JPH10333912A - Fuzzy rule preparation method and device therefor - Google Patents

Abstract

PROBLEM TO BE SOLVED: To facilitate the setting change of the sample data of a new belonging class by performing self-organized learning for the respective belonging classes of sample patterns and storing information based on the result of the self- organized learning in a memory for the respective belonging classes. SOLUTION: The sample patterns stored in a sample pattern class information storage device 14 are divided into the respective belonging classes and respectively stored in a sample pattern division storage part 26. By the stored sample patterns, in a learning part 28, the self-organized learning is successively performed for the respective belonging classes of the sample patterns. The parameters of a membership function are successively calculated for the respective belonging classes from weighting vectors calculated for the respective belonging classes as the result of the self-organized learning and the sample patterns by a function parameter calculating part 30 and then, the function parameters are stored in a function parameter storage part 32 as the memory for the respective belonging classes.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

The present invention relates to a fuzzy rule used for identifying a class to which an unknown pattern such as character recognition or voice recognition belongs by performing self-organizing learning of a Kohonen-type neural network on sample data. A method and apparatus for making.

[0002]

2. Description of the Related Art An example of a conventional method for creating a fuzzy rule for pattern recognition using sample data prior to the recognition of an unknown pattern is described in Reference 1: "11th Fuzzy System Symposium Lecture Papers, pp815-pp. 81
8, 1995 ", entitled" Extraction of Fuzzy Rules by Fuzzy Neural Network Using Structure Learning and Forgetting Learning ". According to the technique disclosed in this document, first, as described as a comparative example in the section of embodiments of the invention described later, first, a Kohonen-type neural network is used by using a sample pattern of each belonging class included in sample data. Perform self-organizing learning of Then, an initial fuzzy rule is created using the weight vector obtained as a result of the self-organizing learning. Further, according to the technology disclosed in this document, the fuzzy rule for pattern recognition is created by performing forgetting structure learning of the fuzzy neural network for the initial fuzzy rule.

[0003] Reference 2: "The 12th Fuzzy System Symposium Lecture Papers, pp 193-196, 19
No. 96 "discloses a method of creating a fuzzy rule which can be divided into a typical rule and an exceptional rule as a method of creating a fuzzy rule which can obtain a higher correct answer rate.

[0004]

[0005] However, Document 1
And the conventional fuzzy rule creation method disclosed in Document 2, the structure of the Kohonen-type neural network and the structure of the fuzzy neural network are determined depending on both dimensions of input / output vectors of sample data. Therefore, when adding a new belonging class to be identified, the creation of the fuzzy rule had to be started from the beginning.

In addition, according to the conventional fuzzy rule creation method, since self-organizing learning and forgetting structure learning are each performed several thousand times, it takes a lot of calculation time to create fuzzy rules. For this reason, it is difficult for the conventional fuzzy rule creation method to easily adapt to a change in the setting of the sample data such as a change in the sample pattern of a new belonging class.

For this reason, a fuzzy rule creating method which can shorten the processing time and can easily adapt to a change in setting of sample data such as an additional change of a sample pattern of a new belonging class. The realization of the device has been desired.

Further, according to the conventional fuzzy rule creation method disclosed in Reference 2, trial and error are required in a method of combining a typical fuzzy rule and an exceptional fuzzy rule.

For this reason, it has been desired to realize a new method and apparatus for subdividing a part of the fuzzy rules to create more appropriate fuzzy rules.

[0009]

The inventor of the present application has made various studies and calculations, and as a result, if self-organizing learning of a Kohonen-type neural network is performed for each class to which a sample pattern belongs, a fuzzy rule is created. When adding a sample pattern belonging to a new belonging class after performing the above, the result of the self-organizing learning for the sample pattern belonging to the new belonging class is added to the result of the self-organizing learning already obtained. The present invention has been found that the fuzzy rule can be created by the above-mentioned method, that is, there is no need to redo the self-organizing learning for all the sample patterns.

Therefore, according to the fuzzy rule creating method of the first aspect of the present invention, self-organizing learning of a Kohonen-type neural network is performed for a plurality of sample patterns, and the result of the self-organizing learning is used by using the result of the self-organizing learning. In creating a fuzzy rule, self-organizing learning is performed for each class to which a sample pattern belongs, and information based on the result of the self-organizing learning is stored in a memory for each class to which the self-organizing learning belongs.

As described above, if the self-organizing learning is performed for each class to which the sample pattern belongs, it can be easily adapted to the case where an additional sample pattern belonging to a new belonging class is added after the fuzzy rule is created. That is,
A fuzzy rule can be created by adding the result of self-organizing learning for an additional sample pattern belonging to a new belonging class to the already obtained result of self-organizing learning.

For example, self-organizing learning of a Kohonen-type neural network is performed for each sample pattern belonging to either class A or class B, and fuzzy rules are created using the results. A case where a fuzzy rule is created by adding an additional sample pattern belonging to C will be described. In this case, the self-organizing learning need only be performed for the additional sample pattern. That is, it is not necessary to repeat self-organizing learning for sample patterns belonging to classes A and B. A fuzzy rule created by adding an additional sample pattern can be used to classify an unknown pattern into one of classes A to C.

Therefore, according to the fuzzy rule creating method of the present invention, it is possible to easily adapt to the setting change of the sample data such as the addition and change of the sample pattern of the new belonging class.

Further, according to the fuzzy rule creation method of the present invention, since the self-organizing learning of the Kohonen-type neural network is performed for each belonging class, the self-organizing learning of different belonging classes is performed in parallel, that is, simultaneously. By proceeding, the processing time required for self-organizing learning can be reduced. As a result, the processing time required for fuzzy rule creation can be reduced.

In the fuzzy rule creating method of the present invention, preferably, as a result of the self-organizing learning,
The weight vector of each competitive layer unit of the Kohonen-type neural network is calculated for each belonging class, and from the weight vector and the sample pattern, the parameter of the antecedent membership function of the fuzzy rule is calculated. It is preferable to calculate a certainty parameter corresponding to the consequent real value of the fuzzy rule from the subject part membership function, the sample pattern, and the class information indicating the class to which the sample pattern belongs.

In calculating the parameters of the antecedent membership function, more preferably, the parameters of the antecedent membership function are calculated for each class.
The calculated parameters are preferably stored in a memory for each belonging class as information based on the result of self-organizing learning.

As described above, if the parameters of the membership function of the antecedent part are also calculated for each belonging class, it is easier to add an additional sample pattern belonging to a new belonging class after creating a fuzzy rule. Can be adapted to. That is, for the additional sample pattern, not only the self-organizing learning but also the calculation results of the parameters of the antecedent membership function can be used to create a fuzzy rule in addition to the parameters of the antecedent membership function already obtained. .

For example, self-organizing learning of a Kohonen-type neural network is performed for each of the sample patterns belonging to either class A or class B for each belonging class. A case where a fuzzy rule is created by adding an additional sample pattern belonging to class C after calculating and creating a fuzzy rule will be described. In this case, the self-organizing learning and the calculation of the parameters of the antecedent part membership function need only be performed for the additional sample pattern. That is, it is not necessary to repeat the self-organizing learning and the calculation of the parameters of the antecedent membership function for the sample patterns belonging to the classes A and B. A fuzzy rule created by adding an additional sample pattern can be used to classify an unknown pattern into one of classes A to C.

Therefore, according to the fuzzy rule creation method of the present invention, it is possible to more easily adapt to a change in setting of sample data such as an additional change of a sample pattern of a new belonging class.

Furthermore, when not only the self-organizing learning of the Kohonen-type neural network but also the calculation of the parameter (function parameter) of the membership function of the antecedent part is performed for each belonging class, the self-organizing learning of the belonging classes different from each other is performed. If the calculation of the function parameters is performed in parallel, that is, simultaneously, the processing time required for the self-organizing learning and the calculation of the function parameters can be reduced. As a result, it is possible to further reduce the processing time required for fuzzy rule creation.

In calculating the parameters (function parameters) of the membership function of the antecedent part, more preferably, the weight vector calculated for each belonging class is stored in a memory as information based on the result of self-organizing learning. Then, the parameters of the membership function of the antecedent part are preferably calculated for the weight vectors of all the belonging classes and the sample patterns of all the belonging classes.

As described above, when the function parameters are calculated for all the belonging classes, the antecedent part reflecting the tendency of the entire distribution of the sample pattern is better than when the function parameters are calculated for each belonging class. The parameters of the membership function can be calculated. As a result, a more appropriate fuzzy rule can be created.

On the other hand, when calculating the function parameters for each belonging class, the function parameters are calculated based on the local distribution of the sample pattern for each belonging class. May not be accurately reflected.

When classifying unknown patterns by class using fuzzy rules, it is desirable that sample patterns belonging to different classes belong to different fuzzy rules. However, in the created fuzzy rules, the range of the fuzzy rules may extend over the sample patterns of two or more belonging classes. In this case, it is desirable to divide the fuzzy rules into at least fuzzy rules for each class.

If only one class belongs to a sample pattern belonging to a certain fuzzy rule, it belongs to one certain class among certainty parameters corresponding to the consequent real value of the fuzzy rule. Only the value of the certainty parameter indicating the certainty of the other class is higher than the value of the other certainty parameter indicating the certainty belonging to another belonging class. On the other hand, when a certain fuzzy rule extends over sample patterns of two or more belonging classes, the values of two or more certainty parameters of the certainty parameters of the fuzzy rule indicate relatively high values. .

Therefore, in the fuzzy rule creation method of the present invention, preferably, the value of the certainty parameter for two or more specific belonging classes of the belonging classes is
Each fuzzy rule that has a value larger than the reference value is determined as a specific fuzzy rule, and instead of the specific fuzzy rule, a divided fuzzy rule corresponding to each of the specific belonging classes is created. The parameters of the antecedent membership function of the division fuzzy rule for each specific affiliation class are calculated from the specific sample pattern and its class information, and the antecedent membership function calculated for this parameter, the specific sample pattern and From the class information, a certainty parameter corresponding to the consequent real value of each of the divided fuzzy rules may be calculated.

As described above, a specific fuzzy rule extending over two or more belonging classes is determined, and a function parameter is calculated for each class using a specific sample pattern belonging to the determined specific fuzzy rule. If a division fuzzy rule is created instead of a rule, it is possible to divide the specific fuzzy rule by class to which it belongs. If the specific fuzzy rules can be divided for each class, more appropriate fuzzy rules can be obtained.

By the way, even if the above dividing process is performed,
The range of the created divided fuzzy rule may cover two or more belonging classes. For example, when the number of divided fuzzy rules is smaller than the number of belonging classes included in the original specific fuzzy rule, the range of the divided fuzzy rule extends to two or more belonging classes.

Therefore, when the divided fuzzy rule is created, more preferably, the number of the divided fuzzy rules created in place of the specific fuzzy rule is the number of the specific class to which the specific sample pattern belonging to the specific fuzzy rule belongs. If the number is smaller than this, a new fuzzy rule may be created instead of the specific fuzzy rule from the specific sample pattern belonging to the specific fuzzy rule and the class information indicating the class to which the sample pattern belongs.

As described above, if a new fuzzy rule is created instead of a specific fuzzy rule using a specific sample pattern belonging to a specific fuzzy rule extending over two or more belonging classes, the class to which the specific fuzzy rule belongs Another division can be achieved. If the specific fuzzy rules can be divided for each class, more appropriate fuzzy rules can be obtained.

By the way, even if the above-mentioned further division processing is performed, the range of the newly created fuzzy rule may cover two or more belonging classes. For example,
If the number of new fuzzy rules is smaller than the number of belonging classes included in the original specific fuzzy rule, the range of the new fuzzy rule extends to two or more belonging classes.

Therefore, when a new fuzzy rule is created, more preferably, the number of new fuzzy rules created in place of the specific fuzzy rule is determined by the specific sample pattern belonging to the specific fuzzy rule. When the number of classes is smaller than the number of classes, the number of competitors in the Kohonen neural network is increased from the number when a new fuzzy rule is created. It is characterized in that self-organizing learning of a network is performed, and a modified fuzzy rule is created using the result of the self-organizing learning.

As described above, after increasing the number of competitive layers of the Kohonen-type neural network, the self-organizing learning is performed on a specific sample pattern belonging to a specific fuzzy rule extending over two or more belonging classes, whereby the identification is performed. If a modified fuzzy rule is created instead of the fuzzy rule, it is possible to divide the specific fuzzy rule by class to which it belongs. If the specific fuzzy rules can be divided for each class, more appropriate fuzzy rules can be obtained.

Further, when the modified fuzzy rule is created by increasing the number of competitive layers of the Kohonen-type neural network, more preferably, the specific fuzzy rule is added until the number of the modified fuzzy rules becomes equal to or more than the number of the specific belonging class. A modified fuzzy rule may be created each time the number of competing layers is increased by one with the number of sample patterns belonging to the upper limit set. However, the number of competitive layers does not necessarily need to be increased to the upper limit.

According to the fuzzy rule creating apparatus of the second aspect of the present invention, self-organizing learning of a Kohonen-type neural network is performed for a plurality of sample patterns, and the result of the self-organizing learning is used by using the result of the self-organizing learning. In a fuzzy rule creation device that creates fuzzy rules, a Kohonen-type neural network learning unit that performs self-organizing learning for each class to which a sample pattern belongs, and the result of self-organizing learning, as information based on the result of self-organizing learning, Memory to store for each class
A parameter calculating unit for calculating a parameter of the fuzzy rule using a result of the self-organizing learning.

As described above, if the self-organizing learning is performed for each class belonging to a sample pattern in the Kohonen-type neural network learning unit, a fuzzy rule is created and then an additional sample pattern belonging to a new belonging class is added. Easily adaptable. That is, a fuzzy rule can be created by adding the result of the self-organizing learning for the additional sample pattern belonging to the new belonging class to the already obtained result of the self-organizing learning.

For example, self-organizing learning of a Kohonen-type neural network is performed for each of the sample patterns belonging to either class A or class B, and fuzzy rules are created using the results. A case where a fuzzy rule is created by adding an additional sample pattern belonging to C will be described. In this case, the self-organizing learning need only be performed for the additional sample pattern. That is, it is not necessary to repeat self-organizing learning for sample patterns belonging to classes A and B. A fuzzy rule created by adding an additional sample pattern can be used to classify an unknown pattern into one of classes A to C.

Therefore, according to the fuzzy rule creating apparatus of the present invention, it is possible to easily adapt to the setting change of the sample data such as the addition and change of the sample pattern of the new belonging class.

Further, according to the fuzzy rule creation apparatus of the present invention, a Kohonen-type neural network learning unit (hereinafter, also simply referred to as "learning unit") for performing self-organizing learning of the Kohonen-type neural network for each belonging class. Therefore, if the self-organizing learning of different affiliation classes is performed in parallel in the learning unit, that is, in parallel, the processing time required for the self-organizing learning can be reduced. As a result, the processing time required for fuzzy rule creation can be reduced.

Further, in the fuzzy rule creating apparatus of the present invention, preferably, the parameter calculating section is constituted by a membership function parameter calculating section and a certainty factor parameter calculating section. The section calculates the parameters of the membership function of the antecedent part of the fuzzy rule from the weight vector calculated for each class as a result of self-organizing learning and the sample pattern of each competitive layer unit of the Kohonen-type neural network. Then, the confidence parameter calculation unit calculates the confidence parameter corresponding to the consequent part real value of the fuzzy rule from the calculated antecedent membership function of this parameter, the sample pattern and the class information of the sample pattern. good.

Further, in the fuzzy rule creating apparatus according to the present invention comprising the membership function parameter calculating section and the certainty parameter calculating section, more preferably, the membership of the antecedent part is calculated from the weight vector calculated for each belonging class. A membership function parameter calculation unit that calculates function parameters for each belonging class, and a memory that stores the parameters calculated for each belonging class as information based on the result of self-organizing learning for each belonging class. good.

As described above, if the membership function parameter calculating unit for calculating the parameters of the membership function of the antecedent part for each belonging class is provided, the additional sample pattern belonging to the new belonging class is created after the fuzzy rule is created. It can be more easily adapted when adding. That is, for the additional sample pattern, not only the self-organizing learning but also the calculation results of the parameters of the antecedent membership function can be used to create a fuzzy rule in addition to the parameters of the antecedent membership function already obtained. .

For example, self-organizing learning of a Kohonen-type neural network is performed for sample patterns belonging to either class A or class B for each belonging class, and the parameters of the antecedent membership function are classified for each belonging class. A case where a fuzzy rule is created by adding an additional sample pattern belonging to class C after calculating and creating a fuzzy rule will be described. In this case, the self-organizing learning and the calculation of the parameters of the antecedent part membership function need only be performed for the additional sample pattern. That is, it is not necessary to repeat the self-organizing learning and the calculation of the parameters of the antecedent membership function for the sample patterns belonging to the classes A and B. A fuzzy rule created by adding an additional sample pattern can be used to classify an unknown pattern into one of classes A to C.

Therefore, according to the fuzzy rule creation device of the present invention, it is possible to more easily adapt to the setting change of the sample data such as the addition and change of the sample pattern of the new belonging class.

Further, when a membership function parameter calculating unit for calculating the parameters (function parameters) of the membership function of the antecedent part for each belonging class is provided, the learning unit performs self-organizing learning and membership of different belonging classes. If the calculation of the function parameters in the ship function parameter calculation unit is performed in parallel, that is, simultaneously, the processing time required for the self-organizing learning and the calculation of the function parameters can be reduced. As a result, it is possible to further reduce the processing time required for fuzzy rule creation.

Further, in the fuzzy rule creating apparatus according to the present invention comprising a membership function parameter calculating section and a certainty factor parameter calculating section, more preferably, the weight vector calculated for each belonging class is used for self-organizing learning. As information based on the result, a memory for storing, and a membership function parameter calculating unit for calculating parameters of the antecedent part membership function for the weight vectors of all the belonging classes and the sample patterns of all the belonging classes are included. I hope you get it.

As described above, the provision of the membership function parameter calculation unit for calculating the function parameters for all the belonging classes makes it possible to reduce the entire distribution of the sample pattern compared to the case of calculating the function parameters for each belonging class. The parameter of the membership function of the antecedent part which reflects the tendency of the above can be calculated. As a result, a more appropriate fuzzy rule can be created.

When classifying unknown patterns by class using fuzzy rules, it is desirable that sample patterns belonging to different classes belong to different fuzzy rules. However, in the created fuzzy rules, the range of the fuzzy rules may extend over the sample patterns of two or more belonging classes. In this case, it is desirable to divide the fuzzy rules into at least fuzzy rules for each class.

If the sample pattern belonging to a certain fuzzy rule has only one class to which it belongs, of the certainty parameters corresponding to the consequent real values of the fuzzy rule, it belongs to one class to which it belongs. Only the value of the certainty parameter indicating the certainty of the other class is higher than the value of the other certainty parameter indicating the certainty belonging to another belonging class. On the other hand, when a certain fuzzy rule extends over sample patterns of two or more belonging classes, the values of two or more certainty parameters of the certainty parameters of the fuzzy rule indicate relatively high values. .

Therefore, in the fuzzy rule creating apparatus of the present invention, it is preferable that the values of the confidence parameters for two or more specific belonging classes of the fuzzy rules are larger than the reference value. A fuzzy rule determining unit that determines the fuzzy rule that has become a specific fuzzy rule, and a specific sample belonging to this specific fuzzy rule to create a divided fuzzy rule corresponding to each of the specific belonging classes instead of the specific fuzzy rule The parameters of the antecedent membership function of the divided fuzzy rule for each specific belonging class are calculated from the pattern and its class information, and the antecedent membership function calculated for this parameter, the specific sample pattern and its class are calculated. From information, split fuzzy file And a fuzzy rule division unit for calculating each certainty parameter corresponding to the real value of the consequent part of each rule.

As described above, the determining unit that determines the specific fuzzy rule extending over two or more belonging classes and the function parameters are calculated for each class using the specific sample pattern belonging to the specific fuzzy rule determined by the determining unit. Thus, if a fuzzy rule dividing unit for creating a divided fuzzy rule is provided instead of the specific fuzzy rule, division by class to which the specific fuzzy rule belongs can be achieved. If the specific fuzzy rules can be divided for each class, more appropriate fuzzy rules can be obtained.

By the way, even if the above dividing process is performed,
The range of the created divided fuzzy rule may cover two or more belonging classes. For example, when the number of divided fuzzy rules is smaller than the number of belonging classes included in the original specific fuzzy rule, the range of the divided fuzzy rule extends to two or more belonging classes.

Therefore, when the divided fuzzy rule is created, more preferably, the number of the divided fuzzy rules created in place of the specific fuzzy rule is equal to the number of the specific belonging class to which the specific sample pattern belonging to the specific fuzzy rule belongs. A new fuzzy rule instead of the specific fuzzy rule from a divided fuzzy rule determining unit that determines whether the number is less than the specified fuzzy rule, and a class information representing a class to which the sample pattern belongs and a specific sample pattern belonging to the specific fuzzy rule. It is desirable to have a new fuzzy rule creation section for creating rules.

As described above, by using a specific sample pattern belonging to a specific fuzzy rule extending over two or more belonging classes and creating a new fuzzy rule instead of the specific fuzzy rule, the belonging class of the specific fuzzy rule can be obtained. Another division can be achieved. If the specific fuzzy rules can be divided for each class, more appropriate fuzzy rules can be obtained.

By the way, even if the above-described further division processing is performed, the range of the newly created fuzzy rule may cover two or more belonging classes. For example,
If the number of new fuzzy rules is smaller than the number of belonging classes included in the original specific fuzzy rule, the range of the new fuzzy rule extends to two or more belonging classes.

Therefore, when a new fuzzy rule is created, more preferably, the number of new fuzzy rules created in place of the specific fuzzy rule is determined by the specific sample pattern belonging to the specific fuzzy rule. A new fuzzy rule determination unit that determines whether it is less than the number of classes to which it belongs, and the number of competitive layers in the Kohonen-type neural network are increased from the number when a new fuzzy rule is created, and It is preferable that a self-organizing learning of the Kohonen-type neural network is performed for the specific sample pattern to which it belongs, and a modified fuzzy rule creating unit that creates a modified fuzzy rule using the result of the self-organizing learning is preferable.

As described above, after increasing the number of competitive layers of the Kohonen-type neural network, self-organizing learning is performed on a specific sample pattern belonging to a specific fuzzy rule extending over two or more belonging classes, thereby specifying the specific fuzzy rule. If a modified fuzzy rule creating unit for creating a modified fuzzy rule instead of a fuzzy rule is provided, it is possible to divide a specific fuzzy rule by class to which it belongs. If the specific fuzzy rules can be divided for each class, more appropriate fuzzy rules can be obtained.

Further, when the modified fuzzy rules are created by increasing the number of competing layers of the Kohonen-type neural network, more preferably, the specific fuzzy rules are maintained until the number of the modified fuzzy rules becomes equal to or greater than the number of the specific belonging class. It is preferable to provide a modified fuzzy rule creating unit for creating a modified fuzzy rule every time the number of competitive layers is increased by one with the number of sample patterns belonging to the upper limit. However, the number of competitive layers does not necessarily need to be increased to the upper limit.

[0059]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS An example of a fuzzy rule creation method and apparatus according to the present invention will be described below with reference to the accompanying drawings. It should be noted that the drawings referred to merely schematically show the sizes, shapes, and arrangements of the components so that the present invention can be understood. Therefore, the present invention is not limited to the illustrated example.

(First Comparative Example) First, prior to description of an example of a fuzzy rule creating method and apparatus of the present invention, in order to facilitate understanding of the present invention, the fuzzy rule creating method disclosed in the above-mentioned reference 1 will be described. This will be briefly described as a comparative example.

This reference 1 describes a method of creating some fuzzy rules from a large number of sample data by self-organizing learning of a Kohonen type neural network and structural learning with forgetting of a fuzzy neural network, and a pattern identification problem. Application examples are disclosed.

First, for the sake of explanation, when N sample data obtained by combining an n-dimensional input vector and an m-dimensional output vector are given, the input vector and the output vector in the p-th sample data are respectively expressed by Ip =
[ _Ip1 , _ip2 , ..., _ipn ], Op = [ _op1 , _op2 , ...,
o _pm ] (1 ≦ p ≦ N). At this time, if the input variable vector is X = [x ₁ , x ₂ ,..., X _n ] and the output variable vector is Y = [y ₁ , y ₂ ,..., Y _m ], the generated fuzzy rule is Generally, it is given by the following (1).

R ⁱ : If x ₁ is A _i1 and x ₂ is A _i2 and ... and x _n is A _in then y ₁ is b _i1 and y ₂ is b _i2 and ... and y _m is b _im ... ( 1) Here, R ⁱ (1 ≦ i ≦ r) is the label of the i-th fuzzy rule, r is the number of rules, A _ij (1 ≦ j ≦ n) is the fuzzy variable of the antecedent part, b _ik (1 ≦ k ≦ m) is a consequent part real value. In addition, the A _ij membership function A _ij (x
_j ) is defined by a Gaussian function, for example, as in the following equation (2).

A _ij (x _j ) = exp {− ((x _j −c _ij ) / d _ij ) ² } (2) where c _ij and d _ij are respectively A _ij (x _j ) Are parameters that determine the center value and the width of.

The parameters, the number of rules r, the consequent part real value b _ik , and the central value c _{ij in the} above equations (1) and (2)
The width d _ij is determined from the sample data by the following procedures 1) to 5).

1) First, the initial rule number r ₀ is set with the sample data number N as an upper limit. In the above document, a value of about half of N is recommended.

2) Next, the input vector Ip and the output vector Op of each sample data are combined into one (n + m) -dimensional vector Sp, as shown in the following equation (3), and the average of each element of Sp is calculated. And the variance are equal to each other.

Sp = [Ip, Op] = [ _ip1 , _ip2 ,..., _Ipn , _op1 , _op2 ,..., _Opm ] (3) Next, the number of units in the input layer Performs self-organizing learning on Sp in a Kohonen-type neural network in which (n + m) and the number of units in the competitive layer is r ₀ .

[0069] 4) Next, the weight vector W with each competitive layer unit after learning _{i = [w i1, w i2} , ..., w in, w
_An initial fuzzy rule is generated from _{i (n + 1)} , wi _{(n + 2)} ,..., wi _{(n + m)} ] (1 ≦ i ≦ r ₀ ). Specifically, the initial values of the parameters b _ik , c _ij and d _ij are respectively set as in the following equations (4) to (6).

B _ik = _{wi (n + k)} (4) c _ij = w _ij (5) d _ij = max (| u _ij -c _ij |, | l _ij -c _ij | (6) where u _ij and l _ij are the maximum value and the minimum value of the input variable x _j of the sample data belonging to the i-th competitive layer unit, respectively.

5) Next, a fuzzy neural network is constructed based on the obtained initial fuzzy rules, and tuning (adjustment) and pruning (pruning) of each parameter are performed by a back propagation method with forgetting (error back propagation). I do. As a result, final values of the rule number r, the consequent part real value b _ik , the center value c _ij, and the width d _ij are determined.

The fuzzy rules created in this way can be used for pattern recognition of an unknown pattern whose belonging class is unknown by simplified fuzzy inference shown in the following equations (7) and (8).

[0073]

(Equation 1)

Here, μ _i (X) is the input vector X =
.., X _n ] represents the degree of conformity indicating how much the [x ₁ , x ₂ ,..., X _n ] matches the antecedent part of the i-th fuzzy rule.
Y ^* _k is the k-th component of the inference value vector Y ^* = [y ^* ₁ , y ^* ₂ ,..., Y ^* _m ] corresponding to the output of the system. Then, the unknown pattern is classified into a class k in which the component y ^* _k is the maximum in pattern recognition.

(Second Comparative Example) Next, a method of creating a fuzzy rule disclosed in the above-mentioned document 2 will be briefly described as a second comparative example.

Reference 2 discloses a method of dividing a fuzzy rule into a typical fuzzy rule and an exceptional fuzzy rule based on an initial fuzzy rule created by self-organizing learning of a Kohonen-type neural network. Is disclosed. Therefore, in this method, the following processes 1) to 6) are sequentially performed for the structure learning with forgetting of the fuzzy neural network based on the initial fuzzy rule.

1) First, an appropriate number of rules, a correct answer rate, and the number of times of learning are set as learning end conditions for the structure learning with forgetting.

2) Next, the given sample data is subjected to forgetting structure learning a predetermined number of times of learning separately from 1).

3) Next, the sample data is input to the fuzzy neural network being trained, and the sample data is divided into two data groups, ie, a correct answer data group which is correct and an incorrect answer data group which is incorrect.

4) Next, a group of incorrect answer data is removed from the sample data.

5) Next, until any of the learning end conditions set in the processing of 1) above is satisfied, the above-mentioned 2) to
The process of 4) is repeated to create a typical fuzzy rule.

6) Next, the above-mentioned processes 2) to 4) are performed again on the incorrect answer data group removed in the process 3) to create an exceptional fuzzy rule.
However, the learning end condition at this time is only when the number of times of learning set in the above 1) is satisfied.

Next, a description will be given of a case in which pattern recognition is performed on an unknown pattern whose belonging class is unknown using the typical fuzzy rule and the exceptional fuzzy rule created as described above.

In this case, first, the simplified fuzzy inference shown in the above equations (7) and (8) is independently performed on the input unknown pattern for the typical fuzzy rule and the exceptional fuzzy rule.

Next, the class to which the unknown pattern belongs is determined by combining the inference value vector obtained for the typical fuzzy rule and the inference value vector obtained for the exceptional fuzzy rule. In Literature 2, five types of trial and error are performed for this combination.

(Pattern Recognition Apparatus) Next, prior to description of an example of a fuzzy rule creation method and apparatus of the present invention,
To facilitate understanding of the present invention, an example of a general pattern recognition device including a fuzzy rule creation device will be described with reference to FIG.

FIG. 2 is a block diagram for explaining the pattern recognition apparatus 100.

The pattern recognition device 100 includes a pattern input device 10, a switch 12, a sample pattern / class information storage device 14, a fuzzy rule creation device 16, a fuzzy rule storage device 18, a pattern identification device 20, and an identification result output device 22. It is configured as follows.

The pattern input device 10 is a part for inputting sample data or unknown patterns as information relating to a pattern to be subjected to recognition processing to the pattern recognition device 100. As the sample data and the unknown pattern, for example, pattern information such as characters and voices, and information obtained by combining some feature amounts obtained by appropriate feature extraction processing from the information are input.

The switch 12 converts the pattern information input from the pattern input device 10 into the fuzzy rule creation device 16 (A side in FIG. 2) or the pattern identification device 20 (B side in FIG. 2). ) Is a part to select which one to send. For example, when creating a fuzzy rule for recognition prior to an unknown pattern recognition process, the switch 12 is turned on to the A side. As a result, the sample data input from the pattern input device 10 and the sample data including the class information thereof are sent to the sample pattern / class information storage device 14.
On the other hand, when performing the unknown pattern recognition process, the switch 12 is turned on to the B side. As a result, the unknown pattern input from the pattern input device 10 is sent to the pattern identification device 20.

The sample pattern / class information storage device 14 is a portion for storing the multidimensional sample patterns and the class information representing the classes to which the multidimensional sample patterns are sent from the pattern input device 10 via the switch 12.

The fuzzy rule creating device 16 creates a fuzzy rule necessary for recognition of an unknown pattern based on a multidimensional sample pattern and its class information. The configuration and operation of the fuzzy rule creation device 16 will be described in detail in an embodiment described later.

The fuzzy rule storage device 18 is a portion for storing the recognition fuzzy rules created by the fuzzy rule creating device 16.

The pattern identification device 20 is a part for identifying the class to which the unknown pattern belongs, using the recognition fuzzy rule, for the unknown pattern sent from the pattern input device 10 via the switch 12. still,
The details of the unknown pattern recognition process will be described later in the third section.
A description will be given after the embodiment.

The identification result output device 22 is a portion that outputs class information indicating the class to which the unknown pattern belongs, identified by the pattern identification device 20.

(First Embodiment) Next, referring to FIG. 1, as an example of the fuzzy rule creation device 16 of the above-described pattern recognition device 100, the fuzzy rule creation device of the first embodiment of the present invention will be described. An example of the device will be described. FIG.
FIG. 2 is a functional block diagram for describing a fuzzy rule creation device according to the first embodiment.

The fuzzy rule creating device 16 of the first embodiment performs self-organizing learning of a Kohonen-type neural network for a plurality of sample patterns,
The fuzzy rule is created using the result of the self-organizing learning. The fuzzy rule creation device 16 includes a sample pattern division storage unit 26, a Kohonen-type neural network learning unit (learning unit) 2
8. Membership function parameter calculation unit (function parameter calculation unit) 30, membership function parameter storage unit (function parameter storage unit) 32, certainty parameter calculation unit 34, belonging pattern set / class information storage unit 36, and fuzzy recognition It comprises a rule judging / dividing unit 38. The function parameter calculation unit 30 and the certainty parameter calculation unit 34 constitute a parameter calculation unit 40 that calculates fuzzy rule parameters using the results of self-organizing learning.

This sample pattern division storage unit 26
This is a section that divides the sample pattern stored in the sample pattern / class information storage device 14 according to its belonging class and stores each.

Further, the learning unit 28 calculates the sample pattern stored in the sample pattern division
In this section, the self-organizing learning of the Kohonen-type neural network is sequentially performed, particularly for each class to which the sample pattern belongs.

The self-organizing learning for each belonging class is performed by a conventionally known method. Conventionally known self-organizing learning of Kohonen-type neural networks is disclosed in, for example, Reference 3: ("Introduction to Neural Network Architecture" by Judith E. Dayhoff, translated by Hiroshi Katsurai, published by Morikita Publishing, 1992). It has been known. However, in Document 3, “Kohonen type neural network” is described as “Kohonen feature map”.

The function parameter calculating section 30 calculates the weight vector calculated for each class as a result of the self-organizing learning of each competitive layer unit of the Kohonen-type neural network and the sample pattern, based on the pre-fuzzy rule. This is a part for sequentially calculating the parameters of the membership function for each belonging class.

The function parameter storage section 32 is a memory for storing function parameters calculated for each belonging class for each belonging class as a memory for storing information based on the result of self-organizing learning for each belonging class.

The confidence parameter calculating unit 34 calculates the antecedent membership function calculated for this parameter,
This is a part for calculating a certainty parameter corresponding to a consequent real value of the fuzzy rule from the sample pattern and the class information of the sample pattern.

Also, the belonging pattern set / class information storage unit 36 sets the set of sample patterns belonging to each fuzzy rule calculated by the certainty parameter calculating unit 34 as the belonging pattern of each fuzzy rule, Is stored together with the class information of the sample pattern belonging to.

The recognition fuzzy rule judging / dividing section 38 comprises a judging section 38a and a dividing section 38b.

Then, the judging unit 38a determines the fuzzy rules in which the values of the certainty parameters for two or more specific belonging classes among the fuzzy rules are larger than the reference value. This is a part to be determined as a specific fuzzy rule.

Further, the dividing unit 38b creates a divided fuzzy rule corresponding to each of the specific belonging classes instead of the specific fuzzy rule. Calculate the parameters of the antecedent membership function of the divided fuzzy rule for each class, and calculate each of the divided fuzzy rules from the antecedent membership function calculated for this parameter, the specific sample pattern and its class information. The consequent part is a part for calculating each certainty parameter corresponding to the real numerical value.

The output of the recognition fuzzy rule determining / dividing section 38 is sent to the recognition fuzzy rule storage device 18.

Next, the operation of the fuzzy rule creation device according to the first embodiment, that is, a fuzzy rule creation method will be described.

In the first embodiment, the sample pattern / class information storage device 14 stores n (N is a natural number) n
A case where a dimension (n is a natural number) sample pattern is stored together with class information indicating a class to which the sample pattern belongs is described.

In the first embodiment, the p-th (p is a natural number not less than 1 and not more than N) sample pattern among the N sample patterns is represented by Xp. Since this sample pattern Xp is n-dimensional, Xp =
[X _p1 , x _p2 , x _p3 ,..., X _pn ] (1 ≦ p ≦ N). That is, each sample pattern Xp is
It has n-dimensional components x _p1 , x _p2 , x _p3 ,..., X _pn .

It is assumed that each sample pattern Xp is classified into one of the m classes (m is a natural number) belonging to one of the classes Cp (1 ≦ Cp ≦ m).

The sample pattern and its class information read from the sample pattern / class information storage device 14 into the fuzzy rule creation device 16 are input to the sample pattern division storage unit 26 and the certainty parameter calculation unit 34, respectively.

Next, in the sample pattern division storage section 26, the input sample pattern Xp is divided and stored for each class. Here, the number of sample patterns of each belonging class C (1 ≦ C ≦ m) is represented by Nc. The affiliation Class C = 1, 2, 3, ... the number Nc of sample patterns belonging to m, respectively N _1, N
When _{2, N 3 ... N m,} N = N 1 + N 2 + N 3 + ... +
Expressed as N _m.

The stored sample patterns X
p is sequentially read out to the Kohonen-type neural network learning unit 28 for each belonging class.

In the learning section 28, self-organizing learning of the Kohonen-type neural network based on a certain distance scale D is sequentially performed for each sample class Xp for each sample pattern Xp read for each class. As the distance scale D, for example, a Euclidean distance between vectors may be used.

By the self-organizing learning, the number of input layer units of the Kohonen-type neural network becomes the number (ie, the number of dimensions) n of the components of the sample pattern Xp.

On the other hand, as for the number of competitive layer units of the Kohonen type neural network, an appropriate value rc is set for each belonging class. The appropriate value rc may be, for example, a number that is about ５〜 to の of the number Nc of the sample patterns Xp of the class to which it belongs. However, the arrangement of the competitive layer units in the Kohonen-type neural network may be one-dimensional or two-dimensional.

Then, as a result of the self-organizing learning, a weight vector of each competitive layer unit is calculated for each belonging class. That is, rc (the number of competitive layer units) weight vectors are calculated for each of the m types of belonging classes.

Also, among the weight vectors corresponding to a certain belonging class, the weight vector of the i-th competitive layer unit is represented by Wi. The weight vector Wi has components of the number of input layer units, that is, n components, for one class. Therefore, the weight vector Wi is expressed as Wi =
[W _i1 , w _i2 ,..., W _in ] (1 ≦ i ≦ rc).

The learning unit 28 calculates the rc weight vectors Wi calculated for each belonging class together with the Nc sample patterns to which the belonging class belongs each time the self-organizing learning for the belonging class ends. Are sent to the membership function parameter calculation unit 30 one by one.

In this embodiment, a fuzzy rule for use in the process of identifying an unknown pattern is created using these weight vectors as candidates. This fuzzy rule is generally given by the following equation (9).

R ⁱ : If x ₁ is A _i1 and x ₂ is A _i2 and… and x _n is A _in then y ₁ is b _i1 and y ₂ is b _i2 and… and y _m is b _im 9) Here, R ⁱ (1 ≦ i ≦ r) is the label of the i-th fuzzy rule, and r is the number of rules. The line beginning with “If” is the antecedent part, and the line beginning with “then” is the consequent part. A _ij (1 ≦ j ≦ n) is a fuzzy variable of the antecedent part, and b _ik (1 ≦ k ≦ m) is a real value of the consequent part. The membership function of the fuzzy variable A _ij of the antecedent part (the antecedent part membership function) A _ij (x _j ) is defined by a well-known triangular membership function as shown in the following equation (10). I do.

A _ij (x _j ) = max {1− (| x _j −c _ij | / d _ij , 0)} (10) where c _ij and d _ij are members of the antecedent part, respectively. _These parameters determine the center value and width of the ship function A _ij (x _j ).

Further, the degree of conformity μ _i (X) indicating the degree of coincidence of the unknown pattern X with the antecedent part of the fuzzy rule R ⁱ is calculated for each dimension (1 to n dimensions) as shown in the following equation (11).
Is obtained as the algebraic product of the antecedent membership function in.

[0126]

(Equation 2)

Next, the function parameter calculating section 30 calculates rc
Based on the weight vector Wi and the Nc sample patterns, the parameters c _ij and c _ij that determine the central value of the membership function A _ij (x _j ) of the antecedent part of the fuzzy rule for recognition defined by the above equation (10) the d _ij to define the width,
Calculated sequentially for each class. Hereinafter, the processing procedure in the membership function parameter calculation unit (function parameter calculation unit) 30 will be described with reference to the flowchart in FIG.

(A) First, the function parameter calculator 30
Is a sample pattern X belonging to each weight vector Wi.
A set of p, ie, a belonging pattern set Si is calculated (S1 in FIG. 3). Therefore, under the distance scale D used in the self-organizing learning of the Kohonen-type neural network described above, as shown in the following equation (12),
Each sample pattern Xp is assigned to the closest distance weight vector. Then, a set of sample patterns Xp belonging to each weight vector Wi is obtained as a belonging pattern set Si.

[0129]

(Equation 3)

Here, D (Xp, Wi) is a distance scale D
Pattern Xp and weight vector Wi using
And when the distance scale D is a Euclidean distance, it is given by the following equation (13).

[0131]

(Equation 4)

(B) Next, the function parameter calculator 30
Performs pruning of the weight vector Wi (S2 in FIG. 3). That is, first, the belonging pattern set S
Weight vectors for which i is an empty set, that is, all weight vectors for which no sample pattern belongs to the weight vector Wi are deleted. Then, the function parameter calculation unit 30
Selects the remaining weight vector Wi and renumbers it with its belonging pattern set.

[0133] These weight vector Wi is, as a candidate of the above-mentioned equation (9) the antecedent membership function A _ij of fuzzy rules of (x _j), and their antecedent membership function A _ij (x _j) One to one. Here, the number of weight vectors Wi remaining as a result of pruning is r ′ _C (≦ rc).

(C) Next, the function parameter calculator 30
Performs fuzzification of the weight vector Wi (S in FIG. 3).
3). In other words, the function parameter calculation unit 30 uses the weight vector Wi and the sample pattern Xp included in the weight vector Wi to determine the central value of the antecedent membership function in the r ′ _C recognition fuzzy rules. Calculate c _ij and a parameter d _ij that determines the width.

In this calculation, the weight vector Wi
Each time, all elements x _pj (1 ≦ j ≦ n) of the sample pattern Xp included in the belonging pattern set Si are represented by α∈
Grade (0,1) or higher. That is, all the elements x _pj of the sample pattern Xp are expressed as A with respect to the antecedent membership function A _ij (x _pj ).
_ij (x _pj ) ≧ α∈ (0,1). Here, α is a parameter appropriately determined by a designer, and is preferably set as large as possible.

Specifically, a parameter c for determining the center value
_ij and a parameter d _ij that determines the width are respectively expressed by the following (1) using the component w _ij of the weight vector Wi.
It is given by equations 4) and (15).

[0137]

(Equation 5)

In the above equation (15), when the value of the numerator is 0, the membership function is set to a width of 0.
(Non-fuzzy value = real value). Hereinafter, the above equation (14) and (1)
The operation by the expression 5) is called fuzzification.

Then, the function parameter calculating section 30 sequentially sends the parameter c _ij for determining the center value and the parameter d _ij for determining the width, which are sequentially calculated for each belonging class, to the membership function parameter storing section 32.

Membership function parameter storage unit 32
Sequentially stores a parameter c _ij that determines the central value of the antecedent part membership function and a parameter d _ij that determines the width, which are input for each belonging class. After the parameters c _ij and d _ij for all the belonging classes are stored in the storage unit 32, the numbers i of the parameters c _ij and d _ij for each belonging class are stored in all the belonging classes. A new number is added to the serial number. Here, the maximum value r ₁ of the serial number is the sum of the number of fuzzy rules after pruning in each belonging class, that is, the number r ′ _C of weight vectors Wi remaining as a result of pruning. For example, the number of r of the weight vector of the first belonging class _'1, the number of weight vectors in the second Member classes r'
_2, and so on, if the number of weight vectors belongs class of the m r _'m pieces and the maximum value r ₁ of the serial number,
r ₁ = r ′ ₁ + r ′ ₂ +... + r ′ _m

Then, the storage unit 32 sends the parameters c _ij and d _ij to which the serial numbers have been newly added to the certainty factor calculation unit 34.

The confidence parameter calculating section calculates r ₁ recognition fuzzy rules R ⁱ for which the antecedent membership function A _ij (x _j ) is determined by calculating parameters c _ij and d _ij , respectively. From the sample pattern Xp and its class information Cp, a certainty parameter for each belonging class corresponding to the consequent part real value b _ik shown in the above equation (9) is calculated. Hereinafter, the processing procedure in the certainty parameter calculating unit 34 will be described with reference to the flowchart in FIG.

(A) First, the certainty parameter calculating unit 34
Calculates a set of sample patterns Xp belonging to each recognition fuzzy rule R ⁱ , that is, a belonging pattern set Si (S1 in FIG. 4). For this purpose, the certainty parameter calculation unit 34 calculates each sample pattern Xp as
Each recognizable fuzzy rule R is assigned to a recognition fuzzy rule with the maximum fitness μ _i (Xp).
An attribute pattern set Si of the sample pattern Xp belonging to ⁱ is obtained. The belonging pattern set Si is expressed by the following equation (16) using the degree of conformity μ _i (Xp) of the sample pattern Xp to the fuzzy rule for recognition R ⁱ .

[0144]

(Equation 6)

(B) Next, the certainty parameter calculating unit 34
Is the pruning of the recognition fuzzy rule R ⁱ
(S2 in FIG. 4). For this purpose, the certainty parameter calculation unit 34 first calculates all the recognition fuzzy rules R ⁱ in which the belonging pattern set Si is an empty set, that is, all the fuzzy rules R ⁱ for which no sample patterns belong to the recognition fuzzy rules R ^i. delete. Then, the certainty parameter calculation unit 34 selects the remaining recognition fuzzy rule R ⁱ and renumbers both the number i and the number of the belonging pattern set Si. Here, the maximum value of the new number of the recognition fuzzy rule R ⁱ selected as a result of the pruning is r ₂ (≦ r ₁ : the maximum value of the serial number).

Then, the certainty parameter calculating unit 34 calculates
The selected fuzzy rule for recognition R ⁱ (1 ≦ i ≦ r ₂ )
Pattern set Si and its belonging pattern set Si
And sends it to the belonging pattern set / class information storage unit 36 together with the class information of each sample pattern Xp included in.

(C) Next, the certainty parameter calculating unit 34
Calculates a certainty parameter (S3 in FIG. 4).
For this purpose, the certainty parameter calculation unit 34 determines, based on the belonging pattern Xp of each recognition fuzzy rule R ⁱ and its class information Cp, a confidence corresponding to the consequent part real value in the r ₂ recognition fuzzy rules. The degree parameter _bik is calculated.

More specifically, the certainty parameter b _ik is the sample pattern Xp for the recognition fuzzy rule R ⁱ .
Is given by the following equation (17) or (18) using the fitness μ _i (Xp) of

[0149]

(Equation 7)

As described above, the certainty parameter calculating unit 34
Then, the recognition fuzzy rules R ⁱ are created by calculating the certainty parameters b _ik of the r ₂ recognition fuzzy rules R ⁱ .

Then, the certainty parameter calculating section 34 calculates
Each of the created recognition fuzzy rules R ⁱ is sent to the recognition fuzzy rule determination / division unit 38.

Further, the belonging pattern set Si and its class information stored in the belonging pattern set / class information storage section 36 are also sent to the recognition fuzzy rule determining / dividing section 38.

Recognition fuzzy rule judging / dividing section 38
In the determining unit 38a, among the fuzzy rules, the fuzzy rules in which the values of the certainty parameters for two or more specific belonging classes of the belonging classes each have a value larger than the reference value are identified as the specific fuzzy rules. Is determined.

The recognition fuzzy rule judging / dividing section 38 generates a divided fuzzy rule corresponding to each specific belonging class in the dividing section 38b instead of the specific fuzzy rule. Calculate the parameters of the antecedent membership function of the divided fuzzy rule for each specific affiliation class from the specific sample pattern to which it belongs and its class information, and calculate the antecedent membership function and the specific sample From the pattern and its class information, a certainty parameter corresponding to the consequent real value of each of the divided fuzzy rules is calculated.

Hereinafter, referring to the flowchart of FIG.
The processing procedure in the recognition fuzzy rule determination / division unit 38 will be described.

(A) First, a fuzzy rule determination for recognition
The dividing unit 38 evaluates the certainty parameter in the determining unit 38a and determines the fuzzy rule for recognition (S1 in FIG. 5). For this purpose, the recognition fuzzy rule determination and division unit 38
^For each ⁱ , a set Ti of belonging classes whose confidence parameter b _ik is larger than β times its maximum value is obtained. Here, β is a value that satisfies β∈ [0, 1), and for example, approximately β = 0.5 is appropriate.

For a certain fuzzy rule for recognition R ⁱ , the belonging class set Ti in which the certainty parameter b _ik has a value larger than β times the maximum value is, for example, the following (1)
It is given by equation 9).

[0158]

(Equation 8)

Here, it is assumed that the number of elements of the belonging class Ti for a certain recognition fuzzy rule R ⁱ is r ′ _i . When the number of elements r ′ _i is 2 or more, the determination unit 3
8a determines that the recognition fuzzy rule R ⁱ is a specific fuzzy rule that requires division processing.

(B) Next, determination of a fuzzy rule for recognition
The dividing unit 38 recalculates the parameters of the membership function of the specific fuzzy rule in the dividing unit 38b (S2 in FIG. 5). For this purpose, the recognition fuzzy rule determining / dividing unit 38 performs r ′ _i replacements for the specific fuzzy rule R ⁱ only for the specific fuzzy rule R ^{i in} which the number of elements r ′ _i of the belonging class set Ti is 2 or more. A new fuzzy rule R _l ′ (1 ≦ l ≦ r ′ _i ) is newly created. For this purpose, the recognition fuzzy rule determining / dividing unit 38 calculates a parameter _clj that determines the center value of the antecedent membership function A _lj ( _xj ) and a parameter d _lj that determines the width of the antecedent membership function A _lj (x _j ) of the divided fuzzy rule.

Also, a plurality of divided fuzzy rules R _l ′ created in place of one specific fuzzy rule R ⁱ are as follows:
If division processing is successful, which corresponds to the respective plurality of belonging class 1-one contained in the belonging class set Ti for the particular fuzzy rule R ^i. Therefore, the value of the parameter c _lj and d _lj, based on the sample pattern Xp and the class information Cp belonging to attribution pattern set Si in the original specific fuzzy rules R ^i, each of the following for the corresponding belongs class k ( It is given by equations (20) and (21).

[0162]

(Equation 9)

Α in the above equation (21) is the same value as α in the above equation (15).

(C) Next, determination of a fuzzy rule for recognition
The dividing unit 38 recalculates the certainty parameter in the dividing unit 38b (S3 in FIG. 5). Therefore, recognition fuzzy rule determination-dividing unit 38 based on the sample pattern Xp in attributable pattern set Si in the original specific fuzzy rules R ⁱ and its class information Cp, replaces the original specific fuzzy rules R ⁱ split Fuzzy rule R
The confidence parameter _blk of _l 'is calculated. The processing of this calculation is the same as the processing in the certainty parameter calculating unit 34 described above.

In this way, each of the specific fuzzy rules R ⁱ to be divided is replaced with some divided fuzzy rules R _l '. As a result, the total number of fuzzy rules increases from r ₂ before the division to r ₃
(≧ r ₂ ).

The created recognition fuzzy rule is stored in the recognition fuzzy rule storage device 18 together with the rule number r. In this embodiment, r = r ₃ is stored as the rule number r. Thus, the process of creating the fuzzy rule for use in the process of identifying an unknown pattern in the first embodiment is completed.

As in the first embodiment, if the self-organizing learning of the Kohonen-type neural network and the calculation of the parameters of the membership function of the antecedent part are calculated for each belonging class, the fuzzy rule is created. This can be easily applied when adding an additional sample pattern belonging to a new belonging class. That is, the result of the self-organizing learning for the additional sample pattern and the calculation result of the parameters of the antecedent membership function are
Fuzzy rules can be easily created in addition to the parameters of the antecedent membership function that have already been obtained. Therefore, according to the first embodiment, it is possible to easily adapt to a change in the setting of the sample data such as an additional change of the sample pattern of the new belonging class.

In the first embodiment, the self-organizing learning of the Kohonen-type neural network is
Although the learning was performed sequentially for each belonging class, in the present invention, the learning order of the self-organizing learning does not need to be limited to this. For example, the self-organizing learning for each belonging class is performed in parallel with each class proceeding in parallel. May be. If the self-organizing learning of different belonging classes is performed in parallel, that is, simultaneously, the processing time required for the self-organizing learning can be reduced. As a result, the processing time required for fuzzy rule creation can be reduced.

In the first embodiment, the parameters of the antecedent part membership function are also calculated sequentially for each belonging class. However, in the present invention, the parameters of the antecedent part membership function are calculated. The order does not need to be limited to this. For example, the parameters of the antecedent membership function for each belonging class may be calculated in parallel in each class simultaneously. Then, if the calculation of the parameters of the antecedent membership function of different belonging classes is performed in parallel, that is, simultaneously, the processing time required for calculating the parameters of the antecedent membership function can be reduced. . As a result, the processing time required for fuzzy rule creation can be reduced.

(Second Embodiment)
Next, with reference to FIG. 6, an example of the fuzzy rule creation device and method according to the second embodiment of the present invention will be described as an example of the fuzzy rule creation device 16 of the pattern recognition device 100 described above. FIG. 6 is a functional block diagram for explaining a fuzzy rule creation device according to the second embodiment.

The fuzzy rule creation device 16 of the second embodiment performs self-organizing learning of a Kohonen-type neural network for a plurality of sample patterns,
The fuzzy rule is created using the result of the self-organizing learning. The fuzzy rule creation device 16 includes a sample pattern division storage unit 26, a Kohonen type neural network learning unit (learning unit) 2
8, weight vector storage unit 42, membership function parameter calculation unit (function parameter calculation unit) 30a, certainty parameter calculation unit 34, belonging pattern set / class information storage unit 36, and fuzzy rule determination / division unit 3 for recognition
8. The function parameter calculator 30a and the certainty parameter calculator 34 constitute a parameter calculator 40a that calculates a fuzzy rule parameter using the result of the self-organizing learning.

The fuzzy rule creator 16 of the second embodiment has a membership function calculator 30a instead of the membership function calculator 30 of the first embodiment, As for the fuzzy rule creation device of the first embodiment described above, except that a weight vector storage unit 42 is provided instead of the membership function parameter storage unit 32 in the first embodiment. It has the same configuration. For this reason, in the second embodiment, the same components as those in the first embodiment are denoted by the same reference numerals, and detailed description thereof will be omitted.

The weight vector storage section 42 is a memory for storing the weight vectors Wi obtained for each belonging class by the Kohonen-type neural network learning section 28 as information based on the result of self-organizing learning.

The membership function parameter calculating section 30a calculates the parameters of the membership function of the antecedent part with respect to the weight vectors of all the belonging classes and the sample patterns of all the belonging classes.

The sample pattern Xp (1 ≦ p ≦ N) read from the sample pattern / class information storage device 14 into the recognition fuzzy rule creating device 16 and its class information Cp (1 ≦ Cp ≦ m) ) Indicates the sample pattern
The membership function parameter calculation unit 3 as well as the membership division storage unit 26 and the certainty parameter calculation unit 34
0a is also input.

Next, the operation of the fuzzy rule creation device according to the second embodiment, that is, a fuzzy rule creation method will be described.

In the second embodiment, the Kohonen-type neural network learning unit 28 performs self-organizing learning for each belonging class, and as a result, calculates the weight vector for each belonging class. This is the same process as in the embodiment. Therefore, in the second embodiment, a detailed description of the process of calculating the weight vector for each belonging class is omitted.

In the second embodiment, the learning unit 28 sequentially calculates the weight vectors Wi (1
≤ i ≤ rc) is sequentially stored in the weight vector storage unit 42 for each belonging class.

After the weight vectors Wi for all the belonging classes are stored in the weight vector storage unit 42, the weight vectors W for each of the belonging classes are stored.
The number i of i is newly added to the serial number for all the belonging classes. Here, the maximum value r ₀ of the serial number is the sum of the numbers r ′ _C of the weight vectors Wi in each belonging class. For example, the number of r of the weight vector of the first belonging class _'1, the number of weight vectors in the second Member classes r'
_{In the same} way, if the number of weight vectors of the m-th belonging class is r ′ _m , the maximum value r _{0 of the} serial number becomes
r ₀ = r ′ ₁ + r ′ ₂ +... + r ′ _m

The weight vector storage unit 42 stores the weight vector Wi with the newly assigned serial number in the membership function parameter calculation unit (function parameter calculation unit).
Send to 30a.

Next, the function parameter calculating section 30a calculates the N sample patterns Xp and r of all the belonging classes.
_{Based on the zero} weight vectors Wi, a parameter c _ij that determines the center value of the antecedent membership function A _ij (x _j ) of the antecedent fuzzy rule defined by the above equation (10) and a parameter d that determines the width _ij is calculated. The first mentioned above
In the embodiment of the present invention, these parameters c _ij and d _ij are calculated for each belonging class.
In this embodiment, these parameters c _ij and d _ij are calculated for all sample patterns. That is, in the first embodiment, the belonging pattern set Si is determined as shown in the above equation (12), but in the second embodiment, processing is performed for all belonging classes. Therefore, the belonging pattern set Si is given by the following equation (22).

[0182]

(Equation 10)

Next, in the function parameter calculation section 30a, pruning of the weight vector and fuzzification of the weight vector are performed in the same manner as the function parameter calculation processing in the first embodiment described above. Also,
As a result of the pruning of the weight vector, in the second embodiment, the parameters c _ij and d _{ij of the} membership function of the antecedent part in r ₁ (≦ r ₀ ) fuzzy rules
Are calculated respectively.

Next, the certainty parameter calculating section 34 calculates the certainty parameter by performing the same processing as in the first embodiment described above.

Next, in the recognition fuzzy rule judging / dividing section 38, similar to the above-described first embodiment, division processing is performed on a specific fuzzy rule, and the creation of the fuzzy rule is completed. I do.

As described above, in the second embodiment, the function parameters are calculated for all the belonging classes based on the result of the self-organizing learning of the Kohonen-type neural network for each belonging class. For this reason, in the second embodiment, as compared with the case where the function parameters are calculated for each belonging class as in the first embodiment, the antecedent that reflects the tendency of the distribution of the entire sample pattern more. The parameters of the department membership function can be calculated. As a result, a more appropriate fuzzy rule can be created.

Also, in the second embodiment, as in the case of the above-described first embodiment, it is easy to change the setting of the sample data such as adding or changing the sample pattern of a new belonging class. Can be adapted.

In the second embodiment, the self-organizing learning of the Kohonen-type neural network is
Although the learning was performed sequentially for each belonging class, in the present invention, the learning order of the self-organizing learning does not need to be limited to this. For example, the self-organizing learning for each belonging class is performed in parallel with each class simultaneously. May be. If the self-organizing learning of different belonging classes is performed in parallel, that is, simultaneously, the processing time required for the self-organizing learning can be reduced. As a result, the processing time required for fuzzy rule creation can be reduced.

(Third Embodiment) Next, referring to FIG. 7, a fuzzy rule according to a third embodiment of the present invention will be described as an example of the fuzzy rule creating device 16 of the pattern recognition device 100 described above. An example of a rule creation device and a rule creation method will be described. FIG. 7 is a functional block diagram for explaining a fuzzy rule creation device according to the third embodiment.

The fuzzy rule creation device 16 of the third embodiment performs self-organizing learning of a Kohonen-type neural network for a plurality of sample patterns,
The fuzzy rule is created using the result of the self-organizing learning. The fuzzy rule creation device 16 includes a sample pattern division storage unit 26, a Kohonen type neural network learning unit (learning unit) 2
8, weight vector storage unit 42, membership function parameter calculation unit (function parameter calculation unit) 30a, certainty parameter calculation unit 34, belonging pattern set / class information storage unit 36, and fuzzy rule determination / division / reproduction for recognition It is configured with a creating unit 44. The function parameter calculator 30a and the certainty parameter calculator 34 constitute a parameter calculator 40a that calculates a fuzzy rule parameter using the result of the self-organizing learning.

The fuzzy rule creator 16 of the third embodiment is different from the fuzzy rule deciding / dividing unit 38 of FIG. The configuration is the same as that of the fuzzy rule creation device 16 according to the above-described second embodiment. For this reason, in the third embodiment, the same components as those in the second embodiment are denoted by the same reference numerals, and detailed description thereof will be omitted.

Next, the recognition fuzzy rule determining / dividing / reproducing unit 44 will be described with reference to FIG. FIG.
FIG. 4 is a functional block diagram for explaining a fuzzy rule determination / division / re-creation unit 44 for recognition.

The recognition fuzzy rule determining / dividing / recreating unit 44 includes a determining unit 38a, a dividing unit 38b, a divided fuzzy rule determining unit 46, a new fuzzy rule creating unit 48,
A new fuzzy rule determining unit 50 and a modified fuzzy rule creating unit 52 are provided. The determination unit 38a and the division unit 38b are the same as those described in the second embodiment.

Also, the divided fuzzy rule determination unit 46
The division unit 38b determines whether the number of divided fuzzy rules created in place of the specific fuzzy rule is smaller than the number of specific belonging classes to which the specific sample pattern belonging to the specific fuzzy rule belongs.
If the number of divided fuzzy rules is smaller than the number of classes belonging to the original specific fuzzy rule, the range of the divided fuzzy rule extends to two or more belonging classes.

Further, the new fuzzy rule creating section 48 obtains the specific fuzzy rule from the specific sample pattern belonging to the specific fuzzy rule determined by the divided fuzzy rule determining section 46 and the class information indicating the class to which the sample pattern belongs. This is the part that creates a new fuzzy rule instead.

Further, the new fuzzy rule determining unit 50 determines whether the number of new fuzzy rules created in place of the specific fuzzy rule in the new fuzzy rule creating unit 48 belongs to the specific sample pattern belonging to the specific fuzzy rule. This is a part for determining whether or not the number is smaller than the number of belonging classes. If the number of the new fuzzy rules is smaller than the number of the belonging classes included in the original specific fuzzy rule, the range of the new fuzzy rule extends to two or more belonging classes. Therefore, the processing in the new fuzzy rule creating unit 48 is insufficient.

Further, the modified fuzzy rule creation unit 52
The number of competitors in the Kohonen neural network is
The self-organizing learning of the Kohonen-type neural network is performed again for a specific sample pattern belonging to the specific fuzzy rule by increasing the number of times when a new fuzzy rule is created, and using the result of this self-organizing learning, This is where the modified fuzzy rule is created.

Next, the fuzzy rule creation method of the third embodiment, in particular, the operation of the recognition fuzzy rule determination / division / re-creation unit 44 will be described with reference to FIG.
FIG. 9 shows a recognition fuzzy rule determining / dividing / recreating unit 4
14 is a flowchart provided to explain a processing procedure of No. 4;

In the third embodiment, the description up to the point where the dividing fuzzy rule is created in the dividing unit 38b of the recognition fuzzy rule determining / dividing / recreating unit 44 has been described in the first embodiment. Since the processing is the same as the processing of the recognition fuzzy rule determination / division unit 38, a detailed description thereof will be omitted. That is, (a) determination of the certainty parameter, which is the step of S1, and (b) recalculation of the membership function parameter, which is the step of S2, and (c) recalculation of the certainty parameter, which is the step of S3 in FIG. Is the same as the processing described in the first embodiment.

(D) After the division fuzzy rule is created by performing the process of step S3, in the third embodiment, the division fuzzy rule determination unit 46 subsequently proceeds to the specific fuzzy rule in the division unit 38b. Then, it is determined whether or not the number of divided fuzzy rules created is smaller than the number of specific belonging classes to which the specific sample pattern belonging to the specific fuzzy rule belongs (S4 in FIG. 9).

For this purpose, the divided fuzzy rule determining section 4
6, for each specific fuzzy rules R ^i, which is the subject of fragmentation, it is checked whether division fuzzy rules R _'l was actually created what one to replace it. For example, a divided fuzzy rule R ′ created by subdividing a specific fuzzy rule R ⁱ
the number of _l and r _i ", on the other hand, the original specific fuzzy rules R
^When the number of specific belonging classes of _i is r _i ′, r _i ′
If> r _i ”, it is determined that the division process of the specific fuzzy rule R ⁱ is insufficient.

(E) Next, the new fuzzy rule creating section 48
In particular sample pattern belongs to a particular fuzzy rules determined in the divided fuzzy rule determination unit 46 and the class information indicating the belonging to the class of the sample pattern, specific fuzzy rules R ⁱ r _i a new fuzzy rule instead of "Or more are created (S5 in FIG. 9).

In creating a new fuzzy rule, any suitable method may be used. For example, the fuzzy rule creating method described in the first and second embodiments may be used.

(F) Next, the new fuzzy rule judging section 50
It is determined whether or not the number of new fuzzy rules created in place of the specific fuzzy rule in the new fuzzy rule creating unit 48 is smaller than the number of specific belonging classes to which the specific sample pattern belonging to the specific fuzzy rule belongs. The determination is made (S6 in FIG. 9).

For this purpose, the new fuzzy rule judging section 50
, For each specific fuzzy rules R ^i, which is the subject of fragmentation, it is checked whether a new fuzzy rules R _'N was actually created what one to replace it. For example, the specific fuzzy rules R ⁱ new fuzzy rules created by subdivision of R 'the number of _N and r _i ^*, whereas, the number of specific belonging class of the original specific fuzzy rules R ⁱ r _i' was Then r
_i '> r _i ^* if, it is determined that the division processing of the specific fuzzy rules R ⁱ was insufficient. In that case, the range of the new fuzzy rule R _'N will be spans more than one belonging class. Therefore, the processing in the new fuzzy rule creating unit 48 is insufficient.

(G) Next, the modified fuzzy rule creating section 5
In 2, the number of competitors in the Kohonen-type neural network is increased from the number when a new fuzzy rule is created, and the self-organizing learning of the Kohonen-type neural network is performed for a specific sample pattern belonging to a specific fuzzy rule. The process is performed again, and a modified fuzzy rule is created using the result of the self-organizing learning (S7 in FIG. 9).

Here, when creating the modified fuzzy rules by increasing the number of competing layers, the specific fuzzy rules R ⁱ , until the number r _i ^** of the modified fuzzy rules becomes equal to or greater than the number r _i ′ of the specific belonging classes. A modified fuzzy rule is created each time the number of competing layers is increased by one with the number of sample patterns belonging to. For example, a modified fuzzy rule is created instead of a specific fuzzy rule by setting m + 1 competition layers to m + 1. If the specific fuzzy rule subdivision is insufficient, the number of competitive layers is increased to m + 2, and a modified fuzzy rule is created again. The process until r _i ^** ≧ r _i ', creating increased by modified fuzzy rule by the number the range of sample patterns that belong to the number of competitive layer to a particular fuzzy rule R ^i.

When the specific fuzzy rules are sufficiently subdivided, the number r ₃ of all fuzzy rules including the modified fuzzy rules is equal to the number r fuzzy rules created up to the step S3. ₂ or more.

As described above, if a specific fuzzy rule extending over two or more affiliation classes is subdivided and a new fuzzy rule and, if necessary, a modified fuzzy rule are created, division of the specific fuzzy rule into affiliation classes can be performed. Therefore, a more appropriate fuzzy rule can be obtained.

(Pattern Recognition Apparatus) Next, the configuration and operation of the pattern recognition apparatus 20 of the pattern recognition apparatus 100 shown in FIG. 2 will be briefly described with reference to FIG.
FIG. 10 is a functional block diagram for explaining the pattern identification device.

The pattern discriminating apparatus 20 includes a conformity calculating section 54, a certainty parameter calculating section (certainty calculating section) 56.
And a belonging class determination unit 58.

In the matching degree calculating section 54, the matching degree μ _i (X) of each fuzzy rule R ⁱ stored in the recognition fuzzy rule storage device 18 is compared with the unknown pattern X as the pattern to be identified. ) Is calculated using the above equation (11).

Next, the certainty parameter calculating unit 56 calculates the algebraic product of the fitness μ _i (X) of each fuzzy rule R ⁱ and its certainty parameter b _ik ,
The degree of certainty y ^* _k for each class belonging to the unknown pattern X is calculated. Specifically, the confidence y ^* _k is expressed by the following (2)
It is given by equation 3) or equation (24).

[0214]

[Equation 11]

Next, the belonging class determination unit 58 searches for a class C in which the certainty factor y ^* _k for each belonging class of the unknown pattern X is maximized. That is, the following (25)
Search for a certainty factor y ^* _k that satisfies the equation.

[0216]

(Equation 12)

However, in the above equation (25), y ^* _C =
When the value becomes 0, the unknown pattern X cannot be identified. The determination result of the belonging class determination unit 58 is transmitted to the identification result output device 22 and output.

In this way, for an unknown pattern such as a character image read by an optical scanner or a human voice pattern recorded by a microphone, the class to which the unknown pattern belongs is determined. , It is possible to recognize the contents of the characters and voices.

In each of the embodiments described above, only examples in which these inventions are configured under specific conditions have been described. However, these inventions can be subjected to many changes and modifications. For example, in each of the above-described embodiments, in creating the fuzzy rule, after performing self-organizing learning of Kohonen-type neural network for each class, it corresponds to the parameter of the membership function of the antecedent part and the real value of the consequent part. Although the certainty parameters are calculated respectively, in the present invention, the fuzzy rule may be created in any manner after the self-organizing learning of the Kohonen-type neural network is performed for each belonging class. For example, after the self-organizing learning, a fuzzy rule may be created by performing learning with forgetting of a fuzzy neural network as in the technique disclosed in the above-mentioned document 1.

For example, in the above-described embodiment, the self-organizing learning of the Kohonen-type neural network is sequentially performed for each belonging class. However, in the present invention, the learning order of the self-organizing learning is limited to this. For example, self-organization learning for each class may be performed in parallel at the same time for each class. If the self-organizing learning of different belonging classes is performed in parallel, that is, simultaneously, the processing time required for the self-organizing learning can be reduced. As a result, the processing time required for fuzzy rule creation can be reduced.

In the above-described embodiment, the antecedent part membership function is defined as a triangular shape by the above equation (10). However, in the present invention, the type of the antecedent part membership function is There is no need to limit. For example, a bell-shaped or multidimensional membership function defined by the above equation (2) may be used as the antecedent part membership function. As the multidimensional membership function, for example, the one described in the above-mentioned document 2 can be used. In the case of a multidimensional membership function, its value matches the fitness.

Further, in the above-described embodiment, the fitness in each fuzzy rule is calculated as an algebraic product by the above equation (11). However, in the present invention, the method of calculating the fitness is not limited to this. Absent. For example, an arbitrary calculation method described in the above-mentioned reference 3 can be used.

[0223]

According to the fuzzy rule creating method of the first aspect and the fuzzy rule creating apparatus of the second aspect of the present invention, self-organizing learning is performed for each class to which a sample pattern belongs. Therefore, the present invention can be easily applied to a case where an additional sample pattern belonging to a new belonging class is added after the fuzzy rule is created. That is, a fuzzy rule can be created by adding the result of the self-organizing learning for the additional sample pattern belonging to the new belonging class to the already obtained result of the self-organizing learning. Therefore, according to the fuzzy rule creation method of the present invention, it is possible to easily adapt to the setting change of the sample data such as the addition and change of the sample pattern of the new belonging class.

Furthermore, according to the fuzzy rule creation method of the present invention, since the self-organizing learning of the Kohonen-type neural network is performed for each belonging class, the self-organizing learning of different belonging classes is performed in parallel, that is, simultaneously. By proceeding, the processing time required for self-organizing learning can be reduced. As a result, the processing time required for fuzzy rule creation can be reduced.

If the parameters of the membership function of the antecedent part are calculated for each belonging class after the self-organizing learning is performed for each belonging class, if the fuzzy rule is created, the member belongs to a new belonging class. It can be more easily adapted when adding an additional sample pattern. That is, for the additional sample pattern, not only the self-organizing learning but also the calculation results of the parameters of the antecedent membership function can be used to create a fuzzy rule in addition to the parameters of the antecedent membership function already obtained. . Therefore, according to the fuzzy rule creation method of the present invention, it is possible to more easily adapt to a change in setting of sample data such as an addition or change of a sample pattern of a new belonging class.

Further, when not only the self-organizing learning of the Kohonen-type neural network but also the calculation of the parameters (function parameters) of the membership function of the antecedent part are performed for each belonging class, the self-organizing learning of the belonging classes different from each other is performed. If the calculation of the function parameters is performed in parallel, that is, simultaneously, the processing time required for the self-organizing learning and the calculation of the function parameters can be reduced. As a result, it is possible to further reduce the processing time required for fuzzy rule creation.

Further, if the self-organizing learning is performed for each belonging class, and the function parameters are calculated for all the belonging classes, the function parameters of the entire sample pattern can be compared with the case where the function parameters are calculated for each belonging class. It is possible to calculate the parameters of the antecedent membership function that better reflect the tendency of the distribution. As a result, a more appropriate fuzzy rule can be created.

Also, a specific fuzzy rule extending over two or more belonging classes is determined, and a function parameter is calculated for each class using a specific sample pattern belonging to the determined specific fuzzy rule. If a division fuzzy rule is created instead, it is possible to divide a specific fuzzy rule by class to which it belongs. If the specific fuzzy rules can be divided for each class, more appropriate fuzzy rules can be obtained.

When the divided fuzzy rules are not sufficiently subdivided, a new fuzzy rule is used instead of the specific fuzzy rule by using a specific sample pattern belonging to a specific fuzzy rule extending over two or more belonging classes. If a rule is created, it is possible to divide a specific fuzzy rule by class to which it belongs. If the specific fuzzy rules can be divided for each class, more appropriate fuzzy rules can be obtained.

When the new fuzzy rule is not sufficiently subdivided, the number of competitive layers of the Kohonen-type neural network is increased, and then a specific sample belonging to a specific fuzzy rule extending to two or more belonging classes is obtained. If a modified fuzzy rule is created in place of the specific fuzzy rule by performing self-organizing learning on the pattern, it is possible to divide the specific fuzzy rule for each belonging class. If the specific fuzzy rules can be divided for each class, more appropriate fuzzy rules can be obtained.

[Brief description of the drawings]

FIG. 1 is a functional block diagram for describing a fuzzy rule creation device according to a first embodiment;

FIG. 2 is a functional block diagram for explaining a pattern recognition device.

FIG. 3 is a flowchart for explaining a processing procedure of a membership function parameter calculation unit;

FIG. 4 is a flowchart for explaining a processing procedure of a certainty parameter calculating unit;

FIG. 5 is a flowchart for explaining a processing procedure of a fuzzy rule determination / division unit for recognition.

FIG. 6 is a functional block diagram for explaining a fuzzy rule creation device according to a second embodiment;

FIG. 7 is a functional block diagram for explaining a fuzzy rule creation device according to a third embodiment;

FIG. 8 is a functional block diagram for explaining a recognition fuzzy rule determining / dividing / recreating unit;

FIG. 9 is a flowchart for explaining a processing procedure of a recognition fuzzy rule determining / dividing / recreating unit;

FIG. 10 is a functional block diagram for explaining a pattern identification device.

[Explanation of symbols]

100: Pattern recognition device 10: Pattern input device 12: Switch 14: Sample pattern / class information storage device 16: Recognition fuzzy rule creation device 18: Recognition fuzzy rule storage device 20: Pattern identification Apparatus 22: Identification result output apparatus 26: Sample pattern division storage section 28: Kohonen-type neural network learning section (learning section) 30, 30a: Membership function parameter calculation section (function parameter calculation section) 32: Membership function parameter Storage unit (function parameter storage unit) 34: certainty parameter calculation unit 36: membership pattern set / class information storage unit 38: recognition fuzzy rule determination / division unit 38a: determination unit 38b: division unit 40, 40a: parameter Calculator 42: Weight vector storage 44: Fuzzy rule for recognition Fixed / divided / recreated part 46: divided fuzzy rule determination part 48: new fuzzy rule creation part 50: new fuzzy rule determination part 52: modified fuzzy rule creation part 54: conformity calculation part 56: certainty parameter calculation part 58: Affiliation class judgment unit

Claims (16)

[Claims] 1. A method for performing self-organizing learning of a Kohonen-type neural network on a plurality of sample patterns and using the result of the self-organizing learning to create the fuzzy rule, the self-organizing learning is performed using the sample pattern. A method for creating a fuzzy rule, wherein information based on a result of the self-organizing learning is stored in a memory for each belonging class. 2. The fuzzy rule creating method according to claim 1, wherein as a result of the self-organizing learning, a weight vector of each competitive layer unit of the Kohonen-type neural network is calculated for each of the belonging classes. And calculating the parameters of the antecedent membership function of the fuzzy rule from the sample pattern and the fuzzy rule, and calculating the antecedent membership function calculated for the parameter, the sample pattern, and a class representing the class to which the sample pattern belongs. A fuzzy rule creation method, wherein a confidence parameter corresponding to a consequent real value of the fuzzy rule is calculated from the information. 3. The fuzzy rule creating method according to claim 2, wherein a parameter of the membership function of the antecedent part is calculated for each of the belonging classes, and the calculated parameter is based on a result of the self-organizing learning. A fuzzy rule creation method, wherein information is stored in a memory for each of the belonging classes. 4. The fuzzy rule creation method according to claim 2, wherein the weight vector calculated for each belonging class is stored in a memory as information based on a result of the self-organizing learning, and A fuzzy rule creation method, wherein a parameter of a membership function is calculated with respect to the weight vectors of all belonging classes and sample patterns of all the belonging classes. 5. The fuzzy rule creation method according to claim 2, wherein the values of the certainty factor parameters for two or more specific belonging classes of the belonging classes have values larger than a reference value. The fuzzy rule is determined as a specific fuzzy rule.In order to create divided fuzzy rules corresponding to each of the specific belonging classes instead of the specific fuzzy rule, a specific sample pattern belonging to the specific fuzzy rule and its class information are used. The parameters of the antecedent membership function of the divided fuzzy rule for each of the specific affiliation classes are calculated, and from the calculated antecedent membership function, the specific sample pattern and its class information, The consequent part of each of the divided fuzzy rules corresponds to the real value. A fuzzy rule creation method, wherein each confidence factor parameter is calculated. 6. The fuzzy rule creation method according to claim 5, wherein the number of the divided fuzzy rules created in place of the specific fuzzy rule belongs to the specific sample pattern belonging to the specific fuzzy rule. When the number is smaller than the number, a new fuzzy rule is created instead of the specific fuzzy rule from the specific sample pattern belonging to the specific fuzzy rule and the class information indicating the class to which the sample pattern belongs. How to create fuzzy rules. 7. The fuzzy rule creating method according to claim 6, wherein the number of the new fuzzy rules created in place of the specific fuzzy rule is such that a specific sample pattern belonging to the specific fuzzy rule belongs to. If the number of classes is smaller than the number of specific classes, the number of competitors in the Kohonen-type neural network
The self-organizing learning of the Kohonen-type neural network is performed on the specific sample pattern belonging to the specific fuzzy rule by increasing the number when the new fuzzy rule is created, and using the result of the self-organizing learning. A modified fuzzy rule. 8. The fuzzy rule creating method according to claim 7, wherein the number of sample patterns belonging to the specific fuzzy rule is set as an upper limit until the number of the modified fuzzy rules is equal to or greater than the number of the specific belonging class. The number of the competitive layers is 1
A fuzzy rule creation method, wherein a modified fuzzy rule is created each time it is increased. 9. A fuzzy rule creating apparatus for performing self-organizing learning of a Kohonen-type neural network for a plurality of sample patterns and creating the fuzzy rule using a result of the self-organizing learning. A Kohonen-type neural network learning unit that performs the following for each belonging class of the sample pattern, a memory that stores the result of the self-organizing learning for each belonging class, and a fuzzy rule using information based on the result of the self-organizing learning. A fuzzy rule creation device, comprising: a parameter calculation unit for calculating a parameter. 10. The fuzzy rule creation device according to claim 9, wherein the parameter calculation unit includes a membership function parameter calculation unit and a certainty factor parameter calculation unit, and the membership function parameter The calculating unit has, from each of the competitive layer units of the Kohonen-type neural network, the weight vector calculated for each belonging class as a result of the self-organizing learning, and the sample pattern, the antecedent member of the fuzzy rule. Calculating a parameter of the ship function, the confidence parameter calculation unit calculates a consequent part of the fuzzy rule from the membership function of the antecedent part, the sample pattern and the class information of the sample pattern for which the parameter is calculated. Calculates certainty parameters equivalent to numerical values Fuzzy rule creation device. 11. The fuzzy rule creation device according to claim 9, wherein the weight vector calculated for each of the belonging classes is:
A membership function parameter calculator for calculating the parameters of the antecedent membership function for each of the affiliation classes; and the parameter calculated for each of the affiliation classes as information based on the result of the self-organizing learning, for the affiliation class. A fuzzy rule creation device characterized by comprising a memory for separately storing. 12. The fuzzy rule creating device according to claim 9, wherein the weight vector calculated for each belonging class is stored as information based on a result of the self-organizing learning, and the antecedent member A fuzzy rule creation device, comprising: a membership function parameter calculation unit that calculates parameters of a ship function with respect to the weight vectors of all the belonging classes and sample patterns of all the belonging classes. . 13. The fuzzy rule creating device according to claim 9, wherein, among the fuzzy rules, two of the belonging classes are included.
A fuzzy rule determining unit for determining, as a specific fuzzy rule, a fuzzy rule in which the value of the certainty factor parameter for one or more specific belonging classes has a value larger than a reference value, and specifying the specific fuzzy rule instead of the specific fuzzy rule To create a divided fuzzy rule corresponding to each of the belonging classes, respectively, from the specific sample pattern belonging to the specific fuzzy rule and its class information, the parameters of the antecedent part membership function of the divided fuzzy rule for each specific belonging class Is calculated, and from the membership function of the antecedent part calculated for the parameter, the specific sample pattern and its class information, a confidence parameter corresponding to the real value of the consequent part of each of the divided fuzzy rules is calculated. Fuzzy rule division unit to calculate each A fuzzy rule creation device characterized by the following. 14. The fuzzy rule creation device according to claim 13, wherein the number of the divided fuzzy rules created in place of the specific fuzzy rule belongs to the specific belonging pattern to which a specific sample pattern belonging to the specific fuzzy rule belongs. A divided fuzzy rule determining unit for determining whether or not the number is smaller than the number of the specific fuzzy rules; and a specific sample pattern belonging to the specific fuzzy rule and class information indicating a class to which the sample pattern belongs. A new fuzzy rule creation unit for creating a simple fuzzy rule. 15. The fuzzy rule creating device according to claim 14, wherein the number of the new fuzzy rules created in place of the specific fuzzy rule is such that a specific sample pattern belonging to the specific fuzzy rule belongs to. A new fuzzy rule determining unit that determines whether the number is smaller than the number of specific belonging classes, and the number of competitive layers of the Kohonen-type neural network are:
The self-organizing learning of the Kohonen-type neural network is performed on the specific sample pattern belonging to the specific fuzzy rule by increasing the number when the new fuzzy rule is created, and using the result of the self-organizing learning. A modified fuzzy rule creating unit for creating a modified fuzzy rule. 16. The fuzzy rule creating device according to claim 15, wherein the number of sample patterns belonging to the specific fuzzy rule is set as an upper limit until the number of the modified fuzzy rules is equal to or more than the number of the specific belonging class. The number of the competitive layers is 1
A fuzzy rule creation device comprising a modified fuzzy rule creation unit for creating a modified fuzzy rule each time the number is increased.