JPH04104302A - Approximate inference device - Google Patents

Abstract

PURPOSE:To improve the appropriateness of the inference result of approximate inference without giving a large burden to a user by correcting a membership function by means of learning in the actual use process of an approximate inference device. CONSTITUTION:Information whether a conclusion contained as the inference result of approximate inference from a user input means 30 as a conclusion appropriate information input means is the appropriate conclusion or not is given to a membership function learning part 26. Then, information existing in a knowledge base, which is stored in a knowledge base buffer 20, or a value having a possibility to become a threshold with which learning by the input of the user from the user input means 30 is executed are stored in a learning data buffer 28. Then, the membership function is learnt as to the conclusion designated from the user and the conclusion having the possibility more than the threshold in conclusions except for said conclusion. Knowledge altered by learning is written in the knowledge base buffer 20 and it is altered and stored in a knowledge base file 18.

Description

DETAILED DESCRIPTION OF THE INVENTION (Field of industrial application) The present invention relates to an approximate inference device, and more particularly to an approximate inference device that performs approximate inference using membership functions given for each conclusion for an event. It is something.

(Prior art) Approximate inference is based on a knowledge base constructed from the knowledge of experts, and approximate inference is performed by finding the possibility that a conclusion holds for an event using membership functions given for each conclusion. The device is already known as a fuzzy expert inference device, and is applied to machine failure diagnosis and the like.

In an approximate inference device such as the one described above, in order to obtain a correct inference result for an event, the knowledge must be appropriate, and the membership function given for each conclusion for the event must be determined based on this knowledge. Appropriate characteristics are required.

(Problems to be Solved by the Invention) In conventional approximate inference devices, in order to obtain a correct inference result for an event, in other words, in order to improve the appropriateness of the inference result of approximate inference, If there is an error in the inference result derived from the membership function based on the constructed knowledge base, especially if the knowledge base is a combination of knowledge from multiple experts, The user needs to consider whether the home knowledge is incorrect, that is, which expert's knowledge is actually inappropriate, correct the incorrect knowledge, redo the knowledge synthesis, and rebuild the knowledge base.

However, in reality, on the user side, finding incorrect knowledge from the knowledge base is a very difficult and troublesome task, and it takes time to synthesize knowledge and reconstruct the knowledge base. There is.

The present invention has been made by focusing on the problems mentioned above in conventional approximate inference devices, and the present invention is designed so that the membership function is modified through learning during the actual use process of the approximate inference device. It is an object of the present invention to provide an improved approximate inference device that improves the appropriateness of the inference results of approximate inference without imposing a large burden on the user, and allows a more appropriate conclusion to be obtained.

(Means for Solving the Problems) According to the present invention, the purpose of the two series is to provide an approximate inference device that performs approximate inference using membership functions given for each conclusion for an event. , a conclusion appropriateness information input means for inputting information regarding whether or not a conclusion obtained as an inference result of the approximate inference is a proper conclusion; and an inference result of the approximate inference based on the information inputted by the conclusion appropriateness information input means. This is achieved by an approximate inference device characterized by comprising: membership function modification means for modifying the membership function so as to improve the appropriateness of the membership function.

(Operation) According to the two-part configuration, the membership function is automatically corrected by the membership function correction means based on the information regarding the propriety of the inference result from the conclusion propriety information input means, and the appropriateness of the inference result is corrected. Go -1.

Modification of the membership function in order to improve the appropriateness of the inference result of approximate inference may be performed according to a defined algorithm determined according to the membership function creation algorithm. For example, if the membership function is The minimum value of a knowledge base if it is created based on the average value of the minimum value, standard deviation of the minimum value, average value of the maximum value, and standard deviation of the maximum value of the knowledge base by combining the knowledge of multiple experts. The average value, the standard deviation of the minimum value, the average value of the maximum value,
The standard deviation of the maximum value may be modified in accordance with information regarding the appropriateness of the inference result from the conclusion appropriateness information input means, according to a predetermined algorithm for improving the appropriateness of the inference result.

(Example) The present invention will be described in detail below with reference to the figures in appendix f.

FIG. 1 shows an embodiment of an approximate inference device according to the present invention. The approximate inference device includes an inference scheduler section 10, an event value input control section 12, an event value input buffer 14, an approximate inference execution device 16, a knowledge base file 18, a knowledge base buffer 20, and an inference output buffer 22. , a conclusion output control section 24, a membership function learning section 26,
It includes a learning data buffer 28 and user input means 30.

The inference scheduler unit 10 controls the timing of event input and conclusion output for the event value input control unit 12 and conclusion output control unit 24, and generates a knowledge base based on information existing in the knowledge base stored in the knowledge base buffer 20. control for storing data specifying the input light of the event value and the specification of the output light of the conclusion in the event value input buffer 14 and the conclusion output buffer 22, respectively; and controlling the membership function learning for the membership function learning section 26. Timing control is performed for each.

The event value input control unit 12 reads the event value into the event value input buffer 14 according to the timing given by the inference scheduler unit 10, and also reads each input light according to the event value input control unit stored in the event value input buffer 14. It is designed to control reading data from.

The event value input buffer 14 includes a keyboard 32, a PLC (
Various event values can be received from the programmable controller online 34 and the event value file 36, these are stored as a data table, and the event values are output to the approximate inference execution device code 6 as inference inputs.

The approximate inference execution device 16 is configured to execute approximate inference according to timing given by the inference scheduler section 10.

Approximate inference in the approximate inference execution device 16 uses the membership function created for each conclusion for each event from the knowledge stored in the knowledge pace buffer 20. The inference result from this approximate inference is output to the inference output buffer 22.

The knowledge pace buffer 20 is provided with the knowledge of a plurality of experts from the knowledge base file 18, and stores a converted rule definition data table, an event data table, and a fitness data table.

Here, an example of how to synthesize knowledge using interval data of two experts and create a membership function based on this will be described.

In this example, the knowledge base file 18 is
The knowledge rules of X J and expert EX2 are stored in the form shown below.

Expert EXJ: If'20≦F]≦60. 0≦F2≦40 Lhe
n CIIf40≦F1≦80.60≦F2≦100
then C2 expert EX2: If30≦F1≦50.10≦F2≦30 then
CIIf50≦F1≦70.70≦F2≦90 L
hen C2 From now on, Fl is event 1, F
2 may be called event 2, C1 may be called conclusion 1, and C2 may be called conclusion 2.

In the case of two sets of rules, the relationship between events and conclusions stored in the knowledge pace buffer 20 is established for each expert.
This will be as shown in Tables 1 and 2.

Note that Table 1 is for Expert EX, later, and Table 2 is for Expert EX2.

Table 1 Table 2 Synthesizing the knowledge of multiple experts involves, for example, calculating the average value and standard deviation of multiple experts for the maximum and minimum values of each event involved in each conclusion. Then, the membership function for each conclusion is calculated using the Gaussian distribution (normalized distribution) using the average value of the maximum value, the average value of the minimum value, the standard deviation of the maximum value, and the standard deviation of the minimum value. Creating is done.

Next, a specific example of this knowledge synthesis will be explained by taking up an example of a rule for determining conclusion 1 (C1) from event 1 (F, , ).

The knowledge of each expert before synthesis is expressed by the relational expression shown below.

Expert EX: If20≦Fj≦60 then
C1 expert EX: If'30≦F1≦50
then C1 The average value Amax of the maximum value and the average value Am of the minimum value of each event in the knowledge of the two experts as described above
1n is calculated according to the following formula.

Amax = (60+50) /2 =55Amin
-(20+30) /2 = 25The standard deviation of the maximum value of each event S(χ) may and the standard deviation of the minimum value S(χ
) min is calculated according to the following formulas.

When the above-mentioned calculation of combining the expert's knowledge is performed for all the maximum and minimum values of each event involved in each conclusion, a table as shown in Table 3 is obtained.

In approximate inference, a membership function is given to each event, and here, as an example, a method of creating a membership function using a Gaussian distribution will be shown. When creating a membership function using a Gaussian distribution, first use the knowledge synthesis results shown in Table 3 and calculate the degree to which the input data (event value) χ fits the event, that is, the degree of fit, according to the formula below. is calculated.

The input data value (event value) for which the goodness of fit is determined by Φ(χ) S(χ) max and is 0.5 is the average value Am1n, Ama
Determined by x.

Furthermore, Figure 6 shows the event F1 due to the synthetic knowledge shown in Table 3.
7 shows the membership function of the conclusion C1 in the event F1, and FIG. 7 similarly shows the membership function of the conclusion C2 in the event F1.

The conclusion output buffer 22 receives the inference result from the approximate inference execution device 16, creates a conclusion data table, and outputs the inference result to the display 38 and PLC according to the conclusion cutting edge designation.
It is designed to be output online 40 and to a conclusion file 42.

The membership function learning unit 26 uses the user input means 30, which serves as a conclusion appropriateness information input means, as an inference result of approximate inference. The information existing in the knowledge base stored in the knowledge base buffer 20 or the user input means 30 is given information regarding whether the M concluded conclusion is a proper conclusion.
The probability value that becomes the threshold for learning based on the user's input is stored in the learning data buffer 28, and the value specified by the user is stored. Degree to which the data χ fits the event (degree of fit) Gauss (y): Value of the Gaussian distribution at the input y This results in a Gaussian distribution as shown in FIG.

When creating the membership function, use the Gaussian distribution shown in Figure 2, as shown by the symbol G,
Only the right half of the Gaussian distribution is used.

That is, Gauss (z −Amin /S (χ)
min l has a distribution as shown in Figure 3, and G
auss(z-Amax /S (χ) max l
becomes the distribution shown in Figure 4, and the membership function shown in Figure 5 is created by combining the distribution shown in Figure 3 and the distribution shown in Figure 4. Ru.

In this case, the slope of the distribution is the standard deviation S(χ) min, and the membership function is learned for the conclusion whose probability value is equal to or higher than the threshold for other conclusions, and the membership function is changed by learning. The acquired knowledge is written into the knowledge pace buffer 20, and further updated and stored in the knowledge base file 18.

When learning this membership function, as shown in Figure 8, the area between the lower limit value of the event value and the second limit value is divided into seven intervals (a) to (g). Considerations are given for modifying membership functions.

Here, the intervals (a) and (g) each represent the fitness degree ω1j (event value χj in the membership function of the conclusion C1 and the event Fj
This is the interval in which the fitness degree) is 0, and the interval (b) to (f
') is an interval where the fitness degree ω1j is not 0, and among these,
Interval (b) is an interval in which the event value χj is less than or equal to the minimum average value Am1n, and interval (C,) is an interval in which the event value χj is greater than the minimum average value Am1n, but the fitness level 0 months does not become 1. The interval (d) is an interval in which the fitness degree ωjj is the highest value of 1, and the interval (e) is an interval in which the event value χj is smaller than the maximum average value A max but the fitness degree ω1j is 1.
The interval (f') is an interval in which the event value χj is greater than or equal to the maximum average value Amax.

In this embodiment, the conclusion to be learned has a probability value Pi that is greater than or equal to the correct conclusion Co specified by the user and a probability threshold P of other conclusions. Suppose that the conclusion is Cj. Incidentally, Ci≠Co, 1≦i≦n (where n=total number of conclusions), and the probability threshold P is determined by the knowledge base data stored in the knowledge pace buffer 20 or by input from the user. knowledge pace buffer 2
It is stored in 0.

First, the correct conclusion C specified by the user as appropriate.
Learning of the membership function of event Fj related to O will be explained.

Here, the fact that the conclusion and the event are related means that more than one expert has defined rules for the conclusion and the event.

In this learning, events Fj (
1≦j≦I'n, nl - total 1r quadrants), the average value Am1n of the minimum value of this event and the conclusion is set so that the goodness of fit at the event value χj of the event Fj increases.
, the average value Amax of the maximum value, or the standard deviation S of the minimum value
(χ)l11jn, standard deviation of maximum value S(χ) may
is changed as follows depending on which interval of the two series the event value χj falls in.

(1) When the event value χj is within the interval range of (a) or (b), reduce the average value Am1n of the minimum value and change the slope part of the membership function on the minimum side as shown in Fig. 6 or 7. Look at it and move it parallel to the left.

The average value Am1n of the minimum values may be reduced, for example, according to the following formula.

Am1n - Amin - (j - limit value - lower limit value)
5/100) However, the second limit value is the upper limit value of the event value χj, and the lower limit value is the lower limit value of the event value χj.

(2) When the event value χj is within the interval range of (e), the standard deviation S(χ)min of the minimum value is decreased to increase the slope of the membership function on the minimum side.

The standard deviation S(χ) min of the minimum value may be reduced, for example, according to the following formula.

S (χ) min = S (χ) mjn xo, 9
(3) When the event value χJ is within the interval range of (d), the fitness is at the maximum value (-1), so in this case nothing is changed.

(4) When the event value χj is within the interval range of (e), the standard deviation S(χ) max of the maximum value is decreased to increase the slope of the membership function on the maximum side.

The standard deviation S(χ) max of the maximum value may be reduced, for example, according to the following formula.

S (χ) ma, x = S (χ) max xo,
9 (5) If the event value χj is within the interval range of (f) or (g), increase the average value A max of the maximum value and change the slope part of the membership function on the maximum side to that shown in Fig. 6 or 7. Move it parallel to the right side as seen in the figure.

The average value A max of the maximum values may be decreased, for example, according to the following formula.

Amax = Amax 10, Ix limit value - lower limit value)
(5/100) For example, the event F of conclusion CI as shown in Figure 6]
In the membership function, if a value of 70 is given as the event value χj, this is within the interval range of interval (g), so the average value Amax of the maximum value is updated as follows.

Amax - Amax + (J: limit value - lower limit value) X (5/1
．． 00 )=55+ (1,00-0) x (5/1
00 ) = 60 As a result, the average value Amax of the maximum values increases from 55 to 60, and when a membership function is created with this updated average value Amax = 60, this becomes the 12th
As shown by the solid line in the figure, compared to the membership function before learning, which is shown by the broken line in the figure, the slope part on the maximum side moves in parallel to the right in the figure, and the event Value χj=
The fitness at 70 will increase.

Next, in conclusions other than conclusion Co, 1-
Learning of the membership function of an event Fj related to a conclusion C1 having a probability value Pi, that is, an incorrect conclusion will be explained.

In this case too, the fact that the conclusion and the event are related means that more than one expert has defined the rules regarding the conclusion and the event.

In this learning, events Fj (
For the membership function (1≦i≦mSm−total number of events), the average value Avn of the minimum value and the average value of the maximum value in this event and the conclusion are set so that the goodness of fit at the event value χj of the event Fj decreases. A max or standard deviation of the minimum value S(χ) m
in , the standard deviation of the maximum value S(χ) may be changed as follows depending on which interval the event value χj falls within.

(1) When the event value χj is within the interval range of (a) or (g), the fitness is the minimum value (-0), so in this case, nothing is changed.

(2) If the event value χj is within the interval range of (l) 1
The standard deviation S(χ) min of the minimum value is decreased to increase the slope of the membership function on the minimum side.

The standard deviation S(χ) min of the minimum value may be reduced, for example, according to the following formula.

S (Z) min = S (χ) min Xo, 9
(3) The absolute value of the difference between the minimum average value Am1n and the event value χj is smaller than the absolute value of the difference between the maximum average value Amax and the event value χj, and the event value χj is (C) or (d ), the average value Am1n of the minimum values is increased, and the slope portion of the membership function on the minimum side is translated to the right as seen in FIG. 6 or 7.

The average value Am1n of the minimum values may be increased, for example, according to the following formula.

Amjn = Amin + (upper limit value - lower limit value)
(5/100) (4) The absolute value of the difference between the average value Am1n of the minimum values and the event value χj is greater than the absolute value of the difference between the average value Amax of the maximum values and the event value χj, and the event value χj is (d ) or (e), the average value Amax of the maximum value is decreased, and the slope portion of the membership function on the maximum side is translated to the left as seen in FIG. 6 or 7.

The average value Amax of the maximum values may be decreased according to the following formula.

Amax = Amax - (1 - limit value - lower limit value)
5/100) For example, event F1 of conclusion C2 as shown in FIG.
In the membership function, if the event value χj is 70, this is within the interval range of interval (e), so the average value A max of the maximum value is updated according to the following formula.

Amax - Amax - (upper limit value - lower limit value) X (5/1.
00 )-75-(100-0) x (5/1.00
) As a result, the average value A max of the maximum values changes from 75 to 7.
0, and the average value of this updated maximum value Amax=
When the membership function is created with 70, the 13th
As shown by the solid line in the figure, the maximum slope part of the membership function moves in parallel to the left in the figure, compared to the membership function before learning, which is shown by the broken line in the figure. As a result, the fitness at the event value χj=70 decreases.

(5) When the event value χj is within the interval range of (f), the standard deviation S(χ) max of the maximum value is decreased to increase the slope of the membership function on the maximum side.

The standard deviation S(χ) max of the maximum value may be reduced, for example, according to the following formula.

S (χ) max = S (χ) max Xo, 9
Next, a specific algorithm for learning a membership function such as a double series will be explained using FIGS. 9 to 11.

FIG. 9 shows the main routine of the membership function learning algorithm, which is repeatedly executed for each conclusion C1. First, in step 10, the user selects a correct conclusion Co from the user input means 30. is specified.

In step 20, it is determined whether or not the conclusion Ci is designated as the correct conclusion Co.
If it is specified, proceed to step 40;
If not, proceed to step 30.

In step 30, it is determined whether the probability P in the conclusion Ci is less than or equal to the threshold P. P1≧P
If so, that is, a conclusion Ci whose probability P1 is greater than the threshold P even though the conclusion Co is not specified by the user is an inappropriate conclusion C1, and the process proceeds to step 50; otherwise, the process proceeds to step 50. No learning is performed regarding conclusion Ci.

In step 40, proper time processing is performed according to the flowchart as shown in FIG. 10, whereas in step 50, inappropriate time processing is performed according to the flowchart as shown in FIG. will be held.

Next, the appropriate time processing shown in FIG. 10 will be explained.

This proper time processing is repeated for all events Fj, and first, in step 100, it is determined whether or not event Fj is defined by a correct conclusion Co. Only when the event value Fj is defined by the correct conclusion CO does the routine proceed to step 1.10, otherwise the routine ends.

That is, only the events Fj related to the correct conclusion CO are learned.

In step 110, it is determined whether the fitness degree ωjj is 1 or not. When ω1j=1, that is, when the event value χj is within the range of interval (d), this routine ends because nothing is changed; otherwise, the routine proceeds to step 120. .

In step 120, it is determined which is larger, the difference (absolute value) between the average value Am1n of the minimum values and the event value χj, or the difference (absolute value) between the average value Amax of the maximum values and the event value χj. It will be done. If the absolute value of Am1n-χj is smaller than the absolute value of Amax-χj, it is assumed that the event value χj is on the minimum side of the membership function and the process proceeds to step 130; otherwise, the event value χj is on the maximum side of the membership function. If there is, the process proceeds to step 140.

In step 130, it is determined whether the event value χj is smaller than the minimum average value Am1n. χj≦A
m1n, the event value χj is in the interval (a> or (b
), in which case the process advances to step 150; otherwise, the event value χj is within the interval range of interval (C), in which case the process advances to step 160. move on.

In step 140, it is determined whether the event value χj is larger than the maximum average value A max. χj≧
If A11laX, this means that the event value χj is within the interval (f) or (g), and in this case, the process proceeds to step 180; otherwise, the event value χj
Step 1 assumes that is within the interval range of interval (e).
Proceed to 70.

Step 150 is executed when - is double (1), in which case the average value Am1n of the minimum values is decreased.

Step 160 is executed in the case of - double (2), in which the standard deviation of the minimum value S(χ)min is decreased.

Step 170 is executed in the case of ``double (4),''
In this case, the standard deviation of the maximum value S(χ)max is reduced.

Step 180 is executed in the case of - (5), in which the average value Amax of the maximum values is increased.

Next, the inappropriate processing shown in FIG. 11 will be explained. This proper time processing is repeated for all events Fj, and first, in step 200, it is determined whether or not event Fj is defined by a correct conclusion Co. The process proceeds to step 210 only when the event value Fj is defined by the correct conclusion Co, otherwise the routine ends.

That is, only the events Fj related to the correct conclusion Co are learned.

In step 210, it is determined whether the fitness degree ω1j is O. When ω1j=0, that is, when the event value χj is within the interval (a) or (g), this routine ends because nothing is changed,
Otherwise, the process proceeds to step 220.

In step 220, the difference (absolute value) between the average value Am1n of the minimum values and the event value χj and the average value Ama,x of the maximum values are calculated.
A determination is made as to which difference (absolute value) between the event value χj and the event value χj is larger. The absolute value of Am1n-χj is AnlaX-χ
If it is smaller than the absolute value of j, the event value χj is determined to be on the minimum side of the membership function and the process proceeds to step 230; otherwise, the event value χj is determined to be on the maximum side of the membership function and the process proceeds to step 240.

In step 230, it is determined whether the event value χj is smaller than the minimum average value Am1n. χj≦A
If m1n, the event value χj is within the interval range of interval (b). In this case, the process advances to step 250; otherwise, the event value χj is within the interval (C) or (d). In this case, the process proceeds to step 260.

In step 240, it is determined whether the event value χj is larger than the maximum average value Amax. χj≧A
If it is max, it means that the event value χj is within the interval range of the interval (''), and in this case, the process advances to step 280; otherwise, the event value χj is within the interval (d) or (e). If it is within the range, step 270
Proceed to.

Step 250 is executed when the condition (2) is -, in which case the standard deviation S(χ)min of the minimum value is decreased.

Step 260 is executed in case (3) above, in which the average value Am1n of the minimum values is increased.

Step 270 is executed in the case of (4) above,
In this case, the average value Amax of the maximum values is decreased.

Step 280 is executed in case (5) described in lx,
In this case, the standard deviation S(χ)max of the maximum value is reduced.

In the embodiment described in -, the membership function learning unit 26 executes learning if a correct conclusion is specified by the user, and the membership function learning unit 26 stores the specified conclusion and the event value at that time in the learning data buffer. 28, and learning may be executed when the load on the CPU is light. In this case, the amount of negative I on the CPU is reduced.

In the above embodiment, the membership function is trained on conclusions that have a probability of -1 or more than a threshold value between the conclusion specified by the user as correct and other conclusions.
The membership functions to be learned may be a conclusion specified as correct by the user and a conclusion specified as incorrect with respect to the output inference result. In this case, the 14th
As shown in the figure, in addition to specifying the correct conclusion CO, an incorrect conclusion Cuo is specified, and step 30
In this case, instead of determining whether the possibility Pj is greater than or equal to the threshold P, the conclusion Cuo specified that the conclusion Ci is incorrect is determined.
A determination is made as to whether or not C1=Cuo, and if C1=Cuo, the process proceeds to step 50 and an inappropriate processing as shown in FIG. 11 is executed.

(Effects of the Invention) As can be understood from the following two explanations, in the approximate inference device according to the present invention, the information regarding the suitability of the inference result from the conclusion suitability information input means allows the membership function modifying means to The membership function is automatically corrected and the appropriateness of the conclusion is improved, so even if the initially created membership function is somewhat less appropriate, its appropriateness will improve as it is learned through use. This eliminates the need for users to modify membership functions.

This eliminates the need for the user to update or reconstruct the knowledge base when the membership function is created based on a knowledge base that is a combination of knowledge from multiple experts. In cases where the knowledge base is used for failure diagnosis, even if the reliability of the initially created knowledge base is somewhat low, through learning during diagnosis, the knowledge base will develop a reliability that reflects the characteristics of the equipment being diagnosed. The knowledge base is automatically updated to a higher level knowledge base, thus eliminating the need for the user to maintain the knowledge base. For example, if there are several identical machines, even if the same knowledge base is initially used for each machine, this knowledge base will automatically change over time and change depending on the operator's input as each machine is individually diagnosed. The knowledge base is updated to match the characteristics of each machine depending on how it is used, etc., and because it is highly reliable and suitable for each machine, diagnosis becomes possible.

[Brief explanation of the drawing]

Fig. 1 is a schematic configuration diagram showing an embodiment of an approximate inference device according to the present invention, Fig. 2 is a graph showing a Gaussian distribution for creating membership functions, and Figs. Graphs showing the procedure for creating membership functions, Figures 6 and 7 are graphs showing specific examples of membership functions based on synthesized knowledge, and Figure 8 is event values for learning membership functions. 9 is a flowchart showing the main routine of the membership function learning algorithm, FIG. 10 is a flowchart showing the processing routine when it is appropriate, and FIG. 11 is the processing routine when it is inappropriate. flowchart shown)
・, FIG. 12 and FIG. 13 are graphs each showing an example of modification by membership function learning, and FIG. 14 is a flowchart 1 showing another example of the main routine of the membership function learning algorithm. 10...Inference scheduler unit 12...Event value input control unit 14...Event individual buffer 16...Approximate inference execution device 18...Knowledge base file 20...Knowledge base buffer 22...Conclusion Output buffer 24...Conclusion output control unit 26...Membership function learning unit 28...Learning data buffer 30...User input means Patent applicant Omron Corporation Agent Patent attorney Kazutoshi Wakuchi Faction 1 charcoal faction

Claims (1)

[Claims] 1. In an approximate inference device that performs approximate inference using a membership function given to each conclusion for an event, the conclusion obtained as an inference result of the approximate inference is a proper conclusion. a conclusion appropriateness information input means for inputting information regarding whether the conclusion is appropriate or not; and a member for modifying the membership function so as to improve the appropriateness of the inference result of the approximate inference based on the information input by the conclusion appropriateness information input means. An approximate inference device comprising: ship function correction means;