sorting uses the concept of dominance. In this concept, when looking at a certain individual (for convenience of explanation, let's call it the first individual), if there is another individual (let's call it the second individual) whose evaluation value is small on every evaluation value axis, then the first individual is said to be dominated by the second individual. Figure 4 is a diagram showing an overview of non-dominated sorting. In non-dominated sorting, first, as shown on the left side of Figure 4, an individual that is not dominated by other individuals (an individual in which no other individuals exist in the lower left region when looking at a certain individual) is found, and that individual is assigned a rank of 1. Next, as shown in the center of Figure 4, individuals that are not dominated by any other individuals other than those assigned rank 1 are found, and rank 2 is assigned to those individuals. This operation is repeated until all individuals have been assigned a rank. In multi-objective optimization, individuals with rank 1 are called Pareto solutions, and the set of Pareto solutions is called a Pareto front. Once the non-dominant sorting unit 142 has finished assigning ranks to all individuals, it outputs parent-child individuals, rank information, and reference direction information to the reference direction setting unit 143.

[0032] The reference direction setting unit 143 acquires N+M parent-child individuals, rank information, and reference direction information output from the non-controlling sorting processing unit 142, and sets the reference direction based on the parent-child individuals. The reference direction setting unit 143 outputs the parent-child individuals, rank information acquired from the non-controlling sorting processing unit 142, and the set reference direction information to the individual assignment unit 144.

[0033] The reference direction set by the reference direction setting unit 143 will now be explained. In multi-objective optimization, it is necessary to optimize multiple evaluation values ​​simultaneously. These evaluation values ​​often compete with each other, and improving one evaluation value may worsen others. To solve this problem, a reference direction is used. The reference direction is a vector that indicates the importance of each evaluation value to be optimized. This is a weighted representation of how much each evaluation value should be considered, indicating the direction in which the optimization algorithm should proceed, and is used to efficiently advance multi-objective optimization while maintaining the diversity of solutions. For example, if a certain evaluation value is more important than others, the component of the reference direction corresponding to that evaluation value will be larger. Generally, the reference direction is set by finding the minimum and maximum values ​​of each axis from the individuals obtained in each generation and uniformly dividing the space formed by these minimum and maximum values. The method of setting the reference direction is not limited to this. The reference direction may also be set in a region where the user wants a solution.

[0034] Let's explain a specific example of a reference direction. Figure 5 is a diagram illustrating how to set reference directions. Figure 5 shows an example where four reference directions are uniformly set based on the evaluation values ​​of the solution obtained for a problem with two objective functions f1 and f2. In Figure 5, r1, r2, r3, and r4 each point to a reference direction, and the dotted arrow between two adjacent reference directions indicates the boundary of the region for each reference direction. It also indicates that individuals existing within the region of each reference direction belong to that reference direction. In the example shown in Figure 5, the slope of each reference direction indicates the importance of objective function f1 to objective function f2. For example, r1 is a vector that considers only f2, and the two individuals between reference direction r1 and the dotted arrow to its right are individuals belonging to r1 (individuals belonging to r1). The reference direction r2 to its right has a slope of 2, so the ratio of importance is f1:f2 = 1:2, and it is a vector that indicates that objective function f2 is considered twice as important as objective function f1. Furthermore, the four individuals located between the two dotted arrows on either side of the reference direction r2 belong to r2. As can be seen from Figure 5, the reference direction also serves to divide the space in which evaluation values ​​exist.

[0035] The individual assignment unit 144 obtains parent-child individuals, rank information, and reference direction information from the reference direction setting unit 143, and calculates which reference direction each parent-child individual is assigned to. The method for assigning individuals to reference directions involves calculating the vertical distance between each individual and each reference direction for all possible combinations, and assigning each individual to the reference direction with the smallest distance. This operation divides the evaluation value space by reference direction and analyzes where each individual is located within the divided space. The individual assignment unit 144 outputs the parent-child individuals, rank information, and reference direction information obtained from the reference direction setting unit 143, along with assignment information showing the assignment results for each parent-child individual to its respective reference direction, to the relative score calculation unit 145.

[0036] The relative score calculation unit 145 obtains parent-child individuals, rank information, reference direction information, and assignment information from the individual assignment unit 144, and calculates the relative score of each parent-child individual based on the obtained information. The relative score is the goodness of fit in genetic algorithms. The relative score may be calculated using the CCR (Charnes-Cooper-Rhodes) model or the super-CCR model of the envelope analysis method, which are known methods, or it may be calculated from the ratio of projection vectors, as described later. Furthermore, the relative score may be calculated for each population belonging to a region divided by the reference direction, or it may be calculated for all N+M parent-child individuals. The relative score calculation unit 145 outputs the parent-child individuals, rank information, reference direction information, and assignment information obtained from the individual assignment unit 144, along with the calculated relative score of each parent-child individual, to the individual count unit 146.

[0037] This section explains the ultra-CCR model of envelope analysis. In the ultra-CCR model, the efficiency value θ of the o-th individual is defined. o The result is calculated by solving the maximization problem shown in equation (1) below using mathematical programming. If the problem to be solved is a minimization problem, then equation (1) is transformed into a minimization problem.

[0038]

[0039] In equation (1), x i o and y jo represents the evaluation value of each individual. i and j respectively represent the i-th and j-th components of a plurality of evaluation values. u j o and v i o are the weights for each evaluation value. The result obtained by solving the above formula (1) is shown in FIG. 6. FIG. 6 is a diagram showing an example of a method for calculating a relative score by the envelope analysis method. Specifically, it shows an example of a method for calculating a relative score by the super CCR model. On the left side of FIG. 6, it shows that individuals A to E are obtained for a problem having two objective functions f1 and f2. The reference point O is the maximum value of each axis among individuals A to E. When solving the above maximization problem by mathematical programming, the individuals are divided into those with a score less than 1 and those with a score of 1 or more. The larger the score, the better. An individual with a score of 1 or more is a solution where, when relatively evaluated with other individuals, there is no solution better than that individual. In the example of FIG. 6, individuals A, C, and E are individuals with a score of 1 or more. The line connecting the set of individuals with a score of 1 or more (individuals A, C, E) as shown in the center of FIG. 6 is called the efficient frontier in the envelope analysis method.

[0040] When obtaining the score of an individual on the efficient frontier, the efficient frontier excluding that individual is used. For example, when obtaining the score of individual C on the efficient frontier, it is necessary to search for the efficient frontier excluding individual C. In the case of the example shown in FIG. 6, the individuals on the efficient frontier excluding individual C are individuals A, B, and E as shown on the right side of FIG. 6. Also, connect individual C and the reference point O with a line, and let the intersection of the efficient frontier excluding individual C and the line segment OC be C'. In this case, the score of individual C is obtained by the ratio of OC to OC', resulting in formula (2). At this time, from the perspective of the envelope analysis method, individual C is constrained by individuals B and E.

[0041]

[0042] Also, when calculating the score of an individual with a score less than 1, a straight line passing through the individual and the reference point O is drawn, and the intersection point of the straight line and the efficient frontier is used. For example, when calculating the score of individual D, as shown on the right side of Fig. 6, if the intersection point of the straight line OD and the efficient frontier is D', the score of individual D is given by Equation (3). At this time, it is said that individual D is restricted by individuals C and E.

[0043]

[0044] Thus, in calculating the relative score by the super CCR model, the relative score of each individual is obtained by making a relative evaluation between the individual for which the score is to be calculated and the individuals on the efficient frontier of the remaining individuals excluding this individual.

[0045] Next, the calculation of the relative score using the projection vector will be explained. Fig. 7 is a diagram showing an example of the method for calculating the relative score using the projection vector. Similar to the example shown in Fig. 6, Fig. 7 shows that individuals A to E have been obtained for a problem with two objective functions f1 and f2.

[0046] For example, when obtaining the relative score of individual D by calculation using the projection vector, a vector is searched for which the projection vector obtained by projecting the vector OC formed by other individuals and the reference point O with respect to the vector OD shown in Equation (4) is maximized.

[0047]

[0048] In the case of the example shown in Fig. 7, the vector for which the projection vector with respect to the vector OD is maximized is the vector OC shown in Equation (5).

[0049]

[0050] The calculation of the relative score using the projection vector is similar to the calculation of the relative score by the above-described super CCR model, and the relative score of individual D is shown by Equation (6). The relative score of individual D represents the score of the result of the relative evaluation between the vector OD and each vector from the reference point O to other individuals other than individual D.

[0051]

[0052] The affiliated individual counting unit 146 obtains parent and child individuals, rank information, reference direction information, assignment information, and relative score from the relative score calculation unit 145, and counts the number of affiliated individuals for each reference direction. The specific operation will be explained below. First, the affiliated individual counting unit 146 grasps the ranks assigned to N+M parent and child individuals based on the rank information, selects individuals in descending order of rank, and repeats the selection until the total number of selected individuals is N or more. Next, the affiliated individual counting unit 146 excludes the individual with the highest rank from the selected set of individuals, and uses the set of individuals after exclusion to count the number of affiliated individuals for each reference direction. A specific example of the counting process will be explained with reference to Figure 4. For example, if N=5, selecting individuals with ranks 1 and 2 results in a total of 6, which exceeds the value of N, so the affiliated individual counting unit 146 selects individuals with ranks 1 and 2. Furthermore, since the highest rank among them is 2, the affiliated individual counting unit 146 excludes individuals with rank 2, and counts the number of affiliated individuals for each reference direction using only individuals with rank 1. In other words, the affiliated individual counting unit 146 counts how many parent-child individuals of rank 1 belong to the N+M parent-child individuals for each reference direction. The affiliated individual counting unit 146 outputs the parent-child individuals, rank information, reference direction information, assignment information, and relative score obtained from the relative score calculation unit 145, along with count information indicating the count result of the affiliated individual count, to the parent individual selection unit 147.

[0053] The parent individual selection unit 147 obtains parent and child individuals, rank information, reference direction information, assignment information, relative score, and count information from the affiliated individual count unit 146, and selects N individuals from the N or more individuals obtained. The N parent and child individuals selected by the parent individual selection unit 147 become new parent individuals, i.e., the parent individuals of the next generation. How the parent individual selection unit 147 selects parent and child individuals will be explained in the operation description described later. The parent individual selection unit 147 outputs the selected N parent individuals to the parent individual information storage unit 148, and also outputs the selected N parent individuals and the rank information, reference direction information, assignment information, relative score, and count information obtained from the affiliated individual count unit 146 to the cross selection unit 15.

[0054] The parent individual information storage unit 148 stores the N parent individuals output from the parent individual selection unit 147.

[0055] Next, the operation of the parameter adjustment device 1 and the optimal solution search unit 10 will be explained. Specifically, the operation of adjusting the machine parameters by the parameter adjustment device 1 will be explained with reference to Figure 8, and the operation of searching for the optimal solution by the optimal solution search unit 10 will be explained with reference to Figure 9. In addition, the selection operation by the environment selection unit 14 of the optimal solution search unit 10 will be explained with reference to Figure 10.

[0056] (Explanation of operation of parameter adjustment device 1) Figure 8 is a flowchart showing an example of the operation of the parameter adjustment device 1 according to Embodiment 1. The parameter adjustment device 1 adjusts the parameters by executing the processes of steps S1 to S7 shown in Figure 8.

[0057] In the parameter adjustment operation by the parameter adjustment device 1, first, the optimization preference information setting unit 40 sets the optimization preference information (step S1). In step S1, the optimization preference information setting unit 40 receives input from the user as optimization preference information, such as the setting range and constraints for each parameter of the processing machine, the target value and constraints for the evaluation value of the processing result, and preference information related to the genetic algorithm, and sets the input optimization preference information in the optimal solution search unit 10.

[0058] Next, the optimal solution search unit 10 calculates the next parameter to be searched (step S2). The optimal solution search unit 10 calculates the next parameter to be searched using a genetic algorithm.

[0059] Next, the parameter storage unit 50 stores the parameters calculated by the optimal solution search unit 10 (step S3).

[0060] Next, the evaluation value calculation unit 20 calculates evaluation values ​​based on the parameters (step S4). Specifically, the evaluation value calculation unit 20 retrieves the parameters stored in the parameter storage unit 50, simulates the operation of the processing machine using the retrieved parameters, and calculates one or more evaluation values ​​to evaluate the processing results.

[0061] Next, the evaluation value storage unit 30 stores the evaluation value calculated by the evaluation value calculation unit 20 (step S5).

[0062] Next, the display unit 60 displays the optimization preference information, parameters, evaluation value, and generation number output from the optimal solution search unit 10 (step S6).

[0063] Next, the optimal solution search unit 10 checks whether the search termination conditions are met (step S7). If the termination conditions are met (step S7: Yes), the search operation is terminated. If the termination conditions are not met (step S7: No), the process returns to step S2. In step S7, the optimal solution search unit 10 determines whether the search termination conditions, which are included in the optimization preference information set in step S1, are met.

[0064] When the parameter adjustment device 1 determines that the conditions for ending the search have been met and terminates the search operation, the parameters displayed on the display unit 60 at this point become the search results, i.e., the parameters optimized by the optimization process, i.e., the adjusted parameters.

[0065] (Explanation of the operation of the optimal solution search unit 10) Figure 9 is a flowchart showing an example of the operation of the optimal solution search unit 10 provided in the parameter adjustment device 1 according to Embodiment 1. Steps S21 to S33 shown in Figure 9 show the details of the process of step S2 shown in Figure 8, that is, the operation in which the optimal solution search unit 10 calculates the next parameter to be searched. The optimal solution search unit 10 performs the operation according to the flowchart in Figure 9 in each generation of the genetic algorithm.

[0066] In the optimal solution search unit 10, first, the optimization preference information acquisition unit 11 acquires optimization preference information from the optimization preference information setting unit 40 (step S21).

[0067] Next, the genetic algorithm information processing unit 12 sets the genetic algorithm (step S22). Specifically, the genetic algorithm information processing unit 12 sets the genetic algorithm based on the optimization preference information acquired by the optimization preference information acquisition unit 11 in step S21. In setting the genetic algorithm, the genetic algorithm information processing unit 12 sets, for example, parameters related to the reference direction, parameters related to crossover selection, parameters related to crossover, parameters related to mutation, and search termination conditions.

[0068] Next, the genetic algorithm information processing unit 12 checks whether the number of generations of the genetic algorithm is 1, that is, whether it is the first generation (step S23). If the number of generations is 1 (step S23: Yes), it decides to execute the processes in steps S24 to S25. If the number of generations is not 1 (step S23: No), the genetic algorithm information processing unit 12 checks whether the number of generations of the genetic algorithm is 2, that is, whether it is the second generation (step S26). If the number of generations is 2 (step S26: Yes), the genetic algorithm information processing unit 12 decides to execute the processes in steps S27 and S29 to S33. If the number of generations is not 2, that is, if the number of generations is 3 or more (third generation or later) (step S26: No), the genetic algorithm information processing unit 12 decides to execute the processes in steps S28 to S33.

[0069] In the case of the first generation (step S23: Yes), the genetic algorithm information processing unit 12 outputs sampling information to the initial parameter generation unit 13, and the initial parameter generation unit 13 generates initial parameters based on the sampling information (step S24). The initial parameter generation unit 13 also outputs the generated initial parameters to the parameter storage unit 50 (step S25). The parameter storage unit 50 stores the initial parameters output from the initial parameter generation unit 13.

[0070] In the case of the second generation (step S23: No, step S26: Yes), the genetic algorithm information processing unit 12 outputs N individuals, each consisting of a set of N parameters and an evaluation value, to the environment selection unit 14. The environment selection unit 14 acquires the N individuals output from the genetic algorithm information processing unit 12 (step S27). In the case of the second generation, the environment selection unit 14 outputs the acquired N individuals directly to the crossover selection unit 15.

[0071] For the third generation and beyond (step S23: No, step S26: No), the genetic algorithm information processing unit 12 outputs N+M individuals, each consisting of a set of N+M parameters and evaluation values, to the environment selection unit 14. The environment selection unit 14 retrieves the N+M individuals output from the genetic algorithm information processing unit 12 and selects N individuals from the retrieved individuals (step S28). The environment selection unit 14 outputs the selected N individuals to the crossover selection unit 15.

[0072] Next, the crossover selection unit 15 selects L individuals from the N individuals obtained from the environment selection unit 14 (step S29). The crossover selection unit 15 outputs the selected L individuals to the crossover processing unit 16.

[0073] Next, the crossover processing unit 16 performs crossover on the L individuals selected by the crossover selection unit 15 to generate M new individuals (step S30). As mentioned above, the crossover processing unit 16 generates only the parameters as new individuals, not a set of parameters and evaluation values. The crossover processing unit 16 outputs the generated M individuals to the mutation processing unit 17.

[0074] Next, the mutation processing unit 17 performs mutations on the individuals generated by the crossover processing unit 16 (step S31). Specifically, the mutation processing unit 17 performs a process to change the parameters constituting the M new individuals generated by the crossover processing unit 16 with a certain probability, based on the mutation-related parameters output from the genetic algorithm information processing unit 12. The mutation processing unit 17 also outputs the M parameters that have been changed by the mutations to the parameter storage unit 50 (step S32). The parameter storage unit 50 acquires and stores the M new parameters output from the mutation processing unit 17. After the parameter storage unit 50 stores the M new parameters output from the mutation processing unit 17, the stored M new parameters are read out by the evaluation value calculation unit 20, and the evaluation values ​​of the M new parameters are calculated by the evaluation value calculation unit 20.

[0075] Next, the display unit 60 displays the optimized preference information, parameters, evaluation values, and generation number (step S33). In step S33, the display unit 60 displays the optimized preference information obtained by executing step S21, the parameters and evaluation values ​​of the parameters obtained by executing steps S22 to S32, and the generation number of the genetic algorithm when the parameters were obtained. The optimized preference information, parameters, evaluation values, and generation number displayed by the display unit 60 are passed to the display unit 60 via the mutation processing unit 17.

[0076] (Explanation of the operation of the environment selection unit 14 of the optimal solution search unit 10) Figure 10 is a flowchart showing an example of the operation of the environment selection unit 14 of the optimal solution search unit 10 provided in the parameter adjustment device 1 according to Embodiment 1. Steps S41 to S53 shown in Figure 10 show the details of the processing of steps S27 and S28 shown in Figure 9, that is, the operation in which the environment selection unit 14 acquires or selects N individuals and outputs them to the cross selection unit 15.

[0077] In the environment selection unit 14, first, the environment selection setting information acquisition unit 141 acquires environment selection setting information (step S41). Specifically, from the second generation onward, the environment selection setting information acquisition unit 141 acquires environment selection setting information such as parameters, evaluation values, number of generations, and reference direction information from the genetic algorithm information processing unit 12. The environment selection setting information acquisition unit 141 outputs the acquired set of parameters and evaluation values ​​as a child individual to the non-dominant sort processing unit 142.

[0078] Next, the non-controlling sorting processing unit 142 performs a non-controlling sort (step S42). Specifically, the non-controlling sorting processing unit 142 performs a non-controlling sort on N+M parent-child individuals, which are the N parent individuals obtained from the parent individual information storage unit 148 and the M child individuals obtained from the environment selection setting information acquisition unit 141, and ranks the individuals.

[0079] Next, the reference direction setting unit 143 sets the reference direction (step S43). The reference direction setting unit 143 sets the reference direction based on N+M parent-child individuals.

[0080] Next, the individual assignment unit 144 calculates the reference direction to which each individual is assigned (step S44). That is, the individual assignment unit 144 calculates which reference direction each of the N+M parent-child individuals is assigned to.

[0081] Next, the relative score calculation unit 145 calculates the relative score for each of the N+M parent-child individuals (step S45).

[0082] Next, the individual count unit 146 counts the number of individuals belonging to each reference direction (step S46).

[0083] Next, the parent individual selection unit 147 selects individuals based on the rank assigned to each individual in step S42 (step S47). The parent individual selection unit 147 selects N or more individuals based on the rank of each individual.

[0084] Next, the parent individual selection unit 147 determines whether there are N selected individuals, that is, whether the number of individuals selected in step S47 is N (step S48). If there are N selected individuals (step S48: Yes), the parent individual selection unit 147 outputs the selected individuals, etc., to the parent individual information storage unit 148 and the cross selection unit 15 (step S53). Specifically, the parent individual selection unit 147 outputs the N individuals selected in step S47 as new parent individuals to the parent individual information storage unit 148, and outputs the N individuals (parent individuals) selected in step S47, rank information, reference direction information, assignment information, relative score, and count information to the cross selection unit 15.

[0085] In contrast, if the number of selected individuals is not N (Step S48: No), the parent individual selection unit 147 determines whether the reference direction with the fewest number of members has only one member (Step S49). In Step S49, the parent individual selection unit 147 determines whether the reference direction with the fewest number of members has only one member among the reference directions with one or more members, based on the count information showing the count result by the member count unit 146 in Step S46. If the reference direction with the fewest number of members has only one member (Step S49: Yes), the parent individual selection unit 147 selects the individual in question (Step S50) and adds to the count of the reference direction to which the selected individual belongs (Step S52). That is, the parent individual selection unit 147 selects the individual belonging to the reference direction with the fewest number of members and adds 1 to the member count value of the reference direction to which the selected individual belongs. After that, the process returns to Step S48. If there are multiple reference directions with only one individual belonging to each, one individual may be randomly selected from among the individuals belonging to each reference direction with one individual belonging to each, or the individual with the highest relative score may be selected.

[0086] On the other hand, if the reference direction with the fewest number of individuals has two or more individuals (step S49: No), the parent individual selection unit 147 selects an individual based on its relative score (step S51). Specifically, the parent individual selection unit 147 refers to the relative score of each individual belonging to the reference direction with the fewest number of individuals and selects the individual with the highest relative score. After selecting an individual in step S51, the parent individual selection unit 147 proceeds to step S52 and adds 1 to the count value of the number of individuals belonging to the reference direction to which the individual selected in step S51 belongs. Then, it returns to step S48.

[0087] As described above, the parameter adjustment device 1 according to Embodiment 1 accepts input from the user of optimization preference information, such as target values ​​and constraints for evaluation values ​​related to machining results, ranges and constraints for the machining parameters, and setting information related to the multi-objective optimization method, before searching for the parameters of the machining machine, and performs optimization that reflects the input optimization preference information. With the parameter adjustment device 1, the user can obtain parameter adjustment results that match the user's preferences simply by setting the optimization preference information initially, and the workload of the user required for parameter adjustment is reduced. Therefore, the user can obtain parameter values ​​that match their preferences more efficiently than in the past. For example, if the user wants to find good parameter values ​​in a short time, the number of searches is limited, so by setting the internal parameters of the multi-objective optimization method (parameters used in the multi-objective optimization method) to emphasize convergence, optimization that emphasizes convergence is automatically performed, and better parameter values ​​can be obtained with fewer searches. Conversely, if there is ample time for adjustment, by setting the internal parameters of the multi-objective optimization method to emphasize diversity, a variety of candidate parameter values ​​that match the user's preferences can be obtained. Furthermore, if parameter values ​​that match the user's preferences cannot be obtained, the best solution can be selected from the solutions displayed on the display unit 60, and the parameters can be adjusted again to obtain parameter values ​​that match the preferences with minimal effort and time.

[0088] Furthermore, since the parameter tuning device 1 uses a genetic algorithm to simultaneously explore using multiple individuals, it is less likely to fall into local optima compared to methods that explore parameters one set at a time, such as Bayesian optimization, and can obtain a large number of candidate parameter values ​​that match the user's preferences. In addition, when selecting individuals in environment selection, the parameter tuning device 1 performs a non-dominant sort to assign a rank to each individual, selects individuals based on their rank, then assigns each individual to a reference direction, and prioritizes selecting individuals belonging to the reference direction with the fewest number of individuals, thereby ensuring diversity that could not be guaranteed with conventional methods. Moreover, if there are multiple individuals belonging to the reference direction with the fewest number of individuals, the parameter tuning device 1 makes a selection based on relative scores, thereby preserving diversity while leaving individuals with better convergence for the next generation. Since the reference directions play a role in dividing the space, the state where the number of individuals belonging to a reference direction is the fewest indicates that little exploration has been performed in the space around that reference direction, and diversity is maintained by prioritizing the individuals in that space for the next generation. Relative scores evaluate individuals relatively compared to traditional goodness-of-fit methods, making it easier to retain better solutions and improving convergence.

[0089] (Modification 1 of Embodiment 1) Next, Modification 1 of Embodiment 1 will be described. In Modification 1 of Embodiment 1, an operation is added at a certain probability in which individuals with high relative scores cross over within the regions divided by the reference direction.

[0090] In the crossover selection unit 15 of the above-described embodiment 1, when selecting a crossover pair, randomly selects K individuals from the N individuals selected by the environment selection unit 14, and then selects one individual based on some indicator. However, in this method, crossover can occur with any combination of N individuals, and depending on the crossover pair, a good individual may not be generated. Therefore, the crossover selection unit 15 in the modified example 1 of embodiment 1 performs an operation to create a crossover pair by selecting individuals that exist in the same region and have high relative scores, with a certain probability. By crossing individuals with high relative scores that exist in the same region, the possibility of generating a good individual can be increased, and efficient searching becomes possible. In other words, it becomes possible to obtain good processing machine parameters in a short amount of time.

[0091] (Modification 2 of Embodiment 1) Next, Modification 2 of Embodiment 1 will be described. In Modification 2 of Embodiment 1, when the relative score calculation unit 145 of the environment selection unit 14 calculates the relative score, it corrects the individual evaluation values ​​to calculate a more accurate relative score. Methods for correcting the individual evaluation values ​​include converting the distribution of evaluation values ​​from a concave shape to a convex shape, and converting the basis vector of the region in which the individual exists.

[0092] This section describes a method for correcting individual evaluation values ​​by transforming the distribution of evaluation values ​​from a concave shape to a convex shape. Figure 11 shows an example of a method for correcting evaluation values ​​in a parameter adjustment device 1 according to a modified example 2 of Embodiment 1. In Figure 11, the region is divided into three regions, region 1, region 2, and region 3, according to the reference direction, and the individuals belonging to each region are plotted. The left side of Figure 11 shows the distribution of evaluation values ​​before transformation, and the right side shows the distribution of evaluation values ​​after transformation.

[0093] In multi-objective optimization, when the shape of an individual's evaluation value is convex downwards, it is called a convex shape, and when it is convex upwards, it is called a concave shape. One method for calculating relative scores is the envelope analysis method. In the envelope analysis method, a more accurate score can be calculated if the shape of the evaluation value is convex, so an accurate score can be calculated by converting a concave shape to a convex shape. In Figure 11, the distribution of individuals belonging to region 2 is concave, so the relative score calculation unit 145 in the modified example 2 of Embodiment 1 converts the distribution of individuals belonging to region 2 to a convex shape before calculating the relative score. This makes it possible to leave better individuals for the next generation and enables efficient search. In other words, it becomes possible to obtain good machine parameters in a short amount of time.

[0094] This section describes a method for correcting the evaluation value of an individual by transforming the basis vector of the region in which the individual exists. Figure 12 shows an example of another method for correcting the evaluation value in the parameter adjustment device 1 according to a modified example 2 of Embodiment 1. The left side of Figure 12 shows the distribution of the evaluation value before transformation, and the right side shows the distribution of the evaluation value after transformation.

[0095] The relative score calculation unit 145 according to the modified example 2 of Embodiment 1 divides the evaluation value space according to the reference direction. If the space divided by the reference direction r2 is called region R2, the range of region R2 is the space between vectors m1 and m2 represented by dotted lines, and the basis vectors of region R2 are vectors m1 and m2. Transformation of the basis vectors means that when the basis vectors m1 and m2 are moved parallel to each axis of the evaluation value, each evaluation value existing within region R2 is transformed to match the moved basis vectors. In the case of a region divided by the reference direction r1 (the space between vectors r1 and m1, hereinafter referred to as region R1), vectors r1 and m1 become the basis vectors of region R1. In this case, only the basis vector m1 is moved. By moving the basis vectors and transforming the evaluation values, the same effect as when transforming a concave shape to a convex shape can be obtained.

[0096] (Modification 3 of Embodiment 1) Next, Modification 3 of Embodiment 1 will be described. In Modification 3 of Embodiment 1, the reference direction is adaptively changed with each generation. In the parameter adjustment device 1 according to Embodiment 1, the same reference direction is used for every generation until the search termination condition is met once, but using the same reference direction may make the search inefficient. For example, if no individuals belong to a certain reference direction, that reference direction becomes wasted. If such a reference direction exists, the reference direction setting unit 143 according to Modification 3 of Embodiment 1 changes the reference direction to a different direction from the next generation. The method of change may be the same as the known method A-NSGA III (Adaptive Non-dominated Sorting Genetic Algorithm III), or the average value of the relative score of individuals may be calculated for each region corresponding to the reference direction before the change, and the reference direction may be changed to a region near the region with a high average value. This makes it possible to perform robust parameter searches for various optimization problems with different true Pareto front shapes.

[0097] (Modification 4 of Embodiment 1) Next, Modification 4 of Embodiment 1 will be described. In Modification 4 of Embodiment 1, the parent individual information storage unit 148 outputs not only the parent individual of the previous generation, but also individuals from all past generations to the non-controlling sort processing unit 142, so that the environment selection unit 14 can select good solutions that were not selected in past generations.

[0098] In typical genetic algorithms, environment selection involves selecting individuals from a total of N+M individuals (parents and offspring combined) in each generation to reach a population size of N. Individuals not selected are not considered in subsequent generations. However, important individuals may exist among those not selected. To enable the selection of such individuals again, the environment selection unit 14 according to modification 4 of Embodiment 1 stores all past individuals in the parent individual information storage unit 148, allowing for the selection of individuals from all past generations in subsequent generations. The individuals output by the parent individual information storage unit 148 to the non-controlling sort processing unit 142 may be all individuals explored so far, individuals with high relative scores, or individuals assigned a rank of 1 in the non-controlling sort.

[0099] The effects of Modification 4 of Embodiment 1 will now be explained. When assigning individuals equal to the population N in each reference direction and calculating a relative score for each assigned area, the number of individuals present in some areas may be small. Since it may not be possible to calculate an accurate score when there are few individuals, a more accurate score can be calculated by also using individuals from past generations stored in the parent individual information storage unit 148 to calculate the relative score. As a result, it becomes possible to leave good individuals for the next generation, and consequently, it becomes possible to obtain good processing machine parameters in less time.

[0100] (Modification 5 of Embodiment 1) Next, Modification 5 of Embodiment 1 will be described. In Modification 5 of Embodiment 1, when selecting crossover pairs in crossover selection, an operation to pair individuals necessary for calculating the relative score is added with a certain probability.

[0101] In the crossover selection unit 15 of the above-described embodiment 1, when selecting a pair for crossover, it randomly selects K individuals from the N individuals selected by the environment selection unit 14, and then selects one individual based on some indicator. However, in this method, crossover can occur in any combination of the N individuals, and depending on the crossover pair, a good individual may not be generated. Therefore, in the crossover selection unit 15 of modified embodiment 1 5, the operation of pairing individuals necessary for calculating the relative score is performed with a certain probability. Since there are two methods for calculating the relative score: a method using a super CCR model and a method using projection vectors, the method for creating crossover pairs will be explained for each method.

[0102] First, the method using the super CCR model will be explained with reference to Figure 6. When calculating the relative score of individual D shown in Figure 6, the efficient frontier formed by individual C and individual E is used. Therefore, the individuals necessary for calculating the relative score of individual D are individual C and individual E. For this reason, the crossover pairs including individual D will be either the pair of individual D and individual C, or the pair of individual D and individual E. That is, the crossover selection unit 15 according to modified example 5 of Embodiment 1 creates either the pair of individual D and individual C, or the pair of individual D and individual E. The choice of which pair to create may be determined randomly, or the pair may be created based on the higher sum of the relative scores of the two individuals in the pair.

[0103] Next, the method using projection vectors will be explained with reference to Figure 7. When calculating the relative score of individual D shown in Figure 7, the vector OC that maximizes the projection vector onto vector OD is used. Therefore, the crossover pair including individual D is the pair of individual D and individual C. In other words, the crossover selection unit 15 according to the modified example 5 of Embodiment 1 creates the pair of individual D and individual C.

[0104] Both the method using the super CCR model and the method using projection vectors pair up individuals with high relationships for crossover, which increases the likelihood of generating better solutions than normal crossover selection, making it possible to obtain good machine parameters in less time.

[0105] Embodiment 2. In Embodiment 1 described above, the method by which the cross selection unit 15, which constitutes the optimal solution search unit 10 of the parameter adjustment device 1, performs tournament selection using a fixed tournament size set by the user via the optimization preference information setting unit 40 was explained. In contrast, Embodiment 2 describes a parameter adjustment device 1a that performs tournament selection in a way different from the tournament selection described in Embodiment 1.

[0106] The configuration of the parameter adjustment device 1a according to Embodiment 2 is the same as that of the parameter adjustment device 1 according to Embodiment 1, but the configuration and operation of the optimal solution search unit differ from those of Embodiment 1. Hereinafter, the optimal solution search unit provided in the parameter adjustment device 1a according to Embodiment 2 will be referred to as the optimal solution search unit 10a. The cross selection unit that constitutes the optimal solution search unit 10a will be referred to as the cross selection unit 15a.

[0107] The parameter adjustment device 1a according to Embodiment 2 acquires information on the initial value of the tournament size and the adjustment criteria for the tournament size, in addition to the optimization preference information described in Embodiment 1, in the process corresponding to step S1 of the flowchart in Figure 8 showing the operation of the parameter adjustment device 1 according to Embodiment 1, and sets this information in the optimal solution search unit 10a. The optimal solution search unit 10a finds the optimal solution by dynamically adjusting the tournament size based on the tournament size adjustment criteria, performing tournament selection to create crossover pairs, and generating new individuals.

[0108] Figure 13 shows an example of the configuration of the optimal solution search unit 10a included in the parameter adjustment device 1a according to Embodiment 2. In Figure 13, components common to the optimal solution search unit 10 (see Figure 2) included in the parameter adjustment device 1 according to Embodiment 1 are given the same reference numerals as those in the optimal solution search unit 10. The components given the same reference numerals as those in the optimal solution search unit 10 will not be described.

[0109] As shown in Figure 13, the optimal solution search unit 10a according to Embodiment 2 is configured such that the cross selection unit 15 of the optimal solution search unit 10 according to Embodiment 1 is replaced with a cross selection unit 15a, and a tournament size adjustment unit 18 is added between the environment selection unit 14 and the cross selection unit 15a.

[0110] The tournament size adjustment unit 18 adjusts the tournament size for the tournament selection performed when the cross selection unit 15a selects individuals to perform crossing on in the cross processing unit 16, based on the tournament size adjustment criteria. Details of the tournament size adjustment unit 18 will be described later.

[0111] The crossover selection unit 15a selects tournaments according to the tournament size adjusted by the tournament size adjustment unit 18, and the crossover processing unit 16 selects individuals to perform crossovers on.

[0112] The environment selection unit 14 and the cross selection unit 15a, which constitute the optimal solution search unit 10a, may use the method described in Embodiment 1 for individual selection, or they may use methods similar to known evolutionary computation methods for multi-objective optimization such as NSGA2 and AGEMOEA (Adaptive Geometry Estimation based Multi and Many-Objective Evolutionary Algorithms).

[0113] Here, in order to explain the effects of Embodiment 2, the definition of tournament size and the nature of tournament selection will be explained. The definition of tournament size K is an integer between 2 and N, where N is the population. In tournament selection, K individuals are randomly selected from the N individuals selected by the environment selection unit 14, and then one more individual is selected from K individuals based on some criterion. Although it depends on how the selection criteria are set, generally, the closer K is to N, the more convergence of the population being explored tends to increase, and the less diversity tends to be. Conversely, the closer K is to 2, the less convergence of the population being explored tends to be, and the more diversity tends to be.

[0114] Next, the details of the tournament size adjustment unit 18 will be explained. The tournament size adjustment unit 18 obtains information on the tournament size adjustment criteria from the genetic algorithm information processing unit 12 and obtains parameters, evaluation values, and the number of generations from the environment selection unit 14. The tournament size adjustment criteria can be anything. In this embodiment, as an example, we will explain the case in which the tournament size adjustment unit 18 adjusts the tournament size using the number of generations as the adjustment criterion. There are mainly two methods for using the number of generations as the adjustment criterion. The first is to set the initial value of the tournament size to be small and increase the tournament size based on the number of generations. The second is, conversely, to set the initial value of the tournament size to be large and decrease the tournament size based on the number of generations.

[0115] Figure 14 shows an example of a method for adjusting the tournament size by the tournament size adjustment unit 18 of the optimal solution search unit 10a according to Embodiment 2. Figure 14 shows an example where the tournament size is increased according to the generations. In Figure 14, the horizontal axis is the number of generations and the vertical axis is the tournament size, with the initial value of the tournament size being 2. In the adjustment method shown in Figure 14, the tournament size is increased by 10 every time the number of generations increases by 5. The effect of the method of increasing the tournament size according to the generations as shown in Figure 14 will be explained. Generally, in multi-objective optimization problems, individuals can be easily brought closer to the Pareto front in the early stages of the search. Therefore, it is better to set the tournament size to a small value in the early stages of the search to ensure diversity. Also, in the later stages of the search, as the individuals converge, a certain degree of diversity is ensured, but it becomes difficult to bring them closer to the Pareto front. Therefore, it is better to set the tournament size to a larger value to improve convergence.

[0116] Next, let's discuss the effectiveness of the method of gradually decreasing the tournament size. In problems where it is difficult to bring individuals closer to the Pareto front in the initial search, reducing the tournament size in the early stages of the search may require many search iterations. In such cases, it is better to set the tournament size large and gradually decrease it. If the search is conducted with a large tournament size, there is a risk of convergence to a local minimum, so gradually decreasing the tournament size helps to ensure diversity.

[0117] As described above, the optimal solution search unit 10a of the parameter adjustment device 1a according to Embodiment 2 includes a tournament size adjustment unit 18 that adjusts the tournament size when the crossover selection unit 15a selects individuals for crossover by the crossover processing unit 16 through tournament selection. In normal tournament selection, the tournament size K is fixed to the value initially set, and the same tournament size is used for all generations, but the optimal tournament size K is likely to differ for each generation. In contrast, the optimal solution search unit 10a of the parameter adjustment device 1a according to this embodiment includes a tournament size adjustment unit 18 that adaptively adjusts the tournament size for each generation, making it possible to obtain candidate processing machine parameter values ​​that match the user's preferences in a short amount of time.

[0118] (Modification 1 of Embodiment 2) Next, Modification 1 of Embodiment 2 will be described. In Modification 1 of Embodiment 2, diversity and convergence scores are adopted as criteria for adjusting the tournament size. In Embodiment 2 described above, the adjustment of the tournament size according to the number of generations was explained, but the tournament size is not determined by analyzing the explored individuals, and there is a possibility that the optimal tournament size is not set for each generation. For this reason, in the optimal solution search unit 10a of the parameter adjustment device 1a according to Modification 1 of Embodiment 2, the criteria for adjusting the tournament size are diversity and convergence scores obtained by analyzing individuals selected in past searches. The diversity and convergence scores may be calculated using known multi-objective optimization indices such as HV (Hyper Volume) and GD (Generational Distance), or they may be scores calculated by a unique method. In the case of the tournament size adjustment unit 18 of the optimal solution search unit 10a, for example, if GD, which evaluates convergence, is used as the criteria for adjusting the tournament size, the tournament size is increased to improve convergence when the improvement in GD is low, and on the other hand, the tournament size is decreased to improve diversity when the improvement in GD is high. By using diversity and convergence scores as adjustment criteria, it becomes more likely that the optimal tournament size can be set for each generation, improving the diversity and convergence of the resulting solutions. Furthermore, the emphasis on convergence or diversity during the search can be adjusted according to the user's preferences. For example, if the goal is to find good parameters quickly, the number of searches will be limited, so the tournament size can be adjusted to improve the convergence score. This allows for obtaining better parameters with fewer searches. Conversely, if there is ample time for adjustment, the tournament size can be adjusted to emphasize diversity. This allows for obtaining a larger number of candidate parameter values ​​that match the user's preferences.

[0119] (Modification 2 of Embodiment 2) Next, Modification 2 of Embodiment 2 will be described. In Modification 2 of Embodiment 2, the tournament size of each tournament during a single tournament selection is changed. In normal tournament selection and Embodiment 2 described above, L tournaments are held during a single tournament selection performed to select L individuals, and the same tournament size value is used for each tournament. In contrast, in Modification 2 of Embodiment 2, the tournament size is changed for each of the L tournaments. The method for determining the tournament size in each tournament may be determined randomly, based on the degree of fit calculated by the environment selection unit 14, or based on the diversity and convergence score described in Modification 1 of Embodiment 2. When determining the tournament size based on the diversity and convergence score, for example, if the diversity score is high and the convergence score is low, convergence is improved by holding more tournaments with a large tournament size value among the L tournaments. This increases the possibility that the optimal tournament size will be set for each tournament, and a large number of candidate parameter values ​​that match the user's preferences can be obtained.

[0120] (Modification 3 of Embodiment 2) Next, Modification 3 of Embodiment 2 will be described. In Modification 3 of Embodiment 2, the shape of the Pareto front estimated from the individuals acquired by the tournament size adjustment unit 18 is adopted as the adjustment criterion for the tournament size. As the method for estimating the Pareto front, known methods such as AGEMOEA and GFM-MOEA (Generic Front Modeling based Multi and Many-Objective Evolutionary Algorithms) can be used. For example, if the shape of the Pareto front has not converged, the tournament size adjustment unit 18 of the optimal solution search unit 10a will decrease the tournament size to broaden the distribution of individuals (improve diversity), and if the shape has converged, it will increase the tournament size to improve convergence. This increases the possibility of setting an appropriate tournament size according to the shape of the Pareto front, and a variety of candidate parameter values ​​that match the user's preferences can be obtained.

[0121] (Modification 4 of Embodiment 2) Next, Modification 4 of Embodiment 2 will be described. In Modification 4 of Embodiment 2, the user sets an ideal reference point, and the relationship with the reference point is used as the adjustment criterion for the tournament size. If the tournament size adjustment unit 18 of the optimal solution search unit 10a adjusts the tournament size according to the distance to the reference point, for example, it adjusts the tournament size to be small in the initial stages of the search so that the search does not concentrate on local minima, and adjusts the tournament size to be large as the search progresses, thereby improving convergence. As a result, the speed of convergence to the reference point is improved, and candidate parameter values ​​that match the user's preferences can be obtained in less time compared to the conventional method.

[0122] (Modification 5 of Embodiment 2) Next, Modification 5 of Embodiment 2 will be described. In Modification 5 of Embodiment 2, the relative score calculated by the environment selection by the environment selection unit 14 is adopted as the adjustment criterion for the tournament size. An example of using the relative score as an adjustment criterion is a method that utilizes the variability of the relative score. In a genetic algorithm, as the search progresses and individuals converge to the true Pareto front, the individuals tend to be distributed uniformly. When individuals are uniformly distributed, the relative score of each individual is determined by relative evaluation with other individuals, so the variability of the relative score becomes small. For this reason, the variability of the relative score can be considered an evaluation index of the convergence of individuals.

[0123] Figure 15 shows an example of how the tournament size is adjusted by the tournament size adjustment unit 18 of the optimal solution search unit 10a according to modification 5 of Embodiment 2. On the left side of Figure 15, individuals are plotted in three stages: initial, mid-search, and final search. Numerical values ​​indicating the relative score of each individual are also shown near each individual. On the right side of Figure 15, the relationship between the progress of the search and the tournament size is shown on the vertical axis, with the progress of the search on the horizontal axis and the tournament size on the vertical axis. On the left side of Figure 15, the distribution of individuals is scattered in the initial stages, resulting in a large variation in the relative scores of each individual. As the search progresses, the distribution of individuals becomes more uniform, and the variation in relative scores decreases. When the variation in relative scores is large, it indicates that the individuals have not converged to a Pareto front. Therefore, the tournament size adjustment unit 18 of the optimal solution search unit 10a adjusts the tournament size to a small value to avoid falling into a local minimum, thereby ensuring that selection is made with an emphasis on diversity. On the other hand, when the variability of the relative scores decreases, the tournament size adjustment unit 18 of the optimal solution search unit 10a adjusts the tournament size to a larger value to indicate that the individuals are converging to the Pareto front, thereby prioritizing convergence in the selection process. This allows for obtaining candidate parameter values ​​that match the user's preferences in a shorter amount of time.

[0124] Next, the hardware configurations of the parameter adjustment devices 1 and 1a according to the above-described embodiments will be explained. Since the hardware configurations of the parameter adjustment devices 1 and 1a are similar, the hardware configuration of the parameter adjustment device 1 according to Embodiment 1 will be explained as an example.

[0125] Figure 16 shows an example of hardware that realizes the parameter adjustment device 1 according to Embodiment 1. The parameter adjustment device 1 is realized by the input circuit 91 shown in Figure 16, the processor 92 and memory 93 that constitute the control circuit, and the display device 94.

[0126] The input circuit 91 is an electronic circuit that receives signals from the outside. The processor 92 is, for example, a CPU (Central Processing Unit), central processing unit, processing unit, arithmetic unit, microprocessor, microcomputer, DSP (Digital Signal Processor), etc. The memory 93 is, for example, a non-volatile or volatile semiconductor memory such as RAM (Random Access Memory), ROM (Read Only Memory), flash memory, EPROM (Erasable Programmable Read Only Memory), EEPROM (Registered Trademark) (Electrically Erasable Programmable Read Only Memory), magnetic disk, flexible disk, optical disk, compact disk, minidisc, DVD (Digital Versatile Disk), etc. The display device 94 is, for example, a liquid crystal monitor, display, etc.

[0127] The optimal solution search unit 10, evaluation value calculation unit 20, and optimization preference information setting unit 40 of the parameter adjustment device 1 are realized by the processor 92 and memory 93. Specifically, a program for operating as the optimal solution search unit 10, evaluation value calculation unit 20, and optimization preference information setting unit 40 of the parameter adjustment device 1 is stored in memory 93, and the processor 92 reads and executes this program to realize the optimal solution search unit 10, evaluation value calculation unit 20, and optimization preference information setting unit 40. The above program stored in memory 93 may be provided to the user in the form of being written on a storage medium such as a CD (Compact Disc)-ROM or DVD-ROM, or it may be provided to the user via a network.

[0128] The evaluation value storage unit 30 and parameter storage unit 50 of the parameter adjustment device 1 are implemented by the memory 93. The display unit 60 of the parameter adjustment device 1 is implemented by the display device 94.

[0129] Furthermore, the optimal solution search unit 10, evaluation value calculation unit 20, and optimization preference information setting unit 40 of the parameter adjustment device 1, which are implemented by the processor 92 and memory 93 shown in Figure 16, can also be implemented with dedicated hardware. The dedicated hardware that implements the optimal solution search unit 10, evaluation value calculation unit 20, and optimization preference information setting unit 40 of the parameter adjustment device 1 may be, for example, a single circuit, a composite circuit, a programmed processor, a parallel programmed processor, an ASIC (Application Specific Integrated Circuit), an FPGA (Field Programmable Gate Array), or a combination thereof.

[0130] Furthermore, it is also possible to implement parts of the optimal solution search unit 10, evaluation value calculation unit 20, and optimization preference information setting unit 40 of the parameter adjustment device 1 using dedicated hardware, while the remaining parts are implemented by a control circuit consisting of a processor 92 and memory 93.

[0131] The configurations shown in the above embodiments are examples only, and can be combined with other known technologies, combined with other embodiments, and some parts of the configuration can be omitted or modified without departing from the gist of the invention. For example, although each embodiment describes an example of adjusting the parameters of a machine tool (processing machine), the method can also be applied to optimizing the parameters of other devices or systems using a multi-objective optimization method.

[0132] In each embodiment and each modification, for the sake of explanation, the object of multi-objective optimization performed by the parameter adjustment device is described as a parameter. However, in reality, the parameter adjustment device performs multi-objective optimization on multiple types of parameters, i.e., parameter sets.

[0133] 1, 1a Parameter adjustment device, 10, 10a Optimal solution search unit, 11 Optimization preference information acquisition unit, 12 Genetic algorithm information processing unit, 13 Initial parameter generation unit, 14 Environment selection unit, 15, 15a Crossover selection unit, 16 Crossover processing unit, 17 Mutation processing unit, 18 Tournament size adjustment unit, 20 Evaluation value calculation unit, 30 Evaluation value storage unit, 40 Optimization preference information setting unit, 50 Parameter storage unit, 60 Display unit, 141 Environment selection setting information acquisition unit, 142 Non-dominant sort processing unit, 143 Reference direction setting unit, 144 Individual assignment unit, 145 Relative score calculation unit, 146 Number of affiliated individuals count unit, 147 Parent individual selection unit, 148 Parent individual information storage unit.

Claims

An evaluation value calculation unit that simulates the operation of a machining tool and calculates an evaluation value of the machining result corresponding to the parameter set of the machining tool used in the simulation, An optimal solution search unit that uses a multi-objective optimization method to optimize each of the aforementioned evaluation values ​​and searches for candidate parameter sets that match the user's preferences, An optimization preference information setting unit obtains optimization preference information from the user, including user preference information regarding the parameter set and user preference information regarding the processing results, and reflects this in the optimization of the evaluation value. A parameter adjustment device characterized by comprising the following features. 　 A display unit that displays a plurality of candidate parameter sets, the evaluation value corresponding to each of the plurality of candidates, and the optimization preference information. The parameter adjustment device according to claim 1, further comprising the features described above. 　 Using a genetic algorithm as the multi-objective optimization method, The aforementioned optimal solution search unit, A genetic algorithm information processing unit manages the number of generations of the genetic algorithm and specifies the conditions for performing each process of the genetic algorithm based on the number of generations and the optimization preference information. An environment selection unit calculates the degree of fit of the parameter set and, based on the calculated degree of fit, selects N individuals from the N+M individuals (where N and M are integers of 2 or more) that are candidates for the solution of the parameter set. A crossover selection unit selects L individuals (where L is an integer of 2 or more) from the N individuals selected by the environment selection unit to perform crossover based on the conditions, A crossover processing unit performs crossover based on the conditions on the L individuals selected by the crossover selection unit to generate M new individuals, A mutation processing unit that modifies individuals by applying mutations to M individuals generated by the crossover processing unit, The parameter adjustment device according to claim 1 or 2, characterized by comprising the above. 　 The environment selection unit performs the selection using the genetic algorithm on individuals from all generations that have been explored in the past. The parameter adjustment device according to feature 3. 　 The aforementioned environment selection unit is A reference direction setting unit sets one or more reference directions that divide the space in which the aforementioned evaluation value exists, An individual assignment unit calculates the vertical distance between each individual and each reference direction, and assigns each individual to the reference direction that minimizes the calculated vertical distance. A relative score calculation unit performs a relative evaluation of the individual being evaluated compared to other individuals and calculates a relative score as the fitness score that indicates how much better the individual being evaluated is compared to other individuals. A parent individual selection unit that, by referring to the number of individuals belonging to the aforementioned reference direction, grasps how much each region in the space divided by the aforementioned reference direction has been explored, preferentially selects individuals in regions where the exploration is sparse, and further selects individuals based on the aforementioned relative score, The parameter adjustment device according to claim 3 or 4, characterized by being equipped with the following: 　 The aforementioned environment selection unit is A non-dominant sorting processing unit that determines the rank of each individual based on the dominance relationships between individuals, A reference direction setting unit sets one or more reference directions that divide the space in which the aforementioned evaluation value exists, An individual assignment unit calculates the vertical distance between each individual and each reference direction, and assigns each individual to the reference direction that minimizes the calculated vertical distance. A relative score calculation unit performs a relative evaluation of the individual being evaluated compared to other individuals and calculates a relative score as the fitness score that indicates how much better the individual being evaluated is compared to other individuals. A parent individual selection unit selects a number of individuals up to or less than N based on the rank of each individual, and if the number of individuals selected based on the rank is less than N, it selects additional individuals based on the relative score until the number of selected individuals reaches N. The parameter adjustment device according to claim 3 or 4, characterized by being equipped with the following: 　 When the parent individual selection unit selects additional individuals based on the relative score, it refers to the number of individuals belonging to the reference direction to understand how much each region in the space divided by the reference direction has been explored, and prioritizes selecting individuals in regions where the exploration is sparse. The parameter adjustment device according to feature 6. 　 The relative score calculation unit calculates the relative score using the super-Charnes-Cooper-Rhodes model of envelope analysis. The parameter adjustment device according to any one of features 5 to 7. 　 The relative score calculation unit calculates the relative score based on a vector from the reference point to the individual being evaluated, and a projection vector obtained by projecting the vector from the reference point to individuals other than the individual being evaluated onto the vector from the reference point to the individual being evaluated. The parameter adjustment device according to any one of features 5 to 7. 　 The crossover selection unit performs a process with a certain probability to pair individuals with high relative scores from among individuals belonging to the same reference direction. The parameter adjustment device according to any one of features 5 to 9. 　 The relative score calculation unit corrects the evaluation values ​​so that the distribution of the evaluation values ​​of multiple individuals becomes convex or linear if the distribution is not convex or linear, and calculates the relative score based on the corrected evaluation values. The parameter adjustment device according to any one of claims 5 to 10. 　 The relative score calculation unit moves the basis vectors of each space divided by the reference direction so that they are parallel to each axis of the space in which the evaluation value exists, corrects the evaluation value to match the moved basis vector, and calculates the relative score based on the corrected evaluation value. The parameter adjustment device according to any one of claims 5 to 10. 　 The reference direction setting unit adaptively changes the reference direction. The parameter adjustment device according to any one of features 5 to 12. 　 The crossover selection unit selects individuals that are considered to have a high relationship when the relative score calculation unit calculates the relative score, and forms a crossover pair. The parameter adjustment device according to any one of features 5 to 13. 　 The crossover selection unit selects the L individuals to perform the crossover using tournament selection, in which the tournament size is dynamically adjusted. The parameter adjustment device according to any one of features 5 to 14. 　 A tournament size adjustment unit adjusts the tournament size using at least one of the following: the number of generations of the search performed by the optimal solution search unit using the genetic algorithm, the diversity and convergence evaluation results of the individuals in the current generation, the relationship between the evaluation value set by the user and a reference point, the shape of the Pareto front, and the relative score. The parameter adjustment device according to claim 15, characterized by comprising the above. 　 A tournament size adjustment unit adjusts the size of each tournament in the tournament selection that the crossover selection unit performs L times to select L individuals in each generation of the genetic algorithm, using at least one of random sampling, the goodness of fit, and the evaluation results of the diversity and convergence of individuals in the current generation. The parameter adjustment device according to claim 15, characterized by comprising the above. 　 A first step involves simulating the operation of a machining tool and calculating an evaluation value of the machining result corresponding to the parameter set of the machining tool used in the simulation. A second step involves optimizing each of the aforementioned evaluation values ​​using a multi-objective optimization method and searching for candidate parameter sets that match the user's preferences. A third step involves obtaining optimization preference information from the user, including user preference information regarding the parameter set and user preference information regarding the processing results, and reflecting this in the optimization of the evaluation value. A parameter adjustment method characterized by including the following. 　 A first step involves simulating the operation of a machining tool and calculating an evaluation value of the machining result corresponding to the parameter set of the machining tool used in the simulation. A second step involves optimizing each of the aforementioned evaluation values ​​using a multi-objective optimization method and searching for candidate parameter sets that match the user's preferences. A third step involves obtaining optimization preference information from the user, including user preference information regarding the parameter set and user preference information regarding the processing results, and reflecting this in the optimization of the evaluation value. A parameter adjustment program characterized by having a computer execute it.

Images