Design apparatus and design method

Abstract

The present invention provides a design apparatus and design method that enable optimal design even when there is a strong dependency between design variables. [Solution] Multiple design variables are grouped into multiple groups. The behavioral learning model stored in the memory unit of the design device receives design information in order from the first group and outputs design variables in order from the first group. The design information for the nth group includes the design variables of the (n-1)th group output by the behavioral learning model. The design device updates the database by adding a design dataset including the input and output of the behavioral learning model to the database, updates the value learning model using the updated database, and updates the behavioral learning model using the updated database so that the value output by the updated value learning model is higher.

Description

[Technical Field]

[0001] The present invention relates to a design apparatus and a design method. [Background technology]

[0002] Patent Document 1 describes a design system for automatically performing product design. The design system described in Patent Document 1 designs a rotor for a rotating machine, which includes a magnet and an iron core, as a product design. The design system optimizes the dimensions of the magnet and the shape of the iron core by modifying a first design variable related to the dimensions of the magnet and a second variable related to the shape of the iron core, respectively. [Prior art documents] [Patent Documents]

[0003] [Patent Document 1] Japanese Patent Publication No. 2023-136975 [Overview of the Initiative] [Problems that the invention aims to solve]

[0004] When multiple design variables exist, they may influence each other, meaning there can be strong dependencies between them. In such cases, optimization calculations may take a long time or even fail. [Means for solving the problem]

[0005] A design device that solves the above problem is a design device that includes an output unit that outputs a plurality of design variables according to a design requirement, wherein the plurality of design variables are grouped into a plurality of groups, and when design information set for each group is input, an action learning model that is learned to output the design variables of the input group, when a pair of the design information input to the action learning model and the design variables output by the action learning model for the design information is input, a value learning model that is learned to output the value of the pair, a database in which a design dataset including the design requirement and the plurality of design variables is stored, a storage unit that stores the action learning model and the value learning model, and the action learning model is the first The design information is input in order from the group, and the design variables are output in order from the first group, the design information of the first group includes the design requirements, and the design information of the nth group includes at least the design variables of the (n-1)th group output by the behavioral learning model. The model update unit updates the database by adding the design dataset, which includes the design requirements input to the behavioral learning model and the multiple design variables output by the behavioral learning model, to the database as an update process, updates the value learning model using the updated database, and updates the behavioral learning model using the updated database so that the value output by the updated value learning model is higher.

[0006] Multiple design variables are grouped into multiple groups. The behavioral learning model receives design information as input, starting from the first group, and outputs design variables in the same order, starting from the first group. The design information for the nth group input to the behavioral learning model includes the design variables of the (n-1)th group output by the behavioral learning model. Therefore, for multiple design variables, highly related design variables are set in the same group, and the order of the groups is set so that they are arranged in order of priority from highest to lowest. As a result, even when there are strong dependencies between design variables, the design device can output appropriate design variables.

[0007] The behavioral learning model is updated to increase the value output by the value learning model. As a result, the behavioral learning model can output more appropriate design variables. The database stores design datasets. The behavioral learning model and value learning model are updated using the design datasets stored in the database. Therefore, if the design requirements input to the design device are similar to the design requirements in the database, the design device can immediately output multiple design variables. Consequently, design time can be reduced. This is particularly effective when there are many design requirements or when the design requirements are complex, as design time tends to be long in such cases. Thus, the design can be carried out optimally.

[0008] With respect to the above-mentioned design apparatus, the model update unit may repeatedly execute the update process. With this configuration, the behavioral learning model can output even more appropriate design variables. Furthermore, even if the design requirements input to the design device are not similar to the design requirements in the database, the model update unit can repeatedly perform update processing, allowing the design device to output appropriate design variables. In addition, as described above, design datasets are accumulated in the database, and the behavioral learning model and value learning model are updated using the design datasets accumulated in the database, thus increasing the number of design requirements that can be handled.

[0009] In the above-described design apparatus, the storage unit stores a separate behavioral learning model that, when the design information is input, outputs the design variable in the database that corresponds to the input design information and has the highest evaluation value. The model update unit may update the separate behavioral learning model using the updated database, calculate a Q-learning teacher signal using the design variable output by the updated separate behavioral learning model, and update the value learning model by Q-learning the value learning model using the calculated Q-learning teacher signal.

[0010] When Q-training a value learning model to output the value of the nth group of pairs, it is necessary to find the design variables that maximize the value of the (n+1)th group of pairs when calculating the training signal for Q-training.

[0011] In typical reinforcement learning, the agent generates diverse data through trial-and-error actions, thereby advancing the learning of the behavioral and value learning models. This allows the behavioral and value learning models to learn appropriate outputs. However, in the design field, generating diverse data can be difficult because the time required to calculate evaluation values ​​tends to be long. As a result, the behavioral learning model may not be sufficiently trained, and it may not be possible to appropriately determine the design variables that maximize the value of the (n+1)th group pair. In this case, it may not be possible to correctly calculate the Q-learning training signals used for Q-learning in the value learning model.

[0012] In contrast, the above configuration calculates the Q-learning teacher signal using the design variables output by the separate behavioral learning model. The design variables output by the separate behavioral learning model are the design variables in the database that correspond to the input design information and have the highest evaluation value. Therefore, the design variable that maximizes the value of the (n+1)th group pair can be replaced with the design variable output by the separate behavioral learning model. Consequently, the value learning model can be updated without generating diverse data.

[0013] With respect to the above-described design apparatus, the model update unit may perform an improvement process to improve the design variables output by the behavioral learning model so that their evaluation values ​​are higher, and add the design dataset, which includes the improved design variables, to the database.

[0014] According to the above configuration, the behavioral learning model and the value learning model are updated using a database containing improved design variables. Therefore, the behavioral learning model and the value learning model can be updated efficiently.

[0015] With respect to the above-described design apparatus, the model update unit may, in the improvement process, make the range of change of the design variable for the nth group narrower than the range of change of the design variable for the (n+1)th group.

[0016] According to the above configuration, the dependency between the design variables of the nth group and the design variables of the (n+1)th group can be reduced. In the above-described design device, the behavioral learning model may be composed of the same number of segmented behavioral learning models as the number of groups, with the nth segmented behavioral learning model receiving the design information of the nth group as input and outputting the design variables of the nth group; and the value learning model may be composed of the same number of segmented value learning models as the number of groups, with the nth segmented value learning model receiving the pair of the nth group as input and outputting the value of the nth pair.

[0017] Regarding the design device described above, the behavioral learning model may be an integrated behavioral learning model configured to be able to input the design information of any group. With the above configuration, the integrated behavioral learning model can be reused across multiple groups, allowing a single integrated behavioral learning model to output design variables for all groups. Therefore, compared to a case where the behavioral learning model is composed of the same number of segmented behavioral learning models as there are groups, the amount of memory used can be reduced and implementation can be simplified.

[0018] With respect to the above-described design device, the number of input frames for the behavioral learning model for inputting the design information into the integrated behavioral learning model is equal to the sum of the number of design requests and the number of design variables in the first to n-1 groups; the number of output frames for the behavioral learning model for the integrated behavioral learning model to output the design variables is equal to the number of design variables in the group with the largest number of design variables among the multiple groups; and when the design information of the nth group is input into the integrated behavioral learning model, null or 0 may be input to at least the input frames for the behavioral learning model corresponding to the design variables of the n+1th group and beyond among the multiple input frames for the behavioral learning model.

[0019] As described above, the behavioral learning model outputs design variables in order from the first group. Therefore, when the behavioral learning model outputs the design variables for the nth group, the design variables for the (n+1)th and subsequent groups are not output. With the above configuration, when the design information for the nth group is input to the integrated behavioral learning model, null or 0 is entered into the input fields for the behavioral learning model corresponding to the design variables for the (n+1)th and subsequent groups. Thus, when the integrated behavioral learning model outputs the design variables for the nth group, it is possible to suppress the influence of the values ​​entered into the input fields for the behavioral learning model corresponding to the design variables for the (n+1)th and subsequent groups.

[0020] With respect to the above-described design device, the design information input to the integrated behavioral learning model may be assigned a behavioral learning model label indicating which group of the design information it belongs to.

[0021] With the above configuration, the integrated behavioral learning model can more easily recognize which group of design information the input design information belongs to. Regarding the design apparatus described above, the value learning model may be an integrated value learning model configured to accept any pair from any group as input.

[0022] With the above configuration, the integrated value learning model can be reused across multiple groups, allowing a single integrated value learning model to output the values ​​of all group pairs. Therefore, compared to a case where the value learning model is composed of the same number of partitioned value learning models as there are groups, the amount of memory used can be reduced and the implementation can be simplified.

[0023] With respect to the above design device, the number of input frames for the value learning model for inputting the pair into the integrated value learning model is the same as the sum of the number of design requirements and the total number of design variables, the number of output frames for the value learning model for outputting the value of the pair is one, and when the nth group of the pair is input into the integrated value learning model, null or 0 may be input to at least the value learning model input frames corresponding to the design variables of the (n+1)th group and beyond among the multiple value learning model input frames.

[0024] According to the above configuration, when the nth group pair is input to the integrated value learning model, null or 0 is input to the value learning model input frame corresponding to the design variables of at least the (n+1)th group and beyond. Therefore, when the value learning model outputs the value of the nth group pair, it is possible to suppress the influence of the values ​​input to the value learning model input frame corresponding to the design variables of the (n+1)th group and beyond.

[0025] With respect to the above-described design device, the pairs input to the integrated value learning model may be assigned a value learning model label indicating which group the pair belongs to. With the above configuration, the integrated value learning model can more easily recognize which group a given input pair belongs to.

[0026] Regarding the design device described above, the separate behavioral learning model may be an integrated separate behavioral learning model configured to allow input of the design information for any group. This configuration allows the integrated separate behavioral learning model to be reused across multiple groups, enabling the output of design variables for all groups using a single integrated separate behavioral learning model. Therefore, it reduces the amount of memory used and simplifies implementation.

[0027] With respect to the above design device, the number of input frames for the separate behavioral learning model for inputting the design information into the integrated separate behavioral learning model is the same as the sum of the number of design requests and the number of design variables in the 1st to n-1th groups. The number of output frames for the separate behavioral learning model for the integrated separate behavioral learning model to output the design variable in the database that corresponds to the input design information and has the highest evaluation value is the same as the number of design variables in the group with the largest number of design variables among the multiple groups. When the design information of the nth group is input into the integrated separate behavioral learning model, null or 0 may be input to at least the input frames for the separate behavioral learning model corresponding to the design variables of the n+1th group or later among the multiple separate behavioral learning model input frames.

[0028] With this configuration, when the design information for the nth group is input into the integrated separate behavioral learning model, null or 0 is entered into the input fields for the separate behavioral learning model corresponding to the design variables of the (n+1)th and subsequent groups. Therefore, when the integrated separate behavioral learning model outputs the design variables for the nth group, it is possible to suppress the influence of the values ​​entered into the input fields for the separate behavioral learning model corresponding to the design variables of the (n+1)th and subsequent groups.

[0029] With respect to the above-described design device, the design information input to the integrated separate-frame behavioral learning model may be assigned a separate-frame behavioral learning model label indicating which group of the design information it belongs to.

[0030] This configuration makes it easier for the integrated, separate-frame behavioral learning model to recognize which group of design information the input belongs to. The above-described design apparatus is used for designing a rotating electric machine having a rotor core and magnets provided on the rotor core, and the first group may include the design variables relating to the position and shape of the magnets, and the second group may include the design variables relating to the shape of the rotor core.

[0031] Since the appropriate rotor core shape depends on the position and shape of the magnets, design variables related to the position and shape of the magnets have higher priority than design variables related to the shape of the rotor core. According to the above configuration, design variables related to the position and shape of the magnets are grouped into the first group, while design variables related to the shape of the rotor core are grouped into the second group. Therefore, the design of the rotating electric machine can be carried out considering the priority of the design variables.

[0032] The above-described design apparatus may be used in the design of a rotating electric machine having a rotor core and magnets provided on the rotor core, wherein the first group includes the design variables relating to the dimensions of the rotor core, the second group includes the design variables relating to the position and shape of the magnets, and the third group may include the design variables relating to the shape of the rotor core.

[0033] Since the size of the rotor core determines the size of the rotating electric machine, design variables related to the rotor core size have higher priority than design variables related to the position and shape of the magnets, and design variables related to the shape of the rotor core. Also, since the appropriate shape of the rotor core depends on the position and shape of the magnets, design variables related to the position and shape of the magnets have higher priority than design variables related to the shape of the rotor core. According to the above configuration, the design variables related to the rotor core size, the design variables related to the position and shape of the magnets, and the design variables related to the shape of the rotor core are grouped into the first, second, and third groups, respectively. Therefore, the design of the rotating electric machine can be carried out considering the priority of the design variables.

[0034] A design method for solving the above problem is a design method in which a design device outputs a plurality of design variables in response to a design requirement, wherein the plurality of design variables are grouped into a plurality of groups, and the design device includes a storage unit that stores: an action learning model that is trained to output the design variables of the input group when design information set for each group is input, a value learning model that is trained to output the value of a pair when a pair of the design information input to the action learning model and the design variables output by the action learning model for the design information is input, and a database in which a design dataset including the design requirement and the plurality of design variables is stored, wherein the action learning model receives the design information in order from the first group, and the first The update step includes the steps of: outputting the design variables in order from the groups, wherein the design information for the first group includes the design request, and the design information for the nth group includes at least the design variables for the (n-1)th group output by the behavioral learning model, and updating the database by adding the design dataset, which includes the design request input to the behavioral learning model and a plurality of design variables output by the behavioral learning model, to the database; updating the value learning model using the updated database; and updating the behavioral learning model using the updated database so that the value output by the updated value learning model is higher.

[0035] According to the above method, the design can be carried out favorably. Regarding the above design method, the design apparatus may repeatedly perform the update step.

[0036] This method allows the behavioral learning model to output even more appropriate design variables. Furthermore, even if the design requirements input to the design device are not similar to those in the database, the device can still output appropriate design variables. Moreover, the number of design requirements it can handle increases.

[0037] Regarding the above design method, the memory unit stores a separate behavioral learning model that, when the design information is input, outputs the design variable in the database that corresponds to the input design information and has the highest evaluation value. The update step includes the step in which the design device updates the separate behavioral learning model using the updated database. In the step of updating the value learning model using the updated database, the design device may update the value learning model by calculating a Q-learning teacher signal using the design variable output by the updated separate behavioral learning model and Q-learning the value learning model using the calculated Q-learning teacher signal.

[0038] This method allows value learning models to be updated without generating diverse data. [Effects of the Invention]

[0039] According to the present invention, design can be performed effectively even when there is a strong dependence between design variables. [Brief explanation of the drawing]

[0040] [Figure 1] Figure 1 is a schematic diagram of a rotating electric machine. [Figure 2] Figure 2 is a magnified view showing a part of the rotor. [Figure 3] Figure 3 is a diagram showing the configuration of the design system. [Figure 4] Figure 4 is a diagram showing the configuration of the memory unit. [Figure 5] Figure 5 shows the first segmented behavior learning model. [Figure 6] Figure 6 shows the second segmented behavior learning model. [Figure 7] Figure 7 shows the first partitioned value learning model. [Figure 8] Figure 8 shows the second part-value learning model. [Figure 9] Figure 9 shows the first segmented, separate-frame behavioral learning model. [Figure 10] Figure 10 shows the second segmented, separate-frame behavioral learning model. [Figure 11] Figure 11 is a flowchart showing the design method. [Figure 12] Figure 12 shows an integrated behavioral learning model. [Figure 13] Figure 13 shows an example of an integrated behavioral learning model. [Figure 14] Figure 14 shows an integrated value learning model. [Figure 15] Figure 15 shows the integrated value learning model in the modified example. [Figure 16] Figure 16 shows an integrated separate-frame behavioral learning model. [Figure 17] Figure 17 shows an example of an integrated separate behavioral learning model. [Modes for carrying out the invention]

[0041] An embodiment of the design apparatus 21 and design method will be described below with reference to Figures 1 to 14. The design target of this embodiment is the rotor 11 of a rotating electric machine 10. The design apparatus 21 and design method are used for the design of the rotor 11 of the rotating electric machine 10.

[0042] <Rotating electric machine 10> As shown in Figure 1, the rotating electric machine 10 includes a rotor 11. The rotor 11 has a rotor core 12 and a plurality of magnets 13 provided on the rotor core 12.

[0043] The rotor core 12 is cylindrical. The rotor core 12 is provided with a plurality of through holes 14. The plurality of through holes 14 are arranged at equal intervals in the circumferential direction of the rotor core 12. Each through hole 14 penetrates the rotor core 12 in the axial direction.

[0044] The magnet 13 is inserted into the through hole 14. The magnet 13 is embedded in the rotor core 12. When viewed from the axial direction of the rotor core 12, the magnet 13 is rectangular. The longitudinal direction of the magnet 13 when viewed from the axial direction of the rotor core 12 is defined as the width direction of the magnet 13. The short direction of the magnet 13 when viewed from the axial direction of the rotor core 12 is defined as the thickness direction of the magnet 13.

[0045] As shown in Figure 2, the magnet 13 has a first surface 13a, a second surface 13b, a third surface 13c, and a fourth surface 13d. The first surface 13a and the second surface 13b are surfaces that run along the thickness direction of the magnet 13. The second surface 13b is the surface opposite to the first surface 13a. The third surface 13c and the fourth surface 13d are surfaces that run along the width direction of the magnet 13. The third surface 13c is located closer to the center C of the rotor core 12 than the fourth surface 13d.

[0046] The through hole 14 has a first hole portion 14a that extends in the width direction of the magnet 13. The dimensions of the first hole portion 14a in the width direction of the magnet 13 are greater than the dimensions of the magnet 13 in the width direction. The rotor core 12 has a first section screen 12a, a second section screen 12b, a third section screen 12c, and a fourth section screen 12d as surfaces that demarcate the first hole portion 14a. The first section screen 12a and the second section screen 12b face each other in the width direction of the magnet 13. The third section screen 12c and the fourth section screen 12d face each other in the thickness direction of the magnet 13. The third section screen 12c is located closer to the center C of the rotor core 12 than the fourth section screen 12d. A recess 12e is provided in the center of the third section screen 12c in the width direction of the magnet 13. The dimensions of the recess 12e in the width direction of the magnet 13 are approximately the same as the dimensions of the magnet 13 in the width direction. The magnet 13 is positioned between the inner bottom surface 120 of the recess 12e and the fourth section screen 12d.

[0047] The through hole 14 has two second holes 14b connected to the first hole 14a. The two second holes 14b extend from both ends of the first hole 14a in the width direction of the magnet 13 toward the outer circumferential surface of the rotor core 12. The second holes 14b extend in the thickness direction of the magnet 13.

[0048] As shown in Figure 1, the rotating electric machine 10 includes a stator 15. The stator 15 has a stator core 16 and a plurality of coils 17. The stator core 16 has a yoke 16a and a plurality of teeth 16b. The yoke 16a is cylindrical. The plurality of teeth 16b extend radially from the inner circumferential surface of the yoke 16a. The plurality of teeth 16b are arranged at equal intervals in the circumferential direction of the yoke 16a. The coils 17 are wound around the teeth 16b. The stator 15 is located on the outer circumference of the rotor 11.

[0049] <Design System 20> As shown in Figure 3, the design system 20 comprises a design device 21, an input device 22, and a display device 23.

[0050] The design device 21 comprises a processor 24 and a storage unit 25. The processor 24 may include, for example, a CPU (Central Processing Unit) and a DSP (Digital Signal Processor). The storage unit 25 includes RAM (Random access memory), ROM (Read Only Memory), and rewritable non-volatile memory. Examples of non-volatile memory include EEPROM (Electrically Erasable Programmable Read-Only Memory) and flash memory. The storage unit 25 stores program code or instructions configured to cause the processor 24 to execute processing. The storage unit 25, i.e., the computer-readable medium, includes any available medium accessible by a general-purpose or dedicated computer. The design device 21 may also be composed of hardware circuits such as an ASIC (Application Specific Integrated Circuit) or an FPGA (Field Programmable Gate Array). The design device 21, which is a processing circuit, may include one or more processors operating according to a computer program, one or more hardware circuits such as an ASIC or FPGA, or a combination thereof.

[0051] The input device 22 and the display device 23 are connected to the design device 21. The input device 22 is, for example, a keyboard and a mouse. The display device 23 is, for example, a monitor. As shown in Figure 4, the memory unit 25 stores the database DB, the behavioral learning model M1, the value learning model M2, and the separate behavioral learning model M3.

[0052] <database DB> The database (DB) contains multiple design datasets. Each design dataset contains one or more design requirements, multiple design variables, and information for calculating evaluation values.

[0053] Design requirements are the conditions and specifications necessary for the design. In this embodiment, the design requirements are the output power of the rotating electric machine 10 and the input voltage of the rotating electric machine 10. In the following description, multiple design requirements may be collectively referred to as design requirement Θ. Design requirement Θ is a vector whose components are multiple design requirements.

[0054] Design variables are variables that define the structure of the object being designed. Design variables are the objects to be optimized in the design. In this embodiment, the design variables are r m , w m h m There are seven design variables: w1, h1, w2, and h2. In the following explanation, multiple design variables may be collectively referred to as design variable χ. Design variable χ is a vector whose components are multiple design variables.

[0055] As shown in Figure 2, r m w is the distance from the center C of the rotor core 12 to the third surface 13c of the magnet 13. m h is the width dimension of the magnet 13. m This is the dimension in the thickness direction of the magnet 13.

[0056] w1 is the distance from the first surface 13a of the magnet 13 to the first divided surface 12a of the rotor core 12 (the distance from the second surface 13b of the magnet 13 to the second divided surface 12b of the rotor core 12). h1 is the distance from the third divided surface 12c to the fourth divided surface 12d. w2 is the dimension of the second hole portion 14b in the width direction of the magnet 13. h2 is the dimension of the second hole portion 14b in the thickness direction of the magnet 13.

[0057] The design variable χ is grouped into a plurality of groups by a user of the design apparatus 21. The design variables χ are grouped based on the correlation between the design variables. Design variables with high correlation are set in the same group. The order of the plurality of groups is set based on priority in design. The order of the groups is set so that groups with higher priority are arranged in sequence. In the following description, the design variable of the n-th group is referred to as the n-th design variable x n and n is an integer of 1 or greater.

[0058] In the present embodiment, r m , w m , and h m are design variables related to the position or shape of the magnet 13, and therefore are grouped into the same group. w1, h1, w2, and h2 are design variables related to the shape of the rotor core 12, more specifically, design variables related to the shape of the through hole 14 provided in the rotor core 12, and therefore are grouped into the same group. An appropriate shape of the rotor core 12 depends on the position and shape of the magnet 13. For this reason, design variables related to the position and shape of the magnet 13 have higher priority than design variables related to the shape of the rotor core 12. Therefore, r m , w m , and h m are grouped into the first group. w1, h1, w2, and h2 are grouped into the second group.

[0059] The information used to calculate the evaluation value f(χ,Θ) for the design variable χ is the information used when calculating the evaluation value f(χ,Θ). This information includes, for example, iron loss and copper loss. The evaluation value f(χ,Θ) is calculated based on the design objective F(χ,Θ) and the design constraint G(χ,Θ).

[0060] The design objective F(χ,Θ) is the value of an index that is improved by optimizing the structure of the object being designed. For example, the design objective F(χ,Θ) is the efficiency of the rotating electric machine 10. When the design objective F(χ,Θ) is the efficiency of the rotating electric machine 10, a larger value of the design objective F(χ,Θ) is preferable.

[0061] A design constraint G(χ,Θ) is a condition that must be satisfied when optimizing the structure being designed. For example, a design constraint G(χ,Θ) might be the size of the rotating electric machine 10. The number of design constraints G(χ,Θ) is not limited to one; there may be multiple constraints.

[0062] The evaluation value f(χ,Θ) is a function in which the evaluation increases as the design objective F(χ,Θ) is improved. The evaluation value f(χ,Θ) is a function in which the evaluation decreases as the number of violations of the design constraints G(χ,Θ) increases.

[0063] The database DB pre-stores existing design datasets. These existing design datasets are, for example, design datasets generated by a designer or a design device other than the design device 21. As will be described in detail later, the database DB will also have design datasets added that include design variables output by the behavioral learning model M1, which will be described later. The database DB will accumulate design datasets.

[0064] <Behavioral Learning Model M1> The behavioral learning model M1 is trained to output the design variables of a given group when it receives design information set for each group. In the following explanation, the design information of the nth group is used for the nth design information s nIt is sometimes written as follows: The behavioral learning model M1 is, for example, a deep neural network. The behavioral learning model M1 is initially trained using an existing design dataset in the database DB.

[0065] The behavioral learning model M1 receives design information in order, starting with the first group. The behavioral learning model M1 outputs design variables in order, starting with the first group. The first design information s1 includes the design requirement Θ. The nth design information s n This is the (n-1)th design variable x output by the behavioral learning model M1. n-1 This includes at least the nth design information s. n This includes the design requirement Θ and the first to (n-1) design variables x1 to x output by the behavioral learning model M1. n-1 That is the case.

[0066] The behavioral learning model M1 of this embodiment is composed of the same number of segmented behavioral learning models as the number of groups. The nth segmented behavioral learning model contains nth design information s n The following is input. The nth split action learning model is determined by the nth design variable x n Outputs.

[0067] As shown in Figure 4, the behavioral learning model M1 of this embodiment is composed of two segmented behavioral learning models M11 and M12. As shown in Figure 5, the first segmented action learning model M11 outputs a first design variable x1 when it receives the first design information s1, which is the design requirement Θ.

[0068] As shown in Figure 6, when the second segmented action learning model M12 receives the design requirement Θ, which is the second design information s2, and the first design variable x1 output by the first segmented action learning model M11, it outputs the second design variable x2.

[0069] <Value Learning Model M2> The value learning model M2 is trained to output the value of a pair of design variables when it receives design information input to the behavior learning model M1 and a pair of design variables output by the behavior learning model M1 in relation to that design information. In the following explanation, the nth design information s input to the behavior learning model M1 is n and the nth design variable x output by the behavioral learning model M1 n pairs {s n ,x n The value of} is the nth value q n It is sometimes written that the value learning model M2 is initially trained using an existing design dataset in the database DB.

[0070] The value learning model M2 of this embodiment is composed of the same number of partitioned value learning models as the number of groups. The nth partitioned value learning model contains the nth design information s n and the nth design variable x output by the nth split action learning model n pairs {s n ,x n} is input. The nth partitioned value learning model is the nth value q n Outputs.

[0071] As shown in Figure 4, the value learning model M2 of this embodiment is composed of two partitioned value learning models M21 and M22. As shown in Figure 7, the first partitioned value learning model M21 is input with a pair {s1,x1} consisting of the first design information s1 and the first design variable x1 output by the first partitioned action learning model M11. The first design information s1 is the design requirement Θ. The first partitioned value learning model M21 outputs the first value q1, which is the value of the pair {s1,x1}.

[0072] As shown in Figure 8, the second partitioned value learning model M22 is input with the pair {s2,x2} of the second design information s2 and the second design variable x2 output by the second partitioned action learning model M12. The second design information s2 is the design requirement Θ and the first design variable x1 output by the first partitioned action learning model M11. The second partitioned value learning model M22 outputs the second value q2, which is the value of the pair {s2,x2}.

[0073] <Separate Behavioral Learning Model M3> The separate behavioral learning model M3 is trained to output the design variable from the database DB that corresponds to the input design information and has the highest evaluation value f(χ,Θ) when design information set for each group is input. The separate behavioral learning model M3 is, for example, a deep neural network. The separate behavioral learning model M3 is initially trained using an existing design dataset in the database DB.

[0074] The separate behavioral learning model M3 receives design information sequentially, starting with the first group. The separate behavioral learning model M3 outputs design variables sequentially, starting with the first group. The first design information s1 includes the design requirement Θ. The nth design information s n This is the (n-1)th design variable x output by the behavioral learning model M1. n-1 This includes at least the nth design information s. n This includes the design requirement Θ and the first to (n-1) design variables x1 to x output by the behavioral learning model M1. n-1 That is the case.

[0075] The separate behavioral learning model M3 is the same as the behavioral learning model M1 in that it receives design information set for each group. The separate behavioral learning model M3 differs from the behavioral learning model M1 in that the design variables output are the design variables in the database DB that correspond to the input design information and have the highest evaluation value f(χ,Θ).

[0076] The separate behavioral learning model M3 of this embodiment is composed of the same number of partitioned separate behavioral learning models as the number of groups. The nth partitioned separate behavioral learning model contains the nth design information s n The following is input. The nth segmented separate behavioral learning model uses the nth design variable x in the database DB. n Among them, the nth design variable x has the highest evaluation value f(χ,Θ). n Outputs.

[0077] As shown in Figure 4, the separate behavioral learning model M3 of this embodiment is composed of two separate behavioral learning models M31 and M32. As shown in Figure 9, the first segmented behavioral learning model M31 is input with the design requirement Θ, which is the first design information s1. The first segmented behavioral learning model M31 outputs the first design variable x1 in the database DB that has the highest evaluation value f(χ,Θ).

[0078] As shown in Figure 10, the second segmented behavioral learning model M32 receives the design requirement Θ, which is the second design information s2, and the first design variable x1 output by the first segmented behavioral learning model M31 as input. The second segmented behavioral learning model M32 outputs the second design variable x2 with the highest evaluation value f(χ,Θ) among the second design variables x2 in the database DB.

[0079] As shown in Figure 2, the design device 21 includes a model update unit 21a and an output unit 21b. The model update unit 21a and the output unit 21b are functional elements that operate when the processor 24 executes a program stored in the memory unit 25. The model update unit 21a updates the behavioral learning model M1, the value learning model M2, and the separate behavioral learning model M3. The output unit 21b outputs multiple design variables according to the design requirements.

[0080] <Design method> This section describes a design method for designing the object to be designed using the design device 21. The design device 21 is pre-configured with multiple design variables and information indicating groups of each design variable. When a design request is input to the design device 21, it executes the design process shown in Figure 11. The user of the design device 21 can input the design request to the design device 21, for example, using the input device 22.

[0081] As shown in Figure 11, the model update unit 21a executes the processes in steps S11 to S16. Steps S11 to S16 are update steps. The processes in steps S11 to S16 are update processes.

[0082] In step S11, the model update unit 21a inputs design information to the initially trained behavioral learning model M1, starting with the first group. The behavioral learning model M1 outputs the design variables, starting with the first group.

[0083] In this embodiment, the model update unit 21a inputs the first design information s1 to the first segmented action learning model M11. The first design information s1 consists of multiple design requests input to the design device 21. The first segmented action learning model M11 outputs the first design variable x1. That is, the first segmented action learning model M11 outputs r m , w m , and h m Outputs.

[0084] Next, the model update unit 21a inputs the second design information s2 to the second segmented action learning model M12. The second design information s2 consists of multiple design requests input to the design device 21 and the first design variable x1 output by the first segmented action learning model M11. The second segmented action learning model M12 outputs the second design variable x2. That is, the second segmented action learning model M12 outputs w1, h1, w2, and h2.

[0085] Next, in step S12, the model update unit 21a performs an improvement process to improve the multiple design variables output by the behavioral learning model M1 in step S11 so that the evaluation value f(χ,Θ) becomes higher. The model update unit 21a improves the multiple design variables by applying an optimization algorithm such as evolutionary computation.

[0086] In this embodiment, the model update unit 21a performs the improvement process on the nth design variable x n The range of change of the (n+1)th design variable x n+1 The range of change is narrowed. Specifically, the model update unit 21a narrows the range of change of the first design variable x1 compared to the range of change of the second design variable x2. If there are three groups, the model update unit 21a, in the improvement process, for example, narrows the range of change of the first and second design variables x1 and x2 compared to the range of change of the third design variable x3.

[0087] Next, in step S13, the model update unit 21a updates the database DB by adding the design dataset to the database DB. In this embodiment, the model update unit 21a adds a design dataset to the database DB that includes the multiple design requirements input to the behavioral learning model M1 in step S11 and the multiple design variables that have been improved in step S12. As described above, the multiple design variables that are improved in step S12 are the multiple design variables output by the behavioral learning model M1 in step S11. Therefore, it can be said that the model update unit 21a adds a design dataset to the database DB that includes the multiple design variables output by the behavioral learning model M1 in step S11.

[0088] Next, in step S14, the model update unit 21a updates the separate behavioral learning model M3 using the database DB updated in step S13. That is, the model update unit 21a updates the separate behavioral learning model M3 using the database DB to which a design dataset containing the design requirements input to the behavioral learning model M1 and multiple design variables output by the behavioral learning model M1 has been added.

[0089] Next, in step S15, the model update unit 21a updates the value learning model M2 using the database DB updated in step S13. In this embodiment, the model update unit 21a updates the value learning model M2 by Q-training the value learning model M2. Specifically, the model update unit 21a calculates the Q-training teacher signal. The model update unit 21a Q-trains the value learning model M2 using the calculated Q-training teacher signal. Q-training teacher signal t of the nth partitioned value learning model n This is expressed by equation (1).

[0090]

number

[0091] In this embodiment, the model update unit 21a uses the design variables output by the separate behavioral learning model M3 updated in step S14 to generate the Q-learning teacher signal t n This is calculated. As mentioned above, the design variables output by the separate behavioral learning model M3 are the design variables in the database DB that correspond to the input design information and have the highest evaluation value f(χ,Θ). Therefore, the (n+1)th value q n+1 The design variable a that maximizes the value is the (n+1)th design variable x output by the (n+1)th separate behavioral learning model M3. n+1 This can be substituted with the (n+1)th design variable x output by the separate behavioral learning model M3. n+1 to a n+1 We can set it as follows. Therefore, the training signal t for learning Q n This is expressed by equation (2).

[0092]

number

[0093]

number

[0094] If the update process has not been executed a predetermined number of times (NO in step S17), the process returns to step S11 and the update process is executed again. If the update process has been executed a predetermined number of times (YES in step S17), the process in step S18 is executed. The predetermined number of times is set in advance by the user of the design device 21. The model update unit 21a repeatedly executes the update process until the update process has been executed a predetermined number of times.

[0095] In step S11 of the second and subsequent update processes, the model update unit 21a inputs design information to the behavioral learning model M1, which was updated in step S16 of the previous update process, starting with the first group. The updated behavioral learning model M1 outputs design variables, starting with the first group.

[0096] In step S18, the output unit 21b outputs the final design variables using the behavioral learning model M1 updated in step S16. Specifically, the output unit 21b inputs design information to the updated behavioral learning model M1 in order from the first group. The updated behavioral learning model M1 outputs the design variables in order from the first group. The output unit 21b outputs the design variables output by the updated behavioral learning model M1 as the final design variables.

[0097] In this embodiment, the output unit 21b outputs the first design variable x1 output by the updated first segmented behavior learning model M11 and the second design variable x2 output by the updated second segmented behavior learning model M12 as the final design variables.

[0098] In this embodiment, the output unit 21b provides the final design variables to the user of the design device 21 by outputting the final design variables to the display device 23. [Operation and Effects of This Embodiment] The operation and effects of this embodiment will now be explained.

[0099] (1) The design device 21 comprises a storage unit 25, a model update unit 21a, and an output unit 21b. The output unit 21b outputs multiple design variables according to the design requirements. The multiple design variables are grouped into multiple groups.

[0100] The memory unit 25 stores the behavioral learning model M1, the value learning model M2, and the database DB. The behavioral learning model M1 is trained to output the design variables of a given group when it receives design information set for each group. The value learning model M2 is trained to output the value of a pair when it receives the design information input to the behavioral learning model M1 and the design variables output by the behavioral learning model M1 for that design information. The database DB stores design datasets containing design requirements and multiple design variables.

[0101] The behavioral learning model M1 receives design information as input, starting with the first group, and outputs design variables as input, starting with the first group. The first design information s1 includes design requirements. The nth design information s n This is the (n-1)th design variable x output by the behavioral learning model M1. n-1 It includes at least [this].

[0102] The model update unit 21a executes the update process. Specifically, the model update unit 21a updates the database DB by adding a design dataset containing the design requirements input to the behavioral learning model M1 and multiple design variables output by the behavioral learning model M1 to the database DB. The model update unit 21a updates the value learning model M2 using the updated database DB. The model update unit 21a updates the behavioral learning model M1 using the updated database DB so that the value output by the updated value learning model M2 is higher.

[0103] Multiple design variables are grouped into multiple groups. The behavioral learning model M1 receives design information as input, starting from the first group, and outputs design variables as input, starting from the first group. The nth design information input to the behavioral learning model M1 is s n This is the (n-1)th design variable x output by the behavioral learning model M1. n-1 This includes the following. Therefore, for multiple design variables, highly related design variables are set in the same group, and the order of the groups is set so that they are arranged in order from the group with the highest priority. As a result, even when there is a strong dependency between design variables, the design device 21 can output appropriate design variables.

[0104] The behavioral learning model M1 is updated to increase the value output by the value learning model M2. As a result, the behavioral learning model M1 can output more appropriate design variables. The database DB stores design datasets. The behavioral learning model M1 and the value learning model M2 are updated using the design datasets stored in the database DB. Therefore, if the design requirements input to the design device 21 are similar to the design requirements in the database DB, the design device 21 can immediately output multiple design variables. Thus, the design time can be shortened. This is particularly effective when there are many design requirements or when the design requirements are complex, as design time tends to be long in such cases. Thus, the design can be carried out optimally.

[0105] (2) The model update unit 21a repeatedly executes the update process. This allows the behavioral learning model M1 to output even more appropriate design variables. Furthermore, even if the design request input to the design device 21 is not similar to the design requests in the database DB, the model update unit 21a repeatedly executes the update process, allowing the design device 21 to output appropriate design variables. In addition, as described above, design datasets are accumulated in the database DB, and the behavioral learning model M1 and value learning model M2 are updated using the design datasets accumulated in the database DB, so the number of design requests that can be handled increases.

[0106] (3) The memory unit 25 stores the separate behavioral learning model M3. The separate behavioral learning model M3 is trained to output the design variable in the database DB that corresponds to the input design information and has the highest evaluation value f(χ,Θ) when design information is input.

[0107] The model update unit 21a updates the separate behavioral learning model M3 using the updated database DB. The model update unit 21a calculates the Q-learning teacher signal using the design variables output by the updated separate behavioral learning model M3. The model update unit 21a updates the value learning model M2 by performing Q-learning on the value learning model M2 using the calculated Q-learning teacher signal.

[0108] The nth value q nIf we Q-train the nth partitioned value learning model to output the following, then the training signal t for Q-training n When calculating the n+1th value q, n+1 We need to find the design variables that maximize [the desired outcome].

[0109] In typical reinforcement learning, the agent generates diverse data through trial-and-error actions, thereby advancing the learning of the behavioral learning model M1 and the value learning model M2. This allows the behavioral learning model M1 and the value learning model M2 to learn appropriate outputs. However, in the field of design, it can be difficult to generate diverse data because the time required to calculate the evaluation value f(χ,Θ) tends to be long. As a result, the learning of the behavioral learning model M1 becomes insufficient, leading to the (n+1)th value q n+1 There is a risk that the design variables that result in the highest value may not be properly determined. In this case, the training signal t for Q learning n It is not possible to calculate this correctly.

[0110] In contrast, in this embodiment, the design variables output by the separate behavioral learning model M3 are used to create the Q-learning teacher signal t n The following is calculated. The design variables output by the separate behavioral learning model M3 are the design variables in the database DB that correspond to the input design information and have the highest evaluation value f(χ,Θ). Therefore, the (n+1)th value q n+1 The design variable a that maximizes the value is the (n+1)th design variable x output by the separate behavioral learning model M3. n+1 This can be substituted. Therefore, the value learning model M2 can be updated without generating diverse data.

[0111] (4) The model update unit 21a performs an improvement process to improve the design variables output by the behavioral learning model M1 so that the evaluation value f(χ,Θ) is higher. The model update unit 21a adds the design dataset containing the improved design variables to the database DB.

[0112] In this configuration, the behavioral learning model M1, the value learning model M2, and the separate behavioral learning model M3 are updated using a database DB containing improved design variables. Therefore, the behavioral learning model M1, the value learning model M2, and the separate behavioral learning model M3 can be updated efficiently.

[0113] (5) The model update unit 21a processes the nth design variable x n The range of change of the (n+1)th design variable x n+1 This narrows the range of change of the nth design variable x. n and the (n+1)th design variable x n+1 This can alleviate the dependence on it.

[0114] (6) The design apparatus 21 and design method of this embodiment are used to design a rotating electric machine 10 having a rotor core 12 and magnets 13 provided on the rotor core 12. The first group is a design variable r relating to the position and shape of the magnets 13. m , w m , and h m The second group includes w1, h1, w2, and h2, which are design variables relating to the shape of the rotor core 12.

[0115] Since the appropriate shape of the rotor core 12 depends on the position and shape of the magnet 13, the design variables related to the position and shape of the magnet 13 have higher priority than the design variables related to the shape of the rotor core 12. In this embodiment, the design variable r related to the position and shape of the magnet 13 is m , w m , and h m The first group is grouped as follows. The design variables w1, h1, w2, and h2, which relate to the shape of the rotor core 12, are grouped as follows. Therefore, the rotating electric machine 10 can be designed taking into account the priority of the design variables.

[0116] [Example of changes] The above embodiment can be implemented with the following modifications. The above embodiment and the following modifications can be combined with each other to the extent that they do not contradict each other technically.

[0117] ○ In the above embodiment, the design apparatus 21 and design method were used to design an IPM type rotating electric machine 10 in which the magnets 13 are embedded in the rotor core 12, but they may also be used to design an SPM type rotating electric machine in which the magnets 13 are arranged on the surface of the rotor core 12. In this case, the design variables relating to the shape of the rotor core 12 are the design variables relating to the shape of the outer circumferential surface of the rotor core 12.

[0118] ○ In the above embodiment, the design apparatus 21 and design method were used to design an inner rotor type rotating electric machine 10 in which the rotor 11 is arranged on the inner circumference of the stator 15, but they may also be used to design an outer rotor type rotating electric machine in which the rotor 11 is arranged on the outer circumference of the stator 15.

[0119] ○ The design apparatus 21 and design method may be used for the design of the stator 15 of the rotating electric machine 10. ○ The design apparatus 21 and design method may also be used for the design of products other than the rotating electric machine 10.

[0120] ○ The number and types of design requirements are not limited to the number and types of the embodiments described above. The number of design requirements may be one or three or more. ○ The number and types of design variables are not limited to those of the above embodiment. The number of design variables only needs to be two or more.

[0121] The design variables are, for example, r m , w m h m , w1, h1, w2, h2, r in , and h in It could also be 9. in h is the distance from the center C of the rotor core 12 to the outer surface of the rotor core 12 (see Figure 1). in This is the axial dimension of the rotor core 12. In this case, the nine design variables are grouped as follows:

[0122] As mentioned above, r m , w m , and h mare design variables related to the position and shape of the magnet 13, and are therefore grouped into the same group. w1, h1, w2, and h2 are design variables related to the shape of the rotor core 12, and are therefore grouped into the same group. r in and h in are design variables related to the size of the rotor core 12, and are therefore grouped into the same group.

[0123] The size of the rotor core 12 determines the size of the rotating electrical machine 10. Therefore, design variables related to the size of the rotor core 12 have higher priority than design variables related to the position and shape of the magnet 13 and design variables related to the shape of the rotor core 12. As described above, an appropriate shape of the rotor core 12 depends on the position and shape of the magnet 13. Therefore, design variables related to the position and shape of the magnet 13 have higher priority than design variables related to the shape of the rotor core 12.

[0124] Accordingly, r m , w m , and h m are grouped into a first group. w1, h1, w2, and h2 are grouped into a second group. r in and h in are grouped into a third group. This allows the rotating electrical machine 10 to be designed in consideration of the priority of design variables.

[0125] The design variables are not limited to dimensions such as r m , w m , h m , w1, h1, w2, h2, r in , and h in . The design variables may be parameters of a parametric method used in topology optimization. For example, the arrangement of the magnets 13 in the rotor core 12 is not limited to the arrangement of the above embodiment, and a plurality of variations are conceivable. The design variables may be parameters that dynamically change the arrangement of the magnets 13 in the rotor core 12.

[0126] ○ In the above embodiment, n-th design information s nThis includes the design requirement Θ and the first to (n-1) design variables x1 to x output by the behavioral learning model M1. n-1 This included, but is not limited to, the nth design information s. n This is the (n-1)th design variable x output by the behavioral learning model M1. n-1 It is sufficient if it contains at least that.

[0127] For example, information s for the nth design n This includes the design requirement Θ and the (n-1)th design variable x output by the behavioral learning model M1. n-1 But that's fine. For example, information s for the nth design n This represents the first to (n-1) design variables x1 to x output by the behavioral learning model M1. n-1 But that's fine.

[0128] For example, information s for the nth design n This is the (n-1)th design variable x output by the behavioral learning model M1. n-1 But that's fine. ○ In the above embodiment, the update process was terminated when it was executed a predetermined number of times, but it is not limited to this. The model update unit 21a may terminate the update process, for example, when the evaluation value f(χ,Θ) of the design variable χ output by the updated behavioral learning model M1 converges to a predetermined range.

[0129] ○ In the above embodiment, the output unit 21b outputs the design variables output by the updated behavioral learning model M1 as the final design variables, but it may also output the design variables output by the updated separate behavioral learning model M3 as the final design variables.

[0130] ○ The model update unit 21a generates diverse design datasets, and if it is possible to learn, the (n+1)th value q n+1 The design variable a that maximizes the value is the (n+1)th design variable x output by the separate behavioral learning model M3. n+1 It is not necessary to substitute it. In this case, the memory unit 25 does not need to store the separate behavioral learning model M3.

[0131] ○ The model update unit 21a does not need to perform improvement processing. In this case, the model update unit 21a adds the design dataset containing the raw design variables output by the behavioral learning model M1 to the database DB.

[0132] ○ The model update unit 21a processes the nth design variable x during the improvement process. n The range of change of the (n+1)th design variable x n+1 It is not necessary to narrow the range of change below the specified range. ○ In the above embodiment, the behavioral learning model M1 was composed of the same number of segmented behavioral learning models as the number of groups, but is not limited to this.

[0133] As shown in Figure 12, the behavioral learning model M1 may be an integrated behavioral learning model M10 configured to allow input of design information for any group. The integrated behavioral learning model M10 has multiple input frames F1 for inputting design information. The number of input frames F1 is equal to the sum of the number of design requirements and the number of design variables in the first to n-1 groups. The integrated behavioral learning model M10 also has multiple output frames F2 for outputting design variables. The number of output frames F2 is equal to the number of design variables in the group with the largest number of design variables among the multiple groups.

[0134] Figure 12 shows the input and output frames F1 and F2 for the behavioral learning model when there are 2 design requirements, 3 design variables for the first design variable x1, 4 design variables for the second design variable x2, and 2 design variables for the third design variable x3. In this case, there are 9 input frames F1 and 4 output frames F2.

[0135] When distinguishing between the nine input frames F1 for the behavioral learning models, they are designated as input frames F1a to F1i for the 1st to 9th behavioral learning models. The input frames F1a and F1b for the 1st and 2nd behavioral learning models can accept the design requirement Θ. The input frames F1c to F1e for the 3rd to 5th behavioral learning models can accept the first design variable x1. The input frames F1f to F1i for the 6th to 9th behavioral learning models can accept the second design variable x2. The input frames F1a to F1i for the 1st to 9th behavioral learning models are arranged in this order.

[0136] When distinguishing between the four behavioral learning model output frames F2, they are designated as F2a for the first behavioral learning model, F2b for the second behavioral learning model, F2c for the third behavioral learning model, and F2d for the fourth behavioral learning model. The output frames F2a, F2b, F2c, and F2d for the first, second, third, and fourth behavioral learning models are arranged in this order.

[0137] nth design information s n When the input is given to the integrated behavioral learning model M10, null or 0 is entered into the behavioral learning model input frame F1 corresponding to the design variable of at least the (n+1)th group among the multiple behavioral learning model input frames F1. The integrated behavioral learning model M10 then outputs the nth design variable x in order from the first behavioral learning model output frame F2a. n Outputs.

[0138] In the example shown in Figure 12, when the design requirement Θ, which is the first design information s1, is input to the integrated behavioral learning model M10, the design requirement Θ is input to the input frames F1a and F1b for the first and second behavioral learning models. Null or 0 is input to the input frames F1c to F1i for the third to ninth behavioral learning models. In this case, the integrated behavioral learning model M10 outputs the first design variable x1 to the output frames F2a to F2c for the first to third behavioral learning models, and outputs null or 0 to the output frame F2d for the fourth behavioral learning model.

[0139] When the design requirement Θ and the first design variable x1 output by the integrated behavioral learning model M10 are input to the integrated behavioral learning model M10 as the second design information s2, the design requirement Θ is input to the input frames F1a and F1b for the first and second behavioral learning models. The first design variable x1 is input to the input frames F1c to F1e for the third to fifth behavioral learning models. Null or 0 is input to the input frames F1f to F1i for the sixth to ninth behavioral learning models. In this case, the integrated behavioral learning model M10 outputs the second design variable x2 to the output frames F2a to F2d for the first to fourth behavioral learning models.

[0140] The third design information s3 includes the design requirement Θ and the first design variable x output by the integrated behavioral learning model M10. 1、 When the second design variable x2 output by the integrated behavioral learning model M10 is input to the integrated behavioral learning model M10, the design requirement Θ is input to the input frames F1a and F1b for the first and second behavioral learning models. The first design variable x1 is input to the input frames F1c to F1e for the third to fifth behavioral learning models. The second design variable x2 is input to the input frames F1f to F1i for the sixth to ninth behavioral learning models. In this case, the integrated behavioral learning model M10 outputs the third design variable x3 to the output frames F2a and F2b for the first and second behavioral learning models, and outputs null or zero to the output frames F2c and F2d for the third and fourth behavioral learning models.

[0141] With this configuration, the integrated behavioral learning model M10 can be reused by multiple groups, allowing all group design variables to be output using a single integrated behavioral learning model M10. Therefore, compared to the case where the behavioral learning model M1 is composed of the same number of segmented behavioral learning models as the number of groups, the amount of memory used by the memory unit 25 can be reduced and implementation can be simplified.

[0142] As described above, the behavioral learning model M1 outputs the design variables in order from the first group. Therefore, when the behavioral learning model M1 outputs the nth design variable x n When outputting, the design variables for the (n+1)th group and beyond are not output. According to the above configuration, the nth design information s nWhen the nth design variable x is input to the integrated behavioral learning model M10, the input box F1 for the behavioral learning model corresponding to the (n+1)th and subsequent groups of design variables will be entered as null or 0. Therefore, the integrated behavioral learning model M10 will input the nth design variable x n When outputting the results, it is possible to suppress the influence of the values ​​entered into the input frame F1 for the behavioral learning model corresponding to the design variables of the (n+1)th group and beyond.

[0143] For example, if the third design information s3 is the design requirement Θ and the second design variable x2 output by the integrated behavioral learning model M10, then null or 0 will be entered into the input frames F1c to F1e for the third to fifth behavioral learning models.

[0144] For example, if the third design information s3 is the first design variable x1 and the second design variable x2 output by the integrated behavioral learning model M10, then null or 0 will be input to the input frames F1a and F1b for the first and second behavioral learning models.

[0145] For example, if the third design information s3 is the second design variable x2 output by the integrated behavioral learning model M10, then null or 0 will be entered into the input frames F1a to F1e for the first to fifth behavioral learning models.

[0146] As shown in Figure 13, the design information input to the integrated behavioral learning model M10 may be assigned a behavioral learning model label L1 indicating which group the design information belongs to. In this case, the integrated behavioral learning model M10 can more easily recognize which group the input design information belongs to.

[0147] ○ In the above embodiment, the value learning model M2 was composed of the same number of partitioned value learning models as the number of groups, but is not limited to this. As shown in Figure 14, the value learning model M2 may be an integrated value learning model M20 configured to accept any pair of design information and design variables for any group.

[0148] The integrated value learning model M20 has multiple value learning model input frames F3 for inputting pairs of design information and design variables. The number of value learning model input frames F3 is equal to the sum of the number of design requirements and the total number of design variables. The integrated value learning model M20 also has a value learning model output frame F4 for outputting value. There is one value learning model output frame F4.

[0149] Figure 14 shows the input and output frames F3 and F4 for the value learning model when there are 2 design requirements, 3 design variables for the first design variable x1, 4 design variables for the second design variable x2, and 2 design variables for the third design variable x3. In this case, there are 11 input frames F3 for the value learning model.

[0150] When distinguishing between the 11 value learning model input frames F3, they are designated as the 1st to 11th value learning model input frames F3a to F3k. The 1st and 2nd value learning model input frames F3a and F3b can accept the design requirement Θ. The 3rd to 5th value learning model input frames F3c to F3e can accept the first design variable x1. The 6th to 9th value learning model input frames F3f to F3i can accept the second design variable x2. The 10th and 11th value learning model input frames F3j and F3k can accept the third design variable x3. The 1st to 11th value learning model input frames F3a to F3k are arranged in this order.

[0151] The nth group pair {s n ,x n When} is input to the integrated value learning model M20, null or 0 is input to the value learning model input frame F3 corresponding to the design variables of at least the (n+1)th group among the multiple value learning model input frames F3. The integrated value learning model M20 outputs the nth value q to the value learning model output frame F4. n Outputs.

[0152] In the example shown in Figure 14, consider the case where the first group pair {s1, x1} is input to the integrated value learning model M20. The first design information s1 is the design requirement Θ. In this case, the design requirement Θ is input to the input frames F3a and F3b for the first and second value learning models. The first design variable x1 is input to the input frames F3c to F3e for the third to fifth value learning models. Null or 0 is input to the input frames F3f to F3k for the sixth to eleventh value learning models. The integrated value learning model M20 outputs the first value q1 to the output frame F4 for the value learning model.

[0153] Consider the case where the second group pair {s2, x2} is input to the integrated value learning model M20. The second design information s2 is the design requirement Θ and the first design variable x1 output by the integrated behavior learning model M10. In this case, the design requirement Θ is input to the input frames F3a and F3b for the first and second value learning models. The first design variable x1 is input to the input frames F3c to F3e for the third to fifth value learning models. The second design variable x2 is input to the input frames F3f to F3i for the sixth to ninth value learning models. Null or 0 is input to the input frames F3j and F3k for the tenth and eleventh value learning models. The integrated value learning model M20 outputs the second value q2 to the value learning model output frame F4.

[0154] Let's consider the case where the third group pair {s3, x3} is input to the integrated value learning model M20. The third design information s3 consists of the design requirement Θ and the first and second design variables x1 and x2 output by the integrated behavior learning model M10. In this case, the design requirement Θ is input to the input frames F3a and F3b for the first and second value learning models. The first design variable x1 is input to the input frames F3c and F3e for the third to fifth value learning models. The second design variable x2 is input to the input frames F3f and F3i for the sixth to ninth value learning models. The third design variable x3 is input to the input frames F3j and F3k for the tenth and eleventh value learning models. The integrated value learning model M20 outputs the third value q3 to the value learning model output frame F4.

[0155] With this configuration, the integrated value learning model M20 can be reused across multiple groups, allowing a single integrated value learning model M20 to output the values ​​for all groups. Therefore, compared to the case where the value learning model M2 is composed of the same number of partitioned value learning models as the number of groups, the amount of memory used in the memory unit 25 can be reduced and the implementation can be simplified.

[0156] Also, the pair of the nth group {s n ,x n When} is input to the integrated value learning model M20, the input box F3 for the value learning model corresponding to the design variables of the (n+1)th group and beyond will be entered as null or 0. Therefore, the integrated value learning model M20 will be the nth value q n When outputting the results, it is possible to suppress the influence of the values ​​entered into the value learning model input frame F3 corresponding to the design variables of the (n+1)th group and beyond.

[0157] As shown in Figure 15, pairs of design information and design variables input to the integrated value learning model M20 may be assigned a value learning model label L2 indicating which group the pair belongs to. In this case, the integrated value learning model M20 can more easily recognize which group the input pair belongs to.

[0158] ○ In the above embodiment, the separate behavioral learning model M3 was composed of the same number of segmented separate behavioral learning models as the number of groups, but is not limited to this. As shown in Figure 16, the separate behavioral learning model M3 may be an integrated separate behavioral learning model M30 configured to allow input of design information for any group.

[0159] The integrated separate behavioral learning model M30 has multiple separate behavioral learning model input frames F5 configured for inputting design information. The number of separate behavioral learning model input frames F5 is equal to the sum of the number of design requirements and the number of design variables in the 1st to n-1th groups. The integrated separate behavioral learning model M30 also has multiple separate behavioral learning model output frames F6 configured for outputting the design variable in the database DB that corresponds to the input design information and has the highest evaluation value f(χ,Θ). The number of separate behavioral learning model output frames F6 is equal to the number of design variables in the group with the largest number of design variables among the multiple groups.

[0160] Figure 16 shows the input frame F5 and output frame F6 for the separate behavioral learning model when there are 2 design requirements, 3 design variables constituting the first design variable x1, 4 design variables constituting the second design variable x2, and 2 design variables constituting the third design variable x3. In this case, there are 9 input frames F5 and 4 output frames F6.

[0161] When distinguishing between the nine separate behavioral learning model input frames F5, they are designated as the 1st to 9th separate behavioral learning model input frames F5a to F5i. The 1st and 2nd separate behavioral learning model input frames F5a and F5b can accept the design requirement Θ. The 3rd to 5th separate behavioral learning model input frames F5c to F5e can accept the first design variable x1. The 6th to 9th separate behavioral learning model input frames F5f to F5i can accept the second design variable x2. The 1st to 9th separate behavioral learning model input frames F5a to F5i are arranged in this order.

[0162] When distinguishing between the four separate behavioral learning model output frames F6, they are designated as the first separate behavioral learning model output frame F6a, the second separate behavioral learning model output frame F6b, the third separate behavioral learning model output frame F6c, and the fourth separate behavioral learning model output frame F6d. The first separate behavioral learning model output frame F6a, the second separate behavioral learning model output frame F6b, the third separate behavioral learning model output frame F6c, and the fourth separate behavioral learning model output frame F6d are arranged in this order.

[0163] nth design information s n When the input is given to the integrated separate behavioral learning model M30, null or 0 is entered into the separate behavioral learning model input frame F5 corresponding to the design variable of at least the (n+1)th group among the multiple separate behavioral learning model input frames F5. The integrated separate behavioral learning model M30 then outputs the nth design variable x in order from the first separate behavioral learning model output frame F6a. n Outputs.

[0164] With this configuration, the integrated separate behavioral learning model M30 can be reused by multiple groups, allowing all groups' design variables to be output with a single integrated separate behavioral learning model M30. Therefore, compared to the case where the separate behavioral learning model M3 is composed of the same number of partitioned separate behavioral learning models as the number of groups, the amount of memory used by the memory unit 25 can be reduced and implementation can be simplified.

[0165] Also, information for the nth design s n When the nth design variable x is input to the integrated separate behavioral learning model M30, null or 0 is entered into the input box F5 for the separate behavioral learning model corresponding to the (n+1)th and subsequent groups of design variables. Therefore, the integrated separate behavioral learning model M30 is input to the nth design variable x n When outputting the results, it is possible to suppress the influence of the values ​​entered into the separate input box F5 for the behavioral learning model corresponding to the design variables of the (n+1)th group and beyond.

[0166] As shown in Figure 17, the design information input to the integrated separate behavioral learning model M30 may be assigned a separate behavioral learning model label L3 indicating which group the design information belongs to. In this case, the integrated separate behavioral learning model M30 can more easily recognize which group the input design information belongs to.

[0167] [Note] The technical concepts that can be understood from the above embodiments and modified examples are described below. <Note 1> A design device comprising an output unit that outputs multiple design variables in accordance with design requirements, wherein the multiple design variables are grouped into multiple groups, and when design information set for each group is input, an action learning model trained to output the design variables of the input group; a value learning model trained to output the value of a pair when a pair of the design information input to the action learning model and the design variables output by the action learning model for said design information is input to the action learning model; a database in which a design dataset including the design requirements and the multiple design variables is stored; a model update unit that updates the action learning model and the value learning model, wherein the action learning model updates the first group in order A design device that receives the design information and outputs the design variables in order from the first group, wherein the design information of the first group includes the design request, and the design information of the nth group includes at least the design variables of the (n-1)th group output by the behavioral learning model, and the model update unit updates the database by adding the design dataset, which includes the design request input to the behavioral learning model and a plurality of design variables output by the behavioral learning model, to the database as an update process, updates the value learning model using the updated database, and updates the behavioral learning model using the updated database so that the value output by the updated value learning model becomes higher.

[0168] <Note 2> The model update unit is the design apparatus described in Appendix 1, which repeatedly performs the update process. <Note 3> The design apparatus according to Appendix 1 or Appendix 2, wherein the memory unit stores a separate behavioral learning model that, when the design information is input, outputs the design variable in the database that corresponds to the input design information and has the highest evaluation value, and the model update unit updates the separate behavioral learning model using the updated database, calculates a Q-learning teacher signal using the design variable output by the updated separate behavioral learning model, and updates the value learning model by Q-learning the value learning model using the calculated Q-learning teacher signal.

[0169] <Note 4> The design apparatus according to any one of the appendices 1 to 3, wherein the model update unit performs an improvement process to improve the design variables output by the behavioral learning model so that their evaluation values ​​are higher, and adds the design dataset containing the improved design variables to the database.

[0170] <Note 5> The design apparatus according to Appendix 4, wherein the model update unit, in the improvement process, makes the range of change of the design variable for the nth group narrower than the range of change of the design variable for the (n+1)th group.

[0171] <Note 6> The design apparatus described in any one of the appendices 1 to 5, wherein the behavioral learning model is composed of the same number of segmented behavioral learning models as the number of groups, the nth segmented behavioral learning model receives the design information of the nth group as input and outputs the design variables of the nth group, and the value learning model is composed of the same number of segmented value learning models as the number of groups, the nth segmented value learning model receives the pair of the nth group as input and outputs the value of the nth pair.

[0172] <Note 7> The design device described in any one of the appendices 1 to 5 is an integrated behavioral learning model configured to allow input of the design information for any group.

[0173] <Note 8> The design apparatus as described in Appendix 7, wherein the number of input frames for the behavioral learning model for inputting the design information into the integrated behavioral learning model is equal to the sum of the number of design requests and the number of design variables in the 1st to the (n-1th)th groups, the number of output frames for the behavioral learning model for the integrated behavioral learning model to output the design variables is equal to the number of design variables in the group with the largest number of design variables among the multiple groups, and when the design information of the nth group is input into the integrated behavioral learning model, null or 0 is input to at least the input frames for the behavioral learning model corresponding to the design variables of the (n+1th)th and subsequent groups among the multiple input frames for the behavioral learning model.

[0174] <Note 9> The design apparatus described in Appendix 8, wherein the design information input to the integrated behavioral learning model is assigned a behavioral learning model label indicating which group of the design information it belongs to.

[0175] <Note 10> The design apparatus described in any one of Appendices 1-5, 7-9, wherein the value learning model is an integrated value learning model configured to accept any pair of the aforementioned groups as input.

[0176] <Note 11> The design apparatus as described in Appendix 10, wherein the number of input frames for the value learning model to input the pair into the integrated value learning model is equal to the sum of the number of design requirements and the total number of design variables, the number of output frames for the value learning model to output the value of the pair is one, and when the nth group of the pair is input into the integrated value learning model, null or 0 is input to at least the value learning model input frames corresponding to the design variables of the (n+1)th group and beyond among the multiple value learning model input frames.

[0177] <Note 12> The design apparatus according to Appendix 11, wherein the pair input to the integrated value learning model is assigned a value learning model label indicating which group the pair belongs to.

[0178] <Note 13> The design apparatus described in Appendix 3 or Appendix 4 to 12, which is an integrated separate behavioral learning model configured to allow input of the design information for any group.

[0179] <Note 14> The design apparatus as described in Appendix 13, wherein the number of input frames for the separate behavioral learning model for inputting the design information into the integrated separate behavioral learning model is equal to the sum of the number of design requests and the number of design variables in the 1st to n-1th groups, and the number of output frames for the separate behavioral learning model for the integrated separate behavioral learning model to output the design variable in the database that corresponds to the input design information and has the highest evaluation value is equal to the number of design variables in the group with the largest number of design variables among the multiple groups, and when the design information of the nth group is input into the integrated separate behavioral learning model, null or 0 is input to at least the input frames for the separate behavioral learning model corresponding to the design variables of the n+1th group or later among the multiple input frames for the separate behavioral learning model.

[0180] <Note 15> The design apparatus described in Appendix 14, wherein the design information input to the integrated separate behavioral learning model is assigned a separate behavioral learning model label indicating which group of the design information it belongs to.

[0181] <Note 16> A design apparatus as described in any one of Appendix 1 to 15, used for designing a rotating electric machine having a rotor core and magnets provided on the rotor core, wherein the first group includes the design variables relating to the position and shape of the magnets, and the second group includes the design variables relating to the shape of the rotor core.

[0182] <Note 17> A design apparatus as described in any one of Appendix 1 to 15, used for designing a rotating electric machine having a rotor core and magnets provided on the rotor core, wherein the first group includes the design variables relating to the physique of the rotor core, the second group includes the design variables relating to the position and shape of the magnets, and the third group includes the design variables relating to the shape of the rotor core.

[0183] <Note 18> A design method in which a design device outputs multiple design variables in response to a design requirement, wherein the multiple design variables are grouped into multiple groups, and the design device includes a storage unit that stores: an action learning model that is trained to output the design variables of the input group when design information set for each group is input; a value learning model that is trained to output the value of a pair when a pair of the design information and the design variables output by the action learning model for the design information is input to the action learning model; and a database in which a design dataset containing the design requirement and the multiple design variables is stored, wherein the action learning model receives the design information in order from the first group, and the group is trained in order from the first group. A design method having an update step, which includes: outputting design variables, wherein the design information for the first group includes the design requirements, and the design information for the nth group includes at least the design variables for the (n-1)th group output by the behavioral learning model, and the design device updates the database by adding the design dataset, which includes the design requirements input to the behavioral learning model and a plurality of design variables output by the behavioral learning model, to the database; updating the value learning model using the updated database; and updating the behavioral learning model using the updated database so that the value output by the updated value learning model is higher.

[0184] <Note 19> The design apparatus is the design method described in Appendix 18, which repeatedly performs the update step. <Note 20> The design method according to Appendix 18 or Appendix 19, wherein the memory unit stores a separate behavioral learning model that, when the design information is input, outputs the design variable in the database that corresponds to the input design information and has the highest evaluation value, the update step includes the step in which the design device updates the separate behavioral learning model using the updated database, and in the step in which the design device updates the value learning model using the updated database, calculates a Q-learning teacher signal using the design variable output by the updated separate behavioral learning model, and updates the value learning model by Q-learning the value learning model using the calculated Q-learning teacher signal. [Explanation of Symbols]

[0185] 10...Rotating electric machine, 12...Rotor core, 13...Magnet, 21...Design device, 21a...Model update unit, 21b...Output unit, 25...Storage unit, DB...Database, F1...Input frame for behavioral learning model, F2...Output frame for behavioral learning model, F3...Input frame for value learning model, F4...Output frame for value learning model, F5...Input frame for separate behavioral learning model, F6...Output frame for separate behavioral learning model, L1...Label for behavioral learning model, L2...Label for value learning model, L3...Label for separate behavioral learning model, M1...Behavioral learning model, M10...Integrated behavioral learning model, M11, M12...Split behavioral learning model, M2...Value learning model, M20...Integrated value learning model, M21, M22...Split value learning model, M3...Separate behavioral learning model, M30...Integrated separate behavioral learning model.

Claims

1. A design device comprising an output unit that outputs multiple design variables according to design requirements, The multiple design variables are grouped into multiple groups, When the design information set for each group is input, an action learning model trained to output the design variables for the input group is used. A value learning model is provided with the design information input to the behavioral learning model, and a pair of design variables output by the behavioral learning model in relation to the design information. The value learning model is trained to output the value of the pair. A database where a design dataset containing the aforementioned design requirements and a plurality of the aforementioned design variables is stored, A memory unit that stores, A model update unit that updates the behavioral learning model and the value learning model, Equipped with, The behavioral learning model receives the design information sequentially from the first group and outputs the design variables sequentially from the first group. The design information of the first group includes the design requirements, The design information for the nth group includes at least the design variables for the (n-1)th group output by the behavioral learning model, The aforementioned model update unit performs the following update process: The database is updated by adding the design dataset, which includes the design requirements input to the behavioral learning model and the multiple design variables output by the behavioral learning model, to the database. The value learning model is updated using the updated database. A design device that updates the behavioral learning model using the updated database so that the value output by the updated value learning model increases.

2. The design apparatus according to claim 1, wherein the model update unit repeatedly performs the update process.

3. The memory unit stores a separate behavioral learning model that, upon inputting the design information, outputs the design variable in the database that corresponds to the input design information and has the highest evaluation value. The aforementioned model update unit is The separate behavioral learning model is updated using the updated database. Using the design variables output by the updated separate behavioral learning model, the teacher signal for Q-learning is calculated. The design apparatus according to claim 1, which updates the value learning model by Q-learning the value learning model using the calculated Q-learning teacher signal.

4. The aforementioned model update unit is An improvement process is executed to improve the design variables output by the behavioral learning model so that their evaluation values ​​increase. The design apparatus according to claim 1, which adds the design dataset containing the improved design variables to the database.

5. The design apparatus according to claim 4, wherein the model update unit, in the improvement process, makes the range of change of the design variable for the nth group narrower than the range of change of the design variable for the (n+1)th group.

6. The behavioral learning model is composed of the same number of segmented behavioral learning models as the number of groups. The nth segmented action learning model receives the design information for the nth group as input and outputs the design variables for the nth group. The value learning model is composed of the same number of partitioned value learning models as the number of groups. The design apparatus according to claim 1, wherein the nth partitioned value learning model receives the pair of the nth group as input and outputs the value of the nth pair.

7. The design apparatus according to claim 1, wherein the behavioral learning model is an integrated behavioral learning model configured to allow input of the design information of any group.

8. The number of input frames for the behavioral learning model for inputting the design information into the integrated behavioral learning model is equal to the sum of the number of design requirements and the number of design variables in the groups from the 1st to the (n-1th)th. The number of output frames for the behavioral learning model for outputting the design variables of the integrated behavioral learning model is the same as the number of design variables in the group with the largest number of design variables among the multiple groups. The design apparatus according to claim 7, wherein when the design information for the nth group is input to the integrated behavioral learning model, null or 0 is input to the behavioral learning model input frame corresponding to the design variable for at least the (n+1)th group among the plurality of behavioral learning model input frames.

9. The design apparatus according to claim 8, wherein the design information input to the integrated behavioral learning model is provided with a behavioral learning model label indicating which group of the design information it belongs to.

10. The design apparatus according to claim 1, wherein the value learning model is an integrated value learning model configured to accept any pair of the group as input.

11. The number of input frames for the value learning model to input the pairs into the integrated value learning model is equal to the sum of the number of design requirements and the total number of design variables. The number of output frames for the value learning model to output the value of the pair is one. The design apparatus according to claim 10, wherein when the pair of the nth group is input to the integrated value learning model, null or 0 is input to the input frame for the value learning model that corresponds to at least the (n+1)th or later group among the plurality of input frames for the value learning model.

12. The design apparatus according to claim 11, wherein the pair input to the integrated value learning model is assigned a label for the value learning model indicating which group the pair belongs to.

13. The design apparatus according to claim 3, wherein the separate behavioral learning model is an integrated separate behavioral learning model configured to allow input of the design information of any group.

14. The number of input frames for the separate behavioral learning model for inputting the design information into the integrated separate behavioral learning model is equal to the sum of the number of design requirements and the number of design variables in the groups from the 1st to the (n-1th)th. The number of output frames for the separate behavioral learning model that output the design variable in the database that corresponds to the input design information and has the highest evaluation value is the same as the number of design variables in the group with the largest number of design variables among the multiple groups. The design apparatus according to claim 13, wherein when the design information for the nth group is input to the integrated separate behavioral learning model, null or 0 is input to the separate behavioral learning model input frame corresponding to the design variables for at least the (n+1)th group among the plurality of separate behavioral learning model input frames.

15. The design apparatus according to claim 14, wherein the design information input to the integrated separate behavioral learning model is provided with a separate behavioral learning model label indicating which group of the design information it belongs to.

16. Used in the design of a rotating electric machine having a rotor core and magnets provided on the rotor core, The first group includes the design variables relating to the position and shape of the magnet, The design apparatus according to claim 1, wherein the second group includes the design variables relating to the shape of the rotor core.

17. Used in the design of a rotating electric machine having a rotor core and magnets provided on the rotor core, The first group includes the design variables relating to the size of the rotor core, The second group includes the design variables relating to the position and shape of the magnet, The design apparatus according to claim 1, wherein the third group includes the design variables relating to the shape of the rotor core.

18. A design method in which a design device outputs multiple design variables in accordance with design requirements, The multiple design variables are grouped into multiple groups, The design apparatus, When the design information set for each group is input, an action learning model trained to output the design variables for the input group is used. A value learning model is provided with the design information input to the behavioral learning model, and a pair of design variables output by the behavioral learning model in relation to the design information. The value learning model is trained to output the value of the pair. A database where a design dataset containing the aforementioned design requirements and a plurality of the aforementioned design variables is stored, It has a memory unit that stores, The behavioral learning model receives the design information sequentially from the first group and outputs the design variables sequentially from the first group. The design information of the first group includes the design requirements, The design information for the nth group includes at least the design variables for the (n-1)th group output by the behavioral learning model, The design device updates the database by adding the design dataset, which includes the design requirements input to the behavioral learning model and a plurality of design variables output by the behavioral learning model, to the database. The design device updates the value learning model using the updated database, The design device updates the behavioral learning model using the updated database so that the value output by the updated value learning model increases; A design method having an update step that includes an update step.

19. The design device is the design method according to claim 18, which repeatedly performs the update step.

20. The memory unit stores a separate behavioral learning model that, upon inputting the design information, outputs the design variable in the database that corresponds to the input design information and has the highest evaluation value. The update step includes the step of the design device updating the separate behavioral learning model using the updated database, The design method according to claim 18, wherein the design device updates the value learning model by, in the step of updating the value learning model using the updated database, calculating a Q-learning teacher signal using the design variables output by the updated separate behavior learning model, and Q-learning the value learning model using the calculated Q-learning teacher signal.

Images