JP2014160423A - Experiment support system, experiment support method, and experiment support program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide an experiment support system capable of predicting conditions of a parameter for satisfying a target performance.SOLUTION: An experiment support system includes a storage device 101 for storing data, and a calculation device 104 for performing: a process of storing in a storage device 101 a collection of existing data including an existing experiment result in which a predetermined target input from an input device 105 is not achieved and a parameter of experiment conditions in the experiment result; a process of reading each collection of the existing data from the storage device 101, calculating difference data among all combinations of the read existing data, and synthesizing new data by adding the calculated difference data to each existing data; and a process of extracting and displaying on an output device 106 a rule of the parameter which satisfies a positive example and does not satisfy a negative example by inductive logic programming from the collection of predicates for expressing the parameter in each of new data, the data of the positive example in which the experiment result satisfies the predetermined target, and the data of the negative example in which the experiment result does not satisfy the predetermined target from among the synthesized data.

Description

  The present invention relates to an experiment support system, an experiment support method, and an experiment support program. Specifically, there is no data that satisfies a target performance as an experiment result, and parameters are unknown to satisfy the target performance. In this case, the present invention relates to a technology that makes it possible to predict parameter conditions that satisfy target performance.

  In order to finally obtain a desired experimental result, it is necessary to appropriately set various parameters that affect the experimental result. In particular, in order to efficiently generate a compound that satisfies a desired performance from a raw material by using a chemical reaction, parameters related to a large number of reaction conditions such as the concentration, flow rate, and temperature of the raw material for generating the compound are measured for each experiment. It is general to obtain an appropriate parameter value by repeatedly performing trial and error such as changing to. In this case, it is practically impossible to cover all possible parameter values by experiment, so appropriate analysis and inference methods are adopted, and an efficient experimental plan is focused on untrialized parameter values. Techniques for planning are proposed.

  Techniques for performing such an efficient experimental design include, for example, statistically significant characteristics in a database of acceptable and unacceptable compounds already determined by logistic regression tree analysis, and target compounds In comparison with the above feature, a technique (see Patent Document 1) that sifts out drug candidates without depending on an experimental test has been proposed.

JP 2001-221792 A

  By the way, when there is data that has already obtained the target performance as an experimental result, it is possible to extract the rule of the parameter for obtaining the target performance by applying an inductive learning method using the data. Under such conditions, a relational expression between the target performance and the parameter can be obtained by regression analysis, but it is impossible to specify which value of the parameter satisfies the target performance. Alternatively, it is possible to learn the relationship between the target performance and the parameter already obtained by the neural network technique, but it is still impossible to specify which value of the parameter satisfies the target performance. In other words, no technology has been proposed that makes it possible to predict the condition of a parameter that satisfies the target performance when the parameter is unknown to satisfy the target performance.

  Therefore, an object of the present invention is to provide a technique that makes it possible to predict the condition of a parameter that satisfies the target performance when there is no data that satisfies the target performance as an experimental result and the parameter is unknown to satisfy the target performance. There is.

  The experiment support system of the present invention that solves the above problems includes a storage device that stores data, an existing experiment result that is not achieved by a predetermined target, and a parameter of an experiment condition in the experiment result that is input from the input device. A first process for storing a set of existing data in the storage device, and reading each set of the existing data from the storage device, calculating difference data between all combinations of the read existing data, and calculating the calculation Second data for synthesizing new data by adding the difference data to each existing data, positive data satisfying a predetermined target among the new data combined, and negative not satisfying the predetermined target From the example data and the set of predicates that represent the parameters in each new piece of data, a positive example is satisfied and a negative example is satisfied by induction logic programming. An arithmetic unit that executes a third process of displaying are in the extracted output device rule parameters, characterized in that it comprises a. According to the present invention, even if there is no data satisfying the target performance as an experimental result and the parameter is unknown to satisfy the target performance, the condition of the parameter satisfying the target performance can be predicted.

  Here, in the experiment support system described above, in the second process, the arithmetic device reads each set of the existing data from the storage device, adds predetermined noise to the read existing data, and further adds a new noise. Combining data, and in the third process, the negative example data further includes the new data to which the predetermined noise is added, and the rule of the parameter that satisfies the positive example and does not satisfy the negative example by induction logic programming. It may be extracted and displayed on the output device. According to the present invention, it is possible to avoid rule extraction based on substantially the same data as existing data.

  In the above-described experiment support system, in the third process, the arithmetic device may include positive data that satisfies a predetermined target and negative data that does not satisfy the predetermined target among the synthesized new data. The MaxSAT problem is generated from the above, and the MaxSAT solver is calculated by giving the MaxSAT solver a hypothesis variable that assumes that the MaxSAT problem and a positive example that is not yet included in the rule is true, and the MaxSAT solver calculates the solution. In the solution, the process of obtaining the variable corresponding to the value on the rule that is true is repeatedly executed while switching the hypothesis variable that is true, thereby extracting the rule of the parameter that satisfies the positive example and does not satisfy the negative example. It may be displayed on the output device. According to this, as in the case of using inductive logic programming, if there is no data that satisfies the target performance as an experimental result and the parameter is unknown to satisfy the target performance, the condition of the parameter that satisfies the target performance Becomes predictable.

  Further, in the experiment support method of the present invention, the information processing apparatus includes a set of existing data including an existing experiment result that the predetermined target is not achieved and an experimental condition parameter in the corresponding experiment result, which are input from the input device Is stored in the storage device, each set of the existing data is read from the storage device, difference data is calculated between all combinations of the read existing data, and the calculated difference data is stored in each existing data A second process for synthesizing new data by adding to the data, a positive example data in which the experimental result satisfies the predetermined target, a negative example data not satisfying the predetermined target, and a new one of the synthesized new data From the set of predicates that express parameters in each data, the rule of parameters that satisfy positive examples and not negative examples is extracted by induction logic programming And executes a third process of displaying on the output device.

  Furthermore, the experiment support program of the present invention includes a set of existing data including an existing experiment result that has not been achieved by a predetermined target and an experiment condition parameter in the experiment result, which are input from the input device to the information processing apparatus. Is stored in the storage device, each set of the existing data is read from the storage device, difference data is calculated between all combinations of the read existing data, and the calculated difference data is stored in each existing data A second process for synthesizing new data by adding to the data, a positive example data in which the experimental result satisfies the predetermined target, a negative example data not satisfying the predetermined target, and a new one of the synthesized new data From the set of predicates that represent the parameters in each of the data sets, the inductive logic programming is used to create a parameter loop that satisfies the positive examples and does not satisfy the negative examples. A third process of displaying the extracted output device, characterized in that for the execution.

  According to the present invention, when there is no data satisfying the target performance as an experimental result and the parameter is unknown to satisfy the target performance, the condition of the parameter satisfying the target performance can be predicted.

It is a block diagram of the experiment support system of this embodiment. It is a figure which shows the example of a synthesis | combination of the new data in the experiment assistance method of this embodiment. It is a flowchart which shows the procedure example 1 of the experiment assistance method in this embodiment. It is a flowchart which shows the procedure example 2 of the experiment assistance method in this embodiment.

--- System configuration example ---
Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings. FIG. 1 is a configuration diagram according to the experiment support system 100 of the present embodiment. The experiment support system 100 shown in FIG. 1 predicts the condition of a parameter that satisfies the target performance when there is no data that satisfies the target performance as an experimental result in the chemical reaction experiment and the parameter is unknown to satisfy the target performance. It is a computer system that makes it possible.

  The hardware configuration of the experiment support system 100 is as follows. The experiment support system 100 includes a storage device 101 configured by an appropriate nonvolatile storage device such as a hard disk drive, a memory 103 configured by a volatile storage device such as a RAM, and a program 102 held in the storage device 101 in the memory 103. An arithmetic unit 104 such as a CPU that performs various operations and control processes while performing overall control of the apparatus itself by executing reading, etc., an input unit 105 that receives key input and voice input from a user, and a display that displays processing data The output device 106 is provided. Of course, the program 102 includes a program for implementing the inductive logic programming function 110.

  Then, the function with which the experiment support system 100 of this embodiment is provided is demonstrated. As described above, the functions described below are performed by executing the program 102 included in the experiment support system 100.

The experiment support system 100 according to the present embodiment stores, in the storage device 101, a set of existing data that is input from the input device 105 and that does not achieve the predetermined target and includes parameters of experiment conditions in the corresponding experiment result. It has a function to store in.
In addition, the experiment support system 100 reads each set of the existing data described above from the storage device 101, calculates difference data between all combinations of the read existing data, and adds the calculated difference data to each existing data. It has a function to synthesize new data.

  Furthermore, the experiment support system 100 includes, among the new data synthesized as described above, positive data in which the experimental result satisfies the predetermined target, negative data that does not satisfy the predetermined target, and parameters in each of the new data. From the set of predicates that express, by inductive logic programming, a rule for parameters that satisfy the positive example but does not satisfy the negative example is extracted and displayed on the output device 106.

  Note that the experiment support system 100 according to the present embodiment reads each set of existing data from the storage device 101 in the process of combining the new data described above, adds predetermined noise to each read existing data, and further adds New data is synthesized, and this is further included as negative example data described above, and by inductive logic programming, a rule for parameters that satisfy the positive example but does not satisfy the negative example is extracted and displayed on the output device 106. It is good.

この表１は、実際の化学実験データである９６個の既存データから適宜抜粋したものをまとめた表であり、Ａ、Ｂ、Ｃ、Ｄが反応物（すなわちパラメータ）、Ｅ、Ｆが生成物（すなわち実験結果）の各データを示している（各数値は対応成分の量を正規化したもの）。また、１レコードが１つの既存データに対応している。このような既存データの集合に対し、従来手法のまま帰納論理プログラミングを適用したとすれば、下に示すルールを抽出することができる。なお、正例となるデータは、Ｅ成分とＦ成分とが共に０．６以上、と指定した。
［Ｒｕｌｅ１］［Ｐｏｓｃｏｖｅｒ＝６Ｎｅｇｃｏｖｅｒ＝０］
ｅｘａｍｐｌｅ（Ｘ）：−ｂ（Ｘ，０．２５），ｄ（Ｘ，０．８）．
［Ｒｕｌｅ２］［Ｐｏｓｃｏｖｅｒ＝４Ｎｅｇｃｏｖｅｒ＝０］ --- Concept of experiment support method ---
Here, in order to explain the concept of the experiment support method of the present embodiment, qualitative parameters are obtained from the actual chemical experiment data (Table 1) using inductive logic programming according to the conventional method and the respective ideas of the present embodiment. A specific example when trying to extract properties will be shown. Among the chemical experiment data, the experiment data (positive example) that satisfies the target performance is explained, and the rule that does not explain the experiment data (negative example) that does not satisfy the performance is extracted by induction logic programming.

This Table 1 is a table summarizing appropriately extracted data from 96 existing data that are actual chemical experiment data. A, B, C, and D are reactants (that is, parameters), and E and F are products. (I.e., experimental results) are shown (each value is a normalized value of the corresponding component). One record corresponds to one existing data. If inductive logic programming is applied to such a set of existing data in the conventional manner, the following rules can be extracted. In addition, the data which becomes a positive example specified that E component and F component were both 0.6 or more.
[Rule1] [Poscover = 6Negcover = 0]
example (X): -b (X, 0.25), d (X, 0.8).
[Rule2] [Poscover = 4Negcover = 0]

  The above first rule “Rule 1” indicates that the target performance is satisfied if the B component is 0.25 and the D component is 0.8. The next rule “Rule 2” indicates that the C component is 0. .7, the D component is 1 and the B component is 0.25 or less, it indicates that the target performance is satisfied. In this way, if there is a positive experimental result, it is possible to extract the parameter conditions to obtain the desired performance even if the conventional method remains as it is, and to apply appropriate rules that can actually be realized in the laboratory. Can be obtained.

  On the other hand, even if the conventional method as described above is applied to “chemical experiment data without a normal example” premised on the experiment support method of the present embodiment, it is not possible to extract parameter conditions for obtaining desired performance. Therefore, in the vicinity of the chemical experiment data, linearity is assumed between the parameter and the performance target, and the experiment support method of this embodiment is applied to virtually synthesize the chemical experiment data to form new data. It was.

  FIG. 2 shows a synthesis example when new data is generated from chemical experiment data (existing data) of only negative examples. Here, the difference between each two data in the chemical experiment data group is calculated, the difference data (b) obtained here is added to the other chemical experiment data (a), and new data (c) is synthesized. doing. Virtual logic data as new data obtained in this way (some of which satisfy the target condition, some of which do not satisfy the target) and the original chemical experiment data group are combined, and inductive logic programming is performed. Apply and rule extraction. The rules thus obtained were evaluated by experimental chemists as being generally valid.

--- Processing example of experiment support method ---
Hereinafter, the actual procedure of the experiment support method in the present embodiment will be described with reference to the drawings. Various operations corresponding to the experiment support method described below are realized by a program 102 that the experiment support system 100 reads into the memory 103 and executes. And this program 102 is comprised from the code | cord | chord for performing the various operation | movement demonstrated below.

  FIG. 3 is a flowchart showing a processing procedure example 1 of the experiment support method in the present embodiment. Here, first, the arithmetic unit 104 of the experiment support system 100 reads out an existing data group including an experimental result obtained from an already performed experiment and parameters at that time from the storage device 101, and based on the read out existing data group, A process of calculating difference data between all combinations of existing data is executed (s100). The part of the calculation procedure relating to the difference data (b) in FIG. 2 already shown corresponds to the step s100.

  Subsequently, the computing device 104 of the experiment support system 100 synthesizes new data by adding the difference data calculated in step s100 described above to each existing data read from the storage device 101 in step s100 (s101). The new data synthesized here is hereinafter referred to as synthesized data. The part of the calculation procedure relating to the new data (c) shown in FIG. 2 already corresponds to the step s101.

  Note that it is preferable that the arithmetic unit 104 of the experiment support system 100 adds predetermined noise to each of the existing data described above and further generates composite data in step s101. Generating synthesized data by adding noise in this way intends to exclude data similar to existing data from the processing target. Data similar to existing data is considered to give a result similar to existing data, but inductive logic programming treats data as completely different data as long as the data values are different from each other. Therefore, preventing generation of meaningless data, that is, data having a value close to existing data, by simply increasing the processing load of the experiment support system 100 leads to an effect of improving processing efficiency.

  The specific range of the noise described above is defined in advance for each parameter. For example, when the two values existing for the parameter can be regarded as almost the same value, the difference between the two values is excluded, and the minimum difference value of the other binary combinations is noted. The absolute value of the difference value is set to the upper limit value of the absolute value of noise. As the noise step value, the smallest absolute value of the difference between the binary combinations of the corresponding parameters is used. In each parameter, a value is shifted by a step value in a section of −upper limit value to + upper limit value, and noise is generated and used for the above-described noise applying process.

  Now, the description returns to the flow of FIG. Next, the arithmetic unit 104 of the experiment support system 100 includes positive data satisfying a predetermined target and negative data not satisfying the predetermined target among the combined data obtained up to step s101 described above. Further, each parameter value described in the predicate logic format with respect to the combined data of each of the positive and negative examples specified here is stored in the storage device 101 (s102). Here, the arithmetic unit 104 stores each composite data as a positive example in a positive example file provided in the storage device 101, and each composite data as a negative example is stored in a negative example file provided in the storage device 101. Each parameter value described in the predicate logic format is stored in a background knowledge file in the storage device 101.

As an example of the above-described positive example file, a file in which each data of the positive example is stored is as follows.
example (no_0_0).
example (no_1_0).
example (no_2_0).
...

Moreover, as an example of the above-mentioned negative example file, it becomes a format which stored each data of the negative example as follows.
example (no_34_0).
example (no_35_0).
example (no_36_0).
...

Similarly, an example of the above-described background knowledge file has a format in which each parameter value is stored as follows.
a (no_0_0, a_51_0).
b (no_0_0, b_3_5).
c (no_0_0, c_20_5).
d (no_0_0, d_4_0).
e (no_0_0, e_7_0).
f (no_0_0, f_5_0).
g (no_0_0, g_0_0).
...

Subsequently, the arithmetic unit 104 of the experiment support system 100 receives the positive example file, the negative example file, and the background knowledge file described above as inputs, and causes the inductive logic programming function 110 to execute the parameters. Rules are extracted and displayed on the output device 106 (s103). An example of the rule obtained in step s103 is shown below.
[Rule1] [Poscover = 6Negcover = 0]
example (X): -b (X, 0.25), d (X, 0.8).
The rule “Rule 1” obtained here indicates that the target performance is satisfied when the B component is 0.25 and the D component is 0.8.

  Note that the process of performing rule extraction using the inductive logic programming function 110 in step s103 described above can be substituted by solving the MaxSAT problem with the MaxSAT solver in the experiment support system 100. Hereinafter, the processing content in that case will be described. The flow of FIG. 4 shows a procedure for performing rule extraction using MaxSAT instead of induction logic programming. Here, it is assumed that the experiment support system 100 holds a program that implements the MaxSAT solver function and data necessary for related processing in the storage device 101 in advance.

  In this case, in the experiment support system 100, a MaxSAT problem is generated using the above-described existing data, and a hypothesis in which the MaxSAT problem and a positive example variable that holds a positive example that is not yet included in a rule holds is True, The solution is given to the MaxSAT solver (s200 to s202).

  Then, in the solution obtained by the MaxSAT solver, a variable corresponding to the value on the rule that becomes True is obtained. In this case, the experiment support system 100 counts the number of p (x) that is True in the solution of the MaxSAT problem, sets the counted value as the number of positive examples that apply to the rule, and sets p (x) that is True. If the number is 2 or more, a rule to be output is generated (s203). For example, in the solution of the MaxSAT problem, r (i (b), j (0.25)) = True, r (i (d), j (0.8)) = True, and other r (i , J) is False, the rule on the second line in FIG. 8 is derived.

  In this way, the experiment support system 100 obtains the same result as that obtained when the inductive logic programming is executed by repeatedly obtaining a solution with the MaxSAT solver while switching the positive example variable to be True as a hypothesis.

この式が（１≦ｉ≦ｌ、１≦ｊ（ａ）≦ｍ（ｉ）、１≦ｊ（ｂ）≦ｍ（ｉ）、ｊ（ａ）≠ｊ（ｂ））を満たす全てのｉ、ｊ（ａ）、ｊ（ｂ）の組み合わせについて成立する。ここで、ｒ（ｉ、ｊ（ａ））はｉ番目のパラメータがｊ（ａ）番目の値となることを示す変数であり、ｒ（ｉ、ｊ（ｂ））はｉ番目のパラメータがｊ（ｂ）番目の値となることを示す変数である。また、ｌはパラメータの総数であり、ｍ（ｉ）はｉ番目のパラメータが取りうる値の総数である。 The method for generating the above MaxSAT problem is as follows. Here, a MaxSAT problem including five types of hard clauses and one type of soft clause is generated. Of these, the concept of the hardware section is as follows.
1. The i-th parameter does not have two values on the rule at the same time. Therefore, the following expression (1) is established.

All i satisfying (1 ≦ i ≦ l, 1 ≦ j (a) ≦ m (i), 1 ≦ j (b) ≦ m (i), j (a) ≠ j (b)) This holds for combinations of j (a) and j (b). Here, r (i, j (a)) is a variable indicating that the i-th parameter becomes the j (a) -th value, and r (i, j (b)) is the i-th parameter j. (B) This is a variable indicating that the value is the first. Also, l is the total number of parameters, and m (i) is the total number of values that the i-th parameter can take.

この式が（１≦ｘ≦ｋ（ｐ））を満たす全てのｘの組み合わせについて成立する。ここで、ｊ（ｉ、ｘ）はｘ番目の正例でのｉ番目のパラメータが取る値がｉ番目のパラメータが取りうる値の中で何番目の値かを示し、１≦ｊ（ｉ、ｘ）≦ｍ（ｉ）が成り立つ整数である。そして、ｒ（ｉ、ｊ（ｉ、ｘ））はルールでｉ番目のパラメータがｘ番目の正例の値となったことを表わす。また、ｐ（ｘ）はｘ番目の正例をルールが満たすことを示す変数である。また、ｌはパラメータの総数であり、ｍ（ｉ）はｉ番目のパラメータが取りうる値の総数であり、ｋ（ｐ）は正例の総数である。 2. If the parameter value of the positive example data is satisfied by the rule, it is assumed that the positive example is satisfied. Therefore, the following expressions (2) and (3) are established.

This formula holds for all combinations of x satisfying (1 ≦ x ≦ k (p)). Here, j (i, x) indicates what value the i-th parameter takes in the x-th positive example among the values that the i-th parameter can take, and 1 ≦ j (i, x) is an integer that satisfies m (i). R (i, j (i, x)) represents that the i-th parameter is the value of the x-th positive example in the rule. P (x) is a variable indicating that the rule satisfies the xth positive example. Further, l is the total number of parameters, m (i) is the total number of values that the i-th parameter can take, and k (p) is the total number of positive examples.

この式が（１≦ｘ≦ｋ（ｐ）、１≦ｊ（ｉ、ｃ）≦ｍ（ｉ）、ｊ（ｉ、ｃ）≠ｊ（ｉ、ｘ））を満たす全てのｉ、ｃ、ｘの組み合わせについて成立する。ここで、ｊ（ｉ、ｘ）はｘ番目の正例でのｉ番目のパラメータが取る値がｉ番目のパラメータが取りうる値の中で何番目の値かを示し、ｊ（ｉ、ｃ）もｉ番目のパラメータが取りうる値の中で何番目の値かを示すが、ｊ（ｉ、ｘ）は除き，１≦ｊ（ｉ、ｃ）≦ｍ（ｉ）が成り立つ整数である。そして、ｒ（ｉ、ｊ（ｉ、ｃ））はルールでｉ番目のパラメータがｃ番目の正例の値となったことを表わす。また、ｐ（ｘ）はｘ番目の正例をルールが満たすことを示す変数である。また、ｌはパラメータの総数であり、ｍ（ｉ）はｉ番目のパラメータが取りうる値の総数であり、ｋ（ｐ）は正例の総数である。

All i, c, x satisfying (1 ≦ x ≦ k (p), 1 ≦ j (i, c) ≦ m (i), j (i, c) ≠ j (i, x)) This holds for combinations of Here, j (i, x) indicates what value the i-th parameter takes in the x-th positive example among the values that the i-th parameter can take, and j (i, c) Indicates the number of values that can be taken by the i-th parameter, except for j (i, x), which is an integer satisfying 1 ≦ j (i, c) ≦ m (i). R (i, j (i, c)) represents that the i-th parameter has become the value of the c-th positive example in the rule. P (x) is a variable indicating that the rule satisfies the xth positive example. Further, l is the total number of parameters, m (i) is the total number of values that the i-th parameter can take, and k (p) is the total number of positive examples.

この式が（１≦ｉ≦ｌ）を満たす全てのiについて成立する。ここで、ｌはパラメータの総数である。

この式が（１≦ｉ≦ｌ、１≦ｊ（ｄ）≦ｍ（ｉ））を満たす全てのｉ、ｊ（ｄ）について成立する。ここで、ｌはパラメータの総数であり、ｍ（ｉ）はｉ番目のパラメータが取りうる値の総数である。 3. It is represented by f (i) that the i-th parameter is unconditional. The following formulas (4) and (5) hold for f (i).

This equation holds for all i satisfying (1 ≦ i ≦ l). Here, l is the total number of parameters.

This equation holds for all i and j (d) that satisfy (1 ≦ i ≦ l, 1 ≦ j (d) ≦ m (i)). Here, l is the total number of parameters, and m (i) is the total number of values that the i-th parameter can take.

この式が（１≦ｙ≦ｋ（ｎ）、Ｑ（ｉ、ｙ）はｒ（ｉ、ｊ（ｉ、ｙ））またはｆ（ｉ）に置き換わる）を満たす全てのｙ、Ｑ（ｉ、ｙ）の組み合わせについて成立する。ここで、ｊ（ｉ、ｙ）はｙ番目の負例でのｉ番目のパラメータが取る値がｉ番目のパラメータが取りうる値の中で何番目の値かを示し、１≦ｊ（ｉ、ｙ）≦ｍ（ｉ）が成り立つ整数である。そして、ｒ（ｉ、ｊ（ｉ、ｙ））はルールでｉ番目のパラメータがｙ番目の負例の値となったことを表わす。また、ｆ（ｉ）はｉ番目のパラメータが無条件であることを示す変数である。また、ｎ（ｙ）はｙ番目の負例をルールが満たすことを示す変数である。また、ｌはパラメータの総数であり、ｍ（ｉ）はｉ番目のパラメータが取りうる値の総数であり、ｋ（ｎ）は負例の総数である。 4). If the parameter value of negative example data is satisfied by the rule, it is assumed that the negative example is satisfied. Therefore, the following equation (6) is established.

All y, Q (i, y) where this expression satisfies (1 ≦ y ≦ k (n), Q (i, y) replaces r (i, j (i, y)) or f (i)) ). Here, j (i, y) indicates what value the i-th parameter takes in the y-th negative example among the values that the i-th parameter can take, and 1 ≦ j (i, y) is an integer that satisfies m (i). R (i, j (i, y)) represents that the i-th parameter is a value of the y-th negative example in the rule. F (i) is a variable indicating that the i-th parameter is unconditional. N (y) is a variable indicating that the rule satisfies the yth negative example. Also, l is the total number of parameters, m (i) is the total number of values that the i-th parameter can take, and k (n) is the total number of negative examples.

この式が（１≦ｙ≦ｋ（ｎ））を満たす全てのｙについて成立する。ここでｎ（ｙ）はｙ番目の負例をルールが満たすことを示す変数である。また、ｋ（ｎ）は負例の総数である。 5. Negative examples cannot hold. Therefore, the following equation (7) is established.

This equation holds for all y satisfying (1 ≦ y ≦ k (n)). Here, n (y) is a variable indicating that the rule satisfies the yth negative example. K (n) is the total number of negative examples.

Next, the concept of the soft section is as follows.
The following formula represents that the positive example is established.
p (x)
This equation exists for all x satisfying (1 ≦ x ≦ k (p)). Here, p (x) is a variable indicating that the rule satisfies the xth positive example. K (p) is the total number of positive examples.

--- Specific application results ---
Next, an actual experiment is performed to obtain data, and based on this data, a specific example will be described in which a parameter extracted by the experiment support system 100 of the present embodiment is tried in a subsequent experiment. . The experiment described here is an experiment in which a water-based paint containing the corresponding polyester resin has an excellent water dispersion stability and the polyester resin composition has been examined so that the coating film exhibits excellent corrosion resistance. . Such a polyester resin is a main resin required for paints for pre-coated metal steel sheets for household appliances and painted steel sheets for automobiles, and is required to reduce VOC without using organic solvents and to have high corrosion resistance during processing.
Regarding the evaluation corresponding to the target performance, the measurement item corresponding to the parameter, and the processing method, the following concept and method were followed. In the following description, “parts” means parts by weight.

(1) Reduced viscosity ηsp / c (dl / g)
A sufficiently dried polyester resin (0.10 g) was dissolved in 25 ml of a mixed solvent of phenol / tetrachloroethane (weight ratio 6/4) and measured at 30 ° C. This reduced viscosity is one of the measures representing the molecular weight of polyester.

(2) Composition analysis of polyester resin The analysis was performed by 400 MHz proton NMR (nuclear magnetic resonance method).

(3) Polyester resin synthesis example In the synthesis of the polyester resin (S1), 100 parts of dimethyl terephthalate, 100 parts of dimethyl isophthalate, and an ethylene oxide adduct of bisphenol A were added to a reaction vessel equipped with a stirrer, a thermometer and a condenser. 176 parts, 100 parts of ethylene glycol, and 0.1 part of tetrabutyl titanate as a catalyst were charged, and a transesterification reaction was carried out for 5 hours while distilling methanol generated at 180 to 230 ° C. out of the system. Subsequently, the inside of the system was gradually brought into a reduced-pressure atmosphere, and a polycondensation reaction was performed at 0.3 mmHg and 255 ° C. for 40 minutes. Subsequently, it cooled to 200 degreeC under normal pressure and nitrogen atmosphere, 4 parts of trimellitic anhydride was added, it heated gradually to 220 degreeC, and trimellitic acid was added, and polyester resin (S1) was obtained.
As a result of analysis such as NMR, the obtained polyester resin was terephthalic acid / isophthalic acid // ethylene oxide adduct of bisphenol A / ethylene glycol // trimellitic acid (post-addition) = 50/50 // 55/45 / / 2 (molar ratio) and a pale yellow transparent resin having a reduced viscosity of 0.43 dl / g.

(4) Synthesis of polyester resins (S2) to (S13) In the same manner as in the synthesis example of polyester resin (S1), polyester resins (S2) to (S13) whose compositions are shown in Table 2 were synthesized. Moreover, the composition analysis and the measurement of the resin characteristic were performed similarly to the polyester resin (S1). Table 2 shows the following.

(5) Preparation of water-based paint A four-necked flask equipped with a stirrer, a thermometer and a condenser was charged with 15 parts of a polyester resin and 85 parts of tetrahydrofuran. After stirring and dissolving the resin at 45 ° C., 135 parts of water If necessary, a predetermined amount of triethylamine for neutralizing carboxylic acid groups is added. After switching the serpentine cooling pipe to the Vigreux fractionating pipe, the solution is heated to 100 ° C., and a predetermined amount of tetrahydrofuran is discharged out of the system. For 150 parts of the polyester resin aqueous dispersion prepared as described above, after cooling the aqueous dispersion, water-dispersed colloidal silica (“OUP-ST” chain colloidal silica manufactured by Nissan Chemical Industries, Ltd., particle size of 40 to 300 nm, solid A water-based paint was prepared by mixing and dispersing 45 parts of a 10% component concentration).

(6) Preparation and deep drawing of coated steel sheet Zinc / nickel-plated steel sheet previously degreased with 75% degreasing agent Surf Cleaner 75N (made by Nippon Paint Co., Ltd.) and pickled with 2% H2SO4 aqueous solution ( ZL material, zinc basis weight 30 g / m 2) was subjected to chromate treatment so as to be 40 to 60 mg / m 2 in terms of chromium, to obtain a base material for a coated steel sheet. Next, the above-mentioned water-based paint was applied with a bar coater so that the dry film thickness was 1.5-2 μm, and baked at a steel plate temperature of 120 ° C. for 20 seconds to prepare a coated steel plate. After cooling this steel plate, it was cylindrically processed by a deep drawing machine so as to be a cup material having an outer diameter of 33 mm and a height of 28 mm. The coating thickness was calculated from the change in weight of the steel sheet before and after coating and drying and the specific gravity of the polyester resin and silica gel.

As described above, the produced water-based paint and deep-drawn painted steel sheet were evaluated as follows.
(7) Water dispersion stability The water-based coating composition was visually judged after storage at room temperature for 1 to 3 months. Judgment criteria are as follows.
5: Good in storage for 3 months 4: Good in storage for 1 month, Small amount of precipitate or gel is generated in storage for 3 months.
3: A small amount of precipitate or gel is generated during storage for one month.
2: Separation and gelation after 1 month storage.
1: A good aqueous dispersion cannot be obtained in the initial stage of production.

(8) Corrosion resistance after processing The deep-drawn coated steel sheet was placed in an SST test tank for a salt spray test, and the surface state of the processed part after 1500 hours was observed. Judgment criteria are as follows.
5: No white rust.
4: Less than 5% white rust 3: 5% or more and less than 20% white rust.
2: White rust 20% or more and less than 50%, red rust less than 5%.
1: White rust 50% or more, red rust 5% or more

  The evaluation results based on the above determination criteria are shown as the total evaluation score at the bottom of Table 2 above. The performance target was a total of 9 or more evaluation points, and each sample did not reach the target.

Therefore, based on the above experimental results (Table 2), the inductive logic programming in the experiment support method of the present embodiment is applied to extract 20 rules of parameters that satisfy the positive example but do not satisfy the negative example. Based on the rule that derived the many examples of positive composition by eliminating the problematic composition and duplicate examples, the experiment plan shown in Table 3 below (T1) to (T3) was established, and the experiment was conducted. As a result, a total of 9 target evaluation points were reached.

On the other hand, after carrying out the evaluation of the polyester resins (S1) to (S13), in view of the knowledge and the like so far, in order to reach the target under the condition that the result of the above inductive logic programming is not recognized at all, Subsequently, different composition experiments were conducted. Here, apart from the polyester resins (S1) to (S13), polyester resins (S14) to (S28) were prepared and evaluated in the same manner as described above. The results are shown in Table 4. The highest point of the total evaluation points increased from 7 points to 8 points, approaching the target evaluation point total of 9 points or more, but still did not reach the target.

Based on a total of 28 experimental examples of the above-described polyester resins (S1) to (S28), 7 rules for deriving positive examples were calculated by inductive logic programming combined with MaxSAT solvers. Among them, the composition of problems in production and the rule that eliminates duplicate examples and derived many positive examples, the experimental plan shown in Table 5 below, the experimental plan of polyester resins (T4) to (T5) When the experiment was carried out, the level of reaching the target was shown against a total of 9 target evaluation points.

Therefore, according to the present embodiment, when there is no data that satisfies the target performance as an experimental result and the parameter is unknown to satisfy the target performance, the condition of the parameter that satisfies the target performance can be predicted.
Although the best mode for carrying out the present invention has been specifically described above, the present invention is not limited to this, and various modifications can be made without departing from the scope of the invention.

100 Experiment Support System 101 Storage Device 102 Program 103 Memory 104 Arithmetic Device 105 Input Device 106 Output Device 110 Inductive Logic Programming Function 125 Positive Example File 126 Negative Example File 127 Background Knowledge File

Claims (5)

A storage device for storing data;
A first process of storing, in the storage device, a set of existing data that is input from an input device and that includes an existing experimental result that does not achieve a predetermined target and a parameter of an experimental condition in the corresponding experimental result;
Read each set of existing data from the storage device, calculate difference data between all combinations of the read existing data, and synthesize new data by adding the calculated difference data to each existing data Two treatments,
Among the synthesized new data, an inductive logic is obtained from positive example data in which the experimental result satisfies a predetermined target, negative example data that does not satisfy the predetermined target, and a set of predicates expressing parameters in each of the new data. An arithmetic device that executes a third process of extracting a rule of a parameter that satisfies a positive example and does not satisfy a negative example by programming and displays the rule on an output device;
An experiment support system comprising: The arithmetic unit is:
In the second process, each set of the existing data is read from the storage device, a predetermined noise is given to the read existing data, and further new data is synthesized,
In the third process, as the negative example data, the new data added with the predetermined noise is further included, and the rules of parameters that satisfy the positive example and do not satisfy the negative example are extracted by induction logic programming to the output device. Is to display,
The experiment support system according to claim 1. The arithmetic unit is:
In the third process, among the synthesized new data, a MaxSAT problem is generated from positive example data in which an experimental result satisfies a predetermined target and negative example data that does not satisfy a predetermined target, and the MaxSAT problem, In the solution calculated by the MaxSAT solver, a hypothesis that a true example not yet included in the rule holds is assumed to be True, and a solution is calculated. In the solution calculated by the MaxSAT solver, the value on the rule that becomes True is obtained. By repeatedly executing the process of obtaining the corresponding variable while switching the hypothesis variable as True, the rule of the parameter that satisfies the positive example and does not satisfy the negative example is extracted and displayed on the output device.
The experiment support system according to claim 1, wherein the system is an experiment support system. Information processing device
A first process of storing, in the storage device, a set of existing data that is input from an input device and that includes an existing experimental result that does not achieve a predetermined target and a parameter of an experimental condition in the corresponding experimental result;
Read each set of existing data from the storage device, calculate difference data between all combinations of the read existing data, and synthesize new data by adding the calculated difference data to each existing data Two treatments,
Among the synthesized new data, an inductive logic is obtained from positive example data in which the experimental result satisfies a predetermined target, negative example data that does not satisfy the predetermined target, and a set of predicates expressing parameters in each of the new data. A third process for extracting a rule of a parameter that satisfies a positive example and does not satisfy a negative example by programming and displays the rule on an output device;
An experiment support method characterized by executing In the information processing device,
A first process of storing, in the storage device, a set of existing data that is input from an input device and that includes an existing experimental result that does not achieve a predetermined target and a parameter of an experimental condition in the corresponding experimental result;
Read each set of existing data from the storage device, calculate difference data between all combinations of the read existing data, and synthesize new data by adding the calculated difference data to each existing data Two treatments,
Among the synthesized new data, an inductive logic is obtained from positive example data in which the experimental result satisfies a predetermined target, negative example data that does not satisfy the predetermined target, and a set of predicates expressing parameters in each of the new data. A third process for extracting a rule of a parameter that satisfies a positive example and does not satisfy a negative example by programming and displays the rule on an output device;
An experiment support program characterized by causing

Images