JPH0353331A - Knowledge acquisition processing method - Google Patents

Abstract

PURPOSE:To eliminate the bottleneck for formation of an expert system and to easily acquire the knowledges from the specialists by acquiring the knowledges based on the guiding interviews prepared by a computer. CONSTITUTION:An initial knowledge base 11 is automatically produced in the processing processes P1 - P6 of a computer 10 consisting of a CPU, a memory, etc. That is, the subject problems are limited and the solution plans are extracted to the problems to extract the attributes of the solution plans. Then a selection standard of solution plans is extracted, and the evaluation value is calculated based on the selection standard. Then an optimum plan obtained based on the evaluation value is turned into a knowledge base. The knowledges are acquired based on the guiding interviews prepared by the computer 10. Thus it is possible to input the knowledges of the specialists just by answering the questions and then to reduce the developing manhour for formation of an expert system.

Description

[Detailed Description of the Invention] [Summary] The present invention relates to a knowledge acquisition processing method for collecting L3 knowledge and knowledge of a problem using a data processing device for knowledge acquisition, and for controlling a knowledge base used in an expert system. The aim is to eliminate the bottleneck and make it easier to acquire knowledge from experts. A processing process for limiting the target problem and a processing process for extracting solutions to the problem. A processing process for extracting the attributes of a solution plan. A processing process for extracting selection criteria for a solution plan; and a processing process for calculating an evaluation value based on the selection criteria for a solution plan.
It is equipped with a processing process that creates a knowledge base of optimal proposals based on evaluation values, and is based on guided interviews using a total of X machines. Constructed to acquire knowledge.

[Industrial application field]

The present invention is. A data processing device for knowledge acquisition recognizes problems and collects knowledge. This paper relates to a knowledge acquisition processing method for controlling the knowledge base used in an expert system.

Currently, the biggest challenge in building an expert system is: It is said to be the process of extracting knowledge from experts. It can be operated directly by experts. A technology that facilitates the work of acquiring knowledge from experts and eliminates the man-hours required to construct an expert system is desired.

[Conventional technology]

In the expert system. A knowledge base is used that represents the knowledge of experts as rules, frames, etc. Conventional. System engineers (SE) and knowledge engineers (KE) acquire the knowledge to create this knowledge base. It was necessary to conduct repeated interviews with experts and organize the knowledge manually.

In addition, in the works of Knowledge Heas, the acquired knowledge can be used with knowledge editors, etc. It needed to be written down in the form of rules and frames.

As a knowledge acquisition tool. There are so-called table designers that use decisigon tables, but it has not always been easy for experts to directly operate them and create a knowledge base.

by the way. PCT (
Personal Construct Theory
), a clinical psychological theory by Kelly (1959) is known.

Figure 8 is. This is a diagram for explaining the PCT. PC
T is. For example, as shown in Figure 8. An object such as a dog, cat, or snake can be judged by some characteristic, such as being cute or scary. It is a methodology of organizing things in a dichotomous manner. In addition, as a problem-solving method in uncertain situations,
ARP (Analytic Hierarchy)
A mathematical theory called ``Process'' is known.

Figure 9 is. FIG. 2 is a diagram for explaining the AHP. A.R.
P is. As shown in Figure 9, objects are compared pairwise and weighted.

In the example shown in Figure 9. The pairwise comparison value for dogs and cats is 2. Larger than cat. It is rather cute.

The pairwise comparison value for the large snake is 5, meaning that the two dogs are much cuter than the snake. On the other hand, the pairwise comparison value of a cat for a dog is 1/2, which is the reciprocal of the pairwise comparison value of a dog for a cat. Cats are better than dogs. It means it's a little less cute.

Conventionally, no technology has been considered that uses a combination of mechanisms based on PCT and AHP theories in the knowledge acquisition process for constructing expert systems.

[Problem that the invention aims to solve] In conventional technology, knowledge is acquired from experts and the knowledge base is modified based on that knowledge. Because it is done manually,
The problem was that building an expert system required a great deal of time and effort.

Experts extract knowledge while directly operating computers. It would be useful to have a tool that automatically generates initial knowledge from the extracted knowledge.

The present invention aims to solve the above problems and eliminates the bottleneck in constructing an expert system. It aims to facilitate the acquisition of knowledge from experts.

(Means for solving problems) FIG. 1 is a diagram explaining the principle of the present invention.

In the figure, 1 10 is a calculation J consisting of the CPU, memory, etc.
M. l1 is the initial knowledge base.

In the present invention, the processing process P by the computer 10 shown in FIG.
1 to P6 automatically as follows. An initial knowledge base 11 is created.

In the processing step Pi, after processing such as initial value setting and screen clearing. Display questions, menus, etc. on the display.
Guided interviews with experts are used to limit the problems targeted by the expert system.

Next, in process P2, a solution to the problem is extracted. In other words, clarify the options that are the ultimate goal of the problem. Prompts you to enter as many result items as you can think of, displays all result items after they have been entered, checks for changes, allows you to rename some result items, remove or add result items Process.

In processing step P3. Extract the attributes of the solution plan.

In this attribute extraction, for example, three items are sequentially extracted from the result items. Divide what you took out into two parts and one part. Ask them to compare this. The experts are. Compare these two groups. Enter similarities and differences. These similarities and differences become the extreme values for evaluating the result item.

Furthermore, it asks for the top attributes of the extracted attributes. Extract superordinate concepts of attributes. The attributes of this superordinate concept are used in the conditional part when creating rules.

In processing step P4. Extract solution selection criteria. In order to narrow down some solutions for the final selection. Some kind of evaluation standard is required. Therefore. For evaluation items, create a table of paired comparisons that reflects the students' values.

Next, in the process P5, based on the selection criteria of the solutions extracted in the process P4. Calculate the evaluation value. Here, the weight of the attribute for each result item is determined using the mathematical theory of AHP. As is known from the mathematical theory of AHP, it is possible to check contradictions by evaluating consistency.

In the processing step P6, processing is performed to convert the optimal plan into a knowledge base based on the evaluation value calculated in the processing step P5. In other words, the rule expressions and frame expressions used by the target expert system are automatically generated, and the initial knowledge
1 and write it to a file.

Each of the above processes is performed based on M',71-type interviews with experts, so even if experts do not have detailed knowledge of rule expressions or frame expressions, an initial knowledge base 1l can be generated with N simple operations. can be admonished.

[Operation] In FIG. 1, the processing process P1 is a process of recognizing a problem, and the processing processes P2 to P4 are . It is a process of collecting knowledge.

In problem recognition and related knowledge extraction, K
Using Elly's (1959) PCT theory, the weights between the extracted knowledge are calculated. We use the theory of AHP to introduce heuristics with uncertainty into the knowledge base.

The fusion of PCT and ARP makes it possible to efficiently acquire knowledge from experts and automate the organization, and also uses guided interviews. Even if you do not have advanced knowledge editor knowledge. For example, if you have knowledge of operating a word processor. It becomes possible to generate the knowledge base necessary for decision support. Example C] Fig. 2 is a diagram for explaining a hardware model for realizing the present invention, Fig. 3 is a hierarchical diagram for explaining the structure of a problem related to an embodiment of the present invention, and Fig. 4 is a diagram for explaining the structure of a problem related to an embodiment of the present invention. An example of extraction of selection criteria according to an embodiment of the present invention, FIG. 5 is an example of evaluation of each vehicle according to an embodiment of the present invention, and FIG. 6 is an example of evaluation results according to an embodiment of the present invention. FIG. 7 shows an example of the total score according to an embodiment of the present invention.

Figure 2 shows a multi-tape Turing machine (hereinafter referred to as MT).
An example of an automaton model of a device that realizes the present invention expressed by M) is shown. In Figure 2, 21~
23 is the input tape. 24 represents an output tape, 25 represents a finite state machine, and 26 represents a work queue.

When formulating MTM! =(K, Σ.", δ8 ζo, [1).

(al　K is a finite set of states).

(b) ``is a finite set of tape symbols. Its special element is blank, which is denoted by Mukatsudori B.

(C) Σ is a subset of ``that does not include B, and is a set of input symbols.

(d) δ is. In the state function. K×" to KX (1'
− − Mapping to (BI X (L, Rl). However, for a certain variable, δ may not be defined.

(el　ζ0 is .K and is called the initial state.

(f) n is. In K. It is called the final state.

The MTM of this model is. One finite state machine 25 for finite control, three manual tapes 21 to 23 with left ends for reading and writing. Work queue 26. It consists of an output tape 24 with a left edge.

The tape is divided into a series of squares. Each has a symbol string consisting of a finite number of characters. The contents of each are. For example, it looks like this:

In the initial state. The ends of the three input tapes 21 to 23 and one output tape 24 are located at the left end of the tapes. This M
TM is the symbol below the tape head and the finite state machine 2
The following actions occur depending on the state in step 5. (1) Read input table 2l. Change the state of the finite state machine 25. (2) As a result, select elements of the problem read from the tape. Write to the work queue 26.

(3) Remove the manual tape 21. Sheep hyperactive one square square to the right.

(4) Repeat steps (1) to (3) until a blank (B) column is found. When a blank column is found, proceed to the next step (5).
Move on to the operation.

(5) Read the input tape 22. Two selection elements of an appropriate problem are read from the work queue 26, and the finite state machine 2
Change the state of 5.

(6) As a result, write the hierarchical structure of the selected elements of the problem into the work key 26.

(7) First, move the head of the manual table 22 one position to the right.

(8) Repeat steps 51 to (7) until a blank (B) column is found. When a blank column is found, proceed to the next step (9).
Let's move on to the operation. (9) Read Human Power Table 123. Read out the problem selection elements and the hierarchical structure between the selection problems from the work queue 26. Change the selection of finite state machine 25.

00) As a result. to the output tape 24. Write an evaluation for each selected element.

01) First, move the hend of the input tape 23 one position to the right. (12) Repeat steps (9) to (11) until a blank (B) column is found. When a blank column is found, the next 0
Move to one action.

03) Write the overall evaluation for each selected element in the output table 124 as a summary.

In addition. There is also a model that derives initial knowledge from this comprehensive evaluation. It can also be easily realized. The equivalent operation is
Needless to say, this can be realized on a general-purpose computer.

Next, embodiments of the present invention will be explained according to specific examples. [Limitation of the target problem 1: For decision making. First, there is ``Problem J,'' and then there are several ``alternatives (solutions)'' that are subject to a final selection. To narrow down to one from the alternatives. There is an "evaluation standard" between the two.

Diagrammatically. It can be expressed as shown in FIG. 3(a).

As a specific example, we will take the problem of selecting a new car.

The problem is "selection of a new car", the evaluation criteria is "price J" fuel efficiency",
It's the "comfort of the ride." There are other criteria such as "vehicle class", but for the sake of simplicity, we will consider these three criteria here. The alternative is car A. These are car types such as car B and car C.

[Extracting solutions to problems] Apply to Figure 3 (a). The problem of selecting a new car can be expressed hierarchically as shown in Figure 3 (b). This diagram breaks down the "problem". It comes down to three evaluation criteria. Furthermore, we will compare and examine the ``candidate cars'' based on each criteria. It is represented by a line connecting the top and bottom.

Note that as the number of evaluation criteria and alternatives increases, the lines connecting the top and bottom become complicated and difficult to see, so when they are displayed on the display screen 30, they are represented as shown in FIG. What does it mean? It is the same as Fig. 3 (b).

[Extracting the attributes of the solution plan] Once the hierarchical structure is completed. For each level of evaluation items. Make a pairwise comparison like this: This comparison will reflect the person's values.

In the example of selecting a new car mentioned above, there were three criteria. "
price". These are "fuel efficiency" and "ride comfort."

first. Compare "price" versus "fuel efficiency". For that comparison. For example, consider the following numbers. 1, 1, 1 and 1: Both items are equally important.

3; The earlier item is slightly more important than the later item.

5: The earlier item is more important than the later item.

[Extracting criteria for selecting solutions] For example, comparing "price" versus "fuel efficiency." For those who consider lower price a little more important than good fuel efficiency,
Set this value to "3". Enter "3" at the intersection of "Price" and "Fuel Economy" in the table displayed on the screen. next,"
In the case of price j vs. ride comfort, let us assume that price is given more importance. At that time, the value of the square will be "5". "price"
The intersection of the pair ``price'' is, of course, r1. This completes the field next to "Price" on the screen. In the vertical column of "Price". Write the reciprocal of the value in the horizontal column. Enter. In other words, "fuel efficiency" versus "price" is "1/3". "Ride comfort" versus "price" is "1/5". This completes the vertical columns of the table. This time it's the column next to "Fuel efficiency.""
First of all, let's look at fuel efficiency J vs. fuel efficiency. Of course I'll include rlJ.

In the case of "fuel efficiency" versus "ride comfort," it is assumed that fuel efficiency is more important. In that case, enter "5". Once you have completed the field next to ``Fuel Economy'': Enter the reciprocal of that number in the vertical field. In this way, a pairwise comparison table such as that shown in FIG. 4, for example, is created. This table reflects the value standards of the evaluator, but this person's preferences are not. It will be a cheap car with good fuel efficiency. [Calculation of evaluation value based on solution selection criteria] Let's proceed with the example of selecting a new car. People who adopt the paired comparison values shown in Figure 4 are looking to buy cheap cars. From this table, the weight (importance) of each criterion item must be determined. This weight calculation formula is determined using the mathematical theory of AHP. (Reference: Kaoru Tone, "Game-like Decision-Making Broad", Nikkei Ren Shuppansha, 1987, pp. 33-37).

In this example, the objective is a 5A car. The task is to decide which is preferable, car B or car C. For that work, the above AHP
According to the mathematical theory of Calculating the weight of each evaluation item.
“Price” accounts for 57%. "Fuel efficiency" is 33%, and "ride comfort" is 10%. Note 2: There is a detailed explanation of how to calculate this weight in the references mentioned above.
The explanation here will be omitted.

Therefore. Cars A, B, and C are compared for each evaluation item. First, regarding the "price", car A and car B. Compare cars C.

The one with the lowest price is highly rated. If you compare car A and car B, and assume that car B is only slightly more expensive, put a ``2'' next to it. A comparison between car A and car C. If car C is slightly more expensive, give it a "3".

In a comparison between car B and car C, if car C is only slightly more expensive, it is set as "2". Let's calculate the weight of each car based on this pairwise comparison table. The geometric mean of the prices for car A is (IX2X3)'' = 1.814 Similarly, the geometric mean of the prices for car B is (1/2XIX2)'' = I. Similarly to OOO, the geometric mean of the price of car C is 5 (1/3 x 1/2 x 1)"3 = 0.550. From this, if you calculate the weight of each car, it will be as shown in Figure 5 (a) Regarding "price", car A is 0.54 and car B is 0.30.
, C car has a weight of 0.16. of course. A larger weight is preferable. A similar comparison regarding fuel efficiency is shown in Figure 5 (opening). Fig. 5 (c) shows the results obtained for ``Ride Comfort''.

The scores (weights) for each evaluation item for cars A, B, and C shown in FIGS. 5(a) to 5(c) are summarized as shown in FIG. 6.

In order to perform a comprehensive weight evaluation, multiply the evaluation result numbers shown in Figure 6 by the weights of each evaluation item. For example, the score for car A's price is 0.5 4
0 x 0.5 7 = 0.3 0 8. Similarly,
The score for car A's fuel efficiency is O. l06x0.33=0
．． 0 3 5. What is the score for ride comfort? 0.540
XO. I O = 0. 0 5 4. , the overall score of car A is. Add these scores together, 0.397
becomes.

Scores are similarly calculated for Car B and Car C. the result. A total score as shown in FIG. 7 is obtained.

This total score is approximately 0.40 for car A and approximately 0.43 for car B.
．． Car C is about O. ] 7 and 3 It can be seen that cars B, A, and C are preferable in this order.

[Knowledge conversion of optimal plan] Based on the results shown in Figure 7, this is converted into rules and frames for extracting the optimal plan. About this algorithm. This is a design matter, and there is no need to explain it. Due to this. The knowledge base of the expert system will be tested. When the expert system is executed, in response to the question ``If ride comfort is important, what kind of car would you like?'' If fuel efficiency is important, what kind of car should I buy?In response to the question J, the system recommends car B.

Also, "If you were to think of an overall good car, what kind of car would it be?"
In response to the question, what is the system? Answer car B.

Although the problem of selecting a new car has been explained as an example, the present invention is not limited to this and can be applied to knowledge used in various expert systems.

〔Effect of the invention〕

As explained above, according to the present invention, a simple operation such as answering questions in the snow can be done. Expert knowledge can be translated into human power. This allows the initial knowledge base to be automatically generated, making it possible to reduce the number of development steps needed to build an expert system.

[Brief explanation of drawings]

Fig. 1 is a diagram for explaining the principle of the present invention, Fig. 2 is a diagram for explaining a hardware model for realizing the present invention, and Fig. 3 is a diagram for explaining the structure of a problem related to an embodiment of the present invention. Hierarchical diagram Fig. 4 is an example of extraction of selection criteria according to an embodiment of the present invention. Fig. 5 is an example of evaluation of each vehicle according to an embodiment of the present invention. Fig. 6 is an example of selection criteria extraction according to an embodiment of the present invention. Example of evaluation results. Figure 7 is an example of the total score according to an embodiment of the present invention, and Figure 8 is a PC
Figure 9 shows a diagram to explain AHP. In the figure, 10 is a computer. 11 represents an initial knowledge base. Patent distributor Fujitsu Ltd. Agent Patent attorney Yoshiyoshi Ogasawara (and two others) Diagram explaining the principles of the present invention Figure 1 (c) Hierarchical diagram 3 Example of extraction of selection criteria Figure 4 (a) ) Example of evaluation of each car Figure 5 Example of evaluation result Figure 6 Figure 30 Example of total profit Figure 7

Claims (1)

[Scope of Claims] A knowledge acquisition processing method in which a knowledge acquisition data processing device recognizes a problem and collects knowledge to create a knowledge base for use in an expert system, the process limiting the target problem. a process (P1), a process for extracting solutions to a problem (P2), a process for extracting attributes of solutions (P3), a process for extracting selection criteria for solutions (P4), and The process includes a process of calculating evaluation values based on the selection criteria of plans (P5), and a process of creating a knowledge base of the optimal plan based on the evaluation values (P6). A knowledge acquisition processing method characterized by acquiring the following information.