JPH0877009A - Knowledge correction device - Google Patents

Abstract

(57) [Abstract] [Purpose] To provide a knowledge correction device capable of correcting confidence by flexibly changing relative strengths of adaptability and constancy in response to changes in approval rating. [Constitution] Inputting a support rate distribution of an opinion and a parameter for changing a certainty factor from the input / output device 1, and convincing the certainty factor of knowledge by a differential equation concerning the certainty factor determined in advance from the input support rate distribution and parameters. Of the certainty factor calculator 5 and the calculation result of the certainty factor calculator 5 and the parameters are input to the stability analyzer 6 via the controller 4, and the calculation time required to obtain the certainty factor distribution is extremely large. A long critical point is obtained by linear stability analysis with respect to the steady solution of the differential equation, and the result of this linear stability analysis is input to the parameter adjuster 7 so that the confidence distribution becomes the critical point at the support rate given by the user. Adjust the parameters to.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a knowledge correction device in an expert system, and more particularly to a knowledge correction device for correcting a knowledge base in response to a change in knowledge approval rate.

[0002]

2. Description of the Related Art An expert system has been developed which stores expert knowledge as a knowledge base and uses the knowledge to solve problems. FIG. 7 shows an outline of the configuration of the expert system. When the user inputs the initial condition of the problem through the user interface 11, the inference mechanism 1
The problem is solved by making an inference using the knowledge stored in the knowledge base 13 by the user 2. The knowledge of the expert is acquired by the expert by the knowledge acquisition mechanism 15 and stored in the knowledge base 13 through the knowledge base management mechanism 14.

By the way, for uncertain and ambiguous knowledge such as "a person with a black hair has dark eyes", "the altitude of Mt. Fuji is high", etc., the certainty is used to express certainty. Then
Furthermore, the inference result is also evaluated by the certainty factor. The certainty factor is described in, for example, Yoshiaki Shirai: Computer Mathematics Series, Theory of Artificial Intelligence (Corona Publishing).

In the acquisition of uncertain knowledge, there is a technique for correcting the certainty by referring to the opinion (true, false, high, low, etc.) on the knowledge. Usually, the opinion with a high approval rating is regarded as valid, and the confidence factor is adjusted so as to be in line with the opinion. Here, when the approval rating changes due to changes in the environment or the situation, the certainty factor is corrected in consideration of the following two properties.

"Adaptability": If the change in approval rate is large,
Modifying the confidence level allows you to adapt to the new environment. "Homeostasis": If the change in approval rating is small, it can be robust against environmental changes by maintaining the conventional certainty factor. Since these properties are contradictory to each other, the certainty factor is modified according to the target problem.

[0006]

It is necessary that the relative strengths of adaptability and homeostasis are not always consistent and can change depending on the situation. For example, it is necessary to strengthen adaptability and facilitate synchronization in response to opinions that have the influence of experts, but in the prior art, which property of adaptability or homeostasis should be emphasized? However, there is a problem that it is difficult to correct the certainty factor that attaches importance to the other property later because the certainty factor is fixed and the certainty factor is corrected.

The present invention has been made in view of the above,
It is an object of the present invention to provide a knowledge correction device capable of correcting a certainty factor by flexibly changing relative strengths of adaptability and constancy in response to a change in approval rating.

[0008]

In order to achieve the above object, a knowledge correction device of the present invention stores knowledge to which a certainty factor expressing the certainty of the knowledge is given by referring to the opinion from the outside. And an input / output device for inputting parameters for changing opinion support distribution and certainty factor and displaying the result of correction of knowledge certainty factor, And a controller for exchanging data, and a confidence factor of knowledge is calculated by a differential equation regarding a certainty factor determined from the support ratio distribution and the parameters input via the controller, and a calculation result is given to the controller. A confidence factor calculator that outputs a calculation result of the certainty factor calculator and the parameter specified by the user via the controller, and the calculation time required to obtain the confidence factor distribution is extremely long. A stability analyzer that obtains points by linear stability analysis for the steady solution of the differential equation and the linear stability analysis result by the stability analyzer are input, and the confidence distribution is critical at the approval ratio given by the user. It is a gist to have a parameter adjuster which adjusts the parameter so that it becomes a point and outputs the adjustment result to the controller.

[0009]

In the knowledge correction device of the present invention, a parameter for changing the opinion support distribution and the certainty factor is input, and knowledge is obtained by a differential equation relating to the certainty factor determined from the input support distribution and parameters. The confidence point is calculated, and the critical point where the computation time required to obtain the confidence distribution is extremely long is obtained by linear stability analysis for the steady solution of the differential equation, and the linear stability analysis result is input, and the The parameters are adjusted so that the confidence distribution becomes a critical point at a given approval rating.

[0010]

Embodiments of the present invention will be described below with reference to the drawings.

FIG. 1 is a block diagram showing the configuration of a knowledge correction device according to an embodiment of the present invention. In the figure,
The input / output device 1 includes an input device 2 and a display device 3.
Further, the confidence factor calculator 5 is connected via the input / output device 1 and the controller 4. Further, the controller 4 and the stability analyzer 6 are connected, the stability analyzer 6 and the parameter adjuster 7 are connected, and the parameter adjuster 7 is further connected to the controller 4.

In the knowledge correction device of this embodiment, when the approval rating changes, a new approval rating is calculated from the new approval rating and the current confidence. When this confidence is obtained by repeating this calculation, the value is taken as the confidence to be recorded in the knowledge base. This final certainty factor can be obtained as a steady solution of a differential equation composed of a parameter (hereinafter referred to as parameter) relating to relative strength of adaptability and constancy, with confidence as a time variable. . When a situation in which one opinion (Opinion 1) is dominant changes to a situation in which another opinion (Opinion 2) is predominant, in the parameter setting that emphasizes adaptability, the opinion 2 Although the certainty factor of is higher than the certainty factor of opinion 1, the setting of the parameter placing importance on the constancy maintains the state of the certainty factor of opinion 1 being higher than the certainty factor of opinion 2.
Such a qualitative relationship between the parameter and the certainty factor enables the user to obtain good results while adjusting the relative strengths of adaptability and homeostasis.

When the constancy is high, if the opinion rating of opinion 2 is increased, the certainty factor of opinion 2 becomes greater than the certainty factor of opinion 1 when the value exceeds a certain value, and the confidence factor distribution becomes It follows the distribution. Here, the fact that the certainty factor distribution follows the approval factor distribution means that the degree of approval factor of each opinion corresponds to the degree of certainty factor of each opinion. The steady-state solution of the differential equation whose confidence distribution greatly changes is hereinafter called a critical point. The steady-state solution, that is, the calculation time required to obtain the certainty distribution can be obtained by linear stability analysis of the steady-state solution of the differential equation. When the steady-state solution is the critical point, this calculation time is extremely long. There are features.

Therefore, even if a parameter having a strong constancy is first set, if the user requests the confidence distribution to follow the support distribution, the calculation time of the steady solution obtained by the linear stability analysis is extremely large. It is possible to meet the user's requirements by automatically adjusting the parameters so that the length becomes longer and setting the steady solution as the critical point. The linear stability analysis is described in, for example, Hiroshi Matano: Iwanami Course Applied Mathematics Series, Differential Equation I (Iwanami Shoten).

Next, the operation of the knowledge correction device shown in FIG. 1 will be described with reference to the flow chart shown in FIG.

The user uses the input device 2 to input the support ratio distribution and parameters (step S11). Controller 4
Outputs the input from the input device 2 to the confidence factor calculator 5,
The confidence calculator 5 calculates the confidence (step S12),
Output to the controller 4. When the controller 4 outputs the calculation result of the certainty factor to the display device 2, the certainty factor distribution is displayed to the user by the display device 2, and the user evaluates the result (step S13). If the result is not good, the user decides whether to re-enter the parameters (step S1).
4). When inputting again, the parameters are input using the input device 2. When the re-input is not performed, the controller 4 outputs the parameter and the certainty factor distribution to the stability analyzer 6, parameter adjustment is performed (step S15), and the certainty factor calculated by the adjusted parameter is displayed on the display unit. Two
Is output to Parameter adjustment will be described later.

Next, the calculation of the certainty factor will be described. The confidence factor calculation is performed by the confidence factor calculator 5 according to the following equation.

[0018]

[Equation 1] Is. The calculation is performed until X _i does not change with time (a steady solution is obtained). In the equation (1), the first term has the effect of bringing the certainty factor X _i closer to the approval rating P _i . The second term relates only to the certainty factor, and an opinion having a higher certainty factor has an effect of increasing the certainty factor. Also, the parameter b _ij
Relates to the evaluation of the difference in confidence between the opinion i and the opinion j. When it is large, an extreme confidence distribution is obtained, and when it is small, an even distribution is obtained. Therefore, the larger the value of c and the value of b _ij with respect to a, the stronger the constancy.

FIG. 3 shows an example of the processing result in this embodiment. In this example, in the case where there are two opinions (Opinion 1 and Opinion 2), when the support rate P ₁ of Opinion 1 is changed from 0 → 1 → 0, the change in the confidence factor X ₁ of Opinion ₁ is changed. This is shown (confidence factor of opinion 2 is X ₂ = 1−X ₁ ). FIG. 3A is an example in which only the effect of adaptability is observed when the parameter c is 0, and it is shown that the confidence distribution that is the same as the approval rating distribution is always obtained. FIG. 3B shows that when the parameter b _ij is 0, a more uniform confidence distribution can be obtained as compared with FIG. 3A. In Fig. 4 (c), when the parameters b _ij and c are increased to enhance the homeostasis, a critical point (indicated by a circle in the figure) appears, and the confidence level changes depending on the change in the approval rating. It shows that a hysteresis phenomenon appears.

When the certainty factor distribution becomes a critical point, there is a feature that the calculation time required to obtain the certainty factor distribution is extremely long, and this calculation time is obtained by linear stability analysis for the steady solution of the differential equation. be able to. When the real part (negative value) of the eigenvalue obtained by the linear stability analysis is small, the calculation time is short, and at the critical point, the eigenvalue is 0 at the maximum and the calculation time is extremely long. FIG. 4 shows a result of performing a linear stability analysis on the confidence distribution in each support distribution shown in FIG. 3C, and the calculation time becomes longer as the confidence distribution approaches the critical point. It shows that you are going.

Such a qualitative relationship between the parameter and the certainty factor allows the user to change the parameter so as to obtain a good result.

Next, the parameter adjustment will be described.
In the parameter adjustment, the parameter c is adjusted so that the steady solution of the equation (1) becomes a critical point in order to make the certainty distribution follow the support distribution.

FIG. 5 is a flow chart showing an outline of parameter adjustment. When the parameters and the confidence distribution are input from the controller 4 to the stability analyzer 6, the stability analyzer 6 performs a linear stability analysis of the confidence distribution (step S2
1), output to the parameter adjuster 7. The parameter adjuster 7 determines whether the maximum value of the real part of the eigenvalue obtained by the linear stability analysis is almost 0 (step S22). If not, the parameter c is reduced to enhance the adaptability ( Step S23), and output to the controller 4. When the controller 4 outputs the parameter and the support rate distribution to the certainty factor calculator 5, the certainty factor calculator 5 calculates the certainty factor using the adjusted parameter c (step S24) and outputs it to the controller 4. The above steps are repeated until the condition of step S22 is satisfied.

FIG. 6 is a diagram showing an example of the processing result of parameter adjustment. This example is shown in FIG.
In the process of increasing the support rate P ₁ of 0 → 1 when P ₁ = 0.6, the parameter is set to 0.07 so that the certainty factor becomes the critical point.
It is the example which adjusted from 0.05 to 0.05. As a result of this adjustment,
The certainty factor X ₁ of opinion ₁ is the certainty factor X _{2 of} opinion ₂ (= 1-
X ₁ ), and the confidence distribution was able to follow the approval rating distribution. FIG. 6A shows a process of adjusting the parameter c, and FIG. 6B shows that P ₁ = 0.x by using the parameter c whose eigenvalue is almost 0 in FIG. 6A. When it is 6, it indicates that the certainty factor distribution becomes a critical point.

As described above, according to the present invention, by handling the properties of adaptability and homeostasis in one framework, the user can flexibly change the relative strength of adaptability and homeostasis while ensuring confidence. It is possible to modify the degree.

Furthermore, by examining the linear stability of the confidence distribution, it is possible to automatically adjust the relative strengths of adaptability and homeostasis.

Although the knowledge correcting device of the present invention has been specifically described above based on the embodiments, the present invention is not limited to the above embodiments, and various modifications can be made without departing from the spirit of the invention. Needless to say.

[0028]

As described above, according to the present invention,
By adjusting the relative strength of adaptability and homeostasis through parameters, it is possible to handle them in one framework. Therefore, by adjusting the parameters for changes in approval rate,
It is possible to flexibly change the relative strength of adaptability and homeostasis to modify the certainty of knowledge.

[Brief description of drawings]

FIG. 1 is a block diagram showing the configuration of a knowledge correction device according to an embodiment of the present invention.

FIG. 2 is a flowchart showing the operation of the embodiment of FIG.

FIG. 3 is a diagram showing an example of a processing result of the embodiment of FIG.

FIG. 4 is a diagram showing an example of a processing result of the embodiment of FIG.

FIG. 5 is a flowchart showing an outline of parameter adjustment of the embodiment of FIG.

FIG. 6 is a diagram showing a processing result of parameter adjustment in the embodiment of FIG.

FIG. 7 is a block diagram showing a configuration of an expert system to which the knowledge correction device is applied.

[Explanation of symbols]

 1 Input / output device 2 Input device 3 Indicator 4 Controller 5 Confidence factor calculator 6 Stability analyzer 7 Parameter adjuster

Claims (1)

[Claims] 1. A knowledge correction device that corrects a knowledge base that stores knowledge that is given a certainty factor that expresses certainty of the knowledge, by referring to an opinion from the outside, and has a distribution of opinion support rates. A parameter for changing the certainty factor is input, an input / output device for displaying the correction result of the certainty factor of the knowledge, a controller for exchanging data with the input / output device, and an input via the controller. A confidence factor calculator that calculates a confidence factor of knowledge by a differential equation regarding a certainty factor determined from the support ratio distribution and the parameter, and outputs a calculation result to the controller, and the confidence factor calculator of the user's designation. The calculation result and the parameters are input via the controller, and the critical point, which requires an extremely long calculation time to obtain the confidence distribution, is obtained by the linear stability analysis for the steady solution of the differential equation. The stability analyzer and the linear stability analysis result by the stability analyzer are input, and the parameters are adjusted so that the certainty factor distribution becomes the critical point in the approval ratio given by the user, and the controller is controlled by the controller. And a parameter adjuster for outputting an adjustment result.