JP2023170227A - Electronic control device - Google Patents

Abstract

To provide an electronic device that can automatically select an optimal feature quantity.SOLUTION: A second ECU generates a candidate feature quantity set including a plurality of feature quantity candidates that are at least some of a plurality of feature quantities and output data in each feature quantity candidate, repeatedly performs changing the feature quantity candidates in the candidate feature quantity set and prediction using the candidate feature quantity set with the changed feature quantity candidates, and selects feature quantity candidates that are involved in improving prediction accuracy (S10-S12). The second ECU generates an improved model obtained by improving a basic model used to predict an objective variable based on the objective variable and an updated feature quantity set including the selected feature quantity candidates and a plurality of items of output data in each feature quantity candidate (S13).SELECTED DRAWING: Figure 2

Description

The present disclosure relates to an electronic control device.

Conventionally, as described in Patent Document 1, there is a learning method in which, in addition to a teacher label and sensor information, a judgment model is trained using factor information representing factors that influence the judgment of the teacher label.

JP 2021-36379 Publication

By the way, in order to improve the prediction accuracy of a machine learning model, it is conceivable to generate new feature quantities using existing feature quantities. However, even if a new feature is added to an existing feature, the prediction accuracy of the model does not necessarily improve. In other words, the prediction accuracy of the model may decrease by adding features that do not contribute to the prediction of the target variable.

Generating new features using existing features has been done through trial and error based on human experience. However, in the future, there is a possibility that data using machine learning will be used to control electronic control devices. In this case, it is predicted that the types and amount of data to be handled will be enormous. Therefore, with the model generation method, it is difficult to generate new feature quantities that can improve prediction accuracy based on human experience.

One disclosed objective is to provide an electronic device that can automatically select optimal features.

The electronic control device disclosed herein is
It is an electronic control device that predicts an objective variable using a basic model that is machine-learned using a feature set that includes multiple feature quantities and multiple output data for each feature quantity, and an objective variable that corresponds to each output data. hand,
A candidate feature set including a plurality of feature candidates that are at least a part of the plurality of feature quantities and output data for each feature candidate is generated, and the feature quantity candidates in the candidate feature set are changed and the feature quantity candidates are changed. a selection unit (S10 to S12) that repeatedly performs prediction using the changed candidate feature set and selects feature candidates that are involved in improving prediction accuracy;
Generates an improved model that improves the basic model used to predict the objective variable based on the objective variable, the feature candidates selected by the selection unit, and an updated feature set that includes multiple output data for each feature candidate. The present invention is characterized by comprising a model generation unit (S13) that generates a model.

According to the electronic control device disclosed herein, an improved model that is an improved basic model used for predicting a target variable is generated based on an updated feature set including selected feature candidates. Furthermore, when selecting feature candidates, the electronic control unit repeatedly changes the feature candidates in the candidate feature set and performs prediction using the candidate feature set with the changed feature candidates. Select feature candidates that are involved in improving accuracy. Therefore, the electronic control device can automatically select the optimal feature amount that is involved in improving prediction accuracy.

The multiple embodiments disclosed in this specification employ different technical means to achieve their respective objectives. The claims and the reference numerals in parentheses described in this section exemplarily indicate correspondence with parts of the embodiment described later, and are not intended to limit the technical scope. The objects, features, and advantages disclosed in this specification will become more apparent by reference to the subsequent detailed description and accompanying drawings.

FIG. 1 is a block diagram showing a schematic configuration of an in-vehicle system. It is a flowchart which shows the outline of the algorithm of a 2nd ECU. It is a flowchart which shows the model-free optimization method of 2nd ECU. It is a flowchart which shows the cost function calculation of 2nd ECU. FIG. 2 is a diagram showing the correspondence between symbols and descriptions in the specification. 3 is a diagram showing Equations 1 to 4. FIG.

Hereinafter, embodiments for implementing the present disclosure will be described with reference to FIGS. 1 to 4.

In this embodiment, as an example, an electronic control device is applied to an in-vehicle system 100 mounted on a vehicle. The in-vehicle system 100 includes three ECUs 11-13 and three sensors 21-23. However, in the present disclosure, one ECU may have the functions of the ECUs 11 to 13. The second ECU 12 and the third ECU 13 correspond to an electronic control device. Additionally, the present disclosure may include four or more sensors. ECU is an abbreviation for Electronic Control Unit.

Each ECU 11 to 13 includes a processor, a memory device, peripheral I/O, and the like. A processor is an arithmetic processing device such as a CPU. Memory devices include volatile memory and nonvolatile memory. The peripheral I/O includes a communication interface, an AD converter, and the like. A processor executes a program stored in a memory device. A processor performs arithmetic processing by executing a program. At this time, the processor performs arithmetic processing using data stored in the memory device, data obtained from peripheral I/O, and the like. The processor then outputs the calculation result from the peripheral I/O.

A first sensor 21, a second sensor 22, and a third sensor 23 are electrically connected to the first ECU 11. The first ECU 11 receives sensor data (sensor values) from each of the sensors 21 to 23. The first ECU 11 controls sensor data output from each sensor 21 to 23 of the vehicle.

The first sensor 21 is a rotational speed sensor that outputs sensor data that is an electrical signal according to the engine rotational speed. The second sensor 22 is a temperature sensor that outputs sensor data that is an electrical signal depending on the engine temperature. The third sensor 23 is an oxygen concentration sensor that outputs sensor data that is an electrical signal depending on the oxygen concentration in the exhaust gas.

However, the present disclosure is not limited thereto, and sensors that output electrical signals according to other physical quantities can also be employed. Note that the engine speed, engine temperature, and oxygen concentration correspond to feature amounts (labels). The sensor data output by each sensor 21 to 23 corresponds to output data.

In the second ECU 12, a processor constructs a model. This model is obtained through supervised learning. To be more specific, the model is machine learned using a feature set including a plurality of feature quantities and a plurality of output data for each feature quantity, and an objective variable corresponding to each output data. It can also be said that the second ECU 12 constructs a model that the third ECU 13 uses for prediction. Further, the second ECU 12 improves the basic model, which is the model used by the third ECU 13 for prediction, to generate an improved model. The feature quantity set and the objective variable are stored in the memory device of the second ECU 12, for example.

The processor of the second ECU 12 generates an improved model by executing the flowcharts shown in FIGS. 2, 3, and 4. Further, the processor of the second ECU 12 calculates Equation 1, Equation 2, Equation 3, and Equation 4 shown in FIG. The processing operation of the second ECU 12 will be explained in detail later.

The model of the third ECU 13 is stored in the memory device. The third ECU 13 uses a model to predict (output) a target variable from the sensor data input to the processor. This prediction is identification or regression. This embodiment employs a model that predicts engine failure from sensor data on engine speed, engine temperature, and oxygen concentration.

Note that this embodiment employs an example in which functions are distributed among three ECUs 11 to 13. However, the present disclosure can be adopted even with a single ECU having the functions of each of the ECUs 11 to 13.

Here, processing operations by the second ECU 12 (processor) will be explained using FIGS. 2 to 6. First, the symbols used below will be explained. X is an existing feature quantity set and is a matrix of n rows and m columns. n is the number of data, and m is the number of features. In other words, the column elements of X are feature amounts. On the other hand, the row elements of X are data (output data) of each feature amount. The features of the existing feature set X can also be said to be existing features. Y is a known target variable set and is a column vector with n rows and 1 column. 1 is the number of objective variables. n is each data of the objective variable, and is the same number as each output data of the feature amount.

x _train and x _test are obtained by dividing the existing feature set X into a training set and a test set. The training set x _train is a matrix with t rows and m columns. t is the number of training data, and m is the number of features. The test set x _test is a matrix with (nt) rows and m columns.

x _all is a candidate feature set. The candidate feature set x _all is a feature set in which a new feature set obtained by combining column elements of the existing feature set x _train is added to the existing feature set x _train . Here, the candidate feature set x _all is a matrix with t rows and (m+l) columns, and l is the number of new features created using existing features. Here, as an example, a training set x _train is used as an existing feature set.

xotr is expressed by the symbol shown in the first row of FIG. xotr is a feature set in which only column vectors of features that are involved in improving the accuracy of prediction of the target variable are extracted from the matrix of the candidate feature set x _all . xotr is an extracted feature set of the training set. xotr is a matrix with t rows and q columns. q is the number of feature amounts (1<q<(m+l)) that are involved in improving the accuracy of prediction of the objective variable.

xote is expressed by the symbol shown in the second row of FIG. xote uses the test set x _test to create a feature set to which new features are added, such _as the candidate feature set It is a quantity set. xote is an extracted feature set of the test set.

y _train and y _test are obtained by dividing the existing target variable set Y into a training set and a test set. y _train is a t-by-1 vector. y _test is a (nt)-by-1 vector.

The policy parameter θ is an input vector to the model-free optimization method shown in step S11. Based on the policy parameter θ, the second ECU 12 determines column elements to be adopted in the intermediate feature set x _temp used in the cost function (S23). The intermediate feature set x _temp is a feature set used to calculate the cost value r _i for the value of the policy parameter θ during rollout in the model-free optimization method.

The optimization parameter θ _opt is a policy parameter θ optimized to minimize the cost function (cost value r _i ). In other words, θ _opt is a parameter that controls which column element is extracted from the candidate feature set x _all . θ _opt is a row vector with 1 row and (m+l) columns. The second ECU 12 calculates xotr based on θ _opt and x _all .

Note that, as shown in the three-tiered plane of FIG. 5, the superscripts (1) to (m+l) of θ _i in Equation 2 and the like are written as θi(k). As shown in the four-stage plane of FIG. 5, the superscripts (1) to (m+l) of _Θi in Equation 3 are written as Θi(k). k is a continuous natural number from 1 to m+l. i is a continuous natural number value of 1 or more until rolling out. The value to be rolled out can be arbitrarily determined, for example, a value at which the cost value r _i can be considered to have converged. Further, as shown in the fifth row of FIG. 5, the superscript (j) of θ _opt in Equation 4 is written as θo(j). j is 1≦j<m+l.

The flowchart of FIG. 2 can also be said to be a processing operation for automatically selecting the feature quantity by the second ECU 12. Moreover, the flowchart of FIG. 2 can also be said to be a method for generating an improved model by the second ECU 12. Furthermore, the flowchart in FIG. 2 can be said to be an outline of an algorithm for generating an improved model. The second ECU 12 starts the flowchart of FIG. 2 in response to an instruction from a user or the like.

In the following, an example of constructing a predictive model for predicting engine failure will be adopted as the basic model and the improved model. First, 10,000 pieces of data are prepared for the output data of each sensor 21 to 23 and the state of the engine (normal, abnormal) when each output data is output. In this case, the number n of output data is 10,000. The feature quantities are engine speed, engine temperature, and oxygen concentration. The number m of feature quantities is three. The data indicating normality and the data indicating abnormality are objective variable data.

Therefore, the existing feature quantity set X is a 10,000-by-3 matrix having the sensor values of sensors 1 to 3 as column elements. The known objective variable set Y is a 10,000-row, 1-column column vector whose column elements are engine states.

Next, the second ECU 12 divides the feature set X into a training set x _train and a test set x _test at a ratio of 9:1. Similarly, the second ECU 12 divides the target variable set Y into a training set y _train and a test set y _test at a ratio of 9:1. The training set x _train is a matrix with 9000 rows and 3 columns. The training set y _train is a matrix of 9000 rows and 1 column. The test set x _test is a matrix of 1000 rows and 3 columns. The test set y _test is a matrix of 1000 rows and 1 column. Note that the ratio of the training set to the test set employed here (9:1) is only an example. The present disclosure can also be employed with other ratios.

Next, the second ECU 12 generates a candidate feature set x _all from the training set x _train (selection unit). In this embodiment, new feature amounts are generated by multiplying output data from each of the sensors 21 to 23. That is, the new feature amount is a combination of column elements of the first sensor 21, the second sensor 22, and the third sensor 23. The new feature amount here is sensor b×sensor c (1≦b≦3, 1≦c≦3). The output data of the new feature amount is data obtained by multiplying the output data corresponding to the combination of each sensor 21 to 23.

Then, the second ECU 12 generates a candidate feature set x _all by adding it to the column elements of the training set x _train . In this case, the candidate feature set x _all is a matrix of 9000 rows and 12 columns. In this way, the second ECU 12 uses a plurality of feature quantity candidates including a plurality of feature quantities and a new feature quantity obtained by combining the plurality of feature quantities, as column elements of the candidate feature quantity set x _all .

Note that the method for generating the candidate feature set x _all is not limited to this. Further, the candidate feature set x _all can be adopted as long as it includes a plurality of feature candidates that are at least part of the existing feature set X and output data for each feature candidate. Therefore, even an existing feature set X can be adopted as the candidate feature set x _all . Furthermore, the candidate feature set x _all may use the average value, variance, or the like of sensor data in existing feature quantities.

Then, the second ECU 12 sets θ=θ _int in step S10 (selection unit). That is, the second ECU 12 assigns the initial value θ _int to the policy parameter θ. The initial value θ _int is set so that only existing feature quantities are selected during the first feature quantity selection. This is to ensure that the improved model does not have lower prediction accuracy than the basic model based on the existing feature set X. For example, the initial value θ _int = [0.1, 0.1, 0.1, -0.2, -0.2, . . . , -0.2]. The initial value θ _int is a row vector with 1 row and 12 rows. Here, by setting a negative value such as -0.2, only existing feature quantities are selected during the first feature quantity selection. Note that when a negative value is used, a value as close to a positive value as possible is preferable.

The second ECU 12 executes a model-free optimization method in step S11 (selection unit). Step S11 will be explained using FIG. 3. Note that θ _i and Θ _i in the following have the relationship shown in Equations 2 and 3.

First, in step S21 of FIG. 3, the second ECU 12 substitutes the input initial value θ ₁ =θ _int . Then, the second ECU 12 performs the first loop process in step S22. The second ECU 12 performs loop processing from the value of 1 to the value of i to be rolled out. That is, the second ECU 12 repeatedly calculates the optimization parameter θ _opt , which is the policy parameter θ that minimizes the cost function defined in FIG. In other words, the second ECU 12 calculates the cost value r _i by calculating the cost function. The second ECU 12 calculates the optimization parameter θ _opt by repeatedly performing calculations so that the cost value r _i becomes close to an appropriate value. It can also be said that the second ECU 12 selects a feature amount candidate whose cost value r _i is close to an appropriate value by determining the optimization parameter θ _opt . xotr includes feature quantity candidates whose cost value r _i is close to an appropriate value.

Here, the first loop process of step S22 will be explained in detail. The second ECU 12 calculates a cost function in step S23. Step S23 will be explained using FIG. 4.

The second ECU 12 calculates Equation 1 in step S31. The second ECU 12 performs calculation by substituting θ _i into the logistic sigmoid function of Equation 1 in order to calculate the intermediate feature set x _temp . In other words, the second ECU 12 performs step S31 in order to select the intermediate feature set x _temp from the candidate feature set x _all in the second loop process of step S32. As will be explained later in step S24, the policy parameter θ _i substituted in step S31 is a value calculated so that the cost value r _i becomes small. The policy parameter θ _i is a continuous value.

Note that the logistic sigmoid function corresponds to a predetermined function. The logistic sigmoid function has the property that the value of Θ _i falls within the range of 0 to 1. A value between 0 and 1 is adopted as the first threshold value to be compared with the value of Θ _i . Thus, the logistic sigmoid function has a limited range of values. Therefore, it is possible to easily set the first threshold value for selecting the intermediate feature quantity set x _temp . However, the present disclosure can also employ a Gaussian function as the predetermined function.

The second ECU 12 performs a second loop process in step S32. The second ECU 12 determines whether Θi(k)>0.5 (first threshold) in step S33. If the second ECU 12 determines that Θi(k)>0.5, the process proceeds to step S34, and if it does not determine that Θi(k)>0.5, it skips step S34. The first threshold corresponds to a threshold.

In step S34, the second ECU 12 adds column elements of column k of the candidate feature set x _all to the intermediate feature set x _temp . The column element of the k column is a feature amount corresponding to Θi(k) having a value larger than 0.5 among the feature amounts of the candidate feature set x _all .

In this way, the second ECU 12 compares the value Θi(k) obtained from the logistic sigmoid function with the first threshold, and selects the value Θi(k) that is larger than the first threshold. Therefore, when selecting feature quantity candidates whose cost value r _i is close to an appropriate value, the second ECU 12 converts continuous values into discrete values by comparing the value Θi(k) obtained from the logistic sigmoid function with the first threshold value. It can be said to convert into a value.

Then, the second ECU 12 selects a feature amount candidate corresponding to Θi(k). It can also be said that the second ECU 12 determines the type of feature to be selected from the candidate feature set x _all by comparing Θi(k) with the first threshold, and extracts only the selected type of feature. It can also be said that the second ECU 12 changes the feature amount candidates in the candidate feature amount set x _all . In addition, the second ECU 12 calculates a cost value r _i that is a prediction result using an intermediate feature set x _temp , which is a candidate feature set in which the feature candidates are changed, and It can be said that feature quantity candidates whose value r _i is close to an appropriate value are selected.

Model-free optimization methods are originally used for continuous value optimization. However, in the present disclosure, the second ECU 12 performs the second loop process S32, thereby making it possible to apply a model-free optimization method to the optimization of discrete values regarding the presence or absence of feature quantity selection.

The second ECU 12 calculates the cost value r _i using the identification method in step S35. Specifically, the second ECU 12 uses the training set y _train and the intermediate feature set x _temp as target variables and the corresponding feature values, and calculates the cost value r _i using the discrimination method to be used. The cost value r _i corresponds to the value of the loss function. Cross validation is used to calculate the loss function. Note that the cost value r _i can also be rephrased as the accuracy and precision of prediction results such as regression and identification. In this embodiment, an example using an identification method is adopted. However, the present disclosure is not limited thereto, and a regression method may be used.

The second ECU 12 returns the cost value r _i as a return value in step S36. In other words, the second ECU 12 uses the return value of the model-free optimization method. Therefore, the cost value r _i is used in step S24.

Here, we return to the explanation of FIG. 3. In step S24, the second ECU 12 calculates a policy parameter θ _i+1 that reduces the cost value r _i . The second ECU 12 calculates the policy parameter θ _i+1 from the cost value r _i that is the return value and the previous policy parameter θ _i .

The second ECU 12 returns the optimized optimization parameter θ _opt as a return value in step S25. The optimization parameter θ _opt is used in step S12.

Here, we return to the explanation of FIG. 2. The second ECU 12 calculates xotr (selection unit) based on the optimization parameter θ _opt and the candidate feature set x _all in step S12. At this time, the second ECU 12 calculates Equation 4 in order to determine whether to extract the feature amount in the j column of the candidate feature set _{x_all} . The second ECU 12 then compares the value of Equation 4 with the second threshold (0.5).

If the value of Equation 4 is greater than 0.5, the second ECU 12 extracts the column element in the candidate feature set x _all that is indicated by the value of j at that time. On the other hand, if the value of Equation 4 is not greater than 0.5, the second ECU 12 does not extract the column element in the candidate feature set x _all indicated by the value of j at that time.

The extracted column elements are regarded as features involved in improving the prediction accuracy of the target variable. On the other hand, column elements that are not extracted are regarded as features that are not involved in improving the prediction accuracy of the target variable. Therefore, the second ECU 12 sets the matrix of the extracted column elements and their output data as xotr. In this way, the second ECU 12 repeatedly calculates and compares Equation 4 while switching the value of j from 1 to m+l to obtain xotr. The column elements extracted here correspond to the selected feature quantity candidates. Furthermore, xotr corresponds to an updated feature set.

As described above, the second ECU 12 repeatedly changes the feature quantity candidates in the candidate feature quantity set x _all and performs prediction using the intermediate feature quantity set x _temp , which is a candidate feature quantity set in which the feature quantity candidates have been changed. , select feature quantity candidates that are involved in improving prediction accuracy. Then, the second ECU 12 obtains xotr including the selected feature amount candidate.

In step S13, the second ECU 12 performs learning using y _train and xotr to obtain a trained model (model generation unit). In other words, the second ECU 12 performs learning using y _train and xotr as objective variables and feature amounts corresponding thereto, and obtains a trained model. In other words, the second ECU 12 generates an improved model that is an improved basic model used for predicting the target variable, based on the xotr calculated in step S12 and the target variable in y _train .

Note that the second ECU 12 may calculate xote based on the optimization parameter θ _opt and the test data x _test . This is to evaluate the improved model. xote can be determined in the same manner as xotr. That is, the second ECU 12 creates a candidate feature set using the test data x _test instead of the training data x _train , and performs the same process as step S12. Then, the second ECU 12 uses y _test and xote as the evaluation set.

Then, the second ECU 12 evaluates the improved model using y _test and xote. In other words, the second ECU 12 determines whether the improved model to which the new feature amount is added is effective. In other words, the second ECU 12 determines whether the improved model has better prediction accuracy than the basic model. The second ECU 12 may formally adopt the improved model determined to be effective. In this case, the third ECU 13 uses the improved model to predict engine failure.

<Effect>
According to the second ECU 12 disclosed herein, an improved model that is an improved basic model is generated based on the updated feature set xotr that includes the selected feature candidates. Furthermore, when selecting feature quantity candidates, the second ECU 12 repeatedly changes the feature quantity candidates in the intermediate feature quantity set x _temp and performs prediction using the intermediate feature quantity set x _temp in which the feature quantity candidates have been changed. , select feature quantity candidates that are involved in improving prediction accuracy. Then, the second ECU 12 obtains xotr including the selected feature amount candidate. Therefore, the second ECU 12 can automatically select the optimal feature quantity that is involved in improving prediction accuracy.

Note that the second ECU 12 can be said to have a function of automating the selection of feature quantities used for prediction such as identification and regression using a model-free optimization method. The second ECU 12 executes a method of minimizing a model-free homemade cost function. This self-made cost function inputs the type of feature to be selected, and uses the value of the cost function when performing prediction such as classification or regression using the selected feature as the return value of the self-made cost function. can also be said to be automatically selected.

The preferred embodiments of the present disclosure have been described above. However, the present disclosure is not limited to the above-described embodiments, and various modifications can be made without departing from the spirit of the present disclosure.

This specification discloses a plurality of technical ideas listed below and a plurality of combinations thereof.

(Technical thought 1)
It is an electronic control device that predicts an objective variable using a basic model that is machine-learned using a feature set that includes multiple feature quantities and multiple output data for each feature quantity, and an objective variable that corresponds to each output data. hand,
A candidate feature set including a plurality of feature candidates that are at least a part of the plurality of feature quantities and output data for each feature candidate is generated, and the feature quantity candidates in the candidate feature set are changed and the feature quantity candidates are changed. a selection unit (S10 to S12) that repeatedly performs prediction using the changed candidate feature set and selects feature candidates that are involved in improving prediction accuracy;
Generates an improved model that improves the basic model used to predict the objective variable based on the objective variable, the feature candidates selected by the selection unit, and an updated feature set that includes multiple output data for each feature candidate. An electronic control device comprising: a model generation unit (S13) for generating a model;

(Technical thought 2)
The selection unit calculates a cost value that is a result of prediction using a candidate feature set with changed feature candidates, and selects a feature candidate whose cost value is close to an appropriate value when changing the feature candidates. The electronic control device according to Selected Technical Idea 1.

(Technical thought 3)
When selecting feature quantity candidates whose cost values are close to appropriate values, the selection unit converts continuous values into discrete values by comparing the value obtained by substituting the continuous value into a predetermined function with a threshold value. The electronic control device according to technical idea 2, which converts and selects feature amount candidates corresponding to the discrete values.

(Technical thought 4)
When selecting feature quantity candidates whose cost values are close to appropriate values, the selection unit converts continuous values into discrete values by comparing the value obtained by substituting the continuous value into a predetermined function with a threshold value. It converts and selects feature quantity candidates corresponding to discrete values.
A function has the property that its value falls between 0 and 1,
The electronic control device according to technical idea 2, wherein the threshold value is between 0 and 1.

(Technical Thought 5)
The electronic control device according to any one of technical ideas 1 to 4, in which the selection unit uses a plurality of feature quantity candidates including a plurality of feature quantities and a new feature quantity obtained by combining the plurality of feature quantities.

Although the present disclosure has been described in accordance with embodiments, it is understood that the present disclosure is not limited to such embodiments or structures. The present disclosure also includes various modifications and equivalent modifications. In addition, although various combinations and configurations are shown in this disclosure, other combinations and configurations involving only one element, more, or fewer elements are also within the scope and spirit of this disclosure. It is something that can be entered.

DESCRIPTION OF SYMBOLS 11... 1st ECU, 12... 2nd ECU, 13... 3rd ECU, 21... 1st sensor, 22... 2nd sensor, 23... 3rd sensor, 100... in-vehicle system

Claims (5)

An electronic control device that predicts the objective variable using a basic model machine-learned using a feature quantity set including a plurality of feature quantities and a plurality of output data for each feature quantity, and an objective variable corresponding to each output data. There it is,
generating a candidate feature set including a plurality of feature quantity candidates that are at least some of the plurality of feature quantities and the output data for each feature quantity candidate, and changing the feature quantity candidates in the candidate feature quantity set; a selection unit (S10 to S12) that repeatedly performs the prediction using the candidate feature set in which the feature quantity candidates are changed, and selects the feature quantity candidates that are involved in improving the accuracy of the prediction;
The basic model used for the prediction of the objective variable based on the objective variable, the feature quantity candidate selected by the selection unit, and an updated feature set including the plurality of output data for each feature quantity candidate. An electronic control device comprising: a model generation unit (S13) that generates an improved model that improves the . The selection unit calculates a cost value that is a result of the prediction using the candidate feature set in which the feature quantity candidates are changed, and when changing the feature quantity candidates, adjusts the cost value to an appropriate value. The electronic control device according to claim 1, wherein the feature quantity candidates that are close to each other are selected. When selecting the feature quantity candidate whose cost value is close to an appropriate value, the selection unit selects the continuous value by comparing a value obtained by substituting the continuous value into a predetermined function with a threshold value. The electronic control device according to claim 2, wherein the feature quantity candidate corresponding to the discrete value is selected by converting into a discrete value. When selecting the feature quantity candidate whose cost value is close to an appropriate value, the selection unit selects the continuous value by comparing a value obtained by substituting the continuous value into a predetermined function with a threshold value. is converted into a discrete value, and the feature amount candidate corresponding to the discrete value is selected,
The function has a property that the value falls between 0 and 1,
3. The electronic control device according to claim 2, wherein the threshold value is between 0 and 1. The electronic control device according to claim 1, wherein the selection unit uses a plurality of feature quantity candidates including a plurality of the feature quantities and a new feature quantity obtained by combining the plurality of feature quantities.

Images