Learning data creation system, learning data creation method, and learning data creation program

Abstract

To provide a learning data creation system, a learning data creation method, and a learning data creation program for accurately and efficiently creating learning data necessary for machine learning.SOLUTION: A calculation device 20 includes an experimental information storage unit 23 configured to record experimental conditions including factors and levels to be used in experiments, a learning information storage unit 24 configured to record learning data to be used in machine learning, and a control unit 21 configured to generate learning data. The control unit 21 creates, by using an experimental design method, an orthogonal table including a plurality of experimental conditions consisting of factors and levels, generates new experimental conditions from successive levels of preceding experimental conditions and succeeding experimental conditions in the orthogonal table, records the new experimental conditions in the experimental information storage unit 23, generates the learning data through experiments using respective experimental conditions recorded in the experimental information storage unit 23, and records the learning data in the learning information storage unit 24.SELECTED DRAWING: Figure 1

Description

[Technical field]

[0001] The present invention relates to a learning data creation system, a learning data creation method, and a learning data creation program for generating learning data to be used in analysis. [Background technology]

[0002] Various analytical methods using computers have been considered (for example, see Patent Document 1). In the technology described in Patent Document 1, a heat load calculation device that calculates the amount of heat passing through a partition material that separates a space has been considered. In this heat load calculation device, the area of ​​an element model is calculated using design data, and the area is used to calculate the amount of heat passing through each partition material element model.

[0003] Furthermore, when machine learning is used in the analysis method, technology for efficiently generating training data to be used in machine learning is also being considered (for example, see Patent Document 2). The training data generation system described in Patent Document 2 performs clustering of multiple classification objects and presents the contents of some of the classification objects that belong to a cluster to an analyst. Then, based on the labels designated by the analyst for the clusters, it generates training data to be trained by a learning model.

[0004] In addition, in the analysis process for calculating explanatory variables with high contributions, learning data using an orthogonal array in experimental design may be used (for example, see Patent Document 3). This orthogonal array is an array table in which all combinations of levels appear the same number of times for any two columns. This property is used to assign factors and levels in experimental design. [Prior art documents] [Patent documents]

[0005] [Patent Document 1] Patent Publication No. 2021-43074 [Patent Document 2] Patent Publication No. 2021-56591 [Patent Document 3] JP 2016-031629 A Summary of the Invention [Problem to be solved by the invention]

[0006] In analysis using experimental design, first, the factors and levels for the experiment are determined. Then, the experiment is assigned using an orthogonal array. Next, the experiment is conducted using the orthogonal array, and learning data is created from the experimental results. Finally, a predictive model is created using machine learning with the learning data.

[0007] However, since the original purpose of experimental design is to reduce the number of experiments and perform efficient experiments, there are cases where the training data obtained by experimental design alone is not enough to obtain the training data required for machine learning. [Means for solving the problem]

[0008] A learning data creation system for solving the above problem includes an experiment information storage unit that records experimental conditions consisting of factors and levels used in an experiment, a learning information storage unit that records learning data used in machine learning, and a control unit that generates learning data. The control unit creates an orthogonal array consisting of a plurality of experimental conditions consisting of the factors and the levels using an experimental design method, generates new experimental conditions using the levels of successive preceding and subsequent experimental conditions in the orthogonal array and records them in the experiment information storage unit, and generates learning data by an experiment using each of the experimental conditions recorded in the experiment information storage unit and records them in the learning information storage unit. Effect of the Invention

[0009] According to the present invention, learning data required for machine learning can be created accurately and efficiently. [Brief description of the drawings]

[0010] [Figure 1] FIG. 1 is an explanatory diagram of a computing device as a learning data creation system according to an embodiment. [Diagram 2] FIG. 2 is an explanatory diagram of a hardware configuration according to the embodiment. [Diagram 3] FIG. 4 is an explanatory diagram of a processing procedure according to the embodiment. [Figure 4] FIG. 4 is an explanatory diagram of a processing procedure according to the embodiment. [Diagram 5] FIG. 4 is an explanatory diagram of an orthogonal table according to the embodiment. [Figure 6] FIG. 4 is an explanatory diagram of experimental conditions in the embodiment. [Figure 7] 5A, B, and C are explanatory diagrams for adding experimental conditions in an embodiment, where FIG. 5A illustrates the addition of a first experimental condition, FIG. 5B illustrates the addition of a second experimental condition, and FIG. 5C illustrates the addition of a third experimental condition. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0011] Hereinafter, an embodiment of the learning data creation system, the learning data creation method, and the learning data creation program will be described with reference to Figs. 1 to 7. In this embodiment, an orthogonal array used for allocating experiments by an experimental design method is decomposed to add experiments, thereby increasing the number of experiments, i.e., the number of learning data. In this orthogonal array, experimental conditions are set in which multiple levels are allocated to multiple factors. Then, the results of an experiment (simulation results) performed using these experimental conditions are used as learning data to generate a prediction model by analysis processing such as machine learning. In this embodiment, a case is assumed in which a simulation of the thermal load of a building generated using BIM (Building Information Modeling) is performed.

[0012] As shown in FIG. 1, a calculation device 20 is used as a learning data creation system. (Hardware configuration example) FIG. 2 shows an example of the hardware configuration of an information processing device H10 that functions as the calculation device 20.

[0013] The information processing device H10 includes a communication device H11, an input device H12, a display device H13, a storage device H14, and a processor H 15. Note that this hardware configuration is an example, and the information processing device H10 may include other hardware.

[0014] The communication device H11 is an interface that establishes a communication path with another device and transmits and receives data, and is, for example, a network interface or a wireless interface.

[0015] The input device H12 is a device that accepts input from a user, such as a mouse, a keyboard, etc. The display device H13 is a display, a touch panel, etc. that displays various information.

[0016] The storage device H14 is a storage device that stores data and various programs for executing various functions of the computing device 20. Examples of the storage device H14 include a ROM, a RAM, and a hard disk.

[0017] The processor H15 uses programs and data stored in the storage device H14 to control each process in the computing device 20 (for example, processes in the control unit 21 described below). Examples of the processor H15 include a CPU and an MPU. The processor H15 loads programs stored in a ROM or the like into a RAM and executes various processes corresponding to the various processes. For example, when an application program of the computing device 20 is started, the processor H15 operates a process that executes each process.

[0018] The processor H15 is not limited to a processor that performs software processing for all the processes it executes. For example, the processor H15 may include a dedicated hardware circuit (e.g., an application specific integrated circuit (ASIC)) that performs hardware processing for at least a part of the processes it executes. The processor H15 may be configured with the following:

[0019] (1) One or more processors that operate according to a computer program (software) (2) One or more dedicated hardware circuits that perform at least some of the processes; or (3) A circuit that includes a combination of these. The processor includes a CPU and memory, such as RAM and ROM, that stores program codes or instructions configured to cause the CPU to execute processes. Memory, i.e., computer-readable media, includes any available media that can be accessed by a general-purpose or special-purpose computer.

[0020] (Functions of the learning data creation system) The computing device 20 includes a control unit 21, a design information storage unit 22, an experiment information storage unit 23, and a learning information storage unit 24.

[0021] The control unit 21 executes the learning data creation program to function as an orthogonal table creation unit 211 and an adjustment unit 212. Furthermore, the control unit 21 functions as a simulator 213 and a learning unit 214.

[0022] The orthogonal table creation unit 211 executes a process of creating an orthogonal table consisting of a plurality of experimental conditions using the levels of each factor, based on an experimental design method. The adjustment unit 212 executes a process of increasing the number of experimental conditions included in the orthogonal array. By adding these experimental conditions, it is possible to increase the amount of learning data, which will be described later.

[0023] The simulator 213 is an experimental device that executes a simulation to calculate the heat load in a structure using the levels included in the experimental conditions. When any one factor is manipulated, the simulator 213 executes a new simulation while maintaining the levels of the other factors.

[0024] The learning unit 214 executes an analysis process to generate a prediction model that predicts factors that affect the heat load using the simulation results. As the analysis process, regression analysis, deep learning, or the like can be used as long as it is machine learning that generates a model that predicts comprehensive information from discrete information. This learning unit 214 holds information about the accuracy reference value for determining whether or not there is a shortage of learning data.

[0025] Design data of a structure designed by BIM is recorded in the design information storage unit 22. This design data is recorded when a structure to be calculated is registered. The experimental data used in the simulation is recorded in the experimental information storage unit 23. This experimental data is recorded when a level is assigned to each factor. In the experimental data, the level of each factor is recorded for each experiment number. A simulation is performed to calculate the heat load using experimental conditions consisting of the levels of each factor.

[0026] The experiment number is an identifier for identifying each experimental condition. A factor is a variable that affects the thermal load of a structure (e.g., window area ratio). The level is the variable value of this factor. For example, the window area ratio can be set to "Level 1: 80%" and "Level 2: 60%."

[0027] The learning information storage unit 24 records learning data generated by simulation. This learning data is recorded when a simulation (experiment) is performed. The learning data is used in an analysis process (machine learning) that generates a prediction model for the heat load. The learning data records the calculation results of a simulation performed at the level of each factor for the experiment number.

[0028] (Analysis processing) Next, the analysis process will be described with reference to FIG. First, the control unit 21 of the calculation device 20 executes a process of creating experimental conditions (step S101). Specifically, the orthogonal table creation unit 211 and the adjustment unit 212 of the control unit 21 generate a plurality of experimental conditions to be used in the simulation. Details will be described with reference to FIG. 4.

[0029] Next, the control unit 21 of the computing device 20 sequentially identifies the experiment numbers to be processed, and repeats the following process for all experimental conditions. First, the control unit 21 of the computing device 20 executes a simulation process (step S102). Specifically, the simulator 213 of the control unit 21 calls up the experimental conditions (levels of each factor) to be processed from the experimental information storage unit 23, and applies them to the design data of the structure recorded in the design information storage unit 22 to perform a heat load calculation simulation.

[0030] Next, the control unit 21 of the calculation device 20 executes a learning data recording process (step S103). Specifically, the simulator 213 of the control unit 21 generates learning data in which the calculation results of the simulation are recorded in association with the experiment number, and records the learning data in the learning information storage unit 24. The control unit 21 of the computing device 20 repeats the above process until it is completed for all experiment numbers.

[0031] Next, the control unit 21 of the calculation device 20 executes a learning process (step S104). Specifically, the learning unit 214 of the control unit 21 generates a prediction model of factors that affect the heat load using the learning data recorded in the learning information storage unit 24. From this prediction model, the contribution of each factor that affects the heat load can be calculated. Furthermore, the learning unit 214 acquires an evaluation index of the accuracy of the prediction model. This evaluation index is information that evaluates the difference (accuracy) between the output value of the prediction model and the experimental result. For example, in the regression analysis, the root mean square error, the mean absolute error, or the coefficient of determination can be used.

[0032] Next, the control unit 21 of the calculation device 20 executes a process of determining whether the learning data is insufficient (step S105). Specifically, the learning unit 214 of the control unit 21 compares the acquired evaluation index with the accuracy reference value and determines whether or not the learning data needs to be added. If the evaluation index is less than the accuracy reference value, it is determined that the learning data is insufficient.

[0033] When it is determined that the learning data is insufficient ("YES" in step S105), the control unit 21 of the calculation device 20 executes a process of adding the number of factors or the number of levels (step S106). Specifically, the learning unit 214 of the control unit 21 instructs the orthogonal array creation unit 211 to add the number of factors or the number of levels. In this case, the orthogonal array creation unit 211 adds, for example, a dummy factor or a new level, and executes the process of creating the experimental conditions (step S101) and subsequent steps.

[0034] (Creating experimental conditions) Next, the process of creating experimental conditions will be described with reference to Fig. 4. Here, the orthogonal array used for the allocation of experiments by the experimental design method is decomposed and added in the following procedure, and the number of experiments, i.e., the amount of learning data, is increased.

[0035] First, the control unit 21 of the calculation device 20 executes an orthogonal array creation process (step S201). Specifically, the orthogonal array creation unit 211 of the control unit 21 uses an experiment plan to generate an orthogonal array in which the levels of each factor are assigned in association with the experiment number.

[0036] 5 shows an orthogonal array 500 assuming two levels for each of seven factors. In this case, experimental conditions for experiment numbers 1 to 8 are set. Next, the control unit 21 of the calculation device 20 sequentially identifies the experiment numbers to be processed and repeats the following process. Here, when n experiment numbers are set, numbers "1" to "n-1" are processed.

[0037] First, the control unit 21 of the calculation device 20 executes a process of identifying factors whose levels are different between the previous experiment and the subsequent experiment (step S202). Specifically, the adjustment unit 212 of the control unit 21 compares the levels of the factors in the previous experiment conditions with the levels of the factors in the subsequent experiment conditions to identify factors whose levels are different (different factors). If there is only one different factor, the process for this experiment number is terminated.

[0038] In the following, a previous experiment (experiment number 4) and a subsequent experiment (experiment number 5) will be described as an example, as shown in Fig. 6. Here, experimental conditions 540 of the previous experiment are compared with experimental conditions 550 of the subsequent experiment. In this case, difference factors 1, 2, 4, and 7 with different levels are identified.

[0039] Next, the control unit 21 of the calculation device 20 executes a process of changing the level of one factor (step S203). Specifically, the adjustment unit 212 of the control unit 21 changes the level of one difference factor. In this embodiment, the levels of the difference factors are changed in the order of the factors (ascending order).

[0040] In FIG. 7(a), experimental condition 541 (first experimental condition) is created by changing the difference factor 1 of experimental condition 540 from "1" to "2", the same as experimental condition 550. Next, the control unit 21 of the computing device 20 executes a process of adding the experiment conditions (step S204). Specifically, the adjustment unit 212 of the control unit 21 records the newly created experiment conditions 541 in the experiment information storage unit .

[0041] Next, the control unit 21 of the computing device 20 executes a process of determining whether or not the level matches the level of the subsequent experiment (step S205). Specifically, the adjustment unit 212 of the control unit 21 determines whether or not the level of the remaining difference factor matches the condition of the subsequent experiment when the level is changed.

[0042] In the experimental condition 541 in FIG. 7(a), when the level of the difference factor 2 is changed, the difference factors 4 and 7 still remain, so that the experimental condition does not match the experimental condition 550 of the subsequent experiment. When it is determined that the result does not match the standard of the subsequent experiment ("NO" in step S205), the control unit 21 of the computing device 20 repeats step S203 and subsequent steps.

[0043] 7(b), in step S203, experimental condition 542 (second experimental condition) is created by changing the difference factor 2 of experimental condition 541 from "2" to "1", the same as experimental condition 550. Then, in step S204, the experimental condition is recorded in the experiment information storage unit 23.

[0044] Next, in step S205, when the level of difference factor 4 in the experimental conditions 542 is changed, difference factor 7 remains, so that it does not match the experimental conditions 550 of the subsequent experiment. Therefore, steps S203 and after are repeated.

[0045] 7(c), in step S203, experimental condition 543 (third experimental condition) is created by changing the difference factor 4 of experimental condition 542 from "2" to "1", the same as experimental condition 550. Then, in step S204, the experimental condition is recorded in the experiment information storage unit 23. In this way, the newly added experimental conditions 541 to 543 are also used in the simulation process (step S102).

[0046] Next, in step S205, when the level of difference factor 7 in the experimental condition 543 is changed, it coincides with the experimental condition 550 of the subsequent experiment. If it is determined that the level matches that of the subsequent experiment (YES in step S205), the process for this experiment number ends.

[0047] Then, the control unit 21 of the calculation device 20 repeats the process for the experiment numbers "1" to "n-1" until the process is completed.

[0048] According to this embodiment, the following effects can be obtained. (1) In this embodiment, the control unit 21 of the computing device 20 executes a process of creating experimental conditions (step S101), a simulation process (step S102), and a process of recording learning data (step S103). This makes it possible to generate a general-purpose prediction model using experimental results under a predetermined number of experimental conditions. In particular, since an orthogonal array is used for the experimental conditions, it is possible to perform analysis processing using well-balanced learning data.

[0049] (2) In this embodiment, the control unit 21 of the calculation device 20 executes a process of creating an orthogonal array (step S201), which makes it possible to set experimental conditions that maintain a balance between the factors and levels of the orthogonal array.

[0050] (3) In this embodiment, the control unit 21 of the calculation device 20 executes a process of identifying factors whose levels are different between the preceding experiment and the subsequent experiment (step S202), a process of changing the level of one factor (step S203), and a process of adding experimental conditions (step S204). This makes it possible to increase the number of experiments while maintaining the balance between the factors and levels of the orthogonal array. Then, learning data to be used in the analysis process can be generated.

[0051] For example, experimental conditions 541 to 543 in Fig. 7(c) are intermediate conditions that maintain a common level between the successive experimental conditions 540 and 550 in the orthogonal array. Therefore, it is possible to increase the number of experiments while maintaining the properties of the experimental conditions 540 and 550 created by the experimental design method.

[0052] (4) In this embodiment, the control unit 21 of the computing device 20 executes a process of changing the level of one factor (step S203). When performing a simulation by changing factors one by one, the levels of the other factors can be maintained, and an experiment can be performed efficiently.

[0053] (5) In this embodiment, when it is determined that the learning data is insufficient ("YES" in step S105), the control unit 21 of the calculation device 20 executes a process of adding the number of factors or the number of levels (step S106). This makes it possible to perform machine learning using the number of experiments and learning data required to improve the accuracy of the prediction model.

[0054] This embodiment can be modified as follows: This embodiment and the following modifications can be combined with each other to the extent that there is no technical contradiction. In the above embodiment, a simulation of the heat load of a building was performed as an experiment, but the experiment is not limited to this. Also, the experiment is not limited to a simulation, and it is sufficient if the experimental results can be obtained.

[0055] In the above embodiment, the calculation device 20 is used as the learning data creation system. The hardware configuration is not limited to this. For example, the simulator 213 as the experimental device and the learning unit 214 that performs machine learning may be other devices (servers). Furthermore, the calculation device 20 may be realized by a server accessible by a user terminal.

[0056] In the above embodiment, the control unit 21 of the calculation device 20 executes a process of changing the level of one factor (step S203). The number of factors whose levels are changed is not limited to one, and the levels of multiple factors may be changed simultaneously.

[0057] In the above embodiment, in the process of creating the experimental conditions, the control unit 21 of the computing device 20 sequentially identifies the experiment numbers to be processed and repeats the process. Here, a target number of experiments may be determined, and the repetition may be terminated when the number of experiments reaches the target number.

[0058] In the above embodiment, the levels of the difference factors are changed in the order of the factors, but the order of the changes is not limited to the order of the factors. For example, in Fig. 6, the ascending and descending order may be reversed, such as 1 → 7 → 2 → 4. Also, when level "1" is changed to "2", priority may be given to changing from level "2" to "1". This allows the number of different levels to be well balanced, and the number of experimental conditions and learning data to be increased. [Explanation of symbols]

[0059] 20...calculation device, 21...control unit, 210...filling management unit, 211...orthogonal table creation unit, 212...adjustment unit, 213...simulator, 214...learning unit, 22...design information storage unit, 23...experiment information storage unit, 24...learning information storage unit.

Claims

1. An experiment information storage unit for recording experimental conditions consisting of factors and levels used in a simulation experiment in which the levels of each factor are set; A learning data creation system including a learning information storage unit that records learning data used in machine learning, and a control unit that creates the learning data, The control unit: Using an experimental design method, an orthogonal table consisting of a plurality of experimental conditions each consisting of the factor and the level is created; In the orthogonal array, the levels of successive preceding and subsequent experimental conditions are compared to identify difference factors having different levels, and the levels of the preceding experimental conditions are sequentially changed for the difference factors to generate new experimental conditions that are closer to the subsequent experimental conditions, and the new experimental conditions are recorded in the experimental information storage unit; A learning data creation system, characterized in that it generates learning data including experimental results of a simulation experiment using each experimental condition recorded in the experimental information storage unit, and records the learning data in the learning information storage unit.

2. The control unit: obtaining an evaluation index for evaluating the accuracy of a prediction model generated by machine learning using the learning data recorded in the learning information storage unit; The learning data creation system according to claim 1 , wherein the necessity for generating the new experimental conditions is determined based on the evaluation index.

3. An experiment information storage unit for recording experimental conditions consisting of factors and levels used in a simulation experiment in which the levels of each factor are set; A method for creating learning data using a learning data creation system including a learning information storage unit that records learning data used in machine learning, and a control unit, comprising: The control unit: Using an experimental design method, an orthogonal table consisting of a plurality of experimental conditions each consisting of the factor and the level is created; In the orthogonal array, the levels of successive preceding and subsequent experimental conditions are compared to identify difference factors having different levels, and the levels of the preceding experimental conditions are sequentially changed for the difference factors to generate new experimental conditions that are closer to the subsequent experimental conditions, and the new experimental conditions are recorded in the experimental information storage unit; A learning data creation method, comprising generating learning data including experimental results of a simulation experiment using each of the experimental conditions recorded in the experiment information storage unit, and recording the generated learning data in the learning information storage unit.

4. An experiment information storage unit for recording experimental conditions consisting of factors and levels used in a simulation experiment in which the levels of each factor are set; A program for creating learning data using a learning data creation system including a learning information storage unit that records learning data used in machine learning and a control unit, The control unit, Using an experimental design method, an orthogonal table consisting of a plurality of experimental conditions each consisting of the factor and the level is created; In the orthogonal array, the levels of successive preceding and subsequent experimental conditions are compared to identify difference factors having different levels, and the levels of the preceding experimental conditions are sequentially changed for the difference factors to generate new experimental conditions that are closer to the subsequent experimental conditions, and the new experimental conditions are recorded in the experimental information storage unit; A learning data creation program characterized by functioning as a means for generating learning data including experimental results of a simulation experiment using each experimental condition recorded in the experimental information storage unit and recording the learning data in the learning information storage unit.

Images