JP7232122B2 - Physical property prediction device and physical property prediction method - Google Patents

Description

The present invention relates to a physical property prediction device and a physical property prediction method, and is suitable for application to a physical property prediction device and a physical property prediction method for predicting physical properties to support design in materials informatics, optimization of manufacturing conditions, etc. be.

Recently, it is expected to shorten the design period and discover new materials and new formulations by utilizing information technology (IT) and artificial intelligence (AI) in materials informatics and optimization of manufacturing conditions. As one of such applications of IT and AI, great expectations are placed on prediction of physical properties based on machine learning.

For example, Patent Literature 1 discloses a physical property prediction device that predicts physical properties of compounds. In the physical property prediction apparatus disclosed in Patent Document 1, among learning samples in which parameter values related to chemical structures and values for prediction items are registered in advance, learning samples whose similarity is equal to or higher than a predetermined threshold are taken out and a sub-sample set is obtained. is constructed, data analysis is performed on the sub-sample set to create a prediction model, and this prediction model is applied to unknown samples to calculate the value of the prediction item (prediction value). At that time, if the number of training samples in the subsample set is less than the number of parameters in the final parameter set, the similarity threshold used to construct the subsample set is changed to increase the reliability of the prediction model. It is disclosed to increase the number of training samples included in the sub-sample set by doing so.

Japanese Patent No. 5083320

However, in the case of the physical property prediction device disclosed in Patent Document 1, the information of the learning sample (training data) whose similarity is less than the threshold is not used for the sub-sample set, so the generalization error of physical property prediction is large. There was a possibility that Moreover, since the physical property prediction apparatus of Patent Document 1 needs to calculate the degree of similarity each time, there is a possibility that the processing capacity may decrease due to an increase in the processing load. In addition, with the physical property prediction apparatus of Patent Document 1, since the statistical degree of risk (degree of blurring) in the prediction is not known, there is also the problem that it is difficult to judge the reliability of the calculated predicted value.

The present invention has been made in consideration of the above points, and intends to propose a physical property prediction apparatus and a physical property prediction method capable of suppressing prediction generalization errors and managing prediction risks in property prediction. be.

In order to solve this problem, in the present invention, a model learning unit that learns a model to be used for physical property prediction, and the model learned by the model learning unit is used for an unknown input vector that is input as unknown sample data. A physical property prediction device including a physical property prediction unit that performs the physical property prediction and a display unit that displays the result of the physical property prediction is provided. In this physical property prediction device, the model learning unit clusters an input set of training data used for learning the model, and includes a representative vector selection unit that selects a representative vector in each cluster; 1 A base model learning unit that learns a base model for predicting physical property values using a predetermined number of the training data, and a second predetermined number of the training data near each of the representative vectors to learn the base model and a correction model learning unit that learns a correction model for predicting the inverse of the residual of the base model for each model. Further, the physical property prediction unit includes, for the unknown input vector, a model search unit for searching the base model and the correction model for the representative vector close to the unknown input vector; a base model prediction unit that calculates a base model prediction value as a prediction value; a correction model prediction unit that calculates a correction model prediction value as a prediction value of the correction model for the unknown input vector; and the base model prediction value and the correction model. Prediction result determination for calculating the predicted value of physical properties for each physical property by summing the predicted value and the value obtained by multiplying the predicted value by a predetermined constant, and calculating the degree of correction indicating the risk of the predicted value of the physical property based on the corrected model predicted value. and The display unit provides a prediction result display screen that displays at least the predicted physical property value and the degree of correction calculated by the prediction result determination unit.

Further, in order to solve such problems, in the present invention, there are provided a model learning step for learning a model to be used for physical property prediction; and a display step of displaying the result of the physical property prediction. In this physical property prediction method, the model learning step includes a representative vector selection step of clustering an input set of training data used for learning the model and selecting a representative vector in each cluster; 1 A base model learning step of learning a base model for predicting physical property values using a predetermined number of the training data, and a second predetermined number of the training data near each of the representative vectors to learn the base model and a correction model learning step of learning a correction model that predicts the inverse of the residual of the base model for each. Further, the physical property prediction step includes, for the unknown input vector, a model search step for searching the base model and the correction model for the representative vector close to the unknown input vector; a base model prediction step of calculating a base model prediction value as a prediction value; a correction model prediction step of calculating a correction model prediction value as a prediction value of the correction model for the unknown input vector; A physical property predicted value is calculated for each physical property by summing the base model predicted value and a value obtained by multiplying the corrected model predicted value calculated in the corrected model prediction step by a predetermined constant, and the corrected and a prediction result determination step of calculating a correction degree indicating a risk of the predicted physical property value based on the model predicted value. In the display step, a prediction result display screen is provided that displays at least the predicted physical property value and the degree of correction calculated in the prediction result determination step.

ADVANTAGE OF THE INVENTION According to this invention, in physical property prediction, while suppressing the general-purpose error of prediction, the risk of prediction can be managed.

BRIEF DESCRIPTION OF THE DRAWINGS It is a block diagram which shows the system configuration|structure and functional configuration of the physical-property prediction apparatus which concerns on one embodiment of this invention. 2 is a configuration diagram showing a hardware configuration of the physical property prediction device shown in FIG. 1; FIG. 4 is a flowchart showing an outline of a processing procedure of model learning processing; FIG. 4 is a diagram showing a data configuration example of an input set of experimental data; FIG. 4 is a diagram showing a data configuration example of an output set of experimental data; It is a figure which shows the specific example of a correction degree adjustment screen. 9 is a flowchart showing a detailed processing procedure example of prediction model learning processing; 8 is a flowchart illustrating an example of a processing procedure of neighborhood definition processing; 6 is a flow chart showing an example of a processing procedure of physical property prediction processing; FIG. 4 is a diagram showing a data configuration example of an input set of unknown input vectors; It is a figure which shows the data structural example of prediction result data. It is a figure which shows the specific example of a prediction result display screen. FIG. 3 is a block diagram showing the process of calculating physical property prediction values and risk values in a basic form; FIG. 11 is a block diagram showing a process of calculating physical property prediction values and risk values in a modified example;

Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.

(1) Configuration of Physical Property Prediction Apparatus FIG. 1 is a configuration diagram showing the system configuration and functional configuration of a physical property prediction apparatus according to an embodiment of the present invention. The physical property prediction apparatus 1 according to the present embodiment learns prediction models (base model and correction model) in the learning phase, and searches based on the input unknown sample data (unknown input vector) in the operation phase. A device that predicts physical properties using a suitable prediction model and outputs the prediction results, and is realized by, for example, a general calculation server (and its peripherals).

As shown in FIG. 1, a physical property prediction device 1 is connected via a network 3 to a terminal 2 used by an engineer so as to be communicable. An engineer means a user who uses the physical property prediction apparatus 1 through the operation of the terminal 2, and is uniformly written as an engineer in the following description. The network 3 is, for example, a LAN (Local Area Network), but is not limited to a specific network and may be wired or wireless. For example, the physical property prediction device 1 and the terminal 2 may be connected via WWW (World Wide Web).

Although details will be described later, the physical property prediction device 1 includes a model learning unit 11, a physical property prediction unit 12, a display unit 13, and a data management unit 14 as functional configurations. The model learning unit 11 includes a representative vector selection unit 15, a base model learning unit 16, and a correction model learning unit 17. The physical property prediction unit 12 includes a model search unit 18, a base model prediction unit 19, a correction model Although it has a prediction unit 20 and a prediction result determination unit 21, it may further have a functional unit (not shown).

FIG. 2 is a configuration diagram showing the hardware configuration of the physical property prediction device shown in FIG. As shown in FIG. 2, the physical property prediction apparatus 1 includes a CPU (Central Processing Unit) 101, a ROM (Read Only Memory) 102, a RAM (Random Access Memory) 103, an external storage device 104, a communication I / F 105, an external input A device 106 and an external output device 107 are provided. The external storage device 104 corresponds to a storage device such as a HDD (Hard Disk Drive) or SSD (Solid State Drive). corresponds to an output device such as a display or a printer.

Regarding the physical property prediction apparatus 1, the relationship between the hardware configuration shown in FIG. 2 and the functional configuration shown in FIG. It is realized by controlling.

Other functional configurations of the physical property prediction apparatus 1 are such that the CPU 101 reads a program stored in the ROM 102 or the external storage device 104 into the RAM 103, reads data stored in the ROM 102, the RAM 103, or the external storage device 104 as necessary. , and controls the communication I/F 105 , the external input device 106 , or the external output device 107 .

In particular, the display unit 13 enables output such as screen display to the external output device 107 under the control of the CPU 101, and outputs output information via the network 3 in a predetermined output format (for example, GUI: Graphic User Interface). can also be provided externally (for example, terminal 2). At this time, the engineer can operate the terminal 2 to access the output information and perform various operations on the GUI, and the CPU 101 of the physical property prediction apparatus 1 can accept the operation result and execute further processing. can. In addition, although "image display" will be described in the following description, this embodiment can be replaced with an output form other than image display (for example, printing or data output).

It should be noted that part of the hardware configuration shown in FIG. Alternatively, it may be replaced with a peripheral device or device externally connected to the calculation server. Also, since the hardware configuration of the terminal 2 can be considered to be the same as that of the physical property prediction device 1, detailed description thereof will be omitted.

Returning to FIG. 1, each functional configuration of the physical property prediction device 1 will be described.

The model learning unit 11 in the physical property prediction device 1 has a model learning function of learning a prediction model (base model and correction model) by executing processing (model learning processing) in the learning phase.

Among the functional components included in the model learning unit 11, the representative vector selection unit 15 has a function of clustering an input set of training data learned in the model learning process and selecting a representative vector of each cluster. The training data learned in the model learning process is selected from past experimental data accumulated in the data management unit 14 .

The base model learning unit 16 also has a function of learning a base model based on the representative vector. Specifically, the base model learning unit 16 selects, for each representative vector selected by the representative vector selection unit 15, a predetermined number ("K" in the present embodiment) of training data near the representative vector. Use (base model training data) to learn a base model that predicts physical property values. A base model learned by the base model learning unit 16 is registered in the data management unit 14 .

Further, the correction model learning unit 17 has a function of learning a correction model. Specifically, the correction model learning unit 17 selects, for each representative vector selected by the representative vector selection unit 15, a predetermined number (in this embodiment, "L") of training data in the vicinity of the representative vector. (correction model training data) is used to learn a correction model that predicts the inverse of the residual of each base model. The correction model learned by the correction model learning unit 17 is registered in the data management unit 14 .

Note that the functional unit of the model learning unit 11 described above is an example of a functional configuration for realizing a representative part of the model learning functions of the model learning unit 11, and the model learning unit 11 , in order to realize "other functions (processing)" included in the model learning function, it may further have a functional unit (not shown) (or any of the above functional units may be substituted). . "Other processing" specifically includes, for example, processing for receiving experimental data (step S11 in FIG. 3), processing for selecting training data from experimental data (step S12 in FIG. 3), and preprocessing of training data. The processing (step S13 in FIG. 3) and the processing for searching for a constant (generalization error minimizing constant) for minimizing the generalization error of the predicted physical property value (step S24 in FIG. 7) correspond to each will be detailed later. In the following description, such processing by a functional unit (not shown) is described with the model learning unit 11 as the processing entity.

In the physical property prediction device 1, the physical property prediction unit 12 executes processing in the operation phase (physical property prediction processing) to perform physical property prediction using a suitable prediction model searched based on the unknown input vector, and the prediction result is It has a physical property prediction function to output to the display unit 13 .

Among the functional components included in the physical property prediction unit 12, the model search unit 18 has a function of searching for a model used for physical property prediction for an unknown input vector. Specifically, the model search unit 18 retrieves a base model (corresponding base model) and a correction model (corresponding correction model) related to a representative vector close to the unknown input vector from among the base models and correction models registered in the data management unit 14. Search for Here, a model "associated with a representative vector" means a model learned based on the representative vector (using training data in the vicinity of the representative vector) in the learning phase.

The base model prediction unit 19 also has a function of calculating a prediction value (base model prediction value) of the corresponding base model for the unknown input vector using the corresponding base model searched by the model search unit 18 .

The correction model prediction unit 20 also has a function of calculating a prediction value (correction model prediction value) of the corresponding correction model for the unknown input vector using the corresponding correction model searched by the model search unit 18 .

The prediction result determination unit 21 has a function of determining a prediction result of physical property prediction for an unknown input vector based on the base model prediction value calculated by the base model prediction unit 19 and the correction model prediction value calculated by the correction model prediction unit 20. have Specifically, first, the prediction result determination unit 21 calculates a prediction value for each physical property (physical property prediction value). Also, the prediction result determination unit 21 calculates the correction degree of each physical property prediction value from the absolute value of the correction model prediction value. The degree of correction of the physical property predicted value is a value indicating the degree of statistical risk (blurring degree) of the predicted value for each physical property. Furthermore, the prediction result determination unit 21 performs a predetermined arithmetic processing on the calculated correction degree of each physical property prediction value, and the degree of risk (blurring degree) of the prediction value in the current physical property prediction as a whole is a “risk value”. Calculate as

The functional unit of the physical property prediction unit 12 described above is an example of a functional configuration for realizing a representative part of the physical property prediction functions of the physical property prediction unit 12, and the physical property prediction unit 12 , in order to realize "other functions (processing)" included in the physical property prediction function, it may further have a functional unit (not shown) (or any of the above functional units may be substituted). . "Other processing" specifically corresponds to, for example, processing for receiving an unknown input vector (step S41 in FIG. 9). In the following description, processing by such a functional unit (not shown) is described with the physical property prediction unit 12 as the main processing entity.

The display unit 13 has a function of displaying instructed information based on an output instruction from the model learning unit 11 or the physical property prediction unit 12 . Specifically, for example, the display unit 13 displays a prediction result display screen (see FIG. 13) showing the result of physical property prediction in the physical property prediction process, or displays the correction degree level on the prediction result display screen in the model learning process. A correction degree adjustment screen (see FIG. 6) that allows adjustment of classification criteria and the like is displayed.

The data management unit 14 has a function of storing and managing various data used in the physical property prediction device 1 . 4, 5, 10 and 11 show examples of data stored and managed by the data management unit 14. FIG.

(2) Learning phase (model learning processing)
A processing procedure in the learning phase of the physical property prediction device 1 will be described with reference to FIG. 3 and the like. FIG. 3 is a flowchart showing an outline of the processing procedure of model learning processing. As described above, the model learning process is mainly executed by each part of the model learning section 11 .

According to FIG. 3, first, when the engineer inputs experimental data from the terminal 2, the model learning unit 11 receives and stores the data in the data management unit 14 (step S11). Here, the experimental data is data related to experiments conducted in the past, and is classified into input information indicating data such as raw materials and conditions used in the experiment, and output information indicating data of physical properties measured in the experiment. can do. In this embodiment, the data management unit 14 collectively manages the input information and the output information of the experiment into an input set and an output set, respectively.

FIG. 4 is a diagram showing a data configuration example of an input set of experimental data, and FIG. 5 is a diagram showing a data configuration example of an output set of experimental data. The input set 110 shown in FIG. 4 includes an experiment number 111 indicating the identifier of the experiment, raw materials 112 and 113 indicating the usage amount (usage rate, etc.) of each raw material used in the experiment, and each condition in the experiment. and conditions 114 and 115 indicating The output set 120 shown in FIG. 5 includes an experiment number 121 indicating the identifier of the experiment, similar to the experiment number 111 in FIG. Configured. The data configurations shown in FIGS. 4 and 5 are examples, and the storage form of experimental data in the present embodiment is not limited to the above examples.

After step S11, the model learning unit 11 selects training data to be learned in the model learning process from the experimental data accumulated in the data management unit 14 in step S11 (step S12). What kind of training data is selected in step S12 is arbitrary, and may be selectable by an engineer, for example. Further, since the training data is selected from the experimental data, the data configuration of the training data input set can be considered to be the same as the experimental data input set 110 shown in FIG. The data organization can be considered similar to the output set 120 of experimental data shown in FIG.

Next, the model learning unit 11 preprocesses the training data selected in step S12 (step S13). The type of preprocessing executed in step S13 is not limited, but includes, for example, one-hot vectorization and normalization of categorical data (qualitative data) included in the training data.

Next, the model learning unit 11 uses the training data preprocessed in step S13 to execute a prediction model learning process for learning a prediction model (base model and correction model) (step S14). Details will be described later with reference to FIG. ) is performed.

Next, the model learning unit 11 associates the representative vector, the learned base model and correction model, and the generalization error minimization constant (constant α) based on the processing result of the prediction model learning processing in step S14. is registered in the data management unit 14 (step S15).

Finally, the model learning unit 11 instructs the display unit 13 to display and output a correction level adjustment screen on which the correction level classification criteria and display mode can be adjusted on the prediction result display screen, and the engineer can operate the screen. If there is, appropriate adjustment is made (step S16), and then the model learning process is terminated.

Here, the correction degree adjustment screen will be supplemented. A specific example is shown in FIG. 13, but when physical property prediction is performed in the operation phase, the physical property prediction device 1 according to the present embodiment displays the predicted value of the physical property on the prediction result display screen that displays the result. It is possible to display the degree of correction (correction level) that indicates the risk of recognizable. Then, in order to enable the engineer to set the display mode of the "correction degree" on the prediction result display screen in advance, a correction degree adjustment screen is displayed in step S16 of the model learning process. In this example, the display of the predicted values of physical properties in a manner in which the degree of correction level can be recognized on the prediction result display screen will be described later in detail. It may be displayed in a similar manner.

FIG. 6 is a diagram showing a specific example of the correction degree adjustment screen. As shown in FIG. 6, the correction degree adjustment screen 210 displays a distribution chart having the average correction degree and the absolute value of the residual on both axes. As shown in legend 211, the display mode of the data in the distribution map differs depending on the correction degree level ("small correction degree", "medium correction degree", and "large correction degree"). Further, on the correction degree adjustment screen 210, slide bars 212 and 213 are provided to allow movement of the correction degree level classification criteria. Therefore, the engineer can freely adjust the correction degree level classification criteria by operating the slide bars 212 and 213 . In this example, different patterns are displayed depending on the correction level, but the present embodiment is not limited to this, as long as the correction level can be identified. For example, they may have different colors and sizes, or they may be combined. Furthermore, it may be possible to select a display mode. The setting information adjusted on the correction degree adjustment screen 210 is registered in the data management unit 14 and reflected in the display settings of the prediction result display screen (see prediction result display screen 220 in FIG. 12).

(2-1) Prediction Model Learning Processing The prediction model learning processing in step S14 of FIG. 3 will be described in detail. FIG. 7 is a flowchart illustrating a detailed processing procedure example of prediction model learning processing.

According to FIG. 7, first, the representative vector selection unit 15 of the model learning unit 11 clusters the input set of training data preprocessed in step S13 of FIG. 3, and selects a representative vector in each cluster (step S21). The selection of the representative vector is not limited to a specific method. For example, it is conceivable to use the k-means method to select a vector indicating the center of gravity or the center of the target cluster as the representative vector of the cluster. Alternatively, for example, a method such as hierarchical clustering or DBSCAN (Density-based Spatial Clustering of Applications with Noise) may be used to select a representative vector.

Next, for each of the plurality of representative vectors selected in step S21, the base model learning unit 16 of the model learning unit 11 selects a predetermined number (K) of training data (base model training data) in the vicinity of the representative vector. ) is used to learn a base model for predicting physical property values (step S22).

A detailed processing procedure example of step S22 will be described. First, the base model learning unit 16 initializes the base model training data to an empty set. Second, the base model learning unit 16 adds training data belonging to the cluster to which the representative vector belongs to the base model training data. In the second procedure, the training data belonging to the cluster are added to the base model training data in order from those in the vicinity of the representative vector. Proceed to step 4. On the other hand, if the number of training data added to the base model training data at the completion of the second procedure is less than K, proceed to the third procedure to add more training data to the base model training data. In the third procedure, the base model learning unit 16 performs base model training on unadded training data in the vicinity of the representative vector until the total reaches K, regardless of whether the representative vector belongs to the cluster to which the base model learning unit 16 belongs. Add to the data and proceed to step 4 after completion. In the fourth procedure, the K base model training data added through the previous procedures are used to learn the base model. By repeating the above processing for each representative vector, the base model learning unit 16 can learn a (corresponding) base model for each representative vector.

Next, for each of the plurality of representative vectors selected in step S21, the correction model learning unit 17 of the model learning unit 11 selects a predetermined number (L) of training data (correction model training data) in the vicinity of the representative vector. ) is used to learn a correction model that predicts the reciprocal of the residual of the base model with respect to the representative vector (step S23). Here, the correction model that predicts the inverse of the residual of the base model is, in other words, a correction model that cancels the deviation of the true value of the base model. It is preferable that the number "L" of training data included in the training data is larger than the number (K) of training data included in the base model training data.

Next, the model learning unit 11 linearly searches for the optimum generalization error minimizing constant (constant α) that minimizes the generalization error of the predicted physical property value for each representative vector (step S24).

A supplementary description of step S24 will be given. In the present embodiment, the predicted values of physical properties are calculated based on the predicted values of the prediction model learned using the training data near the representative vector, that is, the predicted values of the base model and the predicted values of the correction model. At this time, the predicted value of the physical property is represented by the sum of the predicted value of the base model and the predicted value of the correction model multiplied by the "constant α". Then, in step S24, the value of the "constant α" is set between "0" and "1" by linear search so that the generalization error estimated by such predicted physical property values is as small as possible. Explore. Specifically, for example, among the output set 120 (see FIG. 5) indicating the physical property data measured in the experiment, the linear search is performed while referring to the output information corresponding to the input information close to the representative vector of the base model and the correction model. , it is possible to search for the optimal constant α that minimizes the estimated generalization error. The constant α found in this way is called a generalization error minimization constant.

When the process of step S24 is completed, the model learning unit 11 terminates the prediction model learning process.

(2-2) Definition of Neighborhood By the way, in the prediction model learning process described above, in steps S22 and S23, training data to be used for model learning is selected on the basis of being in the "neighborhood" of the representative vector. In the physical property prediction apparatus 1 according to the present embodiment, it is possible to arbitrarily define how to determine this "neighborhood" by using known metric learning or the like.

Specific methods include distance metrics such as Euclidean distance and Mahalanobis distance, metrics defined for one or more variables (dimensions) that contribute significantly (more than a predetermined degree) to physical properties (output), or kernels A metric or the like defined by a function or the like can be used to define the neighborhood.

Also, as another specific method, for example, the neighborhood can be defined using a metric obtained by distance metric learning represented by MLKR (Metric Learning for Kernel Regression). According to this method, the "distance metric" can be treated as having variability and the "distance metric" itself can be learned.

FIG. 8 is a flowchart illustrating an example of a processing procedure for neighborhood definition processing. The neighborhood determination process shown in FIG. 8 is an example of a metric that uses distance metric learning to define a neighborhood. The process shown in FIG. 8 is executed by the model learning unit 11 (eg, the base model learning unit 16 and the correction model learning unit 17) between steps S13 and S14 in FIG. 3, for example. Note that the model learning unit 11 may be considered to have a new functional unit (distance metric learning unit) capable of executing distance metric learning.

According to FIG. 8, first, the model learning unit 11 reduces the dimension of the data set for judging the neighborhood based on the variable importance of the decision tree (step S31). Specifically, for example, in an input set of training data, the dimensions (raw materials and materials) of each training data are reduced. Note that the variable importance of the decision tree is given by the degree of decrease in the degree of irregularity in the decision tree.

Next, the model learning unit 11 performs data conversion by MLKR as distance metric learning (step S32). MLKR can obtain a covariance matrix of Mahalanobis distances that minimizes the generalization error in terms of kernel linear regression. Then, in step S32, by transforming the data using one of the two matrices obtained by decomposing the covariance matrix into two matrices, the "Euclidean distance in the transformed space" can be obtained by "following the original data with the covariance matrix." distance given by the Mahalanobis distance”. Therefore, the model learning unit 11 can determine the neighborhood based on the Mahalanobis distance by using the metric obtained in the distance metric learning in step S32, that is, by determining the neighborhood based on the Euclidean distance in the space. Become.

Note that when "reducing the dimension" in step S31, a technique of principal component analysis (PCA) may be used. Alternatively, the dimensionality may be reduced by MLKR.

As another example, only one of steps S31 and S32 shown in FIG. 8 may be performed.

In addition to the above, other metrics or kernel functions may be defined so that neighbors can be determined using their distances or inner products.

(3) Operation phase (physical property prediction processing)
A processing procedure in the operation phase of the physical property prediction device 1 will be described with reference to FIG. 9 and the like. FIG. 9 is a flowchart illustrating an example of the procedure of physical property prediction processing. As described above, the physical property prediction process is mainly performed by each part of the physical property prediction unit 12. FIG.

According to FIG. 9, first, when an engineer inputs an unknown input vector as unknown sample data from the terminal 2, the physical property prediction unit 12 receives it (step S41). In the present embodiment, the data management unit 14 manages the input unknown input vector as an input set in which the input unknown input vector is associated with the experiment number. The "experiment" in the operation phase means an experiment in which physical properties are predicted for an unknown input vector, and is different from the experiment with experimental data input in the learning phase.

FIG. 10 is a diagram showing a data configuration example of an input set of unknown input vectors. The input set 130 shown in FIG. 10 comprises an experiment number 131 indicating an experiment identifier, and raw materials 132, 133 and conditions 134, 135 of unknown samples input in the experiment.

After step S41, the model search unit 18 of the physical property prediction unit 12 searches for a model to be used for physical property prediction for the unknown input vector received in step S41 (hereinafter simply referred to as the unknown input vector) (step S42 ). Specifically, the model search unit 18 refers to the data management unit 14 to search for one representative vector close to the unknown input vector, and obtains a base model (corresponding base model) and a correction model (corresponding base model) related to the relevant representative vector. correction model).

Next, the base model prediction unit 19 of the physical property prediction unit 12 uses the corresponding base model retrieved in step S42 to calculate the base model prediction value (base model prediction value) for the unknown input vector (step S43). .

Next, the correction model prediction unit 20 of the physical property prediction unit 12 uses the corresponding correction model retrieved in step S42 to calculate the correction model prediction value (correction model prediction value) for the unknown input vector (step S44). .

Next, the prediction result determination unit 21 of the physical property prediction unit 12 predicts each physical property for the unknown input vector based on the base model prediction value calculated in step S43 and the corrected model prediction value calculated in step S44. A value (predicted physical property value) is calculated (step S45). At this time, the physical property predicted value is represented by the sum of the base model predicted value and the corrected model predicted value multiplied by the generalization error minimization constant (constant α). The constant α is searched for in the learning phase and registered in the data management unit 14 (see step S14 in FIG. 3 and step S24 in FIG. 7).

Next, the prediction result determination unit 21 calculates the degree of correction for the physical property predicted value for each physical property calculated in step S43 (step S46). Specifically, an absolute value is calculated for each corrected model predicted value calculated in step S44, and this is taken as the correction degree of the corresponding physical property predicted value. That is, by performing the process of step S46, the physical property predicted value and its correction degree are calculated for each physical property, that is, multidimensionally.

Next, the prediction result determination unit 21 calculates the risk value in the entire physical property prediction experiment of this time based on the elements of each dimension of the correction degree calculated in step S46 (that is, the correction degree for each physical property) ( step S47). Specifically, for example, the sum of the values obtained by multiplying the elements of each dimension of the degree of correction by a constant is calculated, and this sum is used as the risk value. For the "constant" to be multiplied by the elements of each dimension during the above calculation, for example, the reciprocal of the standard deviation of the degree of correction of each dimension with respect to the training data can be set. Alternatively, the degree of importance may be set by the engineer, and furthermore, it may not be limited to a constant. By performing the process of step S47, the degree of risk of the predicted value for each experiment is calculated as a risk value.

Then, the prediction result determination unit 21 registers the prediction result of the current physical property prediction in the data management unit 14 to accumulate prediction result data. Finally, the physical property prediction unit 12 instructs the display unit 13 to output a prediction result display screen based on the prediction result of the physical property prediction registered in the data management unit 14 (step S48), and starts the physical property prediction process. finish.

FIG. 11 is a diagram illustrating a data configuration example of prediction result data. The prediction result data 140 shown in FIG. 11 includes the experiment number 141 indicating the identifier of the physical property prediction experiment corresponding to the experiment number 131 in FIG. 10, the physical property prediction value and the correction degree ( prediction values 142, 144, corrections 143, 145), and experimental unit risk values 146. Note that the data configuration example in FIG. 11 is just an example, and the present invention is not limited to this.

FIG. 12 is a diagram showing a specific example of the prediction result display screen. In the prediction result display screen 220 shown in FIG. 12, the distribution map of the area 223 shows the physical properties (dimensions) specified in the combo boxes 221 and 222 among the physical property prediction values of each dimension calculated in the physical property prediction processing. Data showing the predicted value is displayed. Combo boxes 221 and 222 can change physical properties specified by an engineer's operation. In addition, the points plotted on the distribution map are displayed in different display modes according to the correction level according to the legend 224, thereby facilitating identification of the risk of the predicted value. Then, for points in the selected state in the distribution map (indicated by dashed lines in the figure), the prediction result table 225 displays the predicted value and correction degree of each physical property, and the risk value table 226 displays the risk value. be done.

Although not shown in FIG. 12, when the cursor is placed on a point in the selected state, the unknown input vector used for physical property prediction may be displayed in a pop-up or the like.

(4) Summary As described above, according to the physical property prediction device 1 according to the present embodiment, a base model using a predetermined number (K) of training data in the vicinity of an unknown input vector, and an unknown input By using a correction model that corrects the deviation of the base model using a predetermined number (L) of training data in the vicinity of the vector, and by performing physical property prediction, it is possible to calculate the physical property prediction value with reduced generalization error. Instead, it is possible to predict the risk of physical property prediction values by "correction degree" and "risk value".

Specifically, in the calculation of the physical property prediction value, the physical property prediction device 1 searches for the optimum constant (generalization error minimization constant) that minimizes the generalization error searched while referring to the actual data (training data) in the learning phase. is used to calculate the sum of the output of the base model (predicted value of the base model) and the output of the correction model (predicted value of the corrected model) multiplied by the generalization error minimization constant as the predicted value of the physical property. The value generalization error can be greatly reduced.

In addition, when learning each prediction model, using training data based on the neighborhood also contributes to reducing the generalization error of the prediction value.

More specifically, in the physical property prediction device 1 according to the present embodiment, when selecting base model training data or correction model training data in model learning processing, one or more variables (dimensions) that greatly contribute to physical properties (output), Define the "neighborhood" using a metric defined on the space where the input is compressed to the presented space by principal component analysis, or a metric defined by a kernel function, etc., and train based on the definition of this "neighborhood" Since prediction models (base models and correction models) are learned by selecting data, generalization errors in prediction values of physical property prediction output using the prediction models in the operation phase can be reduced. Further, as described with reference to FIG. 8, in the physical property prediction device 1 according to the present embodiment, distance metric learning represented by MLKR is utilized, and the result of learning the distance metric itself is used to A metric that defines the As a result, it is possible to construct (learn) a learning model using a more accurate distance metric, so that the effect of further reducing the generalization error of the predicted value of physical property prediction can be expected.

Then, according to the physical property prediction device 1, by quantifying and outputting (displaying) the "correction degree" of the physical property unit or the "risk value" of the physical property prediction unit, the degree of risk of the physical property prediction value calculated in the physical property prediction ( The degree of blurring) can be visualized, so engineers can manage the risk of physical property prediction.

Regarding the display of the physical property prediction results, as in the distribution diagram shown in the area 223 of the prediction result display screen 220 in FIG. This makes it easier to identify risks, and enhances the effectiveness of supporting risk management. In addition, regarding the display of association of the degree of correction described above, by providing the correction degree adjustment screen 210 in FIG. 6, it is possible to adjust the display settings according to the engineer's request, which can be expected to improve the serviceability.

Further, in the physical property prediction apparatus 1 according to the present embodiment, a learning phase for learning a prediction model and an operation phase for performing physical property prediction are separated, and the learning phase is executed when experimental data (training data) is updated. do it. That is, according to the physical property prediction device 1 according to the present embodiment, since it is not necessary to learn the prediction model each time the physical property prediction is performed, it is possible to reduce the overall processing load and improve the processing speed. .

In addition, in the physical property prediction device 1 according to the present embodiment, the number of models to be learned in the learning phase can be reduced from the total number of experimental data (training data) used for model learning, so the load of model learning can be reduced to conventional It can be expected that it will be reduced more than the method. Specifically, when the total number of experimental data (training data) used for model learning is N, and the total number of clusters generated by clustering the training data is M, the number of models to be learned in the present embodiment is If one type of base model and one type of correction model are used, the number of models is only 2M at most, and the relationship of "N>2M" is almost certainly established. Furthermore, the number of models to be learned can be reduced as the number of models to be clustered increases.

(5) Modified example (multistage correction model)
A modification of the physical property prediction device 1 according to this embodiment will be described. This modification differs from the basic form of the present embodiment described above in that the correction model used in the physical property prediction process has a multistage configuration. In the following, the features of the modified example will be explained while comparing with the basic form.

FIG. 13 is a block diagram showing the process of calculating physical property prediction values and risk values in the basic form. As shown in FIG. 13, in the physical property prediction process for the unknown input vector 310, a corresponding base model (base model 320) and a corresponding correction model (correction model 330) related to a representative vector close to the unknown input vector 310 are retrieved. Both models are used to calculate physical property prediction values, correction degrees, and risk values. In the case of the basic model, the correction model 330 is composed of a single stage.

13. To explain a specific method of calculating the physical property predicted value, etc. in the configuration of FIG. It is calculated by taking the sum of the value obtained by multiplying the correction model prediction value output by the correction model 330 by the generalization error minimization constant (constant α) (see step S45 in FIG. 9). Further, the degree of correction is calculated by taking the absolute value of the correction model predicted value (see step S46 in FIG. 9). Also, the risk value is calculated by weighting the elements of each dimension of the corrected model prediction values and taking the sum (see step S47 in FIG. 9).

On the other hand, in a modified example of the present embodiment, it is conceivable that the correction model 330 has a multistage configuration. Boosting known as a meta-algorithm of machine learning, for example, can be utilized for multistage correction model 330 .

FIG. 14 is a block diagram showing the process of calculating the predicted physical property value and the risk value in the modified example. In the case of FIG. 14, the correction model 330 is composed of a plurality of multistage correction models (first correction model 331, second correction model 332, and third correction model 333).

It should be noted that the method of creating the multi-stage correction model that constitutes the correction model 330 is not limited to a specific one. For example, in the learning phase, in the manner of boosting, the correction models of each stage are created sequentially, such as from the first correction model 331 to the second correction model 332, from the second correction model 332 to the third correction model 333, and so on. can be

Furthermore, in the learning phase, the number of pieces of training data used for learning the correction model may be sequentially increased according to the number of stages. For example, in the case of the basic model, the number of training data used for learning the base model 320 is K, and the number of training data used for learning the correction model is L, which is larger than K, but in the case of this modification, the first correction model 331 is similarly set to L, more training data than L is used for learning the second correction model 332, and more training data is used for learning the third correction model 333. Data should be used. More specifically, a correction model that predicts the inverse of the residual of the base model 320 is created as the first correction model 331, and the residual of the output of the base model 320 and the constant multiple prediction value of the first correction model 331 is calculated. A correction model that predicts the reciprocal of the difference is created as the second correction model 332, and the correction model that predicts the reciprocal of the residual between the output of the base model 320 and the constant multiple prediction value of the second correction model 332 is the second correction model. 3 correction model 333 can be created. To generalize, the above creation process can be repeated to create up to the N-th correction model (N: a natural number of 2 or more). However, as is generally known, the larger the data used for learning, the more likely it is to overlearn. Preferably, each model is multiplied by a predetermined learning factor.

A method of calculating the predicted physical property values and the like in the modified example shown in FIG. 14 will be specifically described in detail.

The physical property prediction value is a value obtained by multiplying the base model prediction value output by the base model 320 with the unknown input vector 310 as an input and the correction model prediction value output by the first correction model 331 with the unknown input vector 310 as an input multiplied by a constant α1. , the sum of the value obtained by multiplying the corrected model predicted value output by the second correction model 332 by the constant α2 and the value obtained by multiplying the corrected model predicted value similarly output by the third correction model 333 by the constant α3. calculated by

Here, constants α1, α2, and α3 are constants determined by “generalization error minimization constant (constant α)×learning coefficient” and take values between “0” and “1”. The learning coefficient corresponding to each correction model is set to a predetermined value between "0" and "1" so as to prevent over-learning of the prediction model as a whole.

The degree of correction is calculated by taking the sum of the values obtained by multiplying the absolute values of the correction model prediction values output by each stage correction model (first correction model 331 to third correction model 333) by the respective learning coefficients. .

The risk value is the sum of the values obtained by weighting and summing the elements of each dimension of the correction model prediction value in each stage correction model (first correction model 331 to third correction model 333). calculated by taking

The predicted physical property value, degree of correction, and risk value calculated as described above can be confirmed by an engineer through the prediction result display screen as shown in FIG. 12, as in the case of the basic example.

As a further derived example of this modified example, when a plurality of multi-stage correction models can be used in the physical property prediction process, various changes are made to which correction model in the multi-stage configuration is used to perform physical property prediction. You may make it
Specifically, for example, when using only the first correction model 331 in FIG. When used in combination, the predicted physical property value, degree of correction, and risk value may be calculated respectively, and the calculation results may be displayed. In this way, by displaying the calculation results according to the combination of various correction models, the results of changing the data range used for the correction models are displayed, so engineers can see how close the prediction models are. It can be considered whether it should be used to predict physical properties.

As described above, in this modified example, the correction model for correcting the deviation of the base model is configured to have a multi-stage configuration by calculating the physical property prediction value etc. using the prediction value by the multi-stage correction model. Since the range of data used can be changed, in addition to the effect obtained in the case of the basic model, prediction of physical properties using a stronger learning model can be expected. Further, by multiplying the learning coefficient when summing the correction values of these multistage correction models, over-learning can be prevented. In addition, when the combination of correction models can be changed, the prediction risk (correction degree) can be adjusted and selected. As a result, in this modified example, it can be expected to further suppress the generalization error of prediction.

It should be noted that the present invention is not limited to the above-described embodiments and modified examples, and includes various other modified examples. For example, the above-described embodiments have been described in detail in order to explain the present invention in an easy-to-understand manner, and are not necessarily limited to those having all the described configurations. Moreover, it is possible to add, delete, or replace part of the configuration of the embodiment with another configuration.

Further, each of the above configurations, functions, processing units, processing means, and the like may be realized by hardware, for example, by designing a part or all of them using an integrated circuit. Moreover, each of the above configurations, functions, etc. may be realized by software by a processor interpreting and executing a program for realizing each function. Information such as programs, tables, and files that implement each function can be stored in a recording device such as a memory, a hard disk, or an SSD, or a recording medium such as an IC card, an SD card, or a DVD.

Further, the control lines and information lines indicate those considered necessary for explanation, and not all control lines and information lines are necessarily indicated on the product. It may be considered that almost all configurations are interconnected in practice.

1 physical property prediction device 2 terminal 3 network 11 model learning unit 12 physical property prediction unit 13 display unit 14 data management unit 15 representative vector selection unit 16 base model learning unit 17 correction model learning unit 18 model search unit 19 base model prediction unit 20 correction model Prediction unit 21 Prediction result determination unit 101 CPU
102 ROMs
103 RAM
104 external storage device 105 communication I/F
106 external input device 107 external output device 210 correction degree adjustment screen 220 prediction result display screen 310 unknown input vector 320 base model 330 correction model 331 first correction model 332 second correction model 333 third correction model

Claims (15)

a model learning unit that learns a model used for physical property prediction;
a physical property prediction unit that performs the physical property prediction using the model learned by the model learning unit for an unknown input vector that is input as data of an unknown sample;
a display unit for displaying the result of the physical property prediction;
with
The model learning unit
a representative vector selection unit that clusters an input set of training data used for learning the model and selects a representative vector in each cluster;
a base model learning unit that learns a base model for predicting physical property values using a first predetermined number of the training data in the vicinity of each of the representative vectors;
a correction model learning unit that learns a correction model for predicting the inverse of the residual of each base model using a second predetermined number of the training data in the vicinity of each of the representative vectors;
has
The physical property prediction unit is
a model search unit for searching the base model and the correction model for the representative vector close to the unknown input vector;
a base model prediction unit that calculates a base model prediction value as a prediction value of the base model for the unknown input vector;
a correction model prediction unit that calculates a correction model prediction value as a prediction value of the correction model for the unknown input vector;
A physical property predicted value is calculated for each physical property by the sum of the base model predicted value and the corrected model predicted value multiplied by a predetermined constant, and the risk of the physical property predicted value is calculated based on the corrected model predicted value. A prediction result determination unit that calculates the degree of correction indicated;
has
The physical property prediction device, wherein the display unit provides a prediction result display screen that displays at least the physical property prediction value and the degree of correction calculated by the prediction result determination unit. The display unit displays the predicted physical property value calculated by the predicted result determination unit in a display mode in which the correction degree corresponding to the predicted physical property value is associated with the prediction result display screen. Item 1. The physical property prediction device according to item 1. On the prediction result display screen, the physical property prediction value is displayed in a different display mode according to the correction degree level indicating the degree of the corresponding correction degree,
3. The display unit according to claim 2, wherein the display unit further provides a correction degree adjustment screen on which classification criteria for the correction degree levels or a display mode for each correction degree level in the prediction result display screen can be set. Physical property prediction device. The prediction result determination unit further calculates a risk value indicating the risk of the physical property predicted value in the current physical property prediction as a whole, based on the corrected model predicted value corresponding to the physical property predicted value for each physical property. The physical property prediction device according to claim 1. The model learning unit further has a distance metric learning unit that learns the distance metric of the data space by MLKR (Metric Learning for Kernel Regression),
In the learning of the base model by the base model learning unit and the learning of the correction model by the correction model learning unit, which are performed using the training data in the vicinity of the representative vector, the neighborhood is the distance metric learning unit. 2. The physical property prediction device according to claim 1, wherein the distance metric learned in is determined. The correction model learning unit, when learning the correction model, creates a first correction model that predicts the inverse of the residual of the base model, and outputs the output of the base model and the first correction model. Create a second correction model that predicts the reciprocal of the residual from the value obtained by multiplying the predicted value by a constant, and repeat the creation procedure to create up to the Nth correction model (N is a natural number of 2 or more) death,
The physical property prediction device according to claim 1, wherein the physical property prediction unit uses the created first to Nth correction models as the correction models to perform the physical property prediction. The model learning unit linearly searches for a constant that minimizes a generalization error estimated for the input set and the output set of the training data, so that in the calculation of the physical property predicted value by the prediction result determination unit, the 2. The physical property prediction device according to claim 1, wherein the predetermined constant to be multiplied by the corrected model predicted value is set in advance. In learning the base model by the base model learning unit and learning the correction model by the correction model learning unit, which are performed using the training data near the representative vector,
The neighborhood is defined on a given distance metric, a metric defined for one or more dimensions whose contribution to the output is greater than or equal to a given degree, and a space obtained by compressing the input to the presentation space by principal component analysis. 2. The physical property prediction device according to claim 1, wherein the determination is made using either a metric or a metric defined by a kernel function. a model learning step for learning a model used for physical property prediction;
a physical property prediction step of performing the physical property prediction using the model learned in the model learning step for an unknown input vector input as data of an unknown sample;
a display step of displaying the result of the physical property prediction;
with
The model learning step includes:
a representative vector selection step of clustering an input set of training data used for learning the model and selecting a representative vector in each cluster;
a base model learning step of learning a base model for predicting physical property values using a first predetermined number of said training data in the vicinity of each said representative vector;
a correction model learning step of learning, for each base model, a correction model that predicts the inverse of the residual of the base model using a second predetermined number of the training data in the vicinity of each of the representative vectors;
has
The physical property prediction step includes:
a model retrieval step of retrieving the base model and the correction model for the representative vector close to the unknown input vector;
a base model prediction step of calculating a base model prediction value as the prediction value of the base model for the unknown input vector;
a correction model prediction step of calculating a correction model prediction value as a prediction value of the correction model for the unknown input vector;
By taking the sum of the base model prediction value calculated in the base model prediction step and the value obtained by multiplying the correction model prediction value calculated in the correction model prediction step by a predetermined constant, A prediction result determination step of calculating a predicted value and calculating a correction degree indicating a risk of the predicted physical property value based on the correction model predicted value;
has
The physical property prediction method, wherein the display step provides a prediction result display screen that displays at least the physical property prediction value and the degree of correction calculated in the prediction result determination step. In the display step, the predicted physical property value calculated in the predicted result determination step is displayed on the predicted result display screen in a display mode in which the correction degree corresponding to the predicted physical property value is associated. The physical property prediction method according to claim 9 . In the display step, the predicted physical property value is displayed on the prediction result display screen in different display modes according to the corresponding correction level indicating the degree of the correction,
The model learning step includes:
11. The method according to claim 10, further comprising a second display step of displaying a correction degree adjustment screen on which a classification criterion for the correction degree level or a display mode for each correction degree level in the prediction result display screen can be set. physical property prediction method. In the prediction result determination step, further calculating a risk value indicating the risk of the physical property prediction value in the current physical property prediction as a whole, based on the correction model prediction value corresponding to the physical property prediction value for each physical property. The physical property prediction method according to claim 9. The model learning step further includes a distance metric learning step of learning the distance metric of the data space by MLKR (Metric Learning for Kernel Regression),
In learning the base model in the base model learning step and learning the correction model in the correction model learning step using the training data in the vicinity of the representative vector, the neighborhood is the distance metric learning step 10. The physical property prediction method according to claim 9, wherein the distance metric learned in . When learning the correction model in the correction model learning step, creating a first correction model for predicting the inverse of the residual of the base model, and predicting the output of the base model and the first correction model Create a second correction model that predicts the reciprocal of the residual with the value obtained by multiplying the value by a constant, and repeat the creation procedure to create up to the Nth (N is a natural number of 2 or more) correction model. ,
The physical property prediction method according to claim 9, wherein, in the physical property prediction step, the physical property prediction is performed using the created first to Nth correction models as the correction models. In the model learning step, by performing a linear search for a constant that minimizes an estimated generalization error for the input set and the output set of the training data, in the prediction result determination step, the 10. The physical property prediction method according to claim 9, wherein the predetermined constant to be multiplied by the corrected model predicted value is set in advance.

Images