Estimation method and information processing device

Abstract

An estimation method according to the present invention includes: a step (S100) in which a machine learning model is trained by providing multiple datasets, each comprising multiple combinations of variables, as training data to a machine learning model comprising an autoencoder type neural network; a step (S101) in which, from among multiple input variables including variables that are known values and variables to be estimated, input variables in which the variables to be estimated are simulated values are input into the trained machine learning model; and a step (S102) in which, from among the machine learning model outputs with respect to the multiple input variables, the output values corresponding to the variables to be estimated are acquired as estimated values for said variables.

Description

Estimation method and information processing device

[0001] The present invention relates to a method for solving a regression problem using a machine learning model and an information processing device.

[0002] A method for estimating variables is useful in which a machine learning model learns combinations of multiple variables that are related to each other, and the learned model is used to determine the value of an unknown variable in the combination (Patent Document 1).

[0003] FIG. 11 shows an example of a machine learning model for a regression problem. Here, the example shows a case where {X} is the target variable (objective variable) to be estimated among the interrelated variables {A to F, X}. The machine learning model is configured with a neural network consisting of three layers: an input layer 200, an intermediate layer 201, and an output layer 202. During the learning process, a set of {A to F} and {X} is provided as training data, and the parameters of the neural network are adjusted so that the difference between the estimated value obtained by calculation for {X} and the correct value is minimized. The trained machine learning model outputs an estimated value of {X} for a new input {A to F} where {X} is unknown.

[0004] However, in conventional techniques, the estimation target in the variable set is necessarily fixed. Therefore, if the estimation target needs to be changed, a new machine learning model must be prepared and trained again.

[0005] Japanese Patent Application Laid-Open No. 2000-305606

[0006] The present invention has been made to solve the above-mentioned problems, and has an object to provide an estimation method and an information processing device that allow arbitrary selection of variables to be estimated.

[0007] The estimation method of the present invention is characterized by including a first step of providing a plurality of sets of data sets consisting of combinations of a plurality of variables as training data to a machine learning model constituted by an autoencoder neural network and training the machine learning model; a second step of inputting, into the trained machine learning model, an input variable from a plurality of input variables consisting of variables of known values ​​and a variable to be estimated, the input variable having a simulated value of the variable to be estimated; and a third step of obtaining, from outputs of the machine learning model for the plurality of input variables, an output value corresponding to the variable to be estimated as an estimate of this variable.

[0008] The information processing device of the present invention is characterized by comprising: a machine learning model formed by an autoencoder neural network; a learning unit configured to provide the machine learning model with multiple sets of data sets consisting of combinations of multiple variables as training data and to learn the machine learning model; an input unit configured to input, to the trained machine learning model, an input variable, from multiple input variables consisting of variables of known values ​​and variables to be estimated, which has a simulated value for the variable to be estimated; and an acquisition unit configured to acquire, from outputs of the machine learning model for the multiple input variables, an output value corresponding to the variable to be estimated as an estimate of this variable.

[0009] According to the present invention, an estimate of a variable to be estimated can be obtained by training a machine learning model configured with an autoencoder neural network, inputting an input variable, which is a simulated value of the variable to be estimated among a plurality of input variables consisting of variables of known values ​​and a variable to be estimated, into the trained machine learning model, and obtaining an output value corresponding to the variable to be estimated among the outputs of the machine learning model for the plurality of input variables.With this invention, an arbitrary variable can be selected as the variable to be estimated in a single machine learning model, and an estimate of the variable to be estimated can be obtained, without requiring a learning model or learning for each variable.

[0010] FIG. 1 is a block diagram showing the configuration of an information processing device according to an embodiment of the present invention. FIG. 2 is a flowchart illustrating the operation of the information processing device according to an embodiment of the present invention. FIG. 3 is a diagram showing an example of an autoencoder neural network. FIG. 4 is a diagram showing a two-terminal pair circuit to which an estimation method according to an embodiment of the present invention is applied. FIG. 5 is a diagram showing an example of a data set of variables. FIG. 6 is a diagram showing an inductance estimation result using a conventional machine learning model. FIG. 7 is a diagram showing an inductance estimation result using an information processing device according to an embodiment of the present invention. FIG. 8 is a diagram showing a temperature estimation result using an information processing device according to an embodiment of the present invention. FIG. 9 is a diagram showing an S-parameter estimation result using an information processing device according to an embodiment of the present invention. FIG. 10 is a block diagram showing an example of the configuration of a computer that realizes the information processing device according to an embodiment of the present invention. FIG. 11 is a diagram showing an example of a conventional machine learning model.

[0011]

[0023] An embodiment of the present invention will now be described with reference to the accompanying drawings. Fig. 1 is a block diagram showing the configuration of an information processing device according to an embodiment of the present invention. The information processing device comprises a machine learning model 10, a learning unit 11, an input unit 12, and an acquisition unit 13.

[0012] FIG. 2 is a flowchart explaining the operation of an information processing device. An autoencoder neural network is a form of machine learning model that uses the same training data for input and output. FIG. 3 shows an example of an autoencoder neural network. An autoencoder neural network consists of three layers: an input layer 100, an intermediate layer 101, and an output layer 102. The number of nodes in the input layer 100 and the output layer 102 is equal. Note that in the example of FIG. 3, there are seven variables, A to X, but this is not limited to seven, and any autoencoder neural network that inputs and outputs multiple variables may be used.

[0013] During learning, the learning unit 11 provides multiple sets of data of variables {A to X} as training data to the machine learning model 10 configured by an autoencoder neural network. The learning unit 11 then trains the machine learning model 10 by adjusting the parameters of the machine learning model 10 so that an output {A to X} close to the training data is obtained (so that the difference between the input and output is small) (Step S100 in FIG. 2). In this way, the machine learning model 10 can memorize the characteristics of the data, i.e., the relationships when the variables {A to X} are related to each other.

[0014] The input unit 12 provides a new set of input variables {A to X} to the trained machine learning model 10 (step S101 in FIG. 2). At this time, if the input variables {A to X} include simulated values ​​that differ from the original values, it is expected that the output values ​​{A to X} of the machine learning model 10 corresponding to the simulated values ​​will be corrected to reflect the characteristics of the training data. In other words, they will be corrected to approach the original values. The present invention utilizes the properties of such an autoencoder neural network to estimate unknown variable values.

[0015] Therefore, the input unit 12 inputs a data set in which the values ​​of the variables to be estimated by the machine learning model 10, among the multiple input variables {A to X}, are set to simulated values ​​different from the original values, to the trained machine learning model 10. The simulated values ​​may be input by a user of the information processing device or may be automatically generated by the input unit 12.

[0016] Examples of variables {A to X} include physiological indices, circuit parameters, etc. For example, various physiological indices obtained by performing various measurements on a subject are related to each other in the sense that they reflect the state of the same object being observed. In addition, a single physiological indices may be calculated from multiple measured values ​​using a formula or the like.

[0017] For example, a technology has been proposed for estimating a subject's core body temperature based on heart rate, body surface area, skin temperature, ambient temperature, etc. (JP 2020-65823 A). This core body temperature estimation can be applied to the present invention as a regression problem. To apply this to the present invention, first, multiple sets of training data consisting of known values ​​such as heart rate, body surface area, skin temperature, air temperature, and core body temperature are provided to the machine learning model 10, and the machine learning model 10 is trained. By inputting input variables consisting of values ​​such as heart rate, body surface area, skin temperature, and air temperature, as well as a simulated value of core body temperature, into the trained machine learning model 10, an estimated core body temperature can be obtained.

[0018] Additionally, a technology has been proposed for calculating the glucose concentration in a subject's interstitial fluid from the subcutaneous dielectric constant at multiple frequencies measured using a microwave sensor (see, for example, "Masato Nakamura et al., 'Glucose Sensing in Interstitial Fluid Using Dielectric Spectroscopy - In Vivo Verification Using a Coaxial Probe Method', Institute of Electronics, Information and Communication Engineers, Vol. 105, No. 6, pp. 497-501, 2022"). This glucose concentration estimation can be applied to the present invention as a regression problem. To apply this to the present invention, first, multiple sets of training data consisting of the subcutaneous dielectric constant and glucose concentration at multiple frequencies are provided to the machine learning model 10, and the machine learning model 10 is trained. By inputting input variables consisting of the subcutaneous dielectric constant at multiple frequencies and simulated values ​​of the glucose concentration into the trained machine learning model 10, an estimated glucose concentration can be obtained.

[0019] As described above, any variable having a continuous value can be an estimation target of the present invention. Needless to say, the types of variables are not limited to the above examples.

[0020] It is desirable that the values ​​used as simulated values ​​of the variables to be estimated in the input data to the trained machine learning model 10 be within an appropriate range (not deviating significantly from the true value) in consideration of the target problem. For example, the following can be given as simulated values:

[0021] (A) The average value in the training data of the same variable as the estimation target. (B) A value obtained using an estimation formula or the like from other known variables different from the estimation target. (C) The value of another variable different from the estimation target. (D) The value of the same variable as the estimation target in another subject with similar characteristics to the subject.

[0022] Among the above (A) to (D), in the cases of (A), (B), and (D), the input unit 12 can automatically generate simulated values ​​from training data, known variables, or values ​​of other subjects. In the case of (C), the user must input the values.

[0023] The acquisition unit 13 acquires, as an estimate of a variable to be estimated, an output value corresponding to the variable to be estimated from the outputs {A to X} of the machine learning model 10 in response to the input variables {A to X} (step S102 in FIG. 2). For example, if the variable to be estimated is X and the input variables {A to X} including a simulated value of X are given to the machine learning model 10, the value of X output by the machine learning model 10 is acquired as the estimate.

[0024] A specific example of the estimation method of this embodiment will be described below. As shown in FIG. 4, the S parameter S １１ is given by equation (1). 1, 1' are an input terminal pair, 2, 2' are an output terminal pair, L is an inductance, Z ０ is the characteristic impedance of the circuit, and ω (=2πf) is the angular velocity of frequency f.

[0025]

[0026] Assume that the inductance L has a dependency on temperature t [°C] as shown in equation (2). L'(t) is the inductance at temperature t, L ０ is the inductance at the reference temperature (here 25° C.).

[0027]

[0028] Inductance L ０ , temperature t, and the inductance L′(t) at temperature t, the S-parameters S １１By calculating the absolute value and argument of , 14 combinations of variables related to each other are obtained. An example of the combination of 14 variables is shown in FIG. 5. Here, the inductance L ０ By arbitrarily determining the temperature t, 11 sets of data, No. 1 to No. 11, were obtained. ０ was set to 50 Ω.

[0029] The conventional machine learning model was trained using the data No. 1 to No. 7 in Figure 5 as training data. Next, the temperature t and S-parameter S of No. 8 to No. 11 were １１ and are input into the trained machine learning model to calculate the inductance L ０ The estimated results are shown in Figure 6.

[0030] Similarly, the machine learning model 10 of this example was trained using the data of No. 1 to No. 7 as training data. Then, the temperature t and S parameter S of No. 8 to No. 11 were １１ and are input to the trained machine learning model 10 to obtain the inductance L ０ The estimated results are shown in Figure 7. ０ is the estimated value, L ０ＲＥＦ is the correct value (L in Figure 5 ０ 6 and 7, we can see that the values ​​obtained are close to the correct answer.

[0031] On the other hand, in the information processing device of this embodiment, the trained machine learning model 10 is used to ０ It is also possible to obtain estimated values ​​for variables other than the inductance L of No. 8 to No. 11. ０ and S-parameter S １１ and are input to the trained machine learning model 10, and the temperature t is estimated. The results are shown in FIG. 8. In FIG. 8, t is the estimated value, and t ＲＥＦ is the correct value (the value of t in FIG. 5).

[0032] In addition, the S parameters S of No. 8 to No. 11 １１ Of these, the absolute value |S at 8 GHz １１ S parameters S excluding | １１ and inductance L ０and temperature t are input to the trained machine learning model 10, and |S １１ The results of estimating | are shown in Figure 9. １１ | is the estimated value, |S １１ | ＲＥＦ is the correct value (S parameter S at 8 GHz in Figure 5 １１ 8 and 9 show that appropriate estimation is performed.

[0033] As described above, according to this embodiment, it is possible to arbitrarily select a variable to be estimated and have the machine learning model 10 output an estimated value.

[0034] The information processing device described in this embodiment can be realized by a computer equipped with a CPU (Central Processing Unit), a storage device, and an interface, and a program that controls these hardware resources. An example of the configuration of this computer is shown in FIG.

[0035] The computer includes a CPU 300, a storage device 301, and an interface device (I / F) 302. The I / F 302 is connected to, for example, the hardware of the input unit 12 and a display device that displays estimation results. In such a computer, a program for implementing the estimation method of the present invention is stored in the storage device 301. The CPU 300 executes the processing described in this embodiment in accordance with the program stored in the storage device 301.

[0036] Some or all of the above embodiments can be described as, but are not limited to, the following supplementary notes.

[0037] (Supplementary Note 1) The estimation method of the present invention includes a first step of providing a plurality of sets of data consisting of combinations of a plurality of variables as training data to a machine learning model constituted by an autoencoder neural network and training the machine learning model; a second step of inputting, to the trained machine learning model, an input variable from a plurality of input variables consisting of variables of known values ​​and a variable to be estimated, where the value of the variable to be estimated is set as a simulated value; and a third step of obtaining, from outputs of the machine learning model for the plurality of input variables, an output value corresponding to the variable to be estimated as an estimate of this variable.

[0038] (Supplementary Note 2) In the estimation method according to Supplementary Note 1, the variable is a physiological index.

[0039] (Supplementary Note 3) In the estimation method described in Supplementary Note 2, the simulated value is any one of an average value in the training data of the same variable as the estimation target, a value calculated from a variable different from the estimation target, a value of a variable different from the estimation target, and a value of the same variable as the estimation target in a living organism other than the living organism to be estimated.

[0040] (Supplementary Note 4) The information processing device of the present invention comprises a machine learning model formed by an autoencoder neural network, a learning unit configured to provide the machine learning model with multiple sets of data sets consisting of combinations of multiple variables as training data and to learn the machine learning model, an input unit configured to input, to the trained machine learning model, an input variable from multiple input variables consisting of variables of known values ​​and variables to be estimated, where the value of the variable to be estimated is set as a simulated value, and an acquisition unit configured to acquire, from outputs of the machine learning model for the multiple input variables, an output value corresponding to the variable to be estimated as an estimate of this variable.

[0041] The present invention can be applied to regression problems in which values ​​of unknown variables are estimated from input variables.

[0042] 10...machine learning model, 11...learning unit, 12...input unit, 13...acquisition unit, 100...input layer, 101...intermediate layer, 102...output layer

Claims

1. An estimation method comprising: a first step of providing a plurality of sets of data consisting of combinations of multiple variables as training data to a machine learning model constituted by an autoencoder neural network and training the machine learning model; a second step of inputting, into the trained machine learning model, input variables from a plurality of input variables consisting of variables of known values ​​and variables to be estimated, the input variables having simulated values ​​of the variables to be estimated; and a third step of obtaining, from among the outputs of the machine learning model for the plurality of input variables, an output value corresponding to the variable to be estimated as an estimate of this variable.

2. The estimation method according to claim 1, wherein the variable is a physiological index.

3. An estimation method according to claim 2, characterized in that the simulated value is either the average value in the training data of the same variable as the subject to be estimated, a value calculated from a variable different from the subject to be estimated, a value of a variable different from the subject to be estimated, or a value of the same variable as the subject to be estimated in a living organism other than the living organism to be estimated.

4. An information processing device comprising: a machine learning model formed by an autoencoder type neural network; a learning unit configured to provide the machine learning model with a plurality of sets of data sets consisting of combinations of a plurality of variables as training data and to train the machine learning model; an input unit configured to input, to the trained machine learning model, an input variable, from a plurality of input variables consisting of variables of known values ​​and variables to be estimated, which has a value that simulates the value of the variable to be estimated; and an acquisition unit configured to acquire, from outputs of the machine learning model for the plurality of input variables, an output value corresponding to the variable to be estimated as an estimate of this variable.

Images