JPH0573314A - Inference system for artificial intelligence system - Google Patents

Abstract

PURPOSE:To speed up the rule retrieval of an inference operation by selecting the rule from pointer information A and executing the inference while selecting the rule having a conclusion part with the same content successively from pointer information B. CONSTITUTION:In the case of the inference of a conclusion 'take cold tablets' in the diagnosis of a cold, a rule 4 corresponds to the one at the retrieval of the rule having the conclusion of 'take cold tablets'. Then, when the condition of the rule 4 is checked, the only one item of 'there is some doubt about a cold' is found, hereby turning to the evaluation of the condition part immediately. In the case of the establishment of the 'some doubt about a cold' the rule 4 is immediately established, and the inference of 'take cold tablets' ends up in success. In case of the presence or absence of the 'doubt about a cold' can not be decided, a rule 1 corresponds to the case at the retrieval of the rule having the 'doubt about a cold' in the conclusion part. Since the rule 1 consists of three items connected by an OR operator, the conclusion part is established when any of the three items is established. After the evaluation, the propriety of the inference is decided after being returned to the rule 4.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an inference method for an artificial intelligence system in which a rule search during inference execution in the artificial intelligence system and its evaluation method are improved.

[0002]

2. Description of the Related Art Artificial intelligence systems that store the knowledge and know-how of specialists in a computer as a knowledge base in a computer, and the computer makes inferences based on the knowledge base are financial, steel and medical systems. , Has been developed in many fields such as education. The main operation of the artificial intelligence system is to search the knowledge base for a rule that matches the inference target, and interpret and execute the corresponding rule.

However, the above artificial intelligence system generally requires a large knowledge base describing the knowledge of the expert. Therefore, there is a problem that a large amount of memory is required to store a large amount of knowledge, and a large amount of time is required to search for a rule that matches the inference target and interpret and execute the rule during inference.

[0004]

SUMMARY OF THE INVENTION In view of the above problems of the prior art, the present invention mainly aims to perform rule search for inference operation at high speed, and also intends to use memory efficiently. To do.

[0005]

In the inference method of the artificial intelligence system according to the present invention, the above-mentioned problems are solved by limiting the pointer information indicating the location of the rule and the expression of the rule. That is, the present invention searches for a rule having a conclusion part that matches the content of the inference target from among a plurality of predetermined rules composed of a condition part and a conclusion part, and artificial intelligence that verifies the success or failure of the rule. In the inference method of the system, pointer information A indicating the location of a rule having a certain conclusion in the conclusion portion and pointer information B indicating the location of a rule having the same conclusion portion are provided, and a rule is selected by the pointer information A. , Pointer information B
According to the inference method of the artificial intelligence system characterized by sequentially selecting the rules having the conclusion part having the same content and executing the inference, or using the expression form of the conclusion part as one term and the condition part as two terms It is an inference method for an artificial intelligence system, which is characterized by performing.

The knowledge representation in the artificial intelligence system of the present invention is a rule. The rule is expressed in the form of IF <conditional part> THEN <conclusion part>. The condition part and the conclusion part consist of a single item or a plurality of items. In particular, when the condition part is a plurality of items, a logical operator expressing the relation between the items, for example, AND, "or" meaning "and" is used. Meaning OR
And the like are used to construct a logical expression. Since each item is a unit of a logical expression, it is simply expressed as “term” in the present invention. Each term is
It is the smallest unit where the truth can be evaluated logically.

Generally speaking, the reasoning of an artificial intelligence system using rules is based on the assumption of the so-called forward reasoning method for inferring the conclusion of the conclusion part and the conclusion part using rules that match the given conditions. It is roughly classified into a so-called backward inference method that infers whether the condition part of a matching rule is satisfied. The present invention is effective for the latter backward inference.

In the backward inference, first, a rule having a conclusion part that matches the content of the inference target is searched for. In the present invention, for a certain conclusion, pointer information A indicating the location of a rule having the conclusion in the conclusion part is used. The pointer information is information that gives the location, and specifically, the serial number of the rule is preferably used. More specifically, it is also preferable to use the head memory address of the memory in which the rule is stored. Further, it is also used to give the pointer information a rule that does not actually exist to mean that the corresponding rule does not exist. For example, when a rule serial number is adopted, it is preferable to assign 0 or a negative number to this. By using the pointer information A, without searching the conclusion part of the rule,
You can get the whereabouts of the rules.

In the above inference, the number of rules having a certain conclusion in the conclusion part is not limited to one, but is often plural. When a plurality of rules are applicable and the rule obtained by the pointer information A does not succeed in inferring the success or failure of the conclusion part, the inference is continued using another rule having the same conclusion part. .. First, the pointer information B is given to the rule obtained by the pointer information A,
The location of another rule having the same conclusion part is shown. The pointer information B is also given to the rule obtained by the pointer information B, and the rules having the same conclusion part can be sequentially obtained by the pointer information B. When the corresponding rule does not exist, the specific information that does not exist is given.

Pointer information A and pointer information B
By using, it is possible to sequentially select all the rules having a conclusion part that matches the content of the inference target.

In a preferred embodiment of the present invention, the conclusion part is expressed by one term and the condition part is expressed by two terms. A rule with multiple conclusions is decomposed into multiple rules. By limiting the conclusion part to one term, the pointer information B is always limited to one for each rule. These results,
Each rule is composed of a conditional element, an operator, a conclusion element, and certain elements of pointer information B. In order to store such a rule in the memory, if a fixed length symbol is used for each item, a fixed capacity of memory may be prepared for each rule. In other words, it is possible to store in a fixed-length internal representation in the memory, improving the efficiency of memory usage and
Speed up content reference.

Generally, the condition part of a rule is composed of a logical expression consisting of a plurality of terms that logically indicate true or false. In the present invention, a complicated logical expression is decomposed into some simple expressions, and finally an expression up to a binary operation is obtained. Even when the condition part has only one term, in the present invention, the second term is always used as the condition part, and for the other one term, the content meaning that the term is meaningless is substituted or always satisfied. Substitute what you want to do.

Further, when the binary operation is an OR operation, it can be divided into two rules and the binary operation can be limited to AND. In this case, if the condition part is 2 terms, be sure to AND
Since it is an operation, the description of the operator can be omitted.

[0014]

In the inference method of the artificial intelligence system of the present invention, by using the pointer information A and B, it is not necessary to search the rule substantially, and the rule can be obtained according to the pointer information A and B. As a result, the inference operation is significantly speeded up. The effect is particularly remarkable when the number of rules is large. Furthermore, by limiting the expression of the rules and making the total length fixed, the storage in the memory becomes efficient and the inference becomes even faster.

[0015]

EXAMPLES An example of the present invention will be specifically described below. The rules of the present invention are as follows. An example of rules is shown in Table 1 by taking a system for diagnosing a cold as an example. As shown in Table 1, the conclusion part of a rule is generally represented by an expression in which one or more terms are connected by an operator. Therefore, the length of the rule is indefinite, and the evaluation of the conditional part requires the evaluation of the expression including the interpretation of the operator.

The inference operation is as follows.
Take the case of inferring the conclusion "take a cold medicine". First, we search for a rule that has a conclusion part "take a cold medicine." As a result of the search, rule 4 is applicable. Next, the condition part of rule 4 is examined. In this case, the condition part has only one item, "suspicion of cold", so the condition part is immediately evaluated. If the "suspected cold" is established, the rule 4 is immediately established, and the reason for "take the cold medicine" is successfully completed. If the presence or absence of "suspected cold" cannot be determined, a rule having "suspected cold" in the conclusion part is searched. Rule 1 applies at this stage. Rule 1 consists of three terms connected by an OR operator. In this case, the conclusion section is established if any of the three terms is established. As a result of the evaluation as described above, the process returns to the rule 4 again and the success or failure of the inference is determined. The operation described above is a typical example of backward inference. The reasoning operation requires the search of the conclusion part and the evaluation of the expression of the condition part.

In accordance with the present invention, Tables 2 and 3 are provided with pointer information A and pointer information B for the rules of Table 1.
Shown in. Take a search for a rule that has "suspected fever" in the conclusion. First, the corresponding rule number 2 is obtained from the pointer information A. Pointer information B for the second rule
And 3 is shown. That is, it can be seen that the third rule also has the same conclusion part as "suspicion of fever". Further, the pointer information B of the third rule is 0, and in this case, it can be seen that there is no more rule having a conclusion part of "suspicion of heat generation". By sequentially selecting the rules in accordance with the pointer information in this way, it is possible to reliably obtain the corresponding rule and omit unnecessary searches.

In accordance with the present invention, a rewritten version of the rule expressions in Table 1 is shown in Table 4. Since the condition part of rule 1 is three terms connected by OR, it is divided into rules 11 to 13.
Since the rules 5 and 6 are the three terms connected by AND, in this case, the intermediate conclusion 1 is newly provided, the rule 17 is first described, and further expressed as the rules 18 and 19. As described above, the conditional part can be expressed as one term or two terms of the AND operation.

Tables 5 and 6 show an example of storing the rules of Table 4 in a memory together with pointer information. Condition part is 1
Since there are two terms, the two terms are always prepared, and the condition part 2
It can be expressed by the fixed length of the term and the conclusion part 1. Since the operator in the case of the two terms is AND, this can be omitted. Since the second item when the condition part is 1 is unnecessary, leave it blank and ignore it in the actual inference. As a result, if a memory block of a certain size is used for each item, the rule has a fixed length expression of 4 items in the memory, and the conclusion and the pointer information A also have a fixed length expression of 2 items. .. The frame of one item in the table means one memory block. In general, memory is managed in consecutive addresses, so if the amount of memory required by each rule is constant, the memory address at the beginning of the rule can be easily determined simply by knowing what number rule the target rule is. It can be calculated by simple calculation, and the contents of each item can be easily referred. Therefore, the information for managing the correspondence between the memory address and the rule is unnecessary, and the use efficiency of the memory is improved together with the omission of the operator described above, and the inference speed can be increased.

[0020]

[Table 1]

[0021]

[Table 2]

[0022]

[Table 3]

[0023]

[Table 4]

[0024]

[Table 5]

[0025]

[Table 6]

[0026]

As described above in detail, according to the present invention, by using the pointer information A and B, the rule search of the inference operation can be performed at high speed, and the rule can be expressed as a specific expression. This makes it possible to use the memory efficiently and to perform inference at high speed.

Claims (2)

[Claims] 1. An artificial intelligence system that searches for a rule having a conclusion part that matches the content of an inference target from among a plurality of predetermined rules composed of a condition part and a conclusion part, and verifies the success or failure of the rule. In the inference method, the pointer information A indicating the location of the rule having the same content and the pointer information B indicating the location of the rule having the same content are provided and the rule is selected by the pointer information A. An inference method for an artificial intelligence system, characterized in that a rule having a conclusion part having the same content is sequentially selected by the pointer information B to execute inference. 2. The inference method for an artificial intelligence system according to claim 1, wherein the conclusion part is performed by using the expression form of 1 term and the condition part is performed by using the expression form of 2 terms.