Method and control device for configuring preprocessing of supervised learning models

Abstract

An exemplary method includes: defining a function g having a plurality of parameters a-l; placing the function g in front of a supervised learning model f, where the supervised learning model f is trained with training data TDi including instances xi and labels yi, where i is an integer; defining a function f(g(xi)) using the supervised learning model f and the function g; calculating a result Zi by calculating the function f(g(xi)); calculating a difference Li between the result Zi and the label yi; calculating a partial derivative PDi of the difference Li with respect to the plurality of parameters a-l of the function g; updating the plurality of parameters a-l of the function g according to the partial derivative PDi; defining a pseudo-preprocessing h using the plurality of parameters a-l of the function g if the difference Li converges during an iteration between the calculation of the result Zi and the update of the plurality of parameters a-l of the function g; placing the pseudo-preprocessing h in front of the supervised learning model f; and providing the training data TDi to the pseudo-preprocessing h to learn the supervised learning model f.

Description

[Technical Field]

[0001] The present disclosure provides a method for configuring preprocessing used in supervised learning models and Control device In particular, a method for evaluating multiple parameters used in preprocessing and Control device Regarding. [Background technology]

[0002] The object detection device described in Patent Document 1 performs preprocessing on images captured by a camera to detect objects in the images. In order to improve the accuracy of object detection, the object detection device adjusts multiple parameters included in the preprocessing before executing the preprocessing.

[0003] Figure 23 shows a conventional design of preprocessing used for supervised learning models. Figure 23 shows preprocessing PP1 designed by engineer E (A) for the supervised learning model SLM (for 32-bit processing) using training data TD1 to TDn (n is an integer equal to or greater than 2). Here, preprocessing PP1 includes multiple parameters that cannot be reproduced or are unknown to engineer E (B) or other engineers, for example, and engineer E (B) is attempting to convert or transfer the supervised learning model SLM from 32-bit processing to 64-bit processing.

[0004] However, because the parameters used in preprocessing PP1 are unreproducible or unknown, Engineer E (B) is unable to reproduce the parameters used in preprocessing PP1 and constructs another preprocessing PP2 using inappropriate parameters different from those used in preprocessing PP1. As a result, Engineer E (B) is unable to convert or transfer the supervised learning model SLM from 32-bit processing to 64-bit processing. [Prior art documents] [Patent documents]

[0005] [Patent Document 1] Japanese Patent Application Publication No. 2019-219804

[0006] Summary of the Invention

[0007] In order to solve the above problem, one aspect of the present disclosure is to A method performed by a control device including a first definition unit, a first disposer, a second definition unit, a first calculation unit, a second calculation unit, a third calculation unit, an update unit, a third definition unit, a second disposer, and a trainer, wherein the first definition unit: Defining a function g with multiple parameters a to l, The first disposer is Putting a function g in front of a supervised learning model f, The aforementioned A supervised learning model f is trained on training data TDi containing examples xi and labels yi. ( i is an integer ) , The second definition unit is Supervised learning model f and As defined above Define a function f(g(xi)) using functions g and xi. The first calculation unit is Calculating the function f(g(xi)) and calculating the result Zi; The second calculation unit Result Zi and The aforementioned Calculate the difference Li with the label yi, The third calculation unit is For multiple parameters a to l of function g The calculated Calculating the partial derivative PDi of the difference Li; The update unit is of function g The aforementioned Multiple parameters a~l The calculated updating according to the partial derivatives PDi, The third calculation unit Calculate the result Zi and As defined above of function g The aforementioned Multiple parameters a to l The aforementioned During the iteration between updates The aforementioned If the calculated difference Li converges, As defined above of function g The aforementioned Defining a pseudo-preprocessing function h using multiple parameters a to l; The second disposer placing the pseudo-preprocessing h defined above before the supervised learning model f; The trainer Training data TDi As defined above Provided to pseudo-pretreatment h, The aforementioned training a supervised learning model f.

[0008] Another aspect of the present disclosure is a method for providing training data TDi (i is an integer) to a first supervised learning model f1 to obtain a first plurality of combinations COMB_a1 to COMB_l1, each of which includes various values ​​for one of a plurality of parameters a to l of a first function g1 that is disposed before the first supervised learning model f1 that performs feedback to the first function g1; and providing training data TDi to a second supervised learning model f2 to obtain a first plurality of combinations COMB_a1 to COMB_l1, each of which includes various values ​​for one of a plurality of parameters a to l of a second function g2 that is disposed before the second supervised learning model f2 that performs feedback to the second function g2. and obtaining a second plurality of combinations COMB_a2 to COMB_l2 including the first and second supervised learning models f1 and f2, the operations of which are equivalent to each other, calculating a first plurality of distributions D_a1 to D_l1 for each of the first plurality of combinations COMB_a1 to COMB_l1, calculating a second plurality of distributions D_a2 to D_l2 for each of the second plurality of combinations COMB_a2 to COMB_l2, and evaluating which of the first function g1 and the second function g2 is more robust by comparing the first plurality of distributions D_a1 to D_l1 with the second plurality of distributions D_a2 to D_l2, respectively.

[0009] Yet another aspect of the present disclosure provides a method including the steps of: obtaining a plurality of combinations COMB(1) through COMB(n), each including a plurality of values ​​of a plurality of parameters a through l of a function g, by providing training data TDi (n is an integer equal to or greater than 1) to the supervised learning model f, the function g being placed before the supervised learning model f that performs feedback to the function g; arranging the plurality of combinations COMB(1) through COMB(n) in a space having a number of dimensions equal to the number of the plurality of parameters a through l of the function g; obtaining at least one sample point SP in the space; selecting at least one of the plurality of combinations COMB(1) through COMB(n) that is closest to the at least one sample point SP; and assigning a plurality of values ​​included in the selected combination COMB(1) through COMB(n) from the plurality of combinations COMB(1) through COMB(n) to the plurality of parameters a through l of the function g.

[0010] Yet another aspect of the present disclosure includes a first definition unit that defines a function g having a plurality of parameters a to l; a first disposer that places the function g in front of a supervised learning model f; a second definition unit that defines a function f(g(xi)) using the supervised learning model f and the function g; a first calculation unit that calculates the function f(g(xi)) to calculate a result Zi; a second calculation unit that calculates the difference L between the result Zi and the label yi; The present invention provides an apparatus comprising: a third calculation unit that calculates a partial derivative PDi of the difference Li with respect to a number of parameters a to l; an update unit that updates the plurality of parameters a to l of the function g according to the partial derivative PDi; a third definition unit that defines a pseudo-preprocessing h using the plurality of parameters a to l of the function g when the difference Li converges during the repetition of calculating the result Zi and updating the plurality of parameters a to l of the function g; a second disposer that places the pseudo-preprocessing h before the supervised learning model f; and a trainer that provides training data TDi to the pseudo-preprocessing h to train the supervised learning model f.

[0011] Yet another aspect of the present disclosure includes a first acquisition unit (i is an integer) that acquires first combinations COMB_a1 to COMB_l1, each including various values ​​for one of a plurality of parameters a to l of a first function g1, by providing training data TDi to the first supervised learning model f1, and a second acquisition unit (i is an integer) that acquires second combinations COMB_a2 to COMB_l2, each including various values ​​for one of a plurality of parameters a to l of a second function g2, by providing training data TDi to the second supervised learning model f2. The present invention provides an apparatus comprising: a second acquisition unit that acquires training data TDi by providing it (the operation of the first supervised learning model f1 and the operation of the second supervised learning model f2 are equivalent to each other); a first calculation unit that calculates first multiple distributions D_a1 to D_l1 for each of the first multiple combinations COMB_a1 to COMB_l1; a second calculation unit that calculates second multiple distributions D_a2 to D_l2 for each of the second multiple combinations COMB_a2 to COMB_l2; and an evaluation unit that compares the first multiple distributions D_a1 to D_l1 with the second multiple distributions D_a2 to D_l2 and evaluates which of the first function g1 and the second function g2 is more robust.

[0012] Yet another aspect of the present disclosure provides an apparatus including: a first acquisition unit that acquires a plurality of combinations COMB(1) to COMB(n), each including a plurality of values ​​of a plurality of parameters a to l of function g, by providing training data TDi (n is an integer equal to or greater than 2) to the supervised learning model f, the plurality of combinations COMB(1) to COMB(n), the plurality of combinations being placed before the supervised learning model f that realizes feedback to the function g; a disposer that arranges the plurality of combinations COMB(1) to COMB(n) in a space having a plurality of dimensions equal to the number of the plurality of parameters a to l of the function g; a second acquisition unit that acquires at least one sample point SP in the space; a selection unit that selects at least one combination COMB(1) to COMB(n) from the plurality of combinations COMB(1) to COMB(n) that is closest to the at least one sample point SP; and an allocation unit that allocates a plurality of values ​​included in one of the selected plurality of combinations COMB(1) to COMB(n) to the plurality of parameters a to l of the function g. [Brief explanation of the drawings]

[0013] [Figure 1] FIG. 1 is a block diagram showing the configuration of a controlled system CS controlled by a control device CD according to the first embodiment. [Figure 2] FIG. 2 shows the configuration of the function g according to the first embodiment. [Figure 3] FIG. 3 shows the configuration of the training data TD1 to TDn. [Figure 4] FIG. 4 is a block diagram showing the functional configuration of the control device CD according to the first embodiment. [Figure 5] FIG. 5 is a block diagram showing the configuration of the control device CD according to the first embodiment from the viewpoint of hardware. [Figure 6] FIG. 6 is a block diagram showing the configuration of the control device CD according to the first embodiment from the viewpoint of software. [Figure 7] FIG. 7 is a flowchart showing the operation (part 1) of the control device CD according to the first embodiment. [Figure 8] FIG. 8 is a flowchart showing the operation (part 2) of the control device CD according to the first embodiment. [Figure 9] FIG. 9 shows the configuration of the control device CD according to the first embodiment when i reaches 80. [Figure 10] FIG. 10 is a block diagram showing the configuration of a first controlled system CS1 and a second controlled system CS2 controlled by a control device CD according to the second embodiment. [Figure 11] FIG. 11 is a block diagram showing the functional configuration of the control device CD according to the second embodiment. [Figure 12] FIG. 12 is a flowchart showing the operation of the control device CD according to the second embodiment. [Figure 13] FIG. 13 shows a plurality of combinations COMB_a1 to COMB_l1 including various values ​​of a plurality of parameters a to l of the first function g1. [Figure 14] FIG. 14 shows a plurality of combinations COMB_a2 to COMB_l2 including various values ​​of a plurality of parameters a to l of the second function g2. [Figure 15] FIG. 15 shows a plurality of distributions from a distribution D_a1 for the combination COMB_a1 to a distribution D_l1 for the combination COMB_l1, and a plurality of distributions from a distribution D_a2 for the combination COMB_a2 to a distribution D_l2 for the combination COMB_l2. [Figure 16] FIG. 16 is a block diagram showing the configuration of a controlled system CS controlled by a control device CD according to the third embodiment. [Figure 17] FIG. 17 is a block diagram showing the functional configuration of the control device CD according to the third embodiment. [Figure 18] FIG. 18 is a flowchart showing the operation of the control device CD according to the third embodiment. [Figure 19] FIG. 19 shows a number of combinations COMB(1) to COMB(n) that include various values ​​associated with a number of parameters a to l of the function g. [Figure 20] FIG. 20 shows a plurality of combinations COMB(1) to COMB(n) and sample points SP arranged in space. [Figure 21]FIG. 21 shows a plurality of combinations COMB(1) to COMB(n) and unexpected sample points SP. [Figure 22] FIG. 22 shows a plurality of combinations COMB(1) to COMB(n) and some sample points SP1, SP2, and SP3. [Figure 23] FIG. 23 shows a conventional design of preprocessing used in supervised learning models. DETAILED DESCRIPTION OF THE INVENTION

[0014] To explain the present disclosure in more detail, embodiments of the present disclosure will now be described with reference to the accompanying drawings.

[0015] Embodiment 1 Configuration of the first embodiment Hereinafter, the control device according to the first embodiment of the present disclosure will be described with reference to FIGS.

[0016] 1 is a block diagram showing the configuration of a controlled system CS controlled by a control device CD according to embodiment 1. The control device CD controls the configuration, operation, etc. of the controlled system SC.

[0017] As shown in Figure 1, the controlled system CS includes a supervised learning model f and a function g. The configuration of the supervised learning model f has already been completely designed by an engineer E (A) (as shown in Figure 23), and another engineer E (B) wants to train the behavior of the supervised learning model f using training data TD1 to TDn (n is an integer equal to or greater than 2). The function g is an arbitrary function including multiple parameters a to l.

[0018] FIG. 2 shows the configuration of the function g according to the first embodiment. As shown in FIG. 2, the function g, i.e., g(x), is expressed using W, x, and z. As shown in FIG. 2, W is, for example, a matrix for transformation and normalization (for example, a 3-row, 3-column vector), x is an input image (for example, a 3-row, 1-column vector), and z is a vector (for example, a 3-row, 1-column vector). Here, as shown in FIG. 2, W is made up of a plurality of parameters a to i, and z is made up of a plurality of parameters j Since the function g is composed of parameters a to l, we can conclude that the function g, i.e., g(x), is composed of all of these parameters a to l.

[0019] FIG. 3 shows the structure of training data TD1 to TDn. For example, as shown in FIG. 3, training data TD1 has image x1 as a learning example and label y1. Here, label y1 is a correct answer indicating whether image x1, i.e., example x1, is correct when training supervised learning model f. Similar to training data TD1, training data TD2 has image x2 and label y2, training data TD3 has image x3 and label y3, and so on, and training data TDn has image xn and label yn.

[0020] Fig. 4 is a block diagram showing the functional configuration of the control device CD according to embodiment 1. As shown in Fig. 4, the control device CD according to embodiment 1 includes a definition unit 11, a disposer 12, a calculation unit 13, an update unit 14, and a trainer 15. The functions thereof will be described later with reference to the flowcharts of Figs. 7 and 8.

[0021] Fig. 5 is a block diagram showing the configuration of the control device CD according to embodiment 1 from the viewpoint of hardware. As shown in Fig. 5, the control device CD includes an input circuit IC, a processing circuit PC, and an output circuit OC.

[0022] The processing circuit PC is dedicated hardware and realizes the functions of a definition unit 11, a disposer 12, a calculation unit 13, an update unit 14, and a trainer 15 shown in FIG.

[0023] The processing circuit PC may be, for example, a single circuit, a composite circuit, a programmed processor, parallel programmed processors, an ASIC (Application Specific Integrated Circuit), an FPGA (Field Programmable Gate Array), or a combination thereof.

[0024] Fig. 6 is a block diagram showing, from the software perspective, the configuration of the control device CD according to embodiment 1. As shown in Fig. 6, the control device CD includes an input circuit IC, a processor PR, a memory circuit SC, and an output circuit OC.

[0025] The processor PR is, for example, a CPU (including a central processing unit, central processing unit, processing unit, arithmetic unit, microprocessor, microcomputer, and DSP (digital processing)). The processor PR implements the functions from the definition unit 11 to the trainer 15 shown in FIG.

[0026] The processor PR implements the above functions using software, firmware, or a combination of software and firmware, which are written as programs and stored in the storage circuitry SC.

[0027] The processor PR realizes the above functions by reading and executing the program from the storage circuit SC. In other words, the program enables the computer to execute the procedures and methods from the definition unit 11 to the trainer 15.

[0028] Here, the memory circuit SC is, for example, a volatile or non-volatile semiconductor memory such as RAM (random access memory), ROM (read only memory), flash memory, EPROM (erasable programmable read only memory), EEPROM (electrically erasable programmable read only memory), a magnetic disk, a flexible disk, an optical disk, a compact disk, a DVD (digital versatile disk), etc.

[0029] Of the functions from the definition unit 11 to the trainer 15, some functions may be implemented by the processing circuit PC, and other functions may be implemented by the processor PR.

[0030] As mentioned above, the functions of the control device CD from the definition unit 11 to the trainer 15 can be implemented using hardware, software, firmware, or a combination thereof.

[0031] Operation of the First Embodiment FIG. 7 is a flowchart showing the operation (part 1) of the control device CD according to the first embodiment.

[0032] FIG. 8 is a flowchart showing the operation (part 2) of the control device CD according to the first embodiment.

[0033] For ease of explanation and understanding, assume that the integer i can count up to 100, and that the difference Li converges when the integer i reaches 80.

[0034] In step ST11, the definition unit 11 (shown in FIG. 4) defines a plurality of parameters a to l (shown in FIG. 2 ) to define a function g (shown in FIG. 1). Furthermore, definition unit 11 initializes a plurality of parameters a to l. After initialization by definition unit 11, disposer 12 (shown in FIG. 4) places or sets function g before supervised learning model f (shown in FIG. 1), as shown in FIG.

[0035] In step ST12, the definition unit 11 defines a function f(g(x)) using a supervised learning model f and a function g, where the function f(g(x)) is formed in a nested format, i.e., the function g provides an output depending on an input image x (shown in FIG. 2), and the supervised learning model f provides an output depending on the output provided by the function g.

[0036] In step ST13, the definition unit 11 initializes the integer i, that is, sets 1 to the integer i.

[0037] In step ST14, the calculation unit 13 (shown in FIG. 4) calculates the result Zi of the function f(g(xi)).

[0038] In step ST15, the calculation unit 13 calculates the difference Li between the result Zi and the label yi (shown in FIG. 3) by subtracting the label yi from the result Zi.

[0039] In step ST16, the calculation unit 13 calculates the partial derivative PDi of the difference Li with respect to the multiple parameters a to l of the function g.

[0040] In step ST17, the update unit 14 (see FIG. 4) updates the parameters a to l of the function g in accordance with the partial differential PDi.

[0041] As shown in Fig. 7, for example, when integer i is 1, the multiple parameters a to l of function g are updated to a1 to l1, respectively. Similarly, as shown in Fig. 2, when integer i is 2, the multiple parameters a to l of function g are updated to a2 to l2, respectively, when integer i is 3, the multiple parameters a to l of function g are updated to a3 to l3, respectively, and when integer i is n, the multiple parameters a to l of function g are updated to an to ln (not shown), respectively.

[0042] In step ST18, the process from step ST14 to step ST 17 As a result of repeating the above, if it is determined that the difference Li has converged, for example, by comparing it with a predetermined threshold value, step ST 19 On the other hand, if it is determined that the difference Li has not converged, for example, by comparing it with a predetermined value, the process proceeds to step ST 13 Proceed to.

[0043] FIG. 9 shows the configuration of the control device CD according to the first embodiment when i reaches 80.

[0044] Step ST 19In the above, the definition unit 11 defines the pseudo preprocessing h using a plurality of parameters a to l when i reaches 80, that is, a plurality of parameters a80 to l80 as shown in Fig. 9. Furthermore, the disposer 12 replaces the function g with the pseudo preprocessing h, more specifically, as shown in Fig. 9, deletes the function g and places or installs the pseudo preprocessing h before the supervised learning model f.

[0045] Step ST 20 Now, trainer 15 (shown in FIG. 4) trains supervised learning model f using pseudo preprocessing h, more specifically, by providing training data TD1 to TDn to pseudo preprocessing h placed before supervised learning model f, as shown in FIG. 9.

[0046] Steps following a NO return in step ST18 13 In , the integer i is incremented by 1.

[0047] Steps 2 Then, based on the above assumptions, the integer i is compared with 100. If the integer i is smaller than 100, the procedure returns to step ST14; otherwise, the procedure ends.

[0048] Effects of the First Embodiment As described above, the control device CD of the first embodiment defines the pseudo preprocessing h using a plurality of parameters a to l, more specifically, a plurality of parameters a80 to l80 that are given when the difference Li converges, that is, when the integer i reaches 80. This makes it possible to train the supervised learning model f using the pseudo preprocessing h instead of the function g, even if the plurality of parameters a to l of the function g are unknown or cannot be reproduced.

[0049] Embodiment 2 Configuration of the second embodiment Next, a control device according to a second embodiment of the present disclosure will be described with reference to FIGS.

[0050] 10 is a block diagram showing the configurations of a first controlled system CS1 and a second controlled system CS2 controlled by a control device CD according to embodiment 2. The control device CD according to embodiment 2 controls, for example, the configurations and operations of both the first controlled system CS1 and the second controlled system CS2, similar to the control device CD according to embodiment 1.

[0051] As shown in FIG. 10, the first controlled system CS1 includes a first supervised learning model f1 and a first function g1, and the second controlled system CS2 includes a second supervised learning model f2 and a second function g2. The operation of the first supervised learning model f1 and the operation of the second supervised learning model f2 are equivalent to each other. In the first controlled system CS1, the first function g1 is placed or installed before the first supervised learning model f1. Similarly, in the second controlled system CS2, the second function g2 is placed or installed before the second supervised learning model f2. In the second embodiment, an engineer E (B) wants to use training data TD1 to TDn to evaluate which of the first function g1 and the second function g2 is more robust. The evaluation of the robustness of the first function g1 and the second function g2 will be described in detail later.

[0052] Fig. 11 is a block diagram showing the functional configuration of the control device CD according to embodiment 2. As shown in Fig. 11, the control device CD according to embodiment 2 includes an acquisition unit 21, a calculation unit 22, and an evaluation unit 23. The functions thereof will be described later with reference to the flowchart of Fig. 12.

[0053] The control device CD of the second embodiment includes an input circuit IC, a processing circuit PC, and an output circuit OC (shown in FIG. 5), similar to the control device CD of the first embodiment. Alternatively, the control device CD according to the second embodiment may include an input circuit IC, a processor PR, a memory circuit SC, and an output circuit OC (shown in FIG. 6).

[0054] Instead of the functions of the definition unit 11, disposer 12, calculation unit 13, update unit 14, and trainer 15 of embodiment 1 (see Figure 4), a processing circuit PC (see Figure 5) or a processor PR (see Figure 6) realizes the functions of the acquisition unit 21, calculation unit 22, and evaluation unit 23 (see Figure 11).

[0055] Operation of the second embodiment FIG. 12 is a flowchart showing the operation of the control device CD according to the second embodiment.

[0056] FIG. 13 shows a plurality of combinations COMB_a1 to COMB_l1 including various values ​​associated with a plurality of parameters a to l of the first function g1.

[0057] In step ST31, the acquisition unit 21 (shown in FIG. 11) acquires a plurality of combinations COMB_a1 to COMB_l1 shown in FIG. 13. Each of the plurality of combinations COMB_a1 to COMB_l1 includes various values ​​related to one of the plurality of parameters a to l of the first function g1 (shown in FIG. 10).

[0058] For example, as shown in Figure 13, the combination COMB_a1 includes various values ​​of a1(1), a1(2), a1(3), ..., a1(n) for the parameter a of the first function g1, where the values ​​a1(1), a1(2), a1(3), ..., a1(n) are the outputs returned or fed back by the first supervised learning model f1 in response to the training data TD1, TD2, TD3, ..., TD(n).

[0059] For example, the value a1(1) is the output returned by the first supervised learning model f1 in response to the training data TD1, and the value a1(2) is the output returned by the first supervised learning model f1 in response to the training data TD2. 、 The value a1(n) is the output returned by the first supervised learning model f1 in response to the training data TDn.

[0060] FIG. 14 shows a plurality of combinations COMB_a2 to COMB_l2 including various values ​​of a plurality of parameters a to l of the second function g2.

[0061] In step ST32, the acquiring section 21 acquires a plurality of combinations COMB_a2 to COMB_l2. Each of the plurality of combinations COMB_a2 to COMB_l2 includes various values ​​for one of the plurality of parameters a to l of the second function g2.

[0062] For example, as shown in Figure 14, the combination COMB_a2 includes various values ​​a2(1), a2(2), a2(3), ..., a2(n) for the parameter a of the second function g2, where the values ​​a2(1), a2(2), a2(3), ..., a2(n) are the outputs returned or fed back by the second supervised learning model f2 in response to the training data TD1, TD2, TD3, ..., TD(n).

[0063] For example, the value a2(1) is the output returned by the second supervised learning model f2 in response to the training data TD1, and the value a2(2) is the output returned by the second supervised learning model f2 in response to the training data TD2. 、 The value a2(n) is the output returned by the second supervised learning model f2 in response to the training data TDn.

[0064] Figure 15 is 、 A plurality of distributions are shown, from a distribution D_a1 for the combination COMB_a1 to a distribution D_l1 for the combination COMB_l1, and a plurality of distributions are shown, from a distribution D_a2 for the combination COMB_a2 to a distribution D_l2 for the combination COMB_l2.

[0065] In step ST33, the calculation unit 22 (shown in FIG. 11) calculates a plurality of distributions D_l1 (shown in FIG. 15) of the combination COMB_l1 (shown in FIG. 13) from a plurality of distributions D_a1 (shown in FIG. 15) of the combination COMB_a1 (shown in FIG. 13). More specifically, the calculation unit 22 calculates the distribution D_a1 of the combination COMB_a1, which includes, for example, various values ​​a1(1), a1(2), a1(3), . . . , a1(n) (shown in FIG. 13) associated with the parameter a of the first function g1 (shown in FIG. 10). As shown in FIG. 15, the distribution D_a1 of the combination COMB_a1 has a breadth or spread specific to the distribution D_a1.

[0066] The calculation unit 22 calculates the other distributions D_b1 to D_l1 in the same way as when calculating the distribution D_a1. Each of the distributions D_b1 to D_l1 has a specific extent or spread.

[0067] In step ST34, the calculation unit 22 calculates a plurality of distributions D_l2 (shown in FIG. 15) of the combination COMB_l2 (shown in FIG. 14) from a plurality of distributions D_a2 (shown in FIG. 15) of the combination COMB_a2 (shown in FIG. 14). More specifically, the calculation unit 22 calculates the distribution D_a2 of the combination COMB_a2, which includes, for example, various values ​​a2(1), a2(2), a2(3), . . . , a2(n) (shown in FIG. 14) associated with the parameter a of the second function g2 (shown in FIG. 10). As shown in FIG. 15, the distribution D_a2 of the combination COMB_a2 has a spread that is specific to the distribution D_a2.

[0068] In the same way as calculating the distribution D_a2, the calculation unit 22 calculates the other distributions D_b2 to D_l2. Each of the distributions D_b2 to D_l2 has a specific extent or spread.

[0069] In step ST35, the evaluation unit 23 (shown in FIG. 11) evaluates which of the first function g1 (shown in FIG. 10) and the second function g2 (shown in FIG. 10) is more robust by comparing the spread of the multiple distributions D_a1 to D_l1 (shown in FIG. 15) with the spread of the multiple distributions D_a2 to D_l2 (shown in FIG. 15).

[0070] For example, the evaluation unit 23 compares the spread of the distribution D_a1 with the spread of the distribution D_a2. As shown in Fig. 15, the distribution D_a2 is more widely spread than the distribution D_a1, and therefore the evaluation unit 23 evaluates that the second function g2 is more robust than the first function g1 from the viewpoint of the distributions D_a1 and D_a2.

[0071] Similar to the spread of distributions D_a1 and D_a2, the evaluation unit 23 compares the spread of distributions D_b1 and D_b2, D_c1 and D_c2, . . . , and D_l1 and D_l2.

[0072] The evaluation unit 23 evaluates which of the first function g1 and the second function g2 is more robust, taking into consideration all the results of comparing the distributions D_a1 and D_a2, the distributions D_b1 and D_b2, the distributions D_c1 and D_c3, ..., and the distributions D_l1 and D_l2.

[0073] Effects of the second embodiment As described above, the control device CD of embodiment 2 can evaluate which of the first function g1 and the second function g2 is more robust by comparing the extent or spread of the distributions D_a1 to D_l1 with the extent or spread of the distributions D_a2 to D_l2, respectively.

[0074] Embodiment 3 Configuration of the third embodiment Next, a control device according to a third embodiment of the present disclosure will be described with reference to FIGS.

[0075] 16 is a block diagram showing the configuration of a control target system CS controlled by a control device CD according to embodiment 3. The control device CD according to embodiment 3 controls, for example, both the configuration and operation of the control target system CS, similar to the control device CD according to embodiment 1.

[0076] As shown in Fig. 16, a controlled system CS includes a supervised learning model f and a function g, and the function g is placed or installed before the supervised learning model f. In the third embodiment, an engineer E (B) wants to assign various values ​​to multiple parameters a to l of the function g using training data TD1 to TDn. The above assignments will be explained in detail later.

[0077] Fig. 17 is a block diagram showing the functional configuration of the control device CD according to embodiment 3. As shown in Fig. 17, the control device CD according to embodiment 3 includes an acquisition unit 31, a disposer 32, a selection unit 33, and an allocation unit 34. The functions thereof will be described later with reference to the flowchart of Fig. 18.

[0078] The control device CD includes an input circuit IC, a processing circuit PC, and an output circuit OC (shown in FIG. 5), similar to the control device CD of embodiment 1. Alternatively, the control device CD according to embodiment 3 may include an input circuit IC, a processor PR, a memory circuit SC, and an output circuit OC (shown in FIG. 6).

[0079] Instead of the functions of the definition unit 11, disposer 12, calculation unit 13, update unit 14, and trainer 15 of embodiment 1 (see Figure 4), a processing circuit PC (see Figure 5) or a processor PR (see Figure 6) realizes the functions of an acquisition unit 31, disposer 32, selection unit 33, and allocation unit 34 (see Figure 17).

[0080] Operation of the Third Embodiment FIG. 18 is a flowchart showing the operation of the control device CD according to the third embodiment.

[0081] FIG. 19 shows a plurality of combinations COM that include various values ​​related to the parameters a to l of the function g. B (1)~COM B (n) is shown.

[0082] In step ST41, the acquisition unit 31 (shown in FIG. 17) acquires a plurality of combinations COMB(1) to COMB(n) (shown in FIG. 19). Each of the plurality of combinations COMB(1) to COMB(n) includes various values ​​related to a plurality of parameters a to l of the function g.

[0083] For example, as shown in FIG. 19, combination COMB(1) includes various values ​​a(1), b(1), c(1), . . . , l(1) associated with multiple parameters a through l of function g, where the values ​​a(1), b(1), c(1), . . . , l(1) are outputs returned or fed back by supervised learning model f in response to training data TD1, the values ​​a(2), b(2), c(2), . . . , l(2) are outputs returned by supervised learning model f in response to training data TD2, and the values ​​a(n), b(n), c(n), . . . , l(n) are outputs returned by supervised learning model f in response to training data TDn.

[0084] FIG. 20 shows a plurality of combinations COMB(1) to COMB(n) and sample points SP arranged in space.

[0085] In step ST42, the disposer 3 2 (shown in FIG. 17) discards or arranges a plurality of combinations COMB(1) to COMB(n) in a 12-dimensional space, the number of which is the same as the number of parameters a to l of the function g, as shown in FIG.

[0086] In step ST43, the acquisition unit 3 1 acquires sample points SP using a known method or system, for example as shown in FIG.

[0087] In step ST44, the selection unit 33 (shown in FIG. 17) selects the combination COMB(3) that is closest to the sample point SP from among the plurality of combinations COMB(1) to COMB(n) shown in FIG.

[0088] In step ST45, the allocation unit 34 (shown in FIG. 17) allocates the selected combination COMB(3), more specifically, the various values ​​a(3), b(3), c(3), ..., l(3) (shown in FIG. 19) included in the selected combination COMB(3), to the multiple parameters a to l of the function g.

[0089] Effects of the Third Embodiment As described above, the control device CD of the third embodiment selects the combination COMB(3) that is located closest to the sample point SP from among the multiple combinations COMB(1) to COMB(n), and more specifically, selects the combination COMB(3) that can assign various values ​​a(3), b(3), c(3), ..., l(3) that may be appropriate to the multiple parameters a to l of the function g.

[0090] Modification of the third embodiment FIG. 21 shows a plurality of combinations COMB(1) to COMB(n) and unexpected sample points SP.

[0091] The above well-known schemes or methods used for the sample points SP may give unexpected sample points USP (shown in FIG. 21) that are located farther or furthest from some combination groups, e.g., COMB(3), COMB(6), COMB(2), and COMB(1), instead of the sample points SP (shown in FIG. 20) expected for the above groups. If only unexpected sample points USP are available for selection in the above step ST44, the assignment in step ST45 may not give suitable values ​​for the parameters a to l of the function g.

[0092] FIG. 22 shows a plurality of combinations COMB(1) to COMB(n) and sample points SP1, SP2, and SP3.

[0093] To improve the accuracy of both the selection and assignment, in a modification of the third embodiment, as shown in FIG. 22, the acquisition unit 31 (shown in FIG. 17) acquires multiple sample points, such as three sample points SP1, SP2, and SP3, instead of the single sample point SP of the third embodiment (shown in FIG. 20). Even if sample point SP2 is unexpected, at least one of the other sample points SP1 and SP3 may be predictable. This allows the selection unit 33 to select the combination COMB(3) located or placed closest to sample point SP1 or the combination COMB(1) located or placed closest to sample point SP3. This allows various values ​​contained in the combination COMB(3) or COMB(1) that may be appropriate for the multiple parameters a to l of the function g to be assigned, as in the third embodiment.

Claims

1. A method carried out by a control device comprising a first definition unit, a first disposer, a second definition unit, a first calculation unit, a second calculation unit, a third calculation unit, an update unit, a third definition unit, a second disposer, and a trainer, The first definition unit defines a function g having a plurality of parameters a to l; the first disposer placing the defined function g in front of a supervised learning model f, the supervised learning model f being trained with training data TDi including examples x and labels yi, where i is an integer; The second definition unit defines a function f(g(xi)) using the trained supervised learning model f and the defined function g; The first calculation unit calculates the defined function f(g(xi)) to calculate a result Zi; The second calculation unit calculates a difference Li between the result Zi and the label yi; The third calculation unit calculates partial derivatives PDi of the calculated difference Li with respect to a plurality of parameters a to l of the defined function g; the updating unit updates the plurality of parameters a to l of the defined function g according to the calculated partial differential PDi; the third calculation unit defines a pseudo pre-processing h using the plurality of parameters a to l of the defined function g when the calculated difference Li converges during the iteration between the calculation of the result Zi and the update of the plurality of parameters a to l of the defined function g; The second disposer places the defined pseudo-pretreatment h before the trained supervised learning model f; the trainer providing the training data TDi to the defined pseudo-preprocessing h to train the supervised learning model f; A method for providing the above.

2. a first definition part that defines a function g having a plurality of parameters a to l; a first disposer that places a function g in front of a supervised learning model f, where the supervised learning model f is trained with training data TDi that includes examples x and labels yi, where i is an integer; a second definition section that defines a function f(g(xi)) using a supervised learning model f and a function g; a first calculation unit for calculating a function f(g(xi)) to calculate a result Zi; a second calculation unit for calculating the difference Li between the result Zi and the label yi; a third calculation unit that calculates partial derivatives PDi of the difference Li with respect to a plurality of parameters a to l of the function g; an updating unit that updates a plurality of parameters a to l of the function g according to the partial differential PDi; a third definition part for defining a pseudo-preprocessing h using the parameters a to l of the function g when the difference L converges during the iteration between the calculation of the result Zi and the update of the parameters a to l of the function g; a second disposer that places a pseudo-preprocessing h before the supervised learning model f; a trainer that provides training data TDi to a pseudo-preprocessing h to train a supervised learning model f.

Images