JPH0659896A - Method and device for correcting hypothesis in inductive inference - Google Patents

Abstract

PURPOSE:To prevent a hypothesis from being formed in duplicate to the utmost in the process of correcting the hypothesis by using a limit rule so as to select hypothetical and candidate and using the selected candidate as the hypothesis. CONSTITUTION:Let a current hypothesis be H, an area storing succeeding hypothesis candidate be Q and a limit to the hypothesis be R as initial setting (101), and then whether or not the H is opposed to facts having been known so far is checked (102). Whether or not the H proves a fact (correct fact) to be substantially testified is checked (103). When there is any correct fact not testified, an upward operator is applied to the H, the obtained result is selected based on the limit rule R and the result is stored in the area Q (104). Furthermore, whether or not the H proves a fact (negative fact) not substantially to be testified is checked (105). When there is any negative fact to be testified, a downward operator is applied to the H, the obtained result is selected by the limit rule R and the result is stored in the area Q (106). A succeeding hypothetical object is extracted from the area Q and used for a new H (107) and the processing returns to the step (102).

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a hypothesis correction method and device in inductive inference for correcting a hypothesis using a hypothesis correction operator.

[0002]

2. Description of the Related Art A method of inductive inference using a hypothesis correction operator has been proposed by Laird (Laird,
P. D. 'Inductive Experience
byRefinement ', YALEU / DCS
/ TR-376). In this method, inference is advanced using only one hypothesis correction operator. The hypothesis correction operator used at this time needs to have the property (completeness) that the space of the hypothesis can be completely covered by the operator. As the initial value of the hypothesis used in the inference, the most general one or the most special one in the hypothesis space is used.

Kawasaki has extended this method and proposed an inductive reasoning that combines two hypothesis correction operators with a method of analogy (Yoji Kawasaki'Model reasoning using bidirectional refinement operators', AI Society). National Convention (2nd) Proceedings, 111
-114). With this method, the inference can proceed with the initial value of the hypothesis anywhere in the hypothesis space. We also propose that one of the two hypothesis correction operators does not have the property of completeness, so one of them does not have the property of completeness.

[0004]

In the above-mentioned method, the hypothesis correction operator, which does not have the property of completeness, is designed so that its hypothesis correction ability is not lowered as much as possible. However, there was a drawback that duplicates were likely to appear in the hypotheses obtained as a result of the modification. Moreover, since the obtained candidate hypothesis is used as it is as the next hypothesis, it is not possible to exclude an inappropriate hypothesis, which slows down the inference speed.

[0005]

The hypothesis correction method of the first invention is a hypothesis correction method in inductive inference that corrects a hypothesis by using a plurality of hypothesis correction operators. After making it larger than the operator of, and selecting the obtained hypothesis candidates using the restriction rule,
It is characterized by using it as a hypothesis.

The hypothesis correction device of the second invention is a hypothesis correction device for inductive inference that corrects a hypothesis using a hypothesis correction operator, and an old hypothesis storage area for managing an old hypothesis to be corrected and a corrected result. The hypothesis management unit that holds the new hypothesis storage area that manages the obtained new hypothesis, the hypothesis modification rule management unit that manages the rules of the hypothesis modification operator, and the rules that limit the hypothesis, and the old hypothesis A hypothesis correction operator for applying a hypothesis correction operator according to the above, and selecting a hypothesis obtained as a result of the correction according to a restriction rule.

[0007]

The present invention will be described below with reference to the drawings.

In FIG. 2, a hypothesis correction operator in which a logic program is an inference target and an example in which it is applied to a logic program will be described.

Rules 201 and 202 are examples of hypothesis correction operators for logic programs. Rule 201 is a perfect hypothesis correction operator for logic programs. Although the rule 202 is not perfect for the logic program, it is a hypothesis correction operator whose correction amount is larger than that of the rule 201. Each item of the modified rule in the rule 201 makes the logical program more special (the proof ability of the logical program becomes smaller). For this reason, we will call it the downward hypothesis correction operator. Each item of the modified rule in the rule 202 makes the logic program more general (the proof ability of the logic program becomes larger). Therefore, we call it an upward hypothesis correction operator.

In the figure, the old hypothesis (logical program)
Let be e, and let T be the set of facts that were originally to be proved but not proved by e.

The program 210 is an example of a logic program, and is a current hypothesis. p is a predicate symbol, f is a function symbol, a is a constant symbol, and X, Y, and Z are variable symbols. In the following explanation, variables are represented by capital letters. 21
1 is a set of facts that should not be proved by the current hypothesis among the facts that should be proved.

The logic programs 221 to 225 are logic programs obtained by applying the downward hypothesis correction operator definitions (d1) to (d5) to the logic program 210, respectively. 221 removes node (1). 222 adds a derivation clause obtained by using two clauses (1). 223 adds a clause obtained by unifying two variables X and Y appearing in clause (1). 224 adds a clause obtained by substituting the most general term for the variable Y appearing in clause (1). 225 adds a clause obtained by attaching the most general atom to the clause (2).

The logic programs 231 to 234 are logic programs obtained by applying the upward hypothesis correction operator definitions (u1) to (u4) to the logic program 210, respectively. 231 adds a clause consisting of the elements in T. In variable 232, the variable Y appearing in section (1) appears in two places, but one of them is replaced with another variable. 233 replaces the constant term appearing in section (2) with a new variable. 234 removes one atom in section (1).

FIG. 1 is a flow chart showing an embodiment of the first invention. In the description, a logic program will be explained as an inference target.

Step 101 is an initial setting, where H is the current hypothesis, Q is the area for storing candidate hypotheses for the next and subsequent hypotheses,
Let R be the limiting rule of the hypothesis.

In step 102, it is checked whether H is against any facts known so far. If not, the procedure ends. If it is contrary, the following processing is performed depending on how it is against the fact.

In step 103, it is checked whether H does not prove the fact (correct fact) that should be originally proved.

If there is an unproven truth, then proceed to the next step 104, apply the upward operator to H, select the obtained result by the restriction rule R, and then insert Q. If there is no unproved positive fact, the process directly proceeds to step 105.

In step 105, it is checked whether H proves a fact (negative fact) which should not be proved.

If there is a negative fact to be proved, the process proceeds to the next step 106, the downward operator is applied to H, the obtained result is selected by the restriction rule R, and then Q is inserted. If there is no negative fact to prove, the process directly proceeds to step 107.

In step 107, the next hypothesis candidate is taken out of Q, newly set to H, and the process returns to step 102.

FIG. 3 is a block diagram showing an embodiment of the second invention.

Reference numeral 310 is a hypothesis correction rule management unit that manages rules for correcting hypotheses. As the hypothesis correction rule, the rule of the hypothesis correction operator and the rule for selecting the hypothesis obtained as a result of applying the operator are managed.

Reference numeral 330 is a hypothesis management unit for managing hypotheses. It manages the old hypothesis before applying the hypothesis correction operator and the new hypothesis obtained by modifying the old hypothesis.

Reference numeral 320 is a hypothesis correction unit for correcting the hypothesis. First, the hypothesis correction operator application unit 321 obtains information on an old hypothesis from the old hypothesis storage area 331. Next, the hypothesis correction operator application means 321 is the hypothesis correction operator management means 3
The information about the hypothesis correction operator is obtained from 11, and each rule is applied to the old hypothesis to obtain a new hypothesis. The hypothesis restriction unit 322 receives the new hypothesis created by the hypothesis correction operator application unit 321, obtains a restriction rule for the hypothesis from the hypothesis restriction rule management unit 312, and selects the received hypothesis. The obtained hypothesis is transferred to and stored in the new hypothesis storage area 332.

FIG. 4 shows an example of a hypothetical restriction rule in the first and second inventions. Rule 401 is a rule that variables that appear in the body of each clause in the hypothesis must also appear in the head. Using this rule, logic programs 225 and 222 of FIG. 2 are not subject to the new hypothesis.
Rule 402 is a rule that the compound term does not appear in the body of each clause in the hypothesis. Using this rule, the logic program 224 of FIG. 2 is not subject to the new hypothesis.

[0027]

According to the above invention, in inductive inference in which a hypothesis correction operator is used to correct a hypothesis, it is possible to prevent duplicate hypotheses from being generated in the process of correcting hypotheses. In addition, by eliminating inappropriate hypotheses by the restriction rule, it becomes unnecessary to perform extra tests on the hypotheses. The inference speed can be improved by the above two points.

[Brief description of drawings]

FIG. 1 is a flowchart showing an embodiment of the first invention.

[Fig. 2] Hypothesis correction operator in which a logic program is an inference target and an example in which it is applied to a logic program

FIG. 3 is a block diagram showing an embodiment of the second invention.

FIG. 4 is an example of a hypothesis restriction rule in the first and second inventions.

[Explanation of symbols]

101 Initial setting 102 Judgment whether current hypothesis is not true 103 Judgment whether current hypothesis does not prove true fact 104 Processing when current hypothesis does not prove true fact 105 Current hypothesis Whether the current hypothesis does not prove negative facts 106 processing 107 when the current hypothesis does not prove negative facts 201 update processing of the current hypothesis 201 an example of a downward hypothesis correction rule for a logical program 202 an upward direction for a logical program Example of hypothesis correction rule 210 Example of logic program 211 Example of unproven true fact 221 Example of logic program 210 modified by rule (d1) 222 Example of logic program 210 modified by rule (d2) 223 Logical program 210 Example modified by rule (d3) 224 Logic program Example 10 in which 10 is modified by rule (d4) 225 Example in which logical program 210 is modified by rule (d5) 231 Example in which logical program 210 is modified by rule (u1) 232 Example in which logical program 210 is modified by rule (u2) 233 Example where the logic program 210 is modified by the rule (u3) 234 Example where the logic program 210 is modified by the rule (u4) 310 Hypothesis modification rule management unit 311 Hypothesis modification operator management unit 312 Hypothesis restriction rule management unit 320 Hypothesis modification unit 321 Hypothesis Correction operator applying means 322 Hypothesis limiting means 330 Hypothesis management section 331 Old hypothesis storage area 332 New hypothesis storage area 401 Hypothesis restriction rule example 402 Hypothesis restriction rule example

Claims (2)

[Claims] 1. In a hypothesis correction method for inductive inference that corrects a hypothesis using a plurality of hypothesis correction operators, the hypothesis correction amount of one operator is made larger than that of another operator, and the obtained hypothesis is obtained. A hypothesis correction method in inductive inference, which is characterized by selecting and selecting candidates of using a restriction rule and then using them as a hypothesis. 2. A hypothesis correction device for inductive inference that corrects a hypothesis using a hypothesis correction operator, manages an old hypothesis storage area for managing an old hypothesis to be modified, and a new hypothesis obtained as a result of modification. The hypothesis management section that holds the new hypothesis storage area, the hypothesis modification rule management section that manages the rules of the hypothesis modification operator, and the rules that limit the hypothesis, and the application of the hypothesis modification operator to the old hypothesis A hypothesis correction device for inductive inference, comprising: a hypothesis correction unit that selects and selects a hypothesis obtained as a result of correction by a restriction rule.