JP3212695B2 - Knowledge acquisition device for knowledge base system and knowledge correction device - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a knowledge acquisition system for a knowledge base system capable of acquiring knowledge from acquired data and correcting the knowledge in the fields of diagnosis, estimation, prediction, control, measurement, and the like. It relates to a knowledge correction device.

[0002]

2. Description of the Related Art Conventionally, a knowledge base system uses an IF (condition part) -THEN called a production rule.
(Execution unit) There is a knowledge base that stores knowledge in the form, and inferences and the like are performed by combining the knowledge of this knowledge base. Knowledge acquisition and modification of this knowledge base is important in determining the ability of a knowledge base system.

[0003] However, knowledge acquisition and knowledge correction of the knowledge base of this knowledge base system are performed by statistically analyzing the data by another means, even if sufficient data for knowledge acquisition is held. Because knowledge was acquired using the results, storing the knowledge in the knowledge base required a great deal of effort. In addition, knowledge modification in this case also requires a great deal of effort because it does not consider consistency with acquired knowledge.

[0004] For example, knowledge-based rules used in an expert system are obtained by acquiring expert knowledge from an expert, interpreting the knowledge, and creating a rule in a format suitable for the expert system to create a knowledge-based rule. There is a problem that a storing operation is required, and a large amount of time and labor is required to construct a sufficient expert system.

[0005]

As described above, in a conventional knowledge base system, acquisition of knowledge stored in the knowledge base and modification of the knowledge are also important factors that affect the ability of the knowledge base system. Nevertheless, there was a problem that a great deal of time and labor was required.

Therefore, the present invention provides a knowledge acquisition device of a knowledge base system and a knowledge correction device capable of eliminating such problems, automatically acquiring knowledge, and easily modifying knowledge. The purpose is to do.

[0007]

According to a first aspect of the present invention, there is provided a knowledge base system for acquiring knowledge stored in a knowledge base of a knowledge base system from a plurality of data groups having data corresponding to a plurality of attributes. In the knowledge acquisition apparatus, a data classification unit that performs a cluster analysis on a partial data group of the plurality of data groups having a specific attribute among the plurality of attributes and classifies the plurality of data groups into clusters, For each classified cluster, a condition part knowledge obtaining means for obtaining a logical expression of a cluster attribute condition obtained by a predetermined induction learning method as knowledge of a condition part of a rule, and for each cluster classified by the data classification means, Based on a plurality of partial data groups in the cluster including other attributes, an optimal numerical relational expression is determined by a predetermined data analysis method according to a rule execution unit. An execution unit knowledge acquisition unit for acquiring as knowledge, and a rule generation unit for generating the knowledge base rule from the knowledge for each cluster obtained from the condition unit knowledge acquisition unit and the execution unit knowledge acquisition unit. It is characterized by.

A second invention is based on a plurality of data groups having data corresponding to a plurality of attributes, and a partial data group of the plurality of data groups having a specific attribute among the plurality of attributes. Classifying the plurality of data groups into clusters by cluster analysis, and for each of the clusters classified by the data classifying means, a logical expression of a cluster attribute condition obtained by a predetermined induction learning method is used as a rule condition. For each of the clusters classified by the condition class knowledge obtaining means for obtaining as the knowledge of the part, and for each cluster classified by the data classification means, based on a plurality of partial data groups in the cluster including other attributes, an optimum data analysis method is performed by a predetermined data analysis method. Execution unit knowledge obtaining means for obtaining a numerical relational expression as knowledge of an execution unit of a rule; and a class obtained from the condition unit knowledge obtaining means and the execution unit knowledge obtaining means. Rule generation means for generating the rules of the knowledge base from the knowledge for each data, and the knowledge that reserves the use as knowledge of the knowledge base among the knowledge acquired by the condition part knowledge acquisition means as the sub-knowledge. In a knowledge correction device of a knowledge base system for storing knowledge in a knowledge base and correcting the knowledge of the knowledge base with respect to a new data group, the new data group is compared with knowledge of a condition part of a rule stored in the knowledge base. Knowledge determination means for determining whether or not there is something to be satisfied, and when there is no satisfaction by the knowledge determination means, the new data group is compared with the sub-knowledge stored in the knowledge base. Sub-knowledge determining means for determining whether or not there is something to be satisfied, and when the sub-knowledge determining means determines that there is something to be satisfied, A sub knowledge adding means for adding to the knowledge of the condition part of the rule to cancel the reservation of the sub-knowledge, when it is determined that there is to be satisfied by the knowledge judgment means or the sub-knowledge judging means,
An allowable range determining unit that performs a process based on the knowledge of the rule executing unit corresponding to the knowledge of the condition unit of the rule, and determines whether the processing output is within a predetermined allowable range; If determined to be within the range, it is satisfied by the partial knowledge correction means for storing the knowledge of the condition part of the rule and the knowledge of the corrected execution part of the rule as corrected knowledge in the knowledge base, and the sub-knowledge determination means. When it is determined that there is nothing, or when it is determined that the data is not within the allowable range by the allowable range determining unit, an instruction to acquire knowledge again from the plurality of data groups including the new data group is issued. And a knowledge correcting means.

[0009]

[0010]

According to a first aspect of the present invention, there is provided a knowledge acquiring apparatus for a knowledge base system for acquiring knowledge stored in a knowledge base of a knowledge base system from a plurality of data groups having data corresponding to a plurality of attributes. Classifying means for cluster-analyzing a partial data group of the plurality of data groups having a specific attribute among the plurality of attributes to classify the plurality of data groups into clusters; For each of the clusters classified by the means, a logical expression of a cluster attribute condition obtained by a predetermined induction learning method is acquired as knowledge of a condition part of a rule, and the execution unit knowledge acquiring means of the cluster unit knowledge acquiring means comprises: For each of the clusters classified by the classifying means, based on a plurality of partial data groups in the cluster including other attributes, a predetermined data analysis method is used. And a rule generating means of the cluster unit knowledge obtaining means, wherein the rule generating means of the cluster unit knowledge obtaining means obtains the knowledge for each cluster obtained from the condition part knowledge obtaining means and the execution part knowledge obtaining means. From the knowledge base.

According to a second aspect of the present invention, based on a plurality of data groups having data corresponding to a plurality of attributes, a partial data group of the plurality of data groups having a specific attribute among the plurality of attributes is provided. Classifying the plurality of data groups into clusters by cluster analysis, and for each of the clusters classified by the data classifying means, a logical expression of a cluster attribute condition obtained by a predetermined induction learning method is used as a rule condition. For each of the clusters classified by the condition class knowledge obtaining means for obtaining as the knowledge of the part, and for each cluster classified by the data classification means, based on a plurality of partial data groups in the cluster including other attributes, an optimum data analysis method is performed by a predetermined data analysis method. Execution unit knowledge obtaining means for obtaining a numerical relational expression as knowledge of an execution unit of a rule; and a class obtained from the condition unit knowledge obtaining means and the execution unit knowledge obtaining means. Rule generation means for generating the rules of the knowledge base from the knowledge for each data, and the knowledge that reserves the use as knowledge of the knowledge base among the knowledge acquired by the condition part knowledge acquisition means as the sub-knowledge. In a knowledge correction device of a knowledge base system for storing knowledge in a knowledge base and correcting knowledge of the knowledge base for a new data group, a knowledge judging means includes a condition part of the new data group and a rule stored in the knowledge base. It is determined whether there is a satisfactory one by comparing with the knowledge of the sub-knowledge determining means, if there is no satisfying by the knowledge determining means, stored in the new data group and the knowledge base The sub-knowledge adding unit determines whether or not there is any content that is satisfied by comparing the sub-knowledge with the sub-knowledge. If it is determined,
Cancel the reservation of the sub-knowledge and add it to the knowledge of the condition part of the rule, and if the allowable range determining means determines that there is something to be satisfied by the knowledge determining means or the sub-knowledge determining means, A process is performed based on the knowledge of the execution unit of the rule corresponding to the knowledge of the condition unit, and it is determined whether or not the processing output is within a predetermined allowable range. If it is determined, the knowledge of the condition part of the rule and the knowledge of the corrected execution part of the rule are stored in the knowledge base as modified knowledge, and the overall knowledge modifying means is satisfied by the sub-knowledge determining means. When it is determined that there is no data, or when it is determined that the data is not within the allowable range by the allowable range determining means, the plurality of data groups including the new data group are again An instruction to perform the Luo knowledge acquisition.

[0012]

[0013]

An embodiment of the present invention will be described below with reference to the drawings. FIG. 1 is a configuration block diagram of a knowledge base system according to one embodiment of the present invention. Here, a case where the present knowledge base system is applied to a case based reasoning system will be described. What is a case-based reasoning system?
This is to accumulate past problem-solving experiences as cases and to perform inferences that lead to conclusions by searching and correcting similar cases for new problems. Specifically, the present invention will be described in detail below by applying to a sales forecasting system. In other words, in the case of home appliances, retail stores, or chain stores such as convenience stores and restaurants, store past store openings, store similar stores, search for similar stores, and refer to information on similar stores. This is to predict the sales of the store that will open. Here, acquisition of a rule for predicting and calculating sales or modification of the rule will be described.

In FIG. 1, a knowledge base system in the sales forecasting system includes a knowledge acquisition device 1 for acquiring knowledge, a knowledge modification device 2 for modifying knowledge, a knowledge base 3 storing knowledge in IF-THEN format, A knowledge base control unit that controls the case base 4, the knowledge acquisition device 1, the knowledge correction device 2, and the case base 4 in which data serving as a source of knowledge is stored, and performs overall control of acquiring and modifying knowledge of the knowledge base. 5, an inference unit 8 for performing inference based on the knowledge of the knowledge base 3, an input unit 6, and an output unit 7. Further, the knowledge acquisition device 1 performs a cluster analysis on the data stored in the case base 4 by cluster analysis, and inductively learns the condition part of the rule for each of the clusters classified by the cluster analysis part 10 to perform the condition part of the rule. An inductive learning unit 11 for acquiring knowledge of the data, a data analyzing unit 12 for analyzing data in the cluster for each cluster classified by the cluster analyzing unit 10 and acquiring knowledge of a rule executing unit, and an inductive learning unit 11
And IF based on the knowledge acquired by the data analysis unit 12
A rule generation unit 13 that generates a rule in the THEN format.

Further, the knowledge correcting device 2 converts the abstract data of the newly input data and the condition part of all the rules stored in the knowledge base 3 into knowledge including the newly input data. A knowledge determining unit 20 that determines whether there is a rule that satisfies the condition. If there is no condition part of the rule that is satisfied according to the determination result of the knowledge determining unit 20, the sub-knowledge stored in the knowledge base 3; That is, a rule that satisfies the condition of knowledge acquisition of the inductive learning unit 11 but does not satisfy the knowledge application requirement used as knowledge in the knowledge base 3 and includes the condition of the rule stored in the knowledge base. Sub-knowledge determining unit 21 that determines whether or not there is one that satisfies the condition part of the rule. The sub-knowledge adding unit 22 to be added to the condition part of the rule, the execution unit of the rule corresponding to the condition part of the rule determined to be satisfied by the knowledge determination unit 20, and the condition part of the rule added by the sub-knowledge adding unit 22 The allowable range determining unit 23 that performs the process of the execution unit of the corresponding rule to determine whether the output value of the determined factor is within a predetermined allowable range, and satisfies the allowable range by the determination of the allowable range determining unit 23 In this case, the partial knowledge correction unit 24, the sub-knowledge determination unit 21, or the allowable range determination unit 23 that corrects only the rule execution unit by activating the data analysis unit 12
Is satisfied, the knowledge acquisition device 1 instructs the knowledge acquisition device 1 to acquire knowledge for all data including newly input data. Next, the knowledge acquisition device 1 will be described. FIG.
5 is an operation flowchart of the knowledge acquisition device 1. FIG.
, Data is input from the case base 4 (step 201). The input data is subjected to cluster analysis in the cluster analysis unit 10 (step 202). The data used for this cluster analysis is based on a common partial data group. Then, as a result, they are classified into groups, that is, clusters, according to a predetermined similarity (step 203).

Here, a specific cluster analysis of the sales prediction system will be described. First, there are two types of acquired data. One is survey data consisting of store information such as the sales floor area, the number of employees, and the number of parking lots, and regional information such as population density, family composition, and road conditions. One is abstract item data extracted from the area information. This abstract item data may be obtained by a questionnaire or by another extraction method. Alternatively, the local information itself may be used as the abstract item data. Note that the acquired data is obtained as data (case) for each store.

Here, the following eight items are taken as abstract item data, and each item takes a value as shown in parentheses. Note that continuous and discrete values can be considered as values in parentheses, but in the present embodiment, discrete values are used.・ Richness of residents' life (1.Low 2.Slightly low 3.Normal 4.
High 5. Very high ・ Surrounding area situation (1. Not good 2. Not very good)
3.Normal 4.Good 5.Very good ・ Traffic (1.Not good 2.Not very good)
3.Normal 4.Good 5.Very good ・ Customer type (1.Fixed 2.Both 3.Most) ・ Seasonal variability (1.Small 2.Medium 3.Large) ・ Diffusion of OA (1 Small 2.Medium lower 3.Medium 4.Medium upper 5.
Above) ・ Mania (1.Low 2.Normal 3.High) ・ Family (1.Low 2.Normal 3.Strong) Perform cluster analysis based on these eight abstract item data. Cluster analysis is a measure of similarity, that is, similarity, or vice versa.
It is intended to classify into clusters based on distance.

FIGS. 3 and 4 are diagrams showing specific abstract item data. 3 and 4 show data of 68 stores for the above eight abstract items.
It is stored in Case Base 4. The input to the case base 4 is input by the input unit 6.

FIGS. 5 and 6 show the weighted abstract item data. The data shown in FIGS. 5 and 6 are data obtained by weighting the data of FIGS. 3 and 4, and the values of the eight abstract item data are multiplied by the following weighting values. Are weighted. That is, the degree of inhabitants' richness × 0.58 Surrounding area conditions × 0.66 Traffic conditions × 0.85 Customer type × 0.42 Seasonal variability × 0.22 OA adoption degree × 0.39 Maniacity × 0. 27 family nature × 0.34 Therefore, in step 201, the weighted data shown in FIGS. 5 and 6 is input as the input data. For weighting, a dedicated means for weighting may be separately provided.

Now, cluster analysis is performed based on the weighted abstract item data. Although there are various cluster analysis methods, the calculation method in the present embodiment performs the distance calculation by the “standard Euclidean distance” and the cluster combination by the “shortest distance method”. Of course, cluster analysis may be performed by another calculation method.

FIG. 7 is a tree diagram showing the results of the cluster analysis. In the tree diagram of FIG. 7, the horizontal axis represents clusters in which the data numbers of FIGS. 5 and 6 are appropriately arranged, and the vertical axis represents the similarity between clusters. Here, in this embodiment, 16 clusters can be obtained by designating the similarity of the cluster analysis as “a”. That is, the cluster tree that intersects with the line La having the similarity “a” is 1
There are six places, and the data of 68 stores are classified into 16 clusters. Here, the designated similarity can be set arbitrarily. For example, by setting the similarity “b”, the line Lb of the similarity “b” crosses the tree of five clusters, and the five clusters are connected. Obtainable. It is not necessary to set the same similarity for all the data as indicated by the lines La and Lb in FIG. 7, and the clusters may be classified with different similarities. For example, a certain cluster is classified by the similarity crossing at the line La,
Further, a certain cluster may be classified based on the degree of similarity intersecting with the line Lb, and finally, all data may be classified into a certain cluster.

Here, it is assumed that the similarity is "a" and 16 clusters can be obtained. That is, cluster number 1 = {data number 1, 35} cluster number 2 = {data number 2, 3, 5, 6,
7, 8, 10, 12, 13, 18, 20, 26, 27,
28, 32, 34, 36, 37, 40, 41, 42, 4
4,46,47,52,54,60,61,62,6
6,68 {cluster number 3 = {data number 4, 38} cluster number 4 = {data number 9, 33, 43, 6}
7 cluster number 5 = {data number 11, 45} cluster number 6 = {data number 14, 48} cluster number 7 = {data number 15, 31, 49,
65｝ cluster number 8 = ｛data number 16,50｝ cluster number 9 = ｛data number 17,51,58｝ cluster number 10 = ｛data number 19,21,30,
53, 55, 64} cluster number 11 = {data number 22, 56} cluster number 12 = {data number 23} cluster number 13 = {data number 24} cluster number 14 = {data number 25, 59} cluster number 15 = {Data No. 29, 63} Cluster No. 16 = {Data No. 57} In this way, the cluster analysis unit 10 performs the cluster analysis, and classifies the input data into clusters based on the quantified data of the abstract items.

Next, returning to FIG. 2, the description of the flowchart will be continued. The induction learning unit 11 extracts one cluster classified by the cluster analysis unit 10 (step 204) and performs induction learning by a predetermined induction learning method (step 205). Then, a logical expression of the attribute condition obtained by a predetermined induction learning method is obtained as rule part knowledge of the rule (step 206).

On the other hand, the data analysis unit 12 also
A data group related to the acquisition of the execution unit of the rule is extracted from the cluster data extracted in step 4 (step 207),
A multiple regression analysis, which is one of the statistical methods, is performed to determine a numerical relationship between one determined factor and a plurality of determinants (step 208), and parameters of a calculation formula having one numerical relationship are calculated. Acquisition (step 209). And
A calculation formula including this parameter is acquired as the execution part knowledge of the rule (step 210).

First, the operation of the induction learning unit 11 will be described in detail based on a specific example of the sales prediction system. In the present embodiment, the inductive learning unit 11 uses “ID
3 "is adopted. This “ID3” is one of the teaching-type induction algorithms. Here, for example, in the case of performing inductive learning using “ID3” for the cluster number “2”, inductive learning is performed using data of 32 stores belonging to the cluster number “2” as a positive example and data of other stores as a negative example. .

FIG. 8 is a diagram showing an inductive learning result of the inductive learning unit 11. In FIG. 8, eight conditions S1 to S8 that are successful by the inductive learning method “ID3” are obtained. Here, the success by the inductive learning method “ID3” means that the data of 32 stores belonging to the cluster number “2” is a positive example, and the data of the other stores is a negative example by the inductive learning method “ID3”. Successful acquisition means failure,
This refers to a case where acquisition of the condition part knowledge has failed due to the inductive learning method “ID3”. In this case, if the number of data is “1”, the number of data is small, so that it is judged as an exception because the number of data is small and stored in the knowledge base 3 as acquired knowledge. It will not be used as. This knowledge is used when modifying or adding knowledge of the knowledge base 3 described later. This acquired knowledge of the exception is hereinafter referred to as “sub-knowledge” or “sub-condition”.

Therefore, the sub-condition S in which the number of data is one is
1, S3 and S8 are regarded as exceptions, and the knowledge of the condition part of the rule obtained from the cluster number "2" is as follows. That is: ・ Richness of residents 'life (somewhat low) and family (normal) or ・ Richness of residents' life (normal) and traffic conditions ｛(ordinary) OR (good) OR (very good) 又 は or High) and geeky (low).

Next, the specific operation of the data analyzer 12 will be described. The data analysis unit 12 performs multiple regression analysis for each cluster in order to acquire knowledge of the rule execution unit using multiple regression analysis. At this time, the variables use survey data likely to affect sales. Note that it is also possible to pick up data items that have an effect in advance.

Specifically, the following can be considered. In other words, the target variable: the sales amount of the new store Description variable: the sales floor area ratio (the new store sales floor area / similar store sales floor area): the number of employees (the number of new store employees / the number of similar store employees): parking lot accommodation The ratio for the number of vehicles (the number of parking lots of new stores / the number of parking lots of similar stores) is calculated, and the parameters for each variable are calculated. The parameter is a weight value, and a regression coefficient is obtained.
That is, the parameters a, b, c, and d in Expression (1) are obtained. Ps = {(Us / Ur) · a + (Js / Jr) · b + (Cs / Cr) · c + d} · Pr (1) Here, the meanings of the symbols in the expression (1) are as follows. Ps: New store sales Pr: Similar store sales Us: New store sales area Ur: Similar store sales area Js: New store employees Jr: Similar store employees Cs: New store parking lot accommodation Number of vehicles Cr: Number of parking lots accommodated in similar stores a: Regression coefficient of sales area ratio b: Regression coefficient of ratio of number of employees c: Regression coefficient of parking lot ratio d: Constant Although the data ratio is used, a calculation formula such as Expression (1-1) in which the data is substituted as it is without using the data as a pair may be used. In addition, a ', b', c ', and d' in Formula (1-1) are regression coefficients. In this embodiment, the data is treated as a pair in order to take into account the effect of the data of similar stores, that is, the influence of the data of similar stores. Ps = Ur.a '+ Jr.b' + Cr.c '+ d' (1-1) Next, parameters a, b, c, and d are obtained. In this case, since an initial rule is acquired, multiple regression analysis is performed by picking up one case and assuming a new store. For example, the case of data No. 2 is the most similar to the case of data No. 36 from the result of the cluster analysis.
And the data of the similar store is obtained by substituting the data of the data number 36 into the equation (1). Here, the data of the data number 2 as a new store and the data of the data number 36 as a similar store have the following values. Ps = 90 million yen Pr = 79 million yen Us = 40 square meters Ur = 30 square meters Js = 2 persons Jr = 1 person Cs = 10 cars Cr = 10 Similarly, the value of another data pair is expressed by the equation (1). When the multiple regression analysis is performed by substituting into That is, a = 0.0636, b = 0.236, and c = 0.003
6, d = 0.6984, and the following equation (2)
Will be obtained. Ps = {(Us / Ur) · 0.0636 + (Js / Jr) · 0.2236 + (Cs / Cr) · 0.0036 + 0.6984} · Pr (2) FIG. 9 shows a multiple regression analysis of cluster number 2. It is a figure which shows the difference | error of actual sales and expected sales based on a result. In FIG. 9, from the left, the data number of the case, the actual sales amount of the case,
The table shows the expected sales, the error between the actual sales and the expected sales, and the data numbers of the cases picked up as similar stores. In FIG. 9, it can be seen that the error is small overall, and a reasonable sales forecast can be made. In this way, other clusters are similarly processed.

Returning to FIG. 2, the description of the flowchart will be continued. For one cluster, the knowledge of the condition part of the rule acquired by the inductive learning unit 11 and the knowledge of the execution unit of the rule acquired by the data analysis unit 12 are summarized as a rule by the knowledge generation unit 13, and the knowledge base 3
(Step 211). If there are still clusters, the process proceeds to step 204 to acquire knowledge for each cluster, and the process is performed until all clusters are exhausted (step 212).

For example, acquired by cluster number 2,
The knowledge stored in the knowledge base 3 is summarized as follows. IF ・ Resident richness (slightly low) and family (normal) or ・ Resident richness (normal) and traffic conditions (normal) OR (good) OR (very good)｝ or ・ Resident richness (high) ) And mania (low) THEN Ps = {(Us / Ur) .0.0636+ (Js / Jr) .0.2236+ (Cs / Cr) .0.0036 + 0.6984} .Pr Rules are obtained in units and stored in the knowledge base 13.

Next, the modification of the rules obtained as described above will be described. After a rule is acquired, it may be necessary to modify the rule for new data. Hereinafter, a description will be given of a modification of an existing rule for predicting and calculating the sales amount of a store to be newly opened.

FIG. 10 is a flowchart of the operation of the knowledge correction device 2. In FIG. 10, first, new data input at the input unit 6 (step 301) is transmitted to the knowledge determination unit 20 of the knowledge correction device 2 via the knowledge base control unit 5.
, The knowledge determination unit 20 checks the new data against all of the condition parts of the rule of the knowledge base 3 (step 302), and determines whether the matching of the condition part of the rule is successful (step 302). Step 303). If the new data fails to match the condition part of the rule of the knowledge base 3, the sub-knowledge determining unit 21 checks the new data against the sub-knowledge stored in the knowledge base (step 30).
4), it is determined whether the collation is successful (step 30)
5). If the matching is successful in step 305, the sub-knowledge adding unit 22 adds the successfully matched sub-knowledge to the condition part of the rule (step 306).

On the other hand, if the matching of the condition part of the rule is successful in step 303, or if the sub-knowledge is added to the condition part of the rule in step 306, the allowable range determination unit 23 determines whether the matching is successful or not. The processing of the execution unit corresponding to the condition part of the rule to which the knowledge is added is performed, and it is determined whether or not the output value as the processing result satisfies a preset allowable range (step 30).
7). If the allowable range is satisfied, it is further determined whether or not to modify the execution part of the rule (step 3).
08). To correct the rule execution unit, the data analysis unit 12 first performs a multiple regression analysis with new data (step 309), obtains the parameters of the regression equation (step 310), Acquire knowledge of the rule execution unit (step 311). That is,
Only the rule execution unit is corrected by the data analysis unit 12 in the knowledge acquisition device 1. Then, the partial knowledge correction unit 24
Combines the condition part of the rule and the corrected rule execution part corresponding to the rule condition part (step 312).

On the other hand, if it is determined in step 308 that the execution part of the rule is not to be modified, the data including the fact that the tolerance is satisfied is stored in the case base 4 (step 313). The data stored in the case base 4 is then subjected to batch processing by the knowledge base control unit 5 to correct the rule execution unit by activating the data analysis unit 12.

If the allowable range is not satisfied in step 307, or if the matching of the condition part of the rule including the sub-knowledge fails in step 305, the entire rule stored in the knowledge base is corrected. The overall knowledge modification instructing unit 25 determines whether or not to perform the modification (step 314).
(Step 315).
That is, the process from cluster analysis of all data to knowledge storage in the knowledge base is performed. On the other hand, when the correction process is not performed, the data is stored in the case base 4 as in step 313 (step 316). That is, Case Base 4
Then, the entire knowledge base 3 is corrected by batch processing by the knowledge base control unit 5 for the data stored in the knowledge base 3.

Next, a specific rule modification of the sales prediction system will be described. As described above, there are three major cases for modifying the rule. First, there is a case where the condition part of the applicable rule does not exist in the knowledge base even in consideration of the sub-knowledge. In this case, the knowledge acquisition device 1 performs all the rules by performing a cluster analysis.
Second, there is a case where the condition part of the applicable rule exists in the knowledge base, but the processing result of the rule execution part does not satisfy the allowable range. Third, there is a case where the condition part of the applicable rule exists in the knowledge base and the processing result of the rule execution part satisfies an allowable range. In this case, the execution of the rule corresponding to the condition part of the rule is performed. Only data analysis is performed to correct parts.

The following description will focus on the third case in which the matching of the condition part of the rule fails but the matching of the condition part of the rule in consideration of the sub-knowledge is satisfied. First, when the weighted new data is input to the knowledge determination unit 20 of the knowledge correction device 2, the knowledge determination unit 20
It matches with all of the condition parts of the rules of the knowledge base 3. The collation between the new data and the condition part of a plurality of rules is repeated, and it is determined whether the collation of the condition part of the rule is successful. If the matching between the new data and the condition part of the rule of the knowledge base 3 fails, the sub-knowledge determining part 21 compares the condition part of the rule as the sub-knowledge stored in the knowledge base with the new data. For example, FIG. 8 shows the result of acquiring knowledge using inductive learning as described above, and it can be seen that eight pieces of knowledge have been acquired. However, considering the number of cases when acquiring the condition part of the rule,
If the number of data is one, it is regarded as an exception and becomes sub-knowledge. This is because a single case cannot be regarded as a typical rule at this stage.

Here, the sub-condition matching is repeated, and if the matching of all the sub-conditions fails, the data does not belong to any condition. Therefore, the cluster analysis is started again. The determination of whether or not to start over from the cluster analysis is arbitrary. For example, the timing may be set on a computer, and the processing may be automatically started when the number of data reaches a predetermined number, or Processing is performed by the user's specification. Here, the processing is performed according to the user's specification, and when processing the data according to the user's specification, the processing is performed again from the cluster analysis. When data processing is not performed, the data is stored in the case base and the processing is performed when specified. If the matching of the data sub-condition is successful, the satisfied sub-condition is added to the knowledge of the condition part of the rule. For example, when the new data is as follows: ・ Richness of inhabitants' lives (2. Slightly low) ・ Surrounding area (3.Normal) ・ Traffic (3.Normal) ・ Customer type (2.Both) -Seasonal variability (1. Small)-OA penetration (2. Middle and lower)-Mania (1. Low)-Family (3. Strong), if already cluster number 2 in Fig. 8 Although it does not satisfy the knowledge of the condition part of the acquired rules, it satisfies the sub-condition of "richness of inhabitants' lives (somewhat low) and family (strong)". Therefore, the knowledge of the condition part of the rule in FIG. 8 is changed as follows. That is: ・ Richness of resident life (low) and family (normal)… sub-condition ・ Richness of resident life (slightly low) and family (normal)… Rule condition ・ Richness of resident life (slightly low) and family ( Strong)… Rule conditions ・ Richness of resident life (normal) and traffic conditions (normal)… Rule conditions ・ Richness of resident lives (normal) and traffic conditions (good)… Rule conditions ・ Rich condition of residents and richness (normal) and traffic conditions (Very good)… Rule condition ・ Richness of resident life (high) and mania (low)… Rule condition ・ Richness of resident life (high) and mania (normal) and spread of OA (medium)… sub Condition As a result, the sub-condition "richness of residents' life (slightly low) and family (strong)" is changed to the rule condition and added to the condition part of the rule. That.

Next, when the new data satisfies the condition part of the rule, when this new data is substituted into the calculation formula of the execution part of the rule, it is determined whether the calculated output value satisfies the allowable range. to decide. Here, the allowable range is a value in which an output value includes an error. Although it is possible to set an allowable range only for the value obtained by substituting into the calculation formula, that is, for the predicted value, it is possible to make a judgment. However, in this embodiment, since multiple regression analysis is used, Therefore, a confidence interval of the regression coefficient, which is a parameter of the calculation formula, is provided to determine whether or not the tolerance is satisfied. This can be determined according to the probability value of the correlation between the determinant data and each determinant. here,
The probability value is a reliability derived from a data analysis result. That is, at the time of making a prediction, the likelihood of the prediction hitting, that is, the significance level is expressed by “%”.

For example, with respect to the determined factor "sales ratio", the reliability of the determining factors "sales floor area ratio", "parking lot accommodation ratio" and "employee ratio" is 99% in order. 90
% And 90% or less, the width of the range can be determined according to the reliability. This is because those with strong correlations have a large influence and the range of data is narrow,
Conversely, data having a weak correlation can have a wider range of data than data having a strong correlation because the influence is small. It is also important to limit the maximum and minimum values of determinants (generally called explanatory variables in multiple regression analysis). That is, whether the input determinant data falls between the maximum value and the minimum value of the determinant data when multiple regression analysis is performed using past data. If the input data does not fall within the limit value, it becomes a meaningless predicted value even if it is applied to a calculation formula for determining the determined factor.

If the new data does not satisfy the allowable range, the cluster analysis is started again.
This processing is performed in the same manner as when the sub-condition is not satisfied.

If the data satisfies the allowable range, the input data satisfies the existing rules, so there is no problem. However, a multiple regression analysis can be performed in order to correct the stricter rules. The timing for making this correction is determined based on whether or not the execution part of the rule is to be corrected. When data processing is not performed, this data is stored in the case base, and processing is performed when specified. That is, a batch process is performed. When processing the data, add the stored data, perform multiple regression analysis in the same way as when acquiring the rules, obtain the parameters of the calculation formula, obtain the execution unit for each cluster, Modify it as a new rule execution part and compile it into rules.

For example, the existing rule of cluster number 2 for which the initial rule has been acquired by the knowledge acquiring device 1
Get the modified result as follows: -The sub-conditions that have been successfully matched are added to the condition part of the rule.-The results of the new multiple regression analysis: a = 0.0642, b = 0.2228, c = 0.002
5, d = 0.7001 Then, the existing rule is modified as follows. That is, IF ・ Residents 'richness (slightly low) and family characteristics {(normal) OR (strong)} or ・ Residents' richness (normal) and traffic conditions ｛(normal) OR (good) OR (very good)｝ Or ・ Richness of residents' life rich (high) and mania (low) THEN Ps = {(Us / Ur) · 0.0642 + (Js / Jr) · 0.2228 + (Cs / Cr) · 0.0025 + 0.7001}・ It becomes Pr.

In this manner, the rules in the IF-THEN format can be obtained and corrected.

As described above, a huge amount of data is classified into optimal clusters by the cluster classification means as the first step, and a rule in the IF-THEN format is obtained for each classified cluster as the second step. As a result, knowledge acquisition can be performed automatically, and the correction of the knowledge can be performed easily while considering the consistency with the acquired cluster unit knowledge. be able to. That is, since the acquired knowledge is classified for each cluster, it can be easily modified, added, verified, and variously used later. In addition, knowledge that is difficult to obtain from a specialist can be automatically obtained.

[0047]

As described above, the present invention relates to a knowledge base system for acquiring knowledge stored in a knowledge base of a knowledge base system from a plurality of data groups having data corresponding to a plurality of attributes. In the knowledge acquisition device, the data classification unit performs a cluster analysis on the partial data groups of the plurality of data groups including the specified first attribute group among the plurality of attributes to classify the plurality of data groups into clusters. Then, the condition part knowledge obtaining means obtains, as knowledge of the condition part of the rule, a logical expression of a cluster attribute condition obtained by a predetermined induction learning method for each of the clusters classified by the data classification means, and obtains execution part knowledge. Means for each of the clusters classified by the data classification means, based on a plurality of partial data groups in the clusters, each of which includes the specified second attribute group, An optimal numerical relational expression is acquired as knowledge of an execution unit of a rule by a data analysis method, and a rule generation unit uses the knowledge for each cluster obtained from the condition unit knowledge acquisition unit and the execution unit knowledge acquisition unit to obtain the above-mentioned knowledge. Generate knowledge base rules. If the applicable condition part knowledge does not exist in the knowledge base even in consideration of the sub-knowledge, the data classification is performed by the data classification means, and all the knowledge including the new data is corrected. Applicable conditional knowledge exists in the knowledge base,
Even when the processing result by the execution unit knowledge does not satisfy the allowable range, the data classification unit performs data classification and corrects all knowledge from all data including new data. If the applicable condition part knowledge exists in the knowledge base and the processing result by the execution part knowledge satisfies the allowable range, only the correction of the execution part knowledge corresponding to the condition part knowledge is performed by the execution part knowledge acquisition means. I have to. Therefore, according to the present invention, knowledge acquisition of the knowledge base can be automatically performed, and the correction and update of the knowledge can be easily performed because the acquired knowledge is divided into clusters. This has the advantage that the time and effort for knowledge acquisition and its modification can be reduced.

Further, knowledge can be objectively and easily acquired and corrected from a plurality of data groups without relying on all the specialists, so that the acquired knowledge base knowledge can be added, modified, verified, etc. It has the advantage that it can be utilized.

Further, the acquired knowledge has an advantage that it is possible to acquire in advance knowledge that is difficult to acquire from an expert.

[Brief description of the drawings]

FIG. 1 is a configuration block diagram of a knowledge base system according to an embodiment of the present invention.

FIG. 2 is an operation flowchart of the knowledge acquisition device 1.

FIG. 3 is a diagram showing specific abstract item data.

FIG. 4 is a diagram showing specific abstract item data.

FIG. 5 is a diagram showing weighted abstract item data.

FIG. 6 is a diagram showing weighted abstract item data.

FIG. 7 is a dendrogram showing the results of cluster analysis.

FIG. 8 is a diagram showing an induction learning result of the induction learning unit 11;

FIG. 9 is a diagram showing an error between actual sales and expected sales based on the result of the multiple regression analysis of cluster number 2;

FIG. 10 is an operation flowchart of the knowledge correction device 2.

[Explanation of symbols]

 REFERENCE SIGNS LIST 1 knowledge acquisition device 2 knowledge correction device 3 knowledge base 4 case base 5 knowledge base control unit 6 input unit 7 output unit 8 inference unit 10 cluster analysis unit 11 inductive learning unit 12 data analysis unit 13 knowledge generation unit 20 knowledge determination unit 21 sub Knowledge determination unit 22 Sub-knowledge addition unit 23 Tolerance range determination unit 24 Partial knowledge correction unit 25 Overall knowledge correction instruction unit

Continuation of the front page (56) References JP-A-2-59931 (JP, A) JP-A-3-224054 (JP, A) JP-A-1-259488 (JP, A) JP-A-61-72354 (JP) , A) JP-A-4-596 (JP, A) Mathematical Sciences, Vol. 29, No. 3. Science Co., Ltd., 1991, pp. 27-31 (Patent Office CSDB Document Number: CSN W199900287003) (58) Fields investigated (Int. Cl. ⁷ , DB name) G06F 9/44 G06F 17/60 G06F 19/00 G06F 17/30 JICST file (JOIS) CSDB (Japan Patent Office)

Claims (2)

(57) [Claims] 1. A knowledge acquiring apparatus for a knowledge base system for acquiring knowledge stored in a knowledge base of a knowledge base system from a plurality of data groups having data corresponding to a plurality of attributes, Data classification means for performing a cluster analysis on the partial data groups of the plurality of data groups having the specific attribute, and classifying the plurality of data groups into clusters; for each cluster classified by the data classification means,
Condition part knowledge acquiring means for acquiring a logical expression of a cluster attribute condition obtained by a predetermined induction learning method as knowledge of a condition part of a rule, and for each cluster classified by the data classification means,
Executing unit knowledge obtaining means for obtaining an optimal numerical relational expression as knowledge of an executing unit of a rule by a predetermined data analysis method based on a plurality of partial data groups in the cluster including other attributes; A knowledge acquisition device for a knowledge base system, comprising: rule generation means for generating the rule of the knowledge base from the knowledge for each cluster obtained from the acquisition means and the execution part knowledge acquisition means. 2. A cluster analysis of a partial data group of the plurality of data groups having a specific attribute among the plurality of attributes based on a plurality of data groups having data corresponding to the plurality of attributes. Data classification means for classifying the plurality of data groups into clusters, and for each cluster classified by the data classification means,
Condition part knowledge acquiring means for acquiring a logical expression of a cluster attribute condition obtained by a predetermined induction learning method as knowledge of a condition part of a rule, and for each cluster classified by the data classification means,
Executing unit knowledge obtaining means for obtaining an optimal numerical relational expression as knowledge of an executing unit of a rule by a predetermined data analysis method based on a plurality of partial data groups in the cluster including other attributes; And a rule generating means for generating the rule of the knowledge base from the knowledge for each cluster obtained from the obtaining means and the execution part knowledge obtaining means, wherein the knowledge base is obtained from the knowledge obtained by the condition part knowledge obtaining means. A knowledge correcting device of a knowledge base system for storing knowledge that reserves use as knowledge in the knowledge base as sub-knowledge and performing knowledge correction of the knowledge base for a new data group, wherein the new data group and the knowledge base Knowledge determining means for comparing the knowledge of the condition part of the rule stored in A sub-knowledge determining unit that compares the new data group with the sub-knowledge stored in the knowledge base to determine whether there is a satisfied one; A sub-knowledge adding unit for canceling the reservation of the sub-knowledge and adding it to the knowledge of the condition part of the rule when the sub-knowledge determining unit determines that there is something to be satisfied; the knowledge determining unit or the sub-knowledge determining unit If it is determined that there is something that satisfies the rule, processing is performed based on the knowledge of the rule execution unit corresponding to the knowledge of the condition part of the rule, and a determination is made as to whether the processing output is within a predetermined allowable range. When the range is determined to be within the allowable range by the range determining means, the knowledge of the condition part of the rule and the knowledge of the corrected execution part of the rule are corrected. Partial knowledge correcting means to be stored in the knowledge base as described above, and if the sub-knowledge determining means determines that there is no satisfaction, or if the allowable range determining means determines that it is not within the allowable range, A knowledge correction device for a knowledge base system, comprising: an overall knowledge correction unit for instructing to acquire knowledge from the plurality of data groups including the new data group.