Additional data generation method, additional data generation device, program, and recording medium

Abstract

To establish a method for selecting additional data to be added to learning data so as to generate a regression model that can improve prediction accuracy across a wide range.SOLUTION: The present disclosure assumes that an acquisition function used to select additional data to be added to learning data from multiple pieces of candidate data is configured to include a calculation value of a probability value calculated on the basis of a target variable value of a regression model and the similarity between the learning data and candidate data, and selects at least the candidate data with the largest calculation value as the additional data.SELECTED DRAWING: Figure 6

Description

[Technical field]

[0001] The present invention relates to an additional data generating method, an additional data generating device, a program, and a recording medium, and relates to a technology that is effective when applied to, for example, a technology for adding additional data to learning data for machine learning of a regression model. [Background technology]

[0002] Japanese Patent Application Publication No. 2020-38495 (Patent Document 1) describes a technique for increasing the number of samples by complementing the number of samples of learning input data through regression analysis. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] JP 2020-38495 A Summary of the Invention [Problem to be solved by the invention]

[0004] For example, in order to improve the prediction accuracy of a regression model, it is effective to increase the number of data included in the training data for machine learning the regression model. However, when increasing the number of data included in the training data, even if the number of similar data is increased, it is difficult to generate a regression model that can improve the prediction accuracy over a wide range by machine learning based on the training data. In other words, the quality of the generated regression model depends on the training data used for machine learning. For this reason, it is desirable to establish a method for selecting additional data to be added to the training data so that a regression model that can improve the prediction accuracy over a wide range can be generated. [Means for solving the problem]

[0005] In one embodiment, a method for generating additional data includes a step of generating additional data to be added to training data for machine learning a regression model by a computer, the method including a step of selecting additional data from a plurality of candidate data based on an acquisition function, the acquisition function including a probability value calculated based on a response variable value of the regression model and a calculation value of a similarity between the training data and the candidate data.

[0006] In one embodiment, the additional data generating device generates additional data to be added to training data for machine learning a regression model. Here, the additional data generating device includes a storage unit that stores a plurality of candidate data serving as candidates for additional data, and an additional data selecting unit that selects additional data from the plurality of candidate data stored in the storage unit based on an acquisition function. In this case, the acquisition function includes a probability value calculated based on a response variable value of the regression model, and a calculated value of a similarity between the training data and the candidate data.

[0007] A program according to an embodiment causes a computer to execute a process of generating additional data to be added to training data for machine learning a regression model. The program includes a process of selecting the additional data from a plurality of candidate data based on an acquisition function. In this case, the acquisition function includes a probability value calculated based on a response variable value of the regression model and a calculated value of a similarity between the training data and the candidate data. Effect of the Invention

[0008] According to one embodiment, a method for selecting additional data to be added to training data can be established so as to generate a regression model that can improve prediction accuracy over a wide range. [Brief description of the drawings]

[0009] [Figure 1] FIG. 1 is a diagram explaining the concept of the basic idea. [Diagram 2] FIG. 10 is a diagram illustrating a method for calculating a probability value. [Diagram 3] FIG. 10 is a diagram illustrating an inter-data distance. [Figure 4] FIG. 2 illustrates an example of a hardware configuration of an additional data generating device. [Diagram 5] FIG. 2 is a functional block diagram of an additional data generating device. [Figure 6] 10 is a flowchart illustrating an operation of the additional data generating device. [Figure 7] FIG. 13 is a diagram for explaining an overview of a method for generating additional data. [Figure 8] 13 is a table showing the search results of the first additional data. [Figure 9] 13 is a table showing the search results for second additional data. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0010] In all the drawings for explaining the embodiments, the same members are generally designated by the same reference numerals, and the repeated explanations thereof will be omitted. In addition, hatching may be applied even to plan views in order to make the drawings easier to understand.

[0011] The technical idea of ​​this embodiment is a technical idea related to training data used to machine-learn a regression model that predicts physical quantities corresponding to the blending ratios in a composite material obtained by compounding multiple types of resins and compounding agents. More specifically, the technical idea is related to a method for selecting additional data to be added to the training data from multiple candidate data that are candidates for the additional data, so that a regression model that can improve prediction accuracy over a wide range can be generated.

[0012] Here, the composite material may be, for example, a wire coating material containing a resin and compounding agents, and the physical quantity may be, for example, the elongation and tensile strength of the composite material.

[0013] The resin may be, for example, a polyolefin such as high density polyethylene, low density polyethylene, or ethylene acrylic acid copolymer, or an elastomer such as chlorinated polyethylene. On the other hand, the compounding agent may be, for example, a filler such as talc, calcium carbonate, or silica, a plasticizer, a crosslinking agent, or a stabilizer. However, the type and number of the composition such as the resin and compounding agent that constitute the composite material are not limited.

[0014] The technical idea in this embodiment is applicable not only to composite materials made by combining multiple types of resins or compounding agents, but also to composite materials made by combining multiple types of magnetic materials, and examples of physical quantities include magnetic susceptibility and magnetic field (magnetic field, magnetic flux density) strength.

[0015] <Consideration of improvements> For example, it has been considered to predict physical quantities of unknown composite materials whose physical quantities are unknown by using a regression model. Here, the regression model is a function that outputs predicted values ​​of physical quantities when compounding information including the names and compounding ratios of the constituent materials contained in the unknown composite material is input. This regression model can be generated by machine learning using training data in which the correspondence between compounding information and physical quantities is known as training data.

[0016] In this regard, the prediction accuracy of the regression model can be improved by expanding the training data. In other words, in order to obtain a regression model with high prediction accuracy, it is important to expand the training data.

[0017] There is a method called Bayesian optimization, which is a method for efficiently adding additional data to training data with the aim of expanding the training data. In this Bayesian optimization, an acquisition function is used to select additional data from multiple candidate data that are candidates for additional data. In this case, in order to be able to select effective additional data, it is necessary to design an appropriate acquisition function. In other words, in order to select effective additional data by Bayesian optimization, it is important to design an appropriate acquisition function.

[0018] Here, the valid additional data is data that contributes to generating a regression model that can improve prediction accuracy over a wide range of mixing ratios, and specifically, is data that satisfies the following conditions: (1) valid additional data is data related to unknown mixing ratios that satisfy the values ​​of required physical quantities, and (2) data that has dissimilar mixing information that is significantly different from the already acquired learning data. It is desirable to design an acquisition function that can select such valid additional data from multiple candidate data that are candidates for additional data, but it is difficult to design an acquisition function that can select valid additional data that satisfies the above conditions.

[0019] Therefore, in this embodiment, a device is devised to design an acquisition function that can select effective additional data that satisfies the above-mentioned conditions.

[0020] The technical idea behind this innovation will be explained below.

[0021] <Basic Concept of the Embodiment> The basic idea of ​​this embodiment is to select, as the additional data, at least the candidate data with the largest calculated value, on the assumption that the acquisition function used to select the additional data to be added to the learning data from a plurality of candidate data is configured to include a probability value calculated based on the objective variable value of the regression model and a calculated value of the similarity between the learning data and the candidate data. According to this basic idea, the additional data to be added to the learning data can be selected so that a regression model capable of improving prediction accuracy in a wide range can be generated. In other words, by machine learning the regression model using the learning data including the additional data selected according to the basic idea as teacher data, a regression model capable of improving prediction accuracy of the value of the physical quantity in a wide range of blending ratios can be generated. In other words, according to the basic idea, it is possible to generate learning data that can generate a regression model that is capable of outputting a predicted value with a high probability of passing the standard and has a wide range of application.

[0022] Here, according to the basic concept, the acquisition function includes a probability value calculated based on the objective variable value of the regression model, and this probability value represents the probability that the predicted value meets the standard. And, according to the basic concept, data with a large probability value is selected from a plurality of candidate data, which means that the candidate data with a high probability that the predicted value output from the regression model meets the standard is selected as the additional data. Therefore, the additional data selected based on the acquisition function in the basic concept is highly likely to meet the requirement that it is data that meets the value of the required physical quantity.

[0023] Furthermore, according to the basic concept, the acquisition function includes the similarity between the learning data and the candidate data, and data with a small similarity is selected. This means that data that is completely different from the existing data included in the learning data is selected as the additional data. Therefore, the additional data selected based on the acquisition function in the basic concept is likely to be dissimilar data that is significantly different from the already acquired learning data.

[0024] From the above, in the basic concept, as an acquisition function for selecting additional data, an acquisition function including a probability value calculated based on the objective variable value of the regression model and a quotient (probability value / similarity) between the learning data and the candidate data is used, and candidate data with a large quotient is selected as additional data from among multiple candidate data. As a result, according to the basic concept, it is possible to prepare learning data that can generate a regression model that is capable of outputting a predicted value with a high probability of passing the standard and has a wide range of applications.

[0025] The important point of the basic idea is that the acquisition function does not only include a probability value calculated based on the objective variable value of the regression model, but is configured to include a calculated value of this probability value and the similarity between the training data and the candidate data.

[0026] For example, when additional data is selected from multiple candidate data using an acquisition function including only probability values, the selected additional data tends to be similar to data already included in the training data. Therefore, even if additional data is selected using an acquisition function including only probability values ​​to increase the number of data included in the training data, the number of similar data increases, making it difficult to generate a regression model with a wide range of applications by machine learning based on the training data.

[0027] In contrast, the basic concept does not include only a probability value, but is configured to include a calculation value of the probability value and the similarity between the learning data and the candidate data, and the candidate data with a large calculation value is selected as additional data from among multiple candidate data. Therefore, in the basic concept, candidate data with a large probability value but a large similarity is not selected, while candidate data with a large probability value and a small similarity is selected. Therefore, according to the basic concept, the number of data with a small similarity can be increased, and as a result, a regression model with a wide range of applications can be generated by machine learning based on the learning data.

[0028] Here, the "calculated value of the probability value and the similarity" includes, for example, the quotient of the probability value and the similarity, or the difference between the probability value and the similarity. In the following, the quotient of the probability value and the similarity (probability value / similarity) will be taken as an example of the "calculated value of the probability value and the similarity" and explained.

[0029] FIG. 1 is a diagram explaining the basic concept.

[0030] In Fig. 1, data is classified into a number of regions A to D based on probability values ​​and similarities. For example, as shown in Fig. 1, data included in region A is data with a small probability value and small similarity. Data included in region B is data with a large probability value and small similarity. On the other hand, data included in region C is data with a small probability value and large similarity. Data included in region D is data with a large probability value and large similarity.

[0031] For example, if data included in region A is selected as additional data, this additional data is unlikely to satisfy the required physical quantity values, but is likely to be dissimilar to the learning data already acquired.

[0032] Furthermore, if data included in region B is selected as additional data, there is a high possibility that this additional data satisfies the required physical quantity values ​​and is also highly likely to be dissimilar to the learning data that has already been acquired.

[0033] Furthermore, if data included in region C is selected as additional data, this additional data is unlikely to be data that satisfies the required physical quantity values, and is likely to be data that is similar to the training data that has already been acquired.

[0034] Furthermore, if data included in region D is selected as additional data, this additional data is highly likely to be data that satisfies the required physical quantity values, but is also highly likely to be data that is similar to the learning data that has already been acquired.

[0035] Here, for example, the larger the probability value of the data, the higher the probability of passing the standard. By generating a regression model through machine learning based on this data with a high probability value, it is possible to obtain a regression model that can output a predicted value with a high probability of passing the standard.

[0036] On the other hand, by generating a regression model through machine learning based on data with low similarity, it is possible to obtain a regression model with a wide range of applications that can improve the prediction accuracy of physical quantity values ​​over a wide range of mixing ratios.

[0037] Therefore, in order to construct learning data that can generate a regression model that can output a predicted value with a high probability of passing the standard and has a wide range of applications, it is desirable to select data included in region B as additional data. In other words, the basic idea is to design an acquisition function so that data included in region B in Figure 1 is selected as additional data.

[0038] For this reason, in the basic concept, an acquisition function is configured to include a quotient between a probability value and the similarity between the learning data and the candidate data, and the candidate data with a large quotient is selected as additional data from among multiple candidate data. As a result, according to the basic concept, it is possible to actively select the candidate data included in area B shown in Fig. 1 as additional data, and as a result, it is possible to construct learning data that is a regression model that can output a predicted value with a high probability of passing the standard and can generate a regression model with a wide range of applications.

[0039] <<How to calculate probability value>> As described above, the basic idea is that the acquisition function includes a probability value calculated based on the response variable value of the regression model. The method of calculating this probability value will be described below.

[0040] In this embodiment, it is assumed that a predicted value of a physical quantity is obtained using a regression model. In this regard, depending on the type of regression model, for example, the output (objective variable) from the regression model may be a predicted value consisting of a pinpoint numerical value, but there is also a regression model that can output a pinpoint numerical value and a standard deviation. In this embodiment, for example, a regression model belonging to the latter type can be used. That is, in this embodiment, an example will be described that assumes that the output from the regression model constitutes a distribution having a width, and that this distribution is used to predict a physical quantity.

[0041] For example, there is a model called a Gaussian process regression model as a regression model. This Gaussian process regression model has a function of outputting a distribution, and can output a predicted value as a pinpoint average value and a standard deviation that determines the spread of the distribution. Therefore, in this embodiment, a Gaussian process regression model is used as the regression model, and a probability value is calculated using the average value (predicted value) and standard deviation output from the Gaussian process regression model.

[0042] FIG. 2 is a diagram showing a schematic diagram of a method for calculating a probability value.

[0043] In FIG. 2, for example, the Gaussian process regression model outputs a predicted value (average value μ) and a standard deviation value (σ) of the physical quantity. As a result, a normal distribution can be constructed based on the predicted value and the standard deviation value. Here, if a lower limit value and an upper limit value are set as the target value of the physical quantity, the area of ​​the gray region shown in FIG. 2 corresponds to the probability value of passing the standard. That is, the probability value can be calculated by obtaining the area of ​​the gray region shown in FIG. 2. For example, by applying the above-mentioned Gaussian process regression model to the candidate data, the predicted value and the standard deviation value of the physical quantity for the candidate data can be obtained. Then, in the normal distribution constructed from the obtained predicted value and the standard deviation value, the probability value in the candidate data can be calculated from the area sandwiched between the lower limit value and the upper limit value.

[0044] The probability value when both the lower limit and the upper limit are set is called "PTR (Probability in Target Range)". The probability value when only the lower limit or only the upper limit is set is called "PI (Probability of Improvement)". The probability value used in this embodiment may be either "PTR" or "PI".

[0045] Furthermore, when there are multiple types of physical quantities, such as a first physical quantity and a second physical quantity, the probability value used in the acquisition function can be the sum or product of a first probability value indicating the probability that the first physical quantity meets the standard and a second probability value indicating the probability that the second physical quantity meets the standard. In this case, each of the first probability value and the second probability value is calculated by the above-mentioned probability value calculation method. The probability value expressed by the sum or product of the first probability value and the second probability value indicates the probability that both or either one of the first physical quantity and the second physical quantity meets the respective standards.

[0046] <<Example of similarity>> In order to construct training data capable of generating a regression model with a wide range of applications, it is desirable to extract, from among multiple candidate data, candidate data that has a small degree of similarity to data already present as training data as additional data.

[0047] In this regard, the similarity is evaluated based on, for example, the data distance between the training data and the candidate data. Specifically, one example of the similarity is the "distance using the k-approximation method." The "distance using the k-approximation method" is the average value of the distance between a specific piece of data among a plurality of pieces of data and k pieces of data in the vicinity of the specific piece of data. There are various distances using the k-approximation method, such as the Euclidean distance, the Manhattan distance, the Mahalanobis distance, or the Tanimoto distance.

[0048] FIG. 3 is a diagram illustrating the inter-data distance.

[0049] In FIG. 3, there are two pieces of data in the vicinity of data A. In this case, the Euclidean distance of data A for k=2 is defined as (L1+L2) / 2. Meanwhile, in FIG. 3, there are two pieces of data in the vicinity of data B. In this case, the Manhattan distance of data B for k=2 is defined as (L3+L4) / 2. According to the data distance defined as above, it is considered that the greater the data distance is, the weaker the correlation between the data is, qualitatively, and therefore it can be said that the greater the data distance is, the weaker the similarity is. Thus, for example, the similarity can be expressed as the inverse of the data distance (1 / data distance).

[0050] For example, FIG. 3 shows an example of two-dimensional data (X1, X2). As the two-dimensional data (X1, X2), for example, data representing a composite material containing a resin and a flame retardant, such as (resin blend amount, flame retardant blend amount), can be considered. In this two-dimensional data, the distance between data is expressed as a two-dimensional distance. This concept of distance between data can be easily generalized. That is, in general n-dimensional data, such as (X1, X2...Xn), the distance between data is expressed as an n-dimensional distance. As the n-dimensional data, for example, data representing a composite material (m+j=n) containing m types of resins and j types of flame retardants, such as (amount of first resin, amount of second resin,...amount of m-th resin, amount of first flame retardant, amount of second flame retardant,...amount of j-th flame retardant), can be considered. In this way, the distance between data can be generally applied not only to two-dimensional data but also to multidimensional data.

[0051] Note that, although an example has been described here in which similarity is expressed as the inverse of the distance between data (1 / distance between data), this is not limiting, and similarity can also be expressed using cosine similarity, Jaccard coefficient, Dice coefficient, or the like.

[0052] <Embodiment> Hereinafter, an embodiment that embodies the above-mentioned basic concept will be described.

[0053] <<Configuration of additional data generator>> <<<Hardware configuration>>> The hardware configuration of the additional data generating device in this embodiment will be described.

[0054] Fig. 4 is a diagram showing an example of the hardware configuration of the additional data generation device 100 in an embodied embodiment. Note that the configuration shown in Fig. 4 is merely an example of the hardware configuration of the additional data generation device 100, and the hardware configuration of the additional data generation device 100 is not limited to the configuration shown in Fig. 4, and may be other configurations.

[0055] 4, the additional data generating device 100 includes a CPU (Central Processing Unit) 101 that executes a program. The CPU 101 is electrically connected to, for example, a ROM (Read Only Memory) 102, a RAM (Random Access Memory) 103, and a hard disk drive 112 via a bus 113, and is configured to control these hardware devices.

[0056] The CPU 101 is also connected to an input device and an output device via a bus 113. Examples of the input device include a keyboard 105, a mouse 106, a communication board 107, and a scanner 111. Examples of the output device include a display 104, a communication board 107, and a printer 110. The CPU 101 may also be connected to, for example, a removable disk device 108 and a CD / DVD-ROM device 109.

[0057] The additional data generating device 100 may be connected to, for example, a network. For example, when the additional data generating device 100 is connected to other external devices via a network, a communication board 107 constituting a part of the additional data generating device 100 is connected to a LAN (Local Area Network), a WAN (Wide Area Network), or the Internet.

[0058] The RAM 103 is an example of a volatile memory, and the recording media, ROM 102, the removable disk device 108, the CD / DVD-ROM device 109, and the hard disk device 112, are examples of non-volatile memory. The storage device of the additional data generation device 100 is configured by these volatile memories and non-volatile memories.

[0059] The hard disk device 112 stores, for example, an operating system (OS) 201, a group of programs 202, and a group of files 203. The programs included in the group of programs 202 are executed by the CPU 101 using the operating system 201. The RAM 103 temporarily stores at least a portion of the programs of the operating system 201 and application programs executed by the CPU 101, and also stores various data necessary for processing by the CPU 101.

[0060] A BIOS (Basic Input Output System) program is stored in the ROM 102, and a boot program is stored in the hard disk drive 112. When the additional data generating device 100 is started up, the BIOS program stored in the ROM 102 and the boot program stored in the hard disk drive 112 are executed, and the operating system 201 is started up by the BIOS program and the boot program.

[0061] The program group 202 stores programs that realize the functions of the additional data generating device 100, and these programs are read and executed by the CPU 101. The file group 203 stores information, data, signal values, variable values, and parameters indicating the results of processing by the CPU 101 as each item of a file.

[0062] The files are recorded on a recording medium such as the hard disk drive 112 or a memory. The information, data, signal values, variable values, and parameters recorded on the recording medium such as the hard disk drive 112 or a memory are read by the CPU 101 into the main memory or cache memory and used for the operations of the CPU 101, such as extraction, search, reference, comparison, calculation, processing, editing, output, printing, and display. For example, during the above-mentioned operations of the CPU 101, the information, data, signal values, variable values, and parameters are temporarily stored in the main memory, registers, cache memory, buffer memory, etc.

[0063] The functions of the additional data generating device 100 may be realized by firmware stored in the ROM 102, or may be realized by software alone, hardware represented by elements, devices, boards, and wiring, a combination of software and hardware, or even a combination with firmware. The firmware and software are recorded as a program on a recording medium represented by the hard disk drive 112, a removable disk, a CD-ROM, a DVD-ROM, and the like. The program is read and executed by the CPU 101. In other words, the program causes the computer to function as the additional data generating device 100.

[0064] Thus, the additional data generation device 100 is a computer equipped with the CPU 101 which is a processing device, the hard disk drive 112 and memory which are storage devices, the keyboard 105, mouse 106, and communication board 107 which are input devices, and the display 104, printer 110, and communication board 107 which are output devices. The functions of the additional data generation device 100 are realized by utilizing the processing device, the storage device, the input device, and the output device.

[0065] <<<Function block configuration>>> Next, the functional block configuration of the additional data generating device 100 will be described.

[0066] FIG. 5 is a functional block diagram of the additional data generating device 100. As shown in FIG.

[0067] The additional data generating device 100 includes an input unit 301 , a probability value calculating unit 302 , an inter-data distance calculating unit 303 , an additional data selecting unit 304 , an output unit 305 , and a data storage unit 306 .

[0068] The input unit 301 is configured to input various data and functions. Specifically, the input unit 301 is configured to input data such as learning data and candidate data, and functions such as regression models and acquired functions. The learning data, candidate data, regression models, and acquired functions input to the input unit 301 are stored in the data storage unit 306. That is, the data storage unit 306 functions as a database that stores the learning data, candidate data, regression models, and acquired functions.

[0069] Here, "learning data" refers to a data group that is composed of a known data group in which the correspondence between the physical quantities and the compounding information, including the names and compounding ratios of the constituent materials contained in the composite material, is known, and serves as training data when training a regression model by machine learning.

[0070] The "candidate data" is data that is a candidate for additional data to be added to the learning data, and is data that is composed of combination information whose corresponding physical quantity values ​​are unknown.

[0071] A "regression model" is a function that outputs a predicted value of a physical quantity when compounding information including the names and compounding ratios of the constituent materials contained in the unknown composite material is input. This regression model can be generated by machine learning using the above-mentioned learning data, in which the correspondence between compounding information and physical quantities is already known, as training data. In particular, in this embodiment, a Gaussian process regression model capable of outputting both the predicted value and the standard deviation value of the physical quantity is adopted as the "regression model".

[0072] The "acquisition function" is a function used to select additional data from multiple candidate data, and is configured to include, for example, a quotient of a probability value calculated based on the objective variable value of a regression model and a similarity between the learning data and the candidate data. In particular, in the embodiment, the "acquisition function" is configured to include a quotient (probability value / (1 / data distance)) of a probability value calculated from a predicted value and a standard deviation value output from a Gaussian process regression model and an inverse of a data distance calculated based on the learning data and the candidate data.

[0073] The probability value calculation unit 302 is configured to input the blending ratio of the candidate data to the Gaussian process regression model, and calculate the probability value of passing the standard based on the predicted value and standard deviation value of the physical quantity output from the Gaussian process regression model. Specifically, the probability value calculation unit 302 calculates the probability value by the method described in "<<Probability value calculation method>>".

[0074] The probability value calculation unit 302 calculates the probability value for each of the multiple candidate data, and the probability value corresponding to each of the multiple candidate data is stored in the data storage unit 306 in association with each candidate data.

[0075] The data distance calculation unit 303 is configured to calculate a data distance based on the learning data and the candidate data (see FIG. 3). For example, the data distance calculation unit 303 is configured to calculate a data distance between the learning data and each of a plurality of candidate data. Furthermore, the data distance calculation unit 303 is configured to obtain the reciprocal of the calculated data distance (1 / data distance).

[0076] The reciprocals of the inter-data distances calculated by the inter-data distance calculation unit 303 are stored in the data storage unit 306 in association with the corresponding candidate data.

[0077] The additional data selection unit 304 is configured to input the corresponding probability value and the inverse of the data distance to the acquisition function for each of the plurality of candidate data, thereby acquiring an output value output from the acquisition function. The additional data selection unit 304 has a function of comparing the output values ​​of the acquisition function obtained for each of the plurality of candidate data, and selecting at least the candidate data having a large output value of the acquisition function as additional data.

[0078] The additional data selected by the additional data selection unit 304 is then stored in the data storage unit 306 as a part of the learning data.

[0079] The output unit 305 has a function of displaying, for example, the additional data selected by the additional data selection unit 304. The additional data generation device 100 is configured as described above.

[0080] The additional data generating device 100 constructs learning data to which valid additional data has been added, and this learning data is used as teacher data for machine learning of a regression model. At this time, the learning data constructed by the additional data generating device 100 contains valid additional data, so that by using this learning data as teacher data to machine-train a regression model, it is possible to generate a regression model that can output a predicted value that has a high probability of passing the standard and has a wide range of applications.

[0081] Here, although not necessarily required, the additional data generating device 100 may be provided with a regression model generating unit that generates a regression model. In this case, the regression model generating unit performs machine learning using learning data including valid additional data as teacher data, thereby generating a regression model that can output a predicted value with a high probability of passing the standard and has a wide range of application. However, the regression model generating unit does not need to be provided in the additional data generating device 100, and may be provided in a computer that functions as a regression model generating device provided separately from the additional data generating device 100.

[0082] In this case, the computer is configured to acquire learning data including valid additional data from the additional data generating device 100, and perform machine learning using the acquired learning data as training data, thereby generating a regression model that is capable of outputting predicted values ​​with a high probability of passing the standards and has a wide range of applications.

[0083] <<Operation of additional data generator>> The additional data generating device 100 in the embodiment is configured as described above, and its operation will be described below.

[0084] FIG. 6 is a flowchart illustrating the operation of the additional data generating device 100. As shown in FIG.

[0085] 6, it is assumed that learning data is stored in advance in the data storage unit 306. Then, a plurality of candidate data are input to the input unit 301 (S101). The plurality of candidate data input to the input unit 301 are stored in the data storage unit 306.

[0086] Next, for example, a regression model is pre-stored in the data storage unit 306, and the additional data generation device 100 inputs candidate data into this regression model (S102), thereby acquiring predicted values ​​and standard deviation values ​​of physical quantities corresponding to the candidate data (S103).

[0087] Next, the probability value calculation unit 302 calculates the probability value of passing the standard based on the above-mentioned predicted value and standard deviation value (S104). Meanwhile, the data distance calculation unit 303 calculates the data distance based on the learning data and the candidate data, and then calculates the reciprocal value of the data distance based on the calculated data distance (S105).

[0088] Next, the additional data selection unit 304 inputs the probability value calculated by the probability value calculation unit 302 and the inverse value of the data distance calculated by the data distance calculation unit 303 into the acquisition function (S106), and stores the output value output from the acquisition function.

[0089] Thereafter, the additional data selection unit 304 judges whether or not there is a next candidate data (S107), and if there is further candidate data, the process returns to step S102. On the other hand, if there is no further candidate data, the additional data is selected (S108). Specifically, the additional data selection unit 304 compares the output value of the acquisition function obtained for each of the multiple candidate data, and selects at least the candidate data having the largest output value of the acquisition function as the additional data. Then, the additional data generation device 100 adds the additional data selected by the additional data selection unit 304 as a part of the learning data (S109). That is, the additional data is stored in the data storage unit 306 as a part of the learning data. In this manner, according to the additional data generation device 100, an operation of adding additional data to the learning data can be performed.

[0090] At this stage, the additional data selected by the additional data selection unit 304 is data composed of combination information whose corresponding physical quantity values ​​are unknown. For this reason, in order to add additional data to learning data composed of a data group whose correspondence between combination information and physical quantities is known, it is necessary to obtain the physical quantities corresponding to the combination information of the additional data by experiment, and it goes without saying that after the physical quantities are obtained by experiment, the additional data is added as part of the learning data.

[0091] <<Additional Data Generation Program>> The additional data generation method performed by the above-mentioned additional data generation device 100 can be realized by an additional data generation program that causes a computer to execute a process of generating additional data to be added to training data, which is training data for machine learning of a regression model.

[0092] For example, in the additional data generating device 100 consisting of a computer shown in Fig. 4, an additional data generating program in the embodied aspect can be introduced as one of the programs 202 stored in the hard disk drive 112. Then, by having the computer that is the additional data generating device 100 execute this additional data generating program, the additional data generating method in the embodied aspect can be realized.

[0093] The additional data generation program that causes a computer to execute each process for creating data related to the additional data generation process can be recorded on a computer-readable recording medium and distributed. Recording media include magnetic storage media such as hard disks and flexible disks, optical storage media such as CD-ROMs and DVD-ROMs, and hardware devices such as non-volatile memories such as ROMs and EEPROMs.

[0094] <<Examples>> FIG. 7 is a diagram for explaining an outline of a method for generating additional data in a specific example.

[0095] In Fig. 7, first, in a specific example, it is assumed that there is training data with the number of data (n) being n = 235. Then, a first regression model is generated by machine learning using this training data as teacher data. Meanwhile, multiple candidate data with the number of data (n) being n = 163800 are prepared based on a mixture (mixture ratio) with unknown physical quantity values.

[0096] Then, for each of the plurality of candidate data, the probability value and the inverse of the data distance are calculated, and the calculated probability value and the inverse of the data distance are input to an acquisition function to obtain an output value output from the acquisition function. That is, an output value from the acquisition function corresponding to each of the plurality of candidate data is obtained. Next, the first additional data is selected based on the output value of the acquisition function obtained corresponding to each of the plurality of candidate data. Specifically, the candidate data with the largest output value of the acquisition function is selected as the first additional data.

[0097] The selected first additional data is incorporated into the learning data, and as a result, learning data with a data number (n) of n=236 is constructed. Then, a second regression model is generated by machine learning using this learning data as teacher data. Meanwhile, for each of the plurality of candidate data, the number of data (n) of which is n=163799 excluding the first additional data from the plurality of candidate data described above, the probability value and the inverse of the distance between the data are calculated, and the calculated probability value and the inverse of the distance between the data are input to the acquisition function, thereby obtaining an output value output from the acquisition function. That is, an output value from the acquisition function corresponding to each of the plurality of candidate data is obtained. Next, the second additional data is selected based on the output value of the acquisition function obtained corresponding to each of the plurality of candidate data. Specifically, the candidate data with the largest output value of the acquisition function is selected as the second additional data. The selected second additional data is incorporated into the learning data, and as a result, learning data with a data number (n) of n=237 is constructed. Then, a third regression model is generated by machine learning using this learning data as teacher data.

[0098] Thereafter, the same process is repeated.

[0099] In this manner, according to the specific example, the learning data is repeatedly expanded by adding useful additional data having the largest output value from the acquisition function ("addition of first additional data" -> "addition of second additional data"). As a result, the regression model is updated, for example, from "first regression model" -> "second regression model" -> "third regression model" by machine learning using the repeatedly expanded learning data as teacher data. This makes it possible to generate a regression model that can output a predicted value with a high probability of passing the standard and has a wide range of application by repeatedly expanding the learning data.

[0100] A specific example will be described below.

[0101] In the specific example, an example in which the candidate data is composed of 19-dimensional data will be described. For example, each of the plurality of candidate data uses three types of resins out of 14 types of resins and one type of flame retardant out of five types of flame retardants. That is, each of the plurality of candidate data is data of a composite material containing three types of resins out of resin A to resin N and one type of flame retardant out of flame retardants O to S in a predetermined blending ratio.

[0102] 1. Search for the first additional data FIG. 8 is a table showing the search results for the first additional data.

[0103] Figure 8 shows the candidate data for the top 10 combinations with the highest output values ​​of the acquisition function.

[0104] In FIG. 8, for example, candidate data #1 having the highest output value of the acquisition function is data containing resin K, resin M, resin N and flame retardant R, with a mixing ratio of resin K:resin M:resin N:flame retardant R=30 parts by mass:20 parts by mass:50 parts by mass:200 parts by mass, and this candidate data #1 is selected as the first additional data.

[0105] 2. Search for the second additional data FIG. 9 is a table showing the search results for the second additional data.

[0106] Figure 9 shows the candidate data for the top 10 combinations with the highest output values ​​of the acquisition function.

[0107] In Figure 9, for example, candidate data #2 with the highest output value of the acquisition function is data that contains resin C, resin F, resin N and flame retardant Q, with a mixing ratio of resin C:resin F:resin N:flame retardant Q = 20 parts by mass:40 parts by mass:40 parts by mass:120 parts by mass, and this candidate data #2 is selected as the second additional data.

[0108] 3. Comparison between Figure 8 and Figure 9 Comparing FIG. 8 showing the first search result with FIG. 9 showing the second search result, candidate data #1 is selected as the first additional data, while candidate data #2 is selected as the second additional data. And, candidate data #2, candidate data #7, and candidate data #10, which have a different combination from candidate data #1 in FIG. 8, are listed at the top in FIG. 9. On the other hand, candidate data #3, candidate data #4, candidate data #5, candidate data #6, candidate data #8, and candidate data #9, which were listed at the top in FIG. 8, are not listed in the top 10 combinations in FIG. 9 because their combinations are similar to candidate data #1 selected as the first additional data.

[0109] From the above, it is possible to select candidate data with low similarity in the mixture (mixture ratio) in Figures 8 and 9, and it can be seen that, according to the specific example, additional data to be added to the learning data can be selected so that a regression model that can improve prediction accuracy over a wide range can be generated. Therefore, according to the specific example, it is supported that it is possible to efficiently expand the learning data (teacher data) that can generate a regression model that is capable of outputting a predicted value with a high probability of passing the standard and has a wide range of application.

[0110] <<Features of the Realization>> Next, features of the embodiment will be described.

[0111] The characteristic feature of the embodiment is that the acquisition function used to select additional data to be added to the learning data serving as teacher data is composed of a quotient of a probability value of passing the standard and an inverse of a data distance indicating the similarity between the data. As a result, for example, by selecting, as additional data, at least the candidate data having the largest output value from the acquisition function among a plurality of candidate data serving as additional data, data having a high probability value and a small similarity can be added to the learning data. As a result, according to the characteristic feature, it is possible to construct learning data that can generate a regression model that is capable of outputting a predicted value having a high probability of passing the standard and has a wide range of application. In other words, according to the embodiment, it is possible to efficiently expand the learning data serving as teacher data so as to realize an improvement in prediction accuracy and an increase in the range of application in the regression model.

[0112] The invention made by the inventor has been specifically described above based on the embodiment thereof. However, it goes without saying that the present invention is not limited to the above embodiment, and various modifications can be made without departing from the spirit of the invention. [Explanation of symbols]

[0113] 100 Additional data generating device 101 CPU 102 ROM 103 RAM 104 Display 105 Keyboard 106 Mouse 107 Communication Board 108 Removable disk device 109 CD / DVD-ROM device 110 Printer 111 Scanner 112 Hard disk drive 113 Bus 201 Operating Systems 202 Programs 203 Files 301 Input section 302 Probability Value Calculation Unit 303 Data Distance Calculation Unit 304 Additional Data Selection Section 305 Output section 306 Data Storage Unit

Claims

1. An additional data generation method in which a computer generates additional data to be added to learning data for machine learning a regression model, comprising the steps of: A step of selecting the additional data from a plurality of candidate data serving as the additional data based on an acquisition function by a computer; Equipped with The additional data generating method, wherein the acquisition function includes a probability value calculated based on a target variable value of the regression model and a calculated value of a similarity between the training data and the candidate data.

2. 2. The additional data generating method according to claim 1, The similarity is evaluated based on a data distance between the training data and the candidate data.

3. 2. The additional data generating method according to claim 1, The additional data generating method, wherein the step of selecting additional data includes selecting, as the additional data, the candidate data having the largest value of the calculated value.

4. 2. The additional data generating method according to claim 1, the regression model is a Gaussian process regression model; The method for generating additional data, wherein the probability value is calculated based on an estimate and a standard deviation value output from the regression model.

5. 2. The additional data generating method according to claim 1, An additional data generation method, in which the regression model is a function that outputs a predicted value of a physical quantity when composition information including the material names and composition ratios of constituent materials contained in an unknown composite material whose physical quantity value is unknown is input.

6. An additional data generating device that generates additional data to be added to learning data for machine learning a regression model, A storage unit that stores a plurality of candidate data items that are candidates for the additional data item; an additional data selection unit that selects the additional data from the plurality of candidate data stored in the storage unit based on an acquisition function; Equipped with The acquisition function includes a probability value calculated based on a target variable value of the regression model and a calculated value of a similarity between the training data and the candidate data.

7. A program for causing a computer to execute a process of generating additional data to be added to learning data for machine learning a regression model, A process of selecting the additional data from a plurality of candidate data serving as the additional data based on an acquisition function; Equipped with The acquisition function includes a probability value calculated based on a target variable value of the regression model and a calculated value of a similarity between the learning data and the candidate data.

8. A computer-readable recording medium having the program according to claim 7 recorded thereon.

Images