Estimation device, estimation method, and program
The estimation device optimizes OSSE calculations by switching between high- and low-cost models based on accuracy thresholds, reducing computational complexity and maintaining high-accuracy estimates.
Patent Information
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- NT T INC
- Filing Date
- 2024-12-05
- Publication Date
- 2026-06-11
Smart Images
Figure JP2024043089_11062026_PF_FP_ABST
Abstract
Description
Estimation Device, Estimation Method, and Program
[0001] The present disclosure relates to an estimation device, an estimation method, and a program.
[0002] An Observing System Simulation Experiment (OSSE) is known. An OSSE constructs a non-existent observing system on a computer and evaluates its behavior. An OSSE may also be used to evaluate a data assimilation system that fuses observations and numerical models.
[0003] Non-Patent Document 1 discloses a system for improving the computational efficiency of data assimilation. In the ensemble calculation in the Ensemble Kalman Filter (EnKF), Non-Patent Document 1 appropriately combines a Multi-Fidelity model and a Low-Fidelity model to reduce the overall computational cost while ensuring the required accuracy. The Multi-Fidelity model has high precision and high computational cost. The Low-Fidelity Model has low precision and low computational cost.
[0004] Jeffrey van der Voort, 2 others, "A Multi-Fidelity Ensemble Kalman Filter (MF-EnKF) with a machine learned surrogate model", June 18th, 2024
[0005] An OSSE comprehensively performs simulations under each condition combining the values of a plurality of variables. There is a problem that the amount of calculation in an OSSE becomes enormous.
[0006] The present disclosure has been made in view of the above circumstances, and an object of the present disclosure is to provide a technology capable of reducing the amount of calculation in an OSSE.
[0007] An estimation device according to one aspect of the present disclosure includes a high-cost model calculation unit that estimates the value of each variable at each time step at high cost using a high-cost model, a low-cost model calculation unit that estimates the value of each variable at each time step at low cost using a low-cost model, and a control unit that controls the processing of the high-cost model calculation unit and the low-cost model calculation unit. The control unit requests the high-cost model calculation unit to process each time step from the start time to the evaluation time, calculates an estimated value of the effect amount of the reference variable at the evaluation time from the value of the reference variable at the evaluation time, and if the estimated value of the effect amount is within an acceptable range, requests the high-cost model calculation unit to process each time step from after the evaluation time to the end time, and if the estimated value of the effect amount is outside an acceptable range, requests the low-cost model calculation unit to process each time step from after the evaluation time to the end time.
[0008] An estimation method in one aspect of the present disclosure comprises a computer comprising: a high-cost model calculation unit that estimates the value of each variable at each time step at high cost using a high-cost model; a low-cost model calculation unit that estimates the value of each variable at each time step at low cost using a low-cost model; and a control unit that controls the processing of the high-cost model calculation unit and the low-cost model calculation unit, wherein the control unit requests the high-cost model calculation unit to process for each time step from the start time to the evaluation time, calculates an estimated value of the effect amount of the reference variable at the evaluation time from the value of the reference variable at the evaluation time, and if the estimated value of the effect amount is within an acceptable range, requests the high-cost model calculation unit to process for each time step from after the evaluation time to the end time, and if the estimated value of the effect amount is outside an acceptable range, requests the low-cost model calculation unit to process for each time step from after the evaluation time to the end time.
[0009] A program in one aspect of the present disclosure causes a computer to function as a high-cost model calculation unit that estimates the value of each variable at each time step at high cost using a high-cost model, a low-cost model calculation unit that estimates the value of each variable at each time step at low cost using a low-cost model, and a control unit that controls the processing of the high-cost model calculation unit and the low-cost model calculation unit. The control unit requests the high-cost model calculation unit to process each time step from the start time to the evaluation time, calculates an estimated value of the effect size of the reference variable at the evaluation time from the value of the reference variable at the evaluation time, and if the estimated value of the effect size is within an acceptable range, requests the high-cost model calculation unit to process each time step from after the evaluation time to the end time, and if the estimated value of the effect size is outside an acceptable range, requests the low-cost model calculation unit to process each time step from after the evaluation time to the end time.
[0010] This disclosure provides a technology that can reduce the computational complexity in OSSE.
[0011] Figure 1 is a diagram illustrating the functional blocks of the estimation device of this disclosure. Figure 2 is a diagram illustrating an example of the data structure of the initial value data. Figure 3 is a diagram illustrating an example of the data structure of the estimated value data. Figure 4 is a flowchart illustrating the estimation method. Figure 5 is a diagram illustrating the hardware configuration of the computer used in the estimation device.
[0012] Embodiments of this disclosure will be described below with reference to the drawings. In the drawings, the same parts are denoted by the same reference numerals and their descriptions are omitted.
[0013] The estimation device 1 of this disclosure will be described with reference to Figure 1.
[0014] The estimation device 1 includes initial value data 11, time condition data 12, judgment condition data 13, and estimated value data 16, as well as the functions of an estimation unit 20, a control unit 30, and an output unit 40. Each piece of data is stored in a storage device such as a memory 902 or storage 903. Each function is implemented in the CPU 901.
[0015] The initial value data 11 includes initial values for each variable, which the estimation unit 20 uses to estimate the value of each variable at each time step. As shown in Figure 2, the initial value data 11 includes initial values for each variable, such as date and time, velocity field, salinity, water temperature, and water vapor content.
[0016] The time condition data 12 includes the evaluation time Tp and the end time Te. The evaluation time Tp and end time Te may be expressed in any format, such as the format of the time to be estimated and the format of the time step to be estimated.
[0017] The evaluation time Tp is the time at which the control unit 30 determines which calculation unit of the estimation unit 20 to perform processing. The time condition data 12 includes one or more evaluation times Tp. The end time Te is the time at which the control unit 30 terminates the estimation performed by the estimation unit 20.
[0018] The determination condition data 13 includes conditions for the control unit 30 to determine which calculation unit of the estimation unit 20 should perform the processing. The determination condition data 13 includes the variable of interest Q and the tolerance error.
[0019] The variable of interest Q is a variable that the control unit 30 focuses on in order to determine which calculation unit of the estimation unit 20 should perform the processing. The tolerance is the allowable error for the estimated value of the variable of interest Q. The error is, for example, the RMSE (Root Mean Squared Error).
[0020] The estimated value data 16 includes the estimated values output by the estimation unit 20. The estimated value data 16 includes estimated values for each time step, as shown in Figure 3, for example. In Figure 3, the estimated values include the estimated values for each variable.
[0021] The estimation unit 20 performs an OSSE data assimilation experiment (simulation). For each time step, the estimation unit 20 estimates the value of each variable using the OSSE estimation model and outputs it to the estimated value data 16. The estimation unit 20 has multiple estimation models.
[0022] The estimation unit 20 includes a high-cost model calculation unit 21 and a low-cost model calculation unit 22.
[0023] The high-cost model calculation unit 21 estimates the values of each variable at each time step at high cost using a high-cost model. The high-cost model is highly accurate and computationally expensive. An example of a high-cost model is the Multi-Fidelity model described in Non-Patent Literature 1.
[0024] The low-cost model calculation unit 22 estimates the values of each variable at each time step at low cost using a low-cost model. The low-cost model has low accuracy and low computational cost. An example of a low-cost model is the Low-Fidelity model described in Non-Patent Literature 1.
[0025] The high-cost model and low-cost model are examples given in Non-Patent Document 1, but are not limited to these. The estimation unit 20 may use any two models, designating the higher-cost model as the high-cost model and the lower-cost model as the low-cost model.
[0026] The control unit 30 controls the processing of the high-cost model calculation unit 21 and the low-cost model calculation unit 22.
[0027] The control unit 30 requests processing from the high-cost model calculation unit 21 for each time step from the start time to the evaluation time Tp. The control unit 30 calculates an estimated effect size of the reference variable at evaluation time Tp from the value of the reference variable P at evaluation time Tp. If the estimated effect size is within the acceptable range, the control unit 30 requests processing from the high-cost model calculation unit 21 for each time step from after evaluation time Tp to the end time Te. On the other hand, if the estimated effect size is outside the acceptable range, the control unit 30 requests processing from the low-cost model calculation unit 22 for each time step from after evaluation time Tp to the end time Te.
[0028] The evaluation time Tp and end time Te are obtained from the time condition data 12. The criterion variable P is identified from the focus variable Q of the judgment condition data 13 by the sensitivity calculation unit 31, which will be described later. The tolerance range is obtained from the judgment condition data 13. The estimated effect size of the criterion variable at evaluation time Tp is calculated by the effect size calculation unit 32, which will be described later.
[0029] The estimation unit 20 calculates estimated values for each time step from the start time to the end time Te, based on predetermined initial value data 11 and other conditions. The control unit 30 calculates estimated values using a high-cost model from the start time to the evaluation time Tp. If, at the evaluation time Tp, the accuracy cannot be expected with the initial value and other conditions, the control unit 30 calculates estimated values using a low-cost model to avoid incurring costs in subsequent estimation processing. This reduces the computational load in OSSE.
[0030] Even when the low-cost model calculation unit 22 calculates estimated values after evaluation time Tp, the control unit 30 may, depending on the error of the reference variable P, have the high-cost model calculation unit 21 calculate estimated values for each time step after evaluation time Tp. Specifically, for each time step from after evaluation time Tp to end time Te, the low-cost model calculation unit 22 estimates the value of each variable, and if the effect size of the reference variable P at end time Te is within an acceptable range, the control unit 30 requests the high-cost model calculation unit 21 to process for each time step from after evaluation time Tp to end time Te.
[0031] If, from after evaluation time Tp until completion time Te, an estimate is calculated using the low-cost model, and then an estimate is calculated using the high-cost model, the estimation device 1 discards the estimate calculated using the low-cost model and retains the estimate calculated using the high-cost model.
[0032] If the prediction accuracy of the reference variable P is estimated to be outside the acceptable range at evaluation time Tp, the estimation device 1 has the low-cost model calculation unit 22 calculate an estimated value at low cost. If the estimated value calculated by the low-cost model calculation unit 22 at low cost at completion time Te is within the acceptable range, this can provide an incentive to calculate an estimated value at high cost. Therefore, if the estimated value calculated by the low-cost model calculation unit 22 at low cost is within the acceptable range, the estimation device 1 calculates an estimated value with high accuracy. This allows for the efficient use of computational resources by enabling the calculation of an estimated value with high accuracy when there is a high probability that an estimated value within the acceptable range will be calculated.
[0033] The control unit 30 includes a sensitivity calculation unit 31, an effect amount calculation unit 32, and a determination unit 33.
[0034] The sensitivity calculation unit 31 identifies a reference variable P that influences the variable Q of interest based on a sensitivity analysis of the variable Q of interest. This disclosure describes a case where the variable Q of interest is specified as input data and the reference variable P is identified from the variable Q of interest, but is not limited to this case. The reference variable P may also be specified as input data.
[0035] The sensitivity calculation unit 31 performs a sensitivity analysis on, for example, precipitation near the center of a typhoon, going back a predetermined time. It calculates the sensitivity of each variable for the variable of interest Q at the predetermined time point. The sensitivity calculation unit 31 identifies a reference variable P from among multiple variables with high sensitivity. High sensitivity may include both positive sensitivity, where the output value increases as the value increases, and negative sensitivity, where the output value increases as the value decreases.
[0036] The effect size calculation unit 32 calculates the prediction accuracy of the reference variable P at evaluation time Tp as an estimated value of the effect size, based on the value of the reference variable P at evaluation time Tp. The prediction accuracy is calculated by the RMSE between the predicted value of the reference variable P at evaluation time Tp and the observed or analyzed value at evaluation time Tp.
[0037] Furthermore, the effect size calculation unit 32 calculates the accuracy of the reference variable P at the end time Te, using the estimated value of the reference variable P at the end time Te calculated by the low-cost model as the effect size of the reference variable P at the end time Te. The prediction accuracy or accuracy calculated by the effect size calculation unit 32 is used for determination by the determination unit 33.
[0038] The determination unit 33 determines whether to continue the calculation by the high-cost model calculation unit 21 at evaluation time Tp. If the prediction accuracy of the reference variable P calculated by the effect size calculation unit 32 is within an acceptable range, the determination unit 33 determines whether to continue the calculation by the high-cost model calculation unit 21 at each time step after evaluation time Tp. If the prediction accuracy of the reference variable P calculated by the effect size calculation unit 32 is outside an acceptable range, the determination unit 33 determines whether to stop the calculation by the high-cost model calculation unit 21 and perform the calculation by the low-cost model calculation unit 22 at each time step after evaluation time Tp.
[0039] Further, when the calculation by the low-cost model calculation unit 22 is performed at each time step after the evaluation time Tp and the accuracy of the reference variable P calculated by the effect amount calculation unit 32 is within the allowable range, the determination unit 33 determines the calculation by the high-cost model calculation unit 21 at each time step after the evaluation time Tp.
[0040] The output unit 40 outputs the estimated value data 16.
[0041] (Estimation method) Referring to FIG. 4, the estimation method by the estimation device 1 of the present disclosure will be described.
[0042] In step S1, the estimation device 1 specifies a reference variable P that affects the target variable Q.
[0043] In step S2, the estimation device 1 calculates the estimated values of each variable using the high-cost model from time step 0 to the evaluation time Tp. In step S3, the estimation device 1 calculates the estimated value of the effect amount of the reference variable P.
[0044] In step S4, the estimation device 1 determines whether the estimated value of the effect amount calculated in step S3 is within the allowable error.
[0045] When the estimated value of the effect amount calculated in step S3 is within the allowable error, in step S5, the estimation device 1 calculates the estimated values of each variable using the high-cost model from the evaluation time Tp to the end time Te.
[0046] When the estimated value of the effect amount calculated in step S3 is outside the allowable error, in step S6, the estimation device 1 calculates the estimated values of each variable using the low-cost model from the evaluation time Tp to the end time Te. When the estimated values of each variable are calculated up to the end time Te, in step S7, the estimation device 1 calculates the effect amount of the reference variable P. When the effect amount calculated in step S7 is within the allowable error, in step S5, the estimation device 1 calculates the estimated values of each variable using the high-cost model from the evaluation time Tp to the end time Te.
[0047] In step S5, if the estimated values of each variable are calculated using the high-cost model from the evaluation time Tp to the end time Te, or in step S8, if the effect size of the reference variable P calculated using the low-cost model is outside the allowable error, the process proceeds to step S9. In step S9, the estimation device 1 outputs the estimated values of each variable calculated in step S2, step S5, or step S6.
[0048] Such an estimation device 1 of the present disclosure calculates an estimated value using a high-cost model, and before reaching the end time, determines whether to continue the calculation by the high-cost model or switch to the calculation by the low-cost model based on whether the effect size is within the allowable range. Thereby, the computational amount in OSSE can be reduced.
[0049] The estimation device 1 according to the present disclosure described above is implemented using, for example, a general-purpose computer system including a CPU (Central Processing Unit), a memory 902, a storage 903 (HDD: Hard Disk Drive, SSD: Solid State Drive), a communication device 904, an input device 905, and an output device 906. In this computer system, each function of the estimation device 1 is realized by the CPU 901 executing a program loaded on the memory 902.
[0050] The estimation device 1 may be implemented by one computer or may be implemented by a plurality of computers. Further, the estimation device 1 may be a virtual machine implemented on a computer.
[0051] The program of the estimation device 1 can be stored in a computer-readable recording medium such as an HDD, an SSD, a USB (Universal Serial Bus) memory, a CD (Compact Disc), a DVD (Digital Versatile Disc), or can be distributed via a network. The computer-readable recording medium is, for example, a non-transitory recording medium.
[0052] This disclosure is not limited to the embodiments described above, and numerous modifications are possible within the scope of its essence.
[0053] 1 Estimation device 11 Initial value data 12 Time condition data 13 Judgment condition data 16 Estimated value data 20 Estimation unit 21 High cost model calculation unit 22 Low cost model calculation unit 30 Control unit 31 Sensitivity calculation unit 32 Effect size calculation unit 33 Judgment unit 40 Output unit 901 CPU 902 Memory 903 Storage 904 Communication device 905 Input device 906 Output device
Claims
1. An estimation device comprising: a high-cost model calculation unit that estimates the value of each variable at each time step at high cost using a high-cost model; a low-cost model calculation unit that estimates the value of each variable at each time step at low cost using a low-cost model; and a control unit that controls the processing of the high-cost model calculation unit and the low-cost model calculation unit, wherein the control unit requests the high-cost model calculation unit to process for each time step from the start time to the evaluation time; calculates an estimated value of the effect amount of the reference variable at the evaluation time from the value of the reference variable at the evaluation time; if the estimated value of the effect amount is within an acceptable range, requests the high-cost model calculation unit to process for each time step from after the evaluation time to the end time; and if the estimated value of the effect amount is outside an acceptable range, requests the low-cost model calculation unit to process for each time step from after the evaluation time to the end time.
2. The estimation apparatus according to claim 1, wherein, for each time step from after the evaluation time to the end time, the low-cost model calculation unit estimates the value of each variable, and if the effect size of the reference variable at the end time is within an acceptable range, the control unit requests the high-cost model calculation unit to perform processing for each time step from after the evaluation time to the end time.
3. The computer comprises a high-cost model calculation unit that estimates the value of each variable at each time step at high cost using a high-cost model, a low-cost model calculation unit that estimates the value of each variable at each time step at low cost using a low-cost model, and a control unit that controls the processing of the high-cost model calculation unit and the low-cost model calculation unit, wherein the control unit requests the high-cost model calculation unit to process for each time step from the start time to the evaluation time, calculates an estimated value of the effect amount of the reference variable at the evaluation time from the value of the reference variable at the evaluation time, requests the high-cost model calculation unit to process for each time step from after the evaluation time to the end time if the estimated value of the effect amount is within an acceptable range, and requests the low-cost model calculation unit to process for each time step from after the evaluation time to the end time if the estimated value of the effect amount is outside an acceptable range.
4. A computer is configured to function as a high-cost model calculation unit that estimates the value of each variable at each time step at high cost using a high-cost model; a low-cost model calculation unit that estimates the value of each variable at each time step at low cost using a low-cost model; and a control unit that controls the processing of the high-cost model calculation unit and the low-cost model calculation unit, wherein the control unit requests the high-cost model calculation unit to process each time step from the start time to the evaluation time; calculates an estimated value of the effect amount of the reference variable at the evaluation time from the value of the reference variable at the evaluation time; if the estimated value of the effect amount is within an acceptable range, requests the high-cost model calculation unit to process each time step from after the evaluation time to the end time; and if the estimated value of the effect amount is outside an acceptable range, requests the low-cost model calculation unit to process each time step from after the evaluation time to the end time.