JPH05241837A - Rule set preparing device - Google Patents

Abstract

PURPOSE:To efficiently, gradually and increasingly prepare the rule of a logical base. CONSTITUTION:This device is equipped with an exercise base 3 which stores a positive exercise, negative exercise, and a background concept necessary for the definition of a target concept, knowledge base 4 which stores a rule for the definition of the target concept of the exercise of the exercise base, and exercise processing part 5 which fetches the exercise inputted from an inputting part 1, and judges whether or not the exercise is the positive or negative one, and retrieves whether or not the rule of the knowledge base 4 explains the positive or negative exercise And also, the device is equipped with a rule preparing part 6 which generates the new rule by referring to the exercise of the exercise base 3 when the inputted exercise is judged to be the positive one which can not be explained by the rule of the knowledge base 4 by the exercise processing part, and a discrepancy processing part 7 which corrects the rule of the knowledge base so that a discrepancy can be canceled when the inputted exercise is judged to be the negative one which can be explained by the rule of the knowledge base 4. and when the inputted exercise is inconsistent with the rule of the knowledge base 4.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a rule set generation device for generating a set of rules for supporting knowledge construction work in a system for solving a problem such as an expert system and software development work for generating a program from required specifications. Regarding

[0002]

2. Description of the Related Art In recent years, with the progress of advanced information society, advanced knowledge and know-how of experts are stored in a knowledge base, and the facts are stored using a general-purpose inference engine while using the knowledge stored in this knowledge base. An expert system has been proposed to infer data and solve problems.

Therefore, in order to develop an expert system, knowledge is acquired from an expert in a target field through an engineer conventionally known as a knowledge engineer through an interview, and each knowledge is expressed in a rule or the like format. It then stored and built knowledge into the knowledge base.

However, in order to construct knowledge in a knowledge base by such a method, it is difficult to extract knowledge about an ambiguous specialized area from an expert, and it is not possible to store all necessary knowledge in advance. Problems such as being possible are pointed out by knowledge engineers, and this is called "bolt neck in knowledge base construction".

In recent years, in order to eliminate this bolt neck, there is a growing interest in inductive reasoning and learning in which information is extracted from a past experience and a set of knowledge is automatically constructed in a knowledge base.

On the other hand, in software development as well, the demand for software development will further increase with the increase of computer usage, and in the near future, the development of necessary software will not be in time and the existing software will not be sufficiently maintained. There are concerns about such problems. This problem is essentially due to the fact that the demand for software development is increasing and the performance of hardware is improving exponentially, while the productivity of software remains low.

[0007] In order to eliminate this crisis, research on an automatic programming technique for automatically generating a program satisfying given requirements has been actively conducted. As one of the techniques of this automatic programming, there is a method by inductive reasoning. This is to give a requirement specification in the form of an example and derive a program satisfying it by inductive inference.

As described above, inductive inference in machine learning is a powerful mechanism for constructing a theory from facts,
The demand for its practical application is increasing. And, in fact, inductive reasoning of Boolean expressions and propositional level rules on decision trees has been applied to several fields of application with considerable success. For example, "ID3 (Quilian, JR: Induction of
Dicision Ttrrs International Journal of Man-Machi
ne Studies 27, 1986 ”, given a set of positive and negative examples of a concept, performs classification by top-down divide-and-conquer method using the entropy minimization principle as a selection function and generates a decision tree. It is a thing.

However, when dealing with a complicated problem, variables are not included in a definite structure such as a decision tree or a structure equivalent to a Boolean expression. Therefore, the limitation of expression at the propositional logic level is too tight. I have to say. Therefore, it is necessary to shift from propositional logic level to predicate logic level expression.

As a typical system of the conventional inductive inference system at the predicate logic level, the inductive inference based on the hypothesis refutation of the negative fact is called "model inference system MIS (Shapiro, E.
Y .: Inductive Inference of Theory from Facts, Techn
ical Report192, Yale University Computer Seience De
pt, 1981) ”. This is called the incremental enumeration algorithm. However, since the solution space that enumerates the rules is extremely large, there are problems that the number of rules that can survive is limited and that there are many examples to converge.

As an induction reasoning system developed as a method for solving this problem, "FOIL (Qilian, JR: Learning Relations: C) based on the best-priority search method of information logic.
omparison of Symbolic and a Connectionist Approac
h, Technical Report 346, Basser Dept.comp.Sc., Univer
sity of Sydney, 1989) "and" CIGOL
(Muggleton, S .: Machine Invention of First-Order Pr
idicates by InvertingResolution, In Fifth Internati
onal Conference on Machine Learning, 1988) ”was proposed. However, this resulted in problems such as inability to perform incremental learning, over-generalization, and the need to provide complete logic trace examples.

[0012] As described above, although the expectations for the knowledge acquisition technology in the construction of a knowledge base such as an expert system and the inductive inference technology as the child programming language technology in the software development are increasing, the calculation amount and the example Since the acquisition method does not have a realistic performance, there is no practical inductive inference system at the predicate expression level.

[0013]

By the way, in order to use the induction inference system as a practical technique, the following requirements must be satisfied.

First, the user is required to be given an example interactively with the system, and to be an interactive system in which the system inquires the user of all the necessary examples for the induction reasoning system at any time. To be done. Second, it is required that the system be able to perform a dynamic inference method, and that if the inductive inference system finds that the logic generated by the input of the example from the user is incorrect, it is corrected. Thirdly, the system is required to narrow the search of the solution space and improve the calculation efficiency, specifically, to positively induce from positive examples and to divert past similar rules. Fourth, it is required that the mechanism of inductive inference of the system is incremental, and that the system can generate more desirable rules by giving an example.

The present invention provides a rule set generation device capable of generating rules efficiently and incrementally by simply adding examples to the system interactively without declaring all the examples in advance. To aim.

[0016]

In order to achieve the above-mentioned object, the present invention provides a background concept that is necessary to define a positive example that is a sample of a goal concept, a negative example that is not a sample, and a goal concept. , A theory base in which the rules that define the goal concept of the example stored in this example base are stored, an input means for inputting the example, and an example input from this input means An example processing means that determines whether the example is a positive example or a negative example, and verifies whether the positive or negative example can be derived or deduced from the theory-based rule, and is input by the example processing means. If the example is determined to be a positive example, and the positive example cannot be explained by the theory-based rule, then a new rule is generated by referring to the example-based example, The rule inputting means for constructing the theory base and the example input by the example processing means are determined to be negative examples, and this negative example is activated when it can be explained by the theory-based rule. When it is determined that the theory-based rule is inconsistent with the theory-based rule, the theory-based rule is modified so that the contradiction is resolved.

In addition to the above configuration, if various logical redundancy is detected in the rules added to the theory base by the rule generation means, the redundancy of the rules and the redundancy between the rules are eliminated. It is provided with a rule refining means.

[0018]

In the rule set generation device having such a configuration, the input means acquires an example of the target concept in an interactive manner, and the examples are sequentially given to the example processing means in a time series, so that the positive example or the negative example is given. If it is judged as a positive example, it is verified whether the theory-based rule can be explained for that example. If it cannot be explained, the rule generation means is activated to enumerate. First, we generate rules by constructive induction based on minimal generalization and specialization, add them to the theory base, and verify whether theory base rules can be explained for negative examples. If it can be explained, the contradiction processing means is activated assuming that the contradiction with the example-based example has occurred, and the theory-based rule is modified so that the contradiction can be resolved. Can improve That. Further, by providing a rule refining means for refining the redundancy of the rules with respect to the knowledge base, when the newly generated rule is added to the theory based theory, the redundancy of the theory and rules is eliminated, A complex theory can be transformed into a logically equivalent and concise theory.

[0019]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS An embodiment of the present invention will be described below with reference to the drawings.

FIG. 1 is a block diagram showing a system configuration example of a rule set generation device according to the present invention. In FIG.
Reference numeral 1 is an input unit for inputting an example, and 2 is a buffer for storing the example input from the input unit 1. 3 defines a positive target concept example (which is simply referred to as a positive example), which is a sample of target concepts (goal concept), a non-sample negative target concept example (which is simply referred to as a negative example), and a target concept. An example base, in which three types of examples of positive background concepts, which are samples of the background concept (background concept) necessary for the, are stored, 4 is sufficient to define the target concept, and It is a consistent theory that does not cause contradiction, that is, a knowledge base that stores a set of rules that define a goal concept as knowledge.

Reference numeral 5 denotes an example processing section, which is a positive / negative example determination section 5-1 and a positive example verification section 5-2.
And a negative example verification section 5-3. Positive,
The negative example determination unit 5-1 sequentially takes in the examples stored in the buffer 2, and refers to the example in the example base 3 to determine whether the example is a positive example or a negative example. When the positive example verification unit 5-2 determines the positive example by the example determination unit 5-1, the positive example verification unit 5-2 verifies whether this example can be explained by the rules of the knowledge base 4, that is, whether it can be covered. When the negative example verification unit 5-3 determines that the example input to the example determination unit 5 is a negative example, whether the negative example is explained by the rules of the knowledge base 4, that is, is covered. Verify that you are there.

Further, 6 is a rule generation unit, and this rule generation unit 6 is composed of a minimum generalization unit 6-1, a rule correction unit 6-2 and a schema memory 6-3. Minimal generalization unit 6-
If the positive example 1 is verified by the positive example verification unit 5-1 that the positive example cannot be explained by the rules of the knowledge base 4, that is, it cannot be covered, a plurality of positive examples related to the goal concept and the background concept from the example base 3 And a rule relating to another previously generated target concept registered in the schema memory 6-3 is referred to, and a minimum generalization operation is performed on the plurality of positive example sets. As a result of performing the minimum generalization operation by the minimum generalization unit 6-1, the rule correction unit 6-2 specializes the rule when the rule is an over-generalized rule, and performs the specialization of the rule. It is stored in the knowledge base 4 from 6-1. Here, the minimum generalization operation means a procedure for generating the most specific generalized logical expression from two logical expressions.

On the other hand, 7 is a contradiction processing section, and this contradiction processing section 7 is composed of a diagnosis section 7-1 and a rule specialization section 7-2. When the diagnosis unit 7-1 determines that the example input to the example determination unit 5-1 is a negative example, whether the negative example is explained by the rule of the knowledge base 4, that is, whether the negative example is covered. If it is verified that it is covered by the rules of knowledge base 4, the rules in knowledge base 4 conflict with the input negative examples or the negative examples in example base 3 Identify the rules,
Retrieve the rule from Knowledge Base 4. Rule specialization section 7-2
Modifies the rules of the knowledge base 4 extracted by the diagnosis unit 7-1 by specializing or rejecting the rules of the knowledge base 4 so as not to cause a logical contradiction with the example base 3. Further, 8 is a rule refining section that simplifies the theory that rules stored in the knowledge base 4 become redundant due to addition of various rules and eliminates redundancy.

Next, the operation of the rule set generation device having the above configuration will be described with reference to the general flow of rule generation shown in FIG. In the initial state when the system is started, the user has not provided a complete example base for generating rules. Therefore, the user inputs an example to the system, or the system positively asks a question to the user to acquire the example in time series. In addition, the system executes the inductive inference procedure while using the example processing unit 5 as a contact point with the user and sequentially calling the three processing units of the rule generation unit 6, the contradiction processing unit 7, and the rule refinement unit 8.

First, when a positive example and a negative example are input from the input unit 1, these examples are temporarily stored in the buffer 2. When the system is started up in such a state, the example determination section 5 of the example processing section 5 stores the positive data stored in the buffer 2,
Negative examples are sequentially taken in, and it is determined whether the example is a positive example or a negative example by referring to the example of the example base 3. In this case, the example processing unit 5 not only receives the example from the user, but also asks the user in the form of an example a question regarding the target concept in order to effectively eliminate the erroneous rule from the knowledge base 4 and inputs it. The logical consistency between the example given above and the rules of knowledge base 4 is verified.

Now, the example determination section 5-1 causes the buffer 2
If it is determined that the input example is a positive example,
The example is taken into the positive example verification unit 5-2. The positive example verification unit 5-1 verifies whether the positive example is covered by the rules of the knowledge base 4.

If it is determined that the rule is not covered, the rule of the knowledge base 4 does not fully define the target concept, so that the rule generation unit 6 tries to add a rule for explaining it. To start. When the rule generation unit 6 is activated, a set of a plurality of positive examples related to the target concept and the background concept is extracted from the example base 3 with respect to the positive example not covered by the minimum generalization unit 6-1. , Perform a minimum generalization operation on these sets. That is, the one most specific literal set on the logical expression that can be variable for the set of a plurality of examples is output to the knowledge base 4 as a rule. In this case, if the rule is over-generalized by the minimum generalization operation, the rule modification unit 6-2 specializes the rule.

Since the solution space is searched by the positive data existing in the example base 3 by such a technique of the minimum generalization, the search efficiency can be improved as compared with the inductive inference based on the refuted solutions. Becomes

In this case, it is difficult to improve the efficiency by extracting the set of all positive examples registered in the example base 3 and performing the minimum generalization. Therefore, the set of positive examples is taken from the example base 3 by the closeness of the concept to the examples to be covered. This is called bias. This bias gives the generalization direction of the example set in the minimum generalization unit 6-1. Further, by analogizing with reference to the skimmer memory 8 in which the rule relating to another target concept generated in the past is registered, the similar rule in the past can be reproduced from the example base 3 and reused.

The procedure of the rule generator 6 is shown in FIG. The rule generation unit 6 generates an effective rule by using the technique of minimum generalization and analogy and the concept of bias as described above, and the details thereof will be described. The rule to be generated can be obtained by performing a minimum generalization while sequentially extracting the set of examples related to the positive examples that are not currently covered from the example base 3 according to the bias. Here, the bias is for guiding the effective search of the solution space, and is set for each level according to the terms and constants occurring in the example to be covered. For example, at level 1, all the example arguments taken from example base 3 occur in the example arguments to be covered. The higher the level, level 2, 3, ...
The degree of relevance on the argument between the example to be covered and the example to be taken out is low. The level means a conceptual distance between the example to be covered and the example in the extracted example base 3.

Further, in the operation of extracting this data, the analogy technique is used. In other words, the past production rules can be reused by extracting a set of examples in which the similarity can be obtained between the schema and the argument. Further, when the number of examples is insufficient and a complete correspondence cannot be obtained, the user is requested to give examples based on analogy.

The minimum generalization operation as shown in FIG. 4 by the minimum generalization unit 6-1 starts from level 1 of the bias, does not cover the negative example, and has a rule that covers the positive example to be covered. It is done until it is generated. If the minimum generalization operation is performed unilaterally during the operation, the rule correction unit 6-2 performs the specialization operation as shown in FIG. The rule correction section 6-2 is equipped with two refinement operators. And when a new rule is generated,
Ask the user for legitimacy of the rule and reject the rule if the legitimacy is not guaranteed.

FIG. 6 shows an example of the bias-guided minimum generalization process. We are thinking of QUICKSORT as a goal concept. A rule generalized with a bias of level m leads to a negative example, and it is rejected because it is not corrected correctly even if the rule is specialized using two refinement operators. However, at a bias of level n (m <n) after level m, it contains the set of correct examples after the minimal generalization operation, and the argument of the level n example is that of the examples to cover more than level m. The conceptual distance from the argument is increasing.

FIG. 7 shows an example of contents registered in the schema base 6-3. Many schemas generated by the system in the past are registered in the schema base 6-3, or a user registers a theory similar to the theory defining the target concept in the initial state of system activation. It is not uncommon for an operator to generate a rule for a new goal concept by analogy with past experience.
If a schema that expresses past experiences exists, efficient inductive reasoning is possible using the analogy idea. here,
SORT of the concept generated in the past and the abstracted concept Pre
Are stored as a schema.

FIG. 8 illustrates how a similar schema can be reused for a new goal concept rule. This is a positive example taken from the example base based on the SORT schema.
QUICKSOR, which is a similar goal concept by performing minimum generalization
We are building a rule that defines T.

FIG. 9 shows an application example of a refinement operator for specializing a rule. Since the rule defining the goal concept QUICKSORT was over-generalized, I applied the unification operator to the first rule and the assignment operator to the second rule to avoid covering negative examples.

Next, in the positive / negative example determination section 5-1,
When it is determined that the example input from the buffer 2 is a negative example, the negative example verification unit 5-3 verifies whether or not the example is covered by the rules of the knowledge base 4, and if the example is covered. If it is determined that there is a contradiction processing unit 7
To start. When the diagnosis unit 7-1 of the contradiction processing unit 7 determines that the input negative example causes a contradiction with the rule in the knowledge base 4 or the negative example in the example base 3, the error is erroneous. Identify the rule and retrieve the applicable rule from Knowledge Base 3. Then, this rule is given to the rule specializing unit 7-2 to specialize or reject the rule, and the rule of the knowledge base 4 is corrected so that the rule does not conflict with the example base 3.

10 and 11 show the procedure of the contradiction processing unit 7 and the call of the diagnosis unit 7-1. The contradiction processing unit 7 calls the diagnosis unit 7-1 to identify the rule in the knowledge base 4 which causes the contradiction. This is done by listening to the truth of the example that is the leaf of the explanation tree interactively with the user by using the explanation tree which is the process of covering the negative example by theory. Then, in order to specialize the wrong rule causing the contradiction, the rule specializing section 7-2 is called. The rule specialization unit 7-2 applies the refinement operator to the rule until the rule to be specialized no longer covers a negative example. Operators include a single operator and a term assignment operator.

As described above, the user inputs an example to the system, or the system positively asks a question to the user to acquire the example in a time series. By acquiring the new example, the knowledge base 4 acquires the example. It will be updated to rules that are sufficient and consistent to explain Base 3.

In this case, the rule refining unit 8 is activated for the purpose of eliminating the redundancy of theory accumulated in the knowledge base 4.
FIG. 12 shows the procedure of the rule refining section 8. Here, various theoretical redundancies generated by adding a new rule to the knowledge base 4 are found and solved.
When the redundancy of the conditional part of the rule is found, the rule condition deletion operator is used to reduce the conditional part. If the terms in one rule can be reduced, the term rewriting operator is used to convert the term into a simplified rule. If there is redundancy among a plurality of rules in the knowledge base 4, the rule is deleted by using the rule deletion operator.

FIG. 13 shows an example in which a rule with the first redundant condition is given. In the case of such an example, a non-redundant rule is generated by a modification operation by a rule condition deletion operator, which does not cover a negative example like the rule in the lower line. The second is rewriting of terms by the term rewriting operator to generate an equivalence rule having a simpler expression. In the third rule, since the rules are redundant in the knowledge base 4, rule 2 is eliminated by the rule deletion operator. Further, the rule refining unit 8 defines the same target concept and generates a theoretical noun that generates a new predicate from two rules that share a part of the condition.

Here, examples of the contents of the example base 3 and the knowledge base 4 handled by the system are shown in FIGS. 14 and 15. In this example, QUICKS to sort the list elements in ascending order
ORT is the target concept. As a method of expressing the rules, we are considering the PROLOG horn clause (A: -B1, B2, ... Bn). Example base 3 includes positive and negative goal concept examples related to the goal concept QUICKSORT and PARTITION (greeter) of the background concept.
A positive background conceptual example concerning etc. is registered. Here, the materialized n = 0 Horn clauses (facts) are registered together with the knowledge.

Further, in the knowledge base 4, a rule defining QUICKSORT of the target concept generated recursively by the system is registered. This system dynamically changes the rules of the knowledge base 4 by interacting with the user. For example, if the positive example input by the user cannot be explained by the rule in the knowledge base, a new rule for explaining the example is added to the knowledge base 4. Further, when a contradiction occurs in the knowledge base 4 due to the negative example input by the user, the rule causing the contradiction is taken out and the rule is specialized or rejected. Thus knowledge base 4 is essentially non-monotonic,
The content of the knowledge base 4 is updated by acquiring an example based on the interaction with the user, and sufficiency and consistency are maintained.

As described above, in the present embodiment, the examples obtained by the user inputting the examples from the input unit 1 and the user's request for the examples are stored in the buffer 2, and after the examples are acquired, the examples are chronologically output from the buffer 2. Is taken into the example processing section 5, and the example determination section 5-1 determines whether the example is a positive example or a negative example. If the example is determined to be a positive example, the positive example verification section 5-2. Verifies whether the example can cover the rules of the knowledge base 4, and if it does not, the rule generation unit 6 is activated. If the example determination unit 5-1 determines that the rule is a negative example, the negative example verification unit 5-3 verifies whether the rules of the knowledge base 4 can be covered.
If it is covered, it is determined that the knowledge base 4 and the example base 3 are in conflict, and the conflict processing unit 7 is activated. Then, when the rule generation unit 6 is activated, the minimum generalization unit 6-1 extracts a set of a plurality of positive examples related to the target concept and the background concept for the example from the example base 3, and Generalization operation is performed to generate a new rule, and when this rule becomes a rule that is overgeneralized by the generalization operation, specialization is performed by the rule correction unit 6-2 and the knowledge base 4 is generated. When the contradiction processing unit 7 is added, the diagnosis unit 7-1 identifies the erroneous rule causing the contradiction, extracts the corresponding rule from the knowledge base 4, and adds the rule specialization unit. 7-2, the rule is specialized or rejected and the rule of the knowledge base 4 is modified so as not to cause a contradiction with the example base 3.

Therefore, an example of the target concept is interactively acquired, rules that are erroneously and functionally generated are rejected early, and rules are generated by constructive induction based on minimal generalization and specialization without enumeration. Since, is generated, it is possible to efficiently search the solution space from the input example. In the rule generation unit 6, when the minimum generalization operation is performed by the minimum generalization unit 6-1, a rule similar to the current target concept is calculated from the set of rules generated in the past with reference to the schema memory 6-3. By taking out, transforming and using it, it is possible to improve the generation efficiency and generation probability of correct rules. Furthermore, by providing the rule refining unit 8 for refining the redundancy of the rules with respect to the knowledge base 4, when the newly generated rule is added to the theory of the knowledge base 4, the redundancy of the theory and the rules is eliminated. Eliminate and transform complex theories into logically equivalent and concise theories.

Although the above embodiment has described the case where a knowledge base such as an expert system is constructed,
You may make it apply as a theoretical base which stores the program in the case of developing software. in this case,
When applied as a knowledge base of an expert system, facts and processing results in inference processing are stored as an example in the example base. Moreover, in order to apply it as a logical base when developing software, the required specifications of the software developed on an example basis are stored.

[0047]

As described above, according to the present invention, it is possible to efficiently and incrementally generate rules simply by interactively adding an example to the system without declaring all the examples in advance. A rule set generation device can be provided.

[Brief description of drawings]

FIG. 1 is a block diagram showing a system configuration of an embodiment of a rule set generation device according to the present invention.

FIG. 2 is a flowchart showing a procedure flow of the schematic name system of the embodiment.

FIG. 3 is a diagram for explaining a procedure of a rule generation unit.

FIG. 4 is a diagram for explaining a procedure of a minimum generalization unit.

FIG. 5 is a diagram for explaining a procedure of a rule correction unit.

FIG. 6 is a diagram for explaining a processing process of a minimum generalization unit.

FIG. 7 is a diagram showing an example of a skimmer base.

FIG. 8 is a diagram showing generation of rules by analogy from a schema.

FIG. 9 is a diagram showing an application example of a rule correction unit.

FIG. 10 is a diagram for explaining the procedure of the contradiction processing unit.

FIG. 11 is a diagram for explaining the procedure of the diagnosis unit.

FIG. 12 is a view for explaining the procedure of the rule refining unit.

FIG. 13 is a diagram showing an application example of a regular refining unit.

FIG. 14 is a diagram showing an example of an example base handled by the system.

FIG. 15 is a diagram showing an example of a knowledge base handled by the system.

[Explanation of symbols]

1 ... Input part, 2 ... Buffer, 3 ... Example base, 4
...... Knowledge base, 5 …… Example processing unit, 5-1 …… Positive and negative example determination unit, 5-2 …… Positive example verification unit, 5-3 …… Negative example verification unit, 6 …… Rule Generation unit, 6-1 ... Minimal generalization unit, 6-2 ... Rule correction unit, 6-3 ... Schema memory, 7 ... Contradiction processing unit, 7-1 ... Diagnosis unit, 7-2.
Rule specialization processing section, 8 ... Rule refinement section.

Claims (5)

[Claims] 1. An example base in which a positive example which is a sample of a goal concept, a negative example which is not a sample, and a background concept which is a background necessary for defining a goal concept are stored, and an example stored in this example base. The knowledge base that stores the rules that define the goal concept of, the input means for inputting the example, and the example input from this input means are taken and it is determined whether the example is a positive example or a negative example. An example processing means for verifying whether the positive or negative example can be covered by the theory-based rule, an example input by this example processing means is determined to be a positive example, and this positive example is the theory. Rule generation means that is activated when the rule of the base cannot be explained, generates a new rule by referring to the example of the example base, and constructs the new rule in the theory base,
It is determined that the example input by the example processing means is a negative example and this negative example can be explained by the theory-based rule, and it is determined that the input example is inconsistent with the theory-based rule. And a contradiction processing unit that corrects the theory-based rule so that the contradiction is resolved. 2. An example base in which a positive example that is a sample of a goal concept, a negative example that is not a sample, and a background concept that is a background necessary for defining a goal concept are stored, and an example stored in this example base. The theory base in which the rules that define the goal concept of is stored, the input means for inputting the example, and the example input from this input means are taken in to determine whether the example is a positive example or a negative example, An example processing means for verifying whether the positive or negative example can be covered by the theory-based rule, an example input by this example processing means is determined to be a positive example, and this positive example is the theory. Rule generation means that is activated when the rule of the base cannot be explained, generates a new rule by referring to the example of the example base, and constructs the new rule in the theory base,
It is determined that the example input by the example processing means is a negative example and this negative example can be explained by the theory-based rule, and it is determined that the input example is inconsistent with the theory-based rule. Then, when the logical redundancy is detected in the contradiction processing means for correcting the theory-based rule so that this contradiction is resolved, and the rule added to the theory-based rule by the rule generation means. A rule set generation device comprising rule refinement means for eliminating the redundancy of rules and the redundancy between rules. 3. The rule generation means extracts a plurality of positive examples relating to the target concept and the background concept from the example base, and performs a minimum generalization operation on a set of the plurality of positive examples. 3. The rule set generation device according to claim 1 or 2, further comprising: a rule modification unit that specializes and modifies a rule that has been over-generalized by the minimum generalization unit. 4. The rule set generation device according to claim 3, wherein the minimum generalization means has a function of extracting an example by a bias and estimating a new rule by using similar rules in the past. 5. The inconsistency processing means, when a negative input example is determined to be inconsistent with the theory-based rule, identifies the inconsistent rule, and extracts the corresponding rule from the theory-based rule. A means and a rule specializing unit for specializing the rule extracted by the diagnostic means or for rejecting the rule and modifying the theory-based rule so as not to cause a contradiction with the example-based example. Alternatively, the rule set generation device according to 2.