Parameter Adjustment Device and Parameter Adjustment Method

Abstract

An evaluation value calculation device (2) that obtains the operation result of an adjustment target device (1) that performs operation using one parameter or a plurality of parameters, and calculates an evaluation value for the operation result, acquires data including the value of each parameter and the evaluation value related to the value of each parameter. A data acquisition unit (11) and an elite solution extraction unit (12) that extracts zero or more data as elite solutions from among a plurality of data acquired by the data acquisition unit (11) based on the evaluation values included in each data are provided. The parameter adjustment device (3) is configured. Further, the parameter adjustment device (3) determines the value of the parameter to be used by the adjustment target device (1) in the next operation based on the elite solution extracted by the elite solution extraction unit (12) from among a plurality of parameter values existing in the search space of the parameter values, and includes a parameter value determination unit (14) that outputs the determined parameter value to the adjustment target device (1).

Description

Technical Field

[0001] The present disclosure relates to a parameter adjustment device and a parameter adjustment method.

Background Art

[0002] There is a parameter adjustment method for searching for a plurality of parameters used in the operation of a device to be adjusted. As such a parameter adjustment method, for example, Non-Patent Document 1 discloses a method in which a plurality of optimizers search for parameters. The method disclosed in Non-Patent Document 1 has a reception function for receiving a designation of the number of optimizers. Each optimizer is software or the like for setting one parameter or a plurality of parameters that can be used in the operation of the device to be adjusted.

Prior Art Documents

Non-Patent Documents

[0003]

Non-Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0004] In the method disclosed in Non-Patent Document 1, when setting the value of a parameter (hereinafter referred to as "parameter value"), if the appropriate number of parameter values is unknown, there is a problem that it is difficult for the user to specify an appropriate number as the number of optimizers to be executed. The appropriate number is a number that can find many parameter values that can obtain diverse and high evaluations within the allowable range of deterioration of the search efficiency, which is the efficiency of searching for appropriate parameter values. If the specified number of optimizers is too small compared to the appropriate number, the diversity of parameter values is lost. If the specified number of optimizers is too large compared to the appropriate number, the search efficiency of the parameters deteriorates.

[0005] The present disclosure has been made to solve the above problems, and an object thereof is to obtain a parameter adjustment device capable of searching for diverse parameter values within the allowable range of deterioration of search efficiency even when the appropriate number of parameter values to be found is unknown.

Means for Solving the Problems

[0006] The parameter adjustment device according to the present disclosure acquires the operation result of an adjustment target device that performs an operation using one parameter or a plurality of parameters, and from an evaluation value calculation device that calculates an evaluation value for the operation result, data including the value of each parameter and the evaluation value related to the value of each parameter is obtained. A data acquisition unit, and an elite solution extraction unit that extracts zero or more data as elite solutions based on the evaluation values included in each data from among the plurality of data acquired by the data acquisition unit. Further, the parameter adjustment device determines the value of the parameter to be used by the device to be adjusted in the next operation based on the elite solution extracted by the elite solution extraction unit from among the plurality of parameter values existing in the search space of the parameter value, and a parameter value determination unit that outputs the determined parameter value to the device to be adjusted The parameter value determination unit includes a determination method selection unit that selects either a global search method or a local search method as a method for determining the values of the parameters to be used in the next operation based on the elite solutions extracted by the elite solution extraction unit. If the global search method is selected by the determination method selection unit, it samples the values of one or more parameters according to a probability distribution that uniformly spreads over the entire search space. If the local search method is selected by the determination method selection unit, it samples the values of one or more parameters according to a probability distribution centered on a certain elite solution in the search space. A parameter value determination processing unit calculates the distances between the values of the respective parameters sampled by the sampling unit and the values of the parameters included in the elite solutions extracted by the elite solution extraction unit, and determines the values of the parameters to be used by the device under adjustment in the next operation from among the values of one or more parameters sampled by the sampling unit based on the distances. is provided.

Effects of the Invention

[0007] According to the present disclosure, even if the appropriate number of values of the parameters to be found is unknown, it is possible to search for diverse parameter values within an acceptable range of deterioration in search efficiency.

Brief Description of the Drawings

[0008]

Figure 1

Figure 2

Figure 3

Figure 4

Figure 5

Figure 6

Figure 7

Figure 8

Figure 9

Figure 10

Figure 11

Figure 12

Modes for Carrying Out the Invention

[0009] Hereinafter, in order to explain the present disclosure in more detail, embodiments for implementing the present disclosure will be described with reference to the accompanying drawings.

[0010] Embodiment 1. By finding a combination of values of a plurality of parameters having diversity (hereinafter referred to as "parameter values") at once, development efficiency and the like may be improved. For example, in order to find a simulation scenario in which a vehicle control system under development falls into an unsafe state, a case where the parameter values of scenario parameters are automatically adjusted with the degree of danger as an evaluation value can be considered. At this time, development can be made more efficient by automatically extracting as diverse and dangerous scenarios as possible. In addition, there may be a case where there is an evaluation item that can be evaluated only after parameter adjustment, or a constraint condition or the like. For example, after adjusting the parameter values of control parameters in advance using a simulator of a mechanical device, and then controlling the actual device using the extracted control parameter values, it is a case of checking whether a defect that cannot be reproduced in the simulation occurs. If only one extracted parameter value is used and a defect occurs in the actual device, it is necessary to start over from the automatic adjustment using the simulator. However, if a combination of a plurality of diverse parameter values is extracted, the possibility that any combination of parameter values passes the test by the actual device increases, and as a result, the development efficiency is improved. Even if the appropriate number of parameter values to be found is unknown, a parameter adjustment device 3 that can search for combinations of a plurality of diverse parameter values within an allowable range of deterioration in search efficiency will be described.

[0011] Hereinafter, in order to explain the present disclosure in more detail, embodiments for implementing the present disclosure will be described with reference to the accompanying drawings. FIG. 1 is a configuration diagram showing a system including a parameter adjustment device 3 according to Embodiment 1. FIG. 2 is a hardware configuration diagram showing the hardware of the parameter adjustment device 3 according to Embodiment 1. The system shown in FIG. 1 includes an apparatus to be adjusted 1, an evaluation value calculation apparatus 2, and a parameter adjustment apparatus 3.

[0012] The apparatus to be adjusted 1 performs operation using one parameter or a plurality of parameters, and outputs the operation result to the evaluation value calculation apparatus 2. Examples of the apparatus to be adjusted 1 include a simulator for an automatic driving system or a simulator for an air conditioner. If the apparatus to be adjusted 1 is, for example, a simulator for an air conditioner, examples of the parameters used for the operation of the apparatus to be adjusted 1 include an operation pattern of the frequency in a compressor or an operation pattern of an expansion valve.

[0013] The evaluation value calculation apparatus 2 is an evaluation value calculation apparatus that acquires the operation result of the apparatus to be adjusted 1 related to each parameter value from the apparatus to be adjusted 1 that has performed operation using the parameter value of each parameter, and calculates an evaluation value for the operation result. If the parameter used for the operation of the apparatus to be adjusted 1 is, for example, an operation pattern of the frequency in a compressor, as an evaluation value for the operation result, for example, based on the environmental conditions of the air conditioner set in advance, it is evaluated how long it takes for the temperature of the space to be air-conditioned by the air conditioner to reach the set temperature. The shorter the time until the temperature of the space to be air-conditioned reaches the set temperature, the higher the evaluation value for the operation result. The evaluation value calculation apparatus 2 outputs data including each parameter value and the evaluation value related to each parameter value to the parameter adjustment apparatus 3. In the system shown in FIG. 1, the evaluation value calculation apparatus 2 is provided outside the parameter adjustment apparatus 3. However, this is only an example, and the evaluation value calculation apparatus 2 may be provided inside the parameter adjustment apparatus 3.

[0014] The parameter adjustment apparatus 3 includes a dataset storage unit 10, a data acquisition unit 11, an elite solution extraction unit 12, an evaluation value prediction unit 13, a learning model 13a, and a parameter value determination unit 14. The parameter adjustment device 3 is a device that searches for a plurality of parameter values used for the operation of the device to be adjusted 1. The dataset storage unit 10 is realized by, for example, the dataset storage circuit 20 shown in FIG. 2. The dataset storage unit 10 acquires data including each parameter value and the evaluation value related to each parameter value from the evaluation value calculation device 2, and stores, as a dataset, all the data acquired in the past combined. The dataset includes one or more pieces of data.

[0015] The data acquisition unit 11 is realized by, for example, the data acquisition circuit 21 shown in FIG. 2. The data acquisition unit 11 acquires a dataset from the dataset storage unit 10. The data acquisition unit 11 outputs the dataset to each of the elite solution extraction unit 12 and the evaluation value prediction unit 13.

[0016] The elite solution extraction unit 12 is realized by, for example, the elite solution extraction circuit 22 shown in FIG. 2. The elite solution extraction unit 12 acquires a dataset from the data acquisition unit 11. The elite solution extraction unit 12 extracts zero or one or more pieces of data as elite solutions based on the evaluation values included in each piece of data in the dataset. Specifically, the elite solution extraction unit 12 extracts, as elite solutions, the data in the dataset whose included evaluation value is equal to or higher than the lower limit value. The elite solution extraction unit 12 outputs zero or more elite solutions to the parameter value determination unit 14.

[0017] The evaluation value prediction unit 13 is realized by, for example, the evaluation value prediction circuit 23 shown in FIG. 2. The evaluation value prediction unit 13 predicts the evaluation value of the operation result of the device to be adjusted 1 when the device to be adjusted 1 operates using each parameter value sampled by the sampling unit 14b (described later) of the parameter value determination unit 14. Specifically, when the evaluation value prediction unit 13 receives a parameter value from a parameter value determination processing unit 14c (to be described later) of the parameter value determination unit 14, it gives the parameter value to the learning model 13a, and obtains at least an evaluation value corresponding to the parameter value as a prediction result of the evaluation value for the operation result from the learning model 13a. The evaluation value prediction unit 13 outputs the prediction result of the evaluation value to the parameter value determination processing unit 14c.

[0018] The learning model 13a is realized, for example, by a Gaussian process regression model, a linear regression model, a neural network, a decision tree, a random forest, or a gradient boosting tree. During learning, when the learning model 13a is given the parameter value included in each data acquired by the data acquisition unit 11 and the evaluation value included in each data, it learns the evaluation value corresponding to each parameter value. The evaluation value is teacher data. During inference, when the learning model 13a is given a parameter value from the evaluation value prediction unit 13, it outputs, as a prediction result of the evaluation value, the evaluation value corresponding to the parameter value to the evaluation value prediction unit 13. The parameter adjustment device 3 shown in FIG. 1 incorporates the learning model 13a. However, this is only an example, and the learning model 13a may be provided outside the parameter adjustment device 3.

[0019] The parameter value determination unit 14 is realized, for example, by a parameter value determination circuit 24 shown in FIG. 2. The parameter value determination unit 14 includes a determination method selection unit 14a, a sampling unit 14b, and a parameter value determination processing unit 14c. The parameter value determination unit 14 determines a parameter value to be used by the device under adjustment 1 for the next operation based on the elite solution extracted by the elite solution extraction unit 12 from among a plurality of parameter values existing in the search space of the parameter values. The parameter value determination unit 14 outputs the determined parameter value to the device under adjustment 1.

[0020] The decision method selection unit 14a obtains zero or more elite solutions from the elite solution extraction unit 12. Based on the number of elite solutions extracted by the elite solution extraction unit 12, the decision method selection unit 14a selects either a global search method or a local search method as the method for determining the parameter values to be used in the next operation. The global search method is a search method that samples parameter values according to a probability distribution that uniformly spreads over the entire search space of the parameter values. The local search method is a search method that samples parameter values according to a probability distribution centered on a certain elite solution from the search space of the parameter values. Specifically, if the number of elite solutions extracted by the elite solution extraction unit 12 is 0, the decision method selection unit 14a selects the global search method. If the number of elite solutions extracted by the elite solution extraction unit 12 is 1 or more, the decision method selection unit 14a selects either the global search method or the local search method based on the probability of selecting the local search method. The probability of selecting the local search method is calculated by the decision method selection unit 14a based on the evaluation values included in the elite solutions extracted by the elite solution extraction unit 12.

[0021] If the global search method is selected by the decision method selection unit 14a, the sampling unit 14b samples one or more parameter values according to a probability distribution that uniformly spreads over the entire search space of the parameter values. If the local search method is selected by the decision method selection unit 14a, the sampling unit 14b samples one or more parameter values according to a probability distribution centered on a certain elite solution from the search space of the parameter values. The sampling unit 14b outputs the sampled parameter values to the parameter value determination processing unit 14c.

[0022] The parameter value determination processing unit 14c obtains zero or one or more elite solutions from the elite solution extraction unit 12 and obtains one or more parameter values from the sampling unit 14b. Further, the parameter value determination processing unit 14c acquires the prediction result of the evaluation value from the evaluation value prediction unit 13. The parameter value determination processing unit 14c calculates the distance between each of the acquired parameter values and the parameter values included in each of the acquired elite solutions. Based on the calculated distance and the prediction result of the evaluation value, the parameter value determination processing unit 14c determines one parameter value to be used by the device under adjustment 1 for the next operation from among one or more parameter values sampled by the sampling unit 14b.

[0023] In FIG. 1, it is assumed that each of the dataset storage unit 10, data acquisition unit 11, elite solution extraction unit 12, evaluation value prediction unit 13, and parameter value determination unit 14, which are components of the parameter adjustment device 3, is realized by dedicated hardware as shown in FIG. 2. That is, it is assumed that the parameter adjustment device 3 is realized by a dataset storage circuit 20, a data acquisition circuit 21, an elite solution extraction circuit 22, an evaluation value prediction circuit 23, and a parameter value determination circuit 24. Here, the dataset storage circuit 20 corresponds to, for example, a non-volatile or volatile semiconductor memory such as a RAM (Random Access Memory), ROM (Read Only Memory), flash memory, EPROM (Erasable Programmable Read Only Memory), EEPROM (Electrically Erasable Programmable Read Only Memory), a magnetic disk, a flexible disk, an optical disk, a compact disk, a mini disk, or a DVD (Digital Versatile Disc). Each of the data acquisition circuit 21, the elite solution extraction circuit 22, the evaluation value prediction circuit 23, and the parameter value determination circuit 24 corresponds to, for example, a single circuit, a composite circuit, a programmed processor, a parallel-programmed processor, an ASIC (Application Specific Integrated Circuit), an FPGA (Field-Programmable Gate Array), or a combination thereof.

[0024] The components of the parameter adjustment device 3 are not limited to those realized by dedicated hardware, and the parameter adjustment device 3 may be realized by software, firmware, or a combination of software and firmware. Software or firmware is stored in the memory of a computer as a program. The computer means the hardware that executes the program, and for example, a CPU (Central Processing Unit), a GPU (Graphics Processing Unit), a central processing unit, a processing unit, an arithmetic unit, a microprocessor, a microcomputer, a processor, or a DSP (Digital Signal Processor) corresponds thereto.

[0025] FIG. 3 is a hardware configuration diagram of a computer when the parameter adjustment device 3 is realized by software, firmware, or the like. When the parameter adjustment device 3 is realized by software, firmware, or the like, the dataset storage unit 10 is configured on the memory 31 of the computer. A program for causing the computer to execute each processing procedure in the data acquisition unit 11, the elite solution extraction unit 12, the evaluation value prediction unit 13, and the parameter value determination unit 14 is stored in the memory 31. Then, the processor 32 of the computer executes the program stored in the memory 31.

[0026] 2 shows an example in which each of the components of the parameter adjustment device 3 is realized by dedicated hardware, while Fig. 3 shows an example in which the parameter adjustment device 3 is realized by software, firmware, etc. However, this is merely one example, and some of the components in the parameter adjustment device 3 may be realized by dedicated hardware, and the remaining components may be realized by software, firmware, etc.

[0027] Next, the operation of the system shown in FIG. 1 will be described. FIG. 4 is a flowchart showing a parameter adjustment method which is a processing procedure of the parameter adjustment device 3. The adjustable device 1 has I adjustable parameters p i There exists a set of i = 1,...,I, where I is an integer greater than or equal to 1. The parameter p i The parameter value of x i As shown in the following formula (1), all parameter values ​​x i X is the vector that contains X=(x 1 ,x 2 ,···,x I ) (1) First, the parameter value determination unit 14 determines all parameters p i The initial value x for (i=1, ,I) i,init is set to the adjusted device 1. Parameter p i The initial value x i,init may be determined using pseudorandom numbers, for example. i The initial value x i,init is a parameter p i Parameter value x i If known, such a parameter value may be used.

[0028] The adjusted device 1 operates using the parameter value X. The adjusted device 1 outputs a parameter value X and an operation result corresponding to the parameter value X to the evaluation value calculation device 2. If the device under adjustment 1 is, for example, a simulator for an autonomous driving system, the operation results of the device under adjustment 1 include, for example, the position (or relative position) of the host vehicle and other vehicles, etc. at a plurality of times included in the simulation period, or the speed (or relative speed) of the host vehicle and other vehicles, etc. If the device under adjustment 1 is, for example, a simulator for an air conditioner, the operation results of the device under adjustment 1 include, for example, the state quantity of the air conditioner. Examples of the state quantity of the air conditioner include the temperature of the refrigerant, the pressure of the refrigerant, the heating capacity, the cooling capacity, or the energy saving efficiency.

[0029] The evaluation value calculation device 2 acquires the parameter value X and the operation result corresponding to the parameter value X from the device under adjustment 1. The evaluation value calculation device 2 calculates an evaluation value y for the operation result. Since the calculation process of the evaluation value y by the evaluation value calculation device 2 itself is a known technique, a detailed description thereof is omitted. When the device under adjustment 1 is a simulator for an autonomous driving system, for example, when the host vehicle approaches other vehicles, etc., the closer the distance between the host vehicle and other vehicles, etc. becomes, the lower the evaluation value y for the operation result becomes. When the device under adjustment 1 is a simulator for an air conditioner, for example, based on the preset environmental conditions of the air conditioner, it is evaluated how long it takes for the temperature of the space to be air-conditioned by the air conditioner to reach the set temperature. The shorter the time until the temperature of the space to be air-conditioned reaches the set temperature, the higher the evaluation value for the operation result. Alternatively, the evaluation value may be calculated such that the higher the heating capacity or the cooling capacity, the higher the evaluation value y for the operation result.

[0030] The evaluation value calculation device 2 outputs data D including the parameter value X and the evaluation value y to the parameter adjustment device 3 as shown in the following formula (2). D = (X, y) (2)

[0031] The data acquisition unit 11 of the parameter adjustment device 3 receives the stored M pieces of data D from the evaluation value calculation device 2 1 ~D Mto obtain (step ST1 in FIG. 4). The data acquisition unit 11 outputs M pieces of data D 1 ~D M to each of the elite solution extraction unit 12 and the evaluation value prediction unit 13.

[0032] The elite solution extraction unit 12 obtains M pieces of data D 1 ~D M from the data acquisition unit 11. The elite solution extraction unit 12 selects J pieces of data as elite solutions E 1 ~E M from among the M pieces of data D m (m = 1, ···, M) based on the evaluation value y m contained in them (step ST2 in FIG. 4). j = 1, ···, J, where j does not indicate the power of E but an index. J is an integer between 0 and M inclusive. 1 ~E J The elite solution extraction unit 12 outputs the J elite solutions E ~E 1 ~E J to the parameter value determination unit 14.

[0033] Hereinafter, the extraction process of the elite solutions E j (j = 1, ···, J) by the elite solution extraction unit 12 will be specifically described. FIG. 5 is a flowchart showing the extraction process of the elite solutions E j (j = 1, ···, J) by the elite solution extraction unit 12. First, the elite solution extraction unit 12 sets M pieces of data D 1 ~D M as K elite solution candidates Ec 1 ~Ec K . K is an integer greater than or equal to 0, and at this stage, K is equal to M. The set of elite solution candidates is denoted as the elite solution candidate set Ec = {Ec 1 , …, Ec K}. Next, the elite solution extraction unit 12, as shown in the following equation (3), determines the elite solution E jInitialize the list E including it (step ST11 in FIG. 5).

[0034] TIFF0007690148000001.tif11166 In formula (3), [□] indicates an empty list.

[0035] The elite solution extraction unit 12 compares the evaluation values y 1 ~Ec K contained in the K elite solution candidates Ec 1 ~y K with each other. Based on the comparison result, the elite solution extraction unit 12 extracts the highest evaluation value y 1 ~y K from among the K evaluation values y max (step ST12 in FIG. 5). The elite solution extraction unit 12 compares the highest evaluation value y max with the lower limit value y of the evaluation value L . The lower limit value y L may be stored, for example, in the internal memory of the elite solution extraction unit 12, or may be given from outside the parameter adjustment device 3. If the evaluation value y max is less than the lower limit value y L (in the case of NO in step ST13 in FIG. 5), the elite solution extraction unit 12 ends the extraction process of the elite solution E j . If the evaluation value y max is greater than or equal to the lower limit value y L (in the case of YES in step ST13 in FIG. 5), the elite solution extraction unit 12 adds the elite solution candidate Ec max including the evaluation value y max and the parameter value X corresponding to the evaluation value y k to the end of the list E as the elite solution E j (step ST14 in FIG. 5). Hereinafter, the parameter value X included in the elite solution E j and the elite solution candidate Ec kTo distinguish the parameter values X included in (k = 1, …, K), the elite solution E j The parameters included in are denoted as X E j in this way. Also, for the evaluation value included in the elite solution E j and the evaluation value y included in the elite solution candidate Ec k To distinguish them, the evaluation value included in the elite solution E j is denoted as y E j in this way. Also, here, the elite solution E j contains the parameter value X j and the evaluation value y j However, this is just an example, and the elite solution E j may only contain the parameter value X j .

[0036] The elite solution extraction unit 12 updates the elite solution candidate set Ec with a new elite solution candidate set Ec′ (step ST15 in FIG. 5). The elite solution extraction unit 12 selects a certain elite solution candidate Ec 1 ~Ec K from among the M elite solution candidates Ec k For, the parameter value X included in Ec k and all the elite solutions E included in the list E k The distance d(X j and the parameter value X included in E j is calculated. k ,X E j ) As shown in the following formula (4), the elite solution extraction unit 12 calculates all the calculated distances d(X k ,X E j ), if the minimum distance min d(X k ,X E j ) is greater than or equal to the threshold τ, the elite solution candidate Ec k is included in the new elite solution candidate set Ec′. The elite solution extraction unit 12 calculates the minimum distance min d(Xk , X E j If () is less than the threshold τ, the elite solution candidate Ec k is removed from the new elite solution candidate set Ec'. The elite solution extraction unit 12 similarly processes each elite solution candidate Ec k to generate a new elite solution candidate set Ec', and then sets this Ec' as the elite solution candidate set Ec. The threshold τ may be stored in the internal memory of the elite solution extraction unit 12 or may be given from outside the parameter adjustment device 3. The elite solution extraction unit 12 continues to perform the processes of steps ST12 to ST15 in FIG. 5.

[0037] TIFF0007690148000002.tif18166 In Equation (4), d(X k , X E j ) is a function for calculating the distance between the parameter value X k and the parameter value X E j and the parameter value X.

[0038] As the function d(X k , X E j ), for example, there is a function using the distance in the parameter space. Specifically, for example, the Euclidean distance in the space spanned by any of the parameter vector X, the vector obtained by normalizing the parameter vector X, or the vector obtained by non-linearly transforming the parameter vector X using the kernel method, or the Manhattan distance in the space spanned by any vector can be used as the distance between parameters. Alternatively, the distance may be defined in the same manner using one value or a plurality of values selected from among the parameter value, the operation result, and the evaluation value.

[0039] Here, if the evaluation value y max is less than the lower limit value y L , the elite solution extraction unit 12 selects the elite solution E jThe extraction process has been completed. Even when the number of elite solutions E included in the list E exceeds a predetermined number, the elite solution extraction unit 12 may end the extraction process of the elite solution E. By doing so, it is possible to suppress a decrease in search efficiency due to an excessive increase in the elite solution E. j Here, the lower limit value y of the evaluation value is stored, for example, in the internal memory of the elite solution extraction unit 12. The elite solution extraction unit 12 calculates the lower limit value y based on the highest evaluation value y among the M evaluation values y~y stored in the dataset storage unit 10 as shown in the following formula (5). j j

[0040] y = α×y + β (5) L In formula (5), each of α and β is a constant. 1 M max L L max

[0041] The parameter value determination unit 14 acquires J elite solutions E~E from the elite solution extraction unit 12. 1 J The parameter value determination unit 14 determines the parameter value to be used by the device under adjustment 1 for the next operation based on the J elite solutions E~E from among a plurality of parameter values existing in the search space of the parameter values (step ST3 in FIG. 4). 1 J The parameter value determination unit 14 outputs the determined parameter value to the device under adjustment 1.

[0042] Hereinafter, the parameter determination process by the parameter value determination unit 14 will be specifically described. FIG. 6 is a flowchart showing the selection process of the determination method by the determination method selection unit 14a. The determination method selection unit 14a acquires J elite solutions E~E from the elite solution extraction unit 12. 1 J ​​​​​​​​​​​​​​​​ The decision method selection unit 14a selects either the global search method or the local search method as the method for determining the parameter value to be used in the next operation based on the number of elite solutions E 1 ~E J . Specifically, if the number of elite solutions E extracted by the elite solution extraction unit 12 is 0 (in the case of YES in step ST21 of FIG. 6), the global search method is selected (step ST22 of FIG. 6). Specifically, the decision method selection unit 14a j If the number of elite solutions E extracted by the elite solution extraction unit 12 is 0 (in the case of YES in step ST21 of FIG. 6), the global search method is selected (step ST22 of FIG. 6).

[0043] The decision method selection unit 14a j If the number of elite solutions E extracted by the elite solution extraction unit 12 is 1 or more (in the case of NO in step ST21 of FIG. 6), the probability p of selecting the local search method is calculated (step ST23 of FIG. 6). The method for calculating the probability p of selecting the local search method may be any method, but as a method for calculating the probability p, for example, there is a method using a predetermined probability. Also, the decision method selection unit 14a j May also use a method of calculating the probability p of selecting the local search according to the number of elite solutions E. Specifically, the decision method selection unit 14a j Calculates the probability p so that the higher the number of elite solutions E, the higher the probability p of selecting the local search. In this case, an effect of easily finding a parameter with a higher evaluation can be obtained. Also, the decision method selection unit 14a j Based on the evaluation value y j Contained in, a method of calculating the probability p of selecting the local search method may be used. Specifically, the decision method selection unit 14a 1 ~y J Of the sum value, or the higher the average value of the J evaluation values y 1 ~y J , the probability p of selecting the local search method is calculated so as to be higher.

[0044] The decision method selection unit 14a selects either a global search method or a local search method based on the probability p (step ST24 in FIG. 6). When the decision method selection unit 14a selects the global search method (in the case of NO in step ST25 of FIG. 6), the decision method selection process ends. When the decision method selection unit 14a selects the local search method (in the case of YES in step ST25 of FIG. 6), the probability p that each of the J elite solutions E 1 ~E J becomes the target e of the local search is determined (step ST26 in FIG. 6). e j Specifically, the decision method selection unit 14a determines the probability p such that, for example, the lower the evaluation value y is, the higher the probability p that the elite solution E j with a low evaluation value becomes the target e of the local search. j By doing so, the effect of increasing the evaluation value of the entire set of elite solutions can be obtained. Also, the decision method selection unit 14a determines the probability p such that, the higher the evaluation value y e j is, the higher the probability p that the elite solution E e j with a high evaluation value becomes the target e of the local search. j By doing so, the effect of increasing the maximum value of the evaluation value of the elite solutions can be obtained. j e j e j e j j In addition, the decision method selection unit 14a may determine the probability p that becomes the target e of the local search according to the distribution of the parameters that have already been tried. For example, the larger the distance from each elite solution E e j to the N j -th closest data, the higher the probability p that it becomes the target e of the local search. close e j e j e j By doing so, elite solutions with a sparse neighborhood data distribution are more likely to be selected, and the search for unexplored regions is facilitated.

[0045] The decision method selection unit 14a determines the probability p that each of the J elite solutions E 1 ~E J becomes the target e of local search e j Based on this, from among the J elite solutions E 1 ~E J the elite solution E that becomes the target e of local search is selected (step ST27 in FIG. 6). j Specifically, the decision method selection unit 14a uses pseudo-random numbers to select, from among the J elite solutions E 1 ~E J the elite solution E that becomes the target e of local search with a probability of p j e j Select according to. Selecting one option from a plurality of options according to a predetermined probability using pseudo-random numbers is a known technique, so a detailed description is omitted.

[0046] If the global search method is selected by the decision method selection unit 14a, the sampling unit 14b samples the parameter value X-hat from a uniform distribution over the entire search space of the parameter values, for example, using pseudo-random numbers. Sampling the parameter value X-hat using pseudo-random numbers is a known technique, so a detailed description is omitted. In the text of the specification, due to the relationship of electronic applications, the symbol "^" cannot be attached above the character "x", so it is expressed as "x-hat". Further, the search space of the parameter values is a space that includes a plurality of parameter values that can be used for the operation of the device 1 to be adjusted. The search space of the parameter values may be any search space, but if the search space of the parameter values is, for example, a two-dimensional space, and the first dimension of the two-dimensional space is x 1 and the second dimension is x 2 then, [x 1L , x 2L , [x 1L , x 2H , [x 1H , x 2H,[x 1H ,x 2L A rectangle with [x] as its vertex becomes the search space. x 1L is the parameter value x 1 's lower limit value, x 1H is the parameter value x 1 's upper limit value, x 2L is the parameter value x 2 's lower limit value, x 2H is the parameter value x 2 is the upper limit value. In this case, x 1L ≦x 1 ≦x 1H , and x 2L ≦x 2 ≦x 2H Values that satisfy both of these are the parameter values included in the search space.

[0047] If the local search method is selected by the determination method selection unit 14a, the sampling unit 14b samples one or more parameter values X hat from a normal distribution centered on the elite solution E that is the target e of local search among the search space of the parameter values. The timing at which the sampling unit 14b samples the parameter value X hat is the timing shown in step ST2 of FIG. 7 described later. The covariance matrix indicating the normal distribution may use a predetermined one, or the sampling unit 14b may adjust the covariance matrix of the normal distribution according to the search situation. When the sampling unit 14b adjusts the covariance matrix, for example, the larger the evaluation value y j of the elite solution E that is the target e of local search, the smaller each element of the covariance matrix is adjusted. j of the elite solution E that is the target e of local search, the smaller each element of the covariance matrix is adjusted. j is, the smaller each element of the covariance matrix is adjusted. The sampling unit 14b outputs the sampled parameter value X hat to the parameter value determination processing unit 14c.

[0048] FIG. 7 is a flowchart showing the determination process of the parameter value by the parameter value determination processing unit 14c and the like. The parameter value determination processing unit 14c initializes the candidate point set C with an empty set (step ST31 in FIG. 7). If the list E of elite solutions is empty, that is, if the number J of elite solutions is 0, the parameter value determination processing unit 14c adds the parameter value X hat sampled by the sampling unit 14b to the candidate point set C. If the number J of elite solutions is 1 or more, the parameter value determination processing unit 14c determines the priority elite solution (step ST33 in FIG. 7).

[0049] That is, if the local search method is selected by the determination method selection unit 14a, the parameter value determination processing unit 14c determines the J elite solutions E 1 ~E J in the list E as the priority elite solutions Ep j ~Ep 1 ~E j-1 ~Ep 1 ~Ep N The number N of priority elite solutions is equal to (j - 1) in this case. The elite solution E j that is the target e of local search is the j-th elite solution E j from the top of the list E. The following formula (6) shows the list Ep including the priority elite solutions Ep 1 ~Ep N ~Ep Ep = [Ep 1 , ···, Ep N (6) If the elite solution E j that is the target e of local search is at the top of the list E, the number N of priority elite solutions becomes 0, and the list Ep becomes empty. If the global search method is selected, the parameter value determination processing unit 14c determines the J elite solutions E 1 ~E J in the list E as the priority elite solutions Ep 1 ~Ep N At this time, N is equal to J.

[0050] When the number N of the priority elite solutions is 0, the parameter value determination processing unit 14c adds the parameter value \(\hat{X}\) sampled by the sampling unit 14b to the candidate point set C. When the number N of the priority elite solutions is 1 or more, the parameter value determination processing unit 14c calculates the distances s(\(\hat{X}\), \(X\)) between the extracted parameter value \(\hat{X}\) and the parameter values \(X\) included in each of the priority elite solutions \(E_p\)~ \(E_p\) (n = 1, ···, N) included in the list \(E_p\) (step ST34 in FIG. 7). As the distance s(\(\hat{X}\), \(X\)), the distance in the parameter space can be used in the same manner as the above-described distance d(\(X\), \(X\)). 1 ~ \(E_p\) N and the parameter value \(X\) included in each of them p n (n = 1, ···, N) (step ST34 in FIG. 7). As the distance s(\(\hat{X}\), \(X\)), the distance in the parameter space can be used in the same manner as the above-described distance d(\(X\), \(X\)). p n ) (step ST34 in FIG. 7). As the distance s(\(\hat{X}\), \(X\)), the distance in the parameter space can be used in the same manner as the above-described distance d(\(X\), \(X\)). p n ) As the distance s(\(\hat{X}\), \(X\)), the distance in the parameter space can be used in the same manner as the above-described distance d(\(X\), k , \(X\) E j ) Alternatively, by using machine learning, an evaluation value or a value of an operation result may be predicted, and a distance may be calculated using the prediction result.

[0051] The parameter value determination processing unit 14c compares the minimum value min s(\(\hat{X}\), \(X\)) of the distances between the extracted parameter value \(\hat{X}\) and the parameter values \(X\) included in each priority elite solution \(E\) (n = 1, ···, N) with a threshold value δ. The threshold value δ may be stored in the internal memory of the parameter value determination processing unit 14c or may be given from outside the parameter adjustment device 3. p n and the parameter value \(X\) included in it p n (n = 1, ···, N) with a threshold value δ. The threshold value δ may be stored in the internal memory of the parameter value determination processing unit 14c or may be given from outside the parameter adjustment device 3. p n ) with a threshold value δ. The threshold value δ may be stored in the internal memory of the parameter value determination processing unit 14c or may be given from outside the parameter adjustment device 3. As shown in the following formula (7), if the minimum distance min s(\(\hat{X}\), \(X\)) is smaller than the threshold value δ, the parameter value \(\hat{X}\) is added to the candidate point set C (step ST35 in FIG. 7). p n ) (step ST35 in FIG. 7).

[0052] TIFF0007690148000003.tif13166

[0053] The parameter value determination processing unit 14c determines whether the number of parameter values ​​X included in the candidate point set C is N cnd If the number of samples has not yet reached the number of samples (step ST36 in FIG. 7: NO), the processes of steps ST32 to ST36 are repeatedly executed. However, the process of step ST32 is executed by the sampling unit 14b. cnd may be stored in an internal memory of the parameter value determination processing unit 14c, or may be provided from outside the parameter adjustment device 3. The parameter value determination processing unit 14c determines whether the number of parameter values ​​X included in the candidate point set C is N cnd 7 (step ST36: YES in FIG. 7), each parameter value X hat included in the candidate point set C is output to the evaluation value prediction unit 13.

[0054] The evaluation value prediction unit 13 acquires each parameter value X included in the candidate point set C from the parameter value determination processing unit 14c. The evaluation value prediction unit 13 provides each parameter value X to the learning model 13a, and obtains at least an evaluation value y corresponding to each parameter value X from the learning model 13a. During inference, when a parameter x is given from the evaluation value prediction unit 13, the learning model 13a outputs at least an evaluation value y corresponding to the parameter value x to the evaluation value prediction unit 13 as a prediction result of the evaluation value. The evaluation value prediction unit 13 outputs, as a prediction result of the evaluation value, an evaluation value y hat corresponding to each parameter value X hat to the parameter value determination processing unit 14c. 1, the evaluation value prediction unit 13 provides the parameter value X to the learning model 13a and obtains the evaluation value y corresponding to the parameter value X from the learning model 13a. However, it is sufficient for the evaluation value prediction unit 13 to predict the evaluation value y corresponding to the parameter value X, and it may predict the evaluation value y corresponding to the parameter value X without using the learning model 13a.

[0055] The parameter value determination processing unit 14c acquires, from the evaluation value prediction unit 13, at least the evaluation value ŷ corresponding to each parameter value X̂ (step ST37 in FIG. 7). The parameter value determination processing unit 14c calculates at least the acquisition function value based on each evaluation value ŷ (step ST38 in FIG. 7). The calculation of the acquisition function value by the parameter value determination processing unit 14c uses an acquisition function. Examples of the acquisition function include UCB (Upper Confidence Bound) or EI (Expected Improvement). In a general acquisition function such as UCB or EI, it is assumed that when making a prediction by machine learning, the distribution of the evaluation value, for example, the mean and variance can be obtained, and it is often adopted when using Gaussian process regression as the machine learning method. In the case where only a deterministic single value is obtained in the prediction, an acquisition function that returns the predicted evaluation value as it is may be used.

[0056] The parameter value determination processing unit 14c selects the parameter value X̂ with the highest calculated acquisition function value from among the one or more parameter values X̂ acquired from the sampling unit 14b. The parameter value determination processing unit 14c outputs, to the device under adjustment 1, the parameter value X̂ with the highest acquisition function value as the parameter value to be used by the device under adjustment 1 in the next operation (step ST39 in FIG. 7). In the parameter adjustment device 3 shown in FIG. 1, the parameter value determination processing unit 14c outputs, to the device under adjustment 1, the parameter value X̂ with the highest acquisition function value as the parameter value to be used by the device under adjustment 1 in the next operation. However, the parameter value to be used by the device under adjustment 1 in the next operation is not limited to the parameter value X̂ with the highest acquisition function value, and within a range where there are no practical problems, a parameter value X̂ with an acquisition function value lower than that of the parameter value X̂ with the highest acquisition function value may be output to the device under adjustment 1 as the parameter value to be used by the device under adjustment 1 in the next operation.

[0057] The device to be adjusted 1 performs operation using the parameter value \(\hat{X}\) output from the parameter value determination processing unit 14c, and outputs the operation result to the evaluation value calculation device 2. Hereinafter, the parameter adjustment device 3 repeatedly executes the processes of steps ST1 to ST3 shown in FIG. 4.

[0058] In the above-described first embodiment, the operation result of the device to be adjusted 1 that performs operation using one parameter or a plurality of parameters is obtained, and from the evaluation value calculation device 2 that calculates an evaluation value for the operation result, data including the value of each parameter and the evaluation value related to the value of each parameter is obtained by the data acquisition unit 11, and the elite solution extraction unit 12 that extracts zero or more data as elite solutions based on the evaluation values included in each data among the plurality of data obtained by the data acquisition unit 11. The parameter adjustment device 3 is configured to include. Further, the parameter adjustment device 3 determines the value of the parameter to be used by the device to be adjusted 1 in the next operation based on the elite solution extracted by the elite solution extraction unit 12 from among the values of the plurality of parameters existing in the search space of the parameter values, and the determined parameter value is output to the device to be adjusted 1. It has a parameter value determination unit. Therefore, even when the appropriate number of parameter values to be found is unknown, the parameter adjustment device 3 can search for diverse parameter values within an allowable range of deterioration in search efficiency.

[0059] In the parameter adjustment device 3 shown in FIG. 1, the evaluation value y calculated by the evaluation value calculation device 2 is a scalar value. However, this is only an example, and the evaluation value calculation device 2 may calculate an evaluation vector Y having a dimension of L as shown in the following formula (8). L is an integer of 2 or more.

[0060] TIFF0007690148000004.tif26166

[0061] And the evaluation value calculation device 2 uses, for example, as the evaluation value y, as shown in the following formula (9), the L elements y included in the evaluation vector Y1 ~y L Calculate the weighted sum of them. As described above, by calculating the evaluation vector Y with the dimension number L by the evaluation value calculation device 2, each element of the evaluation vector can be reflected in the distance d, and it becomes possible to search for a plurality of parameter values with diversity also in the evaluation vector space.

[0062] TIFF0007690148000005.tif10166In Equation (9), w 1 , ···, w M Each of them is a weighting coefficient.

[0063] Embodiment 2. In Embodiment 2, a parameter adjustment device 3 including a determination method selection unit 15a that selects a parameter determination method based on the evaluation value included in the elite solution and the upper limit value of the evaluation value will be described for the parameter value determination unit 15.

[0064] FIG. 8 is a configuration diagram showing a system including the parameter adjustment device 3 according to Embodiment 2. In FIG. 8, the same reference numerals as those in FIG. 1 denote the same or corresponding parts, and thus detailed description thereof is omitted. FIG. 9 is a hardware configuration diagram showing the hardware of the parameter adjustment device 3 according to Embodiment 2. In FIG. 9, the same reference numerals as those in FIG. 2 denote the same or corresponding parts, and thus detailed description thereof is omitted. The system shown in FIG. 8 includes an apparatus to be adjusted 1, an evaluation value calculation device 2, and a parameter adjustment device 3.

[0065] The parameter value determination unit 15 is realized by, for example, a parameter value determination circuit 25 shown in FIG. 9. The parameter value determination unit 15 includes a determination method selection unit 15a, a sampling unit 14b, and a parameter value determination processing unit 14c. The parameter value determination unit 15 determines the parameter value to be used by the apparatus to be adjusted 1 for the next operation based on the elite solution extracted by the elite solution extraction unit 12 from among a plurality of parameter values existing in the search space of the parameter values. The parameter value determination unit 15 outputs the determined parameter value to the device under adjustment 1.

[0066] The determination method selection unit 15a obtains J elite solutions E 1 ~E J from the elite solution extraction unit 12. The determination method selection unit 15a selects either a global search method or a local search method as the method for determining the parameter value to be used in the next operation based on the elite solutions E 1 ~E J . Specifically, if the number J of elite solutions extracted by the elite solution extraction unit 12 is 0, the determination method selection unit 15a selects the global search method. The determination method selection unit 15a compares the lowest evaluation value y 1 ~y J included in the J elite solutions E 1 ~y J with the upper limit value y min of the evaluation values. If the evaluation value y H is greater than or equal to the upper limit value y min , the determination method selection unit 15a selects the global search method. H If the evaluation value y min is lower than the upper limit value y H , the determination method selection unit 15a calculates the probability p that the local search method is selected. The determination method selection unit 15a selects either the global search method or the local search method based on the probability p.

[0067] In FIG. 8, it is assumed that each of the data acquisition unit 11, the elite solution extraction unit 12, the evaluation value prediction unit 13, and the parameter value determination unit 15, which are components of the parameter adjustment device 3, is realized by dedicated hardware as shown in FIG. 9. That is, it is assumed that the parameter adjustment device 3 is realized by a data acquisition circuit 21, an elite solution extraction circuit 22, an evaluation value prediction circuit 23, and a parameter value determination circuit 25. Each of the data acquisition circuit 21, the elite solution extraction circuit 22, the evaluation value prediction circuit 23, and the parameter value determination circuit 25 corresponds to, for example, a single circuit, a composite circuit, a programmed processor, a parallel programmed processor, an ASIC, an FPGA, or a combination thereof.

[0068] The components of the parameter adjustment device 3 are not limited to those realized by dedicated hardware, and the parameter adjustment device 3 may be realized by software, firmware, or a combination of software and firmware. When the parameter adjustment device 3 is realized by software or firmware, etc., a program for causing a computer to execute the respective processing procedures in the data acquisition unit 11, the elite solution extraction unit 12, the evaluation value prediction unit 13, and the parameter value determination unit 15 is stored in the memory 31 shown in FIG. 3. Then, the processor 32 shown in FIG. 3 executes the program stored in the memory 31.

[0069] Also, FIG. 9 shows an example in which each of the components of the parameter adjustment device 3 is realized by dedicated hardware, and FIG. 3 shows an example in which the parameter adjustment device 3 is realized by software or firmware, etc. However, this is only an example, and some of the components in the parameter adjustment device 3 may be realized by dedicated hardware, and the remaining components may be realized by software or firmware, etc.

[0070] Next, the operation of the system shown in FIG. 8 will be described. However, except for the determination method selection unit 15a of the parameter value determination unit 15, it is the same as the system shown in FIG. 1. For this reason, here, only the operation of the determination method selection unit 15a will be described.

[0071] The determination method selection unit 15a acquires J elite solutions E 1 ~E J from the elite solution extraction unit 12. The determination method selection unit 15a selects the elite solutions E 1 ~EJ Based on this, as a method for determining parameter values to be used in the next operation, either a global search method or a local search method is selected. Specifically, if the number J of elite solutions extracted by the elite solution extraction unit 12 is 0, the decision method selection unit 15a selects the global search method. The decision method selection unit 15a determines the J elite solutions E 1 ~E J The evaluation values y 1 ~y J Among them, the lowest evaluation value y min And the upper limit value y of the evaluation value H Are compared. If the evaluation value y min Is greater than or equal to the upper limit value y H The decision method selection unit 15a selects the global search method. If the evaluation value y min Is lower than the upper limit value y H The decision method selection unit 15a calculates the probability p that the local search method is selected. Based on the probability p, the decision method selection unit 15a selects either the global search method or the local search method. If the decision method selection unit 15a selects the global search method, the selection process of the decision method ends. If the decision method selection unit 15a selects the local search method, the probability p 1 ~E J That each of the J elite solutions E becomes the target e of local search e j Is determined. In the second embodiment, it is assumed that the probability p j That the evaluation value y included in the elite solution E j Is greater than or equal to the upper limit value y H For the elite solution corresponding to is 0. e j Is 0. The decision method selection unit 15a outputs the selection result of the parameter value determination method to the sampling unit 14b.

[0072] In the above-described Embodiment 2, the parameter determination method selection unit 15a configured the parameter adjustment device 3 shown in FIG. 8 so as to select a parameter determination method based on the evaluation value included in the elite solution and the upper limit value of the evaluation value. Therefore, similar to the parameter adjustment device 3 shown in FIG. 1, the parameter adjustment device 3 shown in FIG. 8 can search for diverse parameter values within an allowable range of deterioration in search efficiency even when the appropriate number of parameter values to be found is unknown. Further, in the parameter adjustment device 3 shown in FIG. 8, the probability that parameter values are sampled from a search space centered on elite solution candidates whose evaluation values are less than or equal to the upper limit value is higher than that of the parameter adjustment device 3 shown in FIG. 1. As a result, the parameter adjustment device 3 shown in FIG. 8 can increase the probability that diverse parameter values are searched as compared with the parameter adjustment device 3 shown in FIG. 1.

[0073] Embodiment 3. In Embodiment 3, if the local search method is selected by the determination method selection unit 14a, among the M pieces of data D 1 ~D M the data in the vicinity of the elite solution E that is the target e of local search is extracted as learning data, and a learning model 16a that learns the evaluation value y corresponding to the parameter value X included in the learning data will be described for the parameter adjustment device 3 provided. j FIG. 10 is a configuration diagram showing a system including the parameter adjustment device 3 according to Embodiment 3. In FIG. 10, the same reference numerals as those in FIGS. 1 and 8 denote the same or corresponding parts, and thus detailed description thereof is omitted.

[0074] FIG. 11 is a hardware configuration diagram showing the hardware of the parameter adjustment device 3 according to Embodiment 3. In FIG. 11, the same reference numerals as those in FIGS. 2 and 9 denote the same or corresponding parts, and thus detailed description thereof is omitted. The system shown in FIG. 10 includes an apparatus to be adjusted 1, an evaluation value calculation device 2, and a parameter adjustment device 3. The evaluation value prediction unit 16 is realized, for example, by the evaluation value prediction circuit 26 shown in FIG. 11.

[0075] ​The evaluation value prediction unit 16 predicts an evaluation value for the operation result of the device under adjustment 1 when the device under adjustment 1 performs an operation using each parameter value sampled by the sampling unit 14b of the parameter value determination unit 14. When the global search method is selected by the determination method selection unit 14a during the learning of the learning model 16a, the evaluation value prediction unit 16 gives each of the M pieces of data D 1 ~D M acquired by the data acquisition unit 11 to the learning model 16a. When the local search method is selected by the determination method selection unit 14a during the learning of the learning model 16a, the evaluation value prediction unit 16, among the M pieces of data D 1 ~D M extracts, as learning data, the data in the vicinity of the elite solution E that is the target e of local search j and gives the learning data to the learning model 16a.

[0076] The learning model 16a is realized, for example, by a Gaussian process regression model, a linear regression model, a neural network, a decision tree, a random forest, or a gradient boosting tree. When the global search method is selected by the determination method selection unit 14a during the learning of the learning model 16a, the learning model 16a acquires, from the evaluation value prediction unit 16, the data D m (m = 1, ···, M) acquired by the data acquisition unit 11. The learning model 16a learns the evaluation value y corresponding to the parameter value X included in the data D m (m = 1, ···, M). When the local search method is selected by the determination method selection unit 14a during the learning of the learning model 16a, the learning model 16a acquires, as learning data, the data in the vicinity of the elite solution E that is the target e of local search among the M pieces of data D 1 ~D M and the elite solution E j of the target e of local search. The learning model 16a learns the evaluation value y corresponding to the parameter value X included in the learning data. During inference, when the parameter value \(\hat{X}\) is given from the evaluation value prediction unit 16 to the learning model 16a, the learning model 16a outputs the evaluation value \(\hat{y}\) corresponding to the parameter value \(\hat{X}\) to the evaluation value prediction unit 16 as the prediction result of the evaluation value.

[0077] The parameter adjustment device 3 shown in FIG. 10 incorporates the learning model 16a. However, this is only an example, and the learning model 16a may be provided outside the parameter adjustment device 3. In the parameter adjustment device 3 shown in FIG. 10, each of the evaluation value prediction unit 16 and the learning model 16a is applied to the parameter adjustment device 3 shown in FIG. 1. However, this is only an example, and each of the evaluation value prediction unit 16 and the learning model 16a may be applied to the parameter adjustment device 3 shown in FIG. 8.

[0078] In FIG. 10, it is assumed that each of the data acquisition unit 11, the elite solution extraction unit 12, the evaluation value prediction unit 16, and the parameter value determination unit 14, which are components of the parameter adjustment device 3, is realized by dedicated hardware as shown in FIG. 11. That is, it is assumed that the parameter adjustment device 3 is realized by a data acquisition circuit 21, an elite solution extraction circuit 22, an evaluation value prediction circuit 26, and a parameter value determination circuit 24. Each of the data acquisition circuit 21, the elite solution extraction circuit 22, the evaluation value prediction circuit 26, and the parameter value determination circuit 24 corresponds to, for example, a single circuit, a composite circuit, a programmed processor, a parallel programmed processor, an ASIC, an FPGA, or a combination thereof.

[0079] The components of the parameter adjustment device 3 are not limited to those realized by dedicated hardware, and the parameter adjustment device 3 may be realized by software, firmware, or a combination of software and firmware. When the parameter adjustment device 3 is realized by software, firmware, or the like, a program for causing a computer to execute the respective processing procedures in the data acquisition unit 11, the elite solution extraction unit 12, the evaluation value prediction unit 16, and the parameter value determination unit 14 is stored in the memory 31 shown in FIG. 3. Then, the processor 32 shown in FIG. 3 executes the program stored in the memory 31.

[0080] Further, FIG. 11 shows an example in which each component of the parameter adjustment device 3 is realized by dedicated hardware, and FIG. 3 shows an example in which the parameter adjustment device 3 is realized by software, firmware, or the like. However, this is merely an example, and some components in the parameter adjustment device 3 may be realized by dedicated hardware and the remaining components may be realized by software, firmware, or the like.

[0081] Next, the operation of the system shown in FIG. 10 will be described. However, except for the evaluation value prediction unit 16 and the learning model 16a, it is the same as the system shown in FIG. 1. For this reason, only the operations of the evaluation value prediction unit 16 and the learning model 16a will be described here.

[0082] The evaluation value prediction unit 16 acquires the selection result of the search method from the determination method selection unit 14a. If the selection result of the search method indicates the global search method, the evaluation value prediction unit 16 acquires M pieces of data D 1 ~D M from the data acquisition unit 11 and gives the M pieces of data D 1 ~D M to the learning model 16a. When given M pieces of data D 1 ~D M from the evaluation value prediction unit 16 during learning, the learning model 16a learns the evaluation value y corresponding to the parameter value X included in the data D m (m = 1, ···, M).

[0083] If the selection result of the search method indicates the local search method, the evaluation value prediction unit 16 extracts, as learning data, data in the vicinity of the elite solution E that is the target e of the local search from among the M pieces of data D 1 ~D M and provides the learning data to the learning model 16a. As the data in the vicinity, for example, data within the 3σ confidence interval of the normal distribution used when the sampling unit 14b performs sampling may be used. j Also, depending on the machine learning method, there are some that can set weights for each data. Therefore, after using all of the M pieces of data as learning data, learning may be performed by giving weights such that the closer the distance between each data and the elite solution, the heavier the weight. During learning, the learning model 16a learns the evaluation value y corresponding to the parameter value X included in the learning data.

[0084] When the evaluation value prediction unit 16 receives the parameter value X hat from the parameter value determination processing unit 14c, it provides the parameter value X hat to the learning model 16a and obtains, from the learning model 16a, the evaluation value y hat corresponding to the parameter value X hat. During inference, when the learning model 16a is given the parameter value X hat from the evaluation value prediction unit 16, it outputs, as the prediction result of the evaluation value, the evaluation value y hat corresponding to the parameter value X hat to the evaluation value prediction unit 16. The evaluation value prediction unit 16 outputs, as the prediction result of the evaluation value, the evaluation value y hat corresponding to the parameter value X hat to the parameter value determination processing unit 14c.

[0085] When local search is selected by the determination method selection unit 14a, the parameter value X hat included in the candidate point set C is a parameter that is distributed only in the vicinity of the elite solution E that is the target e of the local search. Therefore, even if the prediction accuracy at a position far from the elite solution E that is the target e of the local search is low, there is no problem. j j

[0086] ​​​In the above third embodiment, if the local search method is selected by the decision method selection unit 14a, M pieces of data D 1 ~D M Among them, the elite solution E j 10 is configured to include a learning model 16a that learns an evaluation value y corresponding to a parameter value X included in data in the vicinity of the parameter value X. Therefore, the parameter adjustment device 3 shown in Fig. 10 can improve the prediction accuracy of the evaluation value when the local search method is selected by the determination method selection unit 14a more than the parameter adjustment device 3 shown in Fig. 1.

[0087] Embodiment 4 The parameter adjustment device 3 according to the first to third embodiments can be applied to, for example, a simulation scenario generation device for an autonomous driving system. When the parameter adjustment device 3 is applied to a simulation scenario generation device, it becomes possible to generate scenarios in which the autonomous driving system under development falls into unsafe situations, thereby improving the efficiency of safety verification of the autonomous driving system.

[0088] The scenario parameters used when the autonomous driving system executes the above scenario are parameter values ​​X used for the operation of the adjusted device 1. As the evaluation value for the operation result of the adjusted device 1, for example, at least one of the degree of danger or the degree of discomfort is used. The risk level is an index that indicates the degree of danger that the autonomous driving system has caused the vehicle to fall into. For example, in the case of a collision risk between the vehicle and another vehicle, the risk level may be calculated by multiplying the closest inter-vehicle distance between the vehicle and the other vehicle by, for example, -1. Furthermore, the degree of danger when a traffic rule is broken can be, for example, the importance of the broken traffic rule or the degree to which the traffic rule was broken. The risk level may be represented by a scalar value, or may be represented by a vector having a value for each item.

[0089] In an automatic driving system, in addition to avoiding dangerous driving, it is necessary to avoid driving that makes the passengers feel uncomfortable. The discomfort level indicates the degree to which the passengers feel uncomfortable. As a specific example of the discomfort level, for example, the maximum absolute value of the jerk, which is the acceleration of the host vehicle or the jerk of the host vehicle, can be used.

[0090] In Embodiment 4, in the elite solution extraction unit 12, as the distance d that defines the proximity between parameter values, for example, the Euclidean distance in the space spanned by z shown in the following formula (10) can be used.

[0091] TIFF0007690148000006.tif17166 In formula (10), Tc is the time when the host vehicle is closest to the other vehicle, and T is a constant. z is the two-dimensional relative position (X(t), Y(t)) of the other vehicle as seen from the host vehicle arranged.

[0092] Embodiment 5. The parameter adjustment device 3 according to Embodiments 1 to 3 can be used, for example, as a control parameter adjustment device for an air conditioner. Generally, the execution time of an air conditioner simulator has the advantage of being significantly shorter than that of the actual device. On the other hand, since the air conditioner simulator cannot reproduce some behaviors of the actual device, even parameter values with high evaluation on the simulator may have low evaluation on the actual device and may cause problems. Compared with the case where only one parameter value is obtained by using the air conditioner simulator, when a plurality of diverse parameter values are obtained, the possibility of obtaining parameter values that can achieve high evaluation in the actual device increases. In Embodiment 5, by using the air conditioner simulator and the actual device in combination, efficient parameter adjustment is made possible.

[0093] FIG. 12 is a configuration diagram showing a system including a parameter adjustment device 3 for an air conditioner according to Embodiment 5. In FIG. 12, the same reference numerals as those in FIGS. 1, 8, and 10 denote the same or corresponding parts, and thus detailed descriptions thereof are omitted. The system shown in FIG. 12 includes an apparatus to be adjusted 1, an evaluation value calculation device 2, a parameter adjustment device 3, a parameter transmission unit 41, an actual air conditioner 42, and an actual operation result storage unit 43.

[0094] In the system shown in FIG. 12, the apparatus to be adjusted 1 is a simulator of the air conditioner. The apparatus to be adjusted 1 has a function of calculating a change in the state of the air conditioner according to an actuator operation pattern determined based on control parameters. An air conditioning state quantity is obtained as an operation result. Examples of the state quantity include the temperature of the refrigerant, the pressure of the refrigerant, the heating capacity, the cooling capacity, the energy saving efficiency, or the temperature, pressure, and enthalpy of each element constituting the air conditioner. The parameter adjustment device 3 shown in FIG. 12 is the same as the parameter adjustment device 3 shown in FIG. 1. However, this is merely an example, and the parameter adjustment device 3 shown in FIG. 12 may be the same as either the parameter adjustment device 3 shown in FIG. 8 or the parameter adjustment device 3 shown in FIG. 10.

[0095] In Embodiment 5, the control parameter for determining the actuator operation pattern is the parameter value X used for the operation of the apparatus to be adjusted 1. As an evaluation value for the operation result of the apparatus to be adjusted 1, for example, at least one of the temperature of the refrigerant, the pressure of the refrigerant, the heating capacity, the cooling capacity, the energy saving efficiency, or the start-up time is used. The start-up time is the time until the internal state of the air conditioner reaches a steady state.

[0096] In Embodiment 5, in the elite solution extraction unit 12, as the distance d defining the proximity between parameter values, for example, the Euclidean distance in the space spanned by z shown in the following formula (11) can be used.

[0097] In (11) of TIFF0007690148000007, Pd is the pressure at the outlet of the compressor of the air conditioner, and Ps is the pressure at the inlet of the compressor of the air conditioner. Δt is the discrete time width, and N is the number of simulation steps. z is the arrangement of the pressures Pd and Ps at time t.

[0098] The parameter transmission unit 41 acquires the J elite solutions E 1 ~E J extracted by the elite solution extraction unit 12 of the parameter adjustment device 3. The parameter transmission unit 41 transmits the parameter values X included in the elite solution E j (j = 1, ···, J) to the actual air conditioner 42.

[0099] The actual air conditioner 42 receives the parameter values X from the parameter transmission unit 41. The actual air conditioner 42 performs air conditioning operation by setting an actuator operation pattern based on the parameter values X. The actual operation result storage unit 43 acquires the operation state of the actual air conditioner 42 and stores the operation state in association with the parameter values X.

[0100] According to the fifth embodiment, it is possible to verify the operation state of the actual air conditioner 42 using control parameters with high evaluation values and diversity. In addition, by performing parameter adjustment using a simulator, it is possible to significantly shorten the adjustment time and increase the probability of obtaining parameter values with high evaluation values even when applied to the actual air conditioner 42.

[0101] Note that in the present disclosure, free combinations of the respective embodiments, modifications of any components of the respective embodiments, or omissions of any components in the respective embodiments are possible.

Industrial Applicability

[0102] The present disclosure is suitable for a parameter adjustment device and a parameter adjustment method.

Description of Signs

[0103] 1 Device to be adjusted, 2 Simulator (evaluation value calculation device), 3 Parameter adjustment device, 10 Dataset storage unit, 11 Data acquisition unit, 12 Elite solution extraction unit, 13, 16 Evaluation value prediction unit, 13a, 16a Learning model, 14 Parameter value determination unit, 14a Determination method selection unit, 14b Sampling unit, 14c Parameter value determination processing unit, 15 Parameter value determination unit, 15a Determination method selection unit, 20 Dataset storage circuit, 21 Data acquisition circuit, 22 Elite solution extraction circuit, 23, 26 Evaluation value prediction circuit, 24, 25 Parameter value determination circuit, 31 Memory, 32 Processor, 41 Parameter transmission unit, 42 Actual air conditioner, 43 Actual operation result storage unit.

Claims

1. A data acquisition unit that acquires data including the value of each parameter and the evaluation value related to the value of each parameter from an evaluation value calculation device that acquires the operation result of an adjustment target device that performs operation using one parameter or a plurality of parameters and calculates an evaluation value for the operation result; An elite solution extraction unit that extracts zero or more data as elite solutions based on the evaluation values included in each data from among the plurality of data acquired by the data acquisition unit; A parameter value determination unit that determines the value of the parameter to be used by the adjustment target device in the next operation based on the elite solution extracted by the elite solution extraction unit from among the values of the plurality of parameters existing in the search space of the parameter value, and outputs the determined parameter value to the adjustment target device Comprising The parameter value determination unit A determination method selection unit that selects either a global search method or a local search method as a method for determining the value of the parameter to be used in the next operation based on the elite solution extracted by the elite solution extraction unit; If the global search method is selected by the determination method selection unit, sample the values of one or more parameters according to a probability distribution that spreads uniformly over the entire search space, and if the local search method is selected by the determination method selection unit, sample the values of one or more parameters according to a probability distribution centered on a certain elite solution in the search space A sampling unit; A parameter value determination processing unit that calculates the distance between the value of each parameter sampled by the sampling unit and the value of the parameter included in the elite solution extracted by the elite solution extraction unit, and determines the value of the parameter to be used by the adjustment target device in the next operation from among the values of one or more parameters sampled by the sampling unit based on the distance. A parameter adjustment device characterized by comprising

2. The determination method selection unit If the number of elite solutions extracted by the elite solution extraction unit is 1 or more, selects either the global search method or the local search method based on the probability of selecting the local search method. The parameter adjustment device according to claim 1.

3. The determination method selection unit Among the elite solutions extracted by the elite solution extraction unit, if the number of elite solutions containing an evaluation value lower than the upper limit value is one or more, either the global search method or the local search method is selected based on the probability of selecting the local search method. The parameter adjustment device according to claim 1, characterized in that.

4. The determination method selection unit The parameter adjustment device according to claim 2 or claim 3, characterized in that it calculates the probability of selecting the local search method based on the evaluation value included in the elite solution extracted by the elite solution extraction unit.

5. It is provided with an evaluation value prediction unit that predicts an evaluation value regarding the operation result of the device to be adjusted when the device to be adjusted operates using the value of each parameter sampled by the sampling unit. The parameter value determination processing unit The parameter adjustment device according to any one of claims 1 to 3, characterized in that, based on the distance and the prediction result of the evaluation value by the evaluation value prediction unit, it determines the value of the parameter to be used by the device to be adjusted for the next operation from among the values of one or more parameters sampled by the sampling unit.

6. The evaluation value prediction unit When the value of the parameter included in each data acquired by the data acquisition unit and the evaluation value included in each data are given, the value of the parameter sampled by the sampling unit is given to a learning model that learns the evaluation value corresponding to each parameter value, and from the learning model, as the prediction result of the evaluation value regarding the operation result of the device to be adjusted, the evaluation value corresponding to the value of the parameter sampled by the sampling unit is acquired. The parameter adjustment device according to claim 5, characterized in that.

7. The evaluation value prediction unit If the global search method is selected by the determination method selection unit, when the value of the parameter included in each data acquired by the data acquisition unit and the evaluation value included in each data are given, for the learning model that learns the evaluation value corresponding to the value of each parameter, the value of the parameter sampled by the sampling unit is given, and from the learning model, as a prediction result of the evaluation value regarding the operation result of the device to be adjusted, the evaluation value corresponding to the value of the parameter sampled by the sampling unit is acquired. If the local search method is selected by the determination method selection unit, among the data acquired by the data acquisition unit, when a part of the data is extracted as learning data and the value of the parameter included in each learning data and the evaluation value included in each learning data are given, for the learning model that learns the evaluation value corresponding to the value of each parameter, the value of the parameter sampled by the sampling unit is given, and from the learning model, as a prediction result of the evaluation value regarding the operation result of the device to be adjusted, the evaluation value corresponding to the value of the parameter sampled by the sampling unit is acquired. The parameter adjustment device according to claim 5, characterized in that.

8. A data acquisition unit that acquires data including the value of each parameter and the evaluation value related to the value of each parameter from an evaluation value calculation device that acquires the operation result of a device to be adjusted that performs operation using one parameter or a plurality of parameters and calculates the evaluation value regarding the operation result. An elite solution extraction unit that extracts zero or more data as elite solutions based on the evaluation values included in each data from among the plurality of data acquired by the data acquisition unit. A parameter value determination unit that determines the value of the parameter to be used by the device to be adjusted for the next operation based on the elite solution extracted by the elite solution extraction unit from among the values of the plurality of parameters existing in the search space of the parameter values, and outputs the determined parameter value to the device to be adjusted. Comprising The elite solution extraction unit The parameter adjustment device is characterized in that, among the plurality of data acquired by the data acquisition unit, data having an evaluation value greater than or equal to a lower limit value is extracted as an elite solution.

9. The data acquisition unit acquires the operation results of the device to be adjusted that operates using one parameter or a plurality of parameters, and obtains data including the value of each parameter and the evaluation value related to the value of each parameter from an evaluation value calculation device that calculates an evaluation value for the operation results. The elite solution extraction unit extracts zero or more data as elite solutions from among the plurality of data acquired by the data acquisition unit, based on the evaluation values included in each data. The parameter value determination unit determines the value of the parameter to be used by the device to be adjusted in the next operation from among the values of the plurality of parameters existing in the search space of the parameter values, based on the elite solutions extracted by the elite solution extraction unit, and outputs the determined parameter value to the device to be adjusted. The parameter value determination unit includes a determination method selection unit, a sampling unit, and a parameter value determination processing unit. The determination method selection unit selects either a global search method or a local search method as the determination method for the value of the parameter to be used in the next operation, based on the elite solutions extracted by the elite solution extraction unit. If the global search method is selected by the determination method selection unit, the sampling unit samples the values of one or more parameters according to a probability distribution that spreads uniformly over the entire search space. If the local search method is selected by the determination method selection unit, the sampling unit samples the values of one or more parameters according to a probability distribution centered on a certain elite solution within the search space. The parameter value determination processing unit calculates the distance between the value of each parameter sampled by the sampling unit and the value of the parameter included in the elite solution extracted by the elite solution extraction unit, and determines the value of the parameter to be used by the device to be adjusted in the next operation from among the values of one or more parameters sampled by the sampling unit, based on the distance. Parameter adjustment method.

Images