JP7557571B2 - Learning device, learning method, and learning program - Google Patents

Description

The present invention relates to a learning device, a learning method, and a learning program.

A technique known as imitation learning is known in which a machine learning model learns human behavior and then uses the model to teach actions to humans or robots, etc.

There is also a known technique known as the Just-In-Time (JIT) method, which accumulates a large amount of observed data, extracts data near the desired point from the accumulated data, and sequentially trains a model using the extracted data (see, for example, Non-Patent Document 1).

Here, for example, in a chemical plant, environmental changes occur over time, such as deterioration of equipment with age, deterioration of catalysts, and changes to production load plans.

In response to this, it is possible to apply the JIT method to imitation learning, which learns how operators operate equipment in chemical plants, and adapt the model to changes in the environment.

JP 2019-185194 A

Shigeru Yamamoto, "Just-In-Time Predictive Control: Predictive Control Based on Accumulated Data," Instrumentation and Control, Vol. 52, No. 10, October 2013 (https://www.jstage.jst.go.jp/article/sicejl/52/10/52_878/_pdf/-char/ja)

However, when the JIT method is applied to imitation learning, the problem is that as sequential learning is repeated, the operator skills reflected in the model at the initial stage are lost, reducing the validity of the model output.

For example, if equipment is operated according to the output of a model that has learned the operations of an operator at an early stage, it is expected that the operating quality of the equipment and the plant in which it operates will improve, even if the operation is performed by an inexperienced operator. Therefore, the validity of the model's output can be said to be high.

On the other hand, as time passes, the history of actions taken by imitating operators is accumulated as training data and the model is trained sequentially, so the influence of the data learned in the early stages becomes weaker and the validity of the model output may decrease.

If equipment is operated according to such outputs with reduced validity, problems such as a deterioration in the operating quality of the equipment and plant and increased costs may occur.

To solve the above-mentioned problems and achieve the objective, the learning device is characterized by having an assignment unit that assigns a weight to each of first data, which is a combination of explanatory variables and objective variables, which increases as the variance of the index decreases; an extraction unit that preferentially extracts, from the first data, data with small values that increase as the distance between the explanatory variable and a specified explanatory variable increases and that decrease as the weight increases, as second data; and an update unit that uses the second data to update a model that outputs the objective variable from the explanatory variables.

According to the present invention, it is possible to prevent a decrease in the validity of the model output when sequential learning using the JIT method is repeated in imitation learning.

FIG. 1 is a diagram illustrating a plant operation system. FIG. 2 is a diagram illustrating an example of the configuration of a server according to the first embodiment. FIG. 3 is a diagram illustrating an example of the history DB. FIG. 4 is a diagram illustrating the relationship between the prediction target and training data. FIG. 5 is a flowchart showing the flow of the learning process according to the first embodiment. FIG. 6 is a diagram for explaining the classification method. FIG. 7 is a flowchart showing the flow of the learning process according to the second embodiment. FIG. 8 is a diagram illustrating an example of a computer that executes a learning program.

Below, embodiments of the learning device, learning method, and learning program according to the present application will be described in detail with reference to the drawings. Note that the present invention is not limited to the embodiments described below.

[First embodiment]
[Configuration of the first embodiment]
First, a plant operation system will be described with reference to Fig. 1. The plant operation system 1 is a system for managing and controlling a production process of a product in a plant. The plant includes a chemical plant for producing chemical products.

As shown in FIG. 1, the system has a server 10, a terminal device 20, a terminal device 30, and a plant system 40.

The server 10 performs processing related to models for imitation learning. The server 10 can function as a learning device.

The server 10, the terminal device 20, the terminal device 30, and the plant system 40 are connected to each other so as to be able to communicate data with each other via a network N. For example, the network N is the Internet and an intranet.

The terminal device 20 and the terminal device 30 are information processing devices such as personal computers, tablet terminals, and smartphones. The terminal device 30 may also be a dedicated terminal for operating equipment in the plant.

The plant system 40 may include equipment used in the production process and a distributed control system (DCS). For example, the equipment may be a reactor, a cooler, a gas-liquid separator, etc.

The operator is a user who operates the equipment included in the plant system 40 via the terminal device 30. The staff is a user who manages the models used in the server 10 via the terminal device 20.

The processing of each device in the plant operation system 1 will be explained based on Figure 1.

First, the terminal device 20 manages the model in response to operations by the staff member (step S1). For example, the terminal device 20 can instruct the server 10 to change the model and to execute learning and inference processes. The terminal device 20 can also output information acquired from the server 10 and present it to the staff member.

The terminal device 30 operates the equipment of the plant system 40 in response to the operation of the operator (step S2). For example, the terminal device 30 sets the temperature within the equipment, the pressure within the equipment, the target production volume in the production process, the amount of raw materials to be input into the equipment, etc., through the operation.

The plant system 40 operates according to the operation from the terminal device 30 (step S3). Then, the plant system 40 provides the operation history to the server 10 (step S4).

For example, the history includes sensor values from sensors installed at various locations in the plant system 40 and setting values set by operations from the terminal device 30. The history may also be time-series data in which each record is assigned a time (time stamp).

The server 10 performs model learning, inference using the model, and weighting for data extraction (step S5). Each process of the server 10 will be described in detail later.

Furthermore, the server 10 provides the inference result to the operator (step S6). For example, the inference result is the operation content predicted from the situation. The operator operates the plant system 40 according to the operation content provided.

The model learns the operations of the operator through imitation learning. Therefore, other operators can imitate the operations by following the operations obtained as a result of inference by the model.

The server 10 will be described in detail with reference to FIG. 2. FIG. 2 is a diagram showing an example of the configuration of a server according to the first embodiment.

As shown in FIG. 2, the server 10 has a communication unit 11, a memory unit 12, and a control unit 13.

The communication unit 11 communicates data with other devices via a network. For example, the communication unit 11 is a network interface card (NIC).

The storage unit 12 is a storage device such as a hard disk drive (HDD), a solid state drive (SSD), or an optical disk. The storage unit 12 may also be a semiconductor memory in which data can be rewritten, such as a random access memory (RAM), a flash memory, or a non-volatile static random access memory (NVSRAM).

The storage unit 12 stores the OS (Operating System) and various programs executed by the server 10. The storage unit 12 stores model information 121 and a history DB 122.

Model information 121 is information such as parameters for constructing a model. For example, if the model is a neural network, model information 121 is the weights and biases of each layer. Furthermore, model information 121 includes parameters such as the order of preprocessing and the window width (window size) in moving average processing.

History DB122 is information including history provided by plant system 40. FIG. 3 is a diagram showing an example of a history DB. As shown in FIG. 3, history DB122 includes a list of explanatory variables such as time, first temperature, second temperature, first pressure, second pressure, flow rate, set value, CO2 concentration, etc., a set value which is a target variable, and a weight.

The first temperature, second temperature, first pressure, second pressure, and flow rate are sensor values of sensors installed at various locations in the plant system 40.

The first temperature, the second temperature, the first pressure, the second pressure, and the flow rate are explanatory variables of the model and are examples of explanatory variables that represent the situation in the product production process.

The set value is a value that is set by an operation from the terminal device 30. The set value may be a normalized value of an actually set value. The set value corresponds to the objective variable of the model.

The set value is the objective variable of the model and is an example of an objective variable that represents the operation of equipment in the production process.

The CO2 concentration is the concentration of CO2 generated during the production process, and is used as an index in the weighting process described below.

A weight is a value that is assigned to each record in the history DB. The process of assigning weights and the process of extracting data using weights will be described later.

The time is a timestamp indicating the date and time when the first temperature, second temperature, first pressure, second pressure, flow rate, and CO2 concentration were acquired.

For example, Figure 3 shows that at the time "2021/11/5 13:30:01", the first temperature is "40°C", the second temperature is "241°C", the first pressure is "501 hPa", the second pressure is "119 hPa", the flow rate is "12 m3/s", the set value is "0.2", and the CO2 concentration is "700 ppm".

Furthermore, Figure 3 shows that the record with the time "2021/11/5 13:30:01" was assigned a weight of "1.11".

The control unit 13 controls the entire server 10. The control unit 13 is, for example, an electronic circuit such as a CPU (Central Processing Unit), MPU (Micro Processing Unit), or GPU (Graphics Processing Unit), or an integrated circuit such as an ASIC (Application Specific Integrated Circuit) or FPGA (Field Programmable Gate Array).

The control unit 13 also has an internal memory for storing programs and control data that define various processing procedures, and executes each process using the internal memory. The control unit 13 also functions as various processing units by the operation of various programs. For example, the control unit 13 has an extraction unit 131, a calculation unit 132, an update unit 133, an assignment unit 134, and a display control unit 135.

The extraction unit 131 extracts data from the data included in the history DB 122 based on the distance between the explanatory variable and the specified explanatory variable and the weight. The data included in the history DB 122 is an example of first data. The data extracted by the extraction unit 131 is an example of second data.

The specified explanatory variables are called request points. For example, a request point is an explanatory variable (corresponding to each sensor value in history DB122) at a specific time. In addition, the objective variable (set value) at the request point may be unknown.

Here, in the JIT method, data is extracted based on the Euclidean distance between the training data (corresponding to the history DB 122 in this embodiment), which is a multidimensional vector, and the request point, which is also a multidimensional vector. Note that the distance between the training data and the request point is not limited to the Euclidean distance, and may be, for example, the Mahalanobis distance, the cosine similarity, etc.

On the other hand, the server 10 (learning device) of this embodiment extracts data using weights in addition to the distance between the training data and the request point.

First, the extraction unit 131 calculates the Euclidean distance between the request point and each record in the history DB 122. Note that since the request point and each record in the history DB 122 are represented as vectors, the extraction unit 131 may calculate the Euclidean norm by the method described in Non-Patent Document 2.

In conventional JIT methods, k-Nearest Neighbors (k-NN) are extracted, which are the k records (k is an integer) with the smallest calculated Euclidean distance.

On the other hand, the extraction unit 131 extracts records not only based on the calculated Euclidean distance but also by referring to the weights in the history DB 122.

Here, the larger the weight, the more desirable the data is to be extracted. At this time, the extraction unit 131 preferentially extracts data from the history DB 122 that has a small value that increases as the distance increases and decreases as the weight increases.

For example, the extraction unit 131 extracts k (e.g., 1,000) records from the history DB 122 in ascending order of the value obtained by multiplying the Euclidean distance by the inverse of the weight.

For example, the extraction unit 131 may extract only k records from the history DB 122 whose weights are equal to or greater than a threshold value, in order of smallest Euclidean distance.

For example, the extraction unit 131 may extract only k records from the history DB 122 whose Euclidean distance is less than or equal to a threshold, in descending order of weight.

The calculation unit 132 calculates the objective variables by inputting the explanatory variables into the model constructed from the model information 121. In other words, the calculation unit 132 performs inference processing.

The update unit 133 uses the data extracted by the extraction unit 131 to update the model that outputs the objective variable from the explanatory variables.

For example, the update unit 133 calculates an objective function that represents the difference between the objective variable calculated by the calculation unit 132 and the objective variable included in the data extracted by the extraction unit 131, and repeatedly updates the model parameters, i.e., the model information 121, so that the objective function becomes smaller until the learning termination condition is satisfied.

Figure 4 is a diagram explaining the relationship between the prediction target and training data. Training data, which is data up to time t-1, is a record in which setting values have been registered in history DB 122. On the other hand, data of the prediction target at time t corresponds to the request point. For example, the period from time 0 to time t-1 is the target period for searching training data. Therefore, as time passes (as t increases), the target period for searching training data increases.

The assignment unit 134 assigns a weight to each piece of data contained in the history DB 122, which is a combination of an explanatory variable and a target variable.

For example, the weighting unit 134 assigns a weight to each piece of data included in the history DB 122 that increases with the degree of contribution of the data to bringing the index closer to the target value. Note that in this embodiment, the larger the weight, the more desirable the data is to be extracted.

It is believed that by operating the equipment according to the output of a model that has been trained on data extracted based on such weights, it is easier to improve indicators.

For example, the assignment unit 134 assigns a weight to each piece of data included in the history DB 122, which is a combination of explanatory variables representing the situation in the product production process and objective variables representing the operation of equipment in the production process, such that the weight decreases as the concentration of a specified substance discharged in the production process increases.

The assigning unit 134 also assigns a weight obtained from the proximity of an index based on an event observed at a first time to a target value to data included in the history DB 122 that is associated with a second time that is a predetermined time in the past from the first time.

Furthermore, if the index is a positive value and a smaller value is desirable (for example, the target value is 0), the assignment unit 134 assigns the inverse of the index as a weight.

The specified substance emitted in the production process is, for example, CO2. Here, from the perspective of greenhouse gas reduction, it is desirable for the concentration of CO2 emitted to be small.

In addition, in the plant system 40 of this embodiment, it is known that it takes approximately 20 minutes for the effect of the operation of setting the set value (set value) to be reflected in the CO2 concentration.

For example, the weighting unit 134 assigns a weight that is a normalized value of the inverse of the CO2 concentration in the record for the first time to a record for a time 20 minutes before the first time in the history DB 122. Since the CO2 concentration is always equal to or greater than 0, the closer the CO2 concentration is to the target value of 0, the larger the inverse of the CO2 concentration becomes.

In the example of FIG. 3, the assignment unit 134 assigns (adds) the normalized value of the inverse of the CO2 concentration at the time "2021/11/5 14:00:02", 1/900, to the weight of the time "2021/11/5 13:40:02".

Furthermore, in this case, it is considered that the effect on the set value of the CO2 concentration begins to occur 30 minutes before, and the assignment unit 134 assigns (adds) the weight obtained from the CO2 concentration at the time "2021/11/5 14:00:02" to the weight from the time "2021/11/5 13:30:02" to "2021/11/5 13:40:02".

Note that the indicators are not limited to CO2 concentration, but may also be production volume, yield, operating time, energy consumption, etc.

Also, if it is desired that the state is stable, the weight may be a value that increases as the variance of the index decreases.

[Processing of the First Embodiment]
The flow of the learning process according to the first embodiment will be described with reference to Fig. 5. Fig. 5 is a flowchart showing the flow of the learning process according to the first embodiment.

As shown in FIG. 5, first, the server 10 assigns weights to the history records based on the evaluation variables (step S101). The evaluation variables are indices such as CO2 concentration. The evaluation variables may be included in the explanatory variables or the objective variables.

Next, the server 10 extracts records from the history based on the distance from the specified record and the weight (step S102). For example, the server 10 extracts a specified number of records in ascending order of the value obtained by multiplying the Euclidean distance from the record that is the requested point by the inverse of the weight.

The server 10 then updates the model based on the extracted records (step S103). For example, the server 10 updates the model so that the error in the setting value obtained by inputting each sensor value of the extracted records into the model is minimized.

The server 10 can update the model using linear methods such as Ridge and Lasso, or non-linear methods such as deep learning.

[Effects of the First Embodiment]
As described above, the assigning unit 134 assigns a weight to each of the first data, which is a combination of an explanatory variable and a target variable. The extracting unit 131 extracts the second data from the first data based on the distance between the explanatory variable and a specified explanatory variable and the weight. The updating unit 133 uses the second data to update the model that outputs the target variable from the explanatory variable.

In this way, the server 10 can extract data for learning by considering not only the distance from the request point but also the weighting that has been assigned. As a result, according to this embodiment, it is possible to prevent a decrease in the validity of the model output when sequential learning using the JIT method is repeated in imitation learning.

The assigning unit 134 assigns a weight to each piece of first data that increases with the degree of contribution to bringing the index closer to the target value. The extracting unit 131 preferentially extracts data from the first data that has a small value that increases with the distance and decreases with the weight.

This makes it easy to extract data that takes into account both the distance from the desired point and the weight.

The weighting unit 134 assigns a weight obtained from the proximity of an index based on an event observed at a first time to a target value to data of the first data that is associated with a second time that is a predetermined time in the past from the first time.

This makes it possible to accurately extract data that will improve the index, even if there is a time lag before the data is reflected in the index.

When the index is a positive value and a smaller value is more desirable, the assignment unit 134 assigns the inverse of the index as a weight. This reverses the magnitude relationship of the index, making it easy to use it as a weight.

The assignment unit 134 assigns a weight to each of the first data, which is a combination of an explanatory variable representing the situation in the product production process and a target variable representing the operation of equipment in the production process, such that the weight decreases as the concentration of a specified substance discharged in the production process increases.

This allows the model to learn operations that can reduce the emission of substances that are desirable in small concentrations.

The display control unit 135 causes the terminal device 20 to display a screen for instructing the execution of a model learning process, a screen for viewing the contents of the history DB, etc.

Second Embodiment
In the second embodiment, an example will be described in which the server 10 selectively uses a plurality of models depending on the characteristics of the data. The second embodiment is implemented by a server 10 having the same configuration as the first embodiment.

In the second embodiment, the extraction unit 131 extracts data based on the distance between the request point and the explanatory variable from among data belonging to one of multiple clusters into which the history DB 122, which is a combination of explanatory variables and objective variables, is classified.

The extraction unit 131 extracts past records in each cluster using the JIT method. When extracting data, the extraction unit 131 may or may not take into account the weights assigned in the first embodiment.

In addition, the update unit 133 uses the data extracted by the extraction unit 131 to update, among the models corresponding to each of the multiple clusters, the model corresponding to the cluster from which the data was extracted by the extraction unit 131.

The extraction unit 131 classifies each record in the history DB 122 into multiple clusters using the method shown in FIG. 6. FIG. 6 is a diagram explaining the classification method.

The extraction unit 131 may perform classification using a statistical clustering method or may perform rule-based classification.

As shown in FIG. 6, the extraction unit 131 calculates the cross-correlation between two variables at each lag number for a group of records in the history DB 122 whose time falls within a specified period.

For example, the extraction unit 131 calculates the cross-correlation between the first temperature and the second temperature. Furthermore, when the number of lags is 0, the cross-correlation is the cross-correlation between the first temperature and the second temperature at the same time, i.e., in the same record.

For example, when the number of lags is 10, the cross-correlation is the cross-correlation between the first temperature at a certain time and the second temperature 10 seconds after that time.

When the lag number is -10, the cross-correlation is the cross-correlation between the first temperature at a certain time and the second temperature 10 seconds prior to that time.

Then, the extraction unit 131 classifies the record group into either cluster A, cluster B, or cluster C depending on the number of lags at which the cross-correlation peaks.

For example, if the lag number at which the cross-correlation peaks is greater than or equal to -10 and less than 0, the extraction unit 131 classifies the record group into cluster A.

For example, if the lag number at which the cross-correlation peaks is greater than or equal to -20 and less than -10, the extraction unit 131 classifies the record group into cluster B.

For example, if the lag number at which the cross-correlation peaks is greater than or equal to -30 and less than -20, the extraction unit 131 classifies the record group into cluster C.

Here, the models corresponding to each cluster have different learning methods. For example, the explanatory variables and hyperparameters used differ for each model. Hyperparameters include, for example, the rounding width, time width (window size), prediction destination, number of steps, number of layers of DNN, number of nodes, activation function, etc.

The window size is the size of the sliding window when learning time series data.

The window size for training the model corresponding to cluster A is 10 minutes. The window size for training the model corresponding to cluster B is 20 minutes. The window size for training the model corresponding to cluster C is 30 minutes.

[Processing of the second embodiment]
The flow of the learning process according to the second embodiment will be described with reference to Fig. 7. Fig. 7 is a flowchart showing the flow of the learning process according to the second embodiment.

As shown in FIG. 7, first, the server 10 calculates the cross-correlation between two pre-specified variables at each lag number (step S201).

Next, the server 10 clusters the records based on the cross-correlation (step S202). For example, the server 10 performs clustering based on the number of lags at which the cross-correlation peaks.

The server 10 performs model learning using a method defined for each cluster (step S203). The model learning is performed by the calculation unit 132 and the update unit 133.

[Effects of the second embodiment]
As described above, the extraction unit 131 extracts data from among data belonging to any of a plurality of clusters into which the history DB 122, which is a combination of explanatory variables and objective variables, is classified, based on the distance between the request point and the explanatory variable. Furthermore, the update unit 133 uses the data extracted by the extraction unit 131 to update, among the models corresponding to each of the plurality of clusters, a model corresponding to the cluster from which the extraction unit 131 extracted data.

In this way, by preparing multiple models with different learning methods in advance and combining them with the JIT method, it is possible to respond to major changes in the environment (operating conditions, etc.) that require changes to the model.

[System configuration, etc.]
In addition, each component of each device shown in the figure is functionally conceptual, and does not necessarily have to be physically configured as shown in the figure. In other words, the specific form of distribution and integration of each device is not limited to that shown in the figure, and all or a part of it can be functionally or physically distributed or integrated in any unit depending on various loads, usage conditions, etc. Furthermore, each processing function performed by each device can be realized in whole or in part by a CPU (Central Processing Unit) and a program analyzed and executed by the CPU, or can be realized as hardware by wired logic. Note that the program may be executed not only by the CPU but also by other processors such as a GPU.

Furthermore, among the processes described in this embodiment, all or part of the processes described as being performed automatically can be performed manually, or all or part of the processes described as being performed manually can be performed automatically using known methods. In addition, the information including the processing procedures, control procedures, specific names, various data, and parameters shown in the above documents and drawings can be changed as desired unless otherwise specified.

[program]
In one embodiment, the server 10 can be implemented by installing a learning program that executes the above learning process as package software or online software on a desired computer. For example, the above learning program can be executed by an information processing device, causing the information processing device to function as the server 10. The information processing device here includes desktop or notebook personal computers. In addition, the information processing device also includes tablet terminals, smartphones, mobile communication terminals such as mobile phones and PHS (Personal Handyphone System), and slate terminals such as PDA (Personal Digital Assistant).

The server 10 can also be implemented as a server that provides a service related to the above-mentioned learning process to a client, the client being a terminal device used by the user. For example, the server is implemented as a server device that provides a learning service in which the specification of a request point is used as input and a trained model is used as output. In this case, the server may be implemented as a web server, or may be implemented as a cloud that provides a service related to the above-mentioned learning process by outsourcing.

Figure 8 is a diagram showing an example of a computer that executes a learning program. The computer 1000 has, for example, a memory 1010 and a CPU 1020. The computer 1000 also has a hard disk drive interface 1030, a disk drive interface 1040, a serial port interface 1050, a video adapter 1060, and a network interface 1070. Each of these components is connected by a bus 1080.

The memory 1010 includes a read only memory (ROM) 1011 and a random access memory (RAM) 1012. The ROM 1011 stores a boot program such as a basic input output system (BIOS). The hard disk drive interface 1030 is connected to a hard disk drive 1090. The disk drive interface 1040 is connected to a disk drive 1100. A removable storage medium such as a magnetic disk or optical disk is inserted into the disk drive 1100. The serial port interface 1050 is connected to a mouse 1110 and a keyboard 1120, for example. The video adapter 1060 is connected to a display 1130, for example.

The hard disk drive 1090 stores, for example, an OS 1091, an application program 1092, a program module 1093, and program data 1094. That is, the programs that define each process of the server 10 are implemented as program modules 1093 in which computer-executable code is written. The program modules 1093 are stored, for example, in the hard disk drive 1090. For example, a program module 1093 for executing processes similar to the functional configuration of the server 10 is stored in the hard disk drive 1090. The hard disk drive 1090 may be replaced by an SSD (Solid State Drive).

The setting data used in the processing of the above-described embodiment is stored as program data 1094, for example, in memory 1010 or hard disk drive 1090. Then, the CPU 1020 reads the program module 1093 or program data 1094 stored in memory 1010 or hard disk drive 1090 into RAM 1012 as necessary, and executes the processing of the above-described embodiment.

The program module 1093 and the program data 1094 are not limited to being stored in the hard disk drive 1090, but may be stored in, for example, a removable storage medium and read by the CPU 1020 via the disk drive 1100 or the like. Alternatively, the program module 1093 and the program data 1094 may be stored in another computer connected via a network (such as a local area network (LAN) or wide area network (WAN)). The program module 1093 and the program data 1094 may then be read by the CPU 1020 from the other computer via the network interface 1070.

REFERENCE SIGNS LIST 1 Plant operation system 10 Server 20, 30 Terminal device 40 Plant system 11 Communication unit 12 Storage unit 13 Control unit 121 Model information 122 History DB
131 Extraction unit 132 Calculation unit 133 Update unit 134 Assignment unit 135 Display control unit

Claims (3)

an assignment unit that assigns a weight to each of first data, which is a combination of an explanatory variable, a target variable, and an evaluation variable , the weight being larger as the variance of the evaluation variable becomes smaller;
an extraction unit that preferentially extracts, from the first data, data having a small value that increases as the distance between the explanatory variable and a designated explanatory variable increases and decreases as the weight increases, as second data;
an update unit that updates a model that outputs the objective variable from the explanatory variables by using the second data;
A learning device comprising: A learning method performed by a learning device, comprising:
a weighting step of weighting each of first data, which is a combination of an explanatory variable, a response variable, and an evaluation variable , the weighting being increased as the variance of the evaluation variable decreases;
an extraction step of preferentially extracting, from the first data, data having a small value that increases as the distance between the explanatory variable and a designated explanatory variable increases and decreases as the weight increases, as the second data;
an updating step of updating a model that outputs the objective variable from the explanatory variables by using the second data;
A learning method comprising: a weighting step of weighting each of first data, which is a combination of an explanatory variable, a response variable, and an evaluation variable , the weighting increasing as the variance of the evaluation variable decreases;
an extraction step of preferentially extracting, from the first data, data having a small value that increases as the distance between the explanatory variable and a designated explanatory variable increases and decreases as the weight increases, as the second data;
an updating step of updating a model that outputs the objective variable from the explanatory variables by using the second data;
A learning program characterized by causing a computer to execute the above.

Images