Machine Learning Equipment

Abstract

A training unit (11) of a behavioral state learning system comprises an estimation unit (112) and a prediction model update unit (115). The estimation unit (112) estimates a predictive value from a measured value using a prediction model. The prediction model update unit (115) trains, while updating, the prediction model using relation data that indicates a relationship between actual values and predictive values so that the predictive values indicate a biased tendency with respect to the actual values.

Description

[Technical field]

[0001] The present invention relates to a learning technique for a prediction model used when estimating (predicting) a predetermined physical quantity from a measured value, and an estimation technique using the prediction model. [Background technology]

[0002] A behavioral state learning device is described in Patent Document 1. The behavioral state learning device described in Patent Document 1 constructs a model used to estimate a behavioral state from a signal measured by a vibration detection sensor.

[0003] The behavioral state learning device described in Patent Document 1 learns a behavioral state model using a signal measured by a vibration detection sensor and a signal measured by a myoelectric potential sensor. In this case, the behavioral state learning device described in Patent Document 1 learns the behavioral state model so that a predicted value predicted from the signal measured by the vibration detection sensor and an actual value obtained from the signal measured by the myoelectric potential sensor approximately match. [Prior art documents] [Patent documents]

[0004] [Patent Document 1] JP 2011-182824 A Summary of the Invention [Problem to be solved by the invention]

[0005] However, in a conventional learning device such as that shown in Patent Document 1, if the residual (a value obtained by subtracting an actual value from a predicted value) becomes positive, the behavior state model is updated so that a lower predicted value is output for the same input. Also, in a conventional learning device such as that shown in Patent Document 1, if the residual becomes negative, the behavior state model is updated so that a higher predicted value is output for the same input. Therefore, when the finally completed trained behavior state model is used, the positive and negative residuals vary on average.

[0006] In this case, for example, when a user performs muscle training while checking the predicted value output from the above-mentioned behavioral state model, if the residual is positive, a predicted value higher than the actual measured value is output. If the residual is positive for a long time, the time during which the load is less than the user's perception will be longer, resulting in poor training efficiency. On the other hand, if the residual is negative, a predicted value lower than the actual measured value is output. If the residual is negative for a long time, the time during which the user erroneously recognizes that the load is less than expected will be longer, which may reduce motivation for training.

[0007] That is, the predicted value may not be suitable for the user's purpose.

[0008] Therefore, an object of the present invention is to provide a learning device that can set predicted values ​​according to the purpose of the user while referring to actual measured values. [Means for solving the problem]

[0009] The machine learning device of the present invention includes an estimation unit and a prediction model update unit. The estimation unit estimates a predicted value from a measured value using a prediction model. The prediction model update unit learns while updating the prediction model using relational data indicating the relationship between the actual measured value and the predicted value so that the predicted value shows a tendency to be biased with respect to the actual measured value. The tendency of the predicted value to be biased with respect to the actual measured value means, for example, a state in which the number of relational data in a relationship in which the predicted value is smaller than the actual measured value is steadily greater than the number of relational data in a relationship in which the predicted value is equal to or greater than the actual measured value. Alternatively, the tendency of the predicted value to be biased with respect to the actual measured value means, for example, a state in which the number of relational data in a relationship in which the predicted value is greater than the actual measured value is steadily greater than the number of relational data in a relationship in which the predicted value is equal to or less than the actual measured value.

[0010] In this configuration, the predicted values ​​have a specific biased tendency with respect to the prediction curve based on the prediction model, so that the predicted values ​​are outputted with a specific tendency, without being uniformly dispersed with respect to the actual measured values. Effect of the Invention

[0011] According to the present invention, a predicted value according to the purpose of the user can be obtained by referring to an actual measurement value. [Brief description of the drawings]

[0012] [Figure 1] FIG. 1 is a functional block diagram of a behavioral state learning system according to an embodiment of the present invention. [Diagram 2] FIG. 2 is a functional block diagram of a learning unit according to an embodiment of the present invention. [Diagram 3] FIG. 3(A) is a diagram showing a first aspect of a concept for classifying a plurality of relational data, and FIG. 3(B) is a diagram showing a first aspect of an example of setting a loss value. [Figure 4] FIG. 4A is a diagram showing an example of a prediction model according to a conventional method for setting a loss value, and FIG. 4B is a diagram showing an example of a prediction model according to the method for setting a loss value of the present application. [Diagram 5] 5(A), 5(B), and 5(C) are diagrams showing the relationships (relationship data) between predicted values ​​and actual measured values. [Figure 6] FIG. 6 is a flowchart showing a behavioral state learning method according to this embodiment. [Figure 7] FIG. 7(A) is a flowchart showing a method for generating an input vector, and FIG. 7(B) is a flowchart showing a method for generating training data. [Figure 8] FIG. 8 is a flowchart showing a process for determining the end of learning of a predictive model. [Figure 9] FIG. 9(A) is a diagram showing a second aspect of the concept of classifying a plurality of relational data, and FIG. 9(B) is a diagram showing a second aspect of an example of setting a loss value. [Figure 10] FIG. 10A is a diagram showing a third aspect of the concept of classifying a plurality of relational data, and FIG. 10B is a diagram showing a third aspect of an example of setting a loss value. [Figure 11] FIG. 11(A) is a diagram showing a fourth aspect of the concept of classifying a plurality of relational data, and FIG. 11(B) is a diagram showing a fourth aspect of an example of setting a loss value. [Figure 12] FIG. 12(A) is a diagram showing a fifth aspect of the concept of classifying a plurality of relational data, and FIG. 12(B) is a diagram showing a fifth aspect of an example of setting a loss value. [Figure 13] FIG. 13 is a functional block diagram of a behavioral state estimating device according to an embodiment of the present invention. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0013] [Behavioral state learning device] A behavioral state learning device according to an embodiment of the present invention will be described with reference to the drawings. Fig. 1 is a functional block diagram of a behavioral state learning system according to an embodiment of the present invention.

[0014] As shown in Fig. 1, the behavioral state learning system includes a learning unit 11, a prediction model storage unit 12, a displacement detection sensor 101, and a muscle activity detection sensor 102. The learning unit 11 corresponds to the "machine learning device." The displacement detection sensor 101 corresponds to the "first sensor," and the muscle activity detection sensor 102 corresponds to the "second sensor."

[0015] The displacement detection sensor 101 is, for example, at least one of a tremor sensor, an acceleration sensor, and an angular velocity sensor. The displacement detection sensor 101 detects the displacement of an observation target and outputs a measurement value according to the magnitude of the displacement.

[0016] The muscle activity detection sensor 102 is, for example, at least one of an electromyogram sensor, a camera for determining posture, a posture sensor, etc. The muscle activity detection sensor 102 detects muscle movement of the observation target and outputs an actual measurement value according to the magnitude of the movement.

[0017] For example, when observing muscle activity in the lower limbs of a subject, the displacement detection sensor 101 is attached to the ankle of the subject. The muscle activity detection sensor 102 is placed on the surface of the skin where the muscle to be observed is located.

[0018] After performing the learning described below, muscle activity in the subject's lower limbs is observed only by the displacement detection sensor 101 attached to the subject's ankle. Therefore, it is necessary to estimate the amount of muscle activity (e.g., myoelectric potential) from the measurement value output by the displacement detection sensor 101, and for this purpose, the learning described below is performed.

[0019] The learning unit 11 updates and learns a prediction model for the above estimation using the measurement values ​​from the displacement detection sensor 101 and the actual measurement values ​​from the muscle activity detection sensor 102. In general, the learning unit 11 estimates a predicted value from a measured value using the immediately previous prediction model. The learning unit 11 learns while updating the prediction model using relational data indicating the relationship between the actual measurement value and the predicted value so that the predicted value shows a tendency to be biased against the actual measurement value. The learning unit 11 outputs the updated (learned) prediction model to the prediction model storage unit 12.

[0020] The tendency for predicted values ​​to be biased relative to actual measured values ​​means, for example, a state in which the number of pieces of relational data whose predicted values ​​are smaller than the actual measured values ​​is steadily greater than the number of pieces of relational data whose predicted values ​​are equal to or greater than the actual measured values, or, the tendency for predicted values ​​to be biased relative to actual measured values ​​means, for example, a state in which the number of pieces of relational data whose predicted values ​​are larger than the actual measured values ​​is steadily greater than the number of pieces of relational data whose predicted values ​​are equal to or smaller than the actual measured values.

[0021] Furthermore, the tendency for predicted values ​​to be biased relative to actual measured values ​​means, for example, a state in which most of the relational data have a relationship in which the predicted values ​​are smaller than the actual measured values, compared to all the relational data, or the tendency for predicted values ​​to be biased relative to actual measured values ​​means, for example, a state in which most of the relational data have a relationship in which the predicted values ​​are larger than the actual measured values, compared to all the relational data.

[0022] Furthermore, the tendency of predicted values ​​to be biased relative to actual measured values ​​means, for example, a state in which the predicted values ​​are smaller than the actual measured values ​​in all relational data, or the tendency of predicted values ​​to be biased relative to actual measured values ​​means, for example, a state in which the predicted values ​​are larger than the actual measured values ​​in all relational data.

[0023] In the following, as an example, a state will be described in which the number of relational data whose predicted value is smaller than the actual measured value is steadily greater than the number of relational data whose predicted value is equal to or greater than the actual measured value, a state in which the majority of the relational data have predicted values ​​smaller than the actual measured value, or a state in which the predicted values ​​of all the relational data are smaller than the actual measured value.

[0024] The prediction model storage unit 12 stores the prediction model learned by the learning unit 11. The prediction model after learning is used for the above-mentioned estimation.

[0025] (Learning Section 11) Fig. 2 is a functional block diagram of a learning unit according to an embodiment of the present invention. As shown in Fig. 2, the learning unit 11 includes a learning data generating unit 111, an estimating unit 112, a data acquiring unit 113, a learning condition setting unit 114, and a predictive model updating unit 115. The learning unit 11 is realized by a hardware arithmetic processing unit such as a CPU, a program for executing each function of the learning unit 11, and a storage unit that stores the program. The storage unit is also used during processing by the arithmetic processing unit. The arithmetic processing unit executes this program, and the learning unit 11 realizes each of the above-mentioned functions.

[0026] The learning data generating unit 111 receives as input the measured values ​​from the displacement detection sensor 101 and the actual measured values ​​from the muscle activity detection sensor 102 .

[0027] The learning data generating unit 111 generates an input vector for a prediction model from the measurement values. For example, the learning data generating unit 111 detects a feature of the measurement values ​​and generates the feature as an input vector. The feature is, for example, a feature indicating a feature of the time waveform of the measurement values ​​(maximum value, average value, etc.) or a feature indicating a feature of the frequency spectrum of the measurement values ​​(maximum spectrum value, maximum spectrum frequency, etc.). Note that the feature is not limited to this, and may be anything that can be associated with the movement of the observation target location of the subject.

[0028] The learning data generating unit 111 generates teacher data for a prediction model from actual measured values. For example, the learning data generating unit 111 detects a behavioral state component from the actual measured values ​​and generates the behavioral state component as an input vector. The behavioral state component is a component that can be associated with the movement of the observation target part of the subject in the actual measured values.

[0029] The learning data generation unit 111 outputs the input vector and the behavioral state component to the prediction model update unit 115. At this time, the learning data generation unit 111 outputs a pair of the input vector and the behavioral state component obtained from the measured value and the actual measured value at approximately the same time. Note that the measured value and the actual measured value at approximately the same time refer to the measured value and the actual measured value obtained at synchronized sampling timing. In addition, the learning data generation unit 111 outputs the input vector to the estimation unit 112.

[0030] The estimation unit 112 estimates a predicted value from the input vector using a prediction model. The prediction model used at this time is the prediction model obtained by the immediately preceding update. The estimation unit 112 outputs the predicted value to the data acquisition unit 113.

[0031] The data acquisition unit 113 receives the actual measurement value and the predicted value. The data acquisition unit 113 generates relationship data indicating the relationship between the actual measurement value and the predicted value. The relationship data has the actual measurement value and the predicted value as the first and second components of the data, respectively. The predicted value in this case is estimated from a measurement value detected at approximately the same time as the actual measurement value. The time of detection is also associated with the relationship data.

[0032] The data acquiring unit 113 generates relational data in a predetermined cycle from the actual measured values ​​and predicted values ​​that are input in sequence. In this way, the data acquiring unit 113 generates a plurality of relational data (relational data at a plurality of times). The data acquiring unit 113 outputs the plurality of relational data to the learning condition setting unit 114.

[0033] The learning condition setting unit 114 classifies a plurality of relational data, and sets learning conditions to be used for updating and learning the prediction model, using a learning loss value set for each classification.

[0034] More specifically, the learning condition setting unit 114 includes an area setting unit 141 , a loss value setting unit 142 , a number detection unit 143 , an evaluation value calculation unit 144 , and a total value calculation unit 145 .

[0035] The region setting unit 141 sets a region for classifying a plurality of relational data. FIG. 3(A) is a diagram showing a first aspect of the concept of classifying a plurality of relational data, and FIG. 3(B) is a diagram showing a first aspect of a setting example of a loss value. As shown in FIG. 3(A), the region setting unit 141 sets an orthogonal coordinate system with the actual measured value and the predicted value as two orthogonal axes. The region setting unit 141 sets a reference line 90 for the orthogonal coordinate system of the actual measured value and the predicted value. The reference line 90 is set by a collection of points where the actual measured value and the predicted value match, that is, a collection of points where the residual (predicted value-actual measured value) is zero.

[0036] The region setting unit 141 sets a region 91 on the actual value side of the reference line 90 and a region 92 on the predicted value side of the reference line 90. In other words, the region setting unit 141 sets a region 91 where the residual is a negative value and a region 92 where the residual is a positive value.

[0037] The loss value setting unit 142 sets loss values ​​for a plurality of regions 91 and a plurality of regions 92. For example, as shown in FIG. 3(B), the loss value setting unit 142 sets a loss value β1 for the relationship data of region 91, and sets a loss value β2 for the relationship data of region 92. The loss values ​​β1 and β2 are positive values. The loss value β2 is greater than the loss value β1 (β2>β1). Note that the loss value setting unit 142 sets a loss value of 0 for the relationship data on the reference line 90.

[0038] Number detection unit 143 calculates the residual for each of the multiple relational data, and classifies the data into region 91 and region 92. Specifically, if the residual between the predicted value and actual value of the relational data is a negative value, number detection unit 143 determines that the relational data belongs to region 91. If the residual between the predicted value and actual value of the relational data is a positive value, number detection unit 143 determines that the relational data belongs to region 92.

[0039] The number detection unit 143 detects the number of pieces of relational data belonging to region 91 (first number n1) and the number of pieces of relational data belonging to region 92 (second number n2). The number detection unit 143 outputs the first number n1 and the second number n2 to the evaluation value calculation unit 144.

[0040] The evaluation value calculation unit 144 calculates an evaluation value L1 using the loss value β1, the loss value β2, the first number n1, and the second number n2. Specifically, the evaluation value calculation unit 144 calculates an evaluation value L1 for the region 91 (relationship data whose residual is a negative value) by multiplying the first number n1 by the loss value β1 (n1×β1). The evaluation value calculation unit 144 calculates an evaluation value L2 for the region 92 (relationship data whose residual is a positive value) by multiplying the second number n2 by the loss value β2 (n1×β1). The evaluation value calculation unit 144 outputs the evaluation values ​​L1 and L2 to the total value calculation unit 145.

[0041] The sum calculation unit 145 adds the evaluation value L1 and the evaluation value L2 to calculate the sum L. That is, the sum calculation unit 145 performs the calculation of L=L1+L2=n1×β1+n2×β2. The sum calculation unit 145 outputs the sum L to the prediction model update unit 115 as a learning condition.

[0042] In this way, the learning condition setting unit 114 does not set a loss value for each relational data item individually depending on the magnitude of the residual, but sets a loss value for the relational data item depending only on whether the residual is positive or negative, regardless of the absolute value of the residual.

[0043] An input vector based on the measured values ​​and teacher data based on the actual measured values ​​are input to the prediction model update unit 115. In addition, the prediction model update unit 115 also receives a learning condition (total value L).

[0044] The prediction model update unit 115 updates and learns the prediction model so that the learning condition (total value L) is lowered. The prediction model is learned by a known method, for example, using a recognizer capable of multi-class classification, such as SVM, GMM, HMM, neural network, learning Bayesian network, or a configuration of multiple recognizers. As for the recognizer, multiple recognizers may be combined, and for example, a method such as boosting, represented by the method AdaBoost, may be used. In addition, a single recognition method may be configured by configuring multiple stages of recognizers that can only classify one class or recognizers that can classify multiple classes.

[0045] The prediction model update unit 115 updates the prediction model while repeating learning, and when it detects the end of learning, outputs the prediction model at that time to the prediction model storage unit 12.

[0046] Here, the learning condition setting unit 114 sets the loss value β2 to be larger than the loss value β1. As a result, the prediction model takes the following shape, and accordingly, the relationship between the predicted value and the actual measurement value also transitions as follows.

[0047] Fig. 4(A) is a diagram showing an example of a prediction model according to a conventional method for setting a loss value, and Fig. 4(B) is a diagram showing an example of a prediction model according to the method for setting a loss value of the present application. The conventional method for setting a loss value shown in Fig. 4(A) is a method for individually setting a loss value for relational data according to the magnitude of the residual.

[0048] 4(A), in the conventional method of setting a loss value, a plurality of points defined by an input vector and training data are approximately evenly distributed on both sides of a prediction model curve 80. That is, there are approximately the same number of points where the training data is higher than the input vector and points where the training data is lower than the input vector.

[0049] However, as shown in Fig. 4(B), in the loss value setting method of the present application, a plurality of points defined by the input vector and the training data are concentrated on one side of the prediction model curve 80. In other words, the points where the training data is higher than the input vector are collected, and there are almost no points where the training data is lower than the input vector.

[0050] As a result, the relationship between the predicted value and the actual measurement value transitions as shown in Figures 5(A), 5(B), and 5(C). Figures 5(A), 5(B), and 5(C) are diagrams showing the relationship (relationship data) between the predicted value and the actual measurement value. Learning progresses in the order of Figures 5(A), 5(B), and 5(C).

[0051] As shown in Figures 5(A), 5(B), and 5(C), as learning progresses, the number of points (relational data) where the predicted value is greater than the actual measured value decreases. And eventually, regardless of the magnitude of the predicted value and the actual measured value, the predicted value will be smaller than the actual measured value in most cases.

[0052] Therefore, when the prediction model is used to estimate a predicted value from a measured value, the predicted value is smaller than the actual measured value. As a result, for example, when training is performed using the predicted value, the time during which the load is heavier than the user (subject) recognizes is longer, and training efficiency can be improved.

[0053] In addition, if the relationship between the loss value β1 and the loss value β2 is reversed (β2<β1), the predicted value will be larger than the actual measured value when the predicted value is estimated from the measured value using this prediction model. As a result, for example, when training using this predicted value, the time during which the user feels that they are under strain will be longer than expected, which can improve motivation for training.

[0054] In this way, by using the behavioral state learning system of this embodiment, a prediction model can be constructed so as to obtain a desired estimation result while using measured values ​​and actual measured values.

[0055] (Behavioral state learning method) Fig. 6 is a flowchart showing a behavioral state learning method according to this embodiment. Fig. 7(A) is a flowchart showing a method for generating an input vector, and Fig. 7(B) is a flowchart showing a method for generating teacher data. Fig. 8 is a flowchart for determining the end of learning of a prediction model. Note that the specific contents of each process are performed in the above-mentioned configuration, and explanations will be omitted below except for points requiring additional explanation.

[0056] The learning unit 11 of the behavioral state learning system generates an input vector from a measurement value, and generates teacher data from an actual measurement value (S11). More specifically, the learning unit 11 acquires a measurement value (S21), and extracts a feature amount of the measurement value (S22). The learning unit 11 generates an input vector using the feature amount of the measurement value (S23). The learning unit 11 also acquires an actual measurement value (S31), and calculates a behavioral state component amount from the actual measurement value (S32). The learning unit 11 generates teacher data from the behavioral state component amount (S33).

[0057] The learning unit 11 generates relationship data between predicted values ​​and actual measured values ​​(S12). The learning unit 11 calculates the numbers n1 and n2 of pieces of relationship data for each of the regions 91 and 92 (S13). The learning unit 11 calculates evaluation values ​​L1 and L2 for each of the regions 91 and 92 using the loss values ​​β1 and β2 and the numbers n1 and n2 set for each of the regions 91 and 92 (S14).

[0058] The learning unit 11 calculates the total value L using the evaluation values ​​L1 and L2 (S15). The learning unit 11 updates (learns) the prediction model using the total value L as a learning condition (S16).

[0059] When the learning end condition is met (S17: YES), the learning unit 11 outputs the prediction model (S18). If the learning end condition is not met (S17: NO), the learning unit 11 repeats learning of the prediction model.

[0060] More specifically, the learning unit 11 compares the previous total value (learning condition) with the current total value (learning condition) (S41). If the current total value is smaller than the previous total value (S42: YES), the learning unit 11 updates to the currently learned prediction model (S43) and repeats learning of the prediction model. If the current total value is not smaller than the previous total value (S42: NO), the learning unit 11 maintains the prediction model from the previous learning (S44). If the number of times the prediction model has been maintained is less than the threshold number of times (S45: YES), the learning unit 11 continues learning (S46). If the number of times the prediction model has been maintained reaches the threshold number of times (S45: NO), the learning unit 11 ends learning (S47).

[0061] (Alternatives for setting area and loss value) FIG. 9(A) is a diagram showing a second aspect of the concept of classifying a plurality of relational data, and FIG. 9(B) is a diagram showing a second aspect of an example of setting a loss value.

[0062] As shown in Fig. 9(A), as a second mode of the region and the loss value, a region 91A, a region 92A, and a region 93A are set. The region 91A is a region where the residual is a negative value and the absolute value is equal to or greater than the region determination threshold. The region 92A is a region where the residual is a positive value and the absolute value is equal to or greater than the region determination threshold. The region 93A is a region where the absolute value of the residual is less than the region determination threshold.

[0063] A loss value β1 is set in area 91A, and a loss value β2 is set in area 92A. The loss value β1 is smaller than the loss value β2. The loss value in area 93A is set to 0.

[0064] By making such settings, the influence of measurement errors between the measured values ​​and the actual measured values ​​on the learning of the prediction model can be suppressed.

[0065] FIG. 10A is a diagram showing a third aspect of the concept of classifying a plurality of relational data, and FIG. 10B is a diagram showing a third aspect of an example of setting a loss value.

[0066] As shown in Fig. 10(A), as a third mode of the region and the loss value, a region 91B, a region 92B, and a region 93B are set. The region 91B is a region where the residual is a negative value and the absolute value is equal to or greater than the first region determining threshold. The region 92B is a region where the residual is a positive value and the absolute value is equal to or greater than the second region determining threshold. The region 93A is a region where the residual is a negative value and the absolute value is less than the first region determining threshold, or where the residual is a positive value and the absolute value is less than the second region determining threshold.

[0067] A loss value β1 is set in region 91B, and a loss value β2 is set in region 92B. The loss value β1 is smaller than the loss value β2. The loss value in region 93B is set to 0.

[0068] By making such settings, it is possible to suppress the effect of measurement errors between the measured values ​​and the actual measured values ​​on the learning of the prediction model. Also, with this setting, the area where the loss value is 0 on the negative side of the residual is wider than the area where the loss value is 0 on the positive side of the residual. This makes it possible to set a prediction model where the predicted value is more reliably smaller than the actual measured value.

[0069] FIG. 11(A) is a diagram showing a fourth aspect of the concept of classifying a plurality of relational data, and FIG. 11(B) is a diagram showing a fourth aspect of an example of setting a loss value.

[0070] As shown in Fig. 11(A), as a fourth mode of the region and the loss value, a region 91C, a region 92C, and a region 93C are set. The region 91C is a region where the residual is a negative value and the actual measurement value is equal to or greater than the third region determination threshold. The region 92C is a region where the residual is a positive value and the actual measurement value is equal to or greater than the third region determination threshold. The region 93C is a region where the absolute value of the residual is less than the region determination threshold or the actual measurement value is less than the third region determination threshold.

[0071] A loss value β1 is set in area 91C, and a loss value β2 is set in area 92C. The loss value β1 is smaller than the loss value β2. The loss value in area 93C is set to 0.

[0072] By making such settings, it is possible to suppress the influence of small actual measured values ​​on the learning of the prediction model, i.e., it is possible to learn the prediction model assuming only cases where the estimated predicted value is within the range of actual use.

[0073] FIG. 12(A) is a diagram showing a fifth aspect of the concept of classifying a plurality of relational data, and FIG. 12(B) is a diagram showing a fifth aspect of an example of setting a loss value.

[0074] 12(A), as a fifth mode of the region and the loss value, a region 91D and a region 92D are set. The region 91D is a region where the residual is a negative value and the absolute value is equal to or greater than the region determination threshold. The region 92D is a region where the residual is a positive value or a negative value and the absolute value is less than the region determination threshold.

[0075] A loss value β1 is set in the area 91 D, and a loss value β2 is set in the area 92 D. The loss value β1 is smaller than the loss value β2.

[0076] By performing such setting, the region 92D includes the region of the reference line 90 and is wider than the region 91D. This makes it possible to set a prediction model in which the predicted value is more reliably smaller than the actual measured value.

[0077] In each of the above-described aspects, the predicted value is smaller than the actual measured value. However, by applying the above-described concept and appropriately setting each region, it is also possible to make the predicted value larger than the actual measured value.

[0078] (Behavioral state estimation device) 13 is a functional block diagram of a behavior state estimation device according to an embodiment of the present invention. As shown in FIG. 13, the behavior state estimation device 16 includes an input vector generation unit 161 and an estimation calculation unit 162.

[0079] The input vector generation unit 161 generates an input vector by using a measurement value from the displacement detection sensor 101. The method of generating the input vector is similar to that of the above-mentioned learning data generation unit 111. The input vector generation unit 161 outputs the generated input vector to the estimation calculation unit 162.

[0080] The estimation calculation unit 162 estimates a predicted value from an input vector by using the prediction model stored in the prediction model storage unit 12. At this time, the estimation calculation unit 162 uses the prediction model learned as described above.

[0081] With such a configuration, the behavior state estimation device 16 can estimate the predicted values ​​so as to show a biased tendency with respect to the actual measured values. For example, the behavior state estimation device 16 can estimate the predicted values ​​so that the predicted values ​​are smaller than the actual measured values. More specifically, the behavior state estimation device 16 can estimate the predicted values ​​so that all the predicted values ​​are smaller than the actual measured values. Alternatively, the behavior state estimation device 16 can estimate the predicted values ​​so that the number of times the predicted values ​​are smaller than the actual measured values ​​is smaller than the number of times the predicted values ​​are larger than the actual measured values.

[0082] Alternatively, the behavioral state estimation device 16 can estimate the predicted value so that the predicted value is greater than the actual measured value.

[0083] Also, the behavioral state estimation device 16 can select whether the predicted value is made smaller or larger than the actual measurement value. In this case, for example, the behavioral state estimation system includes an operation input unit such as a switch. The user operates the operation input unit according to his / her purpose. This allows the behavioral state estimation device 16 to select whether the predicted value is made smaller or larger than the actual measurement value.

[0084] Although the above example shows an application to muscle training, the present invention can also be applied to rehabilitation aimed at restoring muscle strength. [Explanation of symbols]

[0085] 11: Learning Department 12: Prediction model memory unit 101: Displacement detection sensor 102: Muscle activity detection sensor 111: Learning data generation unit 112: Estimation part 113: Data acquisition section 114: Learning condition setting unit 115: Prediction model update unit 141: Area setting section 142: Loss value setting unit 143: Number detection unit 144: Evaluation value calculation unit 145: Total value calculation unit 16: Behavioral state estimation device 161: Input vector generation unit 162: Estimation calculation section

Claims

1. an estimation unit that estimates a predicted value from a measured value using a prediction model; a prediction model update unit that updates and learns the prediction model using relationship data indicating a relationship between an actual measurement value and the predicted value so that the predicted value shows a biased tendency with respect to the actual measurement value; a data acquisition unit that acquires relationship data indicating a relationship between the predicted value and the actual measurement value a plurality of times; a learning condition setting unit that classifies the plurality of relational data according to residuals and sets learning conditions using a learning loss value that is set for each of the classifications; Equipped with The learning condition setting unit a region setting unit that sets a plurality of regions for a distribution of the relational data acquired a plurality of times as a classification of the plurality of relational data; a loss value setting unit that sets a different loss value for each of the plurality of regions; a number detection unit that detects into which region each of the relational data acquired multiple times is classified using the residuals, and detects the number of the relational data for each of the multiple regions; an evaluation value calculation unit that calculates an evaluation value for each of the plurality of regions using the number and the loss value; a total value calculation unit that calculates a total value of the evaluation values ​​of the plurality of regions; Further equipped with The prediction model update unit, updating the prediction model using the learning condition such that the total value decreases as the learning condition; Machine learning device.

2. The prediction model update unit, repeating the update of the prediction model such that the total value after the update of the prediction model is smaller than the total value before the update of the prediction model. The machine learning device according to claim 1 .

3. The prediction model update unit, If the sum value after updating the prediction model is smaller than the sum value before updating the prediction model, continue updating the prediction model. The machine learning device according to claim 1 or 2.

4. The prediction model update unit, If the sum value after the update of the prediction model is greater than the sum value before the update of the prediction model, the prediction model is not updated. The machine learning device according to claim 1 or 2.

5. The prediction model update unit, When the number of times that the prediction model is not updated reaches a threshold number of times, the update of the prediction model is stopped. The machine learning device according to claim 1 or 2.

6. The estimation unit estimates the predicted value using a previously updated prediction model. The machine learning device according to claim 1 or 2.

7. A first sensor that measures the measurement value; a second sensor that is a different type of sensor from the first sensor and that measures the actual measurement value; Prepare, The machine learning device according to claim 1 or 2.

8. the first sensor includes a tremor sensor; The second sensor is an electromyographic sensor. The machine learning device according to claim 7.

9. The first sensor includes at least one of an acceleration sensor and an angular velocity sensor. The machine learning device according to claim 8.

10. The first sensor is attached to the subject's ankle, The predictive model is a model for predicting muscle activity of the lower limbs; The machine learning device according to claim 7.

11. The loss value setting unit varies the sizes of the plurality of regions. The machine learning device according to claim 1 or 2.

12. The loss value setting unit varies the loss value for each of the plurality of regions. The machine learning device according to claim 1 or 2.

13. The loss value setting unit is A plurality of regions in which the loss value is not 0 and a region in which the loss value is 0 are set. The machine learning device according to claim 1 or 2.

14. The loss value setting unit is setting the loss value only for the relational data whose actual measurement value is equal to or greater than a predetermined value; The machine learning device according to claim 1 or 2.

Images