Method and system for configuring pre-processing for supervised learning model

Abstract

An exemplary method includes defining a function g having a plurality of parameters a to l; setting the function g in front of the supervised learning model f; training a supervised learning model f by using training data TDi comprising an example xi and a tag yi, wherein i is an integer; defining a function f (g (xi)) by using the supervised learning model f and the function g; calculating a result Zi by calculating a function f (g (xi)); calculating the difference Li between the result Zi and the tag yi; calculating a partial differential PDi of the difference Li with respect to a plurality of parameters a to l of the function g; updating a plurality of parameters a to l of the function g according to the partial differential PDi; defining a pseudo pre-process h by using the plurality of parameters a to l of the function g when a process difference Li repeating between the calculation of the result Zi and the update of the plurality of parameters a to l of the function g converges; setting the pseudo preprocessing h before the supervised learning model f; and training the supervised learning model f by providing the training data TDi to the pseudo pre-process h.

Description

Technical Field

[0001] The present disclosure relates to methods and systems for configuring preprocessing for supervised learning models, and more particularly, to methods and systems for evaluating multiple parameters used in preprocessing. Background Art

[0002] The object detection device taught in Japanese Patent Publication No. 2019-219804 detects objects in an image by preprocessing an image captured by a camera. In order to improve the accuracy of detecting objects, the object detection device adjusts multiple parameters included in the preprocessing before performing the preprocessing.

[0003] Fig.23 The general design of preprocessing for supervised learning models is shown in Fig.23 , the preprocessing PP1 designed by the engineer E (A) for the supervised learning model SLM (for 32-bit processing) by using the training data TD1 to TDn (n is an integer equal to or greater than 2) is shown. Here, it is assumed that the preprocessing PP1 includes a plurality of parameters that are, for example, irreproducible or unknown to the engineer E (B) and other engineers, and the engineer E (B) attempts to divert or transfer the supervised learning model SLM from 32-bit processing to 64-bit processing.

[0004] However, since the plurality of parameters are not reproducible or unknown, the engineer E (B) cannot reproduce the plurality of parameters used for the preprocessing PP1, and thus configures another preprocessing PP2 with a plurality of any other inappropriate parameters different from the plurality of parameters used in the preprocessing PP1. As a result, the engineer E (B) cannot convert or transfer the supervised learning model SLM from 32-bit processing to 64-bit processing.

[0005] Reference List

[0006] Patent Literature

[0007] [PTL1]:JP2019-219804A Summary of the invention

[0008] In order to solve the above problems, one aspect of the present disclosure provides a method, which includes the following steps: defining a function g having multiple parameters a to l; setting the function g before a supervised learning model f; training the supervised learning model f using training data TDi including examples xi and labels yi, where i is an integer; defining a function f(g(xi)) by using the supervised learning model f and the function g; calculating a result Zi by calculating the function f(g(xi)); calculating the difference Li between the result Zi and the label yi; calculating the partial differential PDi of the difference Li with respect to multiple parameters a to l of the function g; updating multiple parameters a to l of the function g according to the partial differential PDi; when the difference Li converges in the process of repeating between the step of calculating the result Zi and the step of updating the multiple parameters a to l of the function g, defining a pseudo preprocessing h by using the multiple parameters a to l of the function g; setting the pseudo preprocessing h before the supervised learning model f; and training the supervised learning model f by providing the training data TDi to the pseudo preprocessing h.

[0009] Another aspect of the present disclosure provides a method, the method comprising the following steps: obtaining a first plurality of combinations COMB_a1 to COMB_l1 by providing training data TDi to a first supervised learning model f1, wherein the first plurality of combinations COMB_a1 to COMB_l1 each include various values ​​of one of a plurality of parameters a to l regarding a first function g1, wherein the first function g1 is set before the first supervised learning model f1 implements feedback on the first function g1, and i is an integer; obtaining a second plurality of combinations COMB_a2 to COMB_l2 by providing training data TDi to a second supervised learning model f2, wherein the second plurality of combinations COMB_a2 to COMB_l2 each include Regarding various values ​​of one of the multiple parameters a to l of the second function g2, the second function g2 is set before the second supervised learning model f2 that implements feedback on the 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; calculating the first multiple distributions D_a1 to D_l1 respectively about the first multiple combinations COMB_a1 to COMB_l1; calculating the second multiple distributions D_a2 to D_l2 respectively about the second multiple 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 multiple distributions D_a1 to D_l1 and the second multiple distributions D_a2 to D_l2 respectively.

[0010] Another aspect of the present disclosure provides a method, which includes the following steps: obtaining multiple combinations COMB(1) to COMB(n) by providing training data TDi to a supervised learning model f, each of the multiple combinations COMB(1) to COMB(n) including multiple values ​​of multiple parameters a to l regarding a function g, wherein the function g is set before the supervised learning model f implements feedback on the function g, and n is an integer equal to or greater than 2; setting the multiple combinations COMB(1) to COMB(n) in a space having multiple dimensions, the number of the multiple dimensions being equal to the number of the multiple parameters a to l of the function g; obtaining at least one sampling point SP in the space; selecting at least one combination of the multiple combinations COMB(1) to COMB(n) that is closest to the at least one sampling point SP; and assigning the multiple values ​​included in the selected one of the multiple combinations COMB(1) to COMB(n) to the multiple parameters a to l of the function g.

[0011] Another aspect of the present disclosure provides an apparatus comprising: a first definer, which defines a function g having multiple parameters a to l; a first setter, which sets the function g before a supervised learning model f; training the supervised learning model f using training data TDi including examples xi and labels yi, where i is an integer; a second definer, which defines a function f(g(xi)) using the supervised learning model f and the function g; a first calculator, which calculates a result Zi by calculating the function f(g(xi)); and a second calculator, which calculates the difference between the result Zi and the label yi. Li; a third calculator, which calculates the partial differential PDi of the difference Li with respect to multiple parameters a to l of the function g; an updater, which updates the multiple parameters a to l of the function g according to the partial differential PDi; a third definer, which defines a pseudo-preprocessing h by using multiple parameters a to l of the function g when the difference Li converges in a repeated process between the calculation of the result Zi and the updating of the multiple parameters a to l of the function g; a second setter, which sets the pseudo-preprocessing h before the supervised learning model f; and a trainer, which trains the supervised learning model f by providing training data TDi to the pseudo-preprocessing h.

[0012] Another aspect of the present disclosure provides an apparatus, comprising: a first acquirer, the first acquirer acquiring a first plurality of combinations COMB_a1 to COMB_l1 by providing training data TDi to a first supervised learning model f1, the first plurality of combinations COMB_a1 to COMB_l1 each including various values ​​of one of a plurality of parameters a to l regarding a first function g1, the first function g1 being set before the first supervised learning model f1 implementing feedback on the first function g1, i being an integer; a second acquirer, the second acquirer acquiring a second plurality of combinations COMB_a2 to COMB_l2 by providing training data TDi to a second supervised learning model f2, the second plurality of combinations COMB_a2 to COMB_l2 each including various values ​​of one of a plurality of parameters a to l regarding a second function g1 2, the second function g2 is set before the second supervised learning model f2 that implements feedback on the 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; a first calculator, the first calculator calculates the first multiple distributions D_a1 to D_l1 respectively about the first multiple combinations COMB_a1 to COMB_l1; a second calculator, the second calculator calculates the second multiple distributions D_a2 to D_l2 respectively about the second multiple combinations COMB_a2 to COMB_l2; and an evaluator, the evaluator evaluates which of the first function g1 and the second function g2 is more robust by comparing the first multiple distributions D_a1 to D_l1 and the second multiple distributions D_a2 to D_l2 respectively.

[0013] Another aspect of the present disclosure provides an apparatus, comprising: a first acquirer, which acquires a plurality of combinations COMB(1) to COMB(n) by providing training data TDi to a supervised learning model f, wherein the plurality of combinations COMB(1) to COMB(n) each include a plurality of values ​​of a plurality of parameters a to l regarding a function g, wherein the function g is set before the supervised learning model f implements feedback on the function g, and n is an integer equal to or greater than 2; a setter, which sets the plurality of combinations COMB(1) to COMB(n) in a space having a plurality of dimensions, wherein the number of the plurality of dimensions is equal to the number of the plurality of parameters a to l of the function g; a second acquirer, which acquires at least one sampling point SP in the space; a selector, which selects at least one combination of the plurality of combinations COMB(1) to COMB(n) that is closest to the at least one sampling point SP; and an allocator, which allocates a plurality of values ​​included in a selected one of the plurality of combinations COMB(1) to COMB(n) to the plurality of parameters a to l of the function g. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 is a block diagram showing the configuration of a controlled system CS controlled by a control device CD according to the first embodiment.

[0015] Figure 2 The configuration of the function g according to the first embodiment is shown.

[0016] Figure 3 The configuration of the training data TD1 to TDn is shown.

[0017] Figure 4 : is a block diagram showing the configuration of the control device CD according to the first embodiment from a functional viewpoint.

[0018] Figure 5 : is a block diagram showing the configuration of the control device CD according to the first embodiment from a hardware viewpoint.

[0019] Figure 6 : is a block diagram showing the configuration of the control device CD according to the first embodiment from a software viewpoint.

[0020] Figure 7 is a flowchart showing the operation (Part 1) of the control device CD according to the first embodiment.

[0021] Figure 8 is a flowchart showing the operation (Part 2) of the control device CD according to the first embodiment.

[0022] Fig. 9 The configuration of the control device CD according to the first embodiment when i reaches 80 is shown.

[0023] 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.

[0024] Fig.11 : is a block diagram showing the configuration of the control device CD according to the second embodiment from a functional viewpoint.

[0025] Fig.12 : is a flowchart showing the operation of the control device CD according to the second embodiment.

[0026] Fig.13 A plurality of combinations COMB_a1 to COMB_l1 including various values ​​of a plurality of parameters a to l with respect to the first function g1 are shown.

[0027] Fig.14 A plurality of combinations COMB_a2 to COMB_12 including various values ​​of a plurality of parameters a to l with respect to the second function g2 are shown.

[0028] Fig.15A plurality of distributions from distribution D_a1 of combination COMB_a1 to distribution D_l1 of combination COMB_l1 and a plurality of distributions from distribution D_a2 of combination COMB_a2 to distribution D_l2 of combination COMB_l2 are shown.

[0029] 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.

[0030] Fig.17 : is a block diagram showing the configuration of the control device CD according to the third embodiment from a functional viewpoint.

[0031] Fig.18 : is a flowchart showing the operation of the control device CD according to the third embodiment.

[0032] Fig.19 A plurality of combinations COM(1) to COM(n) including various values ​​associated with a plurality of parameters a to l of the function g are shown.

[0033] Fig. 20 A plurality of combinations COMB(1) to COMB(n) and sampling points SP arranged in space are shown.

[0034] Fig.21 A plurality of combinations COMB(1) to COMB(n) and unexpected sampling points SP are shown.

[0035] Fig. 22 A plurality of combinations COMB(1) to COMB(n) and a number of sampling points SP1, SP2 and SP3 are shown.

[0036] Fig.23 Shows a general design for preprocessing of supervised learning models. DETAILED DESCRIPTION

[0037] In order to further describe the present disclosure in detail, embodiments of the present disclosure will be described below with reference to the accompanying drawings.

[0038] First embodiment

[0039] Configuration of the First Embodiment

[0040] Now refer to Figures 1 to 9 A control device according to a first embodiment of the present disclosure is described.

[0041] Figure 1 1 is a block diagram showing the configuration of a controlled system CS controlled by a control device CD according to the first embodiment. The control device CD controls, for example, the configuration and operation of the controlled system SC.

[0042] like Figure 1As shown, the controlled system CS includes a supervised learning model f and a function g. The configuration of the supervised learning model f has been designed by engineer E(A) (e.g. Fig.23 As shown), another engineer E (B) desires to train the operation of the supervised learning model f by using training data TD1 to TDn (n is an integer equal to or greater than 2). Function g is an arbitrary function including a plurality of parameters a to l.

[0043] Figure 2 The configuration of the function g of the first embodiment is shown. Figure 2 As shown in , the function g is represented by using W, x and z, i.e. g(x). Figure 2 As shown, W is a matrix (e.g., a vector of 3 rows × 3 columns) for conversion and normalization, x is an input image (e.g., a vector of 3 rows × 1 column), and z is a vector (e.g., a vector of 3 rows × 1 column). Figure 2 As shown, W is composed of multiple parameters a to i, and z is composed of multiple parameters i to l, so it can be concluded that the function g (ie g(x)) is composed of all these parameters a to l.

[0044] Figure 3 TD1 to TDn are shown. Figure 3 As shown, for example, training data TD1 has an image x1 and a label y1 used for training as an example. Here, label y1 is the correct answer, which shows whether image x1 (ie, 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, ..., training data TDn has image xn and label yn.

[0045] Figure 4 2 is a block diagram showing the configuration of the control device CD according to the first embodiment from a functional viewpoint. Figure 4 As shown, the control device CD according to the first embodiment includes a definer 11, a setter 12, a calculator 13, an updater 14 and a trainer 15. The functions thereof will be described later with reference to Figures 7 and 8 The flowchart is described.

[0046] Figure 5 2 is a block diagram showing the configuration of the control device CD according to the first embodiment from a hardware viewpoint. Figure 5 As shown, the control device CD comprises an input circuit IC, a processing circuit PC and an output circuit OC.

[0047] The processing circuit PC is a dedicated hardware. The processing circuit PC implements Figure 4 The functions of the definer 11, setter 12, calculator 13, updater 14 and trainer 15 are shown.

[0048] The processing circuit PC is for example a single circuit, a complex circuit, a programmed processor, a parallel programmed processor, an ASIC (Application Specific Integrated Circuit), an FPGA (Field Programmable Gate Array) or a combination thereof.

[0049] Figure 6 2 is a block diagram showing the configuration of the control device CD according to the first embodiment from a software viewpoint. Figure 6 As shown, the control device CD includes an input circuit IC, a processor PR, a storage circuit SC and an output circuit OC.

[0050] The processor PR is, for example, a CPU (including a central processing unit, a central processing device, a processing device, an arithmetic device, a microprocessor, a microcomputer or a DSP (digital processing). The processor PR implements Figure 4 The functions of the definer 11 to the trainer 15 are shown.

[0051] The processor PR implements the above functions by using software, firmware, or a combination of software and firmware. Software and firmware are described as programs, which are stored in the storage circuit SC.

[0052] The processor PR realizes the above functions by reading out the program from the storage circuit SC and executing the program. In other words, the above program enables the computer to execute the processes and methods of the definer 11 to the trainer 15.

[0053] Here, the storage circuit SC is, for example, a volatile or nonvolatile semiconductor memory, which includes RAM (random access memory), ROM (read only memory), flash memory, EPROM (erasable programmable read-only memory) and EEPROM (electrically erasable programmable read-only memory); a magnetic disk; a floppy disk; an optical disk; a compact disk; or a DVD (digital versatile disk).

[0054] Among the functions of the definer 11 to the trainer 15, some functions may be implemented by the processing circuit PC, while other functions may be implemented by the processor PR.

[0055] As described above, the functions of the definer 11 to the trainer 15 in the control device CD can be implemented by using hardware, software, firmware, or a combination thereof.

[0056] Operation of the First Embodiment

[0057] Figure 7 is a flowchart showing the operation (Part 1) of the control device CD according to the first embodiment.

[0058] Figure 8 is a flowchart showing the operation (Part 2) of the control device CD according to the first embodiment.

[0059] For ease of explanation and understanding, assume that the integer i can be counted until 100, and when the integer i reaches 80, the difference Li converges.

[0060] In step ST11, definer 11 ( Figure 4 ) using multiple parameters a to l ( Figure 3 As shown) to define the function g( Figure 1 In addition, the definer 11 initializes multiple parameters a to l. After initialization by the definer 11, the setter 12 ( Figure 4 As shown) configure or set the function g in the supervised learning model f( Figure 1 shown) before.

[0061] In step ST12, the definer 11 defines a function f(g(x)) by using the supervised learning model f and the function g. Here, the function f(g(x)) is formed in a nested manner, that is, the function g is based on the input image x( Figure 2 As shown), the supervised learning model f provides output based on the output provided from the function g.

[0062] In step ST13, the definer 11 initializes the integer i, that is, sets 1 to the integer i.

[0063] In step ST14, the calculator 13 (such as Figure 4 The result Zi of calculating the function f(g(xi)) is shown in (Fig.

[0064] In step ST15, the calculator 13 calculates the result Zi and the label yi by subtracting the label yi from the result Zi ( Figure 3 The difference between

[0065] In step ST16 , the calculator 13 calculates partial differentials PDi of the difference Li with respect to a plurality of parameters a to l of the function g.

[0066] In step ST17, the updater 14 (such as Figure 4 As shown) multiple parameters a to l of the function g are updated according to the partial differential PDi.

[0067] like Figure 7 As shown, for example, when the integer i is 1, multiple parameters a to l of the function g are updated to a1 to l1 respectively. Similarly, Figure 2 As shown, when the integer i is 2, multiple parameters a to l of function g are updated to a2 to l2 respectively, when the integer i is 3, multiple parameters a to l of function g are updated to a3 to l3 respectively, and when the integer i is n, multiple parameters a to l of function g are updated to an to ln respectively (not shown).

[0068] In step ST18, as a result of repeating steps ST14 to ST20, when it is identified that the difference Li has converged compared to a predetermined threshold, for example, the process proceeds to step ST21. When it is identified that the difference Li has not converged compared to a predetermined threshold, for example, the process proceeds to step ST19.

[0069] Fig. 9 The configuration of the control device CD according to the first embodiment when i reaches 80 is shown.

[0070] In step ST21, the definer 11 defines the pseudo preprocessing h by using a plurality of parameters a to l when i reaches 80 (ie, a plurality of parameters a80 to l80), such as Fig. 9 In addition, the setter 12 replaces the function g with the pseudo preprocessing h, and more specifically, as Fig. 9 As shown, the function g is removed and the pseudo preprocessing h is set or placed before the supervised learning model f.

[0071] In step ST22, the trainer 15 ( Figure 4 As shown in FIG. 1 , the supervised learning model f is trained by using the pseudo preprocessing h, and more specifically, by providing the training data TD1 to TDn to the supervised learning model f. Fig. 9 The pseudo preprocessing h placed before the supervised learning model f is shown to train the supervised learning model f.

[0072] In step 19 following step ST18 in which "No" is given, the integer i is incremented by only 1.

[0073] In step 20, the integer i is compared with 100 according to the above assumption. If the integer i is less than 100, the process returns to step ST14; otherwise, the process ends.

[0074] Effects of the First Embodiment

[0075] As described above, the control device CD according to the first embodiment defines the pseudo preprocessing h by using a plurality of parameters a to l, and more specifically, by using a plurality of parameters a80 to l80 given when the difference Li converges, that is, when the integer i reaches 80. Therefore, even if the plurality of parameters a to l of the function g are unknown or unreproducible, the supervised learning model f can be trained by using the pseudo preprocessing h instead of the function g.

[0076] Second embodiment

[0077] Configuration of the Second Embodiment

[0078] Now refer to Figures 10 to 15 A control device according to a second embodiment of the present disclosure is described.

[0079] Fig.101 is a block diagram showing the configuration of the first controlled system CS1 and the second controlled system CS2 controlled by the control device CD according to the second embodiment. Similar to the control device CD of the first embodiment, for example, the control device CD of the second embodiment controls both the configuration and operation of the first controlled system CS1 and the second controlled system CS2.

[0080] like Fig.10 As shown, 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 set or positioned before the first supervised learning model f1. Similarly, in the second controlled system CS2, the second function g2 is set or positioned before the second supervised learning model f2. In the second embodiment, the engineer E (B) expects to evaluate which of the first function g1 and the second function g2 is more robust by using training data TD1 to TDn. The evaluation of the robustness of the first function g1 and the second function g2 will be described in detail later.

[0081] Fig.11 2 is a block diagram showing the configuration of the control device CD according to the second embodiment from a functional viewpoint. Fig.11 As shown, the control device CD according to the second embodiment includes an acquirer 21, a calculator 22 and an evaluator 23. The functions thereof will be described later. Fig.12 The flowchart is described.

[0082] Similar to the control device CD according to the first embodiment, the control device CD according to the second embodiment includes an input circuit IC, a processing circuit PC and an output circuit OC ( Figure 5 Alternatively, the control device CD according to the second embodiment may include an input circuit IC, a processor PR, a storage circuit SC and an output circuit OC ( Figure 6 shown).

[0083] Instead of the first embodiment ( Figure 4 The functions of the definer 11, the setter 12, the calculator 13, the updater 14 and the trainer 15 in the processing circuit PC ( Figure 5 shown) or processor PR ( Figure 6 21, the calculator 22 and the evaluator 23 (shown) Fig.11 shown).

[0084] Operation of the Second Embodiment

[0085] Fig.12: is a flowchart showing the operation of the control device CD according to the second embodiment.

[0086] Fig.13 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 are shown.

[0087] In step ST31, the acquirer 21 ( Fig.11 Get Fig.13 Each of the plurality of combinations COMB_a1 to COMB_l1 includes a first function g1 ( Fig.10 Various values ​​related to one of multiple parameters a to l (as shown).

[0088] For example, Fig.13 As shown, the combination COMB_a1 includes various values ​​a1(1), a1(2), a1(3), ..., a1(n) about the parameter a of the first function g1. Here, the various values ​​a1(1), a1(2), a1(3), ..., a1(n) are outputs returned or fed back by the first supervised learning model f1 based on the training data TD1, TD2, TD3, ..., TD(n).

[0089] For example, value a1(1) is the output returned by the first supervised learning model f1 based on the training data TD1, value a1(2) is the output returned by the first supervised learning model f1 based on the training data TD2,..., value a1(n) is the output returned by the first supervised learning model f1 based on the training data TDn.

[0090] Fig.14 A plurality of combinations COMB_a2 to COMB_12 including various values ​​of a plurality of parameters a to l with respect to the second function g2 are shown.

[0091] In step ST32, the acquirer 21 acquires a plurality of combinations COMB_a2 to COMB_12. Each of the plurality of combinations COMB_a2 to COMB_12 includes various values ​​of one of the plurality of parameters a to l with respect to the second function g2.

[0092] For example, Fig.14 As shown, the combination COMB_a2 includes various values ​​a2(1), a2(2), a2(3), ..., a2(n) about the parameter a of the second function g2. Here, the various values ​​a2(1), a2(2), a2(3), ..., a2(n) are outputs returned or fed back by the second supervised learning model f2 based on the training data TD1, TD2, TD3, ..., TD(n).

[0093] For example, value a2(1) is the output returned by the second supervised learning model f2 based on the training data TD1, value a2(2) is the output returned by the second supervised learning model f2 based on the training data TD2,..., value a2(n) is the output returned by the second supervised learning model f2 based on the training data TDn.

[0094] Fig.15 A plurality of distributions from distribution D_a1 of combination COMB_a1 to distribution D_l1 of combination COMB_l1 and a plurality of distributions from distribution D_a2 of combination COMB_a2 to distribution D_l2 of combination COMB_l2 are shown.

[0095] In step ST33, the calculator 22 ( Fig.11 As shown) calculate the combination COMB_a1 ( Fig.13 The distribution D_a1( Fig.15 shown) to combination COMB_l1 ( Fig.13 The distribution D_l1( Fig.15 More specifically, the calculator 22 calculates, for example, a distribution D_a1 of a combination COMB_a1 including a combination COMB_a1 and a first function g1 ( Fig.10 Various values ​​a1(1), a1(2), a1(3), ..., a1(n) related to the parameter a (shown) Fig.13 As shown in Fig.15 As shown, the distribution D_a1 of the combination COMB_a1 has a range or extension that is specific to the distribution D_a1.

[0096] Similar to the calculation of the distribution D_a1 , the calculator 22 calculates other distributions D_b1 to D_l1 , wherein each of the distributions D_b1 to D_l1 has its own unique range or extension.

[0097] In step ST34, the calculator 22 calculates the combination COMB_a2 ( Fig.14 The distribution D_a2( Fig.15 shown) to combination COMB_l2 ( Fig.14 The distribution D_l2( Fig.15 More specifically, the calculator 22 calculates, for example, a distribution D_a2 of a combination COMB_a2, the combination COMB_a2 comprising the distribution D_a2 and the second function g2 ( Fig.10 Various values ​​a2(1), a2(2), a2(3), ..., a2(n) related to the parameter a (shown) Fig.14 As shown in Fig.15 As shown, the distribution D_a2 of the combination COMB_a2 has a range or extension that is specific to the distribution D_a2.

[0098] Similar to the calculation of the distribution D_a2 , the calculator 22 calculates other distributions D_b2 to D_l2 , wherein each of the distributions D_b2 to D_l2 has its own unique range or extension.

[0099] In step ST35, the evaluator 23 ( Fig.11 By comparing multiple distributions D_a1 to D_l1 ( Fig.15 range or extension and multiple distributions D_a2 to D_l2 (shown) Fig.15 ) to evaluate the first function g1 ( Fig.10 shown) and the second function g2( Fig.10 Which of the two is more robust?

[0100] For example, the evaluator 23 compares the range or spread of the distribution D_a1 with the range or spread of the distribution D_a2. Fig.15 As shown, the range or spread of the distribution D_a2 is larger than that of the distribution D_a1 , so from the viewpoint of the distributions D_a1 and D_a2 , the evaluator 23 evaluates that the second function g2 is more robust than the first function g1 .

[0101] Similar to the range or spread of the distributions D_a1 and D_a2 , the evaluator 23 compares the range or spread of the distributions D_b1 and D_b2 , D_c1 and D_c2 , ​​. . . , D_l1 and D_l2 .

[0102] The evaluator 23 considers all results of comparing the distributions D_a1 and D_a2 , the distributions D_b1 and D_b2 , the distributions D_c1 and D_c3 , . . . , the distributions D_l1 and D_l2 to evaluate which of the first function g1 and the second function g2 is more robust.

[0103] Effects of the Second Embodiment

[0104] As described above, the control device CD according to the second embodiment compares the range or spread of the distribution D_a1 to D_l1 and the range or spread of the distribution D_a2 to D_l2, respectively, thereby being able to evaluate which of the first function g1 and the second function g2 is more robust.

[0105] Third embodiment

[0106] Configuration of the Third Embodiment

[0107] Now refer to Figures 16 to 22 A control device according to a third embodiment of the present disclosure is described.

[0108] Fig.16is a block diagram showing a configuration of a controlled system CS controlled by a control device CD according to a third embodiment.Similar to the control device CD of the first embodiment, the control device CD of the third embodiment controls both the configuration and operation of the controlled system CS, for example.

[0109] like Fig.16 As shown, the controlled system CS includes a supervised learning model f and a function g, wherein the function g is configured or placed before the supervised learning model f. In the third embodiment, the engineer E (B) desires to assign various values ​​to a plurality of parameters a to l of the function g by using the training data TD1 to TDn. The above assignment will be described in detail later.

[0110] Fig.17 2 is a block diagram showing the configuration of the control device CD according to the third embodiment from a functional viewpoint. Fig.17 As shown, the control device CD according to the third embodiment includes an acquirer 31, a setter 32, a selector 33 and a distributor 34. The functions thereof will be described later. Fig.18 The flowchart is described.

[0111] Similar to the control device CD according to the first embodiment, the control device CD includes an input circuit IC, a processing circuit PC and an output circuit OC ( Figure 5 Alternatively, the control device CD according to the third embodiment may include an input circuit IC, a processor PR, a storage circuit SC and an output circuit OC (as shown in FIG. Figure 6 shown).

[0112] Instead of the first embodiment ( Figure 4 The functions of the definer 11, the setter 12, the calculator 13, the updater 14 and the trainer 15 in the processing circuit PC ( Figure 5 shown) or processor PR ( Figure 6 ) implements the getter 31, the setter 32, the selector 33 and the allocator 34 ( Fig.17 shown).

[0113] Operation of the Third Embodiment

[0114] Fig.18 : is a flowchart showing the operation of the control device CD according to the third embodiment.

[0115] Fig.19 A plurality of combinations COM(1) to COM(n) including various values ​​associated with a plurality of parameters a to l of the function g are shown.

[0116] In step ST41, the acquirer 31 ( Fig.17 As shown) obtain multiple combinations COMB(1) to COMB(n) ( Fig.19Each of the plurality of combinations COMB(1) to COMB(n) includes various values ​​associated with a plurality of parameters a to l of the function g.

[0117] For example, Fig.19 As shown, the combination COMB(1) includes various values ​​a(1), b(1), c(1), ..., l(1) associated with multiple parameters a to l of the function g. Here, the various values ​​a(1), b(1), c(1), ..., l(1) are outputs returned or fed back by the supervised learning model f based on the training data TD1, the various values ​​a(2), b(2), c(2), ..., l(2) are outputs returned by the supervised learning model f based on the training data TD2, ..., and the various values ​​a(n), b(n), c(n), ..., l(n) are outputs returned by the supervised learning model f based on the training data TDn.

[0118] Fig. 20 A plurality of combinations COMB(1) to COMB(n) and sampling points SP arranged in space are shown.

[0119] In step ST42, if Fig. 20 As shown, the setter 42 ( Fig.17 As shown in FIG. 1 , a plurality of combinations COMB(1) to COMB(n) whose number is equal to the number of the plurality of parameters a to l of the function g are configured or positioned in the 12-dimensional space.

[0120] In step ST43, for example, Fig. 20 As shown, the acquirer 41 acquires the sampling point SP by using a well-known scheme or method.

[0121] In step 44, Fig. 20 In the plurality of combinations COMB(1) to COMB(n) shown, the selector 33 ( Fig.17 As shown in FIG. 1 , a combination that is positioned or placed closest to the sampling point SP is selected, namely, combination COMB(3).

[0122] In step 45, the distributor 34 ( Fig.17 As shown) the selected combination COMB(3) (more specifically, the various values ​​a(3), b(3), c(3), ..., l(3) included in the selected combination COMB(3)) (as shown Fig.19 As shown) multiple parameters a to l are assigned to the function g.

[0123] Effects of the Third Embodiment

[0124] As described above, the control device CD according to the third embodiment selects a combination that is placed or disposed closest to the sampling point SP among multiple combinations COMB(1) to COMB(n), and more specifically, selects the combination COMB(3) that enables the assignment of various values ​​a(3), b(3), c(3), ..., l(3) of multiple parameters a to l that may be suitable for the function g.

[0125] Modification of the third embodiment

[0126] Fig.21 A plurality of combinations COMB(1) to COMB(n) and unexpected sampling points SP are shown.

[0127] The above-mentioned known scheme or method for sampling point SP may give an unexpected sampling point USP ( ) which is located or placed farther or farthest from a group of several combinations (e.g., COMB (3), COM (6), COMB (2) and COMB (1)). Fig.21 ), instead of the expected sampling point SP ( Fig. 20 If only the unexpected sampling point USP is available for selection in the above step ST44, the allocation in step ST45 cannot give any values ​​suitable for the plurality of parameters a to l of the function g.

[0128] Fig. 22 A plurality of combinations COMB(1) to COMB(n) and sampling points SP1, SP2, and SP3 are shown.

[0129] In order to improve the accuracy of both the selection and the allocation, in a variation of the third embodiment, Fig. 22 As shown, the acquirer 31 ( Fig.17 As shown in FIG. 1 ), a plurality of sampling points, for example, three sampling points SP1, SP2, and SP3, are obtained to replace the third embodiment ( Fig. 20 Even if the sampling point SP2 is unexpected, at least one of the other sampling points SP1 and SP3 may be expected, thereby enabling the selector 33 to select the combination COMB(3) placed or positioned closest to the sampling point SP1 or the combination COMB(1) placed or positioned closest to the sampling point SP3. Similar to the third embodiment, this allows various values ​​included in the combination COMB(3) or COMB(1) that may be suitable to be assigned to the multiple parameters a to l of the function g.

Claims

1. A method comprising the following steps: Define a function g with multiple parameters a to l; The function g is set before the supervised learning model f; the supervised learning model f is trained using training data TDi including examples xi and labels yi, and i is an integer; Define a function f(g(xi)) by using the supervised learning model f and the function g; Calculate the result Zi by calculating the function f(g(xi)); Calculate the difference between the result Zi and the label yi; Calculating the partial differentials PDi of the difference Li with respect to the plurality of parameters a to l of the function g; Update the plurality of parameters a to l of the function g according to the partial differential PDi; When the difference Li converges in the process of repeating between the step of calculating the result Zi and the step of updating the plurality of parameters a to l of the function g, a pseudo preprocessing h is defined by using the plurality of parameters a to l of the function g; Setting the pseudo preprocessing h before the supervised learning model f; as well as The supervised learning model f is trained by providing the training data TDi to the pseudo preprocessing h.

2. A method comprising the following steps: Obtaining a first plurality of combinations COMB_a1 to COMB_l1 by providing training data TDi to a first supervised learning model f1, wherein each of the first plurality of combinations COMB_a1 to COMB_l1 includes various values ​​of one of a plurality of parameters a to l with respect to a first function g1, the first function g1 being set before the first supervised learning model f1 implementing feedback on the first function g1, i being an integer; Obtaining a second plurality of combinations COMB_a2 to COMB_l2 by providing the training data TDi to a second supervised learning model f2, wherein the second plurality of combinations COMB_a2 to COMB_l2 each include various values ​​of one of a plurality of parameters a to l with respect to a second function g2, the second function g2 being provided before the second supervised learning model f2 implementing feedback on the second function g2, the operation of the first supervised learning model f1 and the operation of the second supervised learning model f2 being equivalent to each other; calculating a first plurality of distributions D_a1 to D_l1 respectively regarding the first plurality of combinations COMB_a1 to COMB_l1; calculating a second plurality of distributions D_a2 to D_l2 respectively regarding the second plurality of combinations COMB_a2 to COMB_l2; and Which of the first function g1 and the second function g2 is more robust is evaluated by comparing the first plurality of distributions D_a1 to D_l1 and the second plurality of distributions D_a2 to D_l2, respectively.

3. A method comprising the following steps: Acquire a plurality of combinations COMB(1) to COMB(n) by providing training data TDi to a supervised learning model f, each of the plurality of combinations COMB(1) to COMB(n) including a plurality of values ​​of a plurality of parameters a to l with respect to a function g, the function g being set before the supervised learning model f implementing feedback on the function g, n being an integer equal to or greater than 2; arranging the plurality of combinations COMB(1) to COMB(n) in a space having a plurality of dimensions, the number of the plurality of dimensions being equal to the number of the plurality of parameters a to l of the function g; Acquire at least one sampling point SP in the space; selecting at least one combination of the plurality of combinations COMB(1) to COMB(n) that is closest to the at least one sampling point SP; as well as The plurality of values ​​included in a selected one of the plurality of combinations COMB(1) to COMB(n) are assigned to the plurality of parameters a to l of the function g.

4. A device, comprising: a first definer defining a function g having a plurality of parameters a to l; A first setter, which sets the function g before a supervised learning model f; the supervised learning model f is trained using training data TDi including examples xi and labels yi, and i is an integer; a second definer that uses the supervised learning model f and the function g to define a function f(g(xi)); A first calculator, wherein the first calculator calculates a result Zi by calculating the function f(g(xi)); a second calculator, which calculates the difference between the result Zi and the label yi; a third calculator for calculating a partial differential PDi of the difference Li with respect to the plurality of parameters a to l of the function g; An updater, the updater updating the plurality of parameters a to l of the function g according to the partial differential PDi; a third definer that defines a pseudo preprocessing h by using the plurality of parameters a to l of the function g when the difference Li converges in a process of repeating between the calculation of the result Zi and the updating of the plurality of parameters a to l of the function g; A second setter, which sets the pseudo preprocessing h before the supervised learning model f; as well as A trainer, which trains the supervised learning model f by providing the training data TDi to the pseudo preprocessing h.

5. A device, comprising: a first acquirer that acquires a first plurality of combinations COMB_a1 to COMB_l1 by providing training data TDi to a first supervised learning model f1, wherein each of the first plurality of combinations COMB_a1 to COMB_l1 includes various values ​​of one of a plurality of parameters a to l with respect to a first function g1 that is set before the first supervised learning model f1 that implements feedback on the first function g1, wherein i is an integer; a second acquirer, which acquires a second plurality of combinations COMB_a2 to COMB_l2 by providing the training data TDi to a second supervised learning model f2, wherein the second plurality of combinations COMB_a2 to COMB_l2 each include various values ​​of one of a plurality of parameters a to l with respect to a second function g2, the second function g2 being set before the second supervised learning model f2 implementing feedback on the second function g2, the operation of the first supervised learning model f1 and the operation of the second supervised learning model f2 being equivalent to each other; a first calculator that calculates a first plurality of distributions D_a1 to D_l1 with respect to the first plurality of combinations COMB_a1 to COMB_l1, respectively; a second calculator that calculates a second plurality of distributions D_a2 to D_l2 respectively regarding the second plurality of combinations COMB_a2 to COMB_l2; as well as An evaluator that evaluates 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 and the second plurality of distributions D_a2 to D_l2, respectively.

6. A device, comprising: a first acquirer, which acquires a plurality of combinations COMB(1) to COMB(n) by providing training data TDi to a supervised learning model f, wherein each of the plurality of combinations COMB(1) to COMB(n) includes a plurality of values ​​of a plurality of parameters a to l with respect to a function g, wherein the function g is set before the supervised learning model f implements feedback on the function g, and n is an integer equal to or greater than 2; a setter for setting the plurality of combinations COMB(1) to COMB(n) in a space having a plurality of dimensions, the number of the plurality of dimensions being equal to the number of the plurality of parameters a to l of the function g; A second acquirer, which acquires at least one sampling point SP in the space; a selector for selecting at least one combination of the plurality of combinations COMB(1) to COMB(n) that is closest to the at least one sampling point SP; as well as A distributor that distributes the plurality of values ​​included in a selected one of the plurality of combinations COMB(1) to COMB(n) to the plurality of parameters a to l of the function g.

Images