JPH07271602A - Method for inputting and editing dynamic knowledge of expert system - Google Patents

Abstract

PURPOSE:To easily and speedily constitute an expert system and improve the productivity of the expert system by generating input candidates by systemized static knowledge, calculating the similarity between the input candidate names and dynamic knowledge names, and displaying them in the decreasing order of the similarity. CONSTITUTION:Knowledge names, versions, etc., are registered (1-1) in a knowledge base every object problem. Then, the static knowledge is inputted and edited (1-2). Those are used for the dynamic knowledge. Then, inference control knowledge determining the structure and inferring method of the dynamic knowledge is inputted and edited (1-3). Next, the dynamic knowledge names is inputted and edited (1-4), and furthermore, a similarity degree calculation is performed (1-5). The dynamic knowledge names are finally inputted and edited (1-6) to constitute the expert system. Consequently, the static knowledge which is previously inputted becomes reusable and selected candidates are weighted to display highly relative candidates first, thereby reducing the possibility of scrolling.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a knowledge input / editing method for constructing an expert system having a large amount of static knowledge and dynamic knowledge, such as an expert system for blast furnace guidance in the steel industry. .

[0002]

2. Description of the Related Art There is already a technology for inputting knowledge of a forward reasoning expert system using a tabular editor (Japanese Patent Laid-Open No. 04-190425). It systematizes the knowledge representation method of the expert system into four types of inference control knowledge, dynamic knowledge, static knowledge and uncertain knowledge, and tabular editor screens (rules, metarules) for inputting knowledge about each. , Frame, fact, fuzzy, constant, data with certainty factor, external function) are used to input / edit knowledge. In input editing, static knowledge is input first, and then static knowledge is input, and then dynamic knowledge is input.

[0003]

DISCLOSURE OF THE INVENTION Problems to be solved by the invention include:
It is necessary to incorporate and describe part of the static knowledge that has already been entered. However, the static knowledge that was already input at this time was input again. That is, after entering static knowledge,
To enter dynamic knowledge that incorporates static knowledge,
I had to write static knowledge again. For example,
About 2,000 knowledge of blast furnace guidance expert system in steel works, static knowledge is about 5 pieces and dynamic knowledge is about 5 pieces.
It is not uncommon to have an expert system with around 000 units. In an expert system that uses a large amount of static knowledge and dynamic knowledge in this way, rewriting static knowledge when inputting dynamic knowledge induces input error or redundancy. Also, just displaying static knowledge
It is very difficult to find a suitable candidate from a large amount of static knowledge, which is inefficient in constructing an expert system.

The present invention is to improve the conventional knowledge input method of the expert system, and it is an object of the present invention to facilitate and speed up the construction of the expert system and improve the productivity of the expert system.

[0005]

[Means for solving the problem] The knowledge representation method of the expert system is systematized into inference control knowledge, dynamic knowledge, static knowledge and uncertain knowledge, and a general-purpose tabular editor is used to construct the expert system. Use the screen.
For inference control knowledge (metarule), enter / edit the rule execution control procedure from the metarule editor screen. Static knowledge (facts, frames, constants, external functions)
The declarative knowledge that describes facts, work data, etc. from the Fact Editor screen, the declarative knowledge that describes a hierarchical data structure with data inheritance from the Frame Editor screen, and the declaration that describes constants from the Constant Editor screen. Input / edit the definition of the external function used in the rule from the external function editor screen. As uncertain knowledge (fuzzy, confidence data), declarative knowledge describing uncertain facts, work data, etc. is entered and edited from the fuzzy editor screen. As dynamic knowledge,
From the Rule Editor screen, enter / edit rules (rules) regarding how to operate static knowledge.

According to the present invention, in the method of constructing an expert system by using the above-mentioned tabular editor screen, when inputting / editing dynamic knowledge including static knowledge, the input / editing operation is performed. Considering that there is a correlation between the name of static knowledge and the name of static knowledge used in the dynamic knowledge, and the static knowledge used in the dynamic knowledge is limited to some extent, (A ) An input candidate is created for each systematic static knowledge, (B) the similarity between the input candidate name and the dynamic knowledge name is calculated, and displayed in descending order of similarity, (C) expert system Let the builder select and enter.

[0007]

According to this, (1) the input candidates are limited by creating the input candidates for each systematized static knowledge, and (2) the input candidate name and the dynamic knowledge name. Since the static knowledge is displayed in descending order of similarity, the expert system builder can discover the input candidates in a short time. (3) Reusing the static knowledge can be done by letting the expert system builder select and input. It will be possible.

[0008] With this, when dynamic knowledge is input to a tabular editor screen, not only can static knowledge input in advance be reused, but the relevance can be improved by weighting selection candidates. The possibility of scrolling can be reduced because candidates with a high score can be displayed at the beginning. In addition, the expert system can be constructed easily and quickly, improving productivity.

[0009]

EXAMPLES Next, examples of the present invention will be described. In this embodiment, the state of the furnace condition of the blast furnace is determined by an expert system. High-speed dynamic knowledge input editing / expansion of the export system targeted by this embodiment
The expression method will be described. The reasoning (facts, constants, external functions, metal rules) for constructing the expert system is based on the conventional technique, and is the same as in Japanese Patent Laid-Open No. 04-190425.

FIG. 1 shows a work procedure. First, register knowledge base such as knowledge name and version for each target problem 1-1
I do. Next, input and edit 1-2 of static knowledge (fact, frame, constant, external function definition). These are used in dynamic knowledge (rules in the form of what to do). Next, input editing 1-3 of inference control knowledge (metarule) that determines the structure of dynamic knowledge and the inference method is performed. Next, the dynamic knowledge name is input and edited 1-4. Further, similarity calculation 1-5 is performed. Finally, the dynamic knowledge (rule) is input and edited 1-6 to construct an expert system. Note that it is not necessary to define the static knowledge that is not used in dynamic knowledge.

Next, detailed input and editing methods of knowledge at each stage will be described. When the expert system construction software of the computer is started, the entire system management screen shown in FIG. 5 is displayed on the CRT. When the items shown therein are designated, the detailed screens as shown in FIGS. 6, 7 and 8 are displayed. That is, by specifying the items on the overall system management screen shown in FIG. 7, registration of a knowledge base (“Knowledge base edit” in the diagram) and input / edit of knowledge (“Editor” in the diagram) described below are performed.
In addition to selecting and starting, you can select and start the conversion software to AI language, verification of the constructed expert system, test function, etc. Below,
Process according to FIG.

Knowledge base registration (1-1 in FIG. 1): When the whole system management screen (knowledge base edit in FIG. 5) is selected, the knowledge base edit screen (FIG. 9) is displayed; “Knowledge base registration” in this screen ", Select the knowledge base name (" test1 "in this example), version number. (The old and new knowledge is expressed by numbers. In this example, it is 1.0.) And global group number (number is assigned to the file group for passing inference results etc. online with the process computer. The two for the actual machine can be specified in this embodiment) and registered. After registration, select Exit to close this screen.

Input and edit of static knowledge (1-2 in FIG. 1): (1) fact, (2) frame
Four types of (3) constant and (4) external function definition are used; these knowledge defines the name of knowledge to be referred to or rewritten by dynamic knowledge (rule) described later and its initial value, etc. To do.

(1) Facts Various knowledge phenomena used in reasoning are expressed in the form "of ~ is 〜" in the form of fact editor screen FIG. 10; fact names defined as facts, The field name, value, etc. are referred to in the conditional part "if ~ of ~ is ~" in the rule described later, and rewritten by the rule execution part "~ of ~ ~ ~" (update,
(Deleted, generated) Used as data for controlling the execution of rules; FIG. 10 is an example of the input, and defines specifications such as marginal total. The name of the fact and the initial value are set such that the latest value of the peripheral total is 1 and the number of times is 0.

(2) Frame This is a representation of various knowledge events used in the inference in the form of "... of ~ is ...", which is input from the frame editor screen shown in FIG. The frame name, slot name, value, etc. defined as a frame are referred to in the conditional part "if ~ ~ ~ is ~" in the rule described later, and rewritten in the rule execution part "~ ~ ~ ~ ~". (Update, delete, generate) and is used as data that controls the execution of rules. The difference between the fact and the frame is that the fact can describe a hierarchical structure with data inheritance; FIG. 11 is an input example thereof and defines specifications such as the hot metal temperature. Before the hot metal temperature, the first TAP was 15
The name of the frame and the initial value are set, such as 35.0 and the first TAP is 1534.0.

(3) Constant A constant value such as a reference value of an event is defined as a constant name in the knowledge, and the constant name can be quoted by other knowledge such as facts, frames and rules. Input from the constant editor shown in Fig. 12; the constant management can be maintained by defining constants in this editor, and the values quoted in facts, rules, etc. by simply changing the values on this screen Will be changed collectively, and development efficiency will also be improved. FIG. 12 is an example of the input, and "reference value 1" defines 0.5 for the real number type and "reference value 2" defines 0.8 for the real number type. When quoting this value on another screen, add "!" At the beginning like "! Standard value 1" and quote with a constant name.

(4) External function definition When the user wants to call a subroutine created by the user in the dynamic knowledge (rule) execution section, the attributes of the subroutine are defined by the external function definition editor screen shown in FIG. FIG. 13 shows an example of the input, “M
AX "is written in C language and defines that the input has two integer arguments and that the output has one integer argument." MAXR "is written in C language and the input has real arguments. Indicates that there are two and there is one real argument in the output.
“MIN” is created in C language and defines that 2 integer arguments and 1 real number argument are returned to the input as a result.
"MAX", "MAXR" and "MIN" are called as external subroutines by the rule execution unit.

Inference control knowledge input and edit (hereinafter, metarule) (1-3 in FIG. 1): A metarule is a dynamic knowledge (rule) described later, which is collected or divided into several knowledge groups. It is for controlling the order of execution of inference in units of groups. As a method of controlling the execution of the knowledge group, FIG.
Also, there is a method of determining in advance on the rule editor screen shown in FIG. 16 and a method of switching the knowledge group during execution in the rule, and both can be used together.

FIG. 14 shows a metal rule editor screen. Input items specify the execution order of the knowledge group, (in this embodiment, execute from the top of the line), the name of the knowledge group, and the conflict resolution method; the knowledge group switching timing is
When there is no rule that holds in the knowledge group during inference,
This is also the case when another knowledge group is called (start operation) in the rule described later; FIG. 14 is an input example thereof, and this knowledge base “test1” is composed of nine knowledge groups.

Input of dynamic knowledge name (1-4 in FIG. 1): Input dynamic knowledge name and priority for inputting dynamic knowledge; FIGS. 15 and 16 (particularly FIG. 16) This is an input example. The knowledge group name is “furnace heat primary judgment”, the rule name is “hot metal temperature in furnace”, and the priority is 0.

Calculation of similarity (1-5 in FIG. 1): The similarity is calculated from the character strings of the dynamic knowledge name and the input candidate name. The contents are shown in FIG. 2; first, the dynamic knowledge name and the input candidate name are read. After that, the character strings of the dynamic knowledge name and the input candidate name are compared from the beginning, and the partial similarity is calculated. After the calculation of the partial similarity, one of the dynamic knowledge name and the input candidate name having the shorter character string is deleted. Return again and calculate the partial similarity. When the shorter one of the character strings is completely deleted, the partial similarities are summed to obtain the similarity of the input candidates (FIG. 2); for example, in the case of FIG. 3, the fourth and seventh characters match. Suppose In this case, the individual similarities are 1, and the partial similarities are 2 by adding them. Also, in the case of FIG. 4, 4, 5 and 6
Suppose the th character matches consecutively. in this case,
The points of the character string are 1, 2 and 3, respectively, and when they are added, the individual similarity is 6. The partial similarity is also 6; the partial similarities are added together to obtain the overall similarity (hereinafter similarity). For example, in the "test1" expert system, "furnace top temperature", "hot metal temperature", "furnace condition rank determination", "elapsed time", etc. have already been input as frames (FIG. 19). In addition, FIG.
The dynamic knowledge name edited in (continuation of FIG. 15) is “furnace temperature determination”. Of the static knowledge already input, the four frames will be exemplified. When the name of the dynamic knowledge and each frame similarity are calculated, [1] In the case of “furnace top temperature”, comparing “furnace temperature determination” and “furnace top temperature” as permutations, the “furnace” matches Therefore, individual similarity = 1,
Since the “temperature” matches, the individual similarity 1 + 2 =
3, therefore partial similarity (total) = 4 When comparing "internal temperature judgment" (1 character deleted, the same below) and "furnace top temperature" as permutations, there is no match, so partial similarity = 0 "temperature judgment , And "furnace top temperature" as permutations, there is no match, so partial similarity = 0 "degree determination" and "furnace top temperature" as permutations
Partial similarity = 0 because there is no match When comparing "judgment" and "furnace top temperature" as permutations, partial similarity = 0 because there is no match Compared "constant" and "furnace top temperature" as permutations Then, since there is no match, the partial similarity = 0 and the similarity = 4.

[2] In the case of "hot metal temperature", when "judgment temperature inside" and "furnace top temperature" are compared as a permutation, since "temperature" matches, individual similarity 1+
2 = 3, therefore partial similarity (total) = 3 When comparing “inner temperature judgment” and “furnace top temperature” as permutations, there is no match, so partial similarity = 0 “temperature judgment” and “furnace top temperature” When comparing as a permutation, there is no match, so partial similarity = 0.
Partial similarity = 0 because there is no match When comparing "judgment" and "furnace top temperature" as permutations, partial similarity = 0 because there is no match Compared "constant" and "furnace top temperature" as permutations Then, since there is no match, the partial similarity = 0 and the similarity = 3.

[3] In the case of "furnace condition rank judgment", when "reactor temperature judgment" and "furnace condition rank judgment" are compared as permutations, since "furnace" matches, individual similarity = 1, Therefore, partial similarity (total) = 1 When "reactor temperature judgment" and "condition rank judgment" are compared as a permutation, since "judgment" matches, individual similarity 1 + 2 = 3, therefore partial similarity (total) = 3 When comparing “in-reactor temperature judgment” and “rank judgment” as a permutation, there is no match, so partial similarity = 0 When “reactor temperature judgment” and “rank judgment” are compared as a permutation, there is a match There is no partial similarity = 0 When comparing the "reactor temperature judgment" and "ku judgment" as permutations, there is no match. Partial similarity = 0 "reactor temperature judgment" and "judgment" are compared as permutations Then it matches Comparing the permutations "constant" and partial similarity = 0 because there is no "furnace temperature determination"
Since there is no match, the partial similarity = 0 and the similarity = 4.

[4] In the case of "elapsed time", when "reactor temperature determination" and "elapsed time" are compared as permutations, there is no match, so partial similarity = 0 "internal temperature determination" and "elapsed time" When "Time" is compared as a permutation, there is no match, so partial similarity = 0 "Temperature judgment" and when "Elapsed time" is compared as a permutation, there is no match, partial similarity = 0 "Degree judgment" Comparing "elapsed time" as a permutation,
Since there is no match, partial similarity = 0 When comparing “judgment” and “elapsed time” as a permutation, because there is no match, partial similarity = 0 “constant” and “elapsed time” are compared as a permutation Since there is no match, the partial similarity = 0 and the similarity = 0.

Dynamic knowledge input editing (1-6 in FIG. 1): While the static knowledge described above defines the initial value of inference, the dynamic knowledge is the knowledge that is the main body of inference. The format is expressed in the form of a conditional part "if ~ of ~ is ~" and an execution part "what to do of ~". Figure 15 (first half)
And on the rule editor screen shown in FIG. 16 (second half); Dynamic knowledge input method: Define the rule contents on the rule editor screen. A rule is composed of a condition part "if-if" and an execution part "what to do". It is not possible to describe a rule without a condition part or an execution part. The details are given below: Conditional part: Describe the constraint condition or judgment condition of the facts necessary to execute the execution part; execute the execution part when this condition part is satisfied. The description procedure is described below; frame name (fact name) Describes the fact to be judged. If the same frame is used in the second and subsequent lines, this item may be blank; slot name (field) name Describe the field corresponding to the above; Index or in the "Index" column when quoting array elements by facts Specify the element number; Describe the value in the following format to judge the value. Integer, real number, character string, constant name (add '!' At the beginning) variable (add '?' At the beginning) and system default value (symbol Nulli, etc.) can be described; It means to bind (apply) to a variable; and to quote the variable on the left side of the calculation formula, in which case it means to bind the calculation result to that variable; , "Greater than", "less than" can be specified; means that if there is no input, it is equal; connective Specifies the connection relation between each line; "is", "is", "to", "and" Specify any one of "Yes", "Not", "Not", "And", "Or", "Does not exist" for each line. "Not" and "Not" are negative "And", "or" represents the connection relation between conditions. Also means that the rule is satisfied in the absence of one is the fact that you described in the conditions.

Calculation formula For the variables used in the condition part and the execution part, it is possible to use a calculation formula that constrains the variables; the formula is formed in the form of variable name = expression, and the right side is four arithmetic operations +, -, *, /
And, it is described in the usual mathematical expression using parentheses () symbols. Also, constant names (adding '!' At the beginning) and library functions can be used on the right side; Figs. 15 and 1
6 is an input example.

Execution part: Describe the contents to be executed when the condition part is satisfied. The description procedure is almost the same as the condition part, but the following functions are executed by selecting and specifying "Update", "Delete", "Create", "Call", "Display", "Start" in the operation column. Yes; 1) Fact data operation 2) “Call” of standard subroutine prepared in advance in the system 3) “Call” of user-created subroutine (external subroutine) 4) “Display” of fact data etc. on console 5) In addition, the switching of the execution right to the knowledge group “start” The description procedure and operation of each function will be described below.

1) Fact data operation When the operation column is "update", the target fact name and field name (in the case of an array element, an index is specified) and the value to be updated are specified. When using a variable, it must be bound by a conditional part or a calculation formula; in the case of "creation", the fact name (specify the index in the case of an array element) and the value are specified. When executed, new fact data is generated; in the case of “delete”, the fact name is specified; when the rule is executed, the specified fact data itself is deleted (all related slot data is deleted). The example is described in FIG.

2) Call processing of standard subroutine When the operation column is "call", the subroutines prepared as standard in this system are prepared as follows in this embodiment. For the description, specify “% standard” in the item of fact, specify the following subroutine name in the item of field, argument in the item of value, and “call” in the operation field; 3) Call processing field of user-created external subroutine If is "call", the user created FOR
Can call TRAN and C language subroutines. For the description, specify "% external" in the item of fact, subroutine name in the item of field, argument in the item of value, and "call" in the operation field. It is necessary to define the attributes of arguments in advance in the external function editor. The argument is
You can use integers, real numbers, character strings, constant names, and variables. If there are multiple numbers, separate them with ','. If the argument attribute is “output” (subroutine → rule), a variable can be specified as the argument and its value can be used in the rule execution unit.

4) Console display of fact data etc. When the operation column is "display", the value of the fact can be displayed on the console window of the CRT. In the description procedure, the fact is specified in the fact section and the field name is specified in the field section. Specify "display" in the operation column. still,
The value column is unnecessary.

5) Switching the execution right to another knowledge group When the operation column is "start", the knowledge group is normally executed in the order defined by the metarule, but the knowledge group specified by the rule execution unit may be executed. I can. After the specified knowledge group is executed, the knowledge groups after the knowledge group defined by the meta-rule are sequentially executed unless changed by another rule. As a description procedure, specify the name of the knowledge group to be activated in the fact section and specify "start" in the operation field.

In the condition part and the execution part of the rule configured as described above, for example, in inputting a frame,
Select the input in the rule editor screen of FIG. 15 and FIG.
When the frame field is picked, the frame input screen of FIG. 17 is displayed. Here, when the frame (fact) button is selected, the selection screen of FIG. 18 is displayed, and when the frame is pulled down, the frame selection screen of FIG. 19 is displayed. At this time, the frames are displayed in descending order of relevance to the name of the dynamic knowledge, that is, the similarity calculated as described above, and the expert system builder can select and input.

In the above-mentioned example, the respective degrees of similarity are “furnace temperature = 4”, “hot metal temperature = 3”, “furnace heat rank judgment =”
4 ”,“ elapsed time = 0 ”are displayed, and the frames are displayed in the order of“ furnace temperature ”,“ furnace heat rank determination ”,“ hot metal temperature ”, and“ elapsed time ”. This is just a simplified example. When the number of static knowledge and dynamic knowledge actually increases, unnecessary static knowledge is not displayed, and those related to dynamic knowledge are displayed first. , Less need for scrolling. The same applies to fact, fuzzy, and confidence data.

The high-speed knowledge input and editing method of the expert system according to the embodiment of the present invention has been described above. The process generation of the input / edited knowledge, the inference execution, the knowledge verification, and the online process generation function are described. Also, the systemization has already been completed on a computer.

[0035]

As described above, according to the present invention, in constructing and maintaining an expert system, 1) it is efficient and easy without requiring specialized knowledge such as AI language related to a high-level expert system. The expert system can be built in the software, and its software productivity and maintainability are greatly improved; 2) Inconsistency between static knowledge and dynamic knowledge can be prevented,
By reducing errors, productivity is improved.

[Brief description of drawings]

FIG. 1 is a flow chart showing a knowledge input / editing procedure according to an embodiment of the present invention.

FIG. 2 is a flowchart showing the contents of “similarity calculation” 1-5 shown in FIG.

FIG. 3 is a diagram showing the calculation of partial similarity in the present invention.
It is a top view which shows an example of the correspondence of the character arrangement between knowledge names.

FIG. 4 is a diagram showing the calculation of partial similarity in the present invention.
It is a top view which shows an example of the correspondence of the character arrangement between knowledge names.

FIG. 5 is a plan view showing an entire display screen for inputting / editing knowledge to the expert system, which is displayed on the CRT in the expert system development computer.

6 is a plan view showing a display screen when the item "editor" shown in FIG. 5 is designated.

FIG. 7 is a plan view showing a display screen when the item “knowledge base name” shown in FIG. 5 is designated.

FIG. 8 is a plan view showing a display screen when the item “utility” shown in FIG. 5 is designated.

9 is a plan view showing a display screen when the item "knowledge base edit" shown in FIG. 5 is designated.

FIG. 10 is a plan view showing a display screen when the item “fact” shown in FIG. 6 is designated, and shows a specific input example of a fact editor screen in knowledge input / editing.

FIG. 11 is a plan view showing a display screen when the item “constant” shown in FIG. 6 is specified, and shows a specific input example of a constant editor screen in knowledge input / editing.

FIG. 12 is a plan view showing a display screen when the item “frame” shown in FIG. 6 is designated, and shows a specific input example of the frame editor screen in knowledge input / editing.

FIG. 13 is a plan view showing a display screen when the item “external function” shown in FIG. 6 is specified, and shows a specific input example of the external function editor screen in knowledge input / editing.

FIG. 14 is a plan view showing a display screen when the item “metarule” shown in FIG. 6 is specified, and shows a specific input example of the metarule editor screen in knowledge input / editing.

FIG. 15 is a plan view showing the left half of the screen displayed by designating the item “rule” shown in FIG. 6 and further designating from the rule list displayed for it, which is a rule editor in knowledge input / editing. A specific input example on the screen will be shown.

FIG. 16 is a plan view showing the right half of the screen displayed by designating the item “rule” shown in FIG. 6 and further designating from the rule list displayed for it, and is a rule editor in knowledge input / editing. A specific input example on the screen will be shown.

17 is a plan view showing a screen displayed by designating the item “fact name, frame name” shown in FIG. 15. FIG.

18] The item "fact, frame" shown in FIG.
It is a top view showing a screen displayed by designating.

FIG. 19 is a plan view showing a screen displayed by designating the item “frame” shown in FIG. 18, and according to the present invention, the order of frames is displayed in descending order of relevance.

 ─────────────────────────────────────────────────── ─── Continuation of the front page (72) Minoru Harada Minoru 20-1 Shintomi, Futtsu-shi, Chiba Nippon Steel Co., Ltd. Technology Development Division (72) Inventor Isao Otani 20-1 Shintomi, Futtsu-shi, Chiba New Nippon Steel Co., Ltd.

Claims (1)

[Claims] 1. An expert system is constructed by systematizing the knowledge representation method of the expert system into inference control knowledge, dynamic knowledge, static knowledge and uncertain knowledge, and using a general-purpose tabular editor screen. Method, when inputting and editing dynamic knowledge including static knowledge, (A) creating input candidates for each systematic static knowledge, and (B) input candidate name and dynamic knowledge A method for inputting / editing dynamic knowledge in an expert system, characterized in that the degree of similarity with a name is calculated and displayed in descending order of similarity, and (C) the expert system builder is made to select and input.