JPH0695884A - Interactive construction supporting system for fuzzy knowledge - Google Patents

Abstract

PURPOSE:To provide a system for easily constructing fuzzy knowledge in a fuzzy inference application system. CONSTITUTION:This system is composed of a data base 1 for preserving raw data, a fuzzy knowledge base 4 for preserving the fuzzy knowledge, a test data storage device 3 for preserving test data or the like, a data preparation part 2 for preparing the test data from the raw data, a fuzzy inference part 5 for performing fuzzy inference, a knowledge learning part 7 for adjusting the fuzzy knowledge and an inference condition display part 6 and a learning condition display part 8 for displaying the progress and the result. By specifying data within displayed contents by an user at the respective display parts 6 and 8, a flag is set to the test data inside the test data storage device 3 corresponding to the data. At the respective display parts 6 and 8 and the data preparation part 2, display data relating to the data to which the flag is set are outputted while a display method is varied from the other display data.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a fuzzy knowledge interactive support construction method suitable for all systems applying fuzzy inference.

[0002]

2. Description of the Related Art A system applying fuzzy reasoning
"Expert Shell for Plant Control" ERIC "" Magazine, Automation, Vol. 33, No. 6, June 1988
Month, pp. 17-21 (Reference 1), "Securities Investment Expert System Using Fuzzy Reasoning", Information Processing, 19
88.8, pp 963-969 (Reference 2), "Fuzzy Expert Construction Shell", Information Processing, 1989.8,
pp. 948-956 (reference 3), and "Fuzzy development support system" LUNA-FuzzyRON "", Journal of the Fuzzy Society of Japan, Volume 3, No. 3, 1991.8 (reference 4). According to Reference 1, the description of rules and the reasoning mechanism are described. It also discloses fuzzy variables. In detail, when setting a fuzzy variable, it is stated that it is necessary to adjust the membership function of the fuzzy variable to one suitable for the situation by repeating trial and error. In addition, there is a description that it is possible to add fuzzy variables as needed. Document 2 recognizes that "even if good results are obtained from past data, there is no guarantee that the rule will be effective in the future."
It discloses a method of reversing the parameter value from the data and registering and modifying the rule with reference to the parameter value. However, the number of variables to be learned in this way is one, and it is limited to the variables included in the if section described in the if-then format. Document 3 discloses a method of making rules and evaluating them by simulation. In particular,
Using the graph display function of the membership function on each input variable, output value, and each output variable which is the result of fuzzy inference at each time of simulation, the effect of the rule is known (by the user), Is disclosed.
Document 4 also shows a method for making rules and a method for evaluating the rules. Regarding the rule evaluation method, it is disclosed that the reasoning result and the result in one case are displayed graphically by being overlapped with fuzzy knowledge.

[0003]

SUMMARY OF THE INVENTION An object of the present invention is to provide a method for easily constructing fuzzy knowledge. Another object of the present invention is to provide a method for easily grasping a fuzzy reasoning situation and a learning situation regarding a certain portion of test data. Furthermore, another object of the present invention is to provide a method for easily and efficiently classifying test data.

The above object is set by the inventor by paying attention to the following problems of the above-mentioned prior art. When the fuzzy knowledge is adjusted for each case separately as in Reference 4, it takes longer time than in the case of performing fuzzy inference for test data of a plurality of cases at once as in Reference 3. .
However, even if some fuzzy knowledge construction with multiple cases works, some cases do not work well. Therefore, in the next fuzzy knowledge construction process, why the teacher data and the inference result do not match only in the case that does not work well? Which rule / membership function is effective in that case? Whether it is effective or not, and which part of the raw data corresponds to that case. Further, when handling a plurality of cases at once, as in Reference 2,
The construction of fuzzy knowledge is facilitated by using the learning of membership functions. However, that alone does not work well for all cases, and even after that, most fuzzy knowledge is adjusted. At this time, it is easy to deal with it if the learning situation of why the membership function cannot be adjusted is known. Further, in this case as well, since the case causing the adjustment failure will be repeatedly examined, it is necessary to separate the case in that case from other cases.
In the process of constructing the fuzzy knowledge base, it is only necessary to adjust some parts of the test data. The test data used here are rarely used as they are as raw data, and most of them are processed raw data. Therefore, it is necessary to recreate the test data. At that time, it is convenient if the data of the case selected when grasping the fuzzy reasoning situation or learning situation can be referred to, but it is convenient. I can't find any literature that touches the creation. Furthermore, it takes time to select each case that does not work out of the test data in each part of the situation grasp of fuzzy reasoning and the situation grasp of learning, so it is necessary to specify it efficiently. The object of the present invention is set based on the recognition of the problems in the prior art as described above. From the above, it will be apparent that the recognition of the problem itself cannot be suggested by the conventional description.

[0005]

In order to achieve the above object, the present invention is as follows. First, in the inference status display section for editing / displaying the status of fuzzy inference, and the learning status display section for displaying the learning status of the membership function,
By instructing the data being displayed,
By providing a case designating means for designating a case and a flag setting means for assigning a code for distinguishing the designated case from other cases, the test data
The purpose of easily classifying data is achieved. When displaying the fuzzy inference status in the inference status display section, in displaying the learning status in the learning status display section, and in displaying the raw data in the test data creation section that supports the creation of test data. Data display means for outputting the case data designated by the case designating means by changing the display method of colors, symbols, etc. from the data of the non-designated case. By providing the, the purpose of easily grasping the fuzzy reasoning situation / learning situation and the purpose of easily constructing the fuzzy knowledge are achieved. In the case designating means, a polygon is created on the display device based on some points designated by the user, and the case corresponding to the data in the polygon is created.
The purpose of efficiently classifying the test data can be achieved by providing a case judging means for judging the case as a case specified by the user. Further, in the data display means, by displaying the frame showing the polygon on the display screen, the case designated by the user and the case not designated by the user can be classified and shown.

[0006]

According to the present invention, the test data of the case specified by the user by the above means is output separately from the other parts, so that the user can output the test data of the specified case. The test data of non-designated cases can be visually grasped, and the inference / learning status of only the test data of the designated case can be noted. You will be able to grasp. As a result, it becomes possible to grasp the part of the fuzzy knowledge that should be corrected, and to make a comprehensive adjustment in consideration of the overall movement. As a result, fuzzy knowledge can be easily constructed. In addition, the present invention specifies the case while watching the contents being displayed,
By specifying several points, it is automatically determined based on the specified points whether or not the case is specified by the user, so that the data can be classified easily and efficiently. Furthermore, regarding the test data of each case in the test data storage device,
Since a symbol for identifying whether or not the case is designated by the user is added, the inference status display section, the learning status display section, and the data creation section operate in cooperation through the symbol. This makes it easier to build fuzzy knowledge.

[0007]

Embodiments of the present invention will be described in detail below with reference to the drawings. FIG. 1 is a process related diagram of the first embodiment for realizing the present invention. In FIG. 1, 1 is a database for storing raw data such as time-series data, and 2 is a dialogue with a user for testing 1 or more cases of test data required for fuzzy inference. A test data storage unit 3 is created while the test data storage unit 3 is created by the data creation unit 2 and stores test data such as input data and teacher data. A fuzzy knowledge base for storing the fuzzy knowledge (rule and membership function) expressing 5 uses the input data of the test data storage device 3, and uses the fuzzy knowledge base 4
The fuzzy inference unit 6 executes fuzzy inference based on the case stored in the test data storage unit 3, and 6 indicates the progress and result of the fuzzy inference unit 5, and the test data storage unit 3 and the fuzzy knowledge base. -The reasoning status display section for editing the reasoning 4 and outputting the reasoning status; and 7 for the fuzzy knowledge base 4
A knowledge learning unit for adjusting the membership function in the test data storage unit 3 to match the teacher data in the test data storage unit, and 8 superimposing the membership function obtained by the knowledge learning unit 7 and its learning status. It is a learning status display unit for outputting. The fuzzy rules in the fuzzy knowledge base 5 are: if {(A ₁ is α ₁₁ ) ∧ (A ₂ is α ₂₁ ) ∧
…} It is described in a form such as then (B ₁ is β ₁₁ ). “A ₁ is α ₁₁ ” is called a proposition, and it is a sentence such as “temperature (A ₁ ) is high (α ₁₁ )”, and its meaning is defined using a membership function. Here, "A ₁ " is called a subject and "α ₁₁ " is called a predicate. In addition, ∧ indicates AND.

The inference status display section 6 displays the inference result and the teacher data.
The inference result display portion 10 for outputting the data, the error distribution display portion 11 for outputting the distribution of the difference (error) between the inference result and the teacher data, and paying attention to a certain rule, the representative of the rule An error / fitness display portion 12 for outputting the relationship between the difference (error) between the value and the teacher data and the representative fitness of the rule, and which rule is used in the case designated by the user. Is displayed in the form of a table showing how much influence is exerted when obtaining the result. Display parts 10, 11, 12,
Detailed display examples of 13 are shown in FIGS. 28, 29, and 3, respectively.
0, shown in FIG. Details of each item will be described later.

Various methods have been proposed for adjusting fuzzy knowledge. In the knowledge learning section 7, "High-speed learning method of membership function in fuzzy inference" is proposed by IPSJ Microcomputer and WWW. Custer
The extended backpropagation method described in the Study Group 66-5 (19991) is used. In the extended backpropagation method, fuzzy inference using fuzzy knowledge of fuzzy knowledge base 4 is represented by a calculation network. Then, the knowledge learning unit 7 reverses the difference (output error) between the inference result obtained by the fuzzy inference and the teacher data in the test data storage device 3 to the normal fuzzy inference on the calculation network. By propagating in the direction, the value obtained by multiplying the partial differential coefficient of each membership function by the error is calculated. This value is called a correction amount, and the parameters of the membership function are corrected so that the average value of the correction amount for each case in the test data storage device 3 becomes zero. Each proposition P _i thus obtained (i = 1, ..., Number of propositions (number of propositions stored in fuzzy knowledge base 4))
, The modified membership function shape is defined as M _ij (i
= 1, ..., the number of propositions, j = 1, ..., the number of parameters),
The correction amount of each case is set to DE _ij (i = 1, ..., Proposition number, j
= 1, ..., The number of cases (the number of cases stored in the test data storage device 3), and the membership function M _ij
The result of fuzzy inference using (i = 1, ..., Proposition number, j = 1, ..., Parameter number) is Y _i (i = 1, ...
..., the number of cases). The learning status display unit 8 collectively outputs the correction amount calculated by the knowledge learning unit 7 and its membership function for the proposition specified by the user.

FIG. 2 shows the operation of the fuzzy inference unit 5.
Although various methods have been proposed for fuzzy inference, a typical method is used here. The fuzzy knowledge base 4 has two rules, two if-part propositions and one-then part proposition. The ambiguous amount used in each proposition is α ₁₁ ~ The motion will be described assuming that the membership functions are defined as α ₂₂ , β _{11 to} β ₂₁ . Now A ₁
Input data of - data is input data of a _1, A ₂ - and data is given a _2. At this time, the proposition of rule ₁ “A ₁ is α
₁₁ "goodness of fit is g _11, Proposition" A ₂ isα ₂₁ next adaptation degree g ₂₁ of "Le - representative goodness G ₁₁ Le 1 adopts the smaller of the two fit, and g ₂₁ . Next, the proposition "B ₁
The membership function 20 of “is β ₁₁ ” is multiplied by the representative fitness G ₁₁ of rule 1 to be transformed into a function 21. Similarly, for rule 2, the proposition "B ₁ is
The membership function 22 of “β ₂₁ ” is transformed into a function 23 by the representative fitness G ₁₂ of the rule 2. Then, the transformed functions 21 and 23 are integrated, and the barycentric value Y of the integrated function 24 is integrated.
₁ (If the integrated function 24 is f (B ₁ ), the center of gravity value Y ₁ = ∫B ₁
・ F (B ₁ ) dB ₁ / ∫f (B ₁ ) dB ₁ ) is obtained,
Inference result. The fuzzy inference unit 5 repeats such inference as many times as the number of cases stored in the test data storage device 3 to obtain the inference result in each case as Y _i (i =
1, ..., Number of cases), and the representative fitness in each case and each rule is G _ij (i = 1, ..., Number of cases, j =
1, ..., Number of rules (number of rules stored in fuzzy knowledge base 4).

FIG. 3 is a block diagram showing an embodiment of a fuzzy knowledge interactive construction support system for realizing the present invention. In FIG. 3, reference numeral 30 is an editor that assists the user in changing fuzzy knowledge, and 31 is raw data and inference status in the data creation section 2, inference status display section 6, and learning status display section 8. , A display device for outputting fuzzy knowledge, etc., 32
The user designates raw data to be used as test data in the data creation unit 2, or the user designates a part of display contents of the inference status display unit 6 and the learning status display unit 8. It is an input device used when doing.

FIG. 4 shows the structure of the data base 1 for storing raw data. In FIG. 4, 40 is a data number, 4
1 is the value of the data.

FIG. 5 shows the table configuration of the test data storage device 3. The test data storage device 3 stores one set (case) or more of test data necessary for fuzzy inference. In FIG. 5, the test data storage device 3 stores data.
A position number 50 for setting the corresponding data number 40 and an input data for storing the input data which is the input value (state value) of the fuzzy inference in order to indicate the position in the database 1 51, a teacher data serving as a model value of the fuzzy inference result
Whether or not the user designated a corresponding case in the teacher data 52 storing the data, the inference status display section 6 or the learning status display section 7,
The flag 53 stores a flag indicating whether or not it is. The flag appearing in the following description is this flag 53.
The number of input data depends on the type of subject in the rule if part, and the number of teacher data depends on the type of subject in the rule then part. Further, the value of each position number 50 is expressed by (i = 1, ..., Case number). The input data belonging to the case number and the teacher data are test data for one case.

FIG. 6 is a flowchart showing a processing procedure of the inference status display unit 6 of the fuzzy knowledge interactive construction support system according to the present invention. First, in the inference status display unit 6, the case i (i.e., i) stored in the test data storage device 3 is stored.
= 1, ..., Case number N, set for each case
O _i, input data - data, teacher de - data T _i, and the flag F _i, fuzzy knowledge base - read the contents of the scan 4 (601). next,
Fuzzy inference is performed based on the fuzzy knowledge base using the input data of the test data storage device 3. That is, the fuzzy inference unit 5 is activated. The fuzzy inference unit 5 repeats inference as many times as the cases stored in the test data storage device 3, and inference results Y _i (i = 1, ...
, ..., the number of cases), and the representative fitness G _ij (i = 1, ...
1, ..., Rule number) are passed to the inference status display unit 6 (60
2). An inference result display portion 10 shown in FIG. 1, an error distribution display portion 11 for displaying error distribution, error / fitness, and effectiveness, which are reference data for efficiently performing inference / learning, and error / fitness display. The user selects from the input device 32 which item is to be output in the portion 12 and the strength indication portion 13 (603). When the user selects "inference result", the inference result display (details in FIG. 7) is executed (604). When the user selects “error distribution”, the error distribution display (details in FIG. 9) is executed (60
5). When the user selects "error / fitness", the error / fitness display (details in FIG. 12) is executed (605).
When the user selects "effectiveness", the effectiveness display (details are shown in FIG. 14) is executed (607). After executing the steps of the inference result display (604), the error distribution display (605), the error / fitness display (606), and the effectiveness indication (607), the inference status display processing is terminated.

FIG. 7 is a flowchart showing the detailed processing of the inference result display (step 604, FIG. 6) in the processing procedure of the inference status display section 6. First, the inference result Y i of each case _i is sequentially output to the display device 31 like the inference result display portion 10. Further, each case i (i = 1, ...
..., the number of cases), the teacher data T _i is displayed in the same manner as the display device 3
(701). Then there Ke - scan I - if a flag F _I of (1 ≦ I ≦ pilonidal number) is standing, investigate whether.
Here, the fact that the flag F _I is set means that the flag F _I is 1.
If the flag F _I is not on, it means that the flag F _I is 0 (702). In step 702, flag F
_{If I} is present, the background color of the part outputting the inference result Y _I and the teacher data T _I on the display device 31 is changed from that of the other part (703), and then all the cases are processed. It is checked whether the steps 702 to 703 have been executed (704). If there is an unexecuted case, the process jumps to step 702 to check the unexecuted case. If the flag F _I is not set in step 702, the process jumps to step 704.

When all the cases have been executed in step 704, the user waits for the input device 32 to specify two points (705). Next, it is checked from the two points designated by the user whether or not it is an end instruction.
When both of the two points are outside the graph of the inference result display portion 10, it is determined that the instruction is an end instruction (706). When it is not the end instruction, the designated area is created based on the input X coordinates of the two points. That is, the coordinates of the two input points are (X1, Y
1) and (X2, Y2) (where X1 ≦ X2), the designated area is set to [X1, X2] (707). Then, the inference result Y _{I of} a certain case I (1 ≦ I ≦ the number of cases)
The coordinates (xI, yI) on the display device 3 at the time of drawing are obtained, and it is checked whether or not xI is within the designated area [X1, X2] (708). When xI is in the designated area [X1, X2], this case is designated, and flag F _I is set for this case, that is, flag F _{I is set} to 1 (70
9), steps 708-709 for all cases
(710) is executed. That is, the flag is reset for the cases in the designated area and the flags are reset for the other cases. If there are unexecuted ones, step 708
Jump to and examine the flags of the unexecuted cases. When it is determined in step 710 that 708 to 709 have been executed for all cases, the process jumps to step 705. When the user gives an end instruction in step 706, the process of displaying the inference result ends. The designated area may be displayed in a frame.

FIG. 8 shows the structure of the frequency table 800. The frequency table 800 has N sets of data, and each set has a region number 801, which indicates the number of regions of data.
An area 802 indicating the error range of the area, a frequency 803 indicating how many cases the error belongs to the area 802, and a case number storing a case number indicating that the error belongs to the area 802. 804 and attention frequency 805. The attention frequency 805 has N sets of data, and each set has an error belonging to the region 802 and a corresponding flag F _I.
(1≤I≤case number) represents the number of cases.

FIG. 9 is a flowchart showing the detailed processing of the error distribution display (step 605, FIG. 6) in the processing procedure of the inference status display section 6. First, each case i (i =
1, ..., Number of cases), the difference between the inference result Y _i and the teacher data T _i is obtained and is defined as the error E _i . And the error E
_The maximum absolute value of _i is set as the maximum error E _max (901). The error table 800 is created using the error E _i and the maximum error E _max , and the frequency 800 table is output as the error distribution display portion 11 (902; details are shown in FIG. 1).
0). Each flag F _i is standing next to case - scan i (i =
1, ..., Case number), an attention degree 805 is created. That is, it is checked which number region 802 of the frequency table 800 the error E _i belongs to (J), and the Jth attention frequency in the attention frequency 805 is incremented by one. The attention frequency 805 thus created is output to the display device 31 in a color different from the frequency table 800 output in step 902 (903).

Then, it waits for the user to specify two points with the input device 32 (904). Next, two points (X1, Y1) and (X2, Y2) specified by the user (where X1≤
Based on X2), it is checked whether or not it is an end instruction (90
5). The determination condition is the same as that in step 706 of the inference result display (FIG. 7). When it is not the end instruction, the designated area is set to [X1, X2], and in the frequency table 800, all of the areas 802 existing inside the designated area [X1, X2] are set as designated areas (906). Next, it is checked whether or not a case number is set in the case number 804 of the corresponding frequency table 800 for a certain designated area (907). When the case number is set, all the set case numbers J (J is the case number 80
4), the corresponding flag F
Build _J (908). It is checked whether steps 907 to 908 have been executed for all designated areas (909). If there is an unexecuted designated area, the process jumps to step 907 to check another designated area. If the case number is not set in step 907, the process jumps to step 909. For example, in FIG.
2, N-1, when the region having the region number of N is the designated area, the flag _{F 3, F 5, F 6} , F 8, F 2, F 7,
Sequentially 1 and the F _9, F _10, sets a respective flag. When it is determined in step 909 that steps 907 to 908 have been executed for all designated areas, step 90
Jump to 4. In step 905, when the user instructs the end, the error distribution display processing is ended.

FIG. 10 shows the frequency table in the error distribution display.
9 is a flowchart showing detailed processing of processing 902 for creating and displaying a bull 800. First, each area 802 of the frequency table 800 is calculated using the maximum error E _max . Specifically, [-E _max , E _max ] is divided into N areas, and the areas are sequentially allocated from area number 1 (1001). Next, each frequency from the 1st to the Nth in the frequency table is 80
3 is set to 0 and initialized (1002). Next, it is checked which number region of the frequency table 800 the error E _{I of} a certain case I (1 ≦ I ≦ case number) belongs to, and the region number is set to J (1 ≦ J ≦). N) (1003) and J
The 1st frequency 803 is incremented by 1 and the case I is set to the case number 80.
It is set to 4 (1004). This step 1003-1
It is checked whether the processing of 004 has been executed for all cases (1005), and if there is an unexecuted case, step 10
Jump to 03 and step 1 for unexecuted cases
Perform 003 to 1004. In step 1005,
When all the cases have been executed, the frequency 803 of the frequency table is output to the display device 31, and the process of creating and displaying the error table 800 is completed.

FIG. 11 shows an error / fitness table 1100 used for displaying the error / fitness. Error / fitness test
The blue 1100 shows the relationship between the error and the representative fitness in a certain rule, and there are sets of data corresponding to the number of cases.
This table is created in the same number as the number of rules. In FIG. 11, the data of each set i (i = 1, ..., Number of cases) includes an error area 1101 for storing an error E _i and a representative fitness G _iR (1 ≦ R ≦ _rule ). Number) is stored in the representative fitness area 1102. The meaning of the table 1100 will be clarified from the following description.

FIG. 12 is a flowchart showing the detailed processing of the error / fitness display (step 606, FIG. 6) in the processing procedure of the inference status display section 6. First, the user designates a rule number R (R = 1, ..., Number of rules) from the input device 32 (1201). Next, th of the Rth rule
The representative value of the membership function used in the en part is calculated and set as D. Specifically, the representative value of the membership function 20 or 22 in FIG. 2 is obtained, and the representative value is the center of gravity of the membership function. Further, as the representative value, the midpoint of the portion having the maximum height in the membership function may be used. Then, the representative fitness G _iR (i = 1, ..., Case number) of the R-th rule of all _cases
Are set in the representative fitness area 1102 in the error / fitness table 1100 (1202). Next, for all cases i (i = 1, ..., Number of cases), the error E _i of the R-th rule is calculated, and the difference between the representative value D and the teacher data T _i is calculated. And the error / fitness table 1100 is set in the error area 1101 (1203).

For a certain case I (1≤I≤case number), it is checked whether or not the flag F _I is on (120).
5). If the flag F _I is not on, the X-axis is the error E _I ,
“X” is plotted at a position where the Y axis indicates the representative goodness of fit G _IR (1206). In step 1205, flag F
_{When I} is present, “Δ” is plotted at a position where the X-axis indicates the error E _I and the Y-axis indicates the representative goodness of fit G _IR (1206).
After the execution of step 1205 or 1206, it is checked whether or not "x" or "△" is drawn for all the cases (1208). If there is an unexecuted case, step 12 is executed.
Jump to 05. The error / fitness relationship is displayed as in the error / fitness display portion 12 in FIG. 1 or in FIG.

In step 1208, all cases
When steps 1205 to 1207 have been executed for the scan, the user waits for the input device 32 to specify two points (1209). Next, the two points (X1,
Y1), (X2, Y2) (where X1 ≦ X2, Y1 ≦
From Y2), it is checked whether or not it is an end instruction (1210).
When the relationship between the two points is other than X1 ≦ X2 and Y1 ≦ Y2, X1 and X2 or Y1 and Y2 are exchanged so as to satisfy this relationship. The determination condition for the end instruction is the same as that in step 706 of the inference result display (FIG. 7). From the two points specified by the user, the specified error area and the specified conformance area are respectively [X1, X2] and [Y1, Y2].
(1211). For a case I (1 ≦ I ≦ number of cases), the error E _I of the error area 1101 belongs to the specified error area [X1, X2], and the representative fitness area 110
It is checked whether the representative fitness level G _IR of 2 belongs to the designated fitness range [Y1, Y2] (1212). Error E _I, when both of the error fit G _IR satisfies the condition of step 1212, a flag F _I, the flag F _I other words 1
(1213) Step 1 for all cases
It is checked whether or not 212 to 1213 are executed (121
4). If there is an unexecuted case, the process jumps to step 1212 to check the unexecuted case. Step 12
If it has been executed in all cases in step 14, step 1
Jump to 209. If at least one of the error E _I and the error conformance G _IR does not satisfy the condition in step 1212, the process jumps to step 1214. If there is an end instruction in step 1210, the error / fitness display processing is ended.

FIG. 13 shows a table configuration example of the counter 1300 used for displaying the effectiveness. Counter 1300 of FIG. 13: C _ij is 2 of ( _rule number) × (rule number)
In the dimensional area, the counter C _ij stores the number of times (frequency) in which the rank of the i-th (vertical axis) rule (horizontal axis) is j. The meaning of the counter 1300 will be clarified from the following description.

FIG. 14 is a flow chart showing the detailed processing of the effectiveness indication (step 607, FIG. 6) in the processing procedure of the inference status display section 6. First, the counter C
_ij (i, j = 1, number of rules) is set to 0 and initialized (1401). Then there Ke - scan I - against (1 ≦ I ≦ Ke number scan), or the flag F _I is set, that is whether the flag F _I is 1, examine whether (1402). When the flag F _I is 1, the representative conformance of each _rule G _Ir (r
= 1, the number of rules) is used to rank the rules.
The final rank number is the same as the total number of rules. The ranking method is in descending order of the representative suitability G _Ir . The rank of the i-th rule is O _i (i = 1, number of rules). For example, in the case of FIG. 2, the representative suitability G _{11 of} rule 1 ≦ rule −
Since the representative fitness G ₁₂ of rule 2 is, the rank O ₁ of rule _{1 is}
Is 2, and the rank O ₂ of rule ₂ is 1 (1403).

Each rule i (i = 1, rule number)
The counter C _ioi is _incremented by 1 using the rank O _i of. For example, in the case of FIG. 2, since the rank O _{1 of the} rule 1 is 2 and the rank O ₂ of the rule ₂ is 1, the counters C ₁₂ and C ₂₁ are set to 1.
Each is added (1404). Whether each processing of steps 1402 to 1405 has been executed for all cases,
It is checked (1405) whether or not there is an unexecuted case, the process jumps to step 1402, and steps 1402 to 1405 are processed for the unexecuted case. If the flag F _{I of a case I} is not set in step 1402, the process jumps to step 1405. In step 1405, when steps 1402 to 1405 have been executed for all the cases, the counter C _ij is displayed as shown in the effectiveness section 13 of the inference status display section 6.
Is output to the display device 31 (1406).

Next, it waits for the user to specify two points with the input device 32 (1407). Two points designated by the user (X1, Y1), (X2, Y2) (where X1 ≦ X
2, Y1 ≦ Y2), it is checked whether or not it is an end instruction. When the relationship between the two points is other than X1 ≦ X2 and Y1 ≦ Y2, X1 and X2 or Y1 and Y2 are exchanged so as to satisfy this relationship. If both points indicate the outside of the table, the end instruction is given (1408). If it is not the end instruction, the process jumps to step 1407. If it is the end instruction in step 1408, the processing for displaying the effectiveness is ended.

FIG. 15 shows a position table 1500 used for displaying the learning status. The position table 1500 is a cable
There are as many sets of data as there are spaces, and some propositional data DT _i
(I = 1, ..., Case number) and correction position EP _i (i =
1, ..., Case number). In FIG.
The position table 1500 is provided for each case i (i = 1, ...,
Data area 150 for setting data DT _i of the number of cases)
1 and a correction area 1502 for setting the correction position EP _i . Here, when the R-th registered proposition in the fuzzy knowledge base 4 is used in the if part of the rule,
The input data corresponding to the proposition is data DT _i , and the correction position EP _i is the fitness and the correction amount DE _Ri when the input data of the i-th data is given to the membership function. The sum of On the other hand, when the R-th proposition is used in the then part of the rule, the inference result Y _i obtained previously is data D
Let T _i be the correction position EP _{i, which} is the sum of the goodness of fit when the inference result Y _i is given to the membership function and the correction amount DE _Ri .

FIG. 16 is a flowchart showing the processing procedure of the learning status display section 8. First, the user inputs the input device 3
From 2, it is specified which proposition to display the learning status by using the number R (R = 1, the number of propositions) (160
1). Next, fuzzy knowledge from fuzzy knowledge base 4
Test data (a set of input data and teacher data) is read from the test data storage device 3 (1602) and the membership function in the fuzzy knowledge base 4 is learned. That is, the knowledge learning unit 7 is activated. In the knowledge learning unit 7, learning is repeated a predetermined number of times, and each proposition i (i = 1, ...
,, membership function M _ij (i =
1, ..., Proposition number, j = 1, ..., Parameter number), and correction amount DE _ij (i = 1, 1) of all cases in each proposition
.., the number of propositions, j = 1, ..., The number of cases) is passed to the learning status display unit 8 (1603). In the learning status display section 8, R
When the second proposition is used in the rule if section, the corrected position EP of each case i (i =, ..., Case number)
The _i, Ke - scan i input data of - a value obtained by adding the fit and correction amount DE _Ri of when given data to membership functions. When the R-th proposition is used in the rule-then part, the sum of the goodness of fit and the correction amount DE _Ri when the inference result Y _i is given to the membership function is set as the correction position EP _i . If the R-th proposition is used in the if-part proposition, the corresponding input data is the data DT _i (i = 1, ...
, And the number of cases), and if the R-th proposition is used in the then-part proposition, the inference result Y _i (i = 1, ...
..., the number of cases) as data DT _i (i = 1, ..., Number of cases), and the data area 1 in the correction position table 1500
Set to 501. In addition, all cases i (i = 1,
The corrected position EP _i (i = 1, ...) In the case number.
..., the number of cases) is set in the correction area 1502 in the correction position table 1500 (1604).

For a case I (1≤I≤case number), it is checked whether or not the flag F _I is on (160).
5). If the flag F _I is not on, the X-axis is the error D
“X” is plotted at the position where the T _I and Y axes are the corrected position EP _I (1606). In step 1605, when the flag F _I is set, “Δ” is plotted at a position where the X axis is the error DT _I and the Y axis is the corrected position EP _I (160).
7). After executing steps 1605 and 1606, it is checked whether or not "x" and "Δ" are drawn for all the cases (1608), and if there are unexecuted cases, step 160
Jump to 4. The learning status display unit 6 displays the learning status as shown in FIG.

At step 1608, all cases
When steps 1605 to 1607 have already been executed for the scan, the user waits for the user to specify two points with the input device 32 (1609). Next, the two points (X1,
Y1), (X2, Y2) (where X1 ≦ X2, Y1 ≦
From Y2), it is checked whether or not it is an end instruction (1610).
From the two points designated by the user, the data area and the designated correction area are designated as [X1, X2] and [Y1, Y2], respectively (1
611). For a case I (1≤I≤case number), the data DT _I of the data area 1501 belongs to the specified data area [X1, X2] and the correction position EP of the error area 1502. _I find out whether they belong to a specified correction area [Y1, Y2] (1612) . Data DT _I , correction position EP _I
When both of the cases satisfy the condition of step 1612, the flag F _I is set, in other words, the flag F _{I is set} to 1 (1613), and step 161 is executed for all cases.
It is checked whether or not steps 2 to 1613 have been executed (161
4). If there is an unexecuted case, the process jumps to step 1612 to check the unexecuted case. Step 16
If it has been executed in all cases in step 14, step 1
Jump to 609. If at least one of the data DT _I and the corrected position EP _I does not satisfy the condition at step 1612, the process jumps to step 1614. If there is an end instruction in step 1610,
The learning status display processing ends.

FIG. 17 is a flow chart showing the processing procedure of the data creating section 2. First, the raw data is read from the data base 1 and the test data is read from the test data storage device 3 (1701), and the data number i is set to 1 (1
702). Next, the i-th raw data is output to the display device 31. Specifically, in this embodiment, the raw data is displayed as a line graph. That is, when the data number i ≧ 2, a line segment connecting the (i−1) th raw data and the i-th raw data is drawn, and when i = 1, nothing is done (170
3). Position number table 50 of test data storage device 3
Position number NO _j (j = 1, number of cases) and data number i
It is checked whether there is a matching j (1704). When there is j such that the data number i = position number NO _j , the j is set as J. Further, it is checked whether the flag F _J is on (1705). Furthermore, when the flag F _J is set, the i-th raw data is output at the position where
“X” is displayed (1706). Next, the data number i
Is smaller than the number of raw data stored in the database 1, the data number i is incremented by 1 (1708) and the process jumps to step 1703. In step 1704, there is no data number i in the position table 50,
Alternatively, if the flag F _J is not set in step 1705, the process jumps to step 1707. When the raw data is drawn in this way, it becomes like the data creation unit 2 in FIG. Then, it waits for the user to specify two points with the input device 32 (1709) and determines whether or not the user gives an end instruction. Here, both of the two points are the display device 3
When pointing outside the graph being displayed above 1, it is judged that the user has instructed the end (1710). When it is not the end instruction, the test data is edited (1711) and step 1
Jump to 709 and wait until the end instruction is given. When it is determined in step 1710 that the instruction is an end instruction, the data creation process is ended.

In this embodiment, the description has been given only to the case where the number of the propositions (membership function) used in the then part is one, but when the number of the membership functions in the then part is two or more, the members of the then part are described. The rules may be divided so that there is one ship function. Further, in the present embodiment, the information about the flagged cases and the information about the non-flagged cases are displayed in different display methods, but only the information about the flagged cases is displayed. However, the information regarding the case where the flag is not set may not be displayed. Further, in the present embodiment, only the case where the case is designated and the flag is set in each part of the inference status display unit 6 and the learning status display unit 8 is described, but in order to enhance the editing function, the flag is set. A flag canceling unit for canceling the flag of the case in which is set may be provided. At this time, in the present embodiment, the case selection process (step 70
5, 904, 1209, 1407), it is possible to realize by providing a process of selecting an editing content such as what to do with the selected case.

Next, a second embodiment for realizing the present invention will be described.
This will be described in detail with reference to the drawings of FIGS. First,
FIG. 18 is a block diagram of a fuzzy chart analysis proto system in which fuzzy knowledge is constructed using the conversational construction support system of the present invention. In FIG. 18, 1801 is time-series data for storing securities prices (stock prices) of all issues, 1802.
Is external information that stores the number of issued shares, capital, profit forecast, etc., 1803 is knowledge that the boundary and range are clear, if-t
Rule base type knowledge base expressed by hen type rules
1804, the rule-based knowledge base 1803 is used to perform pattern matching, conflict resolution, and execution in sequence.
A rule-based type inference unit for obtaining a solution, 1805 is a specific stock extraction result display unit for outputting the execution result of the rule-based type inference unit 1804 to the display device 31, and 1806 is an execution result of the fuzzy inference unit 5. Is displayed on the display device 31. Rule-based knowledge base 18
03 and the function / construction method of the rule-based type inference unit 1804 are described in detail, for example, in the document "Method of Expert System Construction" by Hiroshi Tanaka / Yuichi Shimoi.

FIG. 19 shows a fuzzy knowledge base 4 used in the fuzzy chart analysis proto system. In FIG. 19, the horizontal direction is the subject, the vertical direction is the rule, and the inside thereof is the predicate. For example, rule 1 is “if average line change (current) is flattering & average line change (past)
is falling & average line divergence rate (current) is large & is above average line rationale is declining is then forecast is surely rising ”. 20
The table below shows the structure of a specific brand table. In FIG. 20, 20
01 is a brand table for setting a brand number.

FIG. 21 is a flowchart for explaining the operation of the fuzzy chart analysis proto system. In general, the financial / financial analysis proto system handles
In the securities world, an expert instantaneously and accurately predicts a security price from a past movement and buys or sells a security, and this system simulates the judgment of this expert. In FIG. 20,
First, the contents of the time series data 1801, external information 1802, rule base type knowledge base 1803, and fuzzy knowledge base 4 are read (2101). Next, it is determined whether or not a certain brand is a specific brand. Specifically, a rule-based knowledge base created based on the situation and preference of the user, external information 1802 such as the number of issued shares, capital, expected profit, and time-series data. The rule base type inference unit 1803 is executed using the data 1801 to perform classification by matching the numerical range and symbol data (21).
02). If the brand is a specific brand, the brand number is set in the brand table 2001 (2103). Then, steps 2102-21 for all issues
It is checked whether or not 03 has been executed (2104), and if there is an unexecuted brand, the process jumps to step 2102 to determine whether or not the unexecuted brand is a specific brand. Step two
When it is determined in 102 that the brand is not the specific brand, the process jumps to step 2104. In step 2104, steps 2102-
When 2103 has been executed, the specific brand extraction result display unit 1805 outputs the brand number in the specific brand table 2001 to the display device 31 (2105).

Then, one issue number in the issue table 2001 is taken out and it is determined whether or not fuzzy inference has been executed for that issue (2106). When the stock has not executed fuzzy reasoning, it judges the market view by fuzzy reasoning. That is, the fuzzy reasoning unit 5 uses the input data obtained from the external information 1802, the time series data 1801, etc., based on the fuzzy knowledge base 4 that stores the know-how of the market trend, and raises the stock price. , Or the degree of descending is determined (2107). Then, the market trend judgment result display unit 1806 displays the result on the display device 31.
(2108). It is checked whether steps 2106 to 2108 have been executed for all the stocks (2
109), when there is a specific stock that has not been executed, step 21
Jump to 06. Further, in step 2106, when fuzzy inference has been executed for that brand, the process jumps to step 2109. Step 210
At 9, when the fuzzy reasoning has been executed for all the specific stocks, all the processes related to the fuzzy chart analysis proto system are ended. As a result, the user of this system will decide which stock and how much to buy and sell, in consideration of the prediction result.

22 to 27, an example of how the fuzzy knowledge base 4 used in the fuzzy chart analysis proto system is modified will be described below. FIG. 22 is a screen display example of the data creating unit 2. In FIG. 22, the horizontal axis of the graph is time and the vertical axis is stock price. 2201 represents a period designated by the user. You
When the user designates the range to be the test data with the input device 32, the user uses a predetermined method to test the range such as the moving average line value and the stock price change rate.
Data is calculated and stored in the test data storage device 3. Figure 2
3 is a screen display example of the error distribution display portion 11. Figure 2
3, 3300 indicates a case range designated by the user. By enclosing it in a case frame specified by the user, the case is distinguished from other cases. FIG. 24 is a screen display example of the effectiveness display portion 13. In FIG. 24, 2401
Indicates the effectiveness of the rule. FIG. 25 is a screen display example of the compatibility / error display portion 12. In FIG. 25, information about the case designated by the user is indicated by "x",
Information regarding cases that are not designated is indicated by "○".
FIG. 26 is a screen display example of the learning status display unit 8. However, as the correction amount displayed in FIG. 26, only the correction amount relating to the case designated by the user is displayed. The method of displaying the amount of modification in each membership function uses different symbols (for example, x, ◯, etc.) depending on whether or not the amount of modification depends on the rule specified by the user. In FIG. 26, 2601 is an output part 260 for displaying the membership function and the correction amount of “stock price change (current) is rising” 260
Reference numeral 2 means an output portion displaying the membership function of "average line change (current) is rising" and the correction amount. FIG. 27 shows the fuzzy knowledge base 4 of the fuzzy chart analysis proto system used in the conversational construction support system of the present invention.
The processing procedure of explaining the operation of the modification is shown.

In FIG. 27, first, the fuzzy knowledge base 4 is created using the editor unit 30 (2701).
Next, the test data is created by the data creating unit 2,
It is stored in the test data storage device 3. For example, the data creation unit 2 outputs the contents of the data base 1 on the display device 31 as shown in FIG. The user specifies the period to be the test data while looking at the displayed contents. Now you
When the period specified by the user is the period 2201, the raw data (stock price) of each case belonging to this is edited and the input data is input.
After calculating the data and teacher data, the test data storage device 5
(2702). The knowledge learning unit 7 learns the membership function in the fuzzy knowledge base 4 based on the test data in the test data storage device 5. (270
3) Next, the fuzzy inference result using the membership function after learning is judged. For example, the user checks the difference between the teacher data and the fuzzy inference result by using the inference result display portion 10 and the error distribution display portion 11. When the inference result display portion 10 is used, a case whose error is far from the inference result is selected, and when the error distribution display portion 11 is used,
Select a case with a large error. In FIG. 23, the teacher data
A case (2301) having a large inference result, that is, a case having a large negative error is selected. (2704)
The effect display part 13 and the error / fitness display part 12 are used to check why the case does not match the teacher data, that is, which rule movement is inappropriate. First, the effectiveness indication portion 13 is used to check which rule was effective or not effective in the selected case. For example, step 27
When the movement of each rule in the period 2201 as shown in FIG. 22 is shown in 04 in FIG. 24, it can be seen that the third rule is most effective. With the track record of movement of this rule,
By comparing the images of the users, the difference was large.
Let the rule be an inappropriate rule. The user has already grasped the effective rule and the ineffective rule by the movement of the raw data at the time of creating the rule. What is the image of the user?
Refers to rules that worked when creating rules and rules that did not work. Alternatively, if the movement of the third rule in the case specified in step 2704 is as shown in FIG. 25, the degree of conformity is large when the error is large, so the movement of the third rule is certainly not correct. Appropriate. Further, according to the rule effect display portion 2401 of FIG. 24, in order to make this third rule an appropriate rule, it is necessary to reduce the degree of conformity or the error. I know that (2705).

Then, the selected rule is further examined by using the error / fitness display section 12 and the learning status display section 8 to check specifically which proposition is not appropriate,
Consider how to change it. For example, when the inappropriate rule is judged to be the third rule in step 2705, the membership function and the correction amount of each proposition are shown in FIG. 26, and the correction amount relating to the third rule is “x”.
Therefore, the correction amount for other rules is indicated by "○". At this time, since the correction amount of the third rule is different from the movements of the other rule correction amounts due to the output portion 2601, the “stock price change (current) is used for the third rule is”.
The membership function of "up" may be defined separately from other rules. Also, “change in average line (current) is rising”
In the case of, the amount of modification of the third rule and the amount of modification of other rules cannot be classified, but it is found from the output part 2602 that the input data has not changed, and the input data is also examined. Required (2707). In step 2707, when the membership function to be modified, the rule, and the content of the modification can be specified, the membership function is changed by the editor unit 30. If there is a problem with the input data, the data creating section 2 recreates it. For example, in the example of the output part 2601, the “stock price change (current)” of the third rule
The predicate of "is rising" is changed so as to be distinguishable from other rules such as "rising 1", and the membership function of "change in stock price (current) is rising 1" is redefined. Further, in the example of the output portion 2602, the meaning of the average line change (current) is changed. (2707) Step 2 above
After the execution of 701 to 2707, all the processes regarding the modification example of the fuzzy knowledge base 4 are finished. The fuzzy knowledge base 4 is constructed by repeating such processing. Further, the membership function to be changed may be specified by the rule if portion effect state portion 2403 of the effect degree display portion 13, and the change content may be determined by the error / fitness degree display portion 12. Alternatively, a test data of only the case selected in the inference result display portion 10 or the error distribution display portion 11
It is also possible to create the data and focus the fuzzy knowledge on this part.

From the above, as seen in the second embodiment, as a mode of use of the present invention, the user uses the inference result display portion 10 of the inference status display portion 6 to display the inference result and the teacher data. After selecting a case having a large difference with, check the effectiveness indication portion 13 of the inference situation display portion 6 to see which rule is effective or not effective in the designated case. While selecting the rule that should be noted,
Since the movement of the selected rule with respect to the specified case can be checked in the error / fitness display portion 12 of the inference status display portion 6, the inference status regarding the portion having the test data can be easily grasped. . As a result, it is possible to easily detect the rule to be modified and the part of the rule to be modified. As another usage mode, after the user selects a case having a large error in the error distribution display portion 11 of the inference status display unit 6,
In the same manner as the above-described usage pattern, the inference status regarding the portion having the test data can be easily grasped by the other items of the inference status display unit 6. As a result, it is possible to easily detect the rule to be corrected and the portion of the rule to be corrected.
Further, according to these embodiments, the user selects a case having a similar correction amount on the learning status display section 8 and then the inference result display section of the data creation section 2 or the inference status display section 6. With 10, it is possible to confirm where the selected case is located in the raw data or the test data while grasping the overall tendency, which is useful for adjusting fuzzy knowledge thereafter. For example, in the membership function of the learning status display unit 8 in FIG.
There is a correction amount indicating that the goodness of fit should be high for the same input, and a correction amount indicating that it should be low on the contrary. A membership function satisfying both of these cases is impossible. is there. Therefore, this membership function is 2
Although it is divided into two membership functions, if a case with a similar correction amount is specified, that portion is inferred as the inference result portion 1 of the data preparation unit 2 and the inference status display unit 6.
Since it can be projected to 0, it is useful for examining what kind of membership function should be created. Further, in the data creating section 2, since the case specified by the inference status display section 6 and the learning status display section 8 can be distinguished from other cases, the user can select only the specified case. Use as test data, or use only the part except the specified case as test data.
As described above, the data can be easily classified and effective test data can be created.

[0043]

According to the present invention, the user designates a part of the test data and outputs the inference status and the learning status of the part by changing the display method from the other parts. There is an effect that the inference situation and the learning situation of a part of the test data can be easily grasped comprehensively while observing the situation.

According to the present invention, the test data storage device stores the data of a plurality of cases, and the data of each case is
Since the configuration is made up of input data, teacher data, and a flag for determining whether or not the case is specified by the user, once the learning status display section and the inference status display section are used by the user. When the user designates a case, the designated case is distinguished from the non-designated cases after that, so that the inference status and learning status of a part of the test data can be easily grasped and the data can be created. Since the raw data is also discriminated and outputted even in the section, there is an effect that it is possible to pay attention to a portion having the raw data.

Further, according to the present invention, if only two points on the display device are designated in the inference status section and the learning status section, a rectangle is created based on the points and the case displayed in the rectangle is displayed. Since the data is judged as the case specified by the user, there is an effect that the test data can be efficiently classified.

In addition, according to the present invention, in the data creating section, the case specified by the user and the case not specified by the user are specified.
Since the data is output separately, the user can use the two types of data.
Since the test data can be specified while being aware of the data, there is also an effect that effective test data can be created.

According to the present invention, in the learning status display section and the inference status display section, the user can specify or cancel the case,
Since the case is interactively edited, there is an effect that the test data can be efficiently specified.

Further, according to the present invention, the learning status display section and the inference status display section can output only the information on the case designated by the user, so that the inference status of a part of the test data can be obtained. There is an effect that the learning situation can be easily grasped.

[Brief description of drawings]

FIG. 1 is a process related diagram of an embodiment for realizing the present invention.

FIG. 2 is a diagram for explaining a process of a fuzzy inference unit.

FIG. 3 is a diagram showing a configuration of an embodiment of a fuzzy knowledge interactive construction support system for implementing the present invention.

FIG. 4 is a diagram showing a structure of a data base for storing raw data such as time series data.

FIG. 5 is a diagram showing a configuration of a test data storage device for storing test data necessary for fuzzy inference and debugging of fuzzy knowledge.

FIG. 6 is a flowchart showing the processing in the inference status display section of the fuzzy knowledge interactive construction support system according to the present invention.

FIG. 7 is a flowchart showing the processing in the inference result display portion of the inference status display portion of the fuzzy knowledge interactive construction support system according to the present invention.

FIG. 8 is a diagram showing a configuration of a frequency table and an attention frequency table used for displaying an error distribution in an inference status display unit of a fuzzy knowledge interactive construction support system according to the present invention.

FIG. 9 is a flowchart showing the processing in the error distribution display portion of the inference status display portion of the fuzzy knowledge interactive construction support system according to the present invention.

FIG. 10 is a flowchart showing a process related to creation and display of an error table in an error distribution display portion of an inference status display portion of a fuzzy knowledge interactive construction support system according to the present invention.

FIG. 11 is a diagram showing the configuration of an error / fitness table used in the error / fitness display of the inference status display unit of the fuzzy knowledge interactive construction support system according to the present invention.

FIG. 12 is a flowchart showing the processing in the error / fitness display portion of the inference status display portion of the fuzzy knowledge interactive construction support system according to the present invention.

FIG. 13 is a diagram showing a table configuration of a counter used for displaying the effectiveness of the inference status display section of the fuzzy knowledge interactive construction support system according to the present invention.

FIG. 14 is a flowchart showing the processing in the effectiveness display portion of the inference status display portion of the fuzzy knowledge interactive construction support system according to the present invention.

FIG. 15 is a diagram showing a configuration of a position table used for displaying a learning situation of a fuzzy knowledge conversation type construction support method according to the present invention.

FIG. 16 is a flowchart showing the processing in the learning status display unit of the fuzzy knowledge interactive construction support system according to the present invention.
It is

FIG. 17 is a flowchart showing the processing in the data creating section of the fuzzy knowledge interactive construction support system according to the present invention.

FIG. 18 is a diagram showing the configuration of a fuzzy chart analysis proto system in which fuzzy knowledge is constructed using the conversational construction support system of the present invention.

FIG. 19 is a diagram showing a configuration of a fuzzy knowledge base 4 used in a fuzzy chart analysis proto system.

FIG. 20 is a diagram showing the structure of a specific brand table.

FIG. 21 is a flowchart for explaining the operation of the fuzzy chart analysis proto system.

FIG. 22 is a screen display example of the data creating section in the fuzzy chart analysis proto system.

FIG. 23 is a screen display example of an error distribution display portion in the fuzzy chart analysis proto system.

FIG. 24 is a screen display example of the effect display portion in the fuzzy chart analysis proto system.

FIG. 25 is a screen display example of a goodness of fit / error display portion in the fuzzy chart analysis proto system.

FIG. 26 is a screen display example of the learning status display unit 8 in the fuzzy chart analysis proto system.

FIG. 27 is a diagram showing an example of a process for changing the fuzzy knowledge base 4 of the fuzzy chart analysis proto system using the conversational construction support system of the present invention.

FIG. 28 is a diagram showing a detailed display example of the inference result screen in FIG. 1.

FIG. 29 is a diagram showing a detailed display example of the error distribution screen in FIG. 1.

FIG. 30 is a diagram showing a detailed display example of the error / fitness screen in FIG. 1.

FIG. 31 is a diagram showing a detailed display example of the efficacy screen in FIG. 1.

[Explanation of symbols]

 1 data base 2 data creation unit 3 test data storage device 4 fuzzy knowledge base 5 fuzzy inference unit 6 inference status display unit 7 knowledge learning unit 8 learning status display unit 10 inference result display unit 11 error distribution Display 12 Error / fitness display 13 Effectiveness display

Claims (11)

[Claims] 1. A fuzzy that has an if-then format and describes one or more propositions in the if part and then part.
And a fuzzy knowledge consisting of a membership function that defines the meaning of the proposition, and a method for supporting interactive construction of fuzzy knowledge in a fuzzy inference system that infers about one or more cases of input data. Input required for fuzzy inference from a test data storage device that stores one or more cases of test data consisting of input data and teacher data, and a data base that stores raw data such as time series data. A data creating unit for creating test data composed of data and teacher data, and a fuzzy inference executed by using input data of each case in the test data storage device based on the fuzzy inference. Depending on the inference result display section that outputs the inference result and / or various reference data for efficiently performing inference / learning to the display device, and the learning that automatically adjusts fuzzy knowledge,
A learning situation display unit that outputs the learning result in each case to a display device is provided, and the inference situation display unit and the learning situation display unit display in a certain area on the display screen when the area is selected and designated. Case designating means for designating each case, and the case
Flag setting means for setting a flag in the test data in the test data storage device corresponding to the case designated by the case designating means, and the case in which the flag is set when the information about each case is displayed. An interactive construction support system for fuzzy knowledge, characterized by comprising data display means for changing the display method of information regarding other cases and outputting it to a display device in a manner different from that for information regarding other cases. 2. The fuzzy knowledge interactive construction support method according to claim 1, wherein the test data storage device is used as an input value (state value) when performing fuzzy inference. Input data, one or more teacher data which means a correct value when fuzzy inference based on the input data, and the input data and the teacher data are data databases. A number (numerical value) indicating the order of raw data in the case and a flag (mark) indicating whether or not the case is specified by the case specifying means by the user. An interactive construction support system for fuzzy knowledge, characterized in that one or more pieces of test data each of which consists of four types of data are stored. 3. The fuzzy knowledge interactive construction support method according to claim 1, wherein the case designating means designates a plurality of points on a display screen according to a user's operation. A polygon is created using a plurality of points designated by the area designating means, and a case in which data exists in the polygon is a case designated by the user. An interactive construction support system for fuzzy knowledge, characterized by having a case determination means. 4. The fuzzy knowledge interactive construction support method according to claim 3, wherein the case determination means creates a rectangle based on the two points when the designated points are two. , A fuzzy knowledge interactive construction support system characterized in that a case in which data exists in the rectangle is determined as a case specified by the user. 5. The fuzzy knowledge interactive construction support method according to claim 1, wherein error distribution, error conformance, effectiveness, etc. are used as the various reference data. . 6. The fuzzy knowledge interactive construction support system according to claim 1, wherein said data creating unit stores test data for raw data of each case in the database. It is determined whether or not a flag (mark) is added to the corresponding test data in the device, and the raw data to which the flag (mark) is added is the data specified by the user. As raw data, the raw data which has not been added is judged as raw data not specified by the user, and the raw data of each case in the data base. The raw data output means for changing the display method of the raw data according to whether or not the case is specified by the user in the raw data determination means, and outputting the raw data to the display device. A fuzzy knowledge interactive construction support method characterized by having. 7. The fuzzy knowledge interactive construction support method according to claim 1, wherein a fuzzy inference unit for executing fuzzy inference and a test data are used.
A knowledge learning unit that automatically learns and adjusts fuzzy knowledge in the fuzzy knowledge base to suit the test data of the data storage device, and assists the user in changing the contents of the fuzzy knowledge base. An interactive construction support system for fuzzy knowledge, comprising: an editor unit, wherein the fuzzy inference status display unit activates a fuzzy inference unit and the learning status display unit activates a knowledge learning unit. 8. The fuzzy knowledge interactive construction support method according to claim 1 or 6, wherein a change in color, symbol, line type, line thickness, flicker or the like is used as a change in the display method. A fuzzy knowledge interactive construction support method characterized by the above. 9. The fuzzy knowledge interactive construction support method according to claim 1 or 6, wherein a flag is set when the information regarding each case is displayed.
A fuzzy knowledge interactive construction support method characterized by displaying only the information on the fuzzy knowledge. 10. The fuzzy knowledge interactive construction support method according to claim 1, further comprising flag releasing means for designating a case in which a flag has been set and releasing the flag of the case in the test data storage device. An interactive construction support method for fuzzy knowledge. 11. The fuzzy knowledge interactive construction support method according to claim 1, wherein when a certain area on the display screen is selected and designated, the display means outputs a frame surrounding the area to a display device. A fuzzy knowledge interactive construction support method characterized by the above.