JP2000155680A - Device and method for correcting judgment rule - Google Patents

Abstract

PROBLEM TO BE SOLVED: To partially and quickly correct a judgment rule by extracting a part in which the inference of a judgment rule described in a system is erroneous from a data set, and correcting this part by using an algorithm capable of parameter adjustment. SOLUTION: Events constituted of classified classes corresponding to image file names are stored in an event storing part 11, and judgment rules described in a system which is used at present are stored in a judgment rule storing part 12. Then, the events stored in the vent storing part 11 are inferred by an inferring part 13 according to the judgment rules stored in the judgment rule storing part 12. When the inferred class is different from the classified class corresponding to the stored event, an object to be corrected extracting part 14 extracts a parameter appearing on a path related with the event as the object to be corrected of the judgment rule. A retrieval range to be corrected is stored in a retrieval range storing part 15. At last, retrieval is operated in a range to be corrected by using an adjustable prescribed algorithm by an object to be corrected retrieving part 16 so that the parameter can be attained.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a system for identifying a defect occurring in a product by inputting, for example, image or audio data and using a matching with a product stored in advance as a defective product, or a current control object. By inputting a state quantity indicating a situation, in a system for controlling a control target, a device and a judgment rule for automatically correcting a judgment rule in the system so as to match a diagnosis result given by an expert are automatically set. How to fix.

[0002]

2. Description of the Related Art Conventionally, as optimization or search by machine intelligence, a decision rule is determined from a feature quantity obtained by substituting a parameter into a predetermined function, and when the decision rule is modified, the parameter is determined. The method of adjusting is adopted. One method is the Genetic Algorithm
(Hereinafter referred to as “GA”); References “Genetic Algorithm / Takeharu Yasui, Tomoharu Nagao / Shokodo / 19
1993. " GA is a kind of optimization / search algorithm based on the evolutionary process of living organisms. In GA, chromosome using the operation of exchanging a part of chromosomes (crossing), the operation of inverting the value of some character strings constituting the chromosome (mutation), and the evaluation value (fitness) of each chromosome An operation (selection) of selecting a chromosome from a group is performed to generate a new chromosome group. GA is an algorithm for searching for a chromosome with a high degree of fitness by repeating such generation of a chromosome population for several generations.

For example, as a method of adjusting parameters included in a decision rule using this GA, a reference document “One method for obtaining and expressing fuzzy rules / Satoru Maebashi,
Takehisa Onizawa / Kanto Branch and Tohoku Branch of the Japan Fuzzy Society
9th Workshop / 1997 ". In this method, the parameters included in the judgment rule are represented by a character string obtained by connecting 0 or 1 (binary representation), thereby associating the judgment rule with the chromosome. In addition, the degree of fitness of the chromosome is determined by evaluating how well the control target can be controlled using the determination rule corresponding to the chromosome. In this way, the decision rules are adjusted so that chromosomes with high relevance are eliminated while the chromosomes repeat generational changes.

[0004] However, even if a GA for searching for an optimal solution is used, it takes an enormous amount of time to search all in a normally considered search space. Furthermore, this method also has a problem that cannot be solved practically in a limited time. Also, when it is necessary to search for a point in the search space that gives the true maximum or minimum value, there is no other way than to perform a full search, even if an enormous amount of time is required. there were.

[0005]

As described above, in the prior art, it is premised that all the corrections of the decision rules are regenerated from the beginning. Since the whole is regarded as an object to be modified, there is a problem that it takes a lot of time to modify the rule and the efficiency is low.

SUMMARY OF THE INVENTION The present invention has been made to solve the above-described problem. When inference is performed on given data, if a correct inference result cannot be obtained for the data, the data is inferred. It is an object of the present invention to provide a method and an apparatus for partially and quickly modifying a decision rule so as to lead to a correct inference result.

Another object of the present invention is to improve the size of the search space, which has been a concern in the use of GA, that is, the fact that the search range is too wide and the processing takes an enormous amount of time. Therefore, it is intended to make GA a more easy-to-use search method.

[0008]

In order to solve the above-mentioned problem, a judgment rule correcting device according to the present invention according to claim 1 is provided.
Means for extracting, from a data set having both data and a classification class for identifying the data, a part in which the inference of the judgment rules described in the system is incorrect, and an algorithm capable of adjusting the parameters of the extracted judgment rules. Means for using the correction.

According to a second aspect of the present invention, there is provided a decision rule correcting apparatus according to the present invention, for a data set having both data and a classification class for identifying the data, a part in which the inference of the decision rule described in the system is incorrect. Means for extracting, a means for correcting the extracted decision rule using an algorithm capable of adjusting parameters, and a parameter on a decision rule corresponding to the data set for a data set led to an incorrect inference result. Means for extracting, and for a classification class to which a data set led to an erroneous inference result belongs, means for extracting a parameter on a decision rule corresponding to the classification class.

According to a third aspect of the present invention, there is provided a decision rule correcting device for a data set having both data and a classification class for identifying the data, wherein a part in which the inference of the decision rule described in the system is incorrect is performed. Means for extracting, a means for correcting the extracted decision rule using an algorithm capable of adjusting parameters, and a parameter on a decision rule corresponding to the data set for a data set led to an incorrect inference result. Means for extracting, for a classification class to which a data set led to an erroneous inference result belongs, means for extracting a parameter on a judgment rule corresponding to the classification class, and for a parameter extracted in claim 2, A means for extracting only parameters that need to be modified on an inference route that provides a correct inference result for a data set that leads to an incorrect inference result Comprising.

According to a fourth aspect of the present invention, there is provided a method for correcting a judgment rule according to the present invention, wherein a part of a data set having both data and a classification class for identifying the data in which the inference of the judgment rule described in the system is erroneous. Extracting, and correcting the extracted judgment rule using an algorithm capable of adjusting parameters.

According to a fifth aspect of the present invention, there is provided a method for modifying a decision rule according to the present invention, wherein a part of a data set having both data and a classification class for identifying the data in which the inference of the decision rule described in the system is erroneous. Extracting, correcting the extracted decision rule using an algorithm capable of adjusting parameters; and targeting a data set led to an incorrect inference result to a parameter on a decision rule corresponding to the data set. Extracting, and for a classification class to which a data set derived from an incorrect inference result belongs, extracting a parameter on a determination rule corresponding to the classification class.

According to a sixth aspect of the present invention, there is provided a method for correcting a decision rule according to the present invention, wherein a part of a data set having both data and a classification class for identifying the data, in which the inference of the decision rule described in the system is erroneous. Extracting, correcting the extracted decision rule using an algorithm capable of adjusting parameters; and targeting a data set led to an incorrect inference result to a parameter on a decision rule corresponding to the data set. Extracting, for a classification class to which a data set led to an erroneous inference result belongs, extracting a parameter on a determination rule corresponding to the classification class, and for a parameter extracted in claim 5, Data set that leads to incorrect inference result Only parameters that need correction on the inference path from which the correct inference result is selected Comprising the step of extracting.

In the modification rule judging device and the modification rule judging method according to the present invention, all the parameters have been set as the search range so far, but only the parameters that need to be modified and the parameters around them are described. By limiting the range of correction, it became possible to shorten the search time until the optimal solution was obtained. This not only makes it easier for the user to set search parameters, but also satisfies the conditions under which the GA can withstand a real system.

In the modification rule judging apparatus and the modification rule judging method according to the present invention, one of the methods for extracting a parameter that needs to be modified and parameters around the parameter is one of an incorrect classification class. The parameters are extracted for the classification class to which the data derived to the original belongs, and for the other only for the data derived to the wrong classification class. Then, the union of the parameters obtained by these extractions is set as the correction target range. According to the modification rule judging device and the modification rule judging method of the present invention described in claims 3 and 6, only the part having a high possibility of modification is extracted from the parameters extracted above in relation to the characteristic amount. This further scrutinizes the search range, leading to a reduction in search time.

[0016]

Embodiments of the present invention will be described below in detail with reference to the drawings.

FIG. 1A is a block diagram showing a configuration of an apparatus according to an embodiment of the present invention. As shown in the figure,
The case storage unit 11 stores a case including a classification class corresponding to the image file name shown in FIG. The judgment rule storage unit 12 stores judgment rules described in the currently used system. Then, the case stored in the case storage unit 11 is inferred by the inference unit 13 according to the judgment rule stored in the judgment rule storage unit 12. Correction target extraction unit 1
When the class inferred by the inference unit 13 is different from the classification class corresponding to the case stored in the case storage unit 11, the parameter 4 extracts a parameter appearing on a route related to the case as a correction target of the determination rule. The search range storage unit 15 stores the search range of the correction target extracted by the correction target extraction unit 14. However, the search range storage unit 15 stores the GA
It is used to correct (generate) a judgment rule by a method, and may not be used when another method is used (see FIG. 1B). Finally, the correction target search unit 1
A search is performed within the correction target range by 6 to change the parameter, that is, to correct the determination rule.

Next, the operation of the decision rule correcting device having the above configuration will be described with reference to the flowchart. FIG.
5 is a flowchart showing a flow of a correction according to an embodiment of the judgment rule correction device.

As shown in FIG. 1, a data set having both input data and a classification class for identifying the data is read (step 31). This makes preparations for evaluating whether or not the judgment rule stored in the judgment rule storage unit 12 makes a correct inference with respect to the case stored in the case storage unit 11 of FIG. Here, the case S1 of FIG. 4 is read.

Next, the read case S1 is evaluated according to the judgment rules stored in the judgment rule storage unit 12, and a classification class corresponding to the read case is inferred (step 32). For example, the evaluation is performed using the judgment rule shown in FIG.
The inferred class for the case S1 in FIG. 4 is determined. However, “area”, “density”, and “width” described in the condition part of the determination rule in FIG.
A function area (f which takes ilen and parameters as arguments
ilen, Q1), density (filen, Q
2) and width (filen, Q3) are defined respectively. Further, the search range storage unit 15 stores a searchable range of the parameters shown in FIG. 6 and a current set value. The range has a “maximum value” and a “minimum value” for each parameter, and a “step size” within the range. When evaluating the case S1 in which the image file name is file11 and the class is class 1, Q1 = 4, Q
It has a value of 2 = 4, Q3 = 4. These parameters are substituted into the above function to obtain a feature amount as a solution. The feature amount is an attribute of a case represented by a numerical value. Assuming that the value SA1 shown in FIG. 7 is obtained as a result of substituting the parameters for the case S1, filen has the feature amounts of “area” 8, “density” 5, and “width” 1.

When inferring a case, first, evaluation is performed using the value of the feature amount “area” assigned to the highest-order branch node N1 in the decision rule of FIG. In case S1, "area"
Is 8, the condition of the branch connected to the left side of the branch node N1 is less than 10. Therefore, the case propagates to the left side of the branch node N1 and moves to the branch node N2. Next, evaluation is performed using the value of the feature amount “density” assigned to the branch node N2. In the case of FIG. 4, the feature amount of “density” is 5, so that the condition of 5 or more of the branch connected to the right side of the branch node N2 is satisfied, and the case propagates to the branch node N5. Finally, evaluation is performed using the value of the feature amount “width” assigned to the branch node N5. In the example of FIG. 4, since the feature amount of “width” is 1, the condition of the branch connected to the left side of the branch node N5 is less than 6, and the left side node, that is, the terminal node E4 to which “classification 2” is assigned.
Propagate to As a result, the case is inferred as class "Category 2". The hatched nodes in FIG. 8 indicate how the case S1 is inferred according to the determination rules.

Subsequently, the case S inferred in step 32
The class No. 1 is compared with the classification class stored in the case storage unit 11 before inference (step 33). If the two classes match, the decision rule does not need to be modified, and the algorithm ends. However, if they do not match, it is determined that the determination rule needs to be modified. In the case of the case S1, the class was inferred to be “class 2” in step 32, but “class 1” is assigned in the case storage unit 1 and does not match. Therefore, in this case, it is determined that the determination rule needs to be modified.

When modifying the judgment rule, first, a modification target is extracted (step 34). The extraction has the following three methods.

One is a method of extracting a correction target related to a case class. In other words, an inference route having the same inference result (class of the terminal node) as the class corresponding to the screen file name of the case in which an incorrect inference result is obtained is traced back from the terminal node to the highest branch node, and the parameters on the path Is extracted as a correction target. Specifically, case S1 in FIG.
In other words, the route related to the originally assigned class "class 1" is traced back. The “classification 1” is inferred at the two end nodes E1 and E6 by the determination rule of FIG. The path from the terminal node E1 to the uppermost branch node N1 is a branch connecting the branch node N4 and the terminal node E1, the branch node N4, the branch node N2 and the branch node N4.
, A branch node N2, a branch connecting the branch node N1 and the branch node N2, and a branch node N1. Parameters P5, Q3, P2, Q2, P1, Q
1 corresponds. In the case of the terminal node E6, the branch node N3
And a branch node N3, a branch node N3, a branch connecting the branch node N1 and the branch node N3, and a branch node N1, and the parameters P3, Q3, P1, and Q1 correspond to each other. Therefore, as a correction target related to the class, the parameter P
1, P2, P3, P5, Q1, Q2, and Q3 are extracted.
The nodes surrounded by oblique lines in FIG. 9 indicate how parameters are extracted.

The second is a method of extracting a correction target related to the evaluation of a case. That is, the parameters of the determination path from the highest branch node to the terminal node followed by the case determined to be the correction target are extracted as correction targets related to the evaluation of the case. In the case S1 of FIG. 4, as shown in FIG. 8, the branch node N1, the branch node N1, and the branch node N
2, the branch node N2, the branch between the branch node N2 and the branch node N5, the branch node N5, and the branch between the branch node N5 and the terminal node E4 to propagate to the terminal node E4. The parameters corresponding to these are Q1, P1, Q2, P2, Q
3, P7, which is extracted as a correction target related to the evaluation of the case.

The three methods are based on the premise that the correction targets extracted using the above two methods are further narrowed down. In other words, if the read case becomes an incorrect inference result, an inference route for obtaining a correct inference result (ending at the terminal node) for the case is assumed, and the feature amount of the case and the threshold value on the route are assumed. (Value used as a reference for branching). If the case deviates from the route due to the threshold, the threshold is corrected, that is, the parameters of the threshold are extracted as correction targets to be searched. On the other hand, if the feature value and the threshold value are on the inference route, they are not corrected. In the decision rule shown in FIG. 5, in the case of the case S1 of FIG. 4, the inference only needs to reach E1 or E6 when viewed from the terminal node.
4 or N3. Therefore, case S1
Is 1, the condition of the branch connecting the branch node N4 and the terminal node E1 is less than 4 (= P5), and the condition of the branch connecting the branch node N3 and the terminal node E6 is 2 (=
Since these values are less than P3), there is no need to change these values. That is, the parameters P5 and P3 and Q3 that changes the feature amount of “width” are not to be corrected. Looking further at the upper nodes, the characteristic amount of “density” in case S1 is 5, whereas the condition of the branch connecting branch node N2 and branch node N4 is less than 5 (= P2). P2 is to be corrected. Similarly, while the feature amount of the “area” in the case S1 is 8, the condition of the branch connecting the branch node N1 and the branch node N3 is 10 (= P1) or more, so the parameter P1 is also to be corrected. Further, even if the case reaches the branch node N5, a correct inference result cannot be obtained, and therefore, the parameter P7 is not to be corrected.

Therefore, the final correction targets are the parameters P1, P2, Q1, Q2 except for the parameters that do not need to be changed from the correction targets related to class and case evaluation.
It becomes Q2.

Further, the method considers the correction parameter extraction in the case where the judgment rule shown in FIG. 10 is stored in the judgment rule storage unit 12, for example.

Here, parameters Q4 = 2 and Q5 = 3 are given as parameters for evaluating the case S1 of FIG. 4, and the functions high (filen, Q4) and length
A variant case S2 (not shown) obtained by adding the "height" feature amount 4 and the "length" feature amount 3 by h (filen, Q5) is used. However, the classification class is the same as the case S1.

According to the determination rule shown in FIG. 10, the case S2 obtains the inference result of the terminal node E8, that is, "classification 2". However, the class “Category 1” is originally assigned to the case S2, and the determination rule needs to be modified. Similar to the extraction method in the determination rule of FIG. 5 described above, in the determination rule of FIG. 10, parameters P1, P2, P3, P5, P6, P8,
P9, P10, Q1, Q2, Q3, and Q4 are extracted. Further, parameters Q1, P1, Q2, P2, Q3, and P7 are extracted as correction targets related to the case evaluation. Finally, paying attention to the features, the end nodes E3, E6, and E10, which are assumed to be correct inferences, are branch nodes N6, N7,
The condition of the branch that branches from N3 and connects the branch node N6 and the terminal node E3 is 5 (= P8) for the feature amount 4 of “height”.
It is less than 8 (= P9). Therefore, the parameter P
8, P9 and Q4 that changes the feature amount of “height” are to be corrected. The condition of the branch connecting the branch node N7 and the terminal node E6 is less than 8 (= P10). However, since Q4 is to be corrected, the feature amount of “height” may change. Therefore, P10 is a correction target. "width"
Branching node N3 and terminal node E1
The condition of the branch connecting 0 is less than 2 (= P3), the branch node N4
Is less than 4 (= P5). Therefore, the parameters P3, P5, and Q3 are not to be corrected. Also, the branch node N4 and the branch node N7
Is more than 9 (= P6), but P6 is
Since it cannot be smaller than 5, it does not reach the branch node N7. Therefore, P6 is out of the target. Then, when tracing further higher nodes, the condition of the branch connecting the branch node N2 and the branch node N4 is 5 (=
Since it is less than P2), the parameters P2 and Q2 are to be corrected. Branch node N1 and branch node N on it
2, and the condition of the branch connecting the branch node N1 and the branch node N3 is less than 10 (= P1) on the branch node N2 side and 10 or more (= P1) or more on the branch node N3 side. Are to be corrected.

Even if the case reaches the branch node N5, a correct inference result cannot be obtained, so the parameter P7 is excluded from the correction target.

Therefore, in the case S2 shown in FIG. 10, if the parameters that do not need to be changed are excluded from the correction targets related to the class and the case evaluation, the parameters P1, P2, P8,
P9, P10, Q1, Q2, and Q4 are final correction targets.

Next, the search start position of the correction target range is determined (step 35). In other words, the correction target extraction unit 1
4 to obtain a searchable range of the correction target by the search range storage unit 15 that stores the search range of the correction target obtained in step 34, and further converts the correction target into a character string expressed in bits. Regarding the bit representation, according to the reference document "Genetic algorithm / Takeharu Yasui, Tomoharu Nagao / Shokodo / 1993", when using GA, what kind of symbol is used as a symbol representing a genotype is arbitrary. , Generally 0 and 1. Therefore,
In this embodiment, the parameters are represented by binary numbers using 0 and 1.

Considering the correction target obtained in the case S1 shown in FIG. 4, a range of the minimum value 4, the maximum value 11, and the step size 1 is obtained from the search range storage unit 15 for P1. At this time, since P1 can take eight kinds of values, all the possible values of P1 can be represented by a binary number by allocating 3 bits. That is, the minimum value 4 corresponds to the binary number 000, and the value of 000 also increases by 1 each time the minimum value increases by 1 according to the step size.
If it increases, all the values of P1 can be expressed. Considering the remaining parameters P2, Q1, and Q2 in the same manner, all the possible values from the minimum value to the maximum value can be expressed by assigning 3 bits, 4 bits, and 4 bits respectively. Looking at the entire correction target, one solution can be represented by 14 bits. FIG. 11 is a chromosome expression in which the parameter setting values of P1, P2, Q1, and Q2 in the determination rule of FIG.

When the chromosome that represents the current state of the correction target in bits and the chromosome that generates 0 or 1 using random numbers are combined to form a predetermined number of aggregates, the number of chromosomes is calculated. The search is started with the chromosome population of the zero generation.

FIG. 12 shows the chromosome group of the 0th generation described above.
It consists of 19 chromosomes generated with the value shown in 1 and a random number of 0 or 1. When the fitness of these chromosomes is calculated, for example, in the case of a chromosome expressing the current set values shown in FIG. 11, the cases T1 to T1 of FIG.
If T10 can be determined correctly, the case S shown in FIG.
Only 1 is a case where it cannot be determined correctly with the current set value, and 0.91 (= 10/11) is obtained as the fitness of the chromosome in FIG. On the other hand, for other chromosomes, the parameters corresponding to the chromosomes are determined first, and the values are replaced with the parameters of the determination rule in FIG. 5 to evaluate the cases of T1 to T10 and S1. After each such case is evaluated, the percentage of cases that can be evaluated correctly (the cases that can be correctly evaluated / all stored cases) is determined in the same way as the current setting value, and the fitness of each chromosome is calculated. And

Subsequently, it is determined whether or not the generation of the chromosome population has reached the preset generation, or whether the fitness of the chromosome giving the highest fitness is 1.0 (step 36). As a result, if any of the conditions is satisfied, the judgment rule is corrected (step 38). On the other hand, if neither condition is satisfied, the next search position is determined (step 37). That is, a new search point is arbitrarily created by applying a genetic operation such as crossover / mutation, and a chromosome population at an intermediate stage is generated. In addition, for each chromosome included in the chromosome population in the middle stage, whether or not the case and the judgment rule stored in the case storage unit 11 need to be modified using the judgment rule applying the parameter value corresponding to the chromosome The ratio at which the cases read for the test are correctly identified is determined as the fitness of each chromosome. Furthermore, a probability according to the degree of fitness is assigned to each chromosome, and a chromosome is selected from the chromosome population at an intermediate stage, thereby selecting a chromosome population of the next generation.

For example, as shown in FIG. 12, a first generation chromosome population can be generated from a 0th generation chromosome population by genetic manipulation as follows.

First, in the case of intersection, two chromosomes are extracted from the chromosome population of the 0th generation using random numbers. The cutting position of the extracted chromosome is determined using random numbers. Chromosomes are separated at the cut positions, and a part of the separated chromosomes is exchanged with each other to generate two new chromosomes. This intersection is repeated until there are no more chromosomes contained in the 0th generation, thereby generating the first generation intermediate stage 1. FIG. 13 illustrates a state in which the first generation middle stage 1 is generated by crossing from the 0th generation chromosome population.

Next, in the case of mutation, the first generation
Take out one chromosome from. For each bit of this chromosome, it is determined whether or not to carry out mutation, and the bit determined to be mutated is inverted, that is, converted into 0 → 1 or 1 → 0, and a new Generate chromosomes. This mutation is repeated until the chromosomes included in the first generation middle stage 1 disappear, thereby generating the first generation middle stage 2. FIG. 14 shows a state in which the first generation middle stage 2 is generated by mutation from the first generation middle stage.

In the case of selection, prior to the operation, the degree of fitness of the chromosomes included in stage 2 of the first generation is calculated. That is,
The cases T1 to T10 in FIG. 2 and the case S1 in FIG. 4 are evaluated by applying the values of the parameters corresponding to the chromosomes to the determination rule in FIG. The proportion of cases evaluated correctly at this time is defined as the chromosome fitness. Then, the fitness f of the n-th chromosome
By substituting itnessn into Equation 1, the selection probability probn of the n-th chromosome is obtained.

[0042]

(Equation 1) A chromosome to be selected is determined by a random number reflecting the value of the probn, and a first-generation chromosome population is generated. FIG. 15 shows a state in which a chromosome population of the first generation is generated by selection from the middle stage 2 of the first generation.

Finally, if either of the following conditions is satisfied: the generation of the chromosome population reaches the preset generation, or the chromosome fitness is 1.0, the best fitness among the searched chromosomes is obtained. The chromosome having a high chromosome is taken out and the parameter value for the chromosome is determined. Then, the determination rule stored in the determination rule storage unit 12 of FIG. 1A is updated (corrected) with this parameter value. For example, assume that the chromosome in FIG. 16 is extracted as the chromosome with the highest degree of matching. At this time, the chromosome in FIG. 16 has P1 = 11, P2 = 7,
Since Q1 = 5 and Q2 = 3, the decision rules stored in the decision rule storage unit 12 are modified from FIG. 5 to FIG.

As can be seen from the flow chart described above, according to the present invention, the part relevant to the determination of the case leading to the wrong inference is divided in three ways, namely for the wrong result, for the wrong case, and for the wrong case. By extracting the feature amount of the reason led to the above as a correction target and searching only the vicinity of the correction target, it is possible to correct the determination rule at high speed and partially. Note that the present determination rule correcting device is not limited to the present embodiment.

For example, the decision rule used in the present embodiment is a decision rule in the form of a decision tree. However, the reference document "Generation of fuzzy decision tree by inductive learning; Shigeaki Sakurai, Dai Araki
Author, Transactions of the Institute of Electrical Engineers of Japan, Vol. 113c, no. 74
88-494, 1993 "may be used.

Further, in this embodiment, GA is used as a search method for correcting the parameters of the decision rule, but the reference document "SIMULATED ANNEALI" is used.
NG; edited by Robert Azenc
ott, John Willy & Sons, In
c. ; 1992 "can be used to determine a corrected value of a parameter. The flowchart of FIG. 3 describes a processing procedure when GA is used as a search method.
When a method other than GA is used, the processing procedure is not limited to this.

Further, the chromosome representation of the parameter to be modified shown in step 35 of FIG. 3 by binary bits is not limited to the 14 bits used in this embodiment, but is generally limited to each modification. The number of digits of each bit is determined from the range (number) of values that can be taken by the target parameter, and a chromosome population is generated in units of the sum of the digits.

In the GA, the definition of the number of chromosome populations to be generated is not defined, so that the number is not limited to the number 20 generated in FIG.

In step 38, the judgment rule is modified by converting the value of the chromosome with the highest degree of conformity. However, by converting the values of a plurality of chromosomes with the highest degree of conformity, the judgment rule is modified. Can be generated, and an expert can select a judgment rule from among them.
In addition, various modifications may be made without departing from the technical spirit of the present invention.

[0050]

As described in detail above, according to the present invention,
In the problem of erroneous inference result in the problem of modifying the judgment rule to lead the given data to the correct inference result, extract the cause as the correction target, and set the periphery of the extracted correction target as the search space I do. Then, the set search space is corrected using a search technique.
Extracting the correction target and setting the search space small in this way eliminates useless operations in correcting the determination rule,
It is possible to quickly modify given data into a decision rule that leads to a correct inference result.

Further, a reduction in the time required for correction leads to a reduction in the effect on the operation of processing (for example, inspection of defective products) using the present apparatus.

[Brief description of the drawings]

FIG. 1A is a block diagram illustrating a configuration of a decision rule modifying device according to the present invention, and FIG. 1B is a block diagram illustrating a configuration of a decision rule modifying device according to an embodiment of the present invention.

FIG. 2 is a diagram showing a set of an image file name and a classification class stored in a case storage unit 11 of the present apparatus.

FIG. 3 is a flowchart showing a flow of processing of the present apparatus.

FIG. 4 is a diagram showing a set of an image file name and a classification class for which an inference result cannot be correctly determined according to the determination rule of FIG. 5;

FIG. 5 is a diagram showing a judgment rule stored in a judgment rule storage unit 12 of the present apparatus.

FIG. 6 is a diagram showing a searchable area of a parameter stored in a search range storage unit 15 of the present apparatus.

FIG. 7 shows the image file name and parameters Q1 to Q in FIG.
FIG. 5 is a diagram illustrating a feature amount of the case of FIG. 4 calculated by substituting a value of 3;

8 is a diagram illustrating a state for inferring the classification class of FIG. 4 by the determination rule of FIG. 5;

FIG. 9 is a diagram for explaining how the data in FIG. 4 is inferred by the determination rule in FIG. 3;

FIG. 10 is a diagram illustrating a correction target to be extracted.

FIG. 11 is a chromosome representation of parameters to be corrected.

FIG. 12 is a diagram showing a chromosome population of the 0th generation including the chromosome of FIG. 11;

FIG. 13 is a diagram illustrating a state in which a first generation halfway state 1 is generated by performing crossover on a 0th generation chromosome.

FIG. 14 is a diagram illustrating a state in which a first generation intermediate state 2 is generated by performing mutation on a chromosome population in the first generation intermediate state.

FIG. 15 is a diagram illustrating a state in which a first-generation chromosome population is generated by selecting a chromosome from the chromosome population in the first generation halfway state 2.

FIG. 16 is a diagram showing a chromosome having the highest degree of matching among the searched chromosomes.

FIG. 17 is a diagram illustrating a modified judgment rule.

[Explanation of symbols]

 11 Case storage unit 12 Judgment rule storage unit 13 Inference unit 14 Correction target extraction unit 15 Search range storage unit 16 Correction target search unit

Claims (6)

[Claims] 1. A means for extracting, from a data set having both data and a classification class for identifying the data, a part in which inference of a judgment rule described in the system is incorrect, parameter adjustment of the extracted judgment rule Means for correcting using an algorithm capable of performing the determination. 2. The apparatus according to claim 1, wherein said extracting means extracts a parameter on a decision rule corresponding to the data set from a data set which is derived from an incorrect inference result. And means for extracting a parameter on a decision rule corresponding to the classification class to which a data set derived from an incorrect inference result belongs, which is characterized by the following. 3. The extracting means according to claim 2, further comprising:
It is characterized in that it comprises means for extracting only parameters that need to be corrected on an inference route from which a data set led to an incorrect inference result can obtain a correct inference result. 4. A step of extracting, from a data set having both data and a classification class for identifying the data, a part in which the inference of the decision rules described in the system is incorrect, adjusting the extracted decision rules by parameters. Correcting using an algorithm that can perform the determination. 5. The method according to claim 4, wherein said extracting step is a step of extracting a parameter on a decision rule corresponding to the data set from a data set that leads to an incorrect inference result. And extracting a parameter on a decision rule corresponding to the classification class to which a data set derived from an incorrect inference result belongs. 6. The extracting step according to claim 5, further comprising: targeting the parameters extracted in claim 5,
The method further comprises a step of extracting only parameters that need to be corrected on an inference route from which a data set led to an incorrect inference result can obtain a correct inference result.