Parameter adjustment method, parameter adjustment device, and program

Abstract

A parameter adjustment method comprises: a step (S10) of generating at least 2m + 1 parameter sets on the basis of a deterministic screening design method using m parameters of a control system, m being a natural number, and obtaining respective evaluation values of the generated at least 2m + 1 parameter sets; a step (S20) for generating an evaluation value regression model by decision tree regression on the basis of the 2m + 1 parameter sets and the corresponding evaluation values; and a step (S30) for searching for a trial parameter set of the control system on the basis of the parameter importance of the evaluation value regression model.

Description

Technical Field

[0001] This disclosure relates to parameter adjustment methods, parameter adjustment devices, and procedures. Background Technology

[0002] In DC-DC converters used in electric vehicles, multiple feedback controls are combined to perform complex and precise current and voltage control in order to control charging and discharging according to driving conditions. Furthermore, in production equipment such as component assembly machines and assembly robots used in factories, multiple servo-controlled motors are combined to perform complex and precise motion control. In such devices, the actions of the controlled object are controlled according to multiple control parameters. To obtain the desired performance, the user adjusts the control parameters based on experience.

[0003] In the adjustment process of these control parameters, the operating conditions are diverse and complex, and the more parameters there are, the more difficult the adjustment process becomes. The number of trial runs in the adjustment process also increases, requiring a significant amount of labor and time. Therefore, the demand for automating these adjustment processes is growing.

[0004] To address this problem, experimental design methods are known to be used in the tuning of classical control, namely PID (Proportional-Integral-Differential) control. For example, Patent Document 1 discloses a technique in which parameter combinations tuned using existing tuning rules are evaluated multiple times using experimental design methods, and the optimal parameters are selected based on comprehensive consideration. Furthermore, Non-Patent Document 1 discloses a method called Deterministic Screening Design (DSD) as one of the experimental design methods, which uses a meeting matrix to determine the combination of parameters.

[0005] Alternatively, techniques for parameter search using regression models are known. For example, Patent Document 2 discloses a technique in which a regression model is generated based on parameters that perform actions within a specified range from randomly generated parameters, initial parameter values ​​are calculated based on the generated regression model, and optimal parameters are searched. Furthermore, as one type of black-box optimization, Non-Patent Document 2 discloses a method called SMAC (Sequential Model-based Algorithm Configuration), which uses random forest regression in the regression model.

[0006] Existing technical documents

[0007] Patent documents

[0008] Patent Document 1: Japanese Patent No. 6652398

[0009] Patent Document 2: Japanese Patent No. 6283851

[0010] Non-patent literature

[0011] Non-patent literature 1: L.Xiao, DKJLin, F.Bai, "Constructing DefinitiveScreening Designs Using Conference Matrices", Journal of Quality Technology, Vol.44, No.1, pages 2-8, 2012

[0012] Non-patent literature 2: F.Hutter, H.Hoos, and K.Leyton-Brown, "Sequential model-based optimization for general algorithm configuration", In C.Coello, editor, Proceedings of the Fifth International Conference on Learning and IntelligentOptimization (LION'11), volume 6683 of Lecture Notes in Computer Science, pages 507-523, Springer, 2011 Summary of the Invention

[0013] The problem that the invention aims to solve

[0014] However, when the number of control parameters reaches tens to hundreds, and the number of action conditions that must simultaneously satisfy performance also reaches tens, the number of combinations becomes enormous. Therefore, it may be difficult to efficiently set appropriate search priorities and search ranges. Consequently, it may be difficult to efficiently set appropriate parameter values, i.e., it may be difficult to efficiently adjust the parameters.

[0015] This disclosure provides a parameter adjustment method, a parameter adjustment device, and a program that can efficiently adjust parameters.

[0016] Methods for solving problems

[0017] One aspect of this disclosure involves a parameter adjustment method comprising the following steps: a step of obtaining evaluation values, using m parameters of the control system to generate at least 2m+1 parameter sets based on a deterministic screening design method, and obtaining evaluation values ​​for each of the generated at least 2m+1 parameter sets, where m is a natural number; a step of generating an evaluation value regression model, based on the 2m+1 parameter sets and the corresponding evaluation values, generating an evaluation value regression model through decision tree regression; and a step of searching, based on the parameter importance of the evaluation value regression model, searching for a trial parameter set of the control system.

[0018] One aspect of the parameter adjustment apparatus disclosed herein comprises: an acquisition unit that generates at least 2m+1 parameter sets using m parameters of the control system based on a deterministic screening design method, and acquires evaluation values ​​for each of the generated at least 2m+1 parameter sets, where m is a natural number; a generation unit that generates an evaluation value regression model based on the 2m+1 parameter sets and the corresponding evaluation values ​​through decision tree regression; and a search unit that searches for trial parameter sets of the control system based on the parameter importance of the evaluation value regression model.

[0019] One aspect of this disclosure relates to a program for causing a computer to execute the parameter adjustment method described in the above-described aspect.

[0020] Furthermore, these general or specific methods can be implemented by systems, apparatuses, methods, integrated circuits, computer programs, or recording media, or by any combination of systems, apparatuses, methods, integrated circuits, computer programs, and recording media.

[0021] Invention Effects

[0022] One method disclosed herein allows for efficient parameter adjustment.

[0023] Further advantages and effects of one aspect of this disclosure become apparent from the specification and drawings. These advantages and / or effects are provided by various embodiments and features described in the specification and drawings, but it is not necessary to provide all features in order to obtain one or more identical features. Attached Figure Description

[0024] Figure 1 This is a block diagram illustrating an example of a parameter adjustment system according to an embodiment of the present disclosure.

[0025] Figure 2 This is a block diagram illustrating an example of the hardware configuration of a parameter adjustment device according to an embodiment of the present disclosure.

[0026] Figure 3 This is a block diagram illustrating an example of the functional configuration of a parameter adjustment device according to an embodiment of the present disclosure.

[0027] Figure 4 This is a data structure diagram illustrating an example of the action parameters of a parameter adjustment object.

[0028] Figure 5 This is a data structure diagram illustrating an example of the control parameters of a parameter adjustment object.

[0029] Figure 6A This is a diagram showing the meeting matrix (m=4).

[0030] Figure 6B This is a diagram showing the DSD matrix (m=4).

[0031] Figure 6C This is a diagram illustrating an example of the data structure of the initial parameter set.

[0032] Figure 6D This is a graph showing examples of the evaluation values ​​for each parameter set.

[0033] Figure 7 This is a flowchart illustrating an example of the operation of a parameter adjustment system according to an embodiment of the present disclosure.

[0034] Figure 8 This is a flowchart illustrating an example of the parameter search process of a parameter adjustment system according to an embodiment of the present disclosure.

[0035] Figure 9 This is a flowchart illustrating an example of the display processing of a parameter adjustment system according to an embodiment of the present disclosure.

[0036] Figure 10 This is a diagram illustrating an example of a display screen (UI screen) shown by the UI section according to an embodiment of the present disclosure. Detailed Implementation

[0037] (Summary of this disclosure)

[0038] The parameter adjustment method disclosed in the first aspect includes the following steps: a step of obtaining evaluation values, using m parameters of the control system, generating at least 2m+1 parameter sets based on a deterministic screening design method, and obtaining the evaluation values ​​of each of the generated at least 2m+1 parameter sets, where m is a natural number; a step of generating an evaluation value regression model, generating an evaluation value regression model based on the 2m+1 parameter sets and the corresponding evaluation values ​​through decision tree regression; and a step of searching, searching for trial parameter sets of the control system based on the parameter importance of the evaluation value regression model.

[0039] The DSD method is considered capable of conducting fewer experiments with less effect confounding (correlation). Furthermore, the parameter adjustment method described here efficiently searches for the range of parameters that are likely to contribute to improving the evaluation value because it performs a search based on parameter importance. Therefore, parameter adjustment can be performed efficiently.

[0040] The parameter adjustment method of the second aspect of this disclosure, according to the parameter adjustment method of the first aspect, includes the following steps in the search: an extraction step, in which the parameter set corresponding to one of the at least 2m+1 evaluation values ​​obtained in the step of obtaining evaluation values ​​is extracted as a benchmark parameter set; a step of generating one or more candidate parameter sets, in which one or more candidate parameter sets are generated based on the parameter importance and the extracted benchmark parameter set; a prediction step, in which the evaluation value regression model is used to predict the predicted value of each of the one or more candidate parameter sets; and a decision step, in which the candidate parameter set corresponding to the predicted value that is better than the evaluation value corresponding to the benchmark parameter set among the one or more predicted values ​​is determined as the trial parameter set.

[0041] Therefore, the parameter set that yields evaluation values ​​better than those corresponding to the baseline parameter set is determined as the trial parameter set. By improving the evaluation values, appropriate parameter adjustments can be made.

[0042] The parameter adjustment method of the third aspect of this disclosure, according to the parameter adjustment method of the second aspect, extracts the parameter set corresponding to the optimal evaluation value among the at least 2m+1 evaluation values ​​as the benchmark parameter set in the extraction step.

[0043] Therefore, it is possible to efficiently improve the evaluation value (predicted value) and to make the search and adjustment of parameters more efficient.

[0044] The parameter adjustment method of the fourth aspect of this disclosure, according to the parameter adjustment method of the second or third aspect, in the step of generating one or more candidate parameter sets, generates one or more random parameter sets that follow a normal distribution as the one or more candidate parameter sets, wherein the normal distribution has the reference parameter set as the average value and the parameter importance as the variance.

[0045] Therefore, a set of candidate parameters that can easily contribute to improving the evaluation value (predicted value) can be generated. This enables efficient improvement of the evaluation value (predicted value) and streamlines the search and adjustment of parameters.

[0046] The parameter adjustment method of the fifth aspect of this disclosure, according to the parameter adjustment method of any one of the first to fourth aspects, further acquires the evaluation value of the trial parameter set determined in the search step in the step of acquiring the evaluation value, and the step of generating the evaluation value regression model includes the step of updating the evaluation value regression model based on the evaluation value of the trial parameter set in the step of generating the evaluation value regression model.

[0047] Therefore, the accuracy of evaluation values ​​predicted using evaluation value regression models can be improved. By improving the reproducibility of evaluation values ​​based on the actual actions of the control system, the accuracy of parameter search and adjustment can be increased.

[0048] The parameter adjustment method of the sixth method of this disclosure, according to the parameter adjustment method of the fifth method, involves repeatedly executing the step of obtaining the evaluation value, the step of updating the regression model of the evaluation value, and the search step until the number of repetitions reaches a predetermined number or until the evaluation value of the trial parameter set is better than a predetermined threshold.

[0049] Therefore, it is possible to efficiently obtain the optimal parameter set or a sufficiently suitable parameter set.

[0050] The parameter adjustment method according to the seventh aspect of this disclosure, based on the parameter adjustment method according to any one of the first to sixth aspects, further includes a display step, wherein for each of the m parameters, the parameter importance of the evaluation value regression model and the prediction curve of the evaluation value predicted based on the evaluation value regression model are displayed.

[0051] Therefore, users can visually confirm the importance of parameters and the prediction curve, thus enabling them to adjust parameters efficiently.

[0052] The parameter adjustment method of the eighth aspect of this disclosure, according to the parameter adjustment method of the seventh aspect, further displays a marker in the display step, which overlaps the position of the prediction curve representing the value of the corresponding parameter, and the m prediction curves are updated when the marker moves.

[0053] Therefore, by displaying overlapping markers, users can visually confirm the parameter values ​​used in obtaining the predicted values ​​for the evaluation. Because users can confirm the relationship between the parameter values ​​and the predicted values, they can efficiently adjust the parameters.

[0054] The parameter adjustment method according to the ninth aspect of this disclosure, based on the parameter adjustment method according to the eighth aspect, includes the step of accepting a user's movement operation on the marker.

[0055] Therefore, users can manually move the markers and visually confirm the changes in the predicted values ​​that accompany the marker movement. Because users can confirm the relationship between changes in parameter values ​​and changes in predicted values, they can efficiently adjust the parameters.

[0056] The parameter adjustment method involved in the 10th aspect of this disclosure, according to the parameter adjustment method involved in any one of the 7th to 9th aspects, further shows the decision tree used in the generation of the evaluation value regression model in the shown steps.

[0057] Therefore, users can visually confirm the structure of the decision tree. Because users can confirm the specific structure of the decision tree, they can efficiently adjust the parameters.

[0058] The program involved in the 11th aspect of this disclosure is used to cause a computer to execute the parameter adjustment method described in any one of the 1st to 10th aspects.

[0059] Therefore, similar to the parameter adjustment methods mentioned above, it is possible to adjust parameters efficiently.

[0060] The parameter adjustment apparatus according to the 12th aspect of this disclosure comprises: an acquisition unit that generates at least 2m+1 parameter sets using m parameters of the control system based on a deterministic screening design method, and acquires evaluation values ​​for each of the generated at least 2m+1 parameter sets, where m is a natural number; a generation unit that generates an evaluation value regression model based on the 2m+1 parameter sets and the corresponding evaluation values ​​through decision tree regression; and a search unit that searches for trial parameter sets of the control system based on the parameter importance of the evaluation value regression model.

[0061] Therefore, similar to the parameter adjustment methods mentioned above, it is possible to adjust parameters efficiently.

[0062] Hereinafter, embodiments of the present disclosure will be described in detail with appropriate reference to the accompanying drawings. However, sometimes necessary detailed descriptions are omitted. For example, detailed descriptions of matters already well known or repetitive descriptions of substantially the same structures are sometimes omitted. This is to avoid making the following description unnecessarily lengthy and to facilitate understanding by those skilled in the art.

[0063] Furthermore, the accompanying drawings and the following description are provided to enable those skilled in the art to fully understand this disclosure and are not intended to limit the subject matter described in the technical solutions.

[0064] For example, each figure is a schematic diagram and not necessarily a strict representation. Therefore, for example, the scales in each figure may not be consistent. In addition, substantially identical components are labeled with the same symbols in each figure, and repeated descriptions are omitted or simplified.

[0065] (Implementation Method)

[0066] <Brief Structure of the Parameter Adjustment System>

[0067] Figure 1 This is a block diagram illustrating an example of a parameter adjustment system according to an embodiment of this disclosure. Figure 1 As shown, the parameter adjustment system 10 includes a parameter adjustment device 100, a user interface (UI) unit 120, a database (DB) unit 140, an adjustment object 160, and a sensor 180. Furthermore, the parameter adjustment device 100 is an example of a parameter adjustment device according to this disclosure, and also an example of a computer. The UI unit 120 is an example of a display unit according to this disclosure. The adjustment object 160 is an example of a control system according to this disclosure, and is also referred to as an optimization object. The sensor 180 is used for evaluating the motion of the adjustment object 160.

[0068] The parameter adjustment device 100 is communicatively connected to the UI unit 120, DB unit 140, adjustment target 160, and sensor 180, and acquires and / or provides data and / or information to each of them. Communication is performed via wired or wireless means, or a combination of wired and wireless means. Furthermore, the acquisition of data and / or information can also be described as receiving or inputting data and / or information. The provision of data and / or information can also be described as sending or outputting data and / or information.

[0069] The parameter adjustment system 10 searches and adjusts m control parameters of the object 160 in a manner that the specified action (action condition) to be performed by the object 160 is to satisfy the target value representing the specified performance. Here, m is a natural number.

[0070] Furthermore, in this specification, the parameters used to define the operating conditions are referred to as operating parameters or operating condition parameters. Operating parameters include, for example, input voltage, output voltage, output current, moving speed, moving distance, and stop position. Additionally, the internal parameters of the adjustment object 160 that achieve the desired performance of the operating conditions are referred to as control parameters or control condition parameters. Control parameters include, for example, control gain and filter time constant. The parameter adjustment system 10 searches for m control parameters.

[0071] Here, "search" refers to automatically finding parameter values ​​that satisfy a target value using various well-known algorithms. In this embodiment, the parameter adjustment system 10 searches for a parameter set as a group of m parameter values. The m parameter values ​​are the individual values ​​of the m control parameters. A parameter set is a group of the individual values ​​of the m control parameters, and can be called a control parameter set.

[0072] Adjustment refers to changing parameter values ​​through searching or manual methods. Furthermore, parameters related to the search are recorded as hyperparameters. Hyperparameters include, for example, search priority (order of precedence), search range, number of searches, and search criteria (target value settings).

[0073] The UI unit 120 is implemented, for example, by a UI device or apparatus such as a monitor, keyboard, or mouse. The UI unit 120 receives adjustment items from the user through user operation. These adjustment items include, for example, action conditions, desired performance, priority parameters, and search range. The UI unit 120 provides the received adjustment items to the parameter adjustment device 100 (specifically, the adjustment item setting unit 102 described later) in order to enable the parameter adjustment device 100 to set the adjustment items. The UI unit 120 receiving adjustment items from the user and providing them to the parameter adjustment device 100 can be understood as the user setting the adjustment items.

[0074] The UI unit 120 displays information set in the parameter adjustment device 100 and / or trial history in text and / or graphical form. Here, trial refers to a series of actions by which the parameter adjustment device 100 optimizes the parameter values ​​set on the adjustment object 160 through repeated control parameter adjustment processing. Control parameter adjustment processing refers to the process of evaluating parameter values ​​that have become setting candidates and determining the next parameter value set on the adjustment object 160 as a setting candidate based on the evaluation results.

[0075] Furthermore, although the UI unit 120 is shown separately from the parameter adjustment device 100 in the figure, it can also be assembled into the parameter adjustment device 100. In addition, the UI unit 120 can be a separate device for input and output, or it can be an integrated device.

[0076] DB unit 140 is implemented, for example, by a storage device or apparatus such as a hard disk drive in a data management server or data center. As detailed below, DB unit 140 stores configuration information of adjustment object 160 and / or sensor 180, as well as historical records of previously set parameters.

[0077] In addition, the configuration information of the adjustment object 160 and / or sensor 180 stored in the DB unit 140, as well as the historical records of previously set parameters, are read out by the parameter adjustment device 100 (specifically, the adjustment item setting unit 102 described later).

[0078] Although the DB unit 140 is shown as separate from the parameter adjustment device 100, it can also be assembled into the parameter adjustment device 100.

[0079] The parameter adjustment device 100 uses m control parameters of the adjustment object 160 to generate at least 2m+1 parameter sets based on a deterministic screening design method, and obtains the evaluation values ​​of each of the generated at least 2m+1 parameter sets. Furthermore, based on the 2m+1 parameter sets and their corresponding evaluation values, the parameter adjustment device 100 generates an evaluation value regression model (specifically, the evaluation value prediction model 206 described later) through decision tree regression. Based on the parameter importance of the evaluation value regression model, the parameter adjustment device 100 searches for trial parameter sets for the adjustment object 160.

[0080] Specifically, the parameter adjustment device 100 determines the search range of control parameters for the adjustment object 160 based on the adjustment items accepted by the user and information stored in the database unit 140, thereby performing the search and adjustment of the control parameters for the adjustment object 160. For example, the parameter adjustment device 100 performs the search and adjustment of control parameters in the following manner.

[0081] The parameter adjustment device 100 (specifically the parameter search unit 104 described later) generates a combination of control parameter values ​​(i.e., a parameter set or control parameter set) that improves the target value based on past trial results.

[0082] The parameter adjustment device 100 (specifically, the trial vector setting unit 106 described later) sets a trial vector, consisting of a generated set of control parameters and one or more action parameter values ​​constituting the action conditions for adjustment, and provides it to the adjustment object 160. Providing the trial vector to the adjustment object 160 can be understood as setting the trial vector on the adjustment object 160. Thus, the parameter adjustment device 100 can control the adjustment object 160 to operate according to the trial vector. In this specification, a vector refers to a combination of a set of values ​​for one or more action parameters constituting the action conditions (i.e., an action parameter set) and a set of control parameters. Sometimes only the control parameter set is referred to as a vector, and sometimes only the action parameter set is referred to as a vector.

[0083] In addition, the setting parameters and other information of a certain device or functional unit (processing unit) can refer to one or more of the following: the device or functional unit (processing unit) stores such information in a data table, the device or functional unit (processing unit) performs an operation according to such information, and the device or functional unit (processing unit) is controlled in a manner in which other devices or functional units (processing units) perform an operation according to such information.

[0084] The parameter adjustment device 100 (specifically, the motion evaluation unit 108 described later) receives motion results from the sensor 180, indicating that the object 160 has performed a motion according to the trial vector. The parameter adjustment device 100 evaluates the received motion results and determines whether the target value has been improved. The parameter adjustment device 100 uses this series of trial results as a prior distribution in the next vector generation.

[0085] The parameter adjustment device 100 stores the distribution of trial values ​​(parameter values) of each parameter, the distribution of vectors of each trial (or setting), and the distribution of evaluation values ​​of each trial (or setting) in the DB unit 140. That is, the parameter adjustment device 100 writes each distribution into the DB unit 140.

[0086] Based on these distributions, the parameter adjustment device 100 generates an evaluation value prediction model through regression. The parameter adjustment device 100 generates parameter decision thresholds, parameter importance, and prediction curves for evaluation values ​​in response to parameter changes, all based on the generated evaluation value prediction model. The parameter adjustment device 100 provides information about the evaluation value prediction model and the prediction curves to the UI unit 120 for display.

[0087] The adjustment object 160 is an example of a control system that is adjusted as a control parameter, such as an actual DC-DC converter or a component assembly machine such as production equipment. Alternatively, the adjustment object 160 can also be a model simulating the actions of an actual machine (i.e., a simulation model). When the adjustment object 160 is a simulation model, the simulation model can be implemented in hardware, software, or a combination of both. The adjustment object 160 as a simulation model can also be referred to as a virtual device.

[0088] The adjustment object 160 performs the prescribed action according to the trial vector set by the parameter adjustment device 100.

[0089] Sensor 180 can be a physical device or a model simulating the actions of a sensor (i.e., a simulation model). When sensor 180 is a simulation model, it can be implemented in hardware, software, or a combination of both. Sensor 180 as a simulation model can also be called a virtual sensor.

[0090] Sensor 180 detects the action result of the adjusted object 160 moving according to the set trial vector. Furthermore, the detection performed by sensor 180 can be understood as sensing or measuring. The action result is, for example, a performance value related to the action. Sensor 180 provides the detected action result to parameter adjustment device 100 (specifically, the action evaluation unit 108 described later).

[0091] As performance values ​​related to the action, physical quantities such as voltage, current, displacement, velocity, vibration, noise, and / or heat generation of the object being adjusted can be used. Alternatively, multiple performance values ​​or evaluation values ​​based on evaluation criteria can be normalized or weighted and summed to serve as a target value for the search object. In this specification, performance values, evaluation values, and target values ​​are sometimes used without special distinction.

[0092] In this way, the parameter adjustment system 10 can provide prompts to the user by displaying information set in the parameter adjustment device 100, trial history, etc., on the UI unit 120. Therefore, the user can efficiently set appropriate control parameters by referring to (e.g., comparing) previously set parameters while operating the search conditions. Furthermore, the search and adjustment of control parameters can be performed efficiently. Moreover, the user can easily understand the search or adjustment process through the displayed information, trial history, etc. Thus, even novice or inexperienced users can efficiently set appropriate control parameters.

[0093] Furthermore, the more than one block illustrated can be installed in one device or in different devices. For example, the adjustment object 160 can be installed as a single device including the UI unit 120, DB unit 140, parameter adjustment device 100, and sensor 180. Alternatively, the parameter adjustment device 100 can be implemented in the cloud, the UI unit 120 can be installed on the user's local PC (personal computer), smartphone, tablet, etc., and the blocks can be connected via a communication network (not shown). Furthermore, multiple adjustment objects 160 and / or sensors 180 can be provided.

[0094] Furthermore, the UI unit 120 and / or parameter adjustment device 100 not only set the adjustment items input by the user, but can also operate the GUI (Graphical User Interface) of the prior art adjustment software used for manually adjusting the parameters of the adjustment object 160 using RPA (Robotic Process Automation), thereby setting control parameters and acquiring action results. This RPA uses image recognition and other methods for automatic control. Accordingly, parameter adjustment can be performed even for the prior art adjustment object 160, which does not have a dedicated communication interface.

[0095] <Composition of the parameter adjustment device>

[0096] The parameter adjustment device 100 is implemented by a computing device or information processing device such as a server or PC. The computing device is also called a computer or PC. For example, the parameter adjustment device 100 may have… Figure 2 The hardware configuration is shown. Figure 2This is a block diagram illustrating an example of the hardware configuration of the parameter adjustment device 100 according to an embodiment of this disclosure. For example, such as... Figure 2 As shown, the parameter adjustment device 100 may include a storage device 110, a processing device 112, a UI device 114, and a communication device 116, which are interconnected via a bus 118.

[0097] The program or instructions for implementing the functions and processes described later in the parameter adjustment device 100 are stored, for example, in the storage device 110. Alternatively, these programs or instructions can be downloaded from an external device (e.g., a server) via a network or other means, or provided from a removable storage medium such as a CD-ROM (Compact Disc Read Only Memory) or flash memory.

[0098] Storage device 110 is implemented using RAM (Random Access Memory), flash memory, hard disk drive, etc. In addition to installed programs or instructions, storage device 110 also stores files, data, etc., used to execute programs or instructions. Storage device 110 may also include non-transitory storage medium.

[0099] When DB unit 140 is assembled in parameter adjustment device 100, DB unit 140 can be stored in storage device 110.

[0100] The processing device 112 can be implemented by, for example, a general-purpose processor or controller (circuit), or a dedicated processor or controller (circuit). When implemented by a general-purpose processor, the processing device 112 can also be implemented by one or more CPUs (Central Processing Units), GPUs (Graphics Processing Units), processing circuitry, etc., that can be composed of one or more processor cores. The processing device 112 executes the functions and processing of the parameter adjustment device 100 according to the program or instructions stored in the storage device 110, and data such as parameters used to execute the program or instructions.

[0101] The UI device 114 can consist of input devices such as a keyboard, mouse, camera, and microphone; output devices such as a monitor, speaker, headphones, and printer; and input / output devices such as a touch screen. The UI device 114 provides an interface between the user and the parameter adjustment device 100. For example, the user operates the parameter adjustment device 100 by using a keyboard, mouse, etc., to operate the GUI displayed on the monitor or touch screen.

[0102] When the UI unit 120 is assembled in the parameter adjustment device 100, the UI unit 120 can be implemented by part or all of the UI device 114.

[0103] The communication device 116 can be implemented by various communication circuits that perform communication processing with external devices, the Internet, LAN (Local Area Network), VPN (Virtual Private Network), and other communication networks.

[0104] The hardware configuration of the parameter adjustment device 100 described above is merely an example. The parameter adjustment device 100 disclosed herein can be implemented using other suitable hardware configurations.

[0105] <Functional Composition of the Parameter Adjustment Device>

[0106] Next, use Figure 3 This section explains the specific processing of the functional components of the parameter adjustment device 100.

[0107] Figure 3 This is a block diagram illustrating an example of the functional configuration of the parameter adjustment device 100 according to an embodiment of the present disclosure. For example, Figure 1 and Figure 3 As shown, the parameter adjustment device 100 includes an adjustment item setting unit 102, a parameter search unit 104, a trial vector setting unit 106, and a motion evaluation unit 108. Furthermore, the parameter adjustment device 100 may also include other components. Additionally, the adjustment item setting unit 102, the parameter search unit 104, the trial vector setting unit 106, and the motion evaluation unit 108 may each include various sub-functional units or processing units. The adjustment item setting unit 102 is an example of a display control unit according to this disclosure. The parameter search unit 104 is an example of an acquisition unit, a generation unit, and a search unit according to this disclosure. The motion evaluation unit 108 is an example of an evaluation unit according to this disclosure.

[0108] The adjustment setting unit 102 provides the parameter search unit 104 with the action conditions 201 for parameter adjustment and the adjustment items 202 as the adjustment targets, based on the setting items and other information input from the UI unit 120 and / or the DB unit 140. Specifically, the action conditions 201 are k action parameters. Specifically, the adjustment items 202 are m control parameters and / or hyperparameters used for searching. Both k and m are natural numbers. Providing the action conditions 201 and adjustment items 202 can be understood as setting the action conditions 201 and adjustment items 202 for the parameter search unit 104.

[0109] The parameter search unit 104 searches for control parameters based on the settings of the adjustment item setting unit 102. In the parameter search performed by the parameter search unit 104, the DSD matrix 203 and the evaluation value prediction model 206 are used. The evaluation value prediction model 206 is an example of an evaluation value regression model generated through decision tree regression, and serves as a proxy model for the adjustment object 160, sensor 180, and motion evaluation unit 108. Figure 3 As shown, the parameter search unit 104 includes an initial vector generation unit 204, a model generation unit 205, a candidate vector generation unit 207, and a decision unit 208.

[0110] First, the initial vector generation unit 204 generates an initial vector for generating the evaluation value prediction model 206. The initial vector is a vector summarizing the initial values ​​of both the motion parameters and control parameters. The parameter search unit 104 provides the initial vector generated by the initial vector generation unit 204 to the trial vector setting unit 106. The trial vector setting unit 106 sets the initial vector on the adjustment object 160 and causes it to move. The sensor 180 detects the motion result of the adjustment object 160 based on the initial vector and outputs it to the motion evaluation unit 108. The motion evaluation unit 108 calculates an evaluation value based on the motion result and provides the calculated evaluation value to the parameter search unit 104. In the parameter search unit 104, not only the evaluation value is acquired, but also the vector set by the trial vector setting unit 106 (here, the initial vector).

[0111] In this embodiment, a smaller evaluation value indicates a better evaluation, and a larger evaluation value indicates a worse evaluation. That is, improving the evaluation value means reducing the evaluation value, and the optimal evaluation value is the smallest evaluation value. Furthermore, the relationship between the magnitude of the evaluation value and the quality of the evaluation can also be inverse.

[0112] Furthermore, the initial vector is generated based on deterministic screening design (DSD), selecting vectors that are orthogonal within the search range and have the fewest possible combinations. The initial vector generation unit 204 uses the DSD matrix 203 to determine the combinations of the minimum, intermediate, and maximum values ​​of each of the m control parameters, expands these combinations according to the number of action conditions (k), and combines them with the action parameters to form the initial vector. Specific examples will be used later. Figures 4 to 6D Please provide an explanation.

[0113] Next, the model generation unit 205 generates an evaluation value prediction model 206 using decision tree regression. This evaluation value prediction model 206 can predict the evaluation value obtained based on the initial input vector. Decision tree regression can use Random Forest (RF), which generates multiple decision trees in parallel, or Gradient Boosting Decision Tree (GBDT), which generates multiple decision trees sequentially. By using these decision trees, action parameters and control parameters can be continuous or discrete variables, or even categorical variables, which can be used directly without transformation. Furthermore, it is not necessary to unify the value range and scale of each parameter. Moreover, the importance of the input features can be calculated in decision tree regression when predicting the evaluation value. Here, the features are equivalent to the elements of the input vector, i.e., the action parameters and control parameters. The importance of these parameters is referred to as parameter importance below.

[0114] The parameter search unit 104 searches for control parameters based on the evaluation value prediction model 206. Specifically, the candidate vector generation unit 207 extracts the parameter set corresponding to one of the at least 2m+1 evaluation values ​​as a baseline parameter set, and generates one or more candidate parameter sets based on parameter importance and the extracted baseline parameter set. The parameter search unit 104 uses the evaluation value prediction model 206 to predict the predicted values ​​of the evaluation values ​​for each of the one or more candidate parameter sets. The parameter search unit 104 then selects the candidate parameter set corresponding to the predicted value that is better than the evaluation value corresponding to the baseline parameter set from among the one or more predicted values ​​as the trial parameter set.

[0115] For example, the candidate vector generation unit 207 generates candidate vectors as those near the input vector (hereinafter referred to as the optimal vector) that corresponds to the optimal evaluation value at the computation time. Specifically, the candidate vector generation unit 207 generates multiple random candidate vectors that follow a normal distribution, where the optimal vector at the computation time is used as the average of each parameter and the parameter importance calculated by the evaluation value prediction model 206 is used as the variance of each parameter. The candidate vector generation unit 207 inputs these candidate vectors into the evaluation value prediction model 206 to calculate the predicted evaluation value. The candidate vector generation unit 207 selects the vector that improves the predicted evaluation value and outputs the selected vector as a trial vector to the trial vector setting unit 106.

[0116] The evaluation value prediction model 206 can calculate the importance of parameters based on decision tree regression, thus limiting the search scope. For example, it can be assumed that parameters with high importance are more likely to output better evaluation values, generating random vectors by searching a wider range. Furthermore, it can be assumed that parameters with low importance have little impact on the optimal evaluation value, and the necessity to change their values ​​at computation time is low, allowing for a narrower search range. This avoids treating all parameters the same way and suppresses the increase in search times caused by searching ranges that contribute little to improving evaluation values.

[0117] Furthermore, if no improvement in the evaluation value can be expected near the optimal vector at the computation time, in addition to using the optimal vector at the computation time, the vectors used for averaging can be changed sequentially, such as the second-best vector, the third-best vector, and so on. Alternatively, a random vector can be selected with a specified probability from the region that has not yet been searched, and its vicinity can be searched.

[0118] Furthermore, various well-known algorithms can be used to select such candidate vectors. For example, heuristic algorithms such as Bayesian optimization, genetic algorithms, and particle swarm optimization, as well as mathematical programming and various solution methods in combinatorial optimization, can be employed.

[0119] The trial vector setting unit 106 provides the trial vector input from the parameter search unit 104 to the adjustment object 160. When the trial vector consists of a parameter set of m generated control parameters and multiple action conditions (each containing more than one action parameter value), the trial vector setting unit 106 causes the adjustment object 160 to operate according to each action condition.

[0120] For example, after setting m control parameters to the adjustment object 160, the trial vector setting unit 106 sets the action conditions to the adjustment object 160 one by one, causing the adjustment object 160 to operate sequentially. The sensor 180 detects the action result of the adjustment object 160 according to each action condition.

[0121] Furthermore, for example, when there are multiple identical adjustment objects 160, the trial vector setting unit 106 can set m control parameters for all identical adjustment objects 160, and then set different operating conditions for each identical adjustment object 160, thereby causing the identical adjustment objects 160 to operate in parallel. The sensor 180 obtains the operating result of each operating condition by detecting the respective operating results of the identical adjustment objects 160.

[0122] The motion evaluation unit 108 evaluates the motion results (e.g., performance values) of the adjustment object 160 according to the set trial vector based on the motion results input from the sensor 180. When the trial vector consists of a parameter set of m generated control parameters and multiple motion conditions (each containing more than one motion parameter value), the motion evaluation unit 108 evaluates the motion results under each motion condition. The motion evaluation unit 108 converts the motion results obtained from the sensor 180 into evaluation values ​​for the trial vector according to the hyperparameters and feeds them back to the parameter search unit 104.

[0123] The trial vector and its evaluation value set for the adjustment object 160 are input again into the model generation unit 205. The model generation unit 205 updates the evaluation value prediction model 206 through decision tree regression. In the parameter search unit 104, candidate vectors are repeatedly searched by using the updated evaluation value prediction model 206, thereby improving the accuracy of the evaluation value prediction model 206, which serves as a surrogate model for the adjustment object 160, while narrowing the candidate search range.

[0124] Furthermore, the trial vector and its evaluation value set for the adjustment object 160 are input to the determination unit 208 to determine whether a parameter set of m control parameters that achieves the desired performance (evaluation value) has been obtained. If the parameter search unit 104 obtains a parameter set that achieves the desired performance or has performed a predetermined number of trials, it ends the above search cycle and provides the optimal evaluation value 210 and the optimal vector (adjusted vector 209) at the end time to the adjustment item setting unit 102.

[0125] The parameter adjustment system 10 of this embodiment can display information set in the parameter adjustment device 100, trial history, parameter importance, and changes in predicted evaluation values ​​to the user via the UI unit 120. Therefore, the user can operate the search conditions while referring to (e.g., comparing) previously set parameters, and can efficiently adjust appropriate control parameters. Furthermore, by selecting vectors orthogonal within the search range and having the fewest possible combinations as the initial vector for generating the evaluation value prediction model 206, a wide range of search spaces can be modeled with fewer iterations. Moreover, by generating the evaluation value prediction model 206 through decision tree regression, parameter importance can be easily calculated, and the search range can be automatically adjusted based on this importance. Therefore, the increase in the number of searches due to searching ranges that do not contribute significantly to improving the evaluation value can be suppressed. Thus, the search and adjustment of control parameters can be performed efficiently. Furthermore, the user can easily grasp the process of searching or adjusting by viewing the displayed information and trial history. Therefore, even novice or inexperienced users can efficiently set appropriate control parameters.

[0126] <Data Structures for Action Conditions and Control Parameters>

[0127] Next, use Figure 4 and Figure 5 Explain the data structure for action conditions and control parameters.

[0128] Figure 4 and Figure 5 These are data structure diagrams illustrating an example of the motion parameters and control parameters of the adjustment object 160. In the following description, the case where the adjustment object 160 is a DC-DC converter will be used as an example.

[0129] Figure 4 The diagram illustrates k (k=2) action conditions. Each of the k action conditions contains n action parameters vn representing the characteristics of the action condition. A row vector Vk represents an action condition, containing n (n=3) action parameters as its elements, denoted as Vk=(v1,v2,…,vn).

[0130] For example, operating parameter v1 represents the input voltage of the DC-DC converter. Operating parameter v2 represents the output voltage of the DC-DC converter. Operating parameter v3 represents the output current of the DC-DC converter. Although only three operating parameters are shown here, in addition to... Figure 4 Beyond the example shown, various operating parameters can be set. These parameters refer to operating conditions that determine the action conditions, such as the duration of voltage and / or current fluctuations, and allowable overshoot and / or undershoot. The operating condition parameters for the initial vector, candidate vector, and trial vector are selected from these vectors.

[0131] Figure 5 The diagram shows m distinct control parameters xm. Additionally, the control parameter set X contains m control parameters xm as elements, denoted as X=(x1,x2,…,xm).

[0132] The value range of each control parameter is as follows: Figure 5 The chart shown is defined specifically using three levels: the minimum value Min (-1), the intermediate value Mid (0), and the maximum value Max (1) for each control parameter. Furthermore, when each control parameter is a discrete value, its resolution (step) can also be defined.

[0133] exist Figure 5 In the example shown, four parameters A, B, C, and D are defined as the control parameter xm. Parameter A is defined to take three values: a minimum value of "0", a median value of "10", and a maximum value of "20". Each parameter can have an independent range of values, or its range can be limited by other parameters. For example, the minimum value of parameter E (not shown) can be limited by the set value of parameter D. Furthermore, the scales of each parameter can also be independent. The control parameters are used to search and select the initial vector, candidate vector, and trial vector within these ranges.

[0134] The parameter adjustment system 10 searches for a set of m control parameters X that satisfy these multiple action conditions Vk.

[0135] Additionally, the value range (minimum and maximum) of each parameter can be set to the maximum range of the adjustment object 160, or it can be a range specified by the user within a narrower range. Alternatively, it can be a range used in past adjustment results, or a range adaptively set in the current search loop.

[0136] <Initial Vector>

[0137] Next, use Figures 6A to 6D This section provides examples of data related to the generation of the initial vector (initial parameter set). Furthermore, the following example illustrates the case where the number of control parameters m is 4.

[0138] Figure 6A This is a diagram showing the meeting matrix C (m=4). Figure 6B This is a diagram showing the DSD matrix (m=4). Figure 6C This is a diagram illustrating an example of the data structure of the initial parameter set (initial vector matrix). Figure 6D This is a graph showing examples of the evaluation values ​​for each parameter set.

[0139] Figure 6A The meeting matrix C shown is an m-row, m-column matrix with diagonal elements of 0 and off-diagonal elements of 1 or -1. It satisfies the following condition: C... T C = (m-1)I. Furthermore, I is the identity matrix. T This refers to the transpose. The meeting matrix C has the characteristic that all rows and columns are orthogonal.

[0140] Figure 6B The DSD matrix shown corresponds to Figure 3 The DSD matrix 203 shown contains a meeting matrix C and row vectors with elements of 0. Specifically, the DSD matrix is ​​as follows: [C, -C, 0] T This results in a 2m+1 row, m column matrix connected along the row direction. In deterministic screening design (DSD), a type of experimental design, a three-level experiment is conducted based on this matrix. This involves piloting an experiment that changes a single factor (parameter) to three levels: low, medium, and high. This allows for fewer experiments with less confounding effects (correlation).

[0141] Figure 6C The initial vector matrix shown is based on Figure 6B The DSD matrix shown is used for generation. Specifically, the initial vector generation unit 204 generates the following for each action condition Vk: Figure 6BThe DSD matrix shown is converted into a matrix of control parameters, and all action conditions are connected along the row direction.

[0142] The upper left part of matrix 301 is to be combined with Figure 4 The row vector (48, 12, 50) corresponding to action condition V1 in the action condition diagram shown is connected along the row direction to form a matrix of 2m+1 elements. Part of matrix 301 can be called the action parameter matrix.

[0143] The upper right matrix 302 is the matrix corresponding to the action parameter matrix of action condition V1, that is, it includes the values ​​of the m (m=4) control parameters that perform the action under action condition V1. Partial matrix 302 can be called the control parameter matrix. In the control parameter matrix, Figure 6B The columns of the DSD matrix shown correspond to the control parameters x1, x2, x3, and x4. Specifically, the element "-1" is assigned to the minimum value Min (-1) of each control parameter, the element "0" is assigned to the intermediate value Mid (0), and the element "1" is assigned to the maximum value Max (1). For example, Figure 6B The first row of the DSD matrix shown is (0, 1, 1, 1). Therefore Figure 6C The first row of the control parameter matrix (partial matrix 302) shown is (10, -500, 1.0, 9000). This corresponds to... Figure 5 The definition chart of the control parameters shows the intermediate value of control parameter A as "10", the maximum value of control parameter B as "-500", the maximum value of control parameter C as "1.0", and the maximum value of control parameter D as "9000".

[0144] Similarly, the lower left part of matrix 303 is to be compared with... Figure 4 The row vector (52, 15, 150) corresponding to action condition V2 in the action condition diagram shown is connected along the row direction to form a matrix of 2m+1 rows. Furthermore, the lower right partial matrix 304 is the control parameter matrix corresponding to the action parameter matrix of action condition V2, and is identical to partial matrix 302. Additionally, when the value range of the control parameters for each action condition is different, the two control parameter matrices (i.e., partial matrices 302 and 304) can be different matrices.

[0145] In this way, the (2m+1)×k×k matrix obtained by connecting all the parameter matrices along the row direction for k action conditions (2m+1) rows (n+m) columns is used as the initial vector matrix to generate.

[0146] In order to generate the evaluation value prediction model 206, Figure 6CThe row vectors of the initial vector matrix shown are set one by one on the adjustment object 160 and evaluated. Specifically, the initial vector generation unit 204 provides, for example, the first row (48, 12, 50, 10, -500, 1.0, 9000) of the initial vector matrix to the trial vector setting unit 106. The trial vector setting unit 106 sets the control parameter set (10, -500, 1.0, 9000) to the internal parameters of the adjustment object 160, and makes the adjustment object 160 move based on the action parameter set (48, 12, 50).

[0147] Sensor 180 detects the motion result of the adjustment object 160, and motion evaluation unit 108 generates an evaluation value based on the detected motion result. For example Figure 6D As shown, "12.4" is obtained as the evaluation value corresponding to the first row of the initial vector matrix. The model generation unit 205 connects the evaluation values ​​obtained for each row of the initial vector matrix along the row direction (vertical direction). By repeatedly performing this process on all row vectors of the initial vector matrix, the following is obtained: Figure 6D The evaluation value matrix (also known as the evaluation value vector) is shown as having (2m+1)×k rows.

[0148] The model generation unit 205 inputs the adjustment object 160. Figure 6C In the initial vector matrix shown, the prediction will be... Figure 6D The regression model of the evaluation value vector shown is generated as evaluation value prediction model 206. The generation of the regression model is also called the learning or training of the regression model.

[0149] Additionally, in the search step (described later) Figure 7 The candidate vectors and trial vectors used in S20 to S40 are all the same as those in S20 to S40. Figure 6C The row vectors of the initial vector matrix shown have the same form (specifically, each element has action parameters and control parameters). The trial vectors input to the adjustment object 160 and the output evaluation values ​​are connected to the initial vector matrix and the evaluation value matrix along the row direction (vertical direction) at each trial. The evaluation value prediction model 206 is updated (or regenerated, relearned, or retrained) at each connection, thereby improving prediction accuracy.

[0150] By using such an initial vector matrix to generate the evaluation value prediction model 206, it is possible to start the search from the state of the prediction model that captures a wide range of features including minimum, intermediate, and maximum values ​​in the search space of control parameters constrained by action parameters.

[0151] <Actions of the parameter adjustment system>

[0152] Next, refer to Figure 7 Explain the action of parameter adjustment system 10. Figure 7This is a flowchart illustrating an example of the operation of the parameter adjustment system 10 according to an embodiment of the present disclosure.

[0153] First, in step S10, the parameter adjustment system 10 uses m parameters to generate at least 2m+1 parameter sets based on the DSD method, and obtains evaluation values ​​for each. Specifically, the initial vector generation unit 204 of the parameter adjustment device 100 generates an initial parameter set (initial vector) for generating the evaluation value prediction model 206. The parameter adjustment device 100, for the m parameters of the adjustment object 160, uses the DSD method to calculate a combination of experiments at three levels (specifically, minimum, median, and maximum values), thereby generating 2m+1 control parameter sets. The parameter adjustment device 100 uses the generated control parameter sets as the initial parameter sets and obtains evaluation values ​​for the adjustment object 160 k times, where k is the number of action conditions. Thus, the following is obtained: Figure 6C The initial vector matrix shown is Figure 6D The evaluation value matrix shown.

[0154] In step S20, the parameter adjustment system 10 generates an evaluation value regression model using decision tree regression. Specifically, the model generation unit 205 of the parameter adjustment device 100 generates or updates an evaluation value prediction model 206 based on a combination of previously set vectors (action parameters and control parameters) and obtained evaluation values. The evaluation value prediction model 206 is generated using decision tree regression. As an example of decision tree regression, random forest regression is used.

[0155] In step S30, the parameter adjustment system 10 searches for a trial parameter set based on the parameter importance of the evaluation value regression model. An example of a specific search process is provided using... Figure 8 Please provide an explanation. Figure 8 This is a flowchart illustrating an example of the parameter search process (S30) performed by the parameter adjustment system 10 according to this embodiment.

[0156] like Figure 8 As shown, in step S32, the candidate vector generation unit 207 first extracts the parameter set corresponding to one of the at least 2m+1 evaluation values ​​obtained in step S10 as a baseline parameter set. The baseline parameter set is a set of parameters used to generate candidate vectors for the search using the evaluation value prediction model 206. Specifically, the candidate vector generation unit 207 extracts the parameter set corresponding to the optimal evaluation value among the at least 2m+1 evaluation values ​​as a baseline parameter set. More specifically, the candidate vector generation unit 207 extracts the parameter set of the vectors from the vectors input to the evaluation value prediction model 206 that contain action conditions with similar tendencies to the evaluation value output, and outputs the vectors containing the predicted values ​​near the optimal evaluation value among these action conditions as a baseline parameter set.

[0157] Next, in step S34, the candidate vector generation unit 207 generates one or more candidate parameter sets based on the baseline parameter set and parameter importance. These one or more candidate parameter sets are the candidate vectors being searched. Specifically, the candidate vector generation unit 207 generates one or more random parameter sets (i.e., candidate vectors) that follow a normal distribution with the baseline parameter set as the mean and the parameter importance calculated based on the evaluation value prediction model 206 using decision tree regression as the variance.

[0158] Next, in step S36, the candidate vector generation unit 207 uses the evaluation value regression model (specifically the evaluation value prediction model 206) to predict the predicted values ​​of the evaluation values ​​of one or more candidate parameter sets.

[0159] Next, in step S38, the candidate vector generation unit 207 determines the trial parameter set as the set of candidate parameters corresponding to the predicted values ​​that are better than the evaluation values ​​corresponding to the benchmark parameter set from one or more candidate parameter sets. The trial parameter set can be one or more.

[0160] return Figure 7 In step S40, the parameter adjustment system 10 acquires the evaluation value of the trial parameter set. Specifically, the trial vector setting unit 106 sets a trial vector, which combines the determined trial parameter set with the motion parameters, onto the adjustment object 160 and causes it to move. The sensor 180 detects the motion result of the adjustment object 160, and the motion evaluation unit 108 calculates the evaluation value based on the detected motion result. The parameter search unit 104 acquires the evaluation value calculated by the motion evaluation unit 108.

[0161] Next, in step S50, the parameter adjustment system 10 determines whether to end the parameter adjustment. Specifically, the determination unit 208 ends the parameter adjustment if at least one of the predetermined conditions is met ("Yes" in S50). The determination unit 208 outputs the optimal evaluation value 210 at the end time and the trial vector corresponding to the evaluation value 210 as the adjusted vector 209.

[0162] Here, one or more conditions include, for example, reaching a predetermined number of repetitions, and at least one of the following: the evaluation value of the trial parameter set is better than a predetermined threshold. The number of repetitions corresponds to the number of times a "no" decision is made in step S50, specifically the number of times the processing steps S20 to S40 are performed. Furthermore, the predetermined threshold used in comparing the evaluation values ​​can be, for example, a value predetermined by the user, a newly obtained evaluation value, or the optimal evaluation value among multiple evaluation values ​​obtained after a predetermined number of repetitions.

[0163] If any of the predetermined conditions is not met ("No" in S50), the parameter adjustment system 10 continues without ending the parameter adjustment process. Specifically, it returns to step S20, where the model generation unit 205 updates the evaluation value prediction model 206 based on the evaluation values ​​of the trial parameter set, and repeatedly searches and evaluates (S30, S40). Specifically, it repeatedly executes step S40 of obtaining evaluation values, step S20 of updating the evaluation value prediction model 206, and step S30 of searching until the number of repetitions reaches a predetermined number or until the evaluation values ​​of the trial parameter set are better than a predetermined threshold.

[0164] Display Processing

[0165] Next, the display processing performed by the parameter adjustment system 10 will be explained.

[0166] The parameter adjustment system 10 of this embodiment displays the parameter importance of the evaluation value prediction model 206 and the prediction curve of the evaluation value predicted based on the evaluation value prediction model 206 for each of the m parameters. Furthermore, the parameter adjustment system 10 can also display the parameters used for evaluation value prediction and their decision thresholds.

[0167] The following examples show specific display screens while using... Figure 9 and Figure 10 This illustrates an example of the processing. Figure 9 This is a flowchart illustrating an example of the display processing of the parameter adjustment system 10 according to an embodiment of the present disclosure. Figure 10 This is a diagram showing an example of a display screen 400 displayed on the UI unit 120 according to an embodiment of the present disclosure.

[0168] like Figure 9 As shown, in step S60, the parameter adjustment system 10 displays a display screen 400 on the UI unit 120. This display screen is used to input and output data and information related to parameter adjustment. Figure 10 The display screen 400 shown is not only a screen that prompts the user with the results of parameter adjustments, but also a UI screen that can accept user input.

[0169] like Figure 10 As shown, the display screen 400 includes a tree-shaped display area 410 for the evaluation value prediction model 206, a judgment threshold display area 420 for the evaluation value prediction model 206, a parameter importance display area 430, and a parameter prediction value display area 440.

[0170] The initial state of the values ​​and graphics displayed in each display area is determined based on the action conditions and control parameter values ​​set on the adjustment object 160 at the time of display, and the adjustment results and evaluation value prediction model 206 maintained internally up to the time of display. The user can change the action conditions and / or control parameter values ​​by manipulating each GUI object. According to this change, the evaluation value is re-predicted, and the GUI objects related to the predicted evaluation value are redrawn. Furthermore, GUI objects are objects that can be manipulated by pointing, clicking, or dragging, such as text, markers, cursors, graphic bars, or graphic curves. Additionally, parameters displayed can be set on the adjustment object 160 and evaluated using a menu (not shown) to reflect the results, or a new search can be performed.

[0171] In the tree-shaped display area 410, the structure of the evaluation value prediction model 206 based on decision tree regression is displayed. For example, a schematic diagram showing the decision tree related to the determination of the evaluation value predicted based on the action conditions and control parameters selected at the display time is displayed, as well as its determination path (e.g., Figure 10 (Thick solid line). Additionally, in cases where random forests or GBDTs are used in decision tree regression, multiple decision trees are used to predict evaluation values. In this case, only a portion of the multiple decision trees can be displayed in the tree display area 410, or a set of multiple leaves outputting predicted values ​​can be shown. Accordingly, the user can visually confirm the number of decision trees and the number of layers, etc., which are hyperparameters of the evaluation value prediction model 206.

[0172] In the judgment threshold display area 420, parameters related to the predicted value of the evaluation value predicted using the decision tree regression-based evaluation value prediction model 206 and their judgment thresholds are displayed. For example, information such as which parameter is related to the judgment of the evaluation value predicted based on the action conditions and parameters selected at the time of display, what the judgment threshold was at that time, and what value the evaluation value was thus judged to be, can be displayed. Figure 10 In the example shown, uppercase letters such as "A" and "B" represent parameters. Furthermore, combinations of inequality signs with numerical values, such as "<18.2" and "<-1200", represent decision conditions and decision thresholds. Additionally, the numerical value "0.03" following "value:" represents the evaluation value. Thus, users can confirm the relationship between parameter settings and thresholds. Alternatively, users can set a portion of these thresholds to impose restrictions on the prediction of evaluation values.

[0173] The parameter importance display area 430 shows the parameter importance of the evaluation value prediction model 206 based on decision tree regression. For example, in Figure 10The example shown represents a case where the number of parameters m=10. Here, a bar chart is used to display the importance of the 10 parameters A to J in action condition V2. The longer the bar, the higher the parameter's importance. Figure 10 The display screen 400 shows that parameter B has high importance, while parameters D and H have low importance. Furthermore, it demonstrates a parameter search performed based on these importance levels. Here, the parameter search unit 104 can also be provided with information about the importance of parameters based on the user's domain expertise. For example, if the user wants to increase the importance of parameter C, they can drag the bar for parameter C to the right to increase its importance and perform a re-search.

[0174] The parameter prediction value display area 440 displays the parameter values ​​and prediction curves set at the current time. For example, in Figure 10 The example shown illustrates a scenario with parameter number m=10. Here, marker 442 displays on the x-axis (horizontal axis) the values ​​of the 10 parameters A to J in action condition V2, normalized to a range of 0.0 to 1.0. Marker 442 superimposed on the prediction curve indicates the position of the corresponding parameter value. Marker 442 indicates the parameter value at the time of display.

[0175] Furthermore, by connecting each marker 442 with dashed lines 444, the set of control parameters (row vectors) is visually displayed using the shape of the dashed lines 444. For example, by using solid lines to represent the set of control parameters that become the optimal evaluation value in the search up to the time of display, it is easy to compare them with the set of parameters set at the time of display.

[0176] The y-axis (vertical axis) of the graph for each parameter represents the predicted value of the evaluation. The prediction curve shows the change in the predicted value of the evaluation when other parameters are fixed and the parameter is scanned. For example, the prediction curve for parameter A represents the predicted value of the evaluation obtained by scanning parameter A from its minimum to its maximum value, with other parameters B to J fixed at their respective displayed marker positions. These prediction curves have a stepped shape because they employ thresholding using decision tree regression.

[0177] Additionally, marker 442 can be moved based on manual user input or a pre-determined procedure. For example... Figure 9 As shown, when parameter marker 442 is moved ("Yes" in step S62), m prediction curves are updated (step S64). Whenever a parameter value changes, all prediction curves are recalculated and redrawn. Thus, a multi-dimensional parameter space can be represented by changes in one-dimensional curves, allowing users to visually confirm how the evaluation value changes when a parameter is moved, and how the relationships between parameters change.

[0178] <Effects, etc.>

[0179] In this way, the parameter adjustment system 10 can display to the user the information set in the parameter adjustment device 100, trial history, parameter importance, and changes in the predicted evaluation value. Therefore, the user can operate the search conditions while referring to (e.g., comparing) previously set parameters, and can efficiently set appropriate control parameters. Furthermore, by selecting an orthogonal vector with the fewest possible combinations within the search range as the initial vector for generating the evaluation value prediction model 206, a wide range of search spaces can be modeled with fewer iterations. Moreover, by generating the evaluation value prediction model 206 through decision tree regression, the parameter importance can be easily calculated, and the search range can be automatically adjusted based on this importance. Therefore, the increase in the number of searches due to searching ranges that do not contribute significantly to improving the evaluation value can be suppressed. Thus, the search and adjustment of control parameters can be performed efficiently. Furthermore, by visually confirming the displayed information, trial history, etc., the user can easily grasp the process of searching or adjusting. Therefore, even novice or inexperienced users can efficiently set appropriate control parameters.

[0180] (other)

[0181] The embodiments have been described above with reference to the accompanying drawings, but this disclosure is not limited to the examples described above. Those skilled in the art will be able to conceive of various modifications or alterations within the scope described in the technical solutions. Such modifications or alterations should also be understood to fall within the technical scope of this disclosure. Furthermore, the constituent elements in the embodiments can be arbitrarily combined without departing from the spirit of this disclosure.

[0182] In the above embodiments, one or more of the functional units (processing units) can be integrated as needed, and one functional unit (processing unit) can be divided into multiple sub-functional units (sub-processing units).

[0183] The order of steps in the flowcharts described in the above embodiments is merely an example and is not limited to the order shown in the figures.

[0184] The items recorded in the above embodiments can be appropriately combined as long as they are not contradictory or are not explicitly stated as not being able to be combined.

[0185] Furthermore, the term "...part" in the above embodiments can be replaced with other terms, such as "...circuitry", "...assembly", "...device", "...unit" or "...module".

[0186] This disclosure can be implemented by software, hardware, or software that works with hardware.

[0187] The functional blocks used in the above embodiments are implemented, in part or in whole, as integrated circuits, i.e., LSIs (Large-Scale Integrated Circuits). The processes described in the above embodiments can be controlled, in part or in whole, by a single LSI or a combination of LSIs. An LSI can be composed of individual chips, or it can be composed of a single chip containing some or all of the functional blocks. An LSI can have data input and output sections. Depending on its integration level, an LSI is sometimes also called an IC, a system LSI, a very large LSI, or an extra-large LSI.

[0188] The method of integrating LSIs is not limited to LSIs; it can also be implemented using dedicated circuits, general-purpose processors, or special-purpose processors. Furthermore, LSIs can be used to fabricate post-programmable FPGAs (Field Programmable Gate Arrays), or reconfigurable processors that allow reconfiguration of the connections and / or settings of the circuitry within the LSI. This disclosure can be implemented as digital or analog processing.

[0189] Furthermore, with advancements in semiconductor technology or other derived technologies, when integrated circuit technologies emerge to replace LSIs, these technologies can also be used for the integration of functional blocks. Applications in biotechnology are also possible.

[0190] This disclosure can be implemented in all kinds of devices, apparatuses, and systems with communication capabilities (collectively, communication devices). A communication device may include a wireless transceiver and processing / control circuitry. A wireless transceiver may include a receiving unit and a transmitting unit, or may include them as a function. The wireless transceiver (transmitting unit, receiving unit) may include an RF (Radio Frequency) module and one or more antennas. The RF module may include an amplifier, an RF modulator / demodulator, or similar devices. Non-limiting examples of communication devices include telephones (portable phones, smartphones, etc.), tablet computers, personal computers (PCs) (laptops, desktops, laptops, etc.), cameras (digital still / video cameras, etc.), digital media players (digital audio / video players, etc.), wearable devices (wearable cameras, smartwatches, tracking devices, etc.), game consoles, digital book readers, telemedicine / telemedicine (telemedicine prescription) devices, vehicles or mobile transportation vehicles with communication capabilities (cars, airplanes, ships, etc.), and combinations of the various devices described above.

[0191] Communication devices are not limited to portable or mobile devices, but also include all kinds of devices, equipment, and systems that are difficult to carry or cannot be carried due to being fixed, such as smart home devices (home appliances, lighting equipment, smart meters or measuring devices, control panels, etc.), vending machines, and all other "things" that may exist on the IoT (Internet of Things) network.

[0192] Communication includes not only data communication via cellular systems, wireless LAN systems, and communication satellite systems, but also data communication via a combination of these.

[0193] Furthermore, the communication device also includes devices such as controllers or sensors connected to or linked to a communication device that performs the communication functions described in this disclosure. For example, this includes controllers or sensors that generate control signals or data signals used by the communication device to perform the communication functions of the communication device.

[0194] In addition, communication devices include infrastructure equipment for communicating with or controlling the various devices described above (not limited to those mentioned above), such as base stations, access points, all other devices, equipment, and systems.

[0195] The specific examples of this disclosure have been described in detail above, but these are merely illustrative and do not limit the technical solutions. The technology described in the technical solutions includes various modifications and alterations to the specific examples illustrated above.

[0196] Industrial availability

[0197] This disclosure can be used as a parameter adjustment method for adjusting the parameters of a control system, and is useful for, for example, for the management methods of a control system.

[0198] Explanation of reference numerals in the attached figures

[0199] 10 Parameter Adjustment System

[0200] 100 Parameter Adjustment Device

[0201] 102 Adjustment Project Setting Department

[0202] 104 Parameter Search Department

[0203] 106 Trial Vector Setting Department

[0204] 108 Motion Evaluation Department

[0205] 110 Storage device

[0206] 112 Processing Unit

[0207] 114 UI Devices

[0208] 116 Communication devices

[0209] 118 bus

[0210] 120 User Interface (UI) Department

[0211] 140 Database (DB) Department

[0212] 160 Adjustment Object

[0213] 180 sensor

[0214] 201 Action Conditions

[0215] 202 Adjustment Project

[0216] 203 DSD matrix

[0217] 204 Initial Vector Generation Section

[0218] 205 Model Generation Department

[0219] 206 Evaluation Value Prediction Model

[0220] 207 Candidate Vector Generation Unit

[0221] 208 Judgment Department

[0222] 209 Adjusted Vector

[0223] 210 rating

[0224] Partial matrices 301, 302, 303, and 304

[0225] 400 display screen

[0226] 410 Tree-shaped display area

[0227] 420 Threshold Display Area

[0228] 430 Parameter Importance Display Area

[0229] 440 Parameter Prediction Value Display Area

[0230] 442 mark

[0231] 444 Dashed line.

Claims

1. A parameter adjustment method, comprising the following steps: The steps to obtain evaluation values ​​are as follows: using m parameters of the control system, at least 2m+1 parameter sets are generated based on the deterministic screening design method, and the evaluation values ​​of each of the generated at least 2m+1 parameter sets are obtained, where m is a natural number. The steps to generate the evaluation value regression model are as follows: based on a set of 2m+1 parameters and the corresponding evaluation values, the evaluation value regression model is generated through decision tree regression. as well as The search step involves searching for a trial parameter set for the control system based on the parameter importance of the evaluation value regression model.

2. The parameter adjustment method as described in claim 1, The search steps include: The extraction step involves using the parameter set corresponding to one of the at least 2m+1 evaluation values ​​obtained in the step of obtaining evaluation values ​​as the benchmark parameter set. The step of generating more than one candidate parameter set involves generating more than one candidate parameter set based on the parameter importance and the extracted benchmark parameter set. The prediction step involves using the evaluation value regression model to predict the predicted values ​​of the evaluation values ​​for each of the one or more candidate parameter sets. as well as The decision-making process involves selecting the candidate parameter set corresponding to the predicted value that is better than the evaluation value corresponding to the baseline parameter set from among more than one predicted value, and using this set as the trial parameter set.

3. The parameter adjustment method as described in claim 2, In the extraction step, the parameter set corresponding to the best evaluation value among the at least 2m+1 evaluation values ​​is extracted as the benchmark parameter set.

4. The parameter adjustment method as described in claim 2, In the step of generating one or more candidate parameter sets, one or more random parameter sets that follow a normal distribution are generated as the one or more candidate parameter sets, wherein the normal distribution has the average value of the reference parameter set and the variance of the parameter importance.

5. The parameter adjustment method as described in claim 1, In the step of obtaining the evaluation value, the evaluation value of the trial parameter set determined in the search step is also obtained. The step of generating the evaluation value regression model includes a step of updating the evaluation value regression model based on the evaluation values ​​of the trial parameter set.

6. The parameter adjustment method as described in claim 5, The steps of obtaining the evaluation value, updating the regression model of the evaluation value, and searching are repeatedly performed until the number of repetitions reaches a predetermined number or until the evaluation value of the trial parameter set is better than a predetermined threshold.

7. The parameter adjustment method as described in any one of claims 1 to 6, The parameter adjustment method further includes a display step, in which, for each of the m parameters, the importance of the parameter in the evaluation value regression model and the prediction curve of the evaluation value predicted based on the evaluation value regression model are displayed.

8. The parameter adjustment method as described in claim 7, The displayed step also includes a marker that overlaps the predicted curve to indicate the position of the corresponding parameter value. When the marker moves, the m predicted curves are updated.

9. The parameter adjustment method as described in claim 8, The parameter adjustment method includes the step of accepting a user's movement operation on the marker.

10. The parameter adjustment method as described in claim 7, The steps shown also include the decision tree used in generating the evaluation value regression model.

11. A program for causing a computer to perform the parameter adjustment method according to any one of claims 1 to 6.

12. A parameter adjustment device, comprising: The acquisition unit uses m parameters of the control system to generate at least 2m+1 parameter sets based on the deterministic screening design method, and acquires the evaluation values ​​of each of the generated at least 2m+1 parameter sets, where m is a natural number; The generation department generates an evaluation value regression model based on a set of 2m+1 parameters and their corresponding evaluation values ​​through decision tree regression. as well as The search unit searches for a trial parameter set for the control system based on the parameter importance of the evaluation value regression model.

Images