Design assitance device, design assitance method, and design assitance program
The design support device improves Bayesian optimization efficiency by generating and evaluating prediction models to learn a Bayesian optimization model, addressing inefficiencies in optimizing resin composite materials through effective use of sample data.
Patent Information
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- Filing Date
- 2024-09-25
- Publication Date
- 2026-04-02
AI Technical Summary
Bayesian optimization faces challenges in optimizing compound composition of resin composite materials due to large variations in raw materials usage across samples, leading to inefficient search points and difficulty in training predictive models.
A design support device and method that utilizes machine learning to generate prediction models, evaluate their performance, and learn a Bayesian optimization model using a preferred prediction model to optimize design parameters, focusing on improving initial search efficiency by appropriately utilizing existing sample data.
Enhances optimization efficiency in the initial stages of Bayesian optimization by effectively utilizing existing sample data, enabling suitable representation of measured values and obtaining optimal design parameters for product development.
Smart Images
Figure JP2024034228_02042026_PF_FP_ABST
Abstract
Description
Design support device, design support method, and design support program
[0001] One aspect of this disclosure relates to a design support device, a design support method, and a design support program.
[0002] Product design utilizing machine learning is being researched. In one area of product design, for example, in the design of functional materials, a model is constructed to estimate material properties using machine learning with training data consisting of pairs of raw material mixing ratios and properties for experimentally and already fabricated materials, and the properties for unexperimented raw material mixing ratios are predicted. By planning experiments based on such property predictions, it becomes possible to efficiently optimize design parameters such as material properties and raw material mixing ratios, thereby improving development efficiency. Furthermore, Bayesian optimization is known to be an effective method for such optimization, and techniques for determining design parameters using Bayesian optimization are known (see, for example, Patent Document 1).
[0003] Japanese Patent Publication No. 2023-69703
[0004] In Bayesian optimization, the next search point is proposed by considering both the use of a model learned from existing data and the exploration of unknown areas of the variables. In general Bayesian optimization, in the early stages of exploration when there is little existing data, more emphasis is placed on exploration, and as the exploration progresses, the use of a model learned from existing data becomes more important. For this reason, in the early stages of exploration, a problem known as a cold start occurs, in which seemingly inefficient search points are proposed.
[0005] When sample data consisting of past experimental data is available, it may be possible to address the cold start problem by using it. However, when the compound composition data of resin composite materials, etc., which are made by blending various raw materials, is used as an explanatory variable, if the compound composition data includes many raw materials that are used in some samples but not in many others, the number of raw materials constituting the explanatory variable in the compound composition data becomes very large as the number of samples based on past experiments increases. In such cases, if many of the raw materials used in only a few samples are different from the raw materials that Bayesian optimization aims to optimize, it becomes difficult to train the predictive model for the region of focus in Bayesian optimization. Therefore, it was difficult to optimize the compound composition of the desired raw materials by Bayesian optimization using sample data consisting of past experimental data as is.
[0006] Therefore, the present invention has been made in view of the above problems, and aims to improve the optimization efficiency in the initial stages of search in Bayesian optimization by appropriately using existing sample data.
[0007] A design support device relating to one aspect of this disclosure is a design support device for determining multiple design parameters, which is applied to a method for optimizing design parameters in the design of a product manufactured based on a set of design parameters consisting of multiple design parameters, by repeatedly determining the design parameters and manufacturing the product based on the determined design parameters, and includes: a sample data acquisition unit that acquires multiple sample data, which are pairs of the set of design parameters and measured values of at least one measurement item relating to the product manufactured based on the set of design parameters; a prediction model generation unit that uses the multiple sample data and takes the set of design parameters as input to generate multiple prediction models by machine learning that predict the measured values as a probability distribution; and a system that uses the previously acquired set of design parameters and measured values to determine the initial set of search design parameters, which are a set of search design parameters obtained in the Bayesian optimization process, and the search measured values, which are measured values obtained from the product manufactured based on the set of search design parameters in the Bayesian optimization process. The system includes: an initial data acquisition unit that acquires initial exploratory measurements; a prediction model evaluation unit that acquires each of the measurements obtained by inputting the group of exploratory design parameters obtained in the Bayesian optimization process into each prediction model as a predicted measurement, and acquires evaluation metrics for the prediction of each prediction model based on the predicted measurement and the exploratory measurement; a Bayesian optimization model learning unit that uses a preferred prediction model, which is the prediction model with the best evaluation metrics among multiple prediction models, to output the difference between the predicted measurement and the exploratory measurement from the preferred prediction model, and learns a Bayesian optimization model for Bayesian optimization using the group of exploratory design parameters as input; an exploratory design parameter group acquisition unit that acquires a group of exploratory design parameters obtained by optimizing a first acquisition function, which is an acquisition function constructed based on the learned Bayesian optimization model and the preferred prediction model; and an exploratory design parameter group output unit that outputs the group of exploratory design parameters acquired by the exploratory design parameter group acquisition unit as the next group of exploratory design parameters for manufacturing the product.
[0008] A design support method relating to one aspect of this disclosure is a design support method performed by a design support device that determines multiple design parameters, for application to a method that optimizes design parameters by repeatedly determining design parameters and manufacturing a product based on a set of design parameters comprising at least one processor, the design support method being applied to the design of a product manufactured based on a set of design parameters comprising a set of design parameters comprising at least one processor, the sample data acquisition step of acquiring multiple sample data, which are pairs of a set of design parameters and a measured value of at least one measurement item relating to a product manufactured based on the set of design parameters, the prediction model generation step of generating multiple prediction models by machine learning that use the set of design parameters as input and predict the measured value as a probability distribution, and the initial exploratory design parameter set of the exploratory design parameter set, which is a set of design parameters obtained in the process of Bayesian optimization, and the measurement obtained from the product manufactured based on the exploratory design parameter set in the process of Bayesian optimization, using the previously acquired set of design parameters and measured value, The process includes: an initial data acquisition step in which the exploration measurement values are obtained as initial exploration measurement values; a prediction model evaluation step in which each measurement value obtained by inputting the exploration design parameter group obtained in the Bayesian optimization process into each prediction model is obtained as a prediction measurement value, and an evaluation index for the prediction of each prediction model is obtained based on the prediction measurement value and the exploration measurement value; a Bayesian optimization model learning step in which a Bayesian optimization model is learned using a preferred prediction model, which is the prediction model with the best evaluation index among multiple prediction models, with the difference between the prediction measurement value by the preferred prediction model and the exploration measurement value as output, and the exploration design parameter group as input; an exploration design parameter group acquisition step in which the exploration design parameter group obtained in the exploration design parameter group acquisition step is output as the next exploration design parameter group for manufacturing the product.
[0009] A design support program relating to one aspect of this disclosure is a design support program that causes a computer to function as a design support device for determining multiple design parameters, in order to apply to a method for optimizing design parameters by repeatedly determining design parameters and manufacturing a product based on the determined design parameters in the design of a product manufactured based on a set of multiple design parameters, the program comprising: a sample data acquisition step of acquiring multiple sample data, which are pairs of a set of design parameters and a measured value of at least one measurement item relating to a product manufactured based on the set of design parameters; a prediction model generation step of generating multiple prediction models by machine learning that use the set of design parameters as input and predict the measured value as a probability distribution, and the previously acquired set of design parameters and measured value, which are the initial search design parameters of the search design parameter set obtained in the Bayesian optimization process, and the search measurements obtained from the product manufactured based on the search design parameter set in the Bayesian optimization process. The computer is instructed to perform the following steps: an initial data acquisition step in which the initial exploration measurements are obtained from the exploration measurements; a prediction model evaluation step in which each measurement obtained by inputting the exploration design parameter group obtained in the Bayesian optimization process into each prediction model is obtained as a predicted measurement, and an evaluation index for the prediction of each prediction model is obtained based on the predicted measurement and the exploration measurement; a Bayesian optimization model learning step in which a Bayesian optimization model is learned using a preferred prediction model, which is the prediction model with the best evaluation index among multiple prediction models, with the difference between the predicted measurement and the exploration measurement taken as the output, and the exploration design parameter group as input; an exploration design parameter group acquisition step in which an exploration design parameter group obtained by optimizing a first acquisition function, which is an acquisition function constructed based on the learned Bayesian optimization model and the preferred prediction model, is obtained; and an exploration design parameter group output step in which the exploration design parameter group obtained in the exploration design parameter group acquisition step is output as the next exploration design parameter group for manufacturing the product.
[0010] From this perspective, for example, multiple predictive models are generated based on existing sample data obtained from previously manufactured products. Based on the set of search design parameters and search measurements obtained during the Bayesian optimization search process, evaluation metrics for each predictive model are acquired, and the Bayesian optimization model is trained based on the predictive model with the best acquired evaluation metrics. This makes it possible to obtain a Bayesian optimization model that appropriately reflects the usefulness of the sample data. Then, by optimizing the acquisition function constructed based on the trained Bayesian optimization model, it becomes possible to obtain a suitable set of design parameters for the next product to be manufactured as a set of search design parameters.
[0011] In design support devices relating to other aspects, the Bayesian optimization model may be a Gaussian process regression model, and the Bayesian optimization model learning unit may learn the Bayesian optimization model by applying a suitable prediction model to the mean function in the Gaussian process regression model.
[0012] From this perspective, by constructing a model for Bayesian optimization using a Gaussian process and applying the best predictive model with the best evaluation metric to the mean function in the Gaussian process, it becomes possible to suitably represent the distribution of measured values, which are the dependent variable with the design parameters as explanatory variables, in the model used for Bayesian optimization.
[0013] In design support devices relating to other aspects, a design parameter group output unit may further be provided, which repeats the following until a predetermined termination condition is met: a prediction model evaluation unit acquires a group of exploration design parameters for the next period and evaluation indices for each prediction model based on exploration measurement values obtained from products manufactured based on the group of exploration design parameters for the next period; a Bayesian optimization model learning unit learns a Bayesian optimization model based on a suitable prediction model; an exploration design parameter group acquisition unit acquires a group of exploration design parameters; and an exploration design parameter group output unit outputs a group of exploration design parameters. The system then outputs the group of exploration design parameters when the termination condition is met as a group of design parameters for the product.
[0014] According to such an aspect, the acquisition of evaluation indexes of each prediction model based on a search design parameter group and a search evaluation value, the learning of a model for Bayesian optimization, the acquisition and output of the next search design parameter group are repeated until a predetermined end condition is satisfied, and the search design parameter at the time of satisfaction of the end condition is output, whereby it becomes possible to obtain a suitable design parameter group for the manufactured product.
[0015] In the design support device according to another aspect, the initial data acquisition unit acquires, as an initial search design parameter group, a design parameter group obtained by optimizing a second acquisition function that is a predetermined acquisition function configured based on one prediction model selected from a plurality of prediction models that predict measurement values as probability distributions, and acquires, as initial search measurement values, measurement values obtained from a manufactured product manufactured based on the initial search design parameter group.
[0016] According to such an aspect, it becomes possible to easily obtain a search design parameter group and search measurement values suitable for initial Bayesian optimization.
[0017] In the design support device according to another aspect, the number of dimensions of the search design parameter group may be less than or equal to the number of dimensions of the design parameter group in the sample data.
[0018] According to such an aspect, for the design parameters for which optimization of the design parameter group is desired, it becomes possible to preferably perform optimization by Bayesian optimization.
[0019] In the design support device according to another aspect, the evaluation index may be a rank correlation coefficient.
[0020] According to such an aspect, it becomes possible to apply the best prediction model to Bayesian optimization from among a plurality of prediction models in which the degrees of reflection of sample data are different. Therefore, Bayesian optimization that appropriately reflects sample data according to the degree of benefit can be performed.
[0021] In the design support device according to another aspect, at least one of the plurality of prediction models may be any one of Gaussian process regression, Bayesian linear regression, random forest regression, and gradient boosting.
[0022] According to such an aspect, a prediction model that suitably reflects the relationship between the explanatory variable and the target variable in the sample data can be used as a candidate for the prediction model to be applied to Bayesian optimization. Therefore, in Bayesian optimization, it is possible to appropriately reflect the tendency of the sample data.
[0023] In the design support device according to another aspect, one of the plurality of prediction models may be configured by a function that outputs zero for any input of a group of design parameters.
[0024] According to such an aspect, when the information indicating the relationship between the explanatory variable and the target variable included in the sample data is not beneficial, the prediction model configured by the function that outputs zero is selected as the best prediction model when applied to Bayesian optimization. Therefore, it is possible to appropriately exclude the influence of the sample data in Bayesian optimization.
[0025] According to one aspect of the present disclosure, it is possible to appropriately use existing sample data to improve the optimization efficiency at the initial stage of exploration in Bayesian optimization.
[0026] It is a diagram showing an outline of the process of material design to which the design support device according to the embodiment is applied. It is a block diagram showing an example of the functional configuration of the design support device according to the embodiment. It is a hardware block diagram of the design support device according to the embodiment. It is a diagram showing an example of sample data including a group of design parameters and measured values regarding a manufactured product that has been manufactured and stored in the sample data storage unit. It is a flowchart showing an example of the content of the design support method in the design support device according to the embodiment.
[0027] Hereinafter, embodiments of the present invention will be described in detail with reference to the accompanying drawings. In the description of the drawings, the same or equivalent elements are denoted by the same reference numerals, and duplicate descriptions are omitted.
[0028] Figure 1 is a diagram illustrating an overview of an example of the design process for a manufactured product to which the design support device according to this embodiment is applied. A manufactured product is a concept that includes, for example, resin composite materials, and encompasses products, work in progress, semi-finished products, parts, or prototypes. The design support device 10 of this embodiment can be applied to a method for optimizing design parameters in the design of a manufactured product that is manufactured based on a group of design parameters consisting of multiple design parameters, by repeatedly determining the design parameters and manufacturing the manufactured product based on the determined design parameters. The design support device 10 can be applied, for example, to the design of plant synthesis conditions and composite formulation design related to the manufactured product.
[0029] Specifically, the design support device 10 can be applied not only to the development and design of materials, but also to the design of products such as automobiles and pharmaceuticals, and the optimization of the molecular structure of pharmaceuticals. In this embodiment, as described above, the design support process by the design support device 10 will be explained using an example of material design as an example of product design.
[0030] As shown in Figure 1, the design support processing by the design support device 10 is applied, as an example, to the fabrication and experimentation of materials in the plant and laboratory A. That is, materials are fabricated in the plant and laboratory A according to the set design parameter group x, and measured values y of measurement items indicating the properties of the materials are obtained based on the fabricated materials. Note that the fabrication and experimentation of materials in the plant and laboratory A may be performed by simulation. In this case, the design support device 10 provides the design parameter group x for the execution of the following simulation.
[0031] The design support device 10 optimizes the design parameters based on search data consisting of a design parameter group x and measured values y of the measurement items of the material fabricated based on the design parameter group x. Specifically, the design support device 10 outputs a design parameter group x that may provide more suitable characteristics for the next fabrication and experimentation, based on the design parameter group x and measured values y of the fabricated material.
[0032] For example, the design support device 10 of this embodiment is applied in the design of material products for the purpose of tuning multiple design parameters to optimize the set of design parameters. As an example of material product design, when a material is manufactured by mixing multiple compositions, the design support device 10 is used to tune the set of design parameters, using a set of design parameters such as the blending amounts of each composition as explanatory variables and the measured values of measurement items measured with respect to the manufactured material as the objective variable.
[0033] Figure 2 is a block diagram showing an example of the functional configuration of the design support device 10 according to the embodiment. The design support device 10 is a device for determining multiple design parameters in order to be applied to a method of optimizing design parameters by repeatedly determining design parameters and manufacturing the product based on the determined design parameters in the design of a product manufactured based on a group of design parameters consisting of multiple design parameters. As shown in Figure 2, the design support device 10 includes functional units 11 to 18, a sample data storage unit 21, and a search data storage unit 22 configured in the processor 101. Each functional unit will be described later.
[0034] Figure 3 shows an example of the hardware configuration of the computer 100 that constitutes the design support device 10 according to this embodiment. Note that the computer 100 can constitute the design support device 10.
[0035] As an example, the computer 100 includes a processor 101, a main memory 102, an auxiliary memory 103, and a communication control device 104 as hardware components. The computer 100 that constitutes the design support device 10 may further include an input device 105 such as a keyboard, touch panel, or mouse, and an output device 106 such as a display.
[0036] The processor 101 is a computing unit that executes the operating system and application programs. Examples of the processor 101 include a CPU (Central Processing Unit) and a GPU (Graphics Processing Unit), but the type of processor 101 is not limited to these. For example, the processor 101 may be a combination of dedicated circuits. The dedicated circuits may be programmable circuits such as FPGAs (Field-Programmable Gate Arrays), or other types of circuits.
[0037] The main memory 102 is a device that stores programs for realizing the design support device 10, calculation results output from the processor 101, and the like. The main memory 102 is composed of at least one of ROM (Read Only Memory) and RAM (Random Access Memory).
[0038] The auxiliary storage device 103 is generally a device capable of storing a larger amount of data than the main memory 102. The auxiliary storage device 103 is composed of a non-volatile storage medium such as a hard disk or flash memory. The auxiliary storage device 103 stores the design support program P1 and various data necessary for the computer 100 to function as a design support device 10.
[0039] The communication control device 104 is a device that performs data communication with other computers via a communication network. The communication control device 104 is composed of, for example, a network card or a wireless communication module.
[0040] Each functional element of the design support device 10 is realized by loading the corresponding design support program P1 onto the processor 101 or main memory 102 and having the processor 101 execute the program. The program P1 includes code for realizing each functional element of the corresponding server. The processor 101 operates the communication control device 104 according to the design support program P1 and performs data reading and writing in the main memory 102 or auxiliary storage device 103. Through this process, each functional element of the corresponding server is realized.
[0041] The design support program P1 may be provided by being permanently recorded on a tangible recording medium such as a CD-ROM, DVD-ROM, or semiconductor memory. Alternatively, the design support program P1 may be provided via a communication network as a data signal superimposed on a carrier wave.
[0042] Referring again to Figure 2, the design support device 10 includes a sample data acquisition unit 11, a prediction model generation unit 12, an initial data acquisition unit 13, a prediction model evaluation unit 14, a Bayesian optimization model learning unit 15, a search design parameter group acquisition unit 16, a search design parameter group output unit 17, and a design parameter group output unit 18. These functional units 11 to 18 may be configured in a single design support device as shown in Figure 2, or they may be distributed across multiple devices. The sample data storage unit 21 and the search data storage unit 22 may be configured in the design support device 10 as shown in Figure 2, or they may be configured as other devices accessible from the design support device 10.
[0043] The sample data acquisition unit 11 acquires multiple sample data sets for manufactured materials and other manufactured products. The sample data consists of pairs of design parameter sets and measured values for measurement items. The sample data storage unit 21 is a storage means for storing the sample data and may be configured as, for example, a main memory 102 and an auxiliary memory 103.
[0044] FIG. 4 is a diagram showing an example of sample data stored in the sample data storage unit 21. As shown in FIG. 4, the sample data storage unit 21 stores sample data t (t = 1 to N) of the number of samples N in past material production. The sample data of the number of samples N is the design parameter group x ｔ (t = 1 to N) and the measured value y ｔ (t = 1 to N).
[0045] The design parameter group x includes M design parameters and can constitute vector data of a dimensionality corresponding to the number D of design parameters. As an example, the design parameter group x may include blending amounts A to F, etc. of various raw materials. In addition to those exemplified, the design parameters may be, for example, non-vector data such as molecular structures and images, etc.
[0046] In the sample data, the measured value y of a predetermined measurement item is associated with the design parameter group x. The measured value y is the value of the measurement item measured in the manufactured product manufactured based on the design parameter group x.
[0047] Here, the sample data shown in FIG. 4 is expressed as follows for the following explanation. Sample data D ａｌｌ = {X ａｌｌ , Y ａｌｌ} X ａｌｌ = RM Ｎ×Ｍ Y ａｌｌ = RN Ｎ×１ In the above formula, as described above, M is the number (dimensionality) of design parameters in the design parameter group, and N is the number of samples. Also, R represents a Euclidean space.
[0048] Conventionally, based on the sample data as shown in FIG. 4, it is conceivable to obtain the design parameter group to be used for manufacturing the next manufactured product by Bayesian optimization. That is, it is conceivable to construct an acquisition function based on a model in which the design parameter group and the measured value are used as explanatory variables and objective variables, and to obtain the design parameter group to be used in the next experiment by optimizing it.
[0049] However, in the example of sample data shown in Figure 4, the raw material blending amounts D, E, and F among the design parameters are included in only some of the sample data t (t=1, 2, 3), and not in the other sample data. As a result, the inclusion of sample data t (t=1, 2, 3) in the sample data group increased the dimensionality of the design parameter group that constitutes the explanatory variables.
[0050] In such cases, when many of the raw materials used in only a small number of samples differ from the blending amounts of the raw materials to be optimized in Bayesian optimization (e.g., blending amounts A, B, C), it becomes difficult to train the predictive model for the region of interest in Bayesian optimization. Therefore, it was difficult to optimize the blending composition of the desired raw materials using Bayesian optimization by directly using the sample data. The design support device 10 of this embodiment appropriately uses such sample data to suitably perform Bayesian optimization, and in particular, by using sample data, it aims to improve the optimization efficiency in the initial stages of search in Bayesian optimization.
[0051] The prediction model generation unit 12 uses multiple sample data and takes the design parameter group x as input to generate multiple prediction models that predict the measured value y using machine learning. Specifically, the prediction model generation unit 12 uses sample data D ａｌｌ or its subset D ｓｕｂ Using the design parameter set x Ｍ (∈X) ａｌｌ ) is used as input, and the measured value y (∈Y ａｌｌ Multiple prediction models Model = {Model_1, Model_2, ..., Model_k} are generated by machine learning, with the probability part predicting ). The prediction model generation unit 12 generates multiple prediction models with different algorithms and / or hyperparameters.
[0052] The prediction model generation unit 12 may construct various prediction models using known machine learning techniques. That is, the prediction model generation unit 12 inputs a set of design parameters x into the model and performs machine learning of the prediction model by updating the model so that the error between the corresponding measured value y and the output from the model becomes smaller.
[0053] The algorithms of each of the multiple prediction models are not limited and may be any known machine learning model. That is, any prediction model is capable of predicting measured values as a probability distribution or an approximation or surrogate index thereof, and its type is not limited. At least one of the multiple prediction models may be one of Gaussian process regression, Bayesian linear regression, random forest regression, and gradient boosting. Furthermore, the prediction model may consist of a posterior distribution of predicted values based on Bayesian theory, a distribution of predicted values of the predictors constituting the ensemble, theoretical formulas for the prediction interval and confidence interval of the regression model, and a distribution obtained by Monte Carlo dropout, etc.
[0054] The initial data acquisition unit 13 acquires the previously acquired design parameter group and measurement values as the initial exploratory design parameter group, which is the group of design parameters obtained in the Bayesian optimization process, and as the initial exploratory measurement values, which are the measurement values obtained from the manufactured product based on the exploratory design parameter group in the Bayesian optimization process.
[0055] Specifically, the initial data acquisition unit 13 may acquire a group of design parameters obtained by optimizing a second acquisition function, which is a predetermined acquisition function configured based on one of several prediction models selected from a plurality of prediction models. Here, each of the plurality of prediction models is a model that predicts measured values as a probability distribution, and is a model obtained by machine learning using sample data. The initial data acquisition unit 13 may also acquire measured values obtained from a product manufactured based on the acquired group of design parameters. That is, the initial data acquisition unit 13 performs initial Bayesian optimization.
[0056] Here, let i be the number of iterations of the search and fabrication (experiment) process in Bayesian optimization, and the set of design parameters obtained as a solution through the first i iterations of the search be called the search design parameter set X. ＢＯ Based on the set of search design parameters obtained from each of the first i searches, the measured values obtained from the manufactured product are called the search measurement values Y. ＢＯ Therefore, the search design parameter group X ＢＯ and the search measurement Y ＢＯ This can be expressed as follows: X ＢＯ = R ｉ×ｍ Y ＢＯ = R ｉ×１ The search and design parameters in the search and design parameter set are the design parameters that are the focus and to be optimized in Bayesian optimization. In this embodiment, the number of dimensions m of the search and design parameter set is at least less than or equal to the number of dimensions M of the design parameter set in the sample data, and the number of dimensions m may be smaller than the number of dimensions M (m << M). Also, the data obtained during the search process of Bayesian optimization is called the search data D. ＢＯ Therefore, it will be written as follows: D ＢＯ = {X ＢＯ , Y ＢＯ}
[0057] That is, search data D ＢＯ This is the set of search design parameters and search measurement values obtained during the Bayesian optimization search up to the i-th iteration. According to this notation, the initial data acquisition unit 13 optimizes the second acquisition function and measures the product to obtain the search data D ＢＯ = {X ＢＯ , Y ＢＯ The initial data acquisition unit 13 obtains the search data D in the search data storage unit 22. ＢＯ The data may be stored. The search data storage unit 22 is a storage means for storing the search data.
[0058] The initial data acquisition unit 13 may construct a second acquisition function based on a prediction model arbitrarily selected from a plurality of prediction models. The second acquisition function may be any known acquisition function, for example, one of LCB (Lower Confidence Bound), UCB (Upper Confidence Bound), EI (Expected Improvement), and PI (Probability of Improvement).
[0059] The prediction model evaluation unit 14 uses the search design parameter group X obtained during the Bayesian optimization process. ＢＯ By inputting this into each prediction model {Model_1, Model_2, ..., Model_k}, the probability distribution of the measured values obtained is used to predict the measured value Y. ＊ ＢＯ The predicted measurement value Y is obtained as follows: ＊ ＢＯ and the search measurement Y ＢＯ Based on this, evaluation metrics for the predictions of each prediction model are obtained.
[0060] Here, the evaluation metric may also be the rank correlation coefficient. By using the rank correlation coefficient as the evaluation metric for the prediction model, it becomes possible to apply the best prediction model to Bayesian optimization from among multiple prediction models, each reflecting the sample data to a different degree. Therefore, Bayesian optimization can be performed that appropriately reflects the sample data according to the degree of usefulness.
[0061] Furthermore, one of the multiple prediction models {Model_1, Model_2, ..., Model_k} may be constructed using a function that outputs zero for any set of design parameters. This ensures that when the information regarding the relationship between the explanatory and dependent variables contained in the sample data is not useful, the prediction model constructed using a function that outputs zero is selected as the best prediction model when applied to Bayesian optimization. Thus, it becomes possible to appropriately eliminate the influence of sample data in Bayesian optimization.
[0062] The Bayesian optimization model learning unit 15 uses a preferred prediction model, which is one of several prediction models {Model_1, Model_2, ..., Model_k} that has the best evaluation metric, to learn a Bayesian optimization model for Bayesian optimization. The difference between the predicted measurement value by the preferred prediction model and the search measurement value is output, and the search design parameter set is taken as input.
[0063] Specifically, the model for Bayesian optimization may be a Gaussian process regression model. Generally, when the output has a functional relationship with respect to the input, and the joint distribution of the output for a set of inputs follows a multivariate Gaussian distribution, the relationship between the input and output is said to follow a Gaussian process. A Gaussian process is generally expressed as follows: f(x) ~ GP(m(x), k(x, x')) where the function m(x) is the mean function and k(x, x') is the covariance matrix.
[0064] When the Bayesian optimization model is composed of a Gaussian process regression model that follows a Gaussian process, the Bayesian optimization model learning unit 15 learns the Bayesian optimization model by applying the best prediction model among multiple prediction models {Model_1, Model_2, ..., Model_k} that has the best evaluation metric to the mean function in the Gaussian process regression model.
[0065] In this way, by constructing a model for Bayesian optimization using a Gaussian process and applying the best-in-class prediction model, which has the best evaluation index, to the mean function in the Gaussian process, it becomes possible to suitably represent the distribution of measured values, which are the dependent variable with the design parameters as explanatory variables, in the model used for Bayesian optimization.
[0066] The exploration design parameter acquisition unit 16 acquires a set of exploration design parameters obtained by optimizing a first acquisition function, which is an acquisition function constructed based on a trained Bayesian optimization model and a suitable prediction model. Specifically, the exploration design parameter acquisition unit 16 acquires a set of exploration design parameters x as the optimal solution for use in the production of the next product, obtained by optimizing the first acquisition function. ｉThe following is obtained. Here, the predictive variance used to calculate the first acquisition function may be the predictive variance of the Bayesian optimization model, and the predictive mean may be configured to be the sum of the outputs of the Bayesian optimization model and the preferred prediction model.
[0067] The function used to calculate the first acquisition function may be expressed as follows, for example: f(x) = m(x) + g(x) where m(x) is the predicted value (predicted mean) of the preferred prediction model, and g(x) is the Bayesian optimization model. If g(x) is a Gaussian process regression model, then f(x) is equivalent to the following Gaussian process regression: f(x) ~ GP(m(x), K) where K represents the Gram matrix.
[0068] The first acquisition function may be any known acquisition function, for example, LCB (Lower Confidence Bound), UCB (Upper Confidence Bound), EI (Expected Improvement), and PI (Probability of Improvement).
[0069] The exploration design parameter group output unit 17 outputs the exploration design parameter group acquired by the exploration design parameter group acquisition unit 16 as the next exploration design parameter group for manufacturing the product.
[0070] This allows us to explore a suitable set of design parameters for the next product: design parameter set x ｉ Since it becomes possible to obtain this, the search measurement value y can be obtained by creating a product (experiment) and measuring the measurement items on the product. ｉ This can be obtained. Here, the search data d obtained based on the process of search and creation (experiment) in the i-th Bayesian optimization is obtained. ｉ This can be expressed as follows: d ｉ = {x ｉ , y ｉ}
[0071] The design parameter group output unit 18 outputs the search design parameter group x when the search and product fabrication (experiment) process in Bayesian optimization is repeated until a predetermined termination condition is met. ｉThis is output as a set of design parameters for the manufacture of the product in question.
[0072] Specifically, the design support device 10 repeats the following until a predetermined termination condition is met: the prediction model evaluation unit 14 acquires the next-term exploratory design parameter group and evaluation indicators for each prediction model based on the exploratory measurement values obtained from the product manufactured based on the said exploratory design parameter group; the Bayesian optimization model learning unit 15 learns a Bayesian optimization model based on the optimal prediction model that yields the best evaluation indicators; the exploratory design parameter group acquisition unit 16 acquires the exploratory design parameter group; and the exploratory design parameter group output unit 17 outputs the exploratory design parameter group. The design parameter group output unit 18 then outputs the exploratory design parameter group x when the predetermined termination condition is met. ｉ This outputs the data as a set of design parameters for the production of the product. This makes it possible to obtain a suitable set of design parameters for the product.
[0073] Figure 5 is a flowchart showing an example of the design support method in the design support device 10 according to this embodiment. The design support method is executed when the design support program P1 is loaded into the processor 101 and the program is executed, thereby realizing each of the functional units 11 to 17.
[0074] In step S1, the sample data acquisition unit 11 acquires multiple sample data points for manufactured materials and other manufactured products. In step S2, the prediction model generation unit 12 uses the multiple sample data points and, taking a group of design parameters x as input, generates multiple prediction models using machine learning to predict measured values y as probability distributions.
[0075] In step S3, the initial data acquisition unit 13 acquires a group of design parameters obtained by optimizing a second acquisition function configured based on one prediction model selected from a plurality of prediction models, as an initial group of exploration design parameters. In step S4, the initial data acquisition unit 13 acquires measured values obtained from products manufactured based on the initial group of exploration design parameters, as initial exploration measured values.
[0076] In step S5, the prediction model evaluation unit 14 evaluates the group of search and design parameters X obtained during the Bayesian optimization process. ＢＯ The predicted measurement value Y obtained by inputting this into each prediction model {Model_1, Model_2, ..., Model_k} ＊ ＢＯ and the search measurement Y ＢＯ Based on this, we obtain evaluation metrics for each prediction model.
[0077] In step S6, the Bayesian optimization model learning unit 15 constructs a Bayesian optimization model using the optimal prediction model among multiple prediction models {Model_1, Model_2, ..., Model_k} that has the best evaluation metric.
[0078] In step S7, the Bayesian optimization model learning unit 15 outputs the difference between the predicted measurement value by the prediction model and the search measurement value, and uses the search design parameter group as input to learn the Bayesian optimization model.
[0079] In step S8, the exploration design parameter acquisition unit 16 acquires the next set of exploration design parameters by optimizing a first acquisition function configured based on a trained Bayesian optimization model and a preferred prediction model. In step S9, the exploration design parameter output unit 17 outputs the exploration design parameter group acquired by the exploration design parameter acquisition unit 16 in step S8 as the exploration design parameter group for manufacturing the next product.
[0080] In step S10, the design support device 10 determines whether a predetermined termination condition is met. If it is determined that the termination condition is met, the process proceeds to step S12. If it is determined that the termination condition is not met, the process returns to step S5, and the Bayesian optimization process is repeated.
[0081] In step S12, the design parameter group output unit 18 outputs the search design parameter group obtained when the search and product fabrication (experiment) processes in Bayesian optimization in steps S5 to S11 are repeated until a predetermined termination condition is met, as the design parameter group for fabricating the product.
[0082] According to the design support device 10, design support method, and design support program P1 of this embodiment described above, for example, multiple prediction models are generated based on existing sample data obtained from previously manufactured products. Based on the search design parameter group and search measurement values obtained during the Bayesian optimization search process, evaluation indices for each prediction model are obtained, and the Bayesian optimization model is trained based on the best-suited prediction model with the best-suited evaluation indices. This makes it possible to obtain a Bayesian optimization model that appropriately reflects the sample data according to its usefulness. Then, by optimizing the acquisition function constructed based on the trained Bayesian optimization model, it becomes possible to obtain a suitable set of design parameters for the next product as a search design parameter group.
[0083] The present invention has been described in detail above based on its embodiments. However, the present invention is not limited to the above embodiments. The present invention can be modified in various ways without departing from its spirit.
[0084] The gist of this disclosure is as follows: [1] to
[10] [1] A design support device for determining a plurality of design parameters, which is applied to a method for optimizing design parameters by repeatedly determining design parameters and manufacturing a product based on the determined design parameters in the design of a product manufactured based on a set of design parameters consisting of a plurality of design parameters, the device comprising: a sample data acquisition unit that acquires a plurality of sample data, which are pairs of the set of design parameters and measured values of at least one measurement item relating to the product manufactured based on the set of design parameters; a prediction model generation unit that uses the plurality of sample data and takes the set of design parameters as input to generate a plurality of prediction models by machine learning that predict the measured values as probability distributions; and an initial data acquisition unit that acquires the previously acquired set of design parameters and measured values as the initial exploratory design parameter set, which is the set of design parameters obtained in the process of Bayesian optimization, and the initial exploratory measured values, which are the measured values obtained from the product manufactured based on the set of design parameters in the process of Bayesian optimization. A design support device comprising: a prediction model evaluation unit that acquires each of the measured values obtained by inputting the group of exploration design parameters obtained in the process of Bayesian optimization into each prediction model as a predicted measured value, and acquires an evaluation index for the prediction of each prediction model based on the predicted measured value and the exploration measured value; a Bayesian optimization model learning unit that learns a Bayesian optimization model for Bayesian optimization using a preferred prediction model which is the prediction model among the plurality of prediction models that has the best evaluation index, outputs the difference between the predicted measured value by the preferred prediction model and the exploration measured value, and takes the group of exploration design parameters as input; a search and design parameter group acquisition unit that acquires the group of exploration design parameters obtained by optimizing a first acquisition function which is an acquisition function constructed based on the learned Bayesian optimization model and the preferred prediction model; and a search and design parameter group output unit that outputs the group of exploration design parameters acquired by the search and design parameter group acquisition unit as the next search and design parameter group for manufacturing a product. [2]The design support device according to [1], wherein the Bayesian optimization model is a Gaussian process regression model, and the Bayesian optimization model learning unit learns the Bayesian optimization model by applying the preferred prediction model to the mean function in the Gaussian process regression model. [3] The design support device according to [1] or [2], further comprising: the prediction model evaluation unit acquires the evaluation index of each prediction model based on the next period's exploration design parameter group and the exploration measurement values obtained from the product manufactured based on the next period's exploration design parameter group; the Bayesian optimization model learning unit learns the Bayesian optimization model based on the prediction model that has the best evaluation index; the exploration design parameter group acquisition unit acquires the exploration design parameter group; and the exploration design parameter group output unit outputs the exploration design parameter group until a predetermined termination condition is met, and the design support device according to [1] or [2], further comprising: the prediction model evaluation unit acquires the next period's exploration design parameter group and the exploration measurement values obtained from the product manufactured based on the next period's exploration design parameter group; the Bayesian optimization model learning unit learns the Bayesian optimization model based on the prediction model that has the best evaluation index; the exploration design parameter group acquisition unit acquires the exploration design parameter group; and the exploration design parameter group output unit outputs the exploration design parameter group when the termination condition is met as a design parameter group for the product. [4] The initial data acquisition unit acquires a group of design parameters obtained by optimizing a second acquisition function, which is a predetermined acquisition function configured based on one of the plurality of prediction models selected from the plurality of prediction models that predict the measured values as a probability distribution, as the initial exploration design parameter group, and acquires measured values obtained from the manufactured product manufactured based on the initial exploration design parameter group, as the initial exploration measured values, the design support device according to any one of [1] to [3]. [5] The number of dimensions of the exploration design parameter group is less than or equal to the number of dimensions of the design parameter group in the sample data, the design support device according to any one of [1] to [4]. [6] The evaluation index is the rank correlation coefficient, the design support device according to any one of [1] to [5]. [7] At least one of the plurality of prediction models is one of Gaussian process regression, Bayesian linear regression, random forest regression, and gradient boosting, the design support device according to any one of [1] to [6]. [8] One of the plurality of prediction models is configured as a function that outputs zero for any input of the design parameter group,A design support device as described in any one of [1] to [7]. [9] A design support method performed by a design support device that determines a plurality of design parameters, for application to a method that optimizes design parameters by repeatedly determining design parameters and manufacturing a product based on a set of design parameters comprising a plurality of design parameters, the method comprising: a sample data acquisition step of acquiring a plurality of sample data which are pairs of the set of design parameters and measured values of at least one measurement item relating to the product manufactured based on the set of design parameters; a prediction model generation step of generating a plurality of prediction models by machine learning using the plurality of sample data, with the set of design parameters as input; and an initial data acquisition step of acquiring a previously acquired set of design parameters and measured values as an initial exploratory design parameter group among the exploratory design parameter group which are a set of design parameters obtained in the process of Bayesian optimization, and as an initial exploratory measured value among the exploratory measured values which are measured values obtained from a product manufactured based on the set of design parameters in the process of Bayesian optimization. A prediction model evaluation step in which each of the measured values obtained by inputting the search design parameter group obtained in the Bayesian optimization process into each prediction model is obtained as a predicted measured value, and an evaluation index for the prediction of each prediction model is obtained based on the predicted measured value and the search measured value; a Bayesian optimization model learning step in which a Bayesian optimization model is learned using a preferred prediction model which is the prediction model among the plurality of prediction models that has the best evaluation index, with the difference between the predicted measured value by the preferred prediction model and the search measured value as the output, and the search design parameter group as input; and a search design parameter group acquisition step in which the search design parameter group obtained by optimizing a first acquisition function which is an acquisition function constructed based on the learned Bayesian optimization model and the preferred prediction model,A design support method comprising: an exploratory design parameter group output step, which outputs the exploratory design parameter group obtained in the exploratory design parameter group acquisition step as the next exploratory design parameter group for manufacturing a product.
[10] A design support program for causing a computer to function as a design support device for determining a plurality of design parameters, in order to apply to a method for optimizing design parameters by repeatedly determining design parameters and manufacturing a product based on the determined design parameters, in the design of a product manufactured based on a set of design parameters consisting of a plurality of design parameters, the program comprising: a sample data acquisition step of acquiring a plurality of sample data, which are pairs of the set of design parameters and measured values of at least one measurement item relating to the product manufactured based on the set of design parameters; a prediction model generation step of generating a plurality of prediction models by machine learning that use the set of design parameters as input and predict the measured values as a probability distribution, using the plurality of sample data; and an initial data acquisition step of acquiring the previously acquired set of design parameters and measured values as the initial exploratory design parameter set of the exploratory design parameter set, which is the set of design parameters obtained in the process of Bayesian optimization, and the initial exploratory measured values, which are the measured values obtained from the product manufactured based on the set of design parameters in the process of Bayesian optimization. A prediction model evaluation step in which each of the measured values obtained by inputting the search design parameter group obtained in the Bayesian optimization process into each prediction model is obtained as a predicted measured value, and an evaluation index for the prediction of each prediction model is obtained based on the predicted measured value and the search measured value; a Bayesian optimization model learning step in which a Bayesian optimization model is learned using a preferred prediction model which is the prediction model among the plurality of prediction models that has the best evaluation index, with the difference between the predicted measured value by the preferred prediction model and the search measured value as the output, and the search design parameter group as input; and a search design parameter group acquisition step in which the search design parameter group obtained by optimizing a first acquisition function which is an acquisition function constructed based on the learned Bayesian optimization model and the preferred prediction model,A design support program that causes the computer to perform a search and design parameter output step, which outputs the search and design parameter group obtained in the search and design parameter group acquisition step as the next search and design parameter group for manufacturing a product.
[0085] 10...Design support device, 11...Sample data acquisition unit, 12...Predictive model generation unit, 13...Initial data acquisition unit, 14...Predictive model evaluation unit, 15...Bayesian optimization model learning unit, 16...Exploration design parameter group acquisition unit, 17...Exploration design parameter group output unit, 18...Design parameter group output unit, 21...Sample data storage unit, 21...Sample data storage unit, 22...Exploration data storage unit.
Claims
1. In designing a product manufactured based on a set of design parameters consisting of multiple design parameters, a design support device for determining the multiple design parameters is applied to a method for optimizing the design parameters by repeatedly determining the design parameters and manufacturing the product based on the determined design parameters, comprising: a sample data acquisition unit that acquires multiple sample data sets, each consisting of a set of design parameters and a measured value of at least one measurement item relating to the product manufactured based on the set of design parameters; a prediction model generation unit that uses the multiple sample data sets to generate multiple prediction models by machine learning, taking the set of design parameters as input; an initial data acquisition unit that acquires the previously acquired set of design parameters and measured values as the initial exploratory design parameters, which are the set of design parameters obtained in the Bayesian optimization process, and the initial exploratory measured values, which are the measured values obtained from the product manufactured based on the set of design parameters in the Bayesian optimization process; and a prediction model evaluation unit that acquires each of the measured values obtained by inputting the set of design parameters obtained in the Bayesian optimization process into each prediction model as a predicted measured value, and acquires an evaluation index for the prediction of each prediction model based on the predicted measured value and the exploratory measured value. A design support device comprising: a Bayes optimization model learning unit that learns a Bayes optimization model for Bayes optimization using a preferred prediction model which is the prediction model among the plurality of prediction models that has the best evaluation index, outputs the difference between the predicted measurement value by the preferred prediction model and the search measurement value, and takes the search design parameter group as input; a search design parameter group acquisition unit that acquires the search design parameter group obtained by optimizing a first acquisition function which is an acquisition function constructed based on the learned Bayes optimization model and the preferred prediction model; and a search design parameter group output unit that outputs the search design parameter group acquired by the search design parameter group acquisition unit as the next search design parameter group for manufacturing the product.
2. The design support device according to claim 1, wherein the Bayesian optimization model is a Gaussian process regression model, and the Bayesian optimization model learning unit learns the Bayesian optimization model by applying the preferred prediction model to the mean function in the Gaussian process regression model.
3. The design support device according to claim 1, further comprising: a design parameter group output unit that repeatedly performs the following until a predetermined termination condition is met: acquisition of evaluation indices for each prediction model based on the next period's exploration design parameter group and the exploration measurement values obtained from the manufactured product manufactured based on the next period's exploration design parameter group by the prediction model evaluation unit; learning of the Bayesian optimization model based on the preferred prediction model by the Bayesian optimization model learning unit; acquisition of the exploration design parameter group by the exploration design parameter group acquisition unit; and output of the exploration design parameter group by the exploration design parameter group output unit; and outputs the exploration design parameter group when the termination condition is met as a design parameter group for the manufactured product.
4. The design support device according to claim 1, wherein the initial data acquisition unit acquires a group of design parameters obtained by optimizing a second acquisition function, which is a predetermined acquisition function configured based on one of the plurality of prediction models selected from the plurality of prediction models that predict the measured values as a probability distribution, as the initial exploration design parameter group, and acquires measured values obtained from the manufactured product manufactured based on the initial exploration design parameter group, as the initial exploration measured values.
5. The design support device according to claim 1, wherein the number of dimensions of the exploration design parameter group is less than or equal to the number of dimensions of the design parameter group in the sample data.
6. The design support device according to any one of claims 1 to 5, wherein the evaluation index is the rank correlation coefficient.
7. The design support device according to claim 1, wherein at least one of the plurality of prediction models is one of Gaussian process regression, Bayesian linear regression, random forest regression, and gradient boosting.
8. The design support device according to claim 1, wherein one of the plurality of prediction models is configured as a function that outputs zero for any input of the set of design parameters.
9. A design support method performed by a design support device that determines a plurality of design parameters, for application to a method that optimizes design parameters by repeatedly determining design parameters and manufacturing a product based on a set of design parameters comprising A design support method comprising: a prediction model evaluation step of acquiring each of the measured values obtained by inputting the exploration design parameter group obtained in the Bayesian optimization process into each prediction model as a predicted measured value, and acquiring an evaluation index for the prediction of each prediction model based on the predicted measured value and the exploration measured value; a Bayesian optimization model learning step of learning a Bayesian optimization model for Bayesian optimization using a preferred prediction model which is the prediction model among the plurality of prediction models that has the best evaluation index, outputting the difference between the predicted measured value by the preferred prediction model and the exploration measured value, and taking the exploration design parameter group as input; an exploration design parameter group acquisition step of acquiring the exploration design parameter group obtained by optimizing a first acquisition function which is an acquisition function constructed based on the learned Bayesian optimization model and the preferred prediction model; and an exploration design parameter group output step of outputting the exploration design parameter group acquired in the exploration design parameter group acquisition step as the next exploration design parameter group for manufacturing the product. 10.A design support program for causing a computer to function as a design support device for determining multiple design parameters, in order to apply to a method for optimizing design parameters by repeatedly determining design parameters and manufacturing a product based on the determined design parameters in the design of a product manufactured based on a set of multiple design parameters, comprising: a sample data acquisition step of acquiring multiple sample data, which are pairs of the set of design parameters and measured values of at least one measurement item relating to the product manufactured based on the set of design parameters; a prediction model generation step of generating multiple prediction models by machine learning that use the set of design parameters as input and predict the measured values, using the multiple sample data; and an initial data acquisition step of acquiring the previously acquired set of design parameters and measured values as the initial exploratory design parameter set of the exploratory design parameter set, which is the set of design parameters obtained in the Bayesian optimization process, and the initial exploratory measured values of the exploratory measured values, which are the measured values obtained from the product manufactured based on the exploratory design parameter set in the Bayesian optimization process. A design support program that causes the computer to execute the following steps: a prediction model evaluation step, which involves inputting the set of exploration design parameters obtained during the Bayesian optimization process into each prediction model to obtain each measurement value as a predicted measurement value, and obtaining an evaluation index for the prediction of each prediction model based on the predicted measurement value and the exploration measurement value; a Bayesian optimization model learning step, which uses a preferred prediction model, which is the prediction model with the best evaluation index among the multiple prediction models, to learn a Bayesian optimization model for Bayesian optimization, outputting the difference between the predicted measurement value by the preferred prediction model and the exploration measurement value, and taking the set of exploration design parameters as input; a search and design parameter acquisition step, which involves acquiring the set of exploration design parameters obtained by optimizing a first acquisition function, which is an acquisition function constructed based on the learned Bayesian optimization model and the preferred prediction model; and a search and design parameter output step, which outputs the set of exploration design parameters obtained in the search and design parameter acquisition step as the next set of search and design parameters for manufacturing the product.