Experiment design support device and experiment design support method

Abstract

To provide an explanatory variable used in experimental design based on search results of a causal relation with high validity.SOLUTION: An experiment design support device is composed of a computer having an arithmetic device configured to execute predetermined processing and a storage device connected to the arithmetic device and includes: an input device configured to receive an input of an objective variable and an explanatory variable searching for a causal relation with the objective variable; a database in which an influence degree of the explanatory variable in a process of the experiment is stored; a causal search unit configured to search for the causal relation between the explanatory variables by the arithmetic device; a structure evaluation unit configured to construct a causal graph having a high evaluation value of a causal graph representing an influence degree between the processes by the arithmetic device; and an explanatory variable extraction unit configured to extract the explanatory variable having a high causal relation with the objective variable using the constructed causal graph by the arithmetic device.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present invention relates to an experimental design support device that uses causal discovery. [Background technology]

[0002] When designing experimental conditions for materials development, it is necessary to design the types and levels of explanatory variables. Designing the types and levels of explanatory variables can be done theoretically based on the mechanism of physical property expression, or by using correlation coefficients with target variables from experimental data. In recent years, experimental design techniques using causal analysis, which is more reliable than correlation coefficients, have been proposed. A known technique for experimental design using causal analysis is described in JP 2017-146899 (Patent Document 1). This publication describes a "design support system comprising an input device capable of inputting knowledge regarding causal relationships, a computing device that integrates the knowledge regarding causal relationships input into the input device with causal inference results obtained by causal inference for each group of multiple experimental data, and an output device that displays the integration results from the computing device." [Prior art documents] [Patent documents]

[0003] [Patent Document 1] Japanese Patent Application Laid-Open No. 2017-146899 Summary of the Invention [Problem to be solved by the invention]

[0004] When conducting an experiment involving multiple processes, the results obtained in an early process may change depending on the subsequent process. For example, if a material is processed to a specified size and then heat-treated, the dimensions of the material will change. However, Patent Document 1 does not take into consideration how to deal with causal relationships that do not follow the order of the processes.

[0005] Therefore, an object of the present invention is to search for highly valid causal relationships by taking into account the causal effects of all processes, and to provide explanatory variables to be used in experimental design based on the search results. [Means for solving the problem]

[0006] A representative example of the invention disclosed in the present application is as follows: That is, an experimental design support device is configured by a computer having an arithmetic unit that executes predetermined processing and a storage device connected to the arithmetic unit, and is characterized in that it comprises: an input unit that receives input of explanatory variables for searching for a causal relationship between objective variables and explanatory variables, a database that stores the influence of the explanatory variables in the process of the experiment, a causal search unit that searches for a causal relationship between the explanatory variables, a structure evaluation unit that constructs a causal graph with a high evaluation value representing the influence between the processes, and an explanatory variable extraction unit that extracts explanatory variables that have a high causal relationship with the objective variable using the constructed causal graph. [Effects of the Invention]

[0007] According to one aspect of the present invention, the types of explanatory variables required for experimental design can be determined based on highly reliable information. Objects, configurations, and effects other than those described above will become apparent from the following description of the examples. [Brief explanation of the drawings]

[0008] [Figure 1] FIG. 1 is a diagram illustrating a configuration of an experiment design support device according to a first embodiment. [Figure 2] 3 is a flowchart showing a process executed by the experiment design support device of the first embodiment. [Figure 3] FIG. 2 is a diagram showing the data structure of the experimental data in Example 1. [Figure 4] FIG. 1 is a diagram showing a causal relationship graph according to the first embodiment. [Figure 5] FIG. 2 is a diagram illustrating a data structure of process causal information according to the first embodiment. [Figure 6]10 is a flowchart of an explanatory variable extraction process according to the first embodiment. DETAILED DESCRIPTION OF THE INVENTION

[0009] Hereinafter, an embodiment will be described with reference to the drawings.

[0010] Example 1 (Configuration of experimental design support device) FIG. 1 is a diagram showing the configuration of an experiment design support device 100 according to the first embodiment.

[0011] The experimental design support device 100 of this embodiment is composed of a general-purpose computer and has a central control unit 110, an input device 120 such as a keyboard, an output device 130 such as a display, a main memory device 140, an auxiliary memory device 150, and a communication device 160.

[0012] The central control unit 110 is a processor that executes programs stored in the memory 102. The central control unit 110 executes various programs to realize the various functional units (e.g., a causal search unit 141, a structural evaluation unit 142, an explanatory variable extraction unit 143, etc.) of the experimental design support system 100. Note that some of the processing performed by the processor executing the programs may be executed by other types of computing devices (e.g., hardware such as ASIC and FPGA).

[0013] The main memory device 140 includes a ROM, which is a non-volatile memory element, and a RAM, which is a volatile memory element. The ROM stores unchanging programs (e.g., BIOS), etc. The RAM is a high-speed, volatile memory element such as a DRAM (Dynamic Random Access Memory), and temporarily stores programs executed by the central control device 110 and data used when the programs are executed.

[0014] The auxiliary storage device 150 is a large-capacity, non-volatile storage device such as a magnetic storage device (HDD) or a flash memory (SSD). The auxiliary storage device 150 also stores data used by the central control device 110 when executing a program (e.g., experimental data 151, process causal information 152, structure evaluation value information 153), and programs executed by the central control device 110 (e.g., a causal search program, a structure evaluation program, an explanatory variable extraction program, etc.). That is, the programs are read from the auxiliary storage device 150, loaded into the main storage device 140, and executed by the central control device 110 to realize each function of the experimental design support device 100.

[0015] The input device 120 is an interface such as a keyboard or mouse that receives input from an operator. The output device 130 is an interface such as a display device or a wearable device (e.g., VR goggles) that outputs the results of program execution as images in a format that can be viewed by the user. Note that a terminal (not shown) connected to the experimental design support device 100 via a network may provide the input device and output device. In this case, the experimental design support device 100 may have a web server function, and the terminal may access the experimental design support device 100 using a predetermined protocol (e.g., http).

[0016] The communication device 160 is a network interface device that controls communication with other devices according to a predetermined protocol.

[0017] The programs executed by the central control device 110 are provided to the experimental design support device 100 via removable media (CD-ROM, flash memory, etc.) or a network, and are stored in a non-volatile auxiliary storage device 150, which is a non-transitory storage medium. For this reason, the experimental design support device 100 should preferably have an interface for reading data from removable media.

[0018] The experimental design support system 100 is a computer system configured on a single physical computer or on multiple logically or physically configured computers, and may operate on a virtual computer constructed on multiple physical computer resources. For example, each functional unit may operate on a separate physical or logical computer, or multiple functional units may be combined to operate on a single physical or logical computer.

[0019] (Overall steps) FIG. 2 is a flowchart showing the processing executed by the experiment design support device 100 of the first embodiment.

[0020] First, the causal search unit 141 reads the experimental data 151 (S210). Next, the causal search unit 141 constructs a causal graph to search for causal relationships using the read experimental data 151 (S220). Next, the structure evaluation unit 142 scores the causal relationships searched for in step S220 using the causal relationships between process phases recorded in the process causal information 152, and calculates a structure evaluation value of the causal graph (S230).

[0021] Next, the structure evaluation unit 142 determines whether the structure evaluation value of the causal relationship is equal to or less than a predetermined threshold (S240). If the structure evaluation value of the causal relationship is equal to or less than the predetermined threshold, there is a possibility that an incorrect causal relationship has appeared among the causal relationships. Therefore, the causal search unit 141 tunes the hyperparameters of the algorithm used, returns to step S220, and searches for the causal relationship again. On the other hand, if the structure evaluation value of the causal relationship is equal to or greater than the predetermined threshold, it determines that the search result for the causal relationship is valid, and extracts explanatory variables that have a causal relationship with the target variable from the results of the causal search with a high structure evaluation value (S250).

[0022] Each step will be described in detail below. First, the experimental data read in step S210 will be described.

[0023] (Experimental data) FIG. 3 is a diagram showing the data structure of the experimental data 151 in the first embodiment.

[0024] The experimental data 151 includes items of material identification information 310, material information 320, and process conditions 330. The material identification information 310 may include detailed analytical information such as X-ray diffraction peaks and infrared absorption spectra in addition to material composition information and physical property values ​​such as density. The material composition information may include physical property values ​​before and after the process.

[0025] Next, the cause and effect search unit 141 that executes the process of step S220 will be described.

[0026] (Cause and Effect Exploration Department) The causal relationship search unit 141 searches for causal relationships using experimental data 151 and a predetermined, publicly known causal relationship search algorithm. FIG. 4 is a diagram showing a causal relationship graph that displays the searched causal relationships in a graphical format. The causal relationship graph shown in FIG. 4 is composed of variable names 410 and causal relationships 420. Here, the causal relationships 420 are represented by the direction of an arrow, with the start point of the arrow representing the cause and the end point representing the result. In this embodiment, a Bayesian network based on the K2 score is used as the causal relationship search algorithm. However, a linear non-Gaussian acyclic model (LiNGAM), which is a linear structural model, or a constraint-based Bayesian network may also be used. From the search results for causal relationships, information between processes with causal relationships and information between variables without causal relationships are stored in process causal information 152.

[0027] Next, the process causal information 152 and the structure evaluation unit 142 used in step S230 will be described.

[0028] (Process causal information) FIG. 5 is a diagram illustrating the data structure of the process causal information 152 according to the first embodiment.

[0029] Based on the results of the causal search, if there is a causal relationship between the variable on the vertical axis, which is the cause 510, and the variable on the horizontal axis, which is the result 520, a value of 1 is recorded, and if there is no causal relationship, a value of 0 is recorded. The variables include materials 530 and processes 540, and the process 540 describes the name of the process phase, and the causal score between the variables is defined by the average value of the range 550. By expressing the causal relationship as the average value of the causal score within a predetermined range, singular values ​​of the causal score due to noise can be removed, and a structural evaluation value of the causal graph that accurately represents the causal graph can be calculated.

[0030] The process causal information 152 pre-stores causal scores between processes for various data sets by searching for causal relationships multiple times with different data combinations. The structure evaluation unit 142, which will be described later, uses the causal scores averaged across the data sets.

[0031] (Structural Evaluation Department) The structure evaluation unit 142 uses the structure evaluation value to confirm the validity of the causal relationship found by the causal search unit 141. The structure evaluation value is defined by Equation 1, which includes a penalty term.

[0032]

number

[0033] Here, the paths included in Equation 1 are paths connected via intermediate factors. For example, for variables A, B, and C, if there is a causal relationship from A to B and a causal relationship from B to C, a path from A to C is generated. The total number of edges on the path from A to C is 2, and the path length is 2. If a causal relationship from C to D is added to this, the number of edges on the path from A to D is 3, and the path length is 3. The causal graph shown in Figure 4 has a total of three paths: a path from process 2 (temperature) to process 2 (pressure) with a path length of 1; a path from raw material (composition) to the target variable via raw material (density) with a path length of 2; and a path from process 1 (humidity) to the target variable via process 1 (time) and raw material (density) with a path length of 3.

[0034] The calculation method for Equation 1 will be specifically explained using the causal graph in Figure 4. As mentioned above, the causal graph in Figure 4 has three paths: one connecting humidity to the target variable, one connecting composition to the target variable, and one connecting temperature to pressure. The score for the path connecting humidity to the target variable is calculated. First, for the edge connecting humidity to time, the causal score from Process 2 (humidity) to Process 2 (time) (Figure 5) is referenced. Second, the causal score from Process 2 (time) to Material (density) (Figure 5) is referenced. Third, for the edge connecting Material (density) to the target variable, the causal score from Material (density) to the material (density) (Figure 5) is referenced. Here, if the causal scores are 0.2, 0.3, and 0.4, the average of these scores (0.3) is the score for the path from humidity to the target variable. The score for each path is calculated using this procedure, and the average score is the structural evaluation value. However, if there is an edge in the path of the causal graph whose causal score is equal to or less than a threshold, the value of the penalty term is set to 0 and multiplied by the causal score. Here, the threshold may be determined by k-means or Otsu's binarization method, which divides the histogram of causal scores into two groups. The structural evaluation unit 142 stores the calculated structural evaluation value in the structural evaluation value information 153. Such a penalty term can eliminate singular values ​​in the causal score due to noise, making it possible to calculate a structural evaluation value of a causal graph that accurately represents the causal graph. Furthermore, it is possible to eliminate singular values ​​in the causal score due to noise, making it possible to calculate a structural evaluation value of a causal graph that accurately represents the causal graph, using the path length of the causal relationship and the correlation coefficient.

[0035] Next, the structure evaluation value information 153 stored in step S230 will be described.

[0036] (Structural evaluation value information) The structure evaluation value information 153 stores the structure evaluation values ​​calculated by the structure evaluation unit 142. It also stores statistical values ​​such as the average, mode, and maximum values ​​calculated from the structure evaluation values ​​using a known algorithm. These statistical values ​​are referenced when setting thresholds for later reconstructing the causal graph.

[0037] Next, a repeat method when the structure evaluation value is equal to or less than the threshold value in step S240 will be described.

[0038] (Causal graph reconstruction) In step S250, a causal graph is repeatedly constructed until the structural evaluation value exceeds a preset threshold. The values ​​changed in the reconstruction of the causal graph are hyperparameters for the design of the initial structure in the search and the function design of the probabilistic model. When the structural evaluation value exceeds the threshold, the process proceeds to step S250 for extracting the next explanatory variable.

[0039] Next, a method for extracting explanatory variables in step S250 will be described.

[0040] (Explanatory variable extraction part) When an orthogonal array is used in an experimental design, the number and types of explanatory variables may be determined depending on the number of experiments that can be performed. The explanatory variable extraction unit 143 receives input of the objective variable and the number of explanatory variables to be used in the experimental design, and extracts explanatory variables up to the input number.

[0041] 6 is a flowchart of the explanatory variable extraction process in the embodiment 1. The explanatory variable extraction process will be described with reference to FIG.

[0042] The explanatory variable extraction unit 143 receives input of the dependent variable and the number i of explanatory variables used in the experimental design (S261). The explanatory variable extraction unit 143 identifies the number k1 of "first neighbor variables," which are variables that have a direct causal relationship with the dependent variable (S262). The explanatory variable extraction unit 143 compares the number k1 of first neighbor variables with the number i of explanatory variables used in the experimental design (S263). If the number k1 of first neighbor variables is equal to or greater than the number i of explanatory variables, the explanatory variable extraction unit 143 calculates the correlation coefficient between the first neighbor variables and the dependent variable for each variable, and extracts the i types of first neighbor variables in order from the largest absolute value of the correlation coefficient (S265), and terminates the explanatory variable extraction process.

[0043] On the other hand, if the number k1 of first proximity variables is smaller than the number i of explanatory variables, the explanatory variable extraction unit 143 extracts k1 types of first proximity variables (S264). Then, the explanatory variable extraction unit 143 identifies the number k2 of "second proximity variables," which are variables that connect to the objective variable via one intermediate factor (S266). The explanatory variable extraction unit 143 compares the sum of the number k1 of first proximity variables and the number k2 of second proximity variables with the number i of explanatory variables used in the experimental design (S267). If k1 + k2 is smaller than i, then k2 types of second proximity variables are extracted (S268). In this case, k1 types of first proximity variables and k2 types of second proximity variables are extracted, and the explanatory variable extraction process ends.

[0044] On the other hand, if k1+k2 is greater than or equal to i, the explanatory variable extraction unit 143 calculates the correlation coefficient between the second proximity variable and the objective variable for each variable, and extracts the (i-k1) types of second proximity variables in order from the one with the largest absolute value of the correlation coefficient (S269). In this case, the k1 types of first proximity variables and the (i-k1) types of second proximity variables are extracted, and the explanatory variable extraction process ends.

[0045] For example, if you want to design three experiments using explanatory variables using an L4 orthogonal array, you can extract explanatory variables using the causal graph in Figure 4 as follows: In step S261, the dependent variable is input, and since there are three explanatory variables used in the experimental design, i = 3. In step S262, since density is the only variable directly connected to the dependent variable by an arrow in the causal graph in Figure 4, k1 = 1 is specified. In step S263, since i = 3 is greater than k1 = 1, proceed to step S264. In step S264, density, which is the nearest neighbor variable in the causal graph in Figure 4, is extracted. In step S266, since composition and time are the only two variables connected to the dependent variable via one intermediate factor, density, in the causal graph in Figure 4, k2 = 2 is specified. In step S267, since k1 + k2 = 3 is equal to i = 3, proceed to step S269. In step S269, since i-k1=2, both the composition and time, which are the second-neighbor variables, are extracted. If a causal relationship from pressure to density exists, two variables with a high correlation coefficient with the objective variable are extracted from the three variables of composition, time, and pressure. In this example, three explanatory variables, density, composition, and time, are extracted, and an experiment can be designed using these three factors and an L4 orthogonal array.

[0046] (Display of explanatory variables) The variables extracted by the explanatory variable extraction unit 143 are output from the output device 130 in the form of, for example, display data.

[0047] (Effects of Example 1) In Patent Document 1, restrictions are placed on the connections between processes to improve the validity of causality. On the other hand, in Example 1, restrictions are not placed on the order of processes, so there is no omission of causality, and the quality of data obtained from experiments can be improved. In addition, since no prior hierarchical definition is required, it can be applied to various complex processes. Furthermore, since explanatory variables can be extracted systematically, experimental designs with low dependency on individuals can be made. Furthermore, by referring to the causal score, the validity of the search results for causal relationships can be improved.

[0048] The present invention is not limited to the above-described embodiments, but includes various modifications and equivalent configurations within the spirit and scope of the appended claims. For example, the above-described embodiments have been described in detail to clearly explain the present invention, and the present invention is not necessarily limited to configurations including all of the described configurations. Furthermore, part of the configuration of one embodiment may be replaced with the configuration of another embodiment. Furthermore, the configuration of another embodiment may be added to the configuration of one embodiment. Furthermore, part of the configuration of each embodiment may be added, deleted, or replaced with other configurations.

[0049] Furthermore, the aforementioned configurations, functions, processing units, processing means, etc. may be realized in part or in whole in hardware, for example by designing them as integrated circuits, or may be realized in software by having a processor interpret and execute a program that realizes each function.

[0050] Information such as programs, tables, and files that realize each function can be stored in a storage device such as a memory, a hard disk, or an SSD (Solid State Drive), or in a recording medium such as an IC card, an SD card, or a DVD.

[0051] In addition, the control lines and information lines shown are those that are considered necessary for explanation, and do not necessarily represent all the control lines and information lines that are necessary for implementation. In reality, it can be assumed that almost all components are interconnected. [Explanation of symbols]

[0052] 100 Experimental design support device 102 memory 110 Central Control Unit 120 Input Device 130 Output Device 140 Main storage 141 Causal Exploration Department 142 Structural Evaluation Department 143 Explanatory variable extraction section 150 Auxiliary storage 151 Experimental Data 152 Process Causal Information 153 Structural evaluation value information 160 Communication Equipment

Claims

1. An experiment design support device, The computer is configured by an arithmetic unit that executes predetermined processing and a storage device connected to the arithmetic unit, an input device for receiving input of a response variable and an explanatory variable for exploring a causal relationship with the response variable; a database in which the influence of the explanatory variables on the experimental process is stored; the calculation device has a causal relationship search unit that searches for a causal relationship between the explanatory variables; a structure evaluation unit that constructs a causal graph having a high evaluation value representing the degree of influence between the processes; The experimental design support device is characterized in that the arithmetic device comprises an explanatory variable extraction unit that uses the constructed causal graph to extract explanatory variables that have a strong causal relationship with the target variable.

2. 2. The experimental design support device according to claim 1, The experimental design support device is characterized in that the structural evaluation unit evaluates an evaluation value of the causal graph using the causal scores and penalty terms between the processes.

3. 3. The experimental design support device according to claim 2, The structure evaluation unit calculating an average value by averaging the value obtained by multiplying the causality score between the processes by a penalty term on the edges of the causality graph; The experimental design support device is characterized in that the calculated average value is averaged over the paths of the causal graph to calculate an evaluation value of the causal graph.

4. 2. The experimental design support device according to claim 1, The experimental design support device is characterized in that the structural evaluation unit calculates the causality score between the processes by averaging the causality scores of multiple data sets.

5. 3. The experimental design support device according to claim 2, The experimental design support device is characterized in that the penalty term reduces the causal score of the path of the causal graph to a value smaller than a predetermined threshold, thereby reducing the evaluation value of the causal graph.

6. 2. The experimental design support device according to claim 1, The experimental design support device is characterized in that the explanatory variable extraction unit extracts explanatory variables that have a large contribution to the target variable using the path length of the causal relationship and the correlation coefficient.

7. An experimental design support method executed by an experimental design support device, The experimental design support device includes: The computer is configured by an arithmetic unit that executes predetermined processing and a storage device connected to the arithmetic unit, an input device for receiving input of explanatory variables for exploring a causal relationship between an objective variable and the explanatory variables; and a database for storing the influence of the explanatory variables on the process of an experiment; The experimental design support method includes: a causal relationship search procedure in which the computing device searches for a causal relationship between the explanatory variables; a structural evaluation step in which the calculation device constructs a causal graph having a high evaluation value representing the degree of influence from the process; an explanatory variable extraction procedure in which the computing device uses the constructed causal graph to extract explanatory variables that have a strong causal relationship with the target variable.

Images