Data calculation device, data calculation method and program

The data calculation device employs a hierarchical model structure to expedite parameter value estimation in large-scale simulations by reducing the number of parameters, addressing the time constraints of existing methods and achieving high-accuracy estimates.

JP7582316B2Active Publication Date: 2024-11-13NEC CORP
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
JP2022550307
Authority / Receiving Office
JP · JP
Patent Type
Patents
Current Assignee / Owner
Filing Date
2020-09-18
Publication Date
2024-11-13
Estimated Expiration
2040-09-18

AI Technical Summary

Technical Problem

Existing methods for estimating parameter values in large-scale simulations are time-consuming and may not allow for obtaining estimates within a realistic timeframe due to the exponential increase in time required as the number of parameters increases.

Method used

A data calculation device and method that utilizes a hierarchical model structure, where parameter values of a lower model are estimated based on a posterior distribution of an upper model, allowing for simulations to be performed in a layered approach to reduce the number of parameters and expedite the estimation process.

Benefits of technology

Enables efficient estimation of parameter values in a shorter time frame by reducing the number of parameters, avoiding the inability to obtain estimates within a realistic time due to a large number of model parameters, and allowing for high-accuracy parameter value estimation.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure 0007582316000023
    Figure 0007582316000023
  • Figure 0007582316000024
    Figure 0007582316000024
  • Figure 0007582316000025
    Figure 0007582316000025
Patent Text Reader

Abstract

Provided is a data calculation device comprising: an upper-level posterior distribution data calculation means for calculating data indicating a posterior distribution of parameter values of an upper-level model that simulates a target, the calculated data being based on target data for an output of the upper-level model; and a lower-level posterior distribution data calculation means for calculating data indicating a posterior distribution of parameter values of a lower-level model that simulates a part of the target and outputs a parameter value of the upper-level model, the calculated data being based on the data indicating the posterior distribution of the parameter values of the upper-level model.
Need to check novelty before this filing date? Find Prior Art

Description

[Technical field]

[0001] The present invention relates to a data calculation device, a data calculation method, and program Regarding. [Background technology]

[0002] A technique has been proposed for estimating parameter values ​​that allow a simulator to output target data based on target data corresponding to the output of the simulator. For example, Patent Document 1 describes the use of Kernel Approximate Bayesian Computation (Kernel ABC) and Kernel Herding to calculate the posterior distribution of parameter values ​​according to target data. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] International Publication No. 2020-105468 Summary of the Invention [Problem to be solved by the invention]

[0004] When there are many parameters whose values ​​are to be estimated, such as when the scale of the simulation target is large, it may take a long time to estimate the parameter values. In particular, it may take an exponential amount of time to estimate the parameter values ​​as the number of parameters whose values ​​are to be estimated increases. For this reason, when there are many parameters whose values ​​are to be estimated, it may be impossible to obtain parameter estimates within a realistic amount of time.

[0005] One object of the present invention is to provide a data calculation device, a data calculation method, and program The purpose of this project is to provide [Means for solving the problem]

[0006] According to a first aspect of the present invention, a data calculation device ,versus Imitating elephant parts 、 The above Simulate the target a lower posterior distribution data calculation means for calculating data indicating a posterior distribution of parameter values ​​of a lower model based on data indicating a posterior distribution of parameter values ​​of the higher model for a lower model that outputs parameter values ​​of a higher model; a lower model simulation execution means for executing a simulation by applying to the lower model parameter values ​​based on data showing a posterior distribution of parameter values ​​of the lower model; a higher model simulation execution means for executing a simulation by applying to the higher model sample data of parameter values ​​of the higher model based on a result of the simulation using the lower model; and a higher posterior distribution data calculation means for calculating data showing a posterior distribution of parameter values ​​of the higher model based on target data for the output of the higher model, based on a similarity between sample data of the output of the higher model obtained as a result of the simulation by the higher model simulation execution means and target data for the output of the higher model. Equipped

[0008] The present invention 2 According to this aspect, in the data calculation method, a computer ,versus Imitating elephant parts 、 The above Simulate the target For a lower model that outputs a parameter value of a higher-level model, data indicating a posterior distribution of the parameter values ​​of the lower model based on data indicating a posterior distribution of the parameter values ​​of the higher-level model is calculated. a simulation is performed by applying parameter values ​​based on data indicating a posterior distribution of parameter values ​​of the lower model to the lower model, and a simulation is performed by applying sample data of parameter values ​​of the higher model based on the results of the simulation using the lower model to the higher model, and data indicating the posterior distribution of parameter values ​​of the higher model based on the target data for the output of the higher model is calculated based on the similarity between sample data of the output of the higher model obtained as a result of the simulation using the higher model and target data for the output of the higher model.

[0009] The present invention 3 According to the embodiment, the program is ,versus Imitating elephant parts 、 The above Simulate the target calculating data indicating a posterior distribution of parameter values ​​of a lower model based on data indicating a posterior distribution of parameter values ​​of the higher model for a lower model that outputs parameter values ​​of a higher model; applying parameter values ​​based on data indicating a posterior distribution of parameter values ​​of the lower model to the lower model to perform a simulation; applying sample data of parameter values ​​of the higher model based on a result of the simulation using the lower model to the higher model to perform a simulation; and calculating data indicating a posterior distribution of parameter values ​​of the higher model based on the target data for the output of the higher model based on a similarity between sample data of an output of the higher model obtained as a result of the simulation using the higher model and target data for the output of the higher model. This is the program to execute it. Effect of the Invention

[0010] The above-mentioned data calculation device, data calculation method, and program According to the method, even when there are a large number of parameters whose values ​​are to be estimated based on the output of a simulator, the values ​​can be estimated in a relatively short time. [Brief description of the drawings]

[0011] [Figure 1]1 is a schematic block diagram illustrating a functional configuration of a data calculation device according to an embodiment. [Diagram 2] FIG. 2 is a diagram illustrating an example of a target model used by the data calculation device according to the embodiment. [Diagram 3] 1 is a flowchart illustrating an example of a processing procedure for performing hierarchical data assimilation by a data calculation device according to an embodiment. [Figure 4] 11 is a flowchart illustrating an example of a procedure of a process performed by a data calculation device according to the embodiment for each lower-level model. [Diagram 5] FIG. 1 is a diagram illustrating a configuration of a production system that is an example of a target according to an embodiment. [Figure 6] FIG. 2 is a diagram illustrating an example of a model that simulates a production system according to an embodiment. [Figure 7] FIG. 13 is a diagram illustrating an example of sample parameter values ​​of a top model based on a prior distribution according to an embodiment. [Figure 8] FIG. 11 is a diagram showing a first example of a sample of output values ​​of a higher-level model according to the embodiment. [Figure 9] FIG. 13 is a diagram illustrating an example of sample parameter values ​​of a top model based on a posterior distribution according to an embodiment. [Figure 10] FIG. 13 illustrates an example of sample parameter values ​​for a sub-model based on a prior distribution according to an embodiment. [Figure 11] FIG. 13 is a diagram illustrating an example of a sample of output values ​​of a lower-level model according to the embodiment. [Figure 12] FIG. 13 illustrates an example of sample parameter values ​​for a sub-model based on a posterior distribution according to an embodiment. [Figure 13] FIG. 13 is a diagram illustrating an example of a sample of output values ​​of a lower-level model according to the embodiment. [Figure 14] FIG. 11 is a diagram showing a second example of a sample of output values ​​of a higher-level model according to the embodiment. [Figure 15] FIG. 13 is a diagram showing a third example of sample parameter values ​​of a higher-level model according to the embodiment. [Figure 16] 11A and 11B are diagrams illustrating an example of a result of estimation of parameter values ​​of a lower-level model by the data calculation device according to the embodiment. [Figure 17] FIG. 13 is a diagram illustrating an example of input and output of data in an intermediate layer when the number of layers in the model according to the embodiment is three or more. [Figure 18] FIG. 1 is a diagram illustrating an example of a configuration of a data calculation device according to an embodiment. [Figure 19] FIG. 11 is a diagram illustrating another example of the configuration of the data calculation device according to the embodiment. [Figure 20] 11 is a diagram illustrating an example of a processing procedure in a data calculation method according to the embodiment. [Figure 21] FIG. 1 is a schematic block diagram illustrating a configuration of a computer according to at least one embodiment. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0012] The following describes embodiments of the present invention, but the following embodiments do not limit the scope of the invention. Furthermore, not all of the combinations of features described in the embodiments are necessarily essential to the solution of the invention. Fig. 1 is a schematic block diagram showing a functional configuration of a data calculation device according to an embodiment. In the configuration shown in Fig. 1, the data calculation device 100 includes a communication unit 110, a display unit 120, an operation input unit 130, a storage unit 170, and a control unit 180. The control unit 180 includes an initial setting unit 181, a loop processing unit 182, an upper simulator unit 183, an upper posterior distribution data calculation unit 184, a lower simulator unit 185, and a lower posterior distribution data calculation unit 186. 1 also shows an object 900 that is the subject of simulation by the simulation model. The simulation model of the object 900 is also referred to as a model of the object 900, or simply as a model.

[0013] The data calculation device 100 estimates parameter values ​​of the model using target data for the output of the model so that the output value of the model approaches the value of the target data. The data calculation device 100 may be configured using a computer such as a workstation or a personal computer (PC).

[0014] The data calculation device 100 may be configured as one device or may be configured as a combination of multiple devices. For example, the upper simulator unit 183 and the lower simulator unit 185 may be configured as separate simulator devices for each model. The value of the target data is also called the target value. The process of estimating the parameter values ​​of a model so that the output value of the model approaches the target value is also called data assimilation. Estimating the parameter values ​​in data assimilation is also called parameter value adjustment.

[0015] The target data may be actual measurement data of the object 900. In this case, the data calculation device 100 can estimate parameter values ​​to adjust the model so that the object 900 can be simulated with high accuracy. Moreover, by referring to the parameter value of the estimation result, it is possible to estimate the state of the target 900. For example, when the target 900 is a factory and the parameter of the target to be estimated by the data calculation device 100 is the production volume for each production line of the factory, the parameter value estimated by the data calculation device 100 can be used as the production volume for each production line of the factory estimated from the actual measurement data.

[0016] Alternatively, the target data may be data indicating a target value for the object 900. In this case, a measure for achieving the target can be created by referring to the parameters of the estimation result by the data calculation device 100. Furthermore, the data calculation device 100 can control the object 900, such as controlling each device in a factory based on the created measure.

[0017] For example, when the target 900 is a factory and the parameter to be estimated by the data calculation device 100 is the production volume for each production line of the factory, the parameter value estimated by the data calculation device 100 can be used as the production volume for each production line of the factory to achieve the target value indicated in the target data. According to the data calculation device 100, it is possible to perform control based on the estimation results, such as controlling equipment for each production line in a factory based on the production volume estimated for each production line.

[0018] The data calculation device 100 uses a hierarchical model as a model that simulates the target 900. For example, the data calculation device 100 may use a model that includes a higher-level model that simulates the entire target 900 and one or more lower-level models that simulate a part of the target 900.

[0019] The term "hierarchy" here refers to the ability to define a hierarchical relationship. For example, as described above, based on the relationship between the whole and the parts, a model simulating the entire target 900 can be placed at the top, and a model simulating a part of the target 900 can be placed at the bottom. This hierarchical relationship may be fixed or variable. For example, in a model simulating a company, the accounting department model can be placed at the top in terms of the flow of expenses, but the purchasing department model can be placed at the top in terms of the flow of materials, and so on. The hierarchical relationship may differ depending on the purpose of the data calculation.

[0020] Furthermore, the data calculation device 100 uses a hierarchical model in which a lower model outputs a parameter value of a higher model. As a result, the data calculation device 100 repeats a process of estimating a parameter value of a model based on target data of the model output according to the hierarchical structure of the model.

[0021] For example, the data calculation device 100 estimates parameter values ​​of the higher-level model based on target data of the output of the higher-level model. Then, the data calculation device 100 estimates parameter values ​​of the lower-level model by using the estimated parameter values ​​of the higher-level model as target data of the output of the lower-level model.

[0022] In this way, the data calculation device 100 estimates the parameter values ​​of the models hierarchically according to the hierarchical structure of the models, so that the number of parameters of each model can be reduced compared to the case where the target 900 is simulated by only one model. For example, by using the production volume of each department including multiple production lines as a parameter for the higher-level model, the number of parameters of the higher-level model can be reduced compared to the case where the production volume of each production line is used as a parameter.

[0023] Here, when the number of parameters of a model is large, it may take a long time to estimate the parameter values. In particular, it may take an exponential amount of time to estimate the parameter values ​​as the number of parameters increases. For this reason, when the number of parameters is large, it may be impossible to obtain parameter estimates within a realistic amount of time. In contrast, in the data calculation device 100, by using hierarchical models in which the number of parameters of each model is relatively small as described above, it is expected that the calculation time will be shorter than when one model with a large number of parameters is used.

[0024] The target 900 is a target to be simulated by the data calculation device 100. The target 900 can be anything that can be simulated using a hierarchical simulation model. For example, the target 900 may be a factory. The data calculation device 100 may use a hierarchical model including a higher-level model that simulates the entire factory and one or more lower-level models that simulate a part of the factory. Alternatively, the target 900 may be a supply chain including a production base, a logistics facility, and a sales base. The data calculation device 100 may use a hierarchical model including a higher-level model that simulates the entire supply chain and lower-level models that simulate each of the production base, the logistics facility, and the sales base.

[0025] Alternatively, the target 900 may be a certain region. The data calculation device 100 may use a hierarchical model including a higher-level model that simulates the spread of infection in the entire region and a lower-level model that simulates the spread of infection in a part of the region or in a transportation facility. The object 900 may be real or virtual.

[0026] Fig. 2 is a diagram showing an example of a model of the target 900 used by the data calculation device 100. In the example of Fig. 2, the model 200 includes one higher model 210 and S lower models 220-1 to 220-S. S is an integer S≧1 and indicates the number of lower models 220. The lower models 220-1 to 220-S are also collectively referred to as lower models 220.

[0027] The input to the higher-level model 210 is represented by vector X, the parameters of the higher-level model 210 are represented by vector Z, and the output of the higher-level model 210 is represented by vector Y. Vectors X, Y, and Z are all real-number vectors. The number of dimensions of each of vectors X and Y may be one or more. The number of dimensions of vector Z may be S or more. Vector X is also referred to as input vector X. Vector Y is also referred to as output vector Y. Vector Z is also referred to as parameter vector Z.

[0028] The elements of the input vector X may include elements that take discrete values ​​such as integer values ​​or binary values. Similarly, the elements of the output vector Y may include elements that take discrete values ​​such as integer values ​​or binary values. The elements of the parameter vector Z may include elements that take discrete values ​​such as integer values ​​or binary values. The higher-level model 210 is denoted by f, and is written as f(X, Z) = Y. Alternatively, the notation of the input vector X is omitted and it is written as f(Z) = Y.

[0029] Input to the lower model 220-s is vector X s The parameters of the lower model 220-s are expressed as a vector θ s and the output of the lower model 220-s is represented by vector Zs where s is an integer in the range of 1≦s≦S. Vector X s , Z s , θ s are real vectors. s , vector Z s , θ s The number of dimensions of each of the vectors X s The number of dimensions of may be 0 or more.

[0030] Input vector X s The elements of may be integer or binary, or may be discrete. Similarly, the output vector Z s The elements of the parameter vector θ may include elements that take discrete values ​​such as integer values ​​or binary values. s The elements of may include elements that take discrete values ​​such as integer values ​​or binary values. Lower model 220g s and g s (X s ,θ s )=Z s Alternatively, the input vector X s The notation of g is omitted. s (θ s )=Z s It can also be written as:

[0031] The data calculation device 100 acquires target data indicating a combination of the values ​​of the input vector X and the output vector Y, and adjusts the value of the parameter vector Z so that the output value of the higher-level model 210 approaches the target value. The probability distribution of the parameter vector Z before adjustment by the data calculation device 100 is expressed as a prior distribution π Z and the prior distribution π Z The values ​​of the parameter vector Z are sampled by Z <pri>< / pri> The probability distribution of a vector may be expressed as a combination of the probability distributions for each element of the vector.

[0032] Prior distribution π Zmay be set by the user based on information of the subject 900. Alternatively, the prior distribution π Z A predetermined distribution such as a uniform distribution may be used as the distribution. The probability distribution of the parameter vector Z after adjustment by the data calculation device 100 is expressed as a posterior distribution p(Z|Y), and the value of the parameter vector Z according to the posterior distribution p(Z|Y) is expressed as a sample Z <post>< / post> Here, the posterior distribution p(Z|Y) means the posterior distribution of the value of the parameter vector Z when the target value of the output vector Y is given.

[0033] In addition, the data calculation device 100 calculates the input value (value of the input vector X) to the lower model 220-s based on the input value (value of the input vector X) to the higher model 210. s For example, the input vector X s The vector space of may be a subspace of the vector space of the input vector X. That is, s may be configured using some of the elements of the input vector X. In this case, the data calculation device 100 calculates the input vector X and the input vector X s Depending on the relationship between s Set the element value to . Alternatively, the input vector X s may include calculated values ​​based on the values ​​of the input vector X.

[0034] In addition, the data calculation device 100 calculates sample Z <post>< / post> For example, the output vector Z s The vector space of may be a subspace of the vector space of the parameter vector Z. That is, the output vector Z s may be configured using some of the elements of the parameter vector Z. In this case, the data calculation device 100 calculates the parameter vector Z and the output vector Z s Depending on the relationship between sample Z <post>< / post>The values ​​of the elements of the output vector Z in the target data of the output of the lower model 220-s s Set the element value to .

[0035] Among the elements of the parameter vector Z, the output vector Z s In other words, there may be parameters that are used in the calculations of the higher-level model 210 as parameters of the higher-level model 210, and whose values ​​are not output to the lower-level model 220. The data calculation device 100 calculates sample Z. <post>< / post> Based on the output of the lower model 220-s, the target value of the sample Z <post>< / post> s It is written as follows.

[0036] In this way, the data calculation device 100 can perform hierarchical data assimilation by setting the parameter values ​​of the higher-level model 210 as target data for the output values ​​of the lower-level model 220. Specifically, the data calculation device 100 performs hierarchical data assimilation for the sample Z after adjusting the parameter values ​​in the higher-level model 210. <post>< / post> Based on the target value of the output of the lower model 220 (sample Z <post>< / post> s ), the results of the adjustments made to the higher-level model 210 can be reflected in the adjustments of the parameter values ​​in the lower-level model 220.

[0037] The number of dimensions of the parameter vector Z is the parameter vector θ 1 From θ S Since the number of dimensions is smaller than the total number of dimensions up to the parameter vector θ, the number of parameters of each model can be reduced compared to the case where the object 900 is simulated by only one model. In this respect, the data calculation device 100 can perform data assimilation in a relatively short time, and the parameter vector θ 1 From θ S In particular, it is expected that the data calculation device 100 can avoid the inability to obtain parameter estimates within a realistic time due to the large number of model parameters.

[0038] In addition, the data calculation device 100 calculates an output vector Z s The value of sample Z <post>< / post> s The parameter vector θ s Data assimilation is performed to adjust the value of The parameter vector θ before adjustment by the data calculation device 100 s The probability distribution of is the prior distribution π θs and the prior distribution π θs Parameter vector θ s Sample value of θ <pri>< / pri> s It is written as follows.

[0039] Prior distribution π θs may be set by the user based on information of the subject 900. Alternatively, the prior distribution π θs A predetermined distribution such as a uniform distribution may be used as the distribution. The probability distribution of the parameter vector Z after adjustment by the data calculation device 100 is expressed as the posterior distribution p(θ s |Z s ) and the posterior distribution p(θ s |Z s ) to sample the parameter vector Z values ​​θ <post>< / post> s It is written as follows.

[0040] In addition, the data calculation device 100 calculates the sample θ <post>< / post> s Applying this to the parameters of the lower model 220-s, a simulation is performed, and the input vector X s Output vector Z for s Applying a sample of a parameter to a parameter of a model means setting the sample value as the parameter value. Applying a sample to a parameter of a model is also called applying the sample to the model. <post>< / post> s The output vector Z calculated by the simulation using s Sample value Z <pred>< / pred> s It is written as follows.

[0041] Then, the data calculation device 100 calculates sample Z <pred>< / pred> 1 From sample Z <pred>< / pred> S Based on the values ​​up to , a sample of the value of the parameter vector Z is generated. This sample is used as a sample reflecting the result of data assimilation in the lower model 220 when the data calculation device 100 repeatedly performs data assimilation using the upper model 210.

[0042] Every element of the parameter vector Z corresponds to a sample Z <pred>< / pred> 1 From sample Z <pred>< / pred> S In the case where the element is linked to at most one element among the elements up to <pred>< / pred> 1 From sample Z <pred>< / pred> S The values ​​of the elements up to the point t1 may be used as the values ​​of the elements of the sample to be generated.

[0043] Or, among the elements of the parameter vector Z, sample Z <pred>< / pred> 1 From sample Z <pred>< / pred> S In this case, the data calculation device 100 may select a sample Z <pred>< / pred> 1 From sample Z <pred>< / pred> S Alternatively, the element values ​​of the sample to be generated may be calculated using a predetermined method such as averaging the values ​​of a plurality of elements up to .

[0044] In this way, the data calculation device 100 calculates the sample Z <pred>< / pred> 1 From sample Z <pred>< / pred> S Hierarchical data assimilation can be repeated by generating samples of the parameter vector Z based on the values ​​of Specifically, the data calculation device 100 applies the generated sample values ​​to the parameter vector Z of the higher-level model 210 to execute a simulation. As a result, the data calculation device 100 acquires samples of the output vector Y based on the generated parameter samples. Then, based on the relationship between the obtained samples of the output vector Y and the target value, the data calculation device 100 estimates the value of the parameter vector Z so that the output value of the higher-level model 210 approaches the target value. The data calculation device 100 applies the estimated parameter vector Z value to the sample Z. <post>< / post> As described above, data assimilation is performed in the lower model 220 using the data assimilation result to generate a sample of the parameter vector Z. The data calculation device 100 repeats the hierarchical data assimilation until a predetermined termination condition is met.

[0045] In this way, it is expected that the data calculation device 100 can estimate the parameter values ​​with high accuracy by repeating the hierarchical data assimilation. In particular, the probability distribution of the parameter vector Z and the parameter vector θ 1 From θ S It is expected that each probability distribution from 1 to 1 will converge to a certain parameter value corresponding to the target value. Here, the convergence of the probability distribution to a certain parameter value means that the probability at that value becomes large.

[0046] For example, if data assimilation relatively increases the probability that the value of parameter Z will result in the value of output vector Y being close to the target value, then when data assimilation is performed again, many of the sampled values ​​of output vector Y will be close to the target value. As a result, it is expected that re-data assimilation will further increase the probability that the value of parameter Z will result in the value of output vector Y being close to the target value. Parameter vector θ 1 From θ S It is expected that by repeating data assimilation, the probability of the parameter values ​​corresponding to the target values ​​will increase.

[0047] In the following, an example will be described in which the data calculation device 100 uses kernel ABC and kernel herding as processing for estimating parameter values ​​so as to bring the output value of the model closer to the target value. However, the method by which the data calculation device 100 estimates the parameter value so as to bring the output value of the model closer to the target value is not limited to a specific method. For example, the data calculation device 100 may use Markov Chain Monte Carlo (MCMC) or Approximate Bayesian Computation (ABC) as a process for estimating the parameter value so as to bring the output value of the model closer to the target value.

[0048] The communication unit 110 communicates with other devices. For example, when the data calculation device 100 acquires actual measurement data by a sensor as target data, the communication unit 110 may receive the actual measurement data by the sensor from the sensor installed in the target 900.

[0049] The display unit 120 includes a display screen, such as a liquid crystal panel or an LED (Light Emitting Diode) panel, and displays various images. For example, the display unit 120 may display the estimation results of the parameter values ​​by the data calculation device 100. The operation input unit 130 includes input devices such as a keyboard and a mouse, and receives user operations. For example, the operation input unit 130 may receive a user operation for inputting a target value for the object 900 as target data.

[0050] The storage unit 170 stores various data. For example, the storage unit 170 stores the above-described hierarchical model as a model of the target 900. The storage unit 170 is configured using a storage device included in the data calculation apparatus 100. The control unit 180 performs various processes by controlling each unit of the data calculation device 100. The functions of the control unit 180 may be performed by a CPU (Central Processing Unit) included in the data calculation device 100 reading and executing a program from the storage unit 170.

[0051] The initial setting unit 181 performs initial settings for hierarchical data assimilation. Specifically, the initial setting unit 181 assigns 1 to a variable l that is used as a loop counter for repeating hierarchical data assimilation. In addition, the initial setting unit 181 sets the prior distribution π Z That is, sampling is performed based on the prior distribution π Z Based on sample Z <pri>< / pri> The initial setting unit 181 calculates the sample Z <pri>< / pri> M samples are calculated, where M is an integer greater than or equal to 1 and indicates the number of samples. <pri> (l) < / pri> , {Z <pri> (1) < / pri> m} m=1 M The value in () is the value of the variable l. In other words, the value in () indicates the number of times the hierarchical data assimilation has been performed. <pri> (1) < / pri> m The subscript "m" in "" stands for M samples Z <pri> (1) < / pri> This indicates the identification number that identifies the

[0052] The loop processor 182 performs a loop process that repeats hierarchical data assimilation. Specifically, the loop processor 182 counts up the value of the variable l according to the number of times the hierarchical data assimilation is repeated. The loop processor 182 also determines whether or not the loop termination condition is satisfied. Various conditions can be used as the end condition of the loop here. For example, the end condition of the loop may be that the value of the variable l is equal to or greater than a predetermined threshold. Alternatively, the end condition of the loop may be that the output value of the higher-level model 210 is evaluated to be closer to the target value than a predetermined condition.

[0053] The upper simulator unit 183 executes a simulation using the upper model 210 . In particular, the upper simulator unit 183 calculates the sample Z obtained as a result of the simulation using the lower model 220. <pred>< / pred> 1 From Z <pred>< / pred> S The updated parameter vector Z based on the above is applied to the upper model 210 to perform a simulation. The upper simulator unit 183 corresponds to an example of an upper model simulation execution means.

[0054] Sample Z <pred> (l) < / pred> 1 From Z <pred> (l) < / pred> S Update the parameter vector Z based on the values ​​of the samples Z <pri> (l+1) < / pri> As mentioned above, the value in parentheses is the value of the variable l. The upper simulator unit 183 uses the sample Z obtained in the lth execution of the hierarchical data assimilation. <pred> (l) < / pred> 1 From Z <pred> (l) < / pred> S Based on the sample Z for the l+1th round of hierarchical data assimilation, <pri> (l+1) < / pri> As described above with respect to the data calculation device 100, the upper simulator unit 183 acquires the sample Z <pred> (l) < / pred> s The values ​​of the elements of sample Z <pri> (l+1) < / pri> Alternatively, as described above with respect to the data calculation device 100, the upper simulator unit 183 may use the sample Z <pred>< / pred> 1 From sample Z <pred>< / pred> S Sample Z in a given way, such as taking the average of the values ​​of multiple elements up to <pred> (l+1) < / pred> s The values ​​of the elements of may be calculated.

[0055] The upper simulator section 183 is sample Z <pri> (l+1) < / pri>Applying the above to the top model 210, a simulation is performed, and the sample Y <pri> (l+1) < / pri> Calculate. The upper simulator section 183 is sample Z <pred> (l) < / pred> 1 From Z <pred> (l) < / pred> S Take M combinations up to sample Z <pri> (l+1) < / pri> Calculate M samples and <pri> (l+1) < / pri> The upper simulator unit 183 calculates M samples Z <pri> (l) < / pri> , {Z <pri> (l) < / pri> m} m=1 M The M samples Y calculated by the upper simulator unit 183 are expressed as follows: <pri> (l) < / pri> , {Y <pri> (l) < / pri> m} m=1 M It is written as "Z <pri> (l) < / pri> m "m" in the subscript and "Y <pri> (l) < / pri> m The subscript "m" in each of the subscripts indicates an identification number that identifies the M samples.

[0056] The upper posterior distribution data calculation unit 184 calculates data indicating a posterior distribution of values ​​of the parameter vector Z of the upper model 210 based on the target data for the output vector Y of the upper model 210. The upper posterior distribution data calculation unit 184 corresponds to an example of an upper posterior distribution data calculation means. Specifically, the upper posterior distribution data calculation unit 184 performs data assimilation using the kernel ABC and kernel herding to estimate the value of the parameter vector Z so that the value of the output vector Y approaches the target value. <pri> (l) < / pri> m} m=1 M Depending on the sample Z <post> (l) < / post> As described above, the value in parentheses is the value of the variable l.

[0057] The M samples Z calculated by the upper posterior distribution data calculation unit 184 <post> (l) < / post> , {Z <post> (l) < / post> m} m=1 M It is written as "Z <post> (l) < / post> m The subscript "m" in each of the subscripts refers to the M samples "Z <post> (l) < / post> " indicates an identification number that identifies M samples {Z <post> (l) < / post> m} m=1 M corresponds to an example of data indicating the posterior distribution p(Z|Y) of the values ​​of the parameter vector Z of the higher-level model 210 based on the target data.

[0058] The upper posterior distribution data calculation unit 184 executes the kernel ABC to calculate the target value and {Y <pri> (l) < / pri> m} m=1 M Then, the upper posterior distribution data calculation unit 184 projects each of the sample Y <pri> (l) < / pri> m The upper posterior distribution data calculation unit 184 applies kernel herding to the calculated kernel mean to obtain {Z <post> (l) < / post> m} m=1 M Calculate. The metric in a reproducing kernel Hilbert space is the distance between the samples Y <pri> (l) < / pri> m This is an example of an index that shows the similarity between the sample Y and the target value. <pri> (l) < / pri> m The more similar a is to the target value, the smaller the distance in the reproducing kernel Hilbert space.

[0059] The lower-level simulator unit 185 executes a simulation using the lower-level model 220 . In particular, the lower-level simulator unit 185 calculates the sample θ <post>< / post> s Applying this to the lower model 220-s, a simulation was performed, and the sample Z <pred>< / pred> s The data assimilation in the lower model 220-s calculates the output vector Z s The value of is the target value (sample Z <post>< / post> s ) to approximate the parameter vector θ s This is a data assimilation method to estimate the value of

[0060] The lower simulator unit 185 calculates M samples θ <post>< / post> s Then, we take M samples Z <pred>< / pred> s The lower-level simulator unit 185 corresponds to an example of a lower-level model simulation execution means. The M θ <post>< / post> s θ <post> (l) < / post> m} m=1 M As mentioned above, the value in parentheses is the value of the variable l. <post> (l) < / post> m The subscript "m" in "" refers to M samples θ <post> (l) < / post> This indicates the identification number that identifies the The lower simulator unit 185 calculates M Z <pred>< / pred> s , {Z <pred> (l) < / pred> s,m} m=1 M As mentioned above, the value in parentheses is the value of the variable l. <pred> (l) < / pred> s,m The subscript "m" in "" refers to M samples Z <pred> (l) < / pred> s This indicates the identification number that identifies the

[0061] The lower-level posterior distribution data calculation unit 186 calculates data indicating a posterior distribution of parameter values ​​of the lower-level model 220 based on data indicating a posterior distribution p(Z|Y) of the parameter vector Z of the higher-level model 210. The lower-level posterior distribution data calculation unit 186 corresponds to an example of a lower-level posterior distribution data calculation means. Specifically, the lower-level posterior distribution data calculation unit 186 calculates the output value of the lower-level model 220-s by dividing the output value of the sample Z <post> (l) < / post> The parameter θ s Data assimilation to estimate the value of is performed using kernel ABC and kernel herding. The lower posterior distribution data calculation unit 186 calculates the M samples {θ <post> (l) < / post> m} m=1 M Calculate.

[0062] The control unit 180 may control the target 900, for example, by transmitting a control signal to the target 900 via the communication unit 110. For example, the control unit 180 may control each part of the target 900, such as each device of the target 900, based on the estimation result of the parameters of the model 200 or based on the execution result of a simulation using the model 200. Furthermore, for example, the control unit 180 may control data (sample {θ <post> (l) < / post> m} m=1 M ) based on which the object 900 may be controlled. The control unit 180 corresponds to an example of a control means.

[0063] FIG. 3 is a flowchart showing an example of a processing procedure in which the data calculation device 100 performs hierarchical data assimilation. (Step S101) The initial setting unit 181 initializes the value of a variable l used as a loop counter. Specifically, the initial setting unit 181 sets the value of the variable l to 1. The value of the variable l indicates the number of times the loop from steps S103 to S110 is repeated. After step S101, the process proceeds to step S102.

[0064] (Step S102) The initial setting unit 181 sets the prior distribution π Z M samples from {Z <pri> (1) < / pri> m} m=1 M Calculate. After step S102, the process proceeds to step S103.

[0065] (Step S103) The loop processing unit 182 determines whether or not the loop end condition is met. If the loop processing unit 182 determines that the end condition is not satisfied (step S103: NO), the process proceeds to step S104. On the other hand, if the loop processing unit 182 determines that the end condition is satisfied (step S103: YES), the data calculation device 100 ends the processing of FIG.

[0066] (Step S104) The upper simulator section 183 is sample Z <pri> (l) < / pri> m (m=1, ,M) as the parameters of the upper model 210, and run the simulation. <pri> (l) < / pri> m Calculate Y <pri> (l) < / pri> m is expressed as equation (1).

[0067]

number

[0068] After step S104, the process proceeds to step S105.

[0069] (Step S105) The upper posterior distribution data calculation unit 184 calculates the kernel mean μ (l)Z|Y Calculate the kernel mean μ^ (l) Z|Y is expressed as equation (2).

[0070]

number

[0071] k z indicates the kernel. The "·" in parentheses indicates that the variables of the function in the reproducing kernel Hilbert space are not limited to any particular ones. Weight w m is expressed as equation (3).

[0072]

number

[0073] "I" represents the identity matrix. δ is the regularization constant, δ>0. G y denotes the Gram Matrix, and is expressed as in equation (4).

[0074]

number

[0075] k y (Y) denotes the kernel of y and is expressed as equation (5).

[0076]

number

[0077] The superscript T denotes the transpose of a matrix or vector. After step S105, the process proceeds to step S106.

[0078] (Step S106) The upper posterior distribution data calculation unit 184 calculates the kernel mean μ (l) Z|Y Samples of parameter Z by {Z <post> (l) < / post> m} m=1 M Calculate. Kernel mean μ^ (l) Z|Y is associated one-to-one with the posterior distribution p(Z|Y) when the target data is given as the output vector Y. By performing Kernel Herding processing, samples based on the posterior distribution can be obtained. For example, the upper posterior distribution data calculation unit 184 calculates Z <post> (l) < / post> m Calculate.

[0079]

number

[0080] "argmax Z h j (Z)" is "h j This is a function that outputs the "Z" that maximizes "(Z)". h j is expressed recursively as equation (7).

[0081]

number

[0082] “H” in equation (7) denotes a reproducing kernel Hilbert space. h 0 is expressed as equation (8).

[0083]

number

[0084] After step S106, the process proceeds to step S107.

[0085] (Step S107) The upper posterior distribution data calculation unit 184 calculates the M samples Z <post> (l) < / post> m (m=1, ,M) is the sample Z <post> (l) < / post> For example, the upper posterior distribution data calculation unit 184 aggregates M samples Z <post> (l) < / post> m (m=1, ,M) is the sample Z <post> (l) < / post> Consolidate into.

[0086]

number

[0087] After step S107, the process proceeds to step S108.

[0088] (Step S108) The data calculation device 100 executes the process for each lower model 220. As described above, the data calculation device 100 executes the process for the lower model 220-s by using the sample Z <pred> (l) < / pred> s Calculate. After step S108, the process proceeds to step S109.

[0089] (Step S109) The upper simulator unit 183 transfers the processing result for each lower model 220 to the processing of the upper model 210. Specifically, as described above for the data calculation device 100, the upper simulator unit 183 transfers the sample Z <pred> (l) < / pred> s Based on the value of <pri> (l+1) < / pri> For example, the upper simulator unit 183 sets the value of the sample Z <pri> (l+1) < / pri> Set the value of .

[0090]

number

[0091] After step S109, the process proceeds to step S110.

[0092] (Step S110) The loop processing unit 182 counts up the value of the variable l used as a loop counter. That is, the loop processing unit 182 increments the value of the variable l by 1. After step S110, the process proceeds to step S102.

[0093] 4 is a flowchart showing an example of a procedure of processing performed by the data calculation device 100 for each lower model 220. The data calculation device 100 performs the processing of FIG. 4 in step S108 of FIG. (Step S201) The upper posterior distribution data calculation unit 184 passes the value of the parameter vector Z of the upper model 210 to the lower model 220. Specifically, as described above for the data calculation device 100, the upper posterior distribution data calculation unit 184 passes the value of the parameter vector Z of the upper model 210 to the lower model 220. <post> (l) < / post> Based on sample Z' <post> (l) < / post> s For example, the upper posterior distribution data calculation unit 184 obtains Z' <post> (l) < / post> s Get the.

[0094]

number

[0095] Z <post> (l) < / post> [s,:] is the matrix Z <post> (l) < / post> indicates the row vector in the s-th row of In equation (11), Z <post> (l) < / post> m To distinguish it from the above, a ' is added to Z. Z' <post> (l) < / post> s is expressed as equation (12).

[0096]

number

[0097] Z' <post> (l) < / post> s、m is expressed as equation (13).

[0098]

number

[0099] Z <post> (l) < / post> [s,m] is the matrix Z <post> (l) < / post> indicates the element in the sth row and mth column of In the example of formula (13), Z <post> (l) < / post> is Z' <post> (l) < / post> s、m The upper posterior distribution data calculation unit 184 calculates the upper posterior distribution data as a matrix having elements Z <post> (l) < / post> By reading the s-th row of the data, M samples Z' are obtained to be used as target data for the lower model 220-s. <post> (l) < / post> s、1 ,···,Z' <post> (l) < / post> s、M Get the. After step S201, the process proceeds to step S202.

[0100] (Step S202) The lower simulator unit 185 uses the prior distribution π θs M samples from {θ <pri> (l) < / pri> s,m} m=1 M Calculate. After step S202, the process proceeds to step S203.

[0101] (Step S203) The lower simulator unit 185 detects the sample θ <pri> (l) < / pri> s,m Applying this to the lower model 220-s, a simulation was performed, and the sample Z <pri> (l) < / pri> s,m Calculate the sample Z <pri> (l) < / pri> s,mis expressed as equation (14).

[0102]

number

[0103] After step S203, the process proceeds to step S204.

[0104] (Step S204) The lower posterior distribution data calculation unit 186 calculates the kernel mean μ^ (l) θ|Z,s Calculate the kernel mean μ^ (l) θ|Z,s is expressed as equation (15).

[0105]

number

[0106] k θ indicates the kernel. Weight w s,m is expressed as equation (16).

[0107]

number

[0108] G z,s denotes the Gram matrix and is expressed as in equation (17).

[0109]

number

[0110] K θ,s (Z) denotes the kernel of Z, and is expressed as equation (18).

[0111]

number

[0112] After step S204, the process proceeds to step S205.

[0113] (Step S205) The lower posterior distribution data calculation unit 186 calculates the kernel mean μ^ (l) θ|Z,s Parameter θ by s Sample of {θ <post> (l) < / post> s,m} m=1 M Calculate. For example, the lower posterior distribution data calculation unit 186 calculates θ <post> (l) < / post> s,m Calculate.

[0114]

number

[0115] h s,j is expressed recursively as equation (20).

[0116]

number

[0117] h s,0 is expressed as equation (21).

[0118]

number

[0119] After step S205, the process proceeds to step S206.

[0120] (Step S206) The lower simulator unit 185 detects the sample θ <post> (l) < / post> s,m Set the parameters of the lower model 220-s and run the simulation. <pred> (l) < / pred> s,m Calculate Z <pred> (l) < / pred> s,m is expressed as equation (22).

[0121]

number

[0122] After step S206, the process proceeds to step S207.

[0123] (Step S207) The lower posterior distribution data calculation unit 186 calculates M samples Z <pred> (l) < / pred> s,m (m=1, ,M) is the sample Z <pred> (l) < / pred> s Consolidate into. After step S207, the data calculation device 100 ends the process of FIG.

[0124] The processing performed by the data calculation device 100 will be further described using an example. Fig. 5 is a diagram showing the configuration of a production system 910 corresponding to an example of the target 900. In the configuration shown in Fig. 5, the production system 910 includes a first subsystem 911-1 and a second subsystem 911-2. The first subsystem 911-1 and the second subsystem 911-2 are collectively referred to as the subsystem 911.

[0125] The first subsystem 911-1 includes a first production line 912-1, a second production line 912-2, a third production line 912-3, a fourth production line 912-4, and a fifth production line 912-5. The second subsystem 911-2 includes a sixth production line 912-6, a seventh production line 912-7, an eighth production line 912-8, and a ninth production line 912-9. The first production line 912-1 to the ninth production line 912-9 are collectively referred to as the production lines 912.

[0126] Fig. 6 is a diagram showing an example of a model simulating a production system 910. In the example of Fig. 6, a lower model 220 is provided for each subsystem 911, and the number of lower models 220 is two. In the upper model 210, two parameters z are set according to the two lower models. 1 and z 2 Z=(z 1 ,z 2 )

[0127] In the lower model 220-1, five parameters θ 1,1 From θ 1,5 θ 1 =(θ 1,1 , ,θ 1,5 )

[0128] In the lower model 220-2, four parameters θ 2,1 From θ 2,4 θ 2 =(θ 2,1 , ,θ 2,4 )

[0129] The time required to produce one production unit of each part, such as a product or component, is used as the output and parameter of the model. Parameter z 1 , z 1 indicate the time required for the production of one production unit in the first subsystem 911-1 and the second subsystem 911-2, respectively. Parameter θ 1,1 From θ 1,5 The parameter θ indicates the time required to produce one production unit from the first production line 912-1 to the fifth production line 912-5. 2,1 From θ 2,4The time up to indicates the time required to produce one production unit from the sixth production line 912-6 to the ninth production line 912-9.

[0130] When a target time is set for the target data, this target time and the output from the upper model 210 to the lower model 220 (sample {Z <post> (l) < / post> m} m=1 M ) is considered to indicate a time requirement, such as Takt Time. On the other hand, the output (sample {Z <pred> (l) < / pred> s,m} m=1 M ) is thought to indicate the time required for processing. In this case, it can be seen that the upper model 210 and the lower model 220 exchange data to search for a solution that can meet both the requested time and the necessary time.

[0131] FIG. 7 is a diagram showing an example of sample parameter values ​​of the higher-level model 210 based on a prior distribution. The horizontal axis of FIG. 7 represents the parameter z 1 The vertical axis shows the value of the parameter z 2 Indicates the value of. In area A11, the parameter z 1 In area A12, the prior distribution of the parameter z 2 Each of the points in area A13 has a prior distribution of (z 1 ,z 2 The sample shown in area A13 is an example of a sample of {Z <pri> (1) < / pri> m} m=1 100 This is expressed as: FIG. 7 shows an example of a sample of the parameter Z based on the prior distribution, which is calculated by the initial setting unit 181 in step S102 of FIG.

[0132] Fig. 8 is a diagram showing a first example of samples of output values ​​of higher-level model 210. The horizontal axis of Fig. 8 represents the output value of higher-level model 210. The vertical axis represents the frequency. Fig. 8 shows, in a frequency distribution, the output values ​​of the higher-level model 210 obtained by executing a simulation using the higher-level model 210 for each value of the parameter vector Z shown in Fig. 7. In step S104 in Fig. 3, the higher-level simulator unit 183 performs this simulation. In the processes of steps S105 and S106 in FIG. 5, the upper posterior distribution data calculation unit 184 estimates the value of the parameter Z such that the model output shown in FIG. 8 approaches the actual value "5".

[0133] FIG. 9 is a diagram showing an example of sample parameter values ​​of the higher-level model 210 based on the posterior distribution. The horizontal axis of FIG. 9 represents the parameter z 1 The vertical axis shows the value of the parameter z 2 Indicates the value of. In area A21, the parameter z 1 The posterior distribution of the parameter z is shown in area A22. 2 Each point in the region A23 has a posterior distribution of (z 1 ,z 2 The sample shown in area A23 is an example of a sample of {Z <post> (1) < / post> m} m=1 100 This is expressed as:

[0134] 9 shows an example of parameter values ​​estimated by the upper posterior distribution data calculation unit 184 using the model output and actual values ​​shown in FIG. 8. In the case of a sample of the prior distribution of parameter Z shown in FIG. 7, it is plotted over the entire area A13. In contrast, in the case of a sample of the posterior distribution of parameter Z shown in FIG. 9, z 1 The value of is about 10 to 15, and z 2 Many of the values ​​are plotted in the range of around 10 to 15. It is considered that the value of parameter Z has converged due to data assimilation by the upper posterior distribution data calculation unit 184.

[0135] FIG. 10 is a diagram showing an example of sample parameter values ​​of the sub-model 220 based on the prior distribution. FIG. 10 shows the parameter θ 1,1 From θ 1,5 and the parameters θ of the lower model 220-2 2,1 From θ 2,4 For each of the above, samples based on the prior distribution are shown as a frequency distribution with the parameter value on the horizontal axis and the frequency on the vertical axis. In step S202 of FIG. 4, the lower-level posterior distribution data calculation unit 186 samples parameter values ​​of the lower-level model 220 based on the prior distribution exemplified in FIG.

[0136] Fig. 11 is a diagram showing an example of a sample of output values ​​of the lower model 220. The horizontal axis of Fig. 11 indicates the output value of the lower model 220. The vertical axis indicates the probability. Fig. 11 shows, in the form of a probability distribution, the output values ​​of the lower model 220 obtained by executing a simulation by applying the parameter values ​​shown in Fig. 10 to the lower model 220. In step S203 of Fig. 4, the lower simulator unit 185 performs this simulation.

[0137] FIG. 12 is a diagram showing an example of sample parameter values ​​of the sub-model 220 based on the posterior distribution. FIG. 12 shows the parameter θ 1,1 From θ 1,5 and the parameters θ of the lower model 220-2 2,1 From θ 2,4 For each of the above, samples based on the posterior distribution are shown as a frequency distribution with the parameter value on the horizontal axis and the frequency on the vertical axis.

[0138] In steps S204 and S205 of FIG. 4, the lower-level posterior distribution data calculation unit 186 performs data assimilation to calculate samples of parameter values ​​of the lower-level model 220 based on the posterior distribution. 10 and 12, the parameter values ​​are more unevenly distributed in FIG. 12 than in FIG. 10. 1 and θ 2 It is considered that the value of has converged.

[0139] FIG. 13 is a diagram showing an example of a sample of the output value of the lower model 220. The horizontal axis of FIG. 13 represents the parameter z 1 The vertical axis shows the value of the parameter z 2 Indicates the value of. Fig. 13 shows output values ​​of the lower model 220 when a simulation is performed by applying the parameter values ​​shown in Fig. 12 to the lower model 220. In step S206 of Fig. 4, the lower simulator unit 185 performs this simulation.

[0140] Each point in the area A33 is (z 1 ,z 2 The sample shown in area A33 is an example of a sample of {Z <pred> (1) < / pred> s,m} m=1 100 This is expressed as: In area A31, the parameter z based on the sample shown in area A33 is 1 In area A32, the distribution of parameters z based on the samples shown in area A33 is shown. 2 The distribution of is shown.

[0141] In Fig. 13, the parameter values ​​are more unevenly distributed than in Fig. 9, and z 1 The value of is about 6 to 10, and z 2 Many of the values ​​are plotted in the range from 7 to 15. It is believed that the value of parameter Z is further converged by reflecting the results of data assimilation by the lower-order posterior distribution data calculation unit 186.

[0142] Fig. 14 is a diagram showing a second example of samples of output values ​​of higher-level model 210. The horizontal axis of Fig. 14 represents the output value of higher-level model 210. The vertical axis represents the frequency. Fig. 14 shows, in the form of a frequency distribution, an example of the output of the higher-level model 210 when a simulation is executed by applying the parameter values ​​shown in Fig. 13 to the higher-level model 210. The higher-level simulator unit 183 performs this simulation in step S104 in the second execution of the loop from steps S103 to S110 in Fig. 3. 8. The value of parameter Z converges more in the case of FIG. 13 than in the case of FIG. 7, and accordingly the output of higher level model 210 converges more in the case of FIG. 14 than in the case of FIG.

[0143] FIG. 15 is a diagram showing a third example of sample parameter values ​​of the higher-level model 210. The horizontal axis of FIG. 15 represents the parameter z 1 The vertical axis shows the value of the parameter z 2 Show the value of Fig. 15 shows parameter values ​​of the higher-level model 210 when data assimilation is performed to estimate parameter values ​​so that the output values ​​of the higher-level model 210 approach the actual values ​​based on the output values ​​of the higher-level model 210 shown in Fig. 14. In steps S104 to S105 during the second execution of the loop of steps S103 to S110 in Fig. 3, the higher-level posterior distribution data calculation unit 184 performs this data assimilation.

[0144] Each point in the area A43 is (z 1 ,z 2 The sample shown in area A33 is an example of a sample of {Z <pred> (1) < / pred> s,m} m=1 100 This is expressed as: In area A41, the parameter z based on the sample shown in area A43 is 1 In area A42, the distribution of parameters z based on the samples shown in area A33 is shown. 2 The distribution of is shown.

[0145] In Fig. 15, the parameter values ​​are more unevenly distributed than in Fig. 13, and z 1 The value of is about 9, and z 2 The points are concentrated in the area where the value is around 12. It is considered that the data calculation device 100 has succeeded in estimating an appropriate value for the parameter Z.

[0146] FIG. 16 is a diagram showing an example of the result of estimation of parameter values ​​of the lower-level model 220 by the data calculation device 100. As shown in FIG. Figure 16 shows an example of parameter values ​​of the lower model 220 when data assimilation is performed to estimate the parameter values ​​of the lower model 220 so that the output value of the lower model 220 approaches the target value, with the parameter value on the horizontal axis and the frequency on the vertical axis.

[0147] In steps S204 and S205 in the second execution of the process in FIG. 4, the lower posterior distribution data calculation unit 186 performs this data assimilation. In FIG. 16, the parameter values ​​are more unevenly distributed than in FIG. 12. 1 and θ 2 It is believed that the value of was successfully estimated.

[0148] The model 200 may be configured as a model with three or more layers. In this case, the model in the highest layer is referred to as a high-level model, the model in the lowest layer is referred to as a low-level model, and the models in the other layers are referred to as intermediate models. Even if the number of layers of the model is three or more, the processing for the upper model is the same as that explained with reference to Fig. 3. If the model that directly inputs and outputs data to and from the upper model is an intermediate model, the "lower model" in the explanation of Fig. 3 should be read as the intermediate model.

[0149] Even if the number of layers of a model is three or more, the processing for the lower model is the same as that explained with reference to Fig. 4. If the model that directly inputs and outputs data to and from the lower model is an intermediate model, the "upper model" in the explanation of Fig. 4 should be read as the intermediate model.

[0150] For the intermediate layer, in addition to the process of FIG. 4, the processes of steps S107 to S109 in FIG. 3 are performed between steps S205 and S206. Here, the input and output data in the intermediate layer is expressed as shown in FIG.

[0151] Fig. 17 is a diagram showing an example of data input / output in an intermediate layer when the number of layers of a model is three or more. In the example of Fig. 17, intermediate model 232 inputs / outputs data with upper layer model 231 and lower layer model 233. Upper layer model 231 is a model in the layer immediately above intermediate model 232. Upper layer model 231 may be an upper model or an intermediate model.

[0152] The lower layer model 233 is a model in the layer immediately below the intermediate model 232. The lower layer model 233 may be a lower model or an intermediate model. The number of intermediate models that input and output data to and from one upper layer model 231 may be one or more. The number of lower layer models that input and output data to and from one intermediate model 232 may be one or more. The number of layers of models does not have to be uniform. For example, there may be both an intermediate model and a lower model as models in a layer immediately below a higher-level model. There may also be both an intermediate model and a lower model as models in a layer immediately below an intermediate model.

[0153] Regarding the input vector, the data calculation device 100 calculates the input vector X (u) , and the input vector X (d) This is because the data calculation device 100 sets the input vector X s This corresponds to performing the process of setting the above in a hierarchical manner.

[0154] Regarding the samples of the model parameters, the samples output from the upper layer model 231 to the intermediate model 232 are expressed as Z' <post>< / post> (u)The sample output from the intermediate model 232 to the lower layer model 233 is denoted as Z' <post>< / post> (d) It is written as follows. Z' <post>< / post> (u) and Z' <post>< / post> (d) is Z' in the process of Figure 4. <post> (l) < / post> s、1 ,···,Z' <post> (l) < / post> s、M Equivalent to Z' <post>< / post> (u) is used as the target data in the data assimilation in the intermediate model 232. <post>< / post> (u) is used as target data in data assimilation in the lower layer model 233.

[0155] The sample output from the lower layer model 233 to the intermediate model 232 is Z <pred>< / pred> (d) The sample output from the intermediate model 232 to the upper layer model 231 is denoted as Z <pred>< / pred> (u) It is written as follows. Z <pred>< / pred> (d) and Z <pred>< / pred> (u) is Z in the process of Fig. 3 and the process of Fig. 4 <pred> (l) < / pred> s Equivalent to Z <pred>< / pred> (d) is the sample Z based on the posterior distribution in the intermediate model 232 <post>< / post> (d) The posterior distribution-based samples here are samples obtained by data assimilation, and correspond to parameter samples based on the posterior distribution when target data for the model output values ​​is given.

[0156] If the upper layer model 231 is an intermediate model, Z <pred>< / pred> (u) is the sample Z based on the posterior distribution in the upper layer model 231 <post>< / post> (d)On the other hand, if the upper layer model 231 is a higher model, Z <pred>< / pred> (u) is used to generate a sample based on the prior distribution for the next loop execution of the upper layer model 231.

[0157] The data calculation device 100 calculates the output Z u The value of target data Z <post>< / post> (u) parameter Z (d) As a result, the data calculation device 100 performs data assimilation to estimate the value of sample Z <post>< / post> (d) Calculate.

[0158] Then, the data calculation device 100 performs processing corresponding to steps S107 to S109 in FIG. 3 to generate a sample Z <pred>< / pred> (d) Calculate. In step S109 of FIG. 3, the data calculation device 100 calculates Z (PRED)(l) s Using Z <pri> (l+1) < / pri> On the other hand, in the process for the intermediate model 232, the data calculation device 100 calculates Z <pred>< / pred> (d) Using the posterior distribution, we estimate the sample Z <post>< / post> (d) Update.

[0159] Then, the data calculation device 100 performs processing corresponding to steps S206 to S207 in FIG. 4 to generate the updated sample Z <post>< / post> (d) We apply the intermediate model 232 to perform a simulation and Z <pred>< / pred> (u) Calculate. In this way, in the process of the intermediate model 232, after the data assimilation in the intermediate model 232, the result of the data assimilation in the lower layer model 233 is used as the sample Z <post>< / post> (d)Then, Z <pred>< / pred> (u) In other respects, the process relating to the intermediate model is similar to the process in FIG.

[0160] As described above, the upper posterior distribution data calculation unit 184 calculates data {Z <post> (l) < / post> m} m=1 M The lower posterior distribution data calculation unit 186 calculates data {Z <post> (l) < / post> m} m=1 M The data {θ <post> (l) < / post> m} m=1 M Calculate.

[0161] The lower posterior distribution data calculation unit 186 calculates data indicating the posterior distribution of the parameter values ​​of the lower model 220 using the parameter values ​​of the higher model 210, thereby enabling data assimilation to be performed hierarchically. <post>< / post> The target value of the output of the lower model 220 (sample Z <post>< / post> s ), the result of adjusting the parameter values ​​in the higher model 210 can be reflected in the adjustment of the parameter values ​​in the lower model 220.

[0162] The number of dimensions of the parameter vector Z is the parameter vector θ 1 From θ S Since the number of dimensions is smaller than the total number of dimensions up to the parameter vector θ, the number of parameters of each model can be reduced compared to the case where the object 900 is simulated by only one model. In this respect, the data calculation device 100 can perform data assimilation in a relatively short time, and the parameter vector θ1 From θ S In particular, it is expected that the data calculation device 100 can avoid the inability to obtain parameter estimates within a realistic time period due to the large number of model parameters.

[0163] In addition, there is no direct data exchange between the lower-level models 220, and each of the lower-level models 220 can perform parallel processing. In this respect, the data calculation device 100 is also expected to be able to perform data assimilation in a relatively short time.

[0164] The lower-level simulator unit 185 also calculates data {θ <post> (l) < / post> m} m=1 M Parameter value θ based on <post>< / post> The upper simulator unit 183 applies the result of the simulation using the lower model 220, {Z <pred> (l) < / pred> s,m} m=1 M The parameter values ​​of the top model 210 based on sample data {Z <pri> (l+1) < / pri> m} m=1 M The upper model 210 is applied to perform a simulation. The upper posterior distribution data calculation unit 184 calculates the sample data {Y <pri> (l+1) < / pri> m} m=1 M and the target data, data {Z <post> (l+1) < / post> m} m=1 M Calculate.

[0165] In this way, in the data calculation device 100, the result of data assimilation in the lower model 220 can be reflected in the parameter values ​​of the upper model 210, and this allows hierarchical data assimilation to be repeated. It is expected that the data calculation device 100 can estimate the parameter values ​​with high accuracy by repeating hierarchical data assimilation. In particular, the probability distribution of the parameter vector Z and the parameter vector θ 1 From θ S It is expected that each of the probability distributions from 1 to 1 will converge to a certain parameter value corresponding to the target value. As described above, when a probability distribution converges to a certain parameter value, the probability of that value increases.

[0166] For example, if data assimilation relatively increases the probability that the value of parameter Z will result in the value of output vector Y being close to the target value, then when data assimilation is performed again, many of the sampled values ​​of output vector Y will be close to the target value. As a result, it is expected that re-data assimilation will further increase the probability that the value of parameter Z will result in the value of output vector Y being close to the target value. Parameter vector θ 1 From θ S It is expected that by repeating data assimilation, the probability of the parameter values ​​corresponding to the target values ​​will increase.

[0167] In addition, when the upper simulator unit 183 executes a simulation, the output of the upper model 210 {Y <pri> (l+1) < / pri> m} m=1 M It is considered that, depending on the result of data assimilation in the lower model 220, it is concentrated in the vicinity of a specific value. <pri> (l+1) < / pri> m} m=1 MBy performing data assimilation using the above, parameter values ​​that are a solution for data assimilation are searched for intensively in the vicinity of a specific value of the output of the upper model 210. In this respect, the data calculation device 100 is expected to be able to estimate parameter values ​​more efficiently than a case where a single-layer model is used to uniformly search the parameter space.

[0168] In addition, the upper posterior distribution data calculation unit 184 performs kernel herding based on the kernel average of the sample data of the parameter values ​​of the upper model 210, and calculates data {Z <post> (l) < / post> m} m=1 M Calculate.

[0169] In this way, the upper posterior distribution data calculation unit 184 calculates data indicating the posterior distribution of the parameter values ​​of the upper model 210 based on the kernel mean, so that it is possible to calculate M samples of the posterior distribution, which is the same number as M samples of the prior distribution given. In this respect, the data calculation device 100 can efficiently estimate the parameter values.

[0170] In addition, the lower-level posterior distribution data calculation unit 186 executes kernel herding based on the kernel average of the sample data of the parameter values ​​of the lower-level model 220, and calculates data {θ <post> (l) < / post> m} m=1 M Calculate. In this way, the lower posterior distribution data calculation unit 186 calculates data indicating the posterior distribution of the parameter values ​​of the lower model 220 based on the kernel mean, so that it is possible to calculate M posterior distribution samples, which is the same number as M given prior distribution samples. In this respect, the data calculation device 100 can efficiently estimate the parameter values.

[0171] Furthermore, the control unit 180 controls the target 900 based on data indicating the posterior distribution of the parameter values ​​of the lower model 220 calculated by the lower posterior distribution data calculation unit 186 . As described above, the data calculation device 100 uses a hierarchical model, so that the number of parameters of each model is relatively small. As a result, even if the target 900 is relatively complicated, the lower-level posterior distribution data calculation unit 186 can estimate the parameter values ​​of the lower-level model 220. The control unit 180 can grasp the state of the target 900 and control the target 900 with relatively high accuracy by referring to the estimation result of the parameter values ​​by the lower-level posterior distribution data calculation unit.

[0172] 18 is a diagram illustrating an example of the configuration of a data calculation device according to an embodiment. In the configuration illustrated in FIG. 18, a data calculation device 610 includes an upper posterior distribution data calculation unit 611 and a lower posterior distribution data calculation unit 612. In this configuration, the upper posterior distribution data calculation unit 611 calculates data indicating the posterior distribution of parameter values ​​of the upper model based on target data for the output of the upper model that simulates the target. The lower posterior distribution data calculation unit 612 calculates data indicating the posterior distribution of parameter values ​​of the lower model based on data indicating the posterior distribution of parameter values ​​of the higher model for the lower model that simulates a part of the target and outputs parameter values ​​of the higher model. The upper posterior distribution data calculation unit 611 corresponds to an example of an upper posterior distribution data calculation means, and the lower posterior distribution data calculation unit 612 corresponds to an example of a lower posterior distribution data calculation means.

[0173] The lower posterior distribution data calculation unit 612 calculates data indicating the posterior distribution of parameter values ​​of the lower model using parameter values ​​of the higher model, thereby enabling data assimilation to be performed hierarchically. Specifically, the lower posterior distribution data calculation unit 612 uses a target value of the output of the lower model, which is determined based on a sample of parameter values ​​after data assimilation in the higher model, thereby enabling the result of data assimilation in the higher model to be reflected in data assimilation in the lower model.

[0174] By making the number of parameters of the higher-level model smaller than the total number of parameters of the lower-level models, the number of parameters of each model can be reduced compared to the case where the target is simulated by only one model. In this respect, it is expected that the data calculation device 610 can perform data assimilation in a relatively short time and estimate the parameter values. In particular, it is expected that the data calculation device 610 can avoid the situation where the estimated values ​​of the parameters cannot be obtained within a realistic time due to the large number of parameters of the model.

[0175] In addition, there is no direct data exchange between the lower-level models, and each of the lower-level models can be processed in parallel. In this respect, the data calculation device 610 is also expected to be able to perform data assimilation in a relatively short time.

[0176] The upper posterior distribution data calculation unit 611 can be realized, for example, by using the function of the upper posterior distribution data calculation unit 184 shown in Fig. 1. The lower posterior distribution data calculation unit 612 can be realized, for example, by using the function of the lower posterior distribution data calculation unit 186 shown in Fig. 1.

[0177] 19 is a diagram showing another example of the configuration of a data calculation device according to an embodiment. In the configuration shown in FIG. 19, a data calculation device 620 includes a lower model simulation execution unit 621 and a higher model simulation execution unit 622. In this configuration, the lower model simulation execution unit 621 executes a simulation by applying data showing the distribution of parameter values ​​of the lower model to the lower model that outputs parameter values ​​of the higher model that simulates the target, and calculates data showing the distribution of parameter values ​​of the higher model. The higher model simulation execution unit 622 executes a simulation of the target by applying the calculated data showing the distribution of parameter values ​​to the higher model.

[0178] The lower model simulation execution unit 621 corresponds to an example of a lower model simulation execution means, and the higher model simulation execution unit 622 corresponds to an example of a higher model simulation execution means. According to the data calculation device 620, the estimation results of the parameter values ​​using the lower model can be reflected in the higher model, and in this respect, the target simulation can be executed with high accuracy.

[0179] The lower model simulation execution unit 621 can be realized, for example, by using the functions of the lower simulator unit 185 shown in Fig. 1. The upper model simulation execution unit 622 can be realized, for example, by using the functions of the upper simulator unit 183 shown in Fig. 1.

[0180] Fig. 20 is a diagram showing an example of a processing procedure in a data calculation method according to an embodiment. The data calculation method shown in Fig. 20 includes a step of calculating data showing a posterior distribution for a higher-level model (step S611) and a step of calculating data showing a posterior distribution for a lower-level model (step S612).

[0181] In a step of calculating data showing a posterior distribution for the higher-level model (step S611), data showing a posterior distribution of parameter values ​​of the higher-level model based on target data for the output of the higher-level model simulating the target is calculated. In the process of calculating data indicating the posterior distribution for the lower model (step S612), data indicating the posterior distribution of the parameter values ​​of the lower model, which simulates the target part and outputs the parameter values ​​of the higher model, is calculated based on data indicating the posterior distribution of the parameter values ​​of the higher model.

[0182] According to the data calculation method shown in Fig. 20, data assimilation can be performed hierarchically by calculating data indicating the posterior distribution of parameter values ​​of a lower model using parameter values ​​of a higher model. Specifically, by using a target value of the output of a lower model, which is determined based on a sample of parameter values ​​after data assimilation in the higher model, the result of data assimilation in the higher model can be reflected in data assimilation in the lower model.

[0183] By making the number of parameters of the higher-level model smaller than the total number of parameters of the lower-level models, the number of parameters of each model can be reduced compared to the case where the target is simulated with only one model. In this respect, the data calculation method shown in FIG. 20 is expected to enable data assimilation in a relatively short time and to estimate parameter values. In particular, the data calculation method shown in FIG. 20 is expected to avoid the inability to obtain parameter estimates within a realistic time due to the large number of model parameters.

[0184] In addition, there is no direct data exchange between the lower-level models, and each lower-level model can be processed in parallel. With the data calculation method shown in Fig. 20, it is expected that data assimilation can be performed in a relatively short time.

[0185] FIG. 21 is a schematic block diagram illustrating a configuration of a computer according to at least one embodiment. In the configuration shown in FIG. 21, a computer 700 includes a CPU 710, a main memory device 720, an auxiliary memory device 730, an interface 740, and a non-volatile recording medium 750.

[0186] Any one or more of the data calculation device 100 and the data calculation device 610 or a part thereof may be implemented in the computer 700. In this case, the operation of each of the above-mentioned processing units is stored in the auxiliary storage device 730 in the form of a program. The CPU 710 reads the program from the auxiliary storage device 730, develops it in the main storage device 720, and executes the above-mentioned processing according to the program. The CPU 710 also secures a storage area corresponding to each of the above-mentioned storage units in the main storage device 720 according to the program. Communication between each device and other devices is executed by the interface 740 having a communication function and performing communication according to the control of the CPU 710. The interface 740 also has a port for the non-volatile recording medium 750, and reads information from the non-volatile recording medium 750 and writes information to the non-volatile recording medium 750.

[0187] When the data calculation device 100 is implemented in a computer 700, the operations of the control unit 180 and each of its units are stored in the form of a program in the auxiliary storage device 730. The CPU 710 reads the program from the auxiliary storage device 730, loads it in the main storage device 720, and executes the above-mentioned processing according to the program.

[0188] Furthermore, the CPU 710 allocates a storage area corresponding to the storage unit 170 in the main storage device 720 in accordance with the program. Communication with other devices by the communication unit 110 is performed by an interface 740 having a communication function and operating under the control of the CPU 710. Display on the display unit 120 is performed by having an interface 740 with a display device and displaying various images under the control of the CPU 710. The operation input unit 130 receives a user operation by an interface 740 having input devices such as a keyboard and a mouse, and outputs information indicating the received user operation to the CPU 710.

[0189] When the data calculation device 610 is implemented in the computer 700, the operations of the upper posterior distribution data calculation unit 611 and the lower posterior distribution data calculation unit 612 are stored in the form of a program in the auxiliary storage device 730. The CPU 710 reads the program from the auxiliary storage device 730, loads it in the main storage device 720, and executes the above-mentioned processing according to the program.

[0190] Furthermore, the CPU 710 allocates a storage area in the main storage device 720 for the processing performed by the data calculation device 610 in accordance with the program. Communication between the data calculation device 610 and other devices is performed by the interface 740 having a communication function and operating under the control of the CPU 710. Interaction between the data calculation device 610 and a user is performed by the interface 740 having an input device and an output device, presenting information to the user via the output device under the control of the CPU 710, and accepting user operations via the input device.

[0191] Any one or more of the above-mentioned programs may be recorded in non-volatile recording medium 750. In this case, interface 740 may read the program from non-volatile recording medium 750. Then, CPU 710 may directly execute the program read by interface 740, or may temporarily store the program in main storage device 720 or auxiliary storage device 730 and execute it.

[0192] Note that a program for executing all or part of the processing performed by data calculation device 100 and data calculation device 610 may be recorded in a computer-readable recording medium, and the program recorded in the recording medium may be read into a computer system and executed to perform processing of each part. Note that the term "computer system" here includes hardware such as an OS and peripheral devices. Furthermore, the term "computer-readable recording medium" refers to portable media such as flexible disks, optical magnetic disks, ROMs (Read Only Memory), and CD-ROMs (Compact Disc Read Only Memory), as well as storage devices such as hard disks built into computer systems. The above-mentioned program may be for realizing part of the above-mentioned functions, or may be capable of realizing the above-mentioned functions in combination with a program already recorded in the computer system.

[0193] Although an embodiment of the present invention has been described in detail above with reference to the drawings, the specific configuration is not limited to this embodiment, and designs that do not deviate from the gist of the present invention are also included. [Industrial Applicability]

[0194] The embodiment of the present invention may be applied to a data calculation device, a data calculation method, and a recording medium. [Explanation of symbols]

[0195] 100, 610 Data calculation device 110 Communications Department 120 Display section 130 Operation input section 170 Storage section 180 Control Unit 181 Initial setting section 182 Loop processing part 183 Upper Simulator Section 184, 611 Top posterior distribution data calculation section 185 Lower Simulator Section 186, 612 Sub-posterior distribution data calculation section

Claims

1. A lower model that simulates a part of an object and outputs parameter values ​​of a higher model that simulates the object, said lower model having a posterior distribution data calculation means for calculating data indicating a posterior distribution of parameter values ​​of said lower model based on data indicating a posterior distribution of parameter values ​​of said higher model; a lower-level model simulation execution means for executing a simulation by applying parameter values ​​based on data indicating a posterior distribution of parameter values ​​of the lower-level model to the lower-level model; a higher-level model simulation execution means for executing a simulation by applying sample data of parameter values ​​of the higher-level model based on a result of a simulation using the lower-level model to the higher-level model; an upper-level posterior distribution data calculation means for calculating data indicating a posterior distribution of parameter values ​​of the upper-level model based on target data for the output of the upper-level model, based on a similarity between sample data of the output of the upper-level model obtained as a result of the simulation by the upper-level model simulation execution means and target data for the output of the upper-level model; A data calculation device comprising:

2. the upper-level posterior distribution data calculation means performs kernel herding based on a kernel average of sample data of parameter values ​​of the upper-level model, and calculates data indicating a posterior distribution of parameter values ​​of the upper-level model; The data calculation device according to claim 1 .

3. the lower-level posterior distribution data calculation means performs kernel herding based on a kernel average of sample data of parameter values ​​of the lower-level model, and calculates data indicating a posterior distribution of parameter values ​​of the lower-level model. The data calculation device according to claim 1 or 2.

4. A control means for controlling the object based on data indicating a posterior distribution of parameter values ​​of the lower-level model calculated by the lower-level posterior distribution data calculation means. The data calculation device according to claim 1 , further comprising:

5. The computer For a lower model that simulates a portion of an object and outputs parameter values ​​of a higher-level model that simulates the object, calculating data indicating a posterior distribution of parameter values ​​of the lower model based on data indicating a posterior distribution of parameter values ​​of the higher-level model; Executing a simulation by applying parameter values ​​based on data indicating a posterior distribution of parameter values ​​of the lower-level model to the lower-level model; applying sample data of parameter values ​​of the higher-level model based on a result of a simulation using the lower-level model to the higher-level model to perform a simulation; calculating data indicating a posterior distribution of parameter values ​​of the higher-level model based on the target data for the output of the higher-level model, based on a similarity between sample data of the output of the higher-level model obtained as a result of a simulation using the higher-level model and target data for the output of the higher-level model; Data calculation method.

6. On the computer, calculating data indicating a posterior distribution of parameter values ​​of a lower model based on data indicating a posterior distribution of parameter values ​​of the higher model for a lower model that simulates a portion of the target and outputs parameter values ​​of the higher model that simulates the target; applying parameter values ​​based on data indicating a posterior distribution of parameter values ​​of the lower-level model to the lower-level model to perform a simulation; applying sample data of parameter values ​​of the higher-level model based on a result of a simulation using the lower-level model to the higher-level model to perform a simulation; calculating data indicating a posterior distribution of parameter values ​​of the higher-level model based on the target data for the output of the higher-level model, based on a similarity between sample data of the output of the higher-level model obtained as a result of a simulation using the higher-level model and target data for the output of the higher-level model; A program for executing.

Citation Information

Patent Citations

  • Simulation device, simulation method, and memory medium

    WO2016031174A1

  • Information processing device, information processing system, information processing method, and non-transitory computer-readable medium having program stored therein

    WO2020105468A1