JP2023001429A - Learning method of model parameter - Google Patents

Abstract

To provide a learning method of a model parameter capable of accurately learning the model parameter of a prediction model for predicting a feeling of a formulation.SOLUTION: A learning apparatus 1 acquires an actual value of a feeling as teacher data (STEP 11), acquires a feature quantity including a physical property value of a formulation (STEP 13, 15, and 16), creates learning data using the feature quantity and the teacher data (STEP 17), and performs learning of a model parameter of a prediction model according to a type of the feeling by using a learning algorithm to which a machine learning algorithm with a teacher is applied and the learning data (STEP 19, 21, and 22).SELECTED DRAWING: Figure 5

Description

The present invention relates to a method of learning model parameters of a predictive model that predicts the feel of a formulation.

The applicant of the present application has already proposed the method described in Patent Document 1 as a method for learning model parameters of a prediction model. In this learning method, the model parameters of a prediction model whose objective variable is the physical property value of the formulation are learned using a predetermined learning algorithm.

Japanese Patent Application No. 2020-126447 (unpublished)

When predicting the properties of formulations, there is a demand for predicting properties other than physical properties. For example, if the formulation is a cosmetic product, the feel of the product (feeling on the skin) is an important factor directly linked to its commercial value. It is However, in the case of this method, the number of man-hours and manufacturing costs increase, so it is desired to accurately predict the feel of formulations such as cosmetics using prediction models.

The present invention has been made to solve the above problems, and an object of the present invention is to provide a model parameter learning method capable of accurately learning the model parameters of a prediction model for predicting the feel of a formulation. .

In order to achieve the above object, the invention according to claim 1 provides model parameter learning in which a learning device learns model parameters of a prediction model that predicts the feel of a formulation prepared by blending a plurality of raw materials as an objective variable. A method, wherein the learning device includes a teacher data acquisition step of acquiring measured values of objective variables as teacher data, a feature quantity acquisition step of acquiring feature quantities including physical property values of formulations, and using the feature quantities and teacher data. a learning data creation step of creating learning data; and a learning step of learning model parameters of a prediction model using a learning algorithm to which a supervised machine learning algorithm is applied and the learning data. and

According to this model parameter learning method, learning data is created using teacher data, which is the feature value including the physical property value of the formulation and the measured value of the objective variable, and the learning algorithm and the supervised machine learning algorithm are applied. Learning of model parameters of the prediction model is performed using the training data. On the other hand, as will be described later, in the experiments of the present applicant, when learning the model parameters of a prediction model that predicts the feel of a formulation as an objective variable, in the case of a feel such as "thickness", It was confirmed that the learning accuracy of the model parameters was improved by using only physical property values as feature quantities. Therefore, according to this model parameter learning method, it is possible to improve the learning accuracy when learning the model parameters of a prediction model that predicts the feel of a formulation such as "thickness" as an objective variable.

The invention according to claim 2 is the model parameter learning method according to claim 1, wherein the formulation is a one-phase formulation, and in the feature amount acquisition step, only physical property values of the formulation are acquired as feature amounts. 1 acquisition method, a second acquisition method that acquires the first selection data selected from the blending ratio of multiple raw materials and the physical property value of the formulation as a feature amount, and the blend ratio of multiple raw materials and the physical property value of the formulation Any one of the third acquisition methods that acquires the second selected data selected from among as a feature amount is selected according to the type of feeling, and in the learning step, the prediction model is a regression model and a classification One of the models is selected according to the type of touch, and the learning algorithm is an algorithm that applies a multivariate prediction method to a supervised machine learning algorithm, and one of the supervised machine learning algorithms is the type of touch. is selected according to

As will be described later, according to the experiments of the present applicant, when learning the model parameters of a prediction model that predicts the feel of a single-phase formulation as an objective variable, depending on the type of feel, the above-mentioned first to third acquisition methods Select any one method to acquire feature values, select one of a regression model and a classification model as a prediction model, and apply a multivariate prediction method to a supervised machine learning algorithm as a learning algorithm, and supervised machine learning algorithms, it was confirmed that the learning accuracy of the model parameters was improved.

Therefore, according to this model parameter learning method, by selecting one of the first to third acquisition methods according to the type of feeling and selecting the prediction model and learning algorithm as described above, It is possible to improve the learning accuracy of the model parameters of a prediction model that predicts the texture of a single-phase formulation as an objective variable.

The invention according to claim 3 is the model parameter learning method according to claim 2, wherein in the second acquisition method, a supervised machine using a feature amount that is a blending ratio of a plurality of raw materials and teacher data as learning data The learning algorithm learns the model parameters of the machine learning model, calculates the importance of the feature amount using the machine learning model after learning, and selects the feature amount in descending order of importance. By applying the first filtering process to the step and the feature amount, which is the mixing ratio of a plurality of raw materials, the correlation between the objective variable and the feature amount is obtained, and the feature amount is correlated with the objective variable. and a second filtering process different from the first filtering process is applied to the feature amount, which is the mixing ratio of a plurality of raw materials, to obtain the feature amount for the objective variable. A third selection step of acquiring the contribution, selecting the feature quantity in order from the highest contribution, calculating the average value of the ranking of the feature quantities selected in the first to third selection steps, and calculating the average value and a fourth selection step of selecting the feature amount from the first place to a predetermined rank as the first selection data when the highest feature amount of is ranked first.

According to this model parameter learning method, the first to fourth selection steps are executed in the second acquisition method. In this first selection step, learning of model parameters of a machine learning model is performed by a supervised machine learning algorithm using a feature amount that is a blending ratio of a plurality of raw materials and teacher data as learning data, and machine learning after learning The model is used to obtain the importance of feature quantities, and the feature quantities are selected in descending order of importance. Therefore, the feature quantities can be selected in order from the one that can be regarded as the most important in the learning of the supervised machine learning model.

Further, in the second selection step, by applying the first filtering process to the feature amount, which is the mixing ratio of a plurality of raw materials, the correlation between the objective variable and the feature amount is obtained, and the feature amount is Selection is made in descending order of correlation with the objective variable. Therefore, feature quantities can be selected in descending order of correlation with the objective variable.

Furthermore, in the third selection step, the contribution of the feature amount to the objective variable is obtained by applying a second filter process different from the first filter process to the feature amount, which is the mixing ratio of the plurality of raw materials. In addition, the feature quantities are selected in descending order of contribution. Therefore, feature quantities can be selected in descending order of contribution to the objective variable.

Then, in the fourth selection step, the average value of the ranks of the feature quantities selected in the first to third selection steps is calculated, and the feature quantity with the highest average value is ranked first. Since the feature amounts up to a predetermined rank are selected as the first selection data, the first selection data are appropriately selected while suppressing the effects of noise and variations in the selection results of the feature amounts in the first to third selection steps. be able to. Thereby, the eligibility and selection accuracy of the first selection data can be improved.

The invention according to claim 4 is the model parameter learning method according to claim 2 or 3, wherein in the third acquisition method, feature amounts and teacher data, which are blending ratios and physical property values of a plurality of raw materials, are used as learning data. Then, the model parameters of the machine learning model are learned by a supervised machine learning algorithm, the importance of the feature quantity is calculated using the machine learning model after learning, and the feature quantity is ranked in order from the highest importance By applying the third filtering process to the fifth selection step of selecting and the feature amount that is the blending ratio and physical property value of a plurality of raw materials, the correlation between the objective variable and the feature amount is obtained, and the feature A sixth selection step of selecting quantities in descending order of correlation with the objective variable; By applying, the contribution of the feature amount to the objective variable is obtained, and the feature amount selected in the fifth to seventh selection steps is selected in descending order of contribution. and an eighth selection step of calculating the average value of the ranks and selecting, as second selection data, feature quantities from the first rank to a predetermined rank when the feature quantity with the highest average value is ranked first. characterized by being

According to this model parameter learning method, the fifth to eighth steps are executed in the third acquisition method. In this fifth selection step, learning of the model parameters of the machine learning model is performed by a supervised machine learning algorithm using the feature amount and teacher data, which are the mixing ratio and physical property values of a plurality of raw materials, as learning data. The importance of the feature quantity is obtained using the machine learning model of , and the feature quantity is selected in descending order of importance. Therefore, the feature quantities can be selected in order from the one that can be regarded as the most important in the learning of the supervised machine learning model.

Further, in the sixth selection step, the correlation between the objective variable and the feature amount is obtained by applying the third filtering process to the feature amount, which is the mixing ratio and physical property value of the plurality of raw materials, and Feature quantities are selected in descending order of correlation with the objective variable. Therefore, feature quantities can be selected in descending order of correlation with the objective variable.

Furthermore, in the seventh selection step, by applying a fourth filtering process different from the third filtering process to the feature amount, which is the mixing ratio and physical property value of a plurality of raw materials, the contribution of the feature amount to the objective variable is As well as being acquired, the feature quantities are selected in descending order of contribution. Therefore, feature quantities can be selected in descending order of contribution to the objective variable, and the eligibility and selection accuracy of the feature quantity selection results can be improved.

Then, in the eighth selection step, the average value of the ranks of the feature quantities selected in the fifth to seventh selection steps is calculated, and the feature quantity with the highest average value is ranked first. Since the feature amounts up to a predetermined rank are selected as the second selection data, the second selection data are appropriately selected while suppressing the influence of noise and variations in the selection results of the feature amounts in the fifth to seventh selection steps. be able to. Thereby, the eligibility and selection accuracy of the second selection data can be improved.

The invention according to claim 5 is the model parameter learning method according to claim 1, wherein the formulation is an emulsified formulation, and in the feature amount acquisition step, only physical property values of the formulation are acquired as feature amounts. 1 acquisition method, a second acquisition method that acquires the first selection data selected from the compounding ratio of a plurality of raw materials and the manufacturing method of the formulation and the physical property value of the formulation as a feature amount, and the compounding ratio of a plurality of raw materials, Any one of the third acquisition methods for acquiring the second selected data selected from the physical property values of the formulation and the manufacturing method of the formulation as a feature amount is executed according to the type of feeling, and in the learning step, As a prediction model, one of a regression model and a classification model is selected according to the type of touch, and as a learning algorithm, an algorithm applying a multivariate prediction method to a supervised machine learning algorithm, and a supervised machine It is characterized in that one of the learning algorithms is selected according to the type of touch.

As will be described later, according to experiments by the present applicant, when learning the model parameters of a prediction model that predicts the feel of an emulsified formulation as an objective variable, depending on the type of feel, the above 1st to 3rd acquisition methods Select any one method to acquire feature values, select one of a regression model and a classification model as a prediction model, and apply a multivariate prediction method to a supervised machine learning algorithm as a learning algorithm, and supervised machine learning algorithms, it was confirmed that the learning accuracy of the model parameters was improved.

Therefore, according to this model parameter learning method, by selecting one of the first to third acquisition methods according to the type of feeling and selecting the prediction model and learning algorithm as described above, It is possible to improve the learning accuracy of the model parameters of a prediction model that predicts the feel of an emulsified formulation as an objective variable.

The invention according to claim 6 is the model parameter learning method according to claim 5, wherein in the second acquisition method, the feature amount and teacher data, which are the blending ratio of a plurality of raw materials and the manufacturing method of the formulation, are used as learning data. , Execute learning of the machine learning model by a supervised machine learning algorithm, calculate the importance of the feature amount using the machine learning model after learning, and select the feature amount in order from the highest importance First A selection step, and applying a first filtering process to the feature amounts that are the compounding ratio of a plurality of raw materials and the manufacturing method of the formulation, thereby selecting the feature amounts in descending order of correlation with the objective variable. 2. Acquire the degree of contribution of the feature amount to the target variable by applying a second filter process different from the first filter process to the feature amount, which is the blending ratio of a plurality of raw materials and the manufacturing method of the formulation, in a selection step. In addition, a third selection step for selecting the feature quantity in descending order of contribution, calculating the average rank of the feature quantity selected in the first to third selection steps, and calculating the feature quantity with the highest average value and a fourth selection step of selecting, as the first selection data, the feature amount from the first place to the predetermined rank when the is the first place.

According to this model parameter learning method, the first to fourth steps are executed in the second acquisition method. In this first selection step, learning of the model parameters of the machine learning model is performed by a supervised machine learning algorithm using the feature amount and teacher data, which are the compounding ratio of a plurality of raw materials and the manufacturing method of the formulation, as learning data. A later machine learning model is used to obtain the importance of the feature quantities, and the feature quantities are selected in descending order of importance. Therefore, the feature quantities can be selected in order from the one that can be regarded as the most important in the learning of the supervised machine learning model.

Further, in the second selection step, the correlation between the objective variable and the feature amount is obtained by applying the first filtering process to the feature amount, which is the compounding ratio of the plurality of raw materials and the manufacturing method of the formulation. , the feature quantity is selected in descending order of correlation with the objective variable. Therefore, feature quantities can be selected in descending order of correlation with the objective variable.

Furthermore, in the third selection step, by applying a second filter process different from the first filter process to the feature amount, which is the blending ratio of a plurality of raw materials and the manufacturing method of the formulation, the contribution of the feature amount to the objective variable is obtained, and the feature quantities are selected in descending order of contribution. Therefore, feature quantities can be selected in descending order of contribution to the objective variable.

Then, in the fourth selection step, the average value of the ranks of the feature quantities selected in the first to third selection steps is calculated, and the feature quantity with the highest average value is ranked first. Since the feature amounts up to a predetermined rank are selected as the first selection data, the first selection data are appropriately selected while suppressing the effects of noise and variations in the selection results of the feature amounts in the first to third selection steps. be able to. Thereby, the eligibility and selection accuracy of the first selection data can be improved.

The invention according to claim 7 is the model parameter learning method according to claim 5 or 6, wherein in the third acquisition method, feature amounts and teacher data, which are the blending ratio of a plurality of raw materials, physical property values, and the manufacturing method of the formulation, are obtained. Using the data as training data, a machine learning model is trained using a supervised machine learning algorithm, and the machine learning model after training is used to calculate the importance of feature quantities, and the feature quantities are sorted from the highest to the highest. A fifth selection step of sequentially selecting, and applying a third filtering process to the feature amounts, which are the compounding ratio of a plurality of raw materials, the physical property values, and the manufacturing method of the formulation, to determine the correlation between the objective variable and the feature amount. A sixth selection step of selecting the feature amount in order from the one with the highest correlation with the objective variable, and a third By applying the fourth filtering process, which is different from the filtering process, the degree of contribution of the feature amount to the objective variable is obtained, and when the feature amount with the highest degree of contribution is ranked first, the feature amount below the first place is The average value of the ranks of the feature values selected in the seventh selection step and the fifth to seventh selection steps is calculated, and the feature value with the highest average value is ranked first. and an eighth selection step of selecting the feature amounts from the first to the predetermined rank of the event as the second selection data.

According to this model parameter learning method, the fifth to eighth steps are executed in the third acquisition method. In this fifth selection step, learning of the model parameters of the machine learning model is performed by a supervised machine learning algorithm using the feature values and teacher data, which are the compounding ratios of multiple raw materials, physical property values, and formulation manufacturing methods, as learning data. Then, the machine learning model after learning is used to acquire the importance of the feature quantity, and the feature quantity is selected in descending order of importance. Therefore, the feature quantities can be selected in order from the one that can be regarded as the most important in the learning of the supervised machine learning model.

In addition, in the sixth selection step, the correlation between the objective variable and the feature amount is obtained by applying the third filtering process to the feature amount, which is the compounding ratio of a plurality of raw materials, the physical property values, and the manufacturing method of the formulation. and the feature quantities are selected in descending order of correlation with the objective variable. Therefore, feature quantities can be selected in descending order of correlation with the objective variable.

Furthermore, in the seventh selection step, by applying a fourth filter process different from the third filter process to the feature amounts that are the compounding ratio of the plurality of raw materials, the physical property values, and the manufacturing method of the formulation, the feature amount for the objective variable is obtained, and the feature quantities are selected in descending order of contribution. Therefore, feature quantities can be selected in descending order of contribution to the objective variable, and the eligibility and selection accuracy of the feature quantity selection results can be improved.

Then, in the eighth selection step, the average value of the ranks of the feature quantities selected in the fifth to seventh selection steps is calculated, and the feature quantity with the highest average value is ranked first. Since the feature amounts up to a predetermined rank are selected as the second selection data, the second selection data are appropriately selected while suppressing the influence of noise and variations in the selection results of the feature amounts in the fifth to seventh selection steps. be able to. Thereby, the eligibility and selection accuracy of the second selection data can be improved.

1 is a diagram showing a learning device that executes a model parameter learning method according to an embodiment of the present invention; FIG. It is a figure which shows an example of the feature-value of the database used for learning. It is a figure which shows an example of the objective variable of the database used for learning. 4 is a flowchart showing learning processing by a learning device; 4 is a flow chart showing a learning process for a single-phase system; FIG. 10 is a diagram showing an example of a touch that gives a good result when a prediction model is learned. FIG. 10 is a diagram showing the results of learning a predictive model of feeling of “stretching” and “fresh and fresh” with learning patterns 1 to 3; FIG. 10 is a diagram showing results when a regression model is used as a predictive model for "rough" and "warm" sensations. FIG. 10 shows the results when a classification model is used as a predictive model for "rough" and "warm" tactile sensations. FIG. 10 is a diagram showing an example of feeling when a good result is obtained when a prediction model is learned by an algorithm to which a multivariate prediction method is applied; FIG. 10 is a diagram showing the results of learning a prediction model for the feeling of “feeling covered” and “fitting in” using the multivariate prediction method and learning without using the multivariate prediction method.

A model parameter learning method according to an embodiment of the present invention will be described below with reference to the drawings. This embodiment is a prediction model model that predicts the feel of these formulations as objective variables by the learning method described below when a plurality of cosmetics are prepared as a plurality of formulations by stirring a mixture of various raw materials. It performs parameter learning. In the following description, learning the model parameters of the prediction model is appropriately referred to as "learning the prediction model".

Specifically, the learning method of this embodiment is executed by the learning device 1 shown in FIG. This learning device 1 is of a personal computer type, and includes a display 1a, a device main body 1b, an input interface 1c, and the like. The apparatus main body 1b includes a storage such as an HDD, a processor, memories (RAM, ROM, etc.), etc. (none of which are shown).

A database is stored in the memory of the device body 1b. As shown in FIGS. 2 and 3, this database contains the names of various formulations, the characteristic amounts (physical property values, etc.) of the formulations, and the measured values of objective variables (tactile sensations).

In this database, various formulations include single-phase and two-phase formulations such as lotions, creams, and serums. Here, one-phase formulations are formulations such as lotions composed of an aqueous solution or oil, and two-phase formulations are formulations such as emulsion formulations, that is, formulations such as creams composed of a mixture of water and oil. is.

In addition, the feature values include physical property values of various formulations. Specifically, in addition to the viscosity, dry residual ratio, specific gravity and thermal conductivity of various formulations, although not shown, specific gravity, dry residual ratio, shear stress, contact angle, thermal conductivity, peel force and friction coefficient It is included. Here, the dry residual ratio indicates the weight ratio of the preparation before and after drying, and the peel force indicates the peel force per unit area of the preparation.

In addition, when the formulation is an emulsified formulation such as cream, that is, a two-phase formulation, the feature amounts include the blending ratio of raw materials (pure water, 1,3-BG, petrolatum, etc.) and the manufacturing method (stirring method, stirring time, stirring speed and emulsification temperature). These blending ratios are set to values normalized based on the maximum blending values of various raw materials, and these maximum blending values are calculated from past actual values of the present applicant.

On the other hand, the actual measured value of the feel, which is the objective variable, includes data for the following 18 items (A1) to (A18) (only part of which is shown in FIG. 3). These 18 items of tactile measurements were obtained by the CATA (Check-All-That-Apply) method.

(A1) Soft, (A2) Hard, (A3) Smooth, (A4) Moist, (A5) Smooth, (A6) Rough, (A7) Warm, (A8) Cold, (A9) Sticky, (A10) Thick , (A11) firm, (A12) stretchy, (A13) familiar, (A14) penetrating, (A15) filmy, (A16) oily, (A17) fresh, (A18) plump .

The items (A1) to (A10) above are items of general feel based on the tactile dimension, and the items (A11) to (A18) are items of feel frequently used in the preparation of cosmetic formulations. item. In the case of this embodiment, the database is configured as described above, and the reason for this will be described later.

On the other hand, application software for executing learning processing, which will be described later, is installed in the storage of the device main body 1b. The input interface 1c is composed of a keyboard, a mouse, and the like for operating the learning device 1. FIG.

Next, the contents of the learning process executed by the learning device 1 of this embodiment will be described with reference to FIG. In this learning process, the prediction model is learned using the database described above, as described below. This predictive model is for predicting the aforementioned 18 items of feel of the formulation.

As shown in FIG. 4, in this learning process, first, it is determined whether or not the formulation for which learning is to be performed this time is of the one-phase system (FIG. 4/STEP 1). When this determination is affirmative (FIG. 4/STEP1 . . . YES) and the formulation is of the one-phase system, the learning process for the one-phase system is executed (FIG. 4/STEP2). After that, the process ends. Details of the learning process for the one-phase system will be described later.

On the other hand, when the above determination is negative (FIG. 4/STEP 1 . . . NO) and the formulation is of the two-phase system, the learning process for the two-phase system is executed (FIG. 4/STEP 3). After that, the process ends. Details of the learning process for the two-phase system will be described later.

Next, the contents of the learning process for the one-phase system described above will be described with reference to FIG. As shown in the figure, first, a prediction model determination process is executed (FIG. 5/STEP 10). In this determination process, the prediction model targeted for the current learning process is determined to be one of the 18 prediction models that predict the 18 items of feeling (A1) to (A18) described above as objective variables. .

This determination is specifically performed based on the operation result of the input interface 1c by an operator (not shown). In the following description, the feel of the objective variable of the prediction model that is the target of the current learning process is referred to as the "learning target feel".

Next, a teaching data reading process is executed (FIG. 5/STEP 11). Specifically, the actually measured value of the feel of the object to be learned determined by the determination process is read from the database described above as teacher data.

Next, it is determined whether or not the feel to be learned is the feel of pattern 1 (FIG. 5/STEP 12). The feel of pattern 1 is the type of feel that maximizes the prediction accuracy of the prediction model by executing the learning method of pattern 1 below. corresponds to

Here, the learning method of pattern 1 is to acquire only the physical property values of the formulation from the database as the feature amount of pattern 1 (first acquisition method), and execute the learning of the prediction model using the feature amount of this pattern 1. method. In the case of pattern 1 feeling, it has been confirmed by experiments described later by the present applicant that the prediction accuracy of the prediction model is most improved by using this pattern 1 learning method.

When this determination is affirmative (FIG. 5/STEP 12 . . . YES) and the feel of the object to be learned is the feel of pattern 1, the feature amount of pattern 1 is read from the database (FIG. 5/STEP 13).

On the other hand, when the above determination is negative (FIG. 5/STEP 12 . . . NO) and the feel to be learned is not pattern 1, it is determined whether or not the feel to be learned is pattern 2 (FIG. 5). / STEP 14).

The feel of pattern 2 is a type of feel that maximizes the prediction accuracy of the prediction model by executing the learning method of pattern 2 below. It corresponds to the feeling of pattern 2.

Here, the learning method of pattern 2 is the first selection data selected by the method of steps (B2) to (B5) described later from the blending ratio of the raw materials of the formulation, and the physical property values of the formulation. 2 (second acquisition method), and the prediction model is learned using the feature amount of pattern 2. Experiments conducted by the present applicant have confirmed that the prediction accuracy of the prediction model is most improved by using the pattern 2 learning method in the case of pattern 2 feeling.

When this determination is affirmative (FIG. 5/STEP 14 . . . YES) and the feel of the object to be learned is the feel of pattern 2, the feature amount of pattern 2 is read from the database (FIG. 5/STEP 15). Next, the processing proceeds to the learning data creation process (FIG. 5/STEP 17), which will be described later.

On the other hand, when the above determination is negative (FIG. 5/STEP 14 . . . NO) and the feel to be learned is not pattern 2, it is assumed that the feel to be learned is pattern 3, and the feature amount of pattern 3 is obtained from the database. It is read out (FIG. 5/STEP16).

This pattern 3 feel is a type of feel that maximizes the prediction accuracy of the prediction model by executing the following pattern 3 learning method. "Sticky" corresponds to the feeling of pattern 3.

Here, the learning method of pattern 3 means that the second selection data selected from the blending ratio of raw materials of the formulation and the physical property values of the formulation by the method of steps (B2) to (B5) described later are used as the characteristics of pattern 3. This is a method of acquiring as a quantity (third acquisition method) and executing learning of a prediction model using the feature quantity of this pattern 3 . In the case of the feel of pattern 3, it has been confirmed by experiments described later by the present applicant that the prediction accuracy of the prediction model is most improved by using the learning method of pattern 3.

After any one of the feature amount readout processes of patterns 1 to 3 is executed, the learning data creation process is executed (FIG. 5/STEP 17). In this creation process, learning data is created so as to include any of the feature values of patterns 1 to 3 read out as described above and the teacher data (actually measured values of feel) described above.

Next, it is determined whether or not the feel of the learning object is the feel of the classification model (FIG. 5/STEP 18). Here, the feeling of the classification model is the feeling that the prediction accuracy can be improved when the classification model is used as the prediction model compared to when the regression model is used as the prediction model. Specifically, the two tactile sensations of "gritty" and "warm" correspond to the tactile sensations of the classification model.

When this determination is affirmative (FIG. 5/STEP 18 . . . YES) and the feel of the learning object is the feel of the classification model, the classification model learning process is executed (FIG. 5/STEP 19). Specifically, learning of the model parameters of the classification model is executed using the learning data described above.

In this case, a model created by applying the XGBoost algorithm is used as the classification model. Note that in the present embodiment, the XGBoost algorithm corresponds to a supervised machine learning algorithm. After the classification model learning process is executed as described above, this process ends.

On the other hand, when the above determination is negative (FIG. 5/STEP 18 . . . NO) and the feel of the learning target is not the feel of the classification model, it is determined whether or not the feel of the learning target is the feel of the regression model (FIG. 5). / STEP 20). Here, the tactile sensation of the regression model is the tactile sensation that ensures high prediction accuracy when the regression model is used as the prediction model. Specifically, tactile sensations such as "stretching", "fresh" and "moisturizing" correspond to the tactile sensations of the regression model.

When this determination is affirmative (FIG. 5/STEP 20 . . . YES) and the feel of the object to be learned is the feel of the regression model, regression model learning processing is executed (FIG. 5/STEP 21). Specifically, learning of the model parameters of the regression model is executed using the learning data described above.

In this case, as the regression model, one created by applying the XGBoost algorithm is used. Note that in the present embodiment, the XGBoost algorithm corresponds to a supervised machine learning algorithm. After the regression model learning process is executed as described above, this process ends.

On the other hand, when the above determination is negative (FIG. 5/STEP 20 . . . NO) and the feel of the object to be learned is not the feel of the regression model, MP learning processing is executed (FIG. 5/STEP 22). In this MP learning process, learning of a prediction model is executed using a learning algorithm to which a multivariate prediction technique is applied, as described below.

In the following, an example of learning a predictive model for the feel of "a feeling of being covered" will be described. First, as learning step 1, the following learning processes (1A) to (1F) are executed.

(1A) First, using the learning data described above, learning of a predictive model for a feeling of “hardness” is executed by a random forest algorithm. In this embodiment, the random forest algorithm corresponds to the supervised machine learning algorithm.

(1B) Next, learning data 1B is created by adding the score (prediction value) of the feeling of “hardness” in the prediction model for which the learning process of (1A) has been completed as a feature amount to the learning data described above. Then, using this learning data 1B, learning of a prediction model with a feeling of “firmness” is executed by a random forest algorithm.

(1C) Next, learning data 1C is created by adding, as a feature amount, the score of the feeling of "firmness" in the prediction model for which the learning process of (1B) has been completed to the learning data 1B described above. , and using this learning data 1C, the learning of the predictive model of the feeling of “fresh and fresh” is executed by the random forest algorithm.

(1D) Further, learning data 1D is created by adding, as a feature amount, the score of the feeling of "freshness" in the prediction model for which the learning process of (1C) has been completed to the learning data 1C described above, and this learning Using the data 1D, a random forest algorithm is used to train a predictive model for the feeling of “penetration”.

(1E) Next, learning data 1E is created by adding, as a feature amount, the score of the feeling of "permeability" in the prediction model for which the learning process of (1D) has been completed to the learning data 1D described above. , and using this learning data 1E, the random forest algorithm learns a predictive model for the feeling of “cold”.

(1F) Next, learning data 1F is created by adding the score of the feeling of "cold" in the prediction model that has completed the learning of (1E) above as a feature amount to the learning data 1E described above, and this learning Using the data 1F, the random forest algorithm is used to learn a predictive model of the feeling of "a feeling of being covered". A predicted value 1 is set to the predicted value of the touch feeling of “there is a feeling of being covered” by the predicted model after learning.

Following learning step 1 above, as learning step 2, the following learning processes (2A) to (2G) are executed.

(2A) First, using the learning data described above, the random forest algorithm is used to perform learning of a predictive model for the feel of “smooth”.

(2B) Next, learning data 2B is created by adding, as a feature amount, the score of the feeling of “smooth” in the prediction model for which the learning process of (2A) has been completed to the learning data described above, and this learning data 2B, a random forest algorithm is used to train a predictive model of "hard" feel.

(2C) Next, learning data 2C is created by adding, as a feature amount, the score of the feeling of "hardness" in the prediction model for which the learning process of (2B) has been completed to the learning data 2B described above, Using this learning data 2C, the random forest algorithm is used to perform learning of a predictive model with a feeling of “fresh and fresh”.

(2D) Furthermore, learning data 2D is created by adding, as a feature amount, the score of the feeling of "fresh and fresh" in the prediction model for which the learning process of (2C) has been completed to the learning data 2C described above, and this learning Using the data 2D, a random forest algorithm is used to train a predictive model for feeling “cold”.

(2E) Next, learning data 2E is created by adding the score of "cold" feeling in the prediction model for which the learning process of (2D) has ended as a feature amount to the learning data 2D described above, and this learning Using data 2E, a random forest algorithm is used to train a predictive model for the feeling of “fluffy”.

(2F) Next, learning data 2F is created by adding, as a feature amount, the score of the feeling of “fluffy” in the prediction model that has completed the learning of (2E) above to the learning data 2E described above, and this learning Using the data 2F, the learning of the predictive model of the feeling of "firmness" is executed by the random forest algorithm.

(2G) Furthermore, learning data 2G is created by adding, as a feature amount, the score of the feeling of "firmness" in the prediction model for which the learning of (2F) above has been completed to the learning data 2F described above. Using the learning data 2G, learning of a predictive model of the feeling of "a feeling of being covered" is executed by a random forest algorithm. Then, a predicted value 2 is set to the predicted value of the touch feeling of “there is a feeling of being covered” by the prediction model after learning.

Following learning step 2 above, learning steps similar to learning steps 1 and 2 are executed a plurality of times to calculate N predicted values 1 to N (N is plural). Ultimately, the average value of these N prediction values 1 to N is determined as the prediction value of the "feeling covered" by the prediction model. After the MP learning process is executed as described above, this process ends.

In this embodiment, as described above, the learning process for the one-phase system (FIG. 4/STEP 2) is executed. On the other hand, although not shown, the learning process for the two-phase system (FIG. 4/STEP 3) is executed in the same manner as the learning process for the one-phase system shown in FIG. 5 except for the following points.

That is, in the learning process for the two-phase system, in the process corresponding to STEP 13 in FIG. 5, only the physical property values of the formulation are acquired as the feature amount of pattern 1 (first acquisition method).

Further, in the processing corresponding to STEP 15 in FIG. is acquired as a feature quantity of (second acquisition method).

Furthermore, in the processing corresponding to STEP 16 in FIG. It is acquired as a quantity (third acquisition method).

As described above, in the learning process for the two-phase system, the manufacturing method of the formulation is used as the feature quantity of the patterns 2 and 3. This is because formulations for two-phase systems are emulsified formulations, and the production method has a large effect on the texture.

Next, the reason and principle for using the above learning method in this embodiment will be described. First, the applicant conducted an experiment to learn a predictive model of the feel of a formulation by the same method as in Japanese Patent Application No. 2020-126447, as in steps (B1) to (B6) below.

(B1) That is, as a database of feature amounts, a database including the compounding ratio of various raw materials, preparation method, stirring method, stirring time, stirring rotation speed, phase state, and the like was created. In the case of this database, the compounding ratios of various raw materials were normalized, raw materials with little effect on texture were omitted, and feature quantities that cause multicollinearity were omitted.

(B2) Furthermore, by a random forest algorithm, a prediction model (supervised machine learning model) is created by learning all feature quantities, the importance of the feature quantity is calculated using this prediction model, and the feature quantity is were selected in order from highest to lowest. In this case, the importance of the feature quantity is calculated according to the frequency of use of the feature quantity at the branch point of the decision tree, and the higher the frequency of use of the feature quantity, the higher the importance of the feature quantity is calculated.

(B3) Also, by network analysis, the relationship between the feature amount and the objective variable (tactile score) was calculated, and the feature amount was selected in descending order of the relationship with the objective variable.

(B4) In addition, the Relief algorithm was used to calculate the degree of contribution of the feature quantity to the measured value of the target variable, and the feature quantity was selected in descending order of contribution.

(B5) Next, the average value of the ranks of the three types of feature quantities selected as described above is calculated, and when the feature quantity with the highest average rank is set to be the first rank, from this first rank to the predetermined rank was selected.

(B6) Then, the XGBoost algorithm was used to create a prediction model in which feature amounts from the first to the predetermined ranks were learned.

Experimental results show that the prediction model created by the above learning method has low prediction accuracy. Therefore, when learning a predictive model for the texture of a formulation, the applicant believes that the learning accuracy of the prediction model is improved by using the physical property values of the formulation as feature quantities instead of the blending ratio and manufacturing method of various raw materials. Assuming that this is the case, we conducted an experiment to learn a prediction model using the method described below. In this case, the prediction model was created as a regression model.

That is, the applicant created a database as shown in FIGS. 2 and 3 described above, and using this database, learned a prediction model by the XGBoost algorithm.

At that time, in the case of a one-phase formulation, learning was performed using the feature values of patterns 1 to 3 described above. Hereinafter, learning of a prediction model using the feature amounts of patterns 1 to 3 will be referred to as "learning patterns 1 to 3").

That is, in the learning pattern 1, the prediction model was learned using only the physical property values of the formulation as feature amounts.

In addition, in learning pattern 2, the first selection data is selected from the blending ratio of the raw materials by the steps (B2) to (B5) described above, and these first selection data and the physical property values of the formulation are used as feature amounts. , training of the prediction model was carried out.

In this case, steps (B2)-(B5) correspond to the first to fourth selection steps, the network analysis corresponds to the first filtering process, and the Relief algorithm corresponds to the second filtering process.

Furthermore, in learning pattern 3, the second selection data is selected from the physical properties of the formulation and the blending ratio of the raw materials by the steps (B2) to (B5) described above, and these second selection data are used as feature amounts. , training of the prediction model was carried out.

In this case, steps (B2)-(B5) correspond to the fifth to eighth selection steps, the network analysis corresponds to the third filtering process, and the Relief algorithm corresponds to the fourth filtering process.

On the other hand, in the case of two-phase formulations, learning was performed using learning patterns 1 to 3 below. That is, in the learning pattern 1, the prediction model was learned using only the physical property values of the formulation as feature amounts.

Further, in learning pattern 2, the first selection data is selected by the above-described steps (B2) to (B5) from the formulation manufacturing method and the raw material blending ratio, and these first selection data and the physical property values of the formulation are selected. A prediction model was trained using the feature values.

In this case, steps (B2)-(B5) correspond to the first to fourth selection steps, the network analysis corresponds to the first filtering process, and the Relief algorithm corresponds to the second filtering process.

Furthermore, in learning pattern 3, the second selection data is selected from the physical properties of the formulation and the blending ratio of the raw materials by the steps (B2) to (B5) described above, and these second selection data are used as feature amounts. , training of the prediction model was carried out.

In this case, steps (B2)-(B5) correspond to the fifth to eighth selection steps, the network analysis corresponds to the third filtering process, and the Relief algorithm corresponds to the fourth filtering process.

As described above, as a result of learning the texture prediction models for single-phase and two-phase formulations, the scores for multiple textures shown in FIG. %) could be secured.

In the experimental results in the figure, the phase state "AQ" indicates that the formulation is a one-phase system, and the phase state "EM" indicates that the formulation is a two-phase system. As shown in the figure, LOO (Leave One Out) method and 5-fold cross-validation method (5-fold CV) are used as accuracy evaluation methods, and NRMSE (standard root mean square error: Normalized Root Mean Square Error) and coefficient of determination ^R2 were used.

Here, in the case of the feeling of “spreading” and “freshness” in the one-phase formulation, learning results shown in FIG. 7 were obtained when learning patterns 1 to 3 were learned. As shown in the figure, in the case of the feeling of "stretching", as is clear from the comparison of the learning results of learning patterns 1 to 3, it can be seen that NRMSE does not change much even if the learning pattern is changed.

On the other hand, the coefficient of determination ^R2 is clearly improved in the learning result of learning pattern 2 compared to learning patterns 1 and 3, and it can be seen that the prediction accuracy is improved. That is, when learning a prediction model (regression model) that predicts the feeling of “spreading” in a one-phase formulation, learning with learning pattern 2 improves the learning accuracy of the prediction model the most. It turned out to do.

Further, as shown in FIG. 7, when the learning results of learning patterns 1 to 3 for the feel of "fresh and fresh" are compared, it can be seen that NRMSE does not change much even if the learning pattern is changed. On the other hand, the coefficient of determination ^R2 of the learning result of the learning pattern 3 is clearly improved as compared with the learning patterns 1 and 2, and it can be seen that the prediction accuracy is improved.

That is, when learning a predictive model (regression model) that predicts the feeling of "freshness" in a one-phase formulation, the learning accuracy of the predictive model is most improved by learning pattern 3. It turned out to do.

Based on the above experimental results, "stretching", "fresh", "moist", "smooth", "sticky", "simmering", "firm", "permeating" and For predictive models predicting the feeling of “cold” and the feeling of “hard”, “smooth”, “conformable”, “oily”, “moist” and “spready” in two-phase formulations, the regression It has been found that high prediction accuracy can be ensured by learning with the learning pattern shown in FIG. 6 using the model.

On the other hand, in the case of "rough" and "warm" sensations, a regression model was created by the XGBoost algorithm for both one-phase and two-phase formulations, and learning of this regression model was performed, as shown in FIG. A good learning result was obtained.

As is clear from the results of this study, in the case of "rough" and "warm" sensations, both the one-phase and two-phase formulations had low values of determination coefficient ^R2 (0.01 to 0.01). 36), and it can be seen that high prediction accuracy cannot be obtained.

Therefore, instead of the regression model, a classification model for classifying whether or not there is a feeling of "rough" and "warm" was created by the XGBoost algorithm, and when this classification model was learned, the result was as shown in FIG. A learning result was obtained.

As is clear from this learning result, in the case of "rough" and "warm" touches, both the one-phase and two-phase formulations have a Kappa Score of 1, indicating high prediction accuracy. It turns out that it is obtained. That is, in both one-phase and two-phase formulations, when learning a prediction model for the feeling of “rough” and “warm”, using a classification model has higher prediction accuracy than using a regression model. It has been found that it is possible to ensure

Also, when the formulation is a cosmetic, it is well known that several textures in the formulation are correlated with each other. For example, there is a phenomenon that a formulation having a feeling of "thickness" is likely to have a feeling of "moistness".

Therefore, as a result of learning the texture prediction model for single-phase and two-phase formulations using the multivariate prediction method as described above, a predetermined prediction accuracy ( It was possible to secure the prediction error within ±10% of the measured value).

Here, FIG. 11 shows the learning results when a multivariate prediction method is performed when learning a prediction model (regression model) for the feel of “having a coating feeling” and “getting used” in a single-phase formulation. , shows the learning results when the multivariate prediction method was not implemented (that is, when only learning by the XGBoost algorithm was implemented) for comparison.

As is clear from the figure, when the multivariate prediction method is implemented, the coefficient of determination ^R2 clearly increases compared to when the multivariate prediction method is not implemented, and the prediction accuracy is greatly improved. It turns out that you are.

Based on the above experimental results, a prediction model for predicting the feeling of "filming", "oily", "smooth", "fluffy", "hard" and "fitting" in single-phase formulations. In this case, it was found that high prediction accuracy can be ensured by learning with the learning pattern shown in FIG. 10 while using the multivariate prediction method.

Similarly, in two-phase formulations, "film feeling", "fresh", "cold", "hard", "permeable", "sticky", "smooth", "fluffy", In the case of a prediction model that predicts the feeling of “firmness” and “smoothness”, it was found that high prediction accuracy can be ensured by learning with the learning pattern shown in FIG. 10 while using the multivariate prediction method. .

In this embodiment, based on the above experimental results, the learning process for the one-phase system is executed as shown in FIG. is executed as previously described.

At that time, one of the feature amounts of patterns 1 to 3 is selected according to the type of feeling, and furthermore, the learning process of the prediction model is performed using the feature amount thus selected. That is, as the prediction model learning process, one of the classification model learning process, the regression model learning process, and the MP learning process is executed according to the type of touch so that the prediction accuracy after learning is maximized. As a result, high prediction accuracy can be ensured in the prediction model for predicting the 18 types of touch described above.

In addition, the embodiment is an example in which the CATA method is used to obtain the actual measurement value of the feel of the formulation. can be obtained.

In addition, the embodiment is an example of using 18 types of feel data of (A1) to (A18) as actual measurement values of the feel of the formulation, but instead of this, 17 types or less or 19 types or more of feel are used. data may be used as actual measurements of the feel of the formulation.

Furthermore, the embodiment is an example of using the XGBoost algorithm as a supervised machine learning algorithm in the classification model learning process (FIG. 5/STEP19) and the regression model learning process (FIG. 5/STEP21). , a neural network or a support vector machine may be used as a supervised machine learning algorithm.

On the other hand, the embodiment is an example of using a random forest algorithm as a supervised machine learning algorithm in the MP learning process (FIG. 5 / STEP 22), but instead of this, an XGBoost algorithm, a neural network, a support vector machine, etc. may be used.

Further, the embodiment is an example of using the method of selecting the first selection data in steps (B2) to (B5) in the second acquisition method, but instead of this, the following method may be used. . That is, any two of the processes (B2) to (B4) are executed, the average value of the ranks of the two types of feature quantities selected by these two processes is calculated, and the feature quantity with the highest average rank is calculated. is ranked first, the feature quantities from this first rank to a predetermined rank may be selected as the first selection data.

Furthermore, the embodiment is an example of using the method of selecting the second selection data in steps (B2) to (B5) in the third acquisition method, but instead of this, the following method may be used. . That is, any two of the processes (B2) to (B4) are executed, the average value of the ranks of the two types of feature quantities selected by these two processes is calculated, and the feature quantity with the highest average rank is calculated. may be selected as the second selection data.

Furthermore, the embodiment is an example of using a network analysis technique as the first filtering process and the third filtering process, but the first filtering process and the third filtering process of the present invention are not limited to this, and the objective variable and Any device can be used as long as it can acquire the correlation of feature amounts. For example, a correlation analysis algorithm, a variance analysis algorithm, a chi-square score, a Fisher score, or anova may be used as the first filtering process and the third filtering process. Also, different filtering processes may be executed as the first filtering process and the third filtering process.

On the other hand, the embodiment is an example using the Relief algorithm as the second filtering process and the fourth filtering process, but the second filtering process and the fourth filtering process of the present invention are not limited to this, and feature Anything that can acquire the degree of contribution of the amount is acceptable. For example, a correlation analysis algorithm, a variance analysis algorithm, a chi-square score, a Fisher score, or anova may be used as the second filtering process and the fourth filtering process. Also, different filtering processes may be executed as the second filtering process and the fourth filtering process.

Moreover, although the embodiment is an example of using a personal computer type as the learning device, instead of this, a server, a cloud server, or the like may be used as the learning device.

1 learning device

Claims (7)

A model parameter learning method for learning, by a learning device, model parameters of a prediction model that predicts the feel of a formulation prepared by blending a plurality of raw materials as an objective variable, comprising:
The learning device
a teacher data acquisition step of acquiring the measured value of the objective variable as teacher data;
A feature amount acquisition step of acquiring a feature amount including physical property values of the formulation;
a learning data creation step of creating learning data using the feature amount and the teacher data;
A learning step of performing learning of model parameters of the prediction model using a learning algorithm to which a supervised machine learning algorithm is applied and the learning data;
A model parameter learning method characterized by executing In the model parameter learning method according to claim 1,
The formulation is a one-phase formulation,
In the feature amount acquisition step, a first acquisition method for acquiring only the physical property value of the formulation as the feature amount, first selection data selected from among the blending ratios of the plurality of raw materials and the physical property value of the formulation as the feature amount, and second selection data selected from the mixing ratio of the plurality of raw materials and the physical property values of the formulation as the feature amount. any one of the techniques is selected according to the type of touch,
In the learning step, as the prediction model, one of a regression model and a classification model is selected according to the type of the feeling, and as the learning algorithm, the supervised machine learning algorithm is a multivariate prediction method. and the supervised machine learning algorithm are selected according to the type of the tactile sensation. In the model parameter learning method according to claim 2,
In the second acquisition method,
Using the feature amount that is the blending ratio of the plurality of raw materials and the teacher data as learning data, the model parameters of the machine learning model are learned by a supervised machine learning algorithm, and the machine learning model after the learning is used. a first selection step of calculating the importance of the feature amount and selecting the feature amount in descending order of importance;
By applying a first filtering process to the feature amount that is the mixture ratio of the plurality of raw materials, the correlation between the objective variable and the feature amount is obtained, and the feature amount is compared with the objective variable. A second selection step of selecting in order from the highest correlation of
By applying a second filtering process different from the first filtering process to the feature amount that is the mixing ratio of the plurality of raw materials, the contribution of the feature amount to the objective variable is obtained, and the feature amount is obtained. a third selection step of selecting quantities in descending order of contribution;
calculating the average value of the ranks of the feature values selected in the first to third selection steps, and determining the feature value with the highest average value as the first rank from the first rank to a predetermined rank; a fourth selection step of selecting the feature amount as the first selection data;
A model parameter learning method characterized in that is executed. In the model parameter learning method according to claim 2 or 3,
In the third acquisition method,
Using the blending ratio of the plurality of raw materials, the feature amount that is the physical property value, and the teacher data as learning data, learning the model parameters of the machine learning model by a supervised machine learning algorithm, and the machine after the learning A fifth selection step of calculating the importance of the feature amount using the learning model and selecting the feature amount in descending order of importance;
By applying a third filter process to the feature amount, which is the mixture ratio of the plurality of raw materials and the physical property value, the correlation between the objective variable and the feature amount is obtained, and the feature amount is obtained. a sixth selection step of selecting in order from the one with the highest correlation with the objective variable;
Obtaining the degree of contribution of the feature amount to the objective variable by applying a fourth filter process different from the third filter process to the feature amount, which is the mixture ratio of the plurality of raw materials and the physical property value together with a seventh selection step of selecting the feature quantity in order from the one with the highest contribution;
calculating the average value of the ranks of the feature amounts selected in the fifth to seventh selection steps, and determining the feature amount with the highest average value as the first rank from the first rank to a predetermined rank; an eighth selection step of selecting the feature quantity as the second selection data;
A model parameter learning method characterized in that is executed. In the model parameter learning method according to claim 1,
The formulation is an emulsified formulation,
In the feature amount acquisition step, the first selection data selected from a first acquisition method for acquiring only the physical property value of the formulation as the feature amount, the mixing ratio of the plurality of raw materials, and the manufacturing method of the formulation, and the A second acquisition method for acquiring the physical property values of the formulation as the feature amount, and second selection data selected from the compounding ratio of the plurality of raw materials, the physical property values of the formulation, and the manufacturing method of the formulation. as the feature quantity, any one of the third acquisition methods is executed according to the type of the touch,
In the learning step, as the prediction model, one of a regression model and a classification model is selected according to the type of the feeling, and as the learning algorithm, the supervised machine learning algorithm is a multivariate prediction method. and the supervised machine learning algorithm are selected according to the type of the tactile sensation. In the model parameter learning method according to claim 5,
In the second acquisition method,
A machine learning model is learned by a supervised machine learning algorithm using the blending ratio of the plurality of raw materials, the feature amount that is the manufacturing method of the formulation, and the teacher data as learning data, and the machine learning model after the learning. A first selection step of calculating the importance of the feature amount using and selecting the feature amount in descending order of importance;
By applying a first filtering process to the feature amounts that are the compounding ratio of the plurality of raw materials and the manufacturing method of the formulation, the feature amounts are selected in descending order of correlation with the objective variable. a second selection step;
Obtaining the degree of contribution of the feature amount to the objective variable by applying a second filter process different from the first filter process to the feature amount, which is the compounding ratio of the plurality of raw materials and the manufacturing method of the formulation. and a third selection step of selecting the feature quantity in order from the one with the highest contribution;
calculating the average value of the ranks of the feature values selected in the first to third selection steps, and determining the feature value with the highest average value as the first rank from the first rank to a predetermined rank; a fourth selection step of selecting the feature amount as the first selection data;
A model parameter learning method characterized in that is executed. In the model parameter learning method according to claim 5 or 6,
In the third acquisition method,
Using the blending ratio of the plurality of raw materials, the physical property value, the feature amount that is the manufacturing method of the formulation, and the teacher data as learning data, machine learning model learning is performed by a supervised machine learning algorithm, and after the learning A fifth selection step of calculating the importance of the feature amount using the machine learning model and selecting the feature amount in descending order of importance;
By applying a third filter to the feature amount that is the compounding ratio of the plurality of raw materials, the physical property value, and the manufacturing method of the formulation, the correlation between the objective variable and the feature amount is obtained, a sixth selection step of selecting the feature quantity in descending order of correlation with the objective variable;
By applying a fourth filter different from the third filter to the feature amount, which is the compounding ratio of the plurality of raw materials, the physical property value, and the manufacturing method of the formulation, the degree of contribution of the feature amount to the objective variable and a seventh selection step of selecting the feature quantities below the first rank when the feature quantity with the highest contribution is ranked first, in order from the one with the highest contribution;
calculating the average value of the ranks of the feature amounts selected in the fifth to seventh selection steps, and determining the feature amount with the highest average value as the first rank from the first rank to a predetermined rank; an eighth selection step of selecting the feature quantity as the second selection data;
A model parameter learning method characterized in that is executed.

Images