JPH05265757A - Inference control system - Google Patents

Abstract

PURPOSE:To omit the undesired inferences, to prevent the increase of the undesired processing time, and to omit the undesired inference results by teaching previously the rules that should be and should not be satisfied during the inferences to an inference system. CONSTITUTION:An inference part 180 stores the contents of the initial data 100 in its built-in working memory and evaluates the rule conditional part included in a knowledge base 104 based on the information on the data 100. The rule having its satisfied conditional part is sent to an inference control part 118. The part 118 checks whether the received rule should be collated with the rule written into the control information 122 or not. If so, the part 118 instructs the part 108 to disuse the inference. If not, the part 108 must carry on the inference. Then, the part 118 teaches the rules that should be and should not be satisfied during the inferences to an inference system together with the time sequential relation. Thus, the inference is disused if it is known that the conditional part of a relevant rule or the rule itsels is not satisfied during the inference.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an inference control method for a system that infers temporal changes, and in particular, it speeds up inference by reducing unnecessary inference by using facts known in advance, so that inference results can be obtained. It relates to an inference control method for reducing ambiguity.

[0002]

2. Description of the Related Art In a forward reasoning such as a rule-based reasoning system, a condition regarding a quantity is written in a condition part of a rule, and it is not known whether the condition is satisfied or not satisfied due to lack of information. In that case, the most common way is to issue an error. OPS83 (artificial intelligence language OPS83, CLFo) which is a typical expert system
Such inference mechanism is adopted in ES / KERNEL (Hitachi Creative Workstation 2050ES / KERNEL Grammar, Hitachi, 2050-3-602-20) by rgy, Personal Media Corporation.

However, in a more advanced inference system, inference is performed by branching into a case where a condition that cannot be evaluated is satisfied and a case where the condition is not satisfied. For example, suppose the following rule is written.

IF X> 0 then Z = X + Y When it is found that X> 0 cannot be evaluated using the information in the working memory, the working memory is copied,
Assuming that X> 0 holds and X = <0, that is, X> 0 does not hold, the inference is branched and the inference is continued. If the inference continues and a contradiction occurs, the inference is rejected. However, if no contradiction occurs, the reasoning is not rejected and the inference is continued until the evaluation of the conditional part of all rules is confirmed. Similar to the case where the condition relating to the quantity is not evaluated in the conditional part of the rule, the same thing occurs when a plurality of conditions are simultaneously fired and the inference is performed in parallel without performing conflict resolution. However, in some applications and debugging, it may be possible or predictable in advance that a certain rule should be established or a certain rule should not be established. Traditional rule-based reasoning systems do not use such information when it is available. Therefore, there arises a problem that it is not possible to reject useless inference on the way and it takes extra processing time, or a plurality of inference results are output and the result becomes ambiguous.

One method of inference control using the fact that a certain rule should not be established is considered in the framework of hypothetical reasoning. In TMS and ATMS (recognition and learning, pages 282 to 290, Iwanami Shoten), which are typical of conventional hypothetical reasoning systems, cause inconsistency and reject useless reasoning according to conditions specified in advance. Was being conducted.

For example, in ATMS, the combination of the node a representing the proposition a and the node b representing the proposition b is represented as a contradiction like node a and node b-> contradiction. Given such information,
The system generates a database such as NOGOOD {{a, b}}, and if the conditions of the inference rule include a and b, then reject the inference using that rule. Since both TMS and ATMS are systems built on propositional logic, for conditions that include variables,
Cannot specify or detect the combination of inconsistencies. For example, a combination of a node a representing the predicate a (X) and a node b representing the predicate b (X) cannot be specified as a (X) and b (X) −> contradiction. However, since TMS and ATMS deal with hypotheses explicitly and are based on propositional logic, it is possible to efficiently find a reasonable solution using another hypothesis in addition to rejecting inference.

Prolog, which is one of the backward inference systems
Then, to reject the inference using a certain clause, write a (X) and b (X) − by writing p (a (X), b (Y)):-!, fail.
> I can express that it is close to contradiction. However, even with Prolog, "p (a (X), b (Y)):-!, fail." Needs to be added to the global database, and in order to control each problem differently, You have to change the global database each time you solve. Further, the change of the global database requires an additional work of finding the changed portion, which is a burden on the user as compared with the case where the changes are put together as control information. Moreover, such changes may not be possible to the user if the global database is provided in compiled form for efficient execution. Prolog generally cannot add another section if the global database is compiled without the source program.

On the other hand, for a system that infers temporal changes, a designation method that includes a temporal order relation is conceivable in addition to the designation that a rule should not be established or a rule should be established. However, in the conventional system that infers temporal changes, such a thing was not dealt with.

A qualitative inference system, which is a system for inferring temporal changes, is generally used to perform simulations and predict and explain temporal changes in the system. At that time, before starting the simulation, there are cases where, during inference, an event or rule that should be established including the temporal order relation, or an event or rule that should not be established including the temporal order relation may be known. This is the case when experiments are being conducted at the same time as the simulation, and some facts can be seen from the results of the experiments.

For example, the qualitative inference system Qupras (qualitative inference based on the laws of physics, qualitative inference, Chapter 8, Kyoritsu Publishing)
Then, the state at a certain time consists of some rules corresponding to the fundamental knowledge. A rule corresponding to the physical law that a light bulb shines when electricity flows can be expressed as follows.

Physical lighting objects Light_bulb-light_bulb; conditions temperature @ Light_bulb> = light_generating_temperature @ Light_bulb; relations ddt (light_generation @ Light_bulb) = ddt (temperature @ Light_bulb); end. rela
IF indicating that the expression written in tions is valid
-THEN format rules. This rule applies to light bulbs (ligh
t_bulb) temperature (light_generating_temperature @ Ligh
t_bulb), the light output of the bulb (light_gene
ration @ Light_bulb) change (enclosed by ddt) means that it is proportional to the change in temperature. This physical law is similar to the rule-based reasoning system, but in qualitative reasoning, the target system can also infer temporal changes.

Therefore, when information such as an event that should occur or an event that has not occurred can be obtained from the actual experimental result including the temporal order relation, unnecessary inference can be deleted using the information. it can. For example, there is a case where it can be experimentally determined that the battery has run out after the light bulb shines. Even if such information is known in advance, if the inference system infers a temporal change in which the battery dies before the light bulb shines, such inference is an unnecessary inference and the result is , The processing time increases and the inference result becomes ambiguous. The reason why such a thing occurs in the qualitative inference system is that when the information of the initial data is insufficient and given, the quantity is treated qualitatively.

In the conventional system for inferring temporal change, even if the rule to be established is known in advance, it is not used even when such information is valid. Therefore, there was a problem that unnecessary inference could not be rejected on the way and extra processing time was required, or multiple inference results were issued and the results were ambiguous. The cause of such a problem is that the information about the rule that should be used during the inference cannot be described in advance in the IF-THEN format rule.

Further, in the conventional system for inferring temporal change, there is no means for validating only inference when a specified rule to be established is used. In particular,
In simulations that explain temporal changes, some things may be known from observed data.
For example, it is experimentally found that the battery ran out after the bulb lit as shown above. In this case, as a result of the inference, if it is not inferred that the battery is exhausted after the light bulb lights up, the inference should be rejected, but the inference result is determined by a human being and the selection is made. Also, it may be found in advance that, after a rule fires, no inference can occur that the battery will run out after the bulb shines. Even in such a case, since the conventional inference system does not control the inference using such information, there is a problem that unnecessary inference is performed or processing time is increased. ..

[0015]

In a system that infers changes over time, even if a rule that should not be established or a rule that should be established is known in advance, such information is not used, so it is unnecessary. There is a problem that processing time is increased and unnecessary inference results are obtained.

An object of the present invention is to control inference using such information without changing the knowledge base when a rule that should not be established or a rule that should be established is known or expected in advance. To reduce the processing time and the ambiguity of inference results.

[0017]

A knowledge base, which is a set of rules having a condition part and a consequent part that becomes effective when the condition is satisfied, and a current state A rule (or rule condition) that has a working memory that stores a rule and that should not be satisfied in the inference control method that executes the consequent part, including temporal order relations, if the rule condition is satisfied Is specified, and if a rule that should not be established is established during inference, by adding a means for rejecting that inference, it is possible to prevent unnecessary inference and unnecessary processing. Achieves no increase in time and leads to unnecessary inference results. There are the following three methods for specifying the rule name.

(1) Rule Name This specifies a rule name such as "Rule 1".

(2) Rule name and condition set This is a rule name added to the rule name of (1). For matching the conditions, there are a pattern matching method and a unifying method as shown in (3).

(3) Condition This is not a rule name but a list of conditions used in the rule. For example, write p (X, a), q (X, b) r (a, Y, c). “P (X, a), q (X, b)” is a combination of two terms, and “r (a, Y, c)” is one term. Corresponds to a rule condition that should not hold. Here, capital letters indicate variables. As a method of collating the conditions listed in the control information with the conditions of the rule that should be satisfied or the conditions of the rule that should not be satisfied, unification is possible (Basics of artificial intelligence, pages 120 to 1
(Page 24, Modern Science Co., Ltd.) or inclusive relations are possible. Inclusive relations are possible as long as the listed conditions can be unified with a part of the conditions of the rules that should be satisfied or the conditions of the rules that should not be satisfied.

Further, means for rejecting the inference when the rule to be established including the temporal order relation is specified during inference and it is found that the specified rule to be established is not established during inference. By adding, it is possible to reject the inference in the case where the rule that should originally be established is not established, and to achieve unnecessary inference.

[0022]

[Function] During inference, by telling the inference system rules that should not be satisfied or rules that should be satisfied, including the temporal order relation, the condition part of the rule may be satisfied during inference, or the rule may be satisfied. If it turns out that does not hold, reject the inference.

[0023]

FIG. 2 is a block diagram of an inference control device according to an embodiment of the present invention. An outline of the configuration and operation of the entire device will be described.

The device comprises a control device (CPU) 200, a main memory 206, a magnetic disk 208, a CRT 202 and a keyboard 204. Knowledge base 104 and initial data 10
0 and the control information 122 are stored in the magnetic disk device 208. The programs of the inference unit 108 and the inference control unit 122 are also stored in the magnetic disk device 208. The keyboard 204 is used for program instructions. The CRT 202 is used to display the result.

If necessary, the programs of the inference unit and the inference control unit, the contents of the knowledge base, the initial data, and the control information are loaded into the main memory 206. This operation is performed from the keyboard 204. The keyboard 204 also designates a command for calling a program of the inference unit or the inference control unit into the main memory 206 and an execution start command. The inference process and results are displayed on the CRT 202.

The present invention will be specifically described below. First,
The outline of the embodiment will be described. FIG. 1 shows the entire program and data structure. The inference unit 108 stores the contents of the initial data 100 in a working memory in the inference unit, and uses the information in the working memory to evaluate the conditional part of the rule in the knowledge base. The rule satisfied by the condition section is transmitted to the inference control section 118, and the inference control section 118 checks whether the rule matches the rule written in the control information 122. If so, instruct the inference unit to reject the inference, otherwise to infer.

The final result is found at 112. Knowledge base 1
It is possible to control the inference using the control information 122 without changing 04.

A process of the inference unit for rejecting an inference when a rule that should not be established is specified and the specified rule is established during inference will be described with reference to FIG. The inference unit takes out rules one by one from the knowledge base (302) and evaluates the condition part (304).

If the condition part of the rule is satisfied (306), the process of the inference control part is performed (308). If not satisfied, check 310 if none of the rules in the knowledge base were satisfied. If so, the process ends (314). Otherwise, another rule is retrieved from the knowledge base (302). The instruction from the inference control unit is checked (312), and if the result is rejection, the inference ends. If not, continue reasoning. The reasoning method of the reasoning unit is the same whether the reasoning does not include a temporal change or the temporal change is inferred. The difference between the inclusion and the exclusion of the temporal change is related to whether the knowledge base rule has a description that causes the temporal change and the inference system infers the temporal change according to the description. The inference method of temporal changes is qualitative reasoning (qualitative reasoning based on physical laws, qualitative reasoning, Chapter 8, Kyoritsu Shuppan), and temporal logic (time logic and artificial intelligence, artificial intelligence and Logic, Chapter 6, Kyoritsu Shuppan).

Before showing the means for realizing the inference including the temporal change, the means for realizing the inference not including the temporal change will be described. The process of the inference control unit that performs the process 308 in the inference that does not include a temporal change will be described with reference to FIG. The inference control unit compares with the contents of the control information storing the rule name that the rule passed from the inference unit should not be established (402),
If it can be matched with the items in it, the rejection result is returned (40
6) If not, the result of continuation is returned (404).

In the collation (402) of the condition part, the control information may be not a rule name but a condition including a variable like the type (3) among the three methods described above as the rule name designating method. .. The processing will be described with reference to FIG. 7. Process 700 is process 308 instead of process 400.
Called from. When the rule is established, the processes 308 to 700 are called. In process 702, the conditions of the established rule are extracted. In process 704, it is checked whether a part of the fetched conditions can be unified with the condition in the control information. If so, the rejection result is returned (708), and if not, the continuation result is returned (706). For example, if the following condition is set in the control information, the condition of the rule that p (X, a), q (X, b) is satisfied is rule 1 (p (c, a), q (c , B)) and rule 2 (p (c, a), q (c, b), r (e)), the matching succeeds. On the other hand, if the established rule is rule 1 (p (c, a), q (d, b)), the collation will not succeed. One method of this matching is unification (Fundamentals of Artificial Intelligence, pages 120-124).
Page, Modern Science Company). Further, in the control information that enumerates only the conditions (the type of (2) among the three methods described above as the rule name specifying method), the condition part of the rule is also used for matching in addition to the rule name. Unification is also one of the methods for checking the condition parts in this case.

A rule to be established in inference that does not include a temporal change is designated, and the process of rejecting the inference when it is found that the specified rule is not established during inference is the same as the process of FIGS. 3 and 4. Is similar. The processing of the inference unit will be described with reference to FIG. The processing from 500 to 512 is the same as the processing from 300 to 312. However, the process of 508 calls the inference control unit 600 of FIG. At 514, when all the inferences are completed, the inference control unit 6 is restarted.
Call 00. This is to check whether the rule that should be established at the time when the inference is completed is established. Fifty
The process of the inference control unit that performs the process of 8 will be described with reference to FIG. The inference control unit checks whether or not the rule passed from the inference unit knows that the rule that should be satisfied and the specified rule are not satisfied (602), and if so, returns a rejection result (606). If not, the result of continuation is returned (604). The process of the inference unit of the process of designating the rule to be established in the inference including the temporal change and rejecting the inference when it is found that the specified rule is not established during the inference is the same as that of FIG. ..

As an extension of the means 602 for checking whether or not a rule to be established and a specified rule in inference that does not include temporal change are established, if one rule is established, another rule is not established. If known, a process of using such information as control information will be described with reference to FIG. Control information is control information 1
And the contents of control information 2 are stored. In the control information 1, a rule for control is written such that "when a certain rule is satisfied, another certain rule is not satisfied". The control information 2 describes the conditions to be met. For example, in the control information 1, when the rule 1 is established, the rules 2 and 4 are not established.

It is assumed that such knowledge is written. It is assumed that the control information 2 is stored as rule 2 as a rule to be established. Process 800 is process 6
It is called from the process 508 similarly to 02. In process 802, the rule name of the established rule is extracted. Control information 1
It is checked whether there is a control rule stored in the control information 1 having the rule name established by using as a condition. If not, the result is returned as a continuation. If there is, the rule name written in the consequent part of the control rule is taken out (806). Then, it is checked whether or not there is a rule written in the control information 2 in the extracted rule name (808). if,
If it is specified in it, it can be seen that the rule written in the control information 2 does not hold, so it makes no sense to infer further. Therefore, the rejection result is returned (812). If it is not included therein, the result of continuation is returned (810). As described above, it is possible to reject the inference on the way because it is found that a rule that should be satisfied is not satisfied because a certain rule is satisfied even during inference. In addition to the rule name, the collation here can be performed using the types (2) and (3) of the three methods described above as the rule name designation method. When the control rule written in the control information 1 is established, the rule name written in the consequent part is registered in the control information as a rule that should not be established. This can be used by the method shown in FIG. 4 to check whether a rule that has been established later should not be established and reject the inference.

Next, an example of a system having a mechanism for inferring temporal change will be described. In a system that infers temporal changes, rules that should not be established including temporal order relations are specified, and if the rules that should not be established specified by that order relation are established during inference, the inference is rejected. The means for doing so can be easily realized by expanding the processing of FIG. The processing will be described with reference to FIG. The process 900 is called from the process 308 similarly to the process 400. In process 902, first, the name of the established rule is recorded. Next, it is checked whether the partial sequence of the recorded sequence matches the rule sequence specified in the control information (904).
If the matching is successful, such a sequence has an order relation that should not hold, and the result is returned as a rejection (908). If matching fails, the result is returned as continue (906).

For example, when rule 3 is recorded as a rule in which rule 2 is established and rule 1 is established, it is recorded as a sequence of rules in which rule 3, rule 2 and rule 1 are established. If the control information specifies that the inference that the rule 1 is established after the rule 3 is rejected, the process 90 is executed when the rule 1 is established.
4, the collation succeeds and the rejection result is returned.

In a system having a mechanism for inferring temporal changes, rules to be established including temporal order relations are specified, and rules to be established specified by the order relations are not established during inference. The means of rejecting the inference at the time when it becomes clear can be easily realized by expanding the processing of FIG. 6 or FIG. The process obtained by expanding FIG. 8 will be described with reference to FIG. Process 1000 is process 8
It is called from the process 508 similarly to 00. In process 1002, first, the name of the established rule is recorded. Next, it is checked whether the recorded subsequence of the sequence matches the condition part of the rule specified in the control information 1 (1004). If the matching fails, the result is returned as a continuation (1010). If the matching is successful, the consequent part of the rule is extracted (100
6) It is checked whether or not the extracted rule name has a rule name for which the control information 2 should be satisfied (1008). If so, the rule of the control information 2 is such a control rule that "when a certain rule is established, another certain rule is not established", the rule is not established and the result is rejected ( 1012). If not, the result is 1010. For example, when rule 3 is recorded as a rule in which rule 2 is satisfied and rule 1 is satisfied, it is recorded as a sequence of rules in which rule 3, rule 2, and rule 1 are satisfied. If the control information 1
"If rule 1 is established after rule 3, rule 4
If the rule that "is not satisfied" is written, it is understood that rule 4 is not satisfied. If the rule 4 is described in the control information 2 as a rule to be established, it is understood that the rule 4 is not established when the rule 2 is established, so the inference is rejected. Since it can be seen that rule 4 is not established when rule 1 is established after rule 3, therefore, it is separately registered in the control information as a rule name to be rejected when it is established, as in the process of FIG. This can be used as a method of rejecting inference by the method shown in FIG.

As for the method of designating the rule name in the inference that changes with time, only the rule name, the rule name and the condition, the combination of the conditions, and the like can be specified as in the inference that does not perform the time change.

[0039]

According to the present invention, the knowledge base is not changed and the inference system is informed of the rules that should not be established or the rules that should be established during the inference as control information. If a rule is established or it is found that the rule is not established, rejecting the inference at that point makes it possible to avoid unnecessary inference, and not to increase unnecessary processing time. It is possible to avoid leading inference results.

[Brief description of drawings]

FIG. 1 is a block diagram of an inference system showing an embodiment of the present invention.

FIG. 2 is a configuration diagram of an inference control device of the present invention.

FIG. 3 is a flowchart showing a rule evaluation process.

FIG. 4 is a flowchart for collating rules that should not be established in rule-based reasoning.

FIG. 5 is a flowchart showing a process of evaluating rules of an inference unit in rule-based inference.

FIG. 6 is a flowchart for checking rules to be established.

FIG. 7 is a flowchart when a condition is used to check a rule that should not be established.

FIG. 8 is a flowchart for checking a rule to be established using a control rule.

FIG. 9 is a flowchart for collating rules that should not be established in inference that changes with time.

FIG. 10 is a flowchart for checking a rule to be established using a control rule in inference that changes with time.

Front page continuation (72) Masahiro Abe, Inventor Masahiro Abe 1-280, Higashi Koigokubo, Kokubunji, Tokyo Inside the Central Research Laboratory, Hitachi, Ltd.

Claims (3)

[Claims] 1. A knowledge base, which is a set of rules having a mechanism for inferring a temporal change and having a condition part and a consequent part which becomes effective when the condition is satisfied, and a working for storing a current state. If there is a memory and the condition of the rule is satisfied, in the inference control method that executes the consequent part, the method that specifies the rule that should not hold including the temporal order relation and the order relation An inference control method having means for rejecting inference when a rule that should not be established is established during inference. 2. A knowledge base, which is a set of rules having a mechanism for inferring a temporal change and having a condition part and a consequent part which becomes effective when the condition is satisfied, and a working for storing the current state. If there is a memory and the condition of the rule is satisfied, in the inference control method that executes the consequent part, the means that specifies the rule that should be satisfied including the temporal order relation and the order specified by the order relation An inference control method having means for rejecting inference when it is found that a power rule is not established during inference. 3. In the inference control method according to claim 1 or 2, as means for designating a rule that should be established or a rule that should not be established during inference, (1) rule name, (2) rule name And its condition part or (3) condition can be specified. (1) Matching by rule name in case of specifying by rule name, (2) Rule name in case of specifying by rule name and its condition Match and match condition,
(3) An inference control system characterized by having a means for collating a part of the conditions of the rule to be collated in the case of designation by a condition.