JPH01187637A - Rule acquisition system - Google Patents

Abstract

PURPOSE:To facilitate the derivation of a correct rule even in a case that an example contains an error by pointing out and checking the possibility that the inputted example is erroneous and excluding the error in the case that it is erroneous. CONSTITUTION:An error deciding part 107 compares the example which was inputted hitherto and stored in an example information storage part 104, with the example which is inputted now and stored in an example buffer 106, and decides the possibility that the example now inputted is erroneous. In the case that the possibility that the example inputted now is erroneous exists, an error inquiring part 108 displays it on an input/output terminal 102, and confirms again the decision whether the example is truly erroneous or not to a person who gave the example. When it is erroneous, it comes possible to make the example not to be inputted to a recursive rule acquiring means, and only the correct example is inputted to the recursive rule acquiring means. Thus, even in the case that the error is contained in the example given at first, the correct rule can be easily derived.

Description

DETAILED DESCRIPTION OF THE INVENTION [Field of Industrial Application] The present invention relates to a rule-based system, and particularly to a rule acquisition system suitable for automatically generating rules from examples.

[Conventional technology]

A rule-based system is an expert system.

Describing and changing the logic of decision-making support systems, flexible O1 to system control systems, etc.
It is widely used as an effective system architecture because it can be easily expanded.

Here, in conventional rule-based systems, rules are created manually, so they are large-scale.

For complex targets, it takes a lot of man-hours to create a large number of rules without fail.

To address this problem, Machine Learning An Artificial Intelligence Approach Volume 1 (1983) pp. 41-8
1 page (Machine Learning AnArt
ificj, al Intelliga mouth ce A
pproach, VoQ, 1° (1983
), a method of inductive rule acquisition has been proposed, as discussed in pp. 41-81). According to this method,
By simply providing a few examples consisting of pairs of specific target states and conclusions to be executed at that time, general rules with conditions and conclusions that can be applied to many target states can be automatically created without contradiction.

[Problem to be solved by the invention]

The above-mentioned conventional technology does not take into consideration the case where the example provided is incorrect, and there is a problem in that if an incorrect example is provided, the correct rule cannot be derived. That is,
In the conventional technology, the contents of the rule conditions are changed to general ones (expressions that satisfy many conditions) so that all the target states of a given example can be satisfied, but it is also possible to change the contents of the rule conditions to general ones (expressions that satisfy many conditions). If a generalization is made that satisfies the target state, and when actually using the derived rule, the rule may be applied even in a state where the conclusion of the rule should not be executed, resulting in unexpected failures. arise.

An object of the present invention is to provide a rule acquisition system that can derive correct rules even when examples include errors.

[Means to solve the problem]

In general, examples are given as concrete contents based on logic at the policy level, such as ``If there is something to lift and there is a place to lower it, lift it up'' or ``If the object is stable, turn on the normal lamp.''

From this, examples that follow a certain policy are
Things that are similar in some respects are expected to be different in other ways.

In the present invention, paying attention to this point, an error example extraction means is provided for extracting an example with a target state that has a large difference from other examples as an example with a high possibility of being incorrect among examples with the same conclusion, The above object is achieved by configuring the conventional inductive rule acquisition means to check the example before inputting it.

[Effect]

That is, since the error example extracting means points out an example that has a possibility of error, the person who gave the example confirms whether the example is really wrong or not, and if it is wrong, the example is removed. Since it is possible to avoid inputting information to the inductive rule acquisition means, and only correct examples are input to the inductive rule acquisition means, it is easy to derive the correct rule even if there is an error in the initially given example. be able to.

〔Example〕

An embodiment of the present invention will be described below with reference to the drawings.

The block configuration of this embodiment is shown in FIG.

The rule acquisition unit 101 acquires rules into the rule storage unit 103 based on the example input from the input/output terminal 102. The given example is stored in the example information storage unit 104.

The rule acquisition unit 101 further includes an example input unit 105, an example buffer 106, an error determination unit 107, and an error inquiry unit 1.
08. It is composed of a rule generation section 109.

The example input unit 105 receives examples from the input/output terminal 102 and stores them in the example buffer 106 .

The error determination unit 107 compares the examples input so far stored in the example information storage unit 104 with the examples that have been subjected to intervention and are stored in the example buffer 106, and determines whether the examples to which intervention has been applied are erroneous. ).

If there is a possibility of an error in the input example, the error inquiry unit 108 displays it on the input/output terminal 1.0243 and receives the user's decision.

The rule generation unit 109 creates a generalization rule for the example in the example buffer 106.

In the above description, an example is a character example that describes the correspondence between a system state that occurs in an actual object and the action that should be taken when the system state occurs.

An example showing specific control instructions in a certain workshop is shown in FIG. The system status is written on the left side of the arrow, and the action to be taken (hereinafter referred to as conclusion) is written on the right side. The system state is further expressed as a combination of states of individual elements of the system. The status and conclusion of each element are described in character strings. Here, individual characters and strings are shown in parentheses 11 (I+
, 11) Use I+ to distinguish. Further, parameters may be written in the character string by enclosing them in square brackets "['11]I+."

A parameter is a part of a string whose contents change (other parts do not change).

In the example shown in Fig. 2, for example, there is a workpiece of type A on the carry-in line, and there is nothing on the carry-out line. The type and validity of existence change. In addition, "mon" may be written in the parameter part,
this is.

Depending on the system state, the parameter may have no meaning (if there is no work on the light, the type has no meaning), and in such cases, there is no specific value that applies to that parameter (it is empty). This clearly indicates the following.

On the other hand, a rule is a character string that describes the correspondence between conditions that the target system state should satisfy and actions that should be taken when the conditions are satisfied.

Generally, there are multiple system states that satisfy the rule conditions.

FIG. 4 shows an example of rules representing control logic in the same workshop as the second example. The condition to be satisfied is on the left side of the arrow, and the position P is on the right side! , (conclusion) is written here.

Rule conditions are expressed as a combination of several individual conditions. The conditions of the rule are said to be satisfied when all the individual conditions are satisfied. The descriptions of individual conditions and conclusions are the same as in the example, but the rules also include an X in the parameter section.

Variables such as Y, or concrete values separated by commas II and I+ ((A, F3) in the figure, etc.) can be written. When a variable is described, it indicates that the condition is satisfied no matter what specific value is applied to that part.

Further, when specific values are described separated by commas ",", it indicates that the condition is satisfied if one of the values is satisfied.

Next, each storage unit and buffer will be explained.

The contents of the example information storage section 104 are shown in FIG.

This storage unit includes a plurality of example tables 301. It consists of an error determination threshold area 302.

Each example table 301 further includes one conclusion area 3011. It has multiple system status areas 3012. All examples having the same conclusion are stored in one example table 301. The conclusion of each example is in the conclusion area 30.
11. Also.

Each system status is shown in the system status area 3o1.
2 respectively. Error determination threshold area 302
The numerical value of the threshold value used for error determination is stored in .

The contents of the rule storage section 103 are shown in FIG.

This storage unit consists of a plurality of rule tables 501.

Each rule table 501 further includes one conclusion area 50L1. It has a plurality of condition areas 5012. All rules having the same conclusion are stored in one rule table 501. The conclusion of each rule is stored in a conclusion area 5011. Further, each condition is stored in a single condition area 5012.

An example buffer 106 is shown in FIG. Example buffer 1
06 is the conclusion area 1061. System status area 10
62, each of which stores the actual conclusion input from the input/output terminal 102 and the system status.

The operation of this embodiment will be described below with reference to FIG.

When the rule acquisition unit 101 starts operating, it first clears the example table 301 and the rule table 501 (step 701), and repeats the following until enough rules are generated (step 702).

First, clear the example buffer 106 in step 70.
3) The example input unit 105 reads the example from the input/output terminal 102 into the example buffer 106 (step 704
). Thereafter, the error determining unit 107 determines the possibility of an error in the read example (step 705, details will be described later). If there is no possibility of an error, the rule generation unit 109 directly generates a rule (steps 706 and 7).
09 rule generation method will be described later), if there is a possibility of an error, the error inquiry unit 108 sends the example buffer 106
The contents are displayed to the user and checked (step 707). If it is found that there is no error as a result of the check (step 708), a rule is generated as is (step 709); if it is found to be an error (step 708), the next example is input without generating the rule.

Furthermore, determination of the possibility of error in an example (operation of the error determination unit 107) will be described in detail.

FIG. 8 shows the operation of the error determining section 107.

The error determination unit 107 first reads the example table 30 in which the same conclusion as the conclusion of the example in the example buffer 106 is stored.
1 is found (step 801).

If not, this is the first example of the conclusion, so for now.

In order to reconfirm with the user, the return code is determined to be incorrect and returns (steps 802 and 805). If there is, the difference between the found examples in example table 3 and O1 is calculated as follows (step 803).

(i) Count the number of places where the value of the parameter part in the state of each corresponding element differs between the system state of the example in the example buffer 106 and the system state of one example in the example table 301, and (Example: 9th
As shown in the figure. There is a total difference of 3 between the first and second lines of the system status).

(n) The above difference is obtained for all examples in the example table 301, and the minimum value is taken as the difference between the example in the example buffer 106 and the current input example.

If the calculated difference is greater than the threshold stored in the error determination threshold area 302.

The return code is set to Error Possible and the process returns (Steps 804, 850). If it is below the threshold,
The return code is set to no possibility of error and the process returns (steps 804 and 806).

Finally, the rule generation method (operation of the rule generation unit 109)
Let's talk about.

FIG. 10 shows the operation of the rule generation unit 109.

The rule generation unit 109 first searches for the rule table 501 in which the same conclusion as the conclusion of the example in the example buffer 106 is stored (step 1001).

If it is not found, a rule with that conclusion has not been generated yet, so a new rule table 501 is created that stores the conclusion of the example in the example buffer 106, and the system state of the example is written as is in the condition area 5012. Store as a rule condition (step 1002.10
10). Furthermore, a new example table 301 is created that stores the conclusions of the examples in the example buffer 106 (there is no such example table yet), and the system status area 3 is
012, stores the actual system status (step 1
011).

If found, we will try to generalize the existing rules as follows.

First, if there are conditions covering the system state of the example in the example buffer 106 (conditions that the system state satisfies) in the rule tables 501 other than the found rule table 501, all of them are deleted (step 100
3). Next, the conditions covering the system state of the instances in the instance buffer 106 are added to the found rule table 50.
1 (step 100).
4).

If it exists, there is no need for a new rule, so the examples in the example buffer 106 are stored in the example table 301 with the same conclusion.
Additionally, step 1009) ends.

If it does not exist, create a new rule as follows.

Note that the following steps are all executed only for the rule table 501 found in step 1o01.

First, among the conditions for which filmization has not been attempted, find the one that is closest to the system state of the example in the example buffer 106 (the number of element states that cannot be covered is the least among the element states that make up the system state). By generalizing the conditions in various ways, we create conditions that can cover the actual system state. Among them, the one closest to the original condition (with the least difference) is searched for (step 1005).

As shown in the example in Figure 11, the conditions can be generalized by increasing the number of parameter parts (the part enclosed in square brackets “[”, ′]”). ) method and individual conditions (brackets ゛j (n,
#) The parts enclosed by #l) are combined in various ways (generalized conditions cover a wider range of system states).

Next, the example table 301 is checked to see if the generalized condition found in step 1005 covers the system state of an example that has a conclusion different from the conclusion of the example currently being handled (step 1005). 1.OO6)
. If all conditions are covered, repeat from step 1005 until generalization is attempted for all conditions (step 10).
06°1007). If not, a consistent generalization condition will be found, so after deleting the conditions (redundant conditions) in the rule table 501 that are covered by that condition (step 1012), the condition will be generalized. It is added to the rule table 501 as a condition of the specified rule (step 1013).

Even if you try to generalize to all the conditions, if no condition is found that does not have ``14'' (step 1007), it is no longer possible to cover the example by generalization, so the example itself is set as a new rule. Add to the table (step 1008).

Finally, the example in the example buffer 106 is added to the corresponding example table 301 (step 1009) and the process returns.

As described above, according to this embodiment, it is possible to point out and check the possibility of an error in the input example, and eliminate it if it is an error, so even if the example contains an error, the correct rule can be derived. becomes possible.

Note that in this embodiment, each time an example is input manually, the difference from the input example is calculated, and errors are measured based on a predetermined threshold value.
The difference between the examples is calculated, and examples with relatively large differences from other examples are pointed out as errors (for example, 30% of the input examples with large differences are likely to be errors). There is also a way to do this.

In addition, as a method for pointing out erroneous examples, multiple examples are input together and some cluster link method is used. 4 (a method of classifying a set of examples into several subsets).

There is also a method of dividing the data into subsets (which may be existing ones) and pointing out subsets with fewer elements as potentially containing erroneous examples.

In addition, the input example is first tentatively entered into the rule generation section, a rule is generated, and some difference is calculated between the rule table before inputting the example and the rule table after the new rule is generated. If the difference is large (e.g., similar to the difference between the examples described in this embodiment), the input example can be pointed out as potentially erroneous.

〔Effect of the invention〕

According to the invention, merely by giving a concrete example.

In a recursive storage rule acquisition system that derives general rules that can be applied to many target states, even if the given examples are incorrect, it is possible to eliminate them and derive the correct rules, making it possible to solve large-scale, complex targets. Even in this case, it has the effect of efficiently creating correct rules.

[Brief explanation of the drawing]

FIG. 1 is a block diagram of an embodiment of the present invention, FIG. 2 is a diagram showing an example of job shop control instructions, FIG. 3 is a detailed diagram of the example information storage unit, and FIG. 4 is a diagram showing job shop control. A diagram showing an example of rules, FIG. 5 is a detailed diagram of the rule storage unit,
FIG. 6 is a content diagram of an example buffer, and FIGS. 7 to 11 are processing flow diagrams and processing examples for explaining the operation of this embodiment. 101... Rule acquisition unit, 102... Input/output terminal,
103... Rule storage section, 104... Example information storage section, 105... Example input section, 106... Example buffer, 107... Error determination section, 108... Error inquiry section, 109. . . . Rule generation unit, 301 . . . Example rule table, 302 . . . Error determination threshold area,
501 No. 1 Figure 1 Q'l IL-1 Shiou] N Sai Kano No. Z Figure Ya; jii 15 feces autopsy Rin 5, pigeon 3rd eye 302% (2) Figure 5 So+ Rule〒-Dingle broom Otsu Figure 茅 l Flash early '8 Figure 9 Figure C Mibu U Kira packing C] (None) 9 3 10 Figure a Weighing 11 11 ] O1+ φ (←2(4)), (Tsurusu

Claims (1)

[Claims] 1. Means for storing examples consisting of a target state and a conclusion or action for that state, based on conditions satisfied by several target states and a conclusion or action when the condition is satisfied. In a processing system comprising means for storing rules consisting of the following, and means for generalizing the above examples to create rules,
A means for inputting examples and pointing out errors in the examples is provided, and before a rule is created by the rule creation means, errors are eliminated by an example error pointing means, and only examples without errors are input to the rule creation means. A rule acquisition system that is characterized by: 2. The rule acquisition system according to item 1, wherein the error pointing means points out errors in the examples based on the degree of difference between the examples. 3. The above error pointing means inputs the example into the above rule creation means, creates a rule, compares the rule before inputting the example with the rule after inputting, and points out errors in the example based on the degree of difference. The rule acquisition system according to item 1, characterized in that: