JP2019144022A - Characteristic prediction method, characteristic prediction program and characteristic prediction device for beverage - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a beverage characteristic prediction method, a characteristic prediction program and a characteristic prediction device for estimating the characteristics of a beverage predicted from input trial brewing conditions with high accuracy. SOLUTION: In the method for predicting the characteristics of a beverage according to the present invention, a step of generating a trial brewing sample using a setting item in the trial brewing condition of the beverage as an explanatory variable and the trial brewing sample are input to a plurality of prediction models. , The step of obtaining the estimated value of the component value of the beverage for each prediction model, and the trial brewing sample are input to the error determination model corresponding to the prediction model, and the probability within the margin of error is calculated for each prediction model. The explanatory variables related to the trial brewing sample and the plurality of probabilities within the margin of error are input to the step, the method selection model, and the component values of the beverage under the trial brewing conditions based on the output values of the method selection model. It is characterized in that the computer executes the step of selecting the estimated value, which is presented as. [Selection diagram] Fig. 1

Description

  The present invention relates to a beverage characteristic prediction method, a characteristic prediction program, and a characteristic prediction apparatus.

  In recent years, consumer preferences have changed in a short period of time, and competition in the beverage market has intensified. Under such circumstances, it is necessary to timely introduce new malt alcoholic beverages (for example, beer and sparkling liquor) and beer-taste beverages that meet consumer demands and have high competitiveness to the market. . In the development of new products, engineers can try to obtain beverages with characteristics such as taste, aroma, color, and alcohol content that meet consumer demand while referring to data such as recipes accumulated in the past. Repeat the brew. It is up to each engineer to decide how to use the past data to determine the brewing conditions. In general, engineers use the senses and know-how cultivated through development experience to determine the conditions for brewing, but it is difficult to actually pass on such senses and know-how as formal knowledge and organizational knowledge. When brewing, it is necessary to use many materials, equipment, and man-hours. As the number of trials increases, not only time but also resource consumption increases.

  Therefore, in order to realize trial production of efficient malt alcoholic beverages and beer-taste beverages, it is conceivable to develop a technique for analyzing characteristics accumulated in the past and predicting characteristics. In the field of food, techniques for predicting the quality of food from past data have been developed. Patent Document 1 discloses a technique for predicting the growth time of microorganisms in food and predicting the shelf life of food. Patent Document 2 discloses a technique for predicting the filterability of a pre-filtration liquid that is an intermediate product of a beer-taste beverage. Patent Document 3 discloses a technique for predicting the flavor stability of a malt alcohol fermented beverage from the wort's nonenal potential.

Japanese Patent No. 6121035 Japanese Patent No. 5894027 Japanese Patent No. 4191390

Clear, John G, et al. K *: An Instance-based Learner Usage an Entropic Distance Measurement. ICML '95 Proceedings of Twelve International Conference on Machine Learning, July, 1995 Wang, Yong, et al. Inducting Model Trees for Continuous Classes, European Conference on Machine Learning, 1997

  The method of Patent Document 1 is intended to predict the growth of microorganisms in food, and does not predict the characteristics of the produced food. In the method of Patent Document 2, the characteristics of the intermediate product in the beverage production process can be estimated, but the characteristics of the beverage finally produced by brewing are not predicted. The method of Patent Document 3 can predict some characteristics of a beverage, but does not predict the general characteristics of a beverage by using a plurality of variables related to brewing conditions as input. That is, in the prior art, the number of conditions that can be input as variables to the model and the properties of the beverage that can be predicted are limited, and it is difficult to directly apply the results to the prediction when the beverage is brewed.

  Based on these issues, in the trial production of malt alcoholic beverages and beer-taste beverages, the data accumulated in the past is analyzed, and the characteristics of the beverages brewed (trial brew) are accurately predicted from the input brewing conditions. Development of technology to do this is desired. The present invention provides a beverage characteristic prediction method, a characteristic prediction program, and a characteristic prediction apparatus that estimate a characteristic of a beverage predicted from input brewing conditions with high accuracy based on past data.

  The method for predicting beverage characteristics according to the present invention includes a step of generating a tasting sample with setting items in beverage tasting conditions as explanatory variables, the tasting sample being input to a plurality of prediction models, and the prediction model A step of obtaining an estimated value of the component value of the beverage every time, a step of inputting the brew sample into an error determination model corresponding to the prediction model, and calculating a probability within tolerance for each prediction model, and a method An explanatory variable related to the tasting sample and a plurality of probabilities within the allowable error are input to the selection model, and are presented as the component values in the brewing conditions of the beverage based on the output value of the method selection model. The computer executes the step of selecting the estimated value.

The figure which shows the structural example of the whole characteristic prediction apparatus which concerns on 1st Embodiment. The figure which showed the example of the cross-validation in a learning process. The figure which showed the example of the cross-validation in a learning process. The figure which showed the example of the brewing data which stored the past brewing example. The figure which showed the example of the brewing data which stored the past brewing example. The figure which showed the example of the learning data produced | generated from the brewing data. The figure which showed the example of the variable used for every component value to predict. The figure which showed the example of the determination table about a certain analysis value. The figure which showed the example of the estimated value comparison table about a certain analysis value. The figure which showed the example of the model data in the case of using M5 '. The figure which showed the example of the model data in the case of using M5 '. The figure which showed the example of the sigmoid function. The flowchart which showed the learning process. The flowchart which showed the learning process. The flowchart which showed the learning process. The flowchart which showed the prediction process of the component value. The flowchart which showed the prediction process of the component value. The figure which showed the example of the extraction method of the related example from feature space. The figure which showed the example of the tasting condition input screen. The figure which showed the 1st example of the prediction result display screen. The figure which showed the 2nd example of the prediction result display screen. The figure which shows the example of the hardware constitutions which concern on a characteristic prediction apparatus.

  Hereinafter, embodiments of the present invention will be described with reference to the drawings. In the drawings, the same components are denoted by the same reference numerals, and description thereof will be omitted as appropriate.

(First embodiment)
FIG. 1 is a diagram illustrating a configuration example of the entire characteristic prediction apparatus according to the first embodiment.

  The characteristic prediction apparatus 1 of FIG. 1 includes a storage unit 2, a learning unit 3, a preprocessing unit 4, a display unit 5, and an operation unit 6. First, an outline of processing executed by the characteristic prediction apparatus 1 will be described.

  The characteristic prediction apparatus 1 outputs the characteristic of a beverage expected when brewing is performed under arbitrary conditions using brewing data including information relating to past brewing cases. The past brewing example includes information on at least one setting item (recipe) of the raw material or the brewing process when the beverage was brewed (trial brewing) in the past, and the component value of the generated beverage. Examples of beverages include wort used in the production of beverages that contain malt as a raw material (malt beverage), fermented beverages that are fermented in the production process, non-fermented non-alcoholic beverages, and beer-taste beverages (with an alcohol content of less than 1%) Beer-flavored sparkling carbonated beverages).

  Examples of fermented beverages include beer, happoshu, a new genre using malt, and malt alcoholic beverages such as whiskey using malt. However, the fermented beverage may be other types of brewed liquor such as wine and sake, or may be distilled liquor accompanied by a distillation step such as whiskey, vodka, shochu, white sake, and brandy. There are no particular restrictions on the raw materials of beverages, the types of yeasts and fungi used in beverage production, the presence or absence of alcohol components in the finally produced beverage, and the alcohol content.

  The characteristics of the malt alcoholic beverage are specified by component values such as final appearance fermentation (AAL), chromaticity, pH, total nitrogen, total amino acid, bitterness (BU), and the like. Below, the case where the component value (characteristic) of the malt alcoholic drink produced | generated on a certain trial brewing condition is estimated when brewing a malt alcoholic drink is demonstrated as an example. The brewing conditions include information relating to the setting items (recipe) of at least one of the raw materials used or the brewing process.

  Examples of setting items related to raw materials of malt alcoholic beverages include malt usage, malt type, malt lot, malt mixing ratio, malt component value, hop usage, hop type, hop lot, hop The mixing ratio, the hop component value, the type of auxiliary material, the amount of auxiliary material used, the type of water treatment agent, the amount of water treatment agent used, the type of enzyme agent, the amount of enzyme agent used, and the like. By adjusting these items in brewing, malt alcoholic beverages with various characteristics and flavors can be produced. The settings relating to these raw materials are examples, and other setting items may be included in the brewing examples and the brewing conditions.

  As the setting items related to the brewing process of malt alcoholic beverages, changes in temperature during saccharification, saccharification time, boiling rate, boiling time, cooling time, cooling temperature, fermentation time, fermentation temperature, aging Examples include time, temperature at which aging is performed, and the type of filter used in filtration. These setting items are examples, and other setting items may be included in the brewing examples and the brewing conditions.

  In the characteristic prediction apparatus according to the present invention, the above-described setting items are set as explanatory variables (features), the component values of the above-mentioned fermented beverages are set as objective variables, and machine learning techniques are applied to set the values of the objective variables. Is estimated. The user can input a setting item related to the brewing condition for which the characteristic (component value) is to be predicted to the characteristic prediction apparatus, and can predict the component value of the beverage generated under the brewing condition.

  As described above, in the field of beverages such as malt alcoholic beverages, it is required to increase the development speed while suppressing time, cost, and resources required for trial production. In such a situation, it is necessary to predict the characteristics of the beverage obtained when brewing (trial brewing) under arbitrary conditions with high accuracy. Moreover, since the characteristic of a drink is specified by the several component value in the field | area of a drink, it is also calculated | required to predict a several component value efficiently.

  In the characteristic prediction apparatus according to the present invention, when each component value is predicted, a machine learning method that minimizes an error between an estimated value and an actual component value in the learning data among a plurality of machine learning methods is used. Presented to the user as a formal predicted value. Here, the estimated value means a component value estimated by each machine learning method.

The deviation between the estimated value by each machine learning method and the actual component value (true value) in the learning data can be examined by cross-validation shown in FIGS. When performing cross-validation, the learning data L is divided into a plurality of blocks L ₁ , L ₂ , L ₃ ,..., L _k in the learning process. For example, when performing learning processing (prediction model and error determination model generation processing) for the component values Y ₁ , Y ₂ , Y ₃ ,..., Y _m , the blocks L ₂ , L _{3 ,.} Is used as training data (model generation). Then, a error in case of using a plurality of machine learning techniques to block L ₁ as a test data.

  The above-described number of block divisions, block assignment of test data, and training data are only examples. If a part of the learning data is set as test data, and the part of the learning data that is not set as test data is used as training data, the number of block divisions, test data, and training data There is no particular limitation on the assignment of. For example, a plurality of blocks may be set as test data. The same test data and training data combination may be used in the learning process (prediction model and error determination model generation process) for each component value (objective variable), or different test data and training data may be used. Combinations may be used.

  Hereinafter, the number of divisions of the learning data L is k, and the number of component values to be predicted is m. There is no particular limitation on the relationship between k and m. In the example of FIG. 2 and FIG. 3, all of the remaining blocks that are not used as test data are used as training data. However, all of the remaining blocks need not be used as training data.

  In the present embodiment, for each component value (objective variable) to be predicted, the first method (first machine learning method) M5 ′ and the second method K * (second machine learning method) are used. To learn the prediction model. And the estimated value of an analytical value is calculated | required using the prediction model by each method. The estimated value based on the prediction model of the first method is referred to as a first estimated value, and the estimated value based on the prediction model of the second method is referred to as a second estimated value. For each component value, the error between the true value in the learning data and the first estimated value is compared with the error between the true value in the learning data and the second estimated value, and the official prediction value of the component value As described above, an estimated value related to a method with less error is used.

  The above-described M5 ′ and K * are examples, and other machine learning methods may be used. Alternatively, an estimated value of the analysis value by a larger number of machine learning methods may be obtained, and a method that minimizes an error between the estimated value and the true value may be selected.

  By performing the learning process as described above, it is possible to use an estimated value by a machine learning method that is estimated to have the smallest error in predicting each component value under the brewing conditions.

  Next, components of the characteristic prediction apparatus 1 will be described.

  The storage unit 2 is a storage area for storing past brewing case data, learning data, data generated by an intermediate process of machine learning, model data generated by machine learning, a characteristic prediction program, and the like. The storage unit 2 includes a brewing database 21, a learning data storage unit 22, a determination data storage unit 23, and a model database 24 as internal components. Details of the brewing database 21, the learning data storage unit 22, the determination data storage unit 23, and the model database 24 will be described later.

  The storage unit 2 may be a volatile memory such as SRAM or DRAM, or a nonvolatile memory such as NAND, MRAM, or FRAM. A storage device such as an optical disk, a hard disk, or an SSD may be used. The storage unit 2 may be built in the characteristic prediction apparatus 1 or may be a storage device outside the characteristic prediction apparatus 1. The storage unit 2 may be a removable storage medium such as an SD card, a MicroSD card, or a USB memory.

  The learning unit 3 outputs a component value of a beverage expected when brewing (trial brewing) is performed under an arbitrary condition. The learning unit 3 includes, as internal components, a variable selection unit 11, an intersection verification unit 12, a model generation unit 13, a first estimation unit 14, a second estimation unit 15, a verification unit 16, and a method selection. Part 17. Details of these components will also be described later.

  The preprocessing unit 4 converts data relating to past brewing cases into a learning data format that can be used by the learning unit 3. Details of processing performed by the preprocessing unit 4 will also be described later.

  The display unit 5 displays a GUI (Graphical User Interface) and CLI (Command Line Interface), various data, a screen for inputting tasting conditions, a screen for prediction results and related cases, etc. used by the user when operating the characteristic prediction device. Display. As the display, for example, a liquid crystal display, an organic electroluminescence display, a projector, an LED display, or the like can be used, but other types of displays may be used. Although the display unit 5 in the example of FIG. 1 is built in the characteristic prediction apparatus 1, the position of the display unit 5 is not particularly limited. The display unit 5 may be installed in a room or a building away from the characteristic prediction device 1 or may be a display of a wireless communication terminal such as a tablet or a smartphone.

  The operation unit 6 is a device that provides operation means of the characteristic prediction device 1 by the user. The operation unit 6 is, for example, a keyboard, a mouse, a switch, a voice recognition device, or the like, but is not limited thereto. The operation unit 6 may be a touch panel integrated with the display unit 5. The position of the operation unit 6 is not particularly limited. The operation unit 6 may be installed in a room or a building away from the characteristic prediction device, or may be a touch panel of a wireless communication terminal such as a tablet or a smartphone.

  Next, each component inside the storage unit 2 will be described.

  The brewing database 21 includes brewing data, which is information on setting items (recipe) such as raw materials and brewing processes in cases where beverages have been brewed in the past (past brewing cases) and component values of the generated beverages. Is stored. The tables of FIGS. 4 and 5 show examples of brewing data. A brewing ID, which is an identifier for uniquely identifying each brewing case, is stored in the column 30 of the table of FIG. For example, an identifier including a plurality of alphanumeric characters can be used as the brew ID, but the format of the identifier is not particularly limited. Each row identified by the brewing ID stores the value of each setting item in the brewing case and the component value (analysis value) of the generated beverage.

  In a column 31 of the table of FIG. 4, values of a plurality of setting items related to the raw material are stored. Column 31 includes the amount of malt A, the amount of malt B, the amount of malt C, the amount of hop D, the amount of hop E, the amount of secondary material F, the amount of secondary material G used. It is. These setting items are examples, and different setting items may be used. For example, when a plurality of raw materials are mixed and used, component values such as total nitrogen and chromaticity related to the mixed raw materials may be used as setting items. When using a component value as a setting item related to the raw material, the component value may be a value before saccharifying wort, or a value after saccharification, and the measurement timing of the component value is particularly It doesn't matter.

  The table in FIG. 5 shows data stored on the right side of the table in FIG. In the column 32 of the table of FIG. 5, values of a plurality of setting items related to the brewing process are stored. The column 32 includes a temperature change pattern, a saccharification time, a boiling rate, and a boiling time. These setting items are examples, setting items different from these may be used, and other setting items may be included.

  A column 33 on the left side of the column 32 stores a plurality of component values of the generated beverage. The column 33 includes component values relating to AAL (final appearance fermentation degree), chromaticity, pH, total nitrogen, amino acid total, and BU. These component values are examples, and component values different from these may be used. Since the characteristic prediction apparatus 1 predicts a value for each component value using a machine learning technique, each component value shown in the column 33a corresponds to an objective variable in machine learning.

  In addition, the component value of the drink produced | generated in brewing data may be the component value of the drink which all the processes completed, and the component value of the intermediate product in wort or an intermediate process may be sufficient as it. That is, the component value of the beverage to be predicted by the characteristic prediction device 1 may be related to the final product or may be related to a product generated during the production, and the time point in the brewing process is particularly questioned. Absent.

The learning data storage unit 22 stores learning data generated from the brew data. The table in FIG. 6 shows an example of learning data. The learning data is uniquely identified by the brewing ID like the brewing data. A column 30 indicates the brewing ID in the learning data. In the brewing data, the data related to each row identified by the brewing ID is called a (brewing) example, but in the learning data, the data related to each row is called a sample. In the table of FIG. 6, a column 34 and a column 35 are further shown. A column 34 is an explanatory variable of learning data. Column 34 includes a n number of explanatory variables from the explanatory variable x ₁ to the explanatory variable x _n in the example of FIG. In a column 35, component values of brewing cases corresponding to the respective samples are stored. In the example of FIG. 6, the column 35 includes _m component values from the component value Y ₁ to the component value Y _m .

  Hereinafter, a sample corresponding to a past brewing example included in the learning data is referred to as a past sample. Each past sample is uniquely identified by a brew ID. A sample generated from the brewing conditions for which the component value is to be predicted is called a brewing sample.

  Next, learning data generation processing (preprocessing) by the preprocessing unit 4 will be described.

  By the preprocessing executed by the preprocessing unit 4, the columns 31 and 32 (setting items) in the brew data are converted to explanatory variables shown in the column 34 of FIG. If data corresponding to brewing data is stored in a plurality of databases, data necessary for pre-processing may be extracted and data combining processing may be performed. In addition, when the accumulated data relating to past brewing cases is not in a format that can be uniquely identified by the brewing ID, information may be aggregated in units of brewing IDs. Further, when abnormal data is included in the brew data, the abnormal data may be removed.

  In the pre-processing, formal knowledge such as engineers' knowledge and documents may be reflected and converted into learning data. For example, when there is a component value that is essential for predicting a specific component value of a product based on experience, the component value can be set to be included in an explanatory variable used in machine learning. Such setting is performed by marking, for example, attribute information of explanatory variables of learning data and metadata storing information about explanatory variables. If it is known that the characteristics of the beverage produced will vary depending on the equipment at which the brewing is performed, the value of the explanatory variable may be corrected according to the equipment used in the brewing case. A user (for example, an engineer, a researcher, etc.) refers to the data displayed on the display unit 5, corrects abnormal data or erroneous data using the operation unit 6, or excludes it from the target of use. May be.

  Conversion to learning data can be performed using information described in the literature. For example, if the literature describes conditions for activating a specific enzyme used in saccharification, for brewing IDs that satisfy the activation conditions, be sure to include information on the enzyme in the explanatory variable. The brewing ID that does not satisfy the activation condition can be set so as to exclude information related to the enzyme from the explanatory variables.

  Moreover, the theoretical formula regarding the component value of a drink may be described in literature. By using a theoretical formula, even when component values are not actually measured in past brewing cases, component values estimated in the case may be derived. You may add the component value derived | led-out from the theoretical formula to learning data as an explanatory variable or an objective variable.

FIG. 7 shows an example of explanatory variables and objective variables used for each component value to be predicted. A table 36 is shown in the upper part of FIG. 7, and a table 37 is shown in the lower part of FIG. Tables 36 and 37 are obtained by extracting a part of the learning data of FIG. The table 36 shows explanatory variables x ₁ , x ₂ , x ₃ used when estimation is performed by setting the component value Y ₁ as the objective variable y ₁ . On the other hand, the table 37 shows explanatory variables x ₄ , x ₅ , x ₆ used when the component value Y ₂ is set to the objective variable y ₂ and estimation is performed. Although three explanatory variables are selected in the tables 36 and 37, the number of selected explanatory variables may be different from this. In addition, there may be duplication of explanatory variables used when estimating values of different objective variables. Variable selection as shown in the examples of the tables 36 and 37 is performed by the variable selection unit 11. Details of the variable selection unit 11 will be described later.

  The determination data storage unit 23 stores data used for selecting a machine learning method used for estimating a component value. Examples of data used for selecting the machine learning method include a determination table and an estimated value comparison table. Below, with reference to FIGS. 8-10, the example of the determination table preserve | saved at the determination data memory | storage part 23 and an estimated value comparison table is demonstrated.

FIG. 8 shows the determination table 38. A determination table is generated for each combination of a prediction model (for example, a machine learning method such as the first method or the second method) and a predicted component value (Y ₁ , Y ₂ ,..., Y _m ). A brewing ID is stored in a column 39 of the determination table 38. The column 40 of the determination table 38 is an explanatory variable corresponding to each brewing ID. Column 41 is the actual component values Y ₁ in each brewing ID (each brewing case), is used as a true value when determining the error of the estimated value of the model. Column 42 is the objective variable y ₁ in the prediction model according to one approach, the estimated value calculated by the predictive model are stored.

  The column 43 stores the determination result within the allowable error. The determination result within the permissible error takes a binary value of “TRUE” or “FALSE”. If the magnitude of the error between the estimated value by the model (column 42) and the actual analysis value (column 41), which is a true value, is within the allowable error, the value in column 43 is “TRUE”. If the error between the estimated value by the model (column 42) and the actual analysis value (column 41), which is a true value, is larger than the allowable error, the value of column 43 is “FALSE”. As an example of the allowable error, values such as 1%, 2%, 5%, and 10% of the true value can be used, but the set value is not particularly limited. As will be described later, the determination result within the allowable error in the column 43 is used as a teacher signal when learning the error determination model.

  In the column 43a, the within-tolerance probability corresponding to each brewing ID is stored. The within-tolerance probability is used for learning a method selection model and selecting an estimated value presented as a formal predicted value. Details of the probability within the allowable error will be described later.

  Columns 39 to 41 in FIG. 8 have the same contents as the learning data in FIG. Therefore, the determination table may be a common table with the learning data. Further, the determination table may not have the data related to the columns 40 and 41, and the corresponding data may be acquired from the table related to the learning data.

FIG. 9 shows an estimated value comparison table 44. An estimated value comparison table is generated for each predicted component value (Y ₁ , Y ₂ ,..., Y _m ). Column 45 stores the brewing ID. The brewing ID stored in the column 45 is related to the sample included in the block used as test data. Since the estimated value comparison table 44 is a table according to the component value Y _1, brewing ID of the samples included in the block L ₁ are stored. The column 46 stores the first estimated value obtained by the prediction model according to the first method. The column 47 stores the second estimated value obtained by the prediction model by the second method. In the example of FIG. 9, estimated values by two methods are shown, but estimated values by a larger number of types of methods may be included in the method comparison table 44. The column 48 stores actual component values that are true values.

The column 49 is selected based on an error (deviation) between the first estimated value based on the prediction model of the first method and the actual component value (true value) of the second estimated value based on the prediction model of the second method. The “preferred method” is stored for the prediction of the component value Y ₁ . When the estimated value is obtained using the test data, the method having the smaller error from the true value can be selected as the “preferred method” for each brewing ID. As will be described later, the “preferred method” in column 49 is used as a teacher signal.

  The model database 24 stores data (model data) of a prediction model that outputs an estimated value, an error determination model, and a method selection model. The prediction model is generated by the model generation unit 13, the error determination model is generated by the verification unit 16, and the method selection model is generated by the method selection unit 17. The contents and storage format of model data vary depending on the method and model used for learning.

FIGS. 10 and 11 show examples of model data when M5 ′ is used as a prediction model learning method (machine learning method). FIGS. 10 and 11 each show a tree having a depth of three. The tree is generated by dividing the learning data. FIG. 10 is a tree generated using explanatory variables used when estimating the objective variable y _{1 in} the learning data. FIG. 11 is a tree generated using the explanatory variables used when estimating the objective variable y _{2 in} the learning data. A regression equation is shown at each node in the trees of FIGS. Each regression equation is generated using the learning data included in each node, and an estimated value of the component value (Y ₁ or Y ₂ ) is output as the objective variable (y ₁ or y ₂ ).

  When K * is used as a technique, model data based on K * can be stored in the model database 24. In K *, an estimated value of an analytical value (objective variable) can be calculated by using the transition probability between the brew sample and the past sample and the component value of the brewing case corresponding to each past sample. . Details of the model generation unit 13, M5 ′, and K * will be described later.

  Next, the internal components of the learning unit 3 will be described.

The variable selection unit 11 selects explanatory variables to be used for estimating the respective target variables y ₁ , y ₂ ,..., Y _m from the learning data stored in the learning data storage unit 22. The variable selection process performed by the variable selection unit corresponds to the process of generating the tables 36 and 37 in FIG. 7 from the table in FIG. The variable selection unit 11 may perform variable selection using a linear dimension reduction method such as principal component analysis, or may narrow down explanatory variables used by other methods. In addition, information related to explanatory variables that are recorded in the attribute information and metadata of the learning data by the preprocessing unit 4 and are essential for use in specific brewing IDs, and explanatory variables that are excluded from the use targets in specific brewing IDs The variable selection may be performed using the information related to. The explanatory variable used by the user may be selected, or the user may change the explanatory variable selected by the variable selection unit 11.

Note that the characteristic prediction apparatus 1 does not necessarily include the variable selection unit 11. That is, the value of the objective variable may be estimated using all the explanatory variables x ₁ , x ₂ ,..., X _m included in the learning data without performing variable selection (dimension reduction). In the following explanation, when it is simply referred to as learning data, the original learning data before variable selection (dimension reduction) and variable selection (dimension reduction) were performed to estimate each objective variable. Both later learning data shall be included.

The cross validation unit 12 divides the learning data L into a plurality of blocks L ₁ , L ₂ , L ₃ ,..., L _k , and component values (Y ₁ , Y ₂ ,. , Y _m ), a block used as test data and a block used as training data are determined in learning of a prediction model and a corresponding error determination model. Details of the processing executed by the cross-validation unit 12 are as described in the description of FIGS. The cross-validating unit 12 may store the information related to the above-described block allocation for each application in the storage unit 2.

  The model generation unit 13 generates a prediction model (machine learning model) that outputs an estimated value of the objective variable based on the learning data. The prediction model generated by the model generation unit 13 is stored in the model database 24. Here, a case where M5 ′ (first method) is used as a machine learning model (method) will be described as an example. However, an estimated value of an objective variable may be calculated using another model such as M5 or a neural network. . Below, the outline | summary of M5 'which is one of the examples of the method which the model production | generation part 13 uses is demonstrated.

ここで、ｍは欠値のないサンプルの数、｜Ｔ｜は、分割前のノードに含まれるサンプル数、β（ｉ）は補正係数、ｓｄ（Ｔ）は分割前のノードにおける標準偏差、｜Ｔ_ｊ｜は分割により作られた一方のノード（左側または右側）に含まれるサンプル数、ｓｄ（Ｔ_ｊ）は分割により作られた一方のノード（左側または右側）における標準偏差である。 In M5 ′ (first method), learning data is sequentially divided on the basis of maximization of a standard deviation reduction (SDR). The SDR in M5 ′ is defined as the following formula (1).

Here, m is the number of samples without missing values, | T | is the number of samples included in the node before division, β (i) is a correction coefficient, sd (T) is the standard deviation at the node before division, | T _j | is the number of samples included in one node (left or right) created by division, and sd (T _j ) is the standard deviation in one node (left or right) created by division.

  The SDR in the formula (1) is an example, and an SDR represented only by a term in parentheses in the formula (1) may be used, such as M5 that is a prototype of M5 ′.

  By dividing the learning data based on the formula (1), a subset of past samples (for example, leaf nodes in FIGS. 10 and 11) belonging to the learning data in which the variance (standard deviation) is minimized is obtained. When statistical significance is no longer recognized in the reduction of standard deviation due to division (for example, when the standard deviation of the data after division is less than 5% of the standard deviation of the original learning data) or included in the node When the number of samples to be reached reaches a lower limit (for example, 3 or more), the division of the learning data can be aborted. When the tree is generated by the division, a regression equation that outputs an estimated value of the objective variable is generated using samples included in each node. The regression equation can be calculated by, for example, linear regression, but the regression method is not particularly limited.

  Once the regression equation is determined for each node in the tree, the tree can be pruned. For the past samples included in all nodes except the root node, an average value of absolute value errors between the estimated value based on the regression equation and the actual component value related to the brewing case corresponding to the past sample is calculated. Next, for each subtree included in the tree, the total sum of the absolute values of the absolute value errors at the branches is compared with the sum of the average values of the absolute value errors at the root node of the subtree. If the former value is larger, the subtree is pruned. Thereby, the tree can be simplified without impairing the prediction accuracy, and the influence of overlearning can be reduced.

  By using M5 ′, a regression equation can be prepared for each subset of past samples obtained by dividing the learning data. Thereby, the prediction accuracy of the regression equation can be improved as compared with the case where one regression equation is applied to the entire learning data.

  The 1st estimation part 14 estimates the component value (objective variable) of a drink using the prediction model (M5 ') by the 1st method produced | generated by the model production | generation part 13. FIG. Hereinafter, the estimated value obtained by the prediction model (M5 ′) by the first method is referred to as a first estimated value, and is distinguished from the estimated value obtained by the prediction model (K *) by the second method.

A calculation process of the first estimated value when M5 ′ is used as the first method will be described. The model generation unit 13 uses the M5 ′ model, divides the learning data, and generates a tree in which a regression equation is associated with each node after division (for example, FIGS. 10 and 11). The first estimation unit 14 refers to the model database 24 and identifies data related to the tree corresponding to the component value Y _i (object variable y _i ) to be estimated. The first estimation unit 14 refers to the explanatory variable value of the tasting sample used for estimating the component value Y _i (objective variable y _i ), and identifies which node of the tasting sample the tree belongs to. When the node to which the sample sample belongs is specified, the value of the objective variable y _i is calculated using the regression equation associated with the node. Then, the first estimation unit 14 outputs the value of the objective variable y _i calculated from the regression equation as the first estimated value.

  The 2nd estimation part 15 estimates the component value (objective variable) of a drink using the prediction model (K *) by the 2nd method produced | generated by the model production | generation part 13. FIG. Hereinafter, the estimated value obtained by the prediction model (K *) by the second method is referred to as a second estimated value, and is distinguished from the first estimated value obtained by the prediction model (M5 ′) by the first method. To do.

  A calculation process of the second estimated value when K * is used as the second method will be described. K * is a case-based technique for obtaining an estimated value of a component value under a brewing condition using a past sample of learning data. Below, the method of performing the predicted value of a component value using K * (2nd method) is demonstrated.

Here, it is assumed that I is a set of samples having continuous values, and T is a set of conversion operations t on I. As shown in the following formula (2), it is assumed that the transition from the sample a to the sample b can be performed by a combination of the conversion operations t.

式（３）はあるサンプルを始点としたときに起こりうる遷移パターンの確率をすべて足し合わせると１になることを示している。 If P is a set of transformations that realize transitions between samples in I, a probability function (Probability Function) that satisfies the relationship shown in the following equation (3) can be defined.

Equation (3) shows that the sum of all transition pattern probabilities that can occur when a certain sample is the starting point is 1.

Equation (4) below shows the sum P * of the probabilities of transformation taking each path from sample a to sample b in the feature space.

Ｋ＊関数は、Ｋ＊学習器における距離関数として用いられる。本実施形態における過去サンプルのように、それぞれのサンプルが複数の説明変数を含む場合、サンプル間の遷移確率Ｐ＊はそれぞれの説明変数について求めた遷移確率の積となる。サンプル間のＫ＊関数は、それぞれの説明変数について求めたＫ＊関数の値の和となる。 The K * function in Equation (5) below represents P * in Equation (4) in a form similar to entropy.

The K * function is used as a distance function in the K * learner. When each sample includes a plurality of explanatory variables as in the past sample in the present embodiment, the transition probability P * between samples is a product of the transition probabilities obtained for the respective explanatory variables. The K * function between samples is the sum of the values of the K * function obtained for each explanatory variable.

Using the transition probability between samples, the component value can be estimated by K *. First, for each explanatory variable x _i, calculate the transition probabilities between each of the past samples of Fermentation samples and learning data. Next, the transition probabilities obtained for each explanatory variable are integrated. As a result, the transition probabilities between the brew sample including all the explanatory variables and the respective past samples are obtained. Finally, the product of the transition probability between the actual component value in the brewing case corresponding to the past sample and the sample sample is calculated for each past sample, and all the data included in the target data (learning data) Add the past samples. By dividing this value by the sum of transition probabilities between the past sample and the brew sample, an estimated value of the component value can be obtained.

  In addition, when checking the error of the second method using the block set in the test data in the learning process, each past sample belonging to the block set in the test data and the past of the training data Calculate transition probabilities between samples.

You can also use K * to classify samples into categories. For example, when obtaining the probability that the sample a belongs to the category C, the sum of the probabilities that the sample a transitions to any of the samples belonging to the category C may be obtained. That is, the following formula (6) can be used as P *.

  When a past sample corresponding to each brewing case is added to a feature space that is classified into a plurality of categories and the above formula (6) is applied, the value of P * is calculated for each category. can do. In this case, it can be determined that the sample brew sample belongs to the category having the maximum P * value.

  The verification unit 16 determines whether or not an error between the estimated value and the true value (actual component value) is within an allowable error when the estimated value is obtained using the test data in the learning process. Then, the determination result within the allowable error is obtained. Then, using the training data, the determination result within the allowable error is used as a teacher signal, and an error determination model for classifying whether the deviation is within the allowable error (TRUE / FALSE) is learned. Finally, the learned error determination model is applied to the test data to determine the within-tolerance probability.

Since the processing by the verification unit 16 described above is executed for each component value to be predicted and each technique used for prediction, the component values (Y ₁ , Y ₂ ,..., Y _m ) to be predicted and machine learning are performed. An error determination model is generated for each combination of methods (for example, the first method and the second method). Next, a process executed by the verification unit 16 in the prediction process will be described.

  In the prediction process, an explanatory variable related to the tasting sample is input to the error determination model, a probability within tolerance is calculated for each method, and a method selection unit 17 (method selection model described later) is calculated based on the probability within tolerance. ) To select a method in which the error is estimated to be small (high accuracy).

  An error determination model that performs binary classification (TRUE / FALSE) of whether or not the deviation between the estimated value and the true value is within an allowable error can be learned by, for example, logistic regression. Below, the outline | summary of the learning process of the error determination model using logistic regression is demonstrated.

ここで、Ｚは目的変数、Ｘ_１〜Ｘ_ｎ＋１は説明変数、ｂ_０は定数、ｂ_１〜ｂ_ｎ＋１は回帰係数である。例えば、説明変数Ｘ_１〜Ｘ_ｎとして試醸サンプルに係る説明変数ｘ_１〜ｘ_ｎを使うことができる。説明変数Ｘ_ｎとして、特徴空間において過去サンプルのうち、試醸サンプルとの距離が最も短いものまでの距離（最近接距離）を使うことができる。距離の例としては、ユークリッド距離、マンハッタン距離、ミンコフスキー距離、チェビシフ距離、マハラノビス距離などがあるが、どの種類の距離を用いてもよい。なお、式（７）のモデルの一例であり、ロジスティックモデルの説明変数に必ず最近接距離を含めなくてもよい。 The logistic model used by the verification unit 16 is expressed as the following equation (7).

Here, Z is an objective variable, X _{1 to} X _{n + 1} are explanatory variables, b ₀ is a constant, and b _{1 to} b _{n + 1} are regression coefficients. For example, explanatory variables x _{1 to} x _n related to the brew sample can be used as the explanatory variables X _{1 to} X _n . As the explanatory variable _Xn , it is possible to use the distance (closest distance) to the shortest distance from the brew sample among the past samples in the feature space. Examples of distances include Euclidean distance, Manhattan distance, Minkowski distance, Chebyshev distance, Mahalanobis distance, etc. Any type of distance may be used. Note that this is an example of the model of Equation (7), and the nearest distance may not necessarily be included in the explanatory variable of the logistic model.

By performing logistic regression using the determination result within the permissible error as a teacher signal, the values of the constant b ₀ and the regression coefficients b _{1 to} b _{n + 1} are calculated. In the numerical calculation, the values of b _{0 to} b _{n + 1} are repeatedly updated so as to take the extreme value of the log likelihood using the steepest descent method or the stochastic gradient descent method. However, the numerical calculation algorithm used is not particularly limited.

When the values of the constant b ₀ and the regression coefficients b _{1 to} b _{n + 1} are obtained, the objective variable Z can be substituted into the sigmoid function represented by the following equation (8) to calculate the within-tolerance probability p.

As shown in FIG. 12, the output value p of the sigmoid function takes a value in the range of (0, 1). When logistic regression analysis is used, it is possible to estimate the probability that an objective variable takes any value when the analysis target is in a binary state of 1 and 0. When obtaining the estimated values by the two machine learning methods, the verification unit 16 determines the in-tolerance probability p ₁ related to the first estimated value and the allowable error related to the second estimated value for each component value to be predicted. Each of the inner probabilities p ₂ is calculated.

  In the above description, the case where the error determination model is learned using logistic regression has been described as an example. However, the error determination model may be learned using other methods. Examples of other methods for performing probabilistic classification include least square probabilistic classification.

  In the learning process, the method selection unit 17 determines an error based on an error (deviation) between an estimated value obtained by inputting test data into a prediction model of each method and an actual component value (true value). A machine learning method with a small (high accuracy) is selected as a “preferred method”. The detailed processing is as described in the description related to FIG. Next, the method selection unit 17 adds the within-tolerance probability (for example, the first in-tolerance probability and the second within-tolerance probability) obtained by the verification unit 16 to the explanatory variable of the learning data, "Is used as a teacher signal to learn the method selection model.

Since the learning of the method selection model is performed for each prediction target component value, a method selection model is generated for each prediction target component value (Y ₁ , Y ₂ ,..., Y _m ). Next, a process executed by the method selection unit 17 in the prediction process will be described.

In the prediction process, the probability within tolerance (for example, probabilities p ₁ , p ₂ ) calculated by the verification unit 16 and the explanatory variable related to the tasting sample are input to the method selection model, and are formalized. An estimated value to be used as a predicted value (for example, when there is a first method and a second method, either the first estimated value or the second estimated value) is selected.

  Next, the learning process of the method selection model will be described. Logistic regression can be used in the same manner as the error determination model when selecting an estimated value of one of the two methods as a formal predicted value, such as the first method and the second method. Below, the case where logistic regression is used for learning of the method selection model will be described as an example.

ここで、Ｚ´は目的変数、Ｘ´_１〜Ｘ´_ｎ＋２は説明変数、ｄ_０は定数、ｄ_１〜ｄ_ｎ＋２は回帰係数である。 In learning of the method selection model, a logistic model such as the following equation (9) is used.

Here, Z ′ is an objective variable, X ′ _{1 to} X ′ _{n + 2} are explanatory variables, d ₀ is a constant, and d _{1 to} dn _{+ 2} are regression coefficients.

Method selection unit 17, for example, describes a variable X'tolerances within probability _{p 1} that a ₁ according to the first estimate, the explanatory variables X'tolerances in probability _{p 2} of the ₂ to the second estimate, explanatory variables _X'3 as ~X' _{n + 2,} it is possible to use the explanatory variable _x 1 ~x _n according to Fermentation samples. That is, the explanatory variable of the logistic regression is obtained by adding the probability within the allowable error of the first estimated value under the brewing condition and the probability within the allowable error of the second estimated value under the brewing condition to the explanatory variable of the brewing sample. . Then, the technique selection unit 17 performs logistic regression using the “preferred technique” shown in the column 49 of FIG. 9 as a teacher signal, and calculates values of the constant d ₀ and the regression coefficients d _{1 to} d _{n + 2} . The numerical calculation method used is not particularly limited.

そして、確率ｐ´の値を２値化する。２値化処理の一例としては、しきい値を使う方法がある。例えば、確率ｐ´がしきい値より大きい場合には出力値を１とし、モデルベースの手法（第１推定値）を選択する。そして確率ｐ´がしきい値以下の場合には出力値を０とし、事例ベースの手法（第２推定値）を選択する。しきい値としては、例えば０．５を使うことができるが、これとは異なる値であってもよい。以降では式（１０）の確率ｐ´を手法変数とよび、許容誤差内確率ｐと区別するものとする。 When the value of the objective variable Z ′ is obtained, the probability p ′ is calculated by substituting Z ′ into the following equation (10) that is a sigmoid function.

Then, the value of the probability p ′ is binarized. As an example of the binarization process, there is a method using a threshold value. For example, when the probability p ′ is larger than the threshold value, the output value is set to 1, and a model-based method (first estimated value) is selected. If the probability p ′ is less than or equal to the threshold value, the output value is set to 0, and a case-based method (second estimated value) is selected. As the threshold value, for example, 0.5 can be used, but a different value may be used. In the following, the probability p ′ in equation (10) is called a technique variable and is distinguished from the tolerance p within the tolerance.

  In the above description, the case where logistic regression is used for learning the method selection model has been described as an example. However, the method selection model may be learned using other methods. For example, least square probabilistic classification may be used, or method selection models may be learned using support vector classification, decision trees, ensemble learning, and the like.

Since a method selection model is generated for each component value (Y ₁ , Y ₂ ,..., Y _m ) to be predicted, the method to be selected differs depending on the component value to be predicted in the actual prediction process. There is a case. In the prediction process, the method selection unit 17 may display the type of the machine learning method selected for the prediction of each component value and the formal prediction value of the component value under the brewing conditions on the display unit 5.

  Next, details of processing executed by the characteristic prediction apparatus 1 will be described. 13 to 15 are flowcharts illustrating the learning process executed by the characteristic prediction apparatus. Before the user predicts a component value under brewing conditions, a learning process for predicting the component value needs to be performed. Hereinafter, the processing will be described with reference to the flowcharts of FIGS.

  First, the preprocessing unit 4 generates learning data from the brewing data, and stores the learning data in the learning data storage unit 22 (step S101). The processing executed in step S101 is as described in the description related to the preprocessing unit 4. Next, the variable selection unit 11 selects a variable to be used for learning for each combination of a component value to be predicted (objective variable) and a technique used for estimating the objective variable (step S102). Details of the processing executed in step S102 are as described in FIGS. 6 and 7 and the description of the variable selection unit 11. At this time, the variable selection unit 11 may store information on the selected variable in the learning data storage unit 22.

Then, the intersection verification unit 12 divides the learning data L into _k and generates a plurality of blocks L ₁ , L ₂ ,..., L _k (step S103). At this time, cross-validation unit 12 Each component value Y _1, Y _2, · · ·, when performing the learning of the prediction model and the error determination model according to Y _m, and a block to the test data, a training data The block allocation is determined. The number of divisions k of the learning data L is not particularly limited. The details of the processing executed in step S103 are as described in FIGS. 2 and 3 and the description related to the cross-validating unit 12.

  When the learning data is divided into a plurality of blocks, a part of the learning data is set as test data, and the remaining blocks are set as training data (step S104). When the block assignment of the test data and the training data in the learning data is determined, the prediction model and the error determination model are learned for each combination of the component value (objective variable) and the machine learning method of each prediction target. Hereinafter, a case where learning of a prediction model and an error determination model is performed by iterative processing will be described as an example. First, 1 is substituted into the loop counter i for repetitive processing (step S105).

  After step S105 is executed, the characteristic prediction apparatus 1 performs the processing of steps S106 to S111 (learning the prediction model and the corresponding error determination model according to the first method) and the processing of steps S112 to S118 (related to the second method). (Prediction model and corresponding error determination model learning) are executed in parallel. Below, the process of step S106-S111 is demonstrated first.

The model generation unit 13 generates a prediction model by the first method (for example, M5 ′) for the component value Y _i (object variable y _i ) estimated using the training data (step S106). The model generation unit 13 stores the prediction model generated by the first method in the model database 24 as model data (step S107). The model data is stored together with information indicating component values (object variables) estimated by the model. The details of the processing executed in step S106 and step S107 are as described in the description related to the model generation unit 13.

Next, the first estimation unit 14 obtains a first estimate of the component values by prediction model with the test data Y _{i (objective} variable y _i). And the verification part 16 calculates | requires and preserve | saves the magnitude | size of the error between the 1st estimated value in each brewing ID, and a true value (actual component value). The verification unit 16 determines whether or not the magnitude of the error is within the allowable error range (step S108).

  The verification unit 16 stores the determination result within tolerance (TRUE / FALSE) for each brew ID included in the test data (step S109). For example, the verification unit 16 can store the determination result within the allowable error in the determination table of FIG. Details of the processing executed in steps S108 and S109 are as described in the description of FIG.

  Then, the verification unit 16 uses the training data, learns an error determination model that estimates the probability that the deviation (error) is within the allowable error range, using the determination result within the allowable error as a teacher signal, and the model of the storage unit 2 Save in the database 24 (step S110). The details of the process executed in step S110 are as described in the explanation relating to the error determination model. Next, the verification unit 16 applies the error determination model to the test data, and obtains the first allowable error probability for each brew ID (step S111). The probability within the first allowable error can be stored in the determination table.

  Next, processing in steps S112 to S118 will be described.

The model generation unit 13 learns the component value Y _i (objective variable y _i ) estimated using the training data by the second method (step S112). When K * is used as the second method, preprocessing for calculating the transition probability may be executed in step S112. In particular, when it is not necessary to execute preprocessing or the like, the processing according to step S112 may be omitted. When a model-based method is used as the second method, another prediction model is generated in step S112.

  Next, the model generation unit 13 stores the data used in the second method as model data in the model database 24 (step S113). In particular, when there is no model data to be stored, the processing according to step S113 may be omitted.

Then, the second estimation unit 15 obtains a second estimated value of the component value Y _i (objective variable y _i ) by the second method using the test data. And the verification part 16 calculates | requires and preserve | saves the magnitude | size of the error between the 2nd estimated value and true value (actual component value) in each brewing ID. The verification unit 16 determines whether or not the magnitude of the error is within the allowable error range (step S114).

  The verification unit 16 stores the determination result within tolerance (TRUE / FALSE) for each brew ID included in the test data (step S115). For example, the verification unit 16 can store the determination result within the allowable error in the determination table of FIG. Details of the processing executed in steps S114 and S115 are as described in the description of FIG.

  Then, the verification unit 16 uses the training data, learns an error determination model that estimates the probability that the deviation (error) is within the allowable error range, using the determination result within the allowable error as a teacher signal, and the model of the storage unit 2 It stores in the database 24 (step S117). The details of the process executed in step S117 are as described in the explanation relating to the error determination model. Next, the verification unit 16 applies an error determination model to the test data, and obtains a second allowable error probability for each brew ID (step S118). The probability within the second allowable error can be stored in the determination table.

When the processing of step S111 and step S118 is completed for the prediction target component value Y _i (objective variable y _i ), i + 1 is substituted into the loop counter i (step S119). Next, the value of the loop counter i is compared with the number m of component values (objective variables) to be predicted. If i ≦ m, the process returns to step S106 and step S112 (parallel processing). If i> m, the process proceeds to step S121 (step S120). Thus, the iterative process is executed until the error determination model corresponding to the prediction model is learned for all the prediction target component values Y _i (objective variable y _i ).

In the example shown in the flowcharts of FIGS. 13 and 14, the learning of the error determination model corresponding to the prediction model related to each prediction target component value Y _i (objective variable y _i ) is realized by the iterative process. Processing parallelism may be different. For example, learning of an error determination model corresponding to a prediction model related to a plurality of component values Y _i (objective variable y _i ) may be performed in parallel, or learning of models related to the first method and the second method may be performed. You may perform sequentially. Parallel learning processing can be realized by using, for example, a plurality of computers (computers) and a plurality of CPUs (Central Processing Units).

  Next, the method selection unit 17 determines, for each objective variable y (component value Y), the error between the first estimated value by the first method and the true value (actual component value), and the second method by the second method. 2 Compare the error between the estimated value and the true value (actual component value). The method selection unit 17 selects a method with a small error for the target variable y as a “preferred method” for estimating the target variable (step S121). Details regarding the “preferred method” selection process are as described in FIG. 9 and the description of the method selection unit 17.

  When “preferred method” is selected for each of the brewing IDs of the learning data, learning of the method selection model is performed without dividing the learning data into training data and training data. In the example of FIG. 15, since learning of the method selection model is repeated for each component value (object variable) to be predicted, 1 is assigned to the loop counter j corresponding to each object variable y (component value Y) (step S122).

First, the method selection unit 17 adds the first allowable error probability and the second allowable error probability to the explanatory variables of the learning data, and uses the “preferred method” as the teacher signal for the objective variable y _j (component value Y _j ). The method selection model is learned (step S123). Details of learning of the method selection model are as described in the description of the method selection unit 17. Then, the technique selection unit 17 stores the technique selection model related to the learned objective variable y _j (component value Y _j ) in the model database 24 of the storage unit 2.

When the process of step S124 is completed for the prediction target component value Y _j (objective variable y _j ), j + 1 is substituted into the loop counter j (step S125). Next, the value of the loop counter j is compared with the number m of component values (objective variables) to be predicted. If j ≦ m, the process returns to step S123. If j> m, all learning processes are terminated (step S126). Thus, the iterative process is executed until the method selection model is learned for all the prediction target component values Y _j (objective variable y _j ).

In the example shown in the flowchart of FIG. 15, learning of the method selection model related to each prediction target component value Y _i (objective variable y _i ) is realized by iterative processing, but a plurality of component values Y _i (objectives) The method selection model relating to the variable y _i ) may be learned in parallel. Parallel learning processing can be realized by using, for example, a plurality of computers (computers) and a plurality of CPUs (Central Processing Units).

  Next, prediction processing by the characteristic prediction apparatus 1 will be described.

16, FIG. 17 is a flowchart showing a prediction processing of a component value Y _p. When the learning process is completed for the prediction target component value (objective variable), the prediction process can be executed for the component value (objective variable). Hereinafter, the processing will be described with reference to the flowcharts of FIGS. 16 and 17.

First, the user determines the brewing conditions and the component value of the beverage to be predicted under the brewing conditions (step S201). The component value of the beverage may be one or plural. Next, the user inputs a component value Y _p to be expected Fermentation conditions using the operating unit 6 in the characteristic prediction apparatus 1 (step S202). Then, the user issues a component value prediction process start command from the operation unit 6.

  When a prediction process start command is issued, the characteristic prediction apparatus 1 determines the explanatory variable for each component value estimation method (for example, a model-based method such as M5 ′ or a case-based method such as K *) from the brewing conditions. An objective variable is selected (step S203). The explanatory variables and objective variables selected here are called tasting samples. In step S203, a brew sample is generated for each component value related to the brew condition that is the prediction target. The details of the processing executed in step S203 are as described in the description related to FIG.

  After the process of step S203 is executed, the characteristic prediction apparatus 1 performs the processes of steps S204 and S205 (estimation by the first method and calculation of the within-tolerance error probability) and the processes of steps S206 and S207 (estimation by the second method). The calculation of probability within tolerance is performed in parallel. Below, the process of step S204, S205 is demonstrated first.

The first estimation unit 14 inputs the brew sample into the prediction model according to the first method for estimating the component value Y _p (objective variable y _p ), and calculates the first estimated value (step S204). When M5 ′ is used as the first method, an estimated value of the component value (objective variable) is obtained from the regression equation at the corresponding node of the tree.

  Next, the verification unit 16 inputs the explanatory variable related to the brew sample to the error determination model corresponding to the prediction model of the first method, and calculates the first allowable error probability (Step S205). The verification unit 16 may store the calculated first allowable error probability in the storage unit 2.

  Next, the processing of steps S206 and S207 will be described.

The second estimation unit 15 inputs the brew sample into the prediction model according to the second method for estimating the component value Y _p (objective variable y _p ), and calculates the second estimated value (step S206). When K * is used as the second method, the product of the transition probability between the brew sample and each past sample and the component value of the brewing case corresponding to each past sample is added, and the transition probability By dividing by the sum, the second estimate can be calculated.

  Next, the verification unit 16 inputs the explanatory variable related to the tasting sample to the error determination model corresponding to the prediction model of the second method, and calculates the second allowable error probability (step S207). The verification unit 16 may store the calculated second allowable error probability in the storage unit 2.

If the process of step S205 and step S207 is performed, the process of step S208 will be performed. In step S <b> 208, the method selection unit 17 determines the method selection model corresponding to the component value Y _p (objective variable y _p ), the explanatory variable relating to the brew sample, the first allowable error probability, and the second allowable error probability. To obtain the output value of the method selection model. When the method selection model is learned by logistic regression, the output value is a value in the range of (0, 1).

Method selection unit 17 based on the output value of the approach selected model, decide whether to present the user either the first estimate or the second estimate as the predicted value of the formal component value Y _p (Step S209) . When the method selection model is learned by logistic regression, the output value is compared with a threshold value (for example, 0.5). For example, when the output value is larger than the threshold value, the first estimated value is selected as the official predicted value. If the output value is less than or equal to the threshold value, the second estimated value is selected as the formal predicted value.

Here, the process of selecting one of the estimates by the two prediction model as the predicted value of the formal component value Y _p is being performed, the judgment result by the formal component values an estimate by both prediction model it may be selected as the predicted value of Y _p. Further, when the estimated value by more than two predictive model is calculated may select an estimate of any one among them as a prediction value of the formal component value Y _p, multiple estimates may be selected as a formal predicted value of component value Y _p.

  Next, the learning unit 3 extracts a past sample to be presented as a related case based on the position of the tasting sample in the feature space of the learning data (step S210). For example, a past sample that is classified into the same category (cluster) as the tasting sample in the feature space or a past sample that is in the vicinity of the tasting sample in the feature space can be extracted as a related case.

  Hereinafter, the related case extraction process will be described with reference to FIG. FIG. 18 shows a feature space related to learning data. In the example of FIG. 18, past samples related to learning data are classified into category # 1, category # 2, and category # 3. Classification processing into categories can be performed by, for example, K *, k-average method, decision tree, ensemble classification, etc., but the method used for classification is not particularly limited.

  In the feature space of FIG. 18, the black circle corresponds to the position of the brew sample 60. When a past sample classified in the same category (cluster) as the sample sample is extracted as a related case, a past sample belonging to category # 3 is extracted as a related case.

  In addition, it is possible to extract related cases in a feature space that is not classified into categories (clusters). The surface of the sphere 61 centering on the brewing sample 60 (black circle) in FIG. 18 shows an equidistant line from the brewing sample 60. In the feature space of FIG. 18, for example, a past sample within the range of the sphere 61 can be extracted as a related case. As the distance in the feature space, Euclidean distance, Manhattan distance, Minkowski distance, Chebyshev distance, Mahalanobis distance, etc. can be used, but any kind of distance may be used.

  In this way, past samples within a certain distance from the tasting sample may be extracted as related cases, or a predetermined number of past samples are extracted as related cases in order from the nearest past sample. May be. Here, for example, values such as 10, 15, and 20 can be used as the predetermined number, but the set value is not particularly limited.

  Hereinafter, the description returns to the flowchart of FIG.

Finally, it presented to the user the relevant case the predicted values and Fermentation conditions formal component value Y _p (step S211). In step S211, it is only necessary that the predicted value of the component value is presented to the user by some method. For example, the predicted value may be displayed on a display, the predicted component value may be printed out on paper and provided to the user, or the predicted component value may be provided to the user's email address. You may transmit and you may notify the expected value of a component value with an audio | voice. Moreover, when the characteristic prediction apparatus 1 has a function such as a web service or an intranet server, the predicted value of the component value in the brewing conditions may be confirmed from a web browser. An example of the screen display in step S211 will be described later.

  The predicted value presented (displayed) to the user in step S211 may be a predicted value based on one prediction model or may be predicted values based on a plurality of prediction models. Further, the number of related cases presented to the user is not particularly limited.

  In the flowcharts of FIG. 16 and FIG. 17, the estimation process by the first method and the calculation process of the probability within the allowable error and the estimation process by the second method and the calculation process of the probability within the allowable error are performed in parallel. May be executed sequentially. Further, when a plurality of component values are predicted, all or some component values may be predicted in parallel, or each component value may be predicted sequentially.

  Next, an example of a screen displayed on the display unit 5 by the characteristic prediction apparatus 1 will be described. FIG. 19 shows an example of a tasting condition input screen. The screen 50 of FIG. 19 is a brewing condition input screen in the “beverage characteristic prediction system”. On the screen 50, the user can modify the past brewing cases and set the brewing conditions. Moreover, about the malt, after combining the malt which concerns a some lot, the component value of the malt at the time of a brew start can be calculated. In screen 50, the type of malt, the amount of each malt used, the type of hop, the amount of each hop used, the type of sub-material, the amount of each sub-material used, the type of water treatment agent, the type of each water treatment agent Fine brewing conditions such as the amount used, type of enzyme agent, amount of each enzyme agent used, and temperature conditions can be set.

  FIG. 20 shows a first example of a prediction result display screen displayed on the display. In the table 51a in the upper part of the screen 51 in FIG. 15, final predicted values relating to the final appearance fermentation degree (AAL), chromaticity, pH, total nitrogen, amino acid total, and a plurality of component values of BU are displayed. A table 51b at the bottom of the screen 51 displays brewing IDs corresponding to a plurality of related cases. The user can refer to more detailed information related to related cases by clicking the table at the bottom of the screen 51. The detailed information of the related cases includes, for example, setting items in the respective related cases and component values of the beverages generated in the respective related cases.

  FIG. 21 shows a second example of the prediction result display screen displayed on the display. In the table of FIG. 21, predicted values of final appearance fermentation degree (AAL), chromaticity, pH, total nitrogen, amino acid total, and BU by the first method and the second method are displayed. Thus, you may display the estimated value by a some machine learning method (prediction model) as a component value of the drink to be predicted.

  By using the characteristic prediction apparatus according to the present invention, using the brewing data accumulated in the past, the characteristics of the beverage obtained when brewing under arbitrary conditions (trial brewing conditions) are predicted with high accuracy. Can do. The more the number of cases included in the brewing data that is the basis of the learning data, the more accurate the component value can be predicted. Moreover, the characteristic prediction apparatus according to the present invention presents the user with related cases having relevance to the brewing conditions input by the user. As a result, the user is freed from the trouble of searching for a brewing case that can be used as a reference for development among a large number of past brewing cases. The user can refer to related cases and adjust or reexamine the brewing conditions (setting items such as raw materials and brewing processes) from which a beverage having a desired characteristic (component value) is obtained.

  New employees and young employees assigned to the R & D department can refer to related cases to find out what raw materials are used and how the brewing process is adjusted to produce beverages with the desired characteristics (component values). You can learn what can be generated. By using the characteristic prediction apparatus according to the present invention, new employees and young employees can efficiently understand the know-how and sense in brewing that engineers have acquired through repeated trial brewing for many years. The characteristic predicting apparatus according to the present invention reduces the time, cost, and resources required for beverage prototyping, increases the development speed, and contributes to education and succession of know-how to employees.

(Second Embodiment)
In the second embodiment, a hardware configuration of the characteristic prediction apparatus according to the present invention will be described. The characteristic prediction apparatus according to the present invention is configured by a computer 100. The computer 100 includes various information processing apparatuses such as a server, a client terminal, a microcomputer of an embedded device, a tablet, a smartphone, a feature phone, and a personal computer. The computer 100 may be realized by a virtual machine (VM) or a container.

  FIG. 17 is a diagram illustrating an example of the computer 100. The computer 100 in FIG. 17 includes a processor 101, an input device 102, a display device 103, a communication device 104, and a storage device 105. The processor 101, the input device 102, the display device 103, the communication device 104, and the storage device 105 are connected to each other via a bus 106.

  The processor 101 is an electronic circuit including a control device and a calculation device of the computer 100. Examples of the processor 101 include a general-purpose processor, a central processing unit (CPU), a microprocessor, a digital signal processor (DSP), a controller, a microcontroller, a state machine, an application-specific integrated circuit, a field programmable gate array (FPGA), and a program Possible logic circuits (PLDs) or combinations thereof can be used.

  The processor 101 performs arithmetic processing based on data or a program input from each device (for example, the input device 102, the communication device 104, and the storage device 105) connected via the bus 106, and outputs a calculation result and a control signal. And output to each device (for example, the display device 103, the communication device 104, and the storage device 105) connected via the bus 106. Specifically, the processor 101 executes an OS (operating system) of the computer 100, a characteristic prediction program, and the like, and controls each device included in the computer 100.

  The characteristic prediction program is a program that causes the computer 100 to execute processing relating to each component of the characteristic prediction apparatus. The characteristic prediction program is stored in a non-transitory tangible computer-readable storage medium. The storage medium is, for example, an optical disk, a magneto-optical disk, a magnetic disk, a magnetic tape, a flash memory, or a semiconductor memory, but is not limited thereto. When the processor 101 executes the characteristic prediction program, the computer 100 can function as a characteristic prediction apparatus.

  The input device 102 is a device for inputting information to the computer 100. The input device 102 is, for example, a keyboard, a mouse, a touch panel, but is not limited thereto. By using the input device 102, the user uses the brewing data preprocessing start command, the learning data change operation, the learning processing start command, the learning data variable selection operation, and the brewing condition prediction for component value prediction. It is possible to input a designation operation, a designation operation for a component value to be predicted, a command to start component value prediction processing, a command to change display contents, and the like.

  The display device 103 is a device for displaying images and videos. The display device 103 is, for example, an LCD (liquid crystal display), a CRT (CRT), an organic EL (organic electroluminescence) display, a projector, or an LED display, but is not limited thereto. As described above, the display device 103 displays the contents of brewing data and learning data, the prediction result of the component value under the brewing conditions, and the like.

  The communication device 104 is a device used by the computer 100 to communicate with an external device wirelessly or by wire. The communication device 104 is, for example, a NIC (Network Interface Card), a communication module, a modem, a hub, or a router, but is not limited thereto. The computer 100 may collect brewing data accumulated in a remote factory or laboratory via the communication device 104. When the computer 100 (characteristic prediction apparatus 1) is a server installed in a data center or a machine room, the computer 100 receives an operation command from a remote terminal or displays a screen via the communication device 104. The contents may be displayed on a remote terminal.

  The storage device 105 is a storage medium that stores the OS of the computer 100, a characteristic prediction program, data necessary for executing the characteristic prediction program, data generated by executing the characteristic prediction program, and the like. The storage device 105 includes a main storage device and an external storage device. The main storage device is, for example, a RAM, a DRAM, or an SRAM, but is not limited thereto. The external storage device is, for example, a hard disk, an optical disk, a flash memory, a magnetic tape, or the like, but is not limited thereto. The brewing database 21, the learning data storage unit 22, the determination data storage unit 23, and the model database 24 described above may be built on the storage device 105, or may be built on an external server or storage.

  Note that the computer 100 may include one or more processors 101, input devices 102, display devices 103, communication devices 104, and storage devices 105, respectively. A peripheral device such as a printer or a scanner may be connected to the computer 100.

  Moreover, the characteristic prediction apparatus may be comprised by the single computer 100, and may be comprised by the information system with which the some computer 100 was mutually connected.

  Furthermore, the characteristic prediction program may be stored in advance in the storage device 105 of the computer 100, may be stored in a storage medium external to the computer 100, or may be uploaded on the Internet. In any case, the function of the characteristic prediction apparatus can be realized by installing the characteristic prediction program in the computer 100 and executing it.

  Although several embodiments of the present invention have been described, these embodiments have been presented by way of example, and are not intended to limit the scope of the invention. Various omissions, replacements, and changes can be made without departing from the scope of the invention. These embodiments and their modifications are included in the scope and gist of the invention, and are also included in the invention described in the claims and the equivalents thereof.

DESCRIPTION OF SYMBOLS 1 Characteristic prediction apparatus 2 Memory | storage part 3 Learning part 4 Preprocessing part 5 Display part 6 Operation part 11 Variable selection part 12 Cross verification part 13 Model production | generation part 14 1st estimation part 15 2nd estimation part 16 Verification part 17 Method selection part 21 Brewing database 22 Learning data storage unit 23 Determination data storage unit 24 Model databases 30, 31, 32, 33, 33a, 34, 35, 39, 40, 41, 42, 43, 43a. 45, 46, 47, 48, 49 Rows 36, 37, 38, 44, 51a, 51b Tables 50, 51, 52 Screen 60 Samples 61 Ball 100 Computer 101 Processor 102 Input device 103 Display device 104 Communication device 105 Storage device 106 Bus

Claims (13)

Generating a tasting sample using the setting items in the beverage brewing conditions as explanatory variables;
Inputting the tasting sample into a plurality of prediction models, and obtaining an estimate of the component value of the beverage for each prediction model;
Inputting the sample sample to an error determination model corresponding to the prediction model, and calculating a probability within tolerance for each prediction model;
The method selection model is input with explanatory variables related to the tasting sample and a plurality of probabilities within the allowable error, and the component values predicted under the tasting conditions of the beverage based on the output value of the method selection model The computer executes the step of selecting the estimated value presented as:
Beverage property prediction method. Including the step of displaying the estimated value by at least one of the prediction models on the display as the component value predicted in the brewing conditions of the beverage,
The beverage characteristic prediction method according to claim 1. Including a past sample with the setting item in the beverage brewing case as an explanatory variable, generating learning data;
Dividing the learning data into a plurality of blocks, setting some blocks as test data, and setting blocks not set as the test data as training data;
Using the training data to estimate the component values of the beverage and generating a plurality of the prediction models;
Obtaining the estimated value of the component value by a plurality of the prediction models using the test data;
Determining whether the magnitude of the error between the estimated value and the true value of the component value in the brewing case is within an allowable error, and storing the result as an allowable error determination result;
The error determination model is generated for each prediction model, using the training data and estimating the probability that the magnitude of the error in the estimated value is within the allowable error using the determination result within the allowable error as a teacher signal. And steps to
Applying a plurality of the error determination models to the test data, and calculating the probability within the allowable error for each prediction model;
For the past sample included in the test data, compare the error between the estimated value and the true value by a plurality of prediction models, select the prediction model with the smallest error, and as a preferred method Saving step;
The method selection model that adds a plurality of probabilities within the allowable error to the explanatory variable of the learning data, and selects the estimated value presented as the component value of the beverage using the preferred method as a teacher signal. And a step of generating a computer,
The beverage characteristic prediction method according to claim 1 or 2. The past sample that is classified in the same category as the tasting sample in the feature space related to the learning data or the past sample that is in the vicinity of the tasting sample in the feature space is extracted, and Including the step of presenting as a related case,
The beverage characteristic prediction method according to claim 3. Including the step of displaying the setting items in the related case and the component values of the beverage generated in the related case on a display.
The beverage characteristic prediction method according to claim 4. Any of the prediction models adds, for a plurality of the past samples, a product of a transition probability between the past samples determined by K * and the component value of the brewing case corresponding to the past samples, The estimated value is calculated by dividing by the sum of the transition probabilities,
The beverage characteristic prediction method according to any one of claims 3 to 5. At least one of the error determination model and the method selection model is generated by logistic regression,
The characteristic prediction method of the drink according to any one of claims 1 to 6. Any of the prediction models is generated by M5 or M5 ′,
The beverage characteristic prediction method according to any one of claims 1 to 7. The beverage is one of beer, happoshu, new genre, and beer-taste beverage,
The method for predicting beverage characteristics according to any one of claims 1 to 8. The beverage is a beverage containing malt as a raw material,
The beverage characteristic prediction method according to any one of claims 1 to 9. The setting item includes at least information related to either the raw material of the beverage or the brewing process of the beverage,
The method for predicting characteristics of a beverage according to any one of claims 1 to 10. Generating a tasting sample using the setting items in the beverage brewing conditions as explanatory variables;
Inputting the tasting sample into a plurality of prediction models, and obtaining an estimate of the component value of the beverage for each prediction model;
Inputting the sample sample to an error determination model corresponding to the prediction model, and calculating a probability within tolerance for each prediction model;
The method selection model is input with explanatory variables related to the tasting sample and a plurality of probabilities within the allowable error, and the component values predicted under the tasting conditions of the beverage based on the output value of the method selection model Selecting the estimated value presented as:
The computer executes a step of displaying at least one of the estimated values on the display as the component value predicted in the brewing conditions of the beverage,
Characteristic prediction program. Generating a tasting sample using the setting items in the beverage brewing conditions as explanatory variables;
Inputting the tasting sample into a plurality of prediction models, and obtaining an estimate of the component value of the beverage for each prediction model;
Inputting the sample sample to an error determination model corresponding to the prediction model, and calculating a probability within tolerance for each prediction model;
The method selection model is input with explanatory variables related to the tasting sample and a plurality of probabilities within the allowable error, and the component values predicted under the tasting conditions of the beverage based on the output value of the method selection model Selecting the estimated value presented as:
Displaying at least one of the estimated values on the display as the component value predicted in the brewing conditions of the beverage,
Characteristic prediction device.

Images