JPH07334371A - Inference device - Google Patents

Abstract

PURPOSE:To efficiently and easily execute debugging or the like by accumulating the evaluated values of respective inference rules. CONSTITUTION:Concerning the inference device equipped with an inference engine 6 for performing inference based on a knowledge base, this device is provided with a control item storage part 10 for storing the evaluated values and the information of action, evaluated value storage part 11 for storing rule identification numbers, accumulated evaluated values and number of times of ignition, ignition rule storage part 12, and evaluation control part 8 for storing the rule identification number in the ignition rule storage part each time any rule is ignited, adding the inputted evaluated value to the accumulated evaluated values in the evaluated value storage part when any evaluated value corresponding to the inference is inputted, and adding the ignition this time to the number of times of ignition. Further, the inference engine is provided with a rule evaluation part 7 for advancing inference processing based on the action information when the evaluated value of the ignited rule satisfies prescribed conditions while referring to the information in the evaluated value storage part and the control item storage part each time the rule is ignited.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an inference device that can be used in expert systems and the like. At present, artificial intelligence technology is used in various fields, and a device using the artificial intelligence technology is commercialized. However, in a reasoning device (reasoning system) having specialized knowledge such as an expert system, there are many cases where the reasoning rules become enormous and the validity of the reasoning rules themselves is ambiguous.

Therefore, a lot of time and labor is spent on development and debugging at present. Against this background of the times, there has been a demand for the development of an inference device that does not require specialized knowledge and can easily perform debugging, customization, or system learning.

[0003]

2. Description of the Related Art Conventionally, various inference devices have been developed as inference devices (inference systems) used in expert systems and the like. Such an inference apparatus generally includes a knowledge base storage unit that stores a knowledge base (inference rules, etc.) and an inference engine that makes an inference based on the knowledge base.

In the inference apparatus, it is necessary to debug the knowledge base (verify inference rules) or customize (make it easy for the user to use). Therefore, conventionally, natural language is used to express the inference rules in the knowledge base to make the inference rules easy to understand (easy to understand for humans).

In addition, when debugging a knowledge base,
While executing each inference rule, the contradiction is extracted in the inference process, or the validity of the inference device (validity as the whole inference device) is tested by using a model (specific event). It was carried out by doing.

[0006]

SUMMARY OF THE INVENTION The above-mentioned conventional device has the following problems. (1): Expressing the inference rules in natural language can make them easier to understand, but does not provide a means for verifying whether the inference rules themselves are valid or not. That is, even if the inference rules are expressed in natural language, the validity of the inference rules themselves cannot be verified by this alone.

(2): When debugging or customizing a knowledge base, since inconsistencies are extracted in the inference process while executing each inference rule, it takes a lot of time and effort to debug.

(3): When debugging or customizing the knowledge base, if the validity of the inference apparatus is tested using a model, it is possible to evaluate the validity of the inference apparatus as a whole. It is difficult to evaluate the validity of each inference rule.

The present invention solves such a conventional problem and accumulates evaluation values for individual inference rules, thereby enabling debugging or customization of an inference device to be performed efficiently and easily. Is intended.

The present invention also applies to rules based on the rules of thumb of experts, intuition, and sensuous rules.
It is an object of the present invention to provide a simple evaluation means for an inference result by automatically accumulating the evaluation value of the inference rule and the cumulative number of firings.

[0011]

FIG. 1 is a diagram for explaining the principle of the present invention. In FIG. 1, 2 is a knowledge base storage unit, 5 is an input device, 6 is an inference engine, 7 is a rule evaluation unit, and 8 is Evaluation control unit, 10 is a control item storage unit, 11 is an evaluation value storage unit, 1
Reference numeral 2 denotes a firing rule storage unit.

In order to achieve the above-mentioned object, the present invention comprises a knowledge base storage unit 2 in which a knowledge base is stored, an inference engine 6 for performing an inference process based on the knowledge base, and an input device 5 for inputting data. In the inference device, a control item storage unit 10 in which information (evaluation value information and action information) of a specific evaluation value, a comparison operator, and an action is stored in advance, a rule identification number, a cumulative evaluation value, and the number of firings. (Evaluation value storage unit 11 that stores (cumulative number of firings), firing rule storage unit 12 that stores the identification number (logging data) of the fired rule, and rule identification of the rule each time the rule fires in the inference process) Number firing rule storage 1
When the evaluation value is input from the input device 5 with respect to the finally derived conclusion while being stored in 2, the input evaluation value is added to the cumulative evaluation value stored in the evaluation value storage unit 11. Add and add this firing to the number of firings (+1)
The evaluation control unit 8 is provided.

Further, the inference engine 6 refers to the information in the evaluation value storage unit 11 and the control item storage unit 10 every time a rule is fired in the process of inference, and thereby the cumulative evaluation value of the fired rule is calculated. However, when the condition of the comparison operator is satisfied, the rule evaluation unit 7 that performs the process of proceeding the inference process based on the action information is provided.

[0014]

The operation of the present invention based on the above construction will be described with reference to FIG. First, the inference engine 6 makes an inference based on the knowledge base (inference rule) stored in the knowledge base storage unit 2. If the rule fires during this inference process,
The inference engine 6 sends the identification number of the fired inference rule to the evaluation control unit 8, and the evaluation control unit 8 stores the identification number in the firing rule storage unit 12 as logging data.

Inference is performed in this way, and when the inference result is finally derived, it is evaluated whether the user or the like is appropriate and an evaluation value is calculated. Then, the user or the like inputs the evaluation value from the input device 5.

After the evaluation control unit 8 receives the input evaluation value, the evaluation value is stored in the evaluation value storage unit 11 according to the rule identification number stored in the firing rule storage unit 12.
Is added to the cumulative evaluation value in, and 1 is added to the number of firings. That is, the cumulative evaluation value and the number of firings are updated for all rules fired by inference.

The evaluation value storage unit 11 thus accumulated
By calculating the cumulative evaluation value / the number of firings based on the cumulative evaluation value and the number of firings, the average validity of the rule can be verified.

When the inference engine 6 makes an inference and the rule fires, the rule evaluation unit 7 causes the evaluation value storage unit 1 to operate.
The cumulative evaluation value and the number of firings stored in 1 are read out, and the cumulative evaluation value / the number of firings is calculated to obtain an average evaluation value for the rule currently firing.

Then, if any of the evaluation values stored in the control item storage unit 10 matches the average evaluation value (evaluation value / firing number) of the rule currently being fired, the corresponding action, for example, Decide whether to apply the rule and proceed with inference.

As described above, the validity of each rule can be evaluated by giving an evaluation value each time the inference result is output and accumulating the evaluation values. It is also possible to verify the importance of the rule by using the number of firings.

Furthermore, the data of the cumulative evaluation value and the cumulative number of firings in the evaluation value storage unit obtained as described above is verified at the time of inference, and the information in the control item storage unit is referred to to determine the rule. It is also possible to determine whether or not to apply (determine based on action information), and learning, debugging, or customization of the inference device can be performed easily and efficiently.

[0022]

Embodiments of the present invention will be described below with reference to the drawings. 2 to 5 are diagrams showing an embodiment of the present invention. In FIGS. 2 to 5, 2 is a knowledge base storage unit, 3 is an inference control unit, 4 is a storage device, 5 is an input device, 6 Is an inference engine, 7 is a rule evaluation unit, 8 is an evaluation control unit, 9 is a work area, 10 is a control item storage unit, 11 is an evaluation value storage unit, and 12 is an evaluation value storage unit.
Indicates a firing rule storage unit.

§1: Description of the configuration of the inference device--see FIG. 2 FIG. 2 is a device configuration diagram of the embodiment. The configuration of the inference device will be described below with reference to FIG. The inference apparatus of this embodiment is
It is an example of an inference device that uses specialized knowledge such as an expert system.

As shown in the figure, the inference apparatus includes a knowledge base storage unit 2 storing a knowledge base (inference rule),
An inference control unit 3 that performs various controls during inference, a storage device 4 that stores various information in the inference process, an input device 5 that inputs various information such as evaluation information, and the like are provided.

The inference control section 3 has a knowledge base (inference rule) stored in the knowledge base storage section 2.
An inference engine 6 that performs various inference processes based on the above, and an evaluation control unit 8 that controls evaluation information for inference results are provided. The inference engine 6 includes a rule evaluation unit 7 that evaluates inference rules. Is provided.

In the storage device 4, a work area 9 used as a work by the inference engine 6 in the process of inference, a control item storage section 10 for storing control items, and inference result evaluation information. An evaluation value storage unit 11 for storing and a firing rule storage unit 12 for storing logging information of a fired inference rule are provided.

As the storage device 4, for example, a magnetic disk device is used, or a storage device in which a magnetic disk device and a semiconductor memory are used in combination is used. In particular, the control item storage unit 10, the evaluation value storage unit 11, the firing rule storage unit 1
The data No. 2 is stored in a storage device (nonvolatile storage medium) such as a magnetic disk device so as not to be lost even when the power supply of the device is cut off.

§2: Explanation of format of each part in storage device ... See FIG. 3 FIG. 3 is an explanatory view of format of each part in storage device.
The figure shows the control item storage section format, the B figure shows the evaluation value storage section format, and the C figure shows the firing rule storage section format. The format of each unit in the storage device shown in FIG. 2 will be described below with reference to FIG.

(1): Explanation of format of control item storage section ... See FIG. A In the control item storage section 10, an item of "evaluation value (real number)" and "comparison operator (equal / inequality sign)" And the control item information in which three items including the item of "action" and the item of "action" are set as a set.

In this case, the evaluation value (real number), the comparison operator, and the action data are data stored in advance before the normal operation of the apparatus, and then read by the rule evaluation unit 7. Used.

In this case, the comparison operator is data in which the condition or the like for the evaluation value (real number) is expressed by an equal sign or an inequality sign, and the action is an action when the condition expressed by the comparison operator is satisfied (for example, , For the rule,
Action information such as unusable, caution, and warning) is stored.

As a concrete example, for example, the evaluation value item is "0.1", the comparison operator item is "≥", and the action item is "unusable". In this example,
When the condition of 0.1 ≧ x is satisfied for the evaluation value x to be compared, the rule of the evaluation value x is disabled.

Further, the evaluation value item is described as “0.5”, the comparison operator item is described as “≧”, and the action item is described as “warning”. In this example, when the condition of 0.5 ≧ x is satisfied for the evaluation value x to be compared, the evaluation value x
Rule is a warning (not appropriate).

The evaluation value item and the comparison operator item are put together, and these items are, for example,
It is also possible to describe “0.1 <x ≦ 0.5” (x: evaluation value) and describe the action item like “warning”. In this example, for the evaluation value x, 0.5 ≧ x> 0.
When the condition of 1 is satisfied, the rule is a warning (not appropriate).

(2): Explanation of format of evaluation value storage unit
.. see FIG. B The evaluation value storage unit 11 has an item of "rule identification number", an item of "cumulative evaluation value", and an item of "number of firings". The item "rule identification number" stores the identification numbers of all rules, the item "cumulative evaluation value" stores the cumulative evaluation value, and the item "number of firings" stores the cumulative evaluation value. Stores the number of firings.

Each data in the evaluation value storage unit 11 is initially cleared (cleared to 0), and then an inference result is obtained by firing the rule.
When the evaluation value is input from the outside, the evaluation control unit 8 updates the data of the cumulative evaluation value and the number of firings.

(3): Explanation of format of firing rule storage unit ... See FIG. C In the firing rule storage unit 12, when a rule is fired in the process of inference, the identification number of the fired rule is used as logging data. It is to be stored sequentially.

For example, if rule A fires for the first time,
For example, 01 is stored as the identification number of the rule A. Then, when the second rule C is fired, for example, 03 is stored as the identification number of the rule C. Thus, each time the rule is fired, its identification number is sequentially stored as logging data.

§3: Operation of the inference apparatus ... See FIG. 2 The operation of the inference apparatus shown in FIG. 2 will be described below. First,
The inference engine 6 sequentially reads out the inference rules stored in the knowledge base storage unit 2 and makes an inference. When the inference rule is fired in the process of this inference, the inference engine 6 sends the identification number of the fired inference rule to the evaluation control unit 8, and the evaluation control unit 8 sends the identification number of the rule to the firing rule storage unit 12 as logging data. Store.

In this way, the inference is sequentially performed, and when the inference result is finally derived, for example, it is evaluated whether or not the user is appropriate, and the evaluation value is calculated. Then, the user inputs the evaluation value from the input device 5.

After the evaluation value inputted is received by the evaluation control unit 8, the evaluation value is stored in the evaluation value storage unit 11 according to the rule identification number stored in the firing rule storage unit 12.
It is added to the cumulative evaluation value in, and 1 is added to the number of firings. That is, the cumulative evaluation value and the number of firings (cumulative firing number) in the evaluation value storage unit 11 are updated for all rules fired by inference.

Evaluation value storage unit 11 accumulated in this way
By calculating the cumulative evaluation value / the number of firings based on the cumulative evaluation value and the number of firings, the average validity of the rule can be verified.

When the inference engine 6 makes an inference and a rule is fired, the rule evaluation unit 7 causes the evaluation value storage unit 1 to operate.
By reading the cumulative evaluation value and the number of firings (cumulative firings) stored in 1 and calculating the cumulative evaluation value / the number of firings (cumulative firings), the average of the rules currently firing is calculated. Obtain an evaluation value.

If any of the evaluation values stored in the control item storage unit 10 matches the average evaluation value (evaluation value / number of firings) of the rules currently being fired, the rule is selected from the corresponding action. To determine whether to proceed with inference. With such processing, the inference device can be automatically learned.

As described above, the validity of each rule can be evaluated by giving an evaluation value each time the inference result is output and accumulating the evaluation values. It is also possible to verify the importance of the rule by using the number of firings.

With the data thus obtained, it becomes possible to decide whether or not to apply the rule by extracting the rule to be debugged and verifying the evaluation value at the time of inference. It is also possible to customize the inference apparatus (to meet the user's request) or to learn.

§4: Processing explanation of the embodiment by a concrete example ...
-Refer to FIG. 4. FIG. 4 is a process explanatory diagram 1 of the embodiment, and FIG. 5 is a process explanatory diagram 2 of the embodiment. The processing of the embodiment according to the concrete example will be described below with reference to FIGS. 4 and 5.

A large number of inference rules are stored as a knowledge base in the knowledge base storage unit 2 shown in FIG.
An identification number is attached to each of these inference rules. For example, as shown in FIG. 4, the inference rules are AG
The rule A has an identification number of 01, the rule B has an identification number of 02, the rule C has an identification number of 03, the rule D has an identification number of 04, the rule E has an identification number of 05, and a rule F.
The identification number is 06, and the identification number of rule G is 07.

The evaluation value storage unit 11 stores all the rule identification numbers, the cumulative evaluation values corresponding to them, and the respective data of the number of firings (cumulative firing number). In FIG.
Data before and after evaluation are shown.

As data before evaluation, in the case of this example,
The cumulative evaluation value of the rule with identification number 01 is 1.0, the number of firings is 1, the cumulative evaluation value of the rule with identification number 02 is 2.4, the number of firings is 2, and the cumulative evaluation value of the rule with identification number 03 is 0.
5, the number of firings is 1, the cumulative evaluation value of the rule of identification number 04 is 0.7, the number of firings is 1, the cumulative evaluation value of the rule of identification number 05 is 3.2, the number of firings is 4, the rule of identification number 06 Has a cumulative evaluation value of 1.1, the number of firings is 2, the rule of identification number 07 has a cumulative evaluation value of 0.3, and the number of firings is 1.

In this state, the inference engine 6
Inference is performed by extracting inference rules from the knowledge base. In this case, for example, rule A, rule C, rule F
Is ignited and a certain conclusion is obtained. In this case, the inference engine 6 sends the identification numbers of the fired inference rules A, C, and F to the evaluation control unit 8, and the evaluation control unit 8 stores them in the firing rule storage unit 12.

That is, at this time, the firing rule storage unit 12
Is 0, which is the identification number of rules A, C, and F that have fired.
1, 03, 06 are stored as logging data. Here, since a certain conclusion is obtained by the firing of the rules A, C, and F, the user or the like evaluates this conclusion. If, for example, the evaluation value is 0.7 according to this evaluation, the user or the like inputs 0.7 as the evaluation value from the input device 5.

The input evaluation value is received by the evaluation control unit 8, and based on this evaluation value 0.7, the evaluation value storage unit 11
Update the evaluation value stored in. This evaluation value is updated as follows. In this case, rule A that fired,
Since C and F are rules for data update, the data of the identification numbers 01, 03, 06 corresponding to these rules will be updated.

Therefore, the evaluation control unit 8 updates the data in the evaluation value storage unit 11 based on the evaluation value 0.7 input from the outside, as shown in the drawing (see data after evaluation). In this case, according to the rule of the identification number 01, a value 1.7 (1.
0 + 0.7 = 1.7) as the cumulative evaluation value after the update, and the value 2 which is the number of firings 1 before the evaluation plus the number of firings 1 newly 2
Let (1 + 1 = 2) be the number of firings after updating.

In the rule of the identification number 03, the value 1.2 which is obtained by adding the new evaluation value 0.7 to the cumulative evaluation value 0.5 before the evaluation.
(0.5 + 0.7 = 1.2) is the cumulative evaluation value after the update, and the value 2 (1 + 1 = 2) obtained by adding 1 to the number of new firings before evaluation is the number of firings after the update. And

According to the rule of the identification number 06, a value 1.8 obtained by adding a new evaluation value 0.7 to the cumulative evaluation value 1.1 before the evaluation.
(1.1 + 0.7 = 1.8) is the cumulative evaluation value after updating, and the value 3 (2 + 1 = 3), which is the number of firings 2 before evaluation, is added to the number of firings 1 newly, is 3 (2 + 1 = 3). And

After that, when the rule is fired and a certain conclusion is reached, the above-described processing is repeated, and then the evaluation is performed and the data in the evaluation value storage unit 11 is updated. For the data of the evaluation value storage unit 11 thus obtained, the average validity of the rules can be verified by, for example, calculating the cumulative evaluation value / the number of firings.

Further, the average validity of the rules can be used to search for a rule to be debugged, and when the threshold of the average validity and a rule matching the threshold are fired. It is also possible to provide an action to be taken by the inference engine 6 in the control item storage unit 10 and to perform inference processing by adding control so that a rule with low validity is not fired.

[0059]

As described above, the present invention has the following effects. (1): When operating the inference apparatus, the user or the like can evaluate each rule by giving an evaluation only to the conclusion without being aware of the inside of the rule. Therefore, it is possible to easily debug or customize the inference apparatus.

(2): The evaluation value of the inference rule and the number of firings can be automatically accumulated even for a rule based on an empirical rule of an expert, an intuition, or a sensory rule. Therefore, a simple evaluation means for the inference result can be provided.

(3): Logging data of the fired rule,
Also, by accumulating the evaluation value for each inference rule and the cumulative number of firings, debugging or customization of the inference device can be efficiently and easily executed. Also,
It is also possible to verify the importance of the rule using the number of firings.

(4): By accumulating the evaluation value of the inference rule and the number of firings, for example, the cumulative evaluation value / the number of firings is calculated to verify the average validity of the rule. You can Also, using the accumulated evaluation value,
It can also be used during debugging, such as extracting a rule of a specific evaluation value, and efficient debugging and customization of the inference device can be easily performed.

(5): When inference is performed and the rule is fired,
Any of the evaluation values stored in the control item storage unit is the average evaluation value of the currently firing rules (evaluation value / number of firings)
If it matches with, it is possible to decide whether to apply the rule and proceed with the inference from the information of the corresponding action. With such processing, learning of the inference apparatus can be automatically and efficiently performed.

(6): It is possible to realize an inference device that does not require specialized knowledge and can easily perform debugging, customization, or system learning.

[Brief description of drawings]

FIG. 1 is a diagram illustrating the principle of the present invention.

FIG. 2 is a device configuration diagram of an embodiment.

FIG. 3 is an explanatory diagram of a format of each unit in the storage device according to the embodiment.

FIG. 4 is a process explanatory diagram 1 of the embodiment.

FIG. 5 is a process explanatory diagram 2 of the embodiment.

[Explanation of symbols]

 2 Knowledge base storage unit 5 Input device 6 Inference engine 7 Rule evaluation unit 8 Evaluation control unit 10 Control item storage unit 11 Evaluation value storage unit 12 Firing rule storage unit

Claims (2)

[Claims] 1. An inference device comprising a knowledge base storage unit (2) storing a knowledge base and an inference engine (6) for performing an inference process based on the knowledge base, which is fired to draw a certain conclusion. A firing rule storage unit (12) for storing the logging data of the rule, and a cumulative evaluation value indicating the validity of the rule for each rule,
And an evaluation value storage unit (11) for storing the cumulative number of firings, and by the inference processing by the inference engine, each time the rule fires, the logging data of the rule is stored in the firing rule storage unit and finally derived. When an evaluation value is input from the outside in response to the conclusion, the input evaluation value is added to the cumulative evaluation value for the data stored in the evaluation value storage unit, and the cumulative number of firings is calculated as follows. An inference apparatus comprising an evaluation control unit (8) for adding (+1) firing. 2. The inference apparatus according to claim 1, wherein the control item storage unit stores specific evaluation value information and information about an action to be performed by the inference engine when a rule satisfying the condition of the evaluation value information is fired. (10) is provided, and the inference engine (6) evaluates the fired rule by referring to the information in the evaluation value storage unit and the control item storage unit every time the rule fires in the process of inference. An inference apparatus, comprising: a rule evaluation unit (7) for performing inference processing based on the action information when a value satisfies the condition of the evaluation value information.