JP2006079326A - Classification support map preparation method, program for executing the method, and classification support map preparation device - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a classification support map preparation method capable of preparing a classification support map useful for the classification of unknown data for correctly classifying unknown data regardless of the learning sequence of learning data. <P>SOLUTION: This classification support map preparation method as a self-organization map preparation method for preparing a self-organization map 3 for supporting the classification of unknown data constituted of vector data comprises a step for preparing an non-learning map 4 including a plurality of non-learning cells 4a which are respectively provided with non-learning vector data, a step for preparing a pseudo map 5 including a plurality of pseudo cells 5a having learning integral values corresponding to the plurality of non-learning cells 4a, a step for making the pseudo map 5 learn the learning data constituted of the vector data and a step for reflecting the learning integral values owned by the pseudo cells 5a of the pseudo map 5 which has learned on the vector data owned by the non-learning cells 4a of the non-learning map 4. <P>COPYRIGHT: (C)2006,JPO&NCIPI

Description

  The present invention relates to a classification support map creation method and a classification support map creation apparatus, and more particularly to a classification support map creation method for creating a classification support map that supports classification of unknown data composed of vector data, and a method for executing the same. The present invention relates to a program and a classification support map creation device.

  Conventionally, a classification support map creation method for creating a classification support map that supports classification of unknown data composed of vector data is known. As the above classification support map, for example, T.M. A self-organizing map (SOM: Self-Organizing Map) proposed by Mr. Kohonen is known. As a method for creating this self-organizing map (SOM), various creation methods have been conventionally proposed (for example, see Patent Document 1).

  In the method of creating a self-organizing map (SOM) proposed in Patent Document 1, a self-organizing map is obtained by causing each cell of an unlearned map including a plurality of cells having unlearned vector data to learn learning data. Have created. When learning data is learned in each cell of the unlearned map, a single vector data having the smallest inter-vector distance (Euclidean distance) from the vector data constituting the learned data from the unlearned map is used. In addition to selecting a cell, learning data is learned for the selected cell and cells in the vicinity thereof.

JP 2001-229362 A

  However, in the method of creating a self-organizing map proposed in Patent Document 1, each time learning data is learned, the vector data of the selected cell and cells in its vicinity are changed, so that the learning data is learned. When sequentially selecting power cells, there is an inconvenience that the selected cells may differ depending on the learning status of the previous learning data to the unlearned map. For this reason, there exists a problem that the self-organization map obtained may differ with the order which learns learning data. Therefore, in the method for creating a self-organizing map of Patent Document 1, a classification support map useful for classification of unknown data that can correctly classify unknown data may not be created depending on the order in which learning data is learned. There is a problem.

  The present invention has been made to solve the above-described problems, and one object of the present invention is to detect unknown data that can correctly classify unknown data regardless of the order of learning data. To provide a classification support map creation method capable of creating a classification support map useful for classification.

Means for Solving the Problems and Effects of the Invention

  In order to achieve the above object, a classification support map creation method according to a first aspect of the present invention is a classification support map creation method for creating a classification support map that supports classification of unknown data composed of vector data. A step of preparing an unlearned map including a plurality of unlearned cells each having unlearned vector data, and a pseudo map corresponding to the plurality of unlearned cells and including a plurality of pseudo cells having predetermined vector data A step of learning the learning data composed of vector data in a pseudo map, and a step of reflecting the vector data held by the pseudo cell of the learned pseudo map in an unlearned cell corresponding to the pseudo cell of the unlearned map And. Note that “unlearned” in the present invention is not only when learning data is not learned at all, but when there is a plurality of learning data, one or more learning data is learned. This includes cases where learning has not been completed.

  In the classification support map creating method according to the first aspect, as described above, learning data consisting of vector data in a pseudo map corresponding to a plurality of unlearned cells and including a plurality of pseudo cells having predetermined vector data. When learning a plurality of pieces of learning data sequentially by reflecting the vector data held by the pseudo cells of the learned pseudo map to the unlearned cells corresponding to the pseudo cells of the unlearned map, While the learning data is learned by the pseudo map, the vector data of each unlearned cell of the unlearned map is not changed. As a result, if a cell in which learning data is to be learned in the unlearned map is selected and then learning data is learned in a cell in the pseudo map corresponding to the cell in the selected unlearned map, Since the vector data of the cells in the unlearned map does not change when selecting the cells in which the learning data is to be learned, there is no inconvenience that the cells selected in the unlearned map differ depending on the order in which the learning data is learned. Thereby, also when learning several learning data, it can suppress that the self-organization map obtained differs by the order which learns learning data. Therefore, a classification support map useful for classification of unknown data that can correctly classify unknown data can be created regardless of the order of learning data.

  In the classification support map creation method according to the first aspect, preferably, the method further includes a step of selecting a first block including a collection of unlearned cells from the unlearned map, and the step of causing the pseudo map to learn the learning data. Learning the learning data in a pseudo cell included in the second block of the pseudo map corresponding to the first block of the unlearned map. With this configuration, the learning data can be learned in a plurality of pseudo cells included in the second block corresponding to the learning data. Therefore, for a single cell corresponding to the learning data and a cell in the vicinity region thereof, Unlike the case of learning the learning data, there is no need to set a neighborhood function for setting the range of the neighborhood region for learning the learning data. Thereby, even a person who is not skilled in setting the neighborhood function can easily learn the learning data, so that the classification support map can be easily created.

  In this case, preferably, the step of selecting the first block from the unlearned map includes a plurality of unlearned items included in the block for each of a plurality of blocks made up of a collection of unlearned cells extracted from the unlearned map. A step of calculating block data as vector data of a block based on unlearned vector data of a cell, and comparing learning data with block data of a plurality of blocks, corresponding to learning data from a plurality of blocks Selecting a first block having block data. If comprised in this way, the 1st block corresponding to learning data can be easily selected from an unlearned map.

  In the configuration including the step of causing the pseudo cell included in the second block to learn the learning data, preferably, the step of causing the pseudo cell included in the second block to learn the learning data includes all the pseudo cells included in the second block. A step of adding vector data based on the learning data to the vector data of the cell. If comprised in this way, learning data can be easily learned by all the pseudo cells contained in a 2nd block.

In this case, preferably, the step of adding the vector data based on the training data, the vector data included in the pseudo cells included in the second block and w _a, the learning data is x, the pseudo cells included in the second block And adding the vector data based on the learning data to the vector data of the pseudo cell included in the second block, based on the formula represented by w _a + x / c, where c is the number of. If comprised in this way, the vector data based on learning data can be easily added to the vector data which all the pseudo cells contained in a 2nd block have, respectively.

  In the configuration including the step of causing the pseudo cell included in the second block to learn the learning data, preferably, the step of causing the pseudo cell included in the second block to learn the learning data includes all the pseudo cells included in the second block. Adding vector data based on the difference between the learning data and the unlearned vector data of the unlearned cell of the first block to the vector data of the cell. If comprised in this way, learning data can be easily learned by the pseudo cell contained in a 2nd block.

In this case, preferably, the step of adding the vector data based on the difference between the learning data and the unlearned vector data included in the unlearned cell of the first block includes the vector data included in the pseudo cell included in the second block. When w _a is set, t is a constant, learning data is x, unlearned vector data of an unlearned cell in the first block is w _b, and the number of pseudo cells included in the second block is c , W _a + t · (x−w _b ) / c Based on the expression represented by the pseudo cell included in the second block, the learning data and the unlearned cell which the unlearned cell of the first block have Adding vector data based on the difference from the vector data. If comprised in this way, the vector data based on the difference of learning data and the unlearned vector data which the unlearned cell of the 1st block will have in the vector data which all the pseudo cells contained in the 2nd block easily have Can be added respectively.

  In the configuration including the second block, preferably, in the step of learning the learning data in the pseudo map, a learning load factor indicating a learning degree of the learning data with respect to the pseudo cell is set to div, and the pseudo cell included in the second block When the number is c, it includes a step of calculating an integrated value of the learning load factor based on an expression represented by div + 1 / c, and the vector data included in the pseudo cell of the pseudo map is converted into a pseudo cell of the unlearned map. The step of reflecting in the corresponding unlearned cell is the vector data of the pseudo cell in which the learning data has been learned by dividing the vector data of the pseudo cell of the pseudo map in which the learning data has been learned by the integrated value of the learning load factor. And calculating vector data of an unlearned cell reflecting the above. If comprised in this way, the vector data which the pseudo cell of the pseudo map by which learning data was learned can be easily reflected in the corresponding unlearned cell of an unlearned map.

  In the classification support map creation method according to the first aspect, preferably, after the step of learning the learning data in the pseudo map is executed for all the learning data, the step of reflecting the vector data of the pseudo map in the unlearned map is executed. To do. If comprised in this way, the vector data of the pseudo cell which learned all the learning data can be easily reflected in the unlearned cell of an unlearned map.

  A step of calculating the block data; a step of selecting a first block; a step of learning learning data in a pseudo cell included in a second block corresponding to the first block; and vector data of the pseudo map as an unlearned map Preferably, the step of calculating block data for each of a plurality of third blocks composed of a set of M or more unlearned cells extracted from the unlearned map, Selecting a first block from among the third blocks, causing a pseudo cell included in a second block corresponding to the first block to learn learning data, and reflecting vector data of the pseudo map in the unlearned map After the step is executed, a set of N (N <M) or more unlearned cells extracted from the unlearned map Calculating a block data for each of a plurality of fourth blocks, a step of selecting a first block from the plurality of fourth blocks, and a pseudo cell included in a second block corresponding to the first block. A step of learning the learning data and a step of reflecting the vector data of the pseudo map in the unlearned map are executed. According to this configuration, after learning data is relatively coarsely learned for each relatively large second block including M or more pseudo cells, a relatively large number of N (N <M) or more pseudo cells is included. The learning data can be learned relatively finely for each small second block. Thus, unlike the case where learning data is learned for each relatively small second block from the beginning, a plurality of regions made up of a plurality of pseudo cells having similar vector data have local clusters on the pseudo map. It is possible to suppress the dispersion and formation. For this reason, it can suppress that the area | region which consists of a several cell which has similar vector data on a classification | category assistance map is distributed and formed in multiple in the state with a local mass.

  A classification support map creating apparatus according to a second aspect of the present invention is a classification support map creating apparatus for creating a classification support map for supporting classification of unknown data consisting of vector data, each of which is unlearned vector data And an unlearned map storage means for storing an unlearned map including a plurality of unlearned cells, and a pseudo for storing a pseudo map corresponding to the plurality of unlearned cells and including a plurality of pseudo cells having predetermined vector data Map storage means, pseudo-map learning means for learning learning data composed of vector data in a pseudo map, and vector data of a pseudo-cell of a learned pseudo map reflected in an unlearned cell corresponding to a pseudo-cell of an unlearned map And an unlearned map reflecting means.

  In the classification support map creating apparatus according to the second aspect, as described above, learning data consisting of vector data is included in a pseudo map corresponding to a plurality of unlearned cells and including a plurality of pseudo cells having predetermined vector data. When learning a plurality of pieces of learning data sequentially by reflecting the vector data held by the pseudo cells of the learned pseudo map to the unlearned cells corresponding to the pseudo cells of the unlearned map, While the learning data is learned by the pseudo map, the vector data of each unlearned cell of the unlearned map is not changed. As a result, if a cell in which learning data is to be learned in the unlearned map is selected and then learning data is learned in a cell in the pseudo map corresponding to the cell in the selected unlearned map, Since the vector data of the cells in the unlearned map does not change when selecting the cells in which the learning data is to be learned, there is no inconvenience that the cells selected in the unlearned map differ depending on the order in which the learning data is learned. Thereby, also when learning several learning data, it can suppress that the self-organization map obtained differs by the order which learns learning data. Therefore, a classification support map useful for classification of unknown data that can correctly classify unknown data can be created regardless of the order of learning data.

  The classification support map creating apparatus according to the second aspect preferably further comprises block selection means for selecting a first block made up of a collection of unlearned cells from the unlearned map, and the pseudo-map learning means comprises an unlearned map Pseudo-cell learning means for causing the pseudo-cell included in the second block of the pseudo-map corresponding to the first block to learn learning data. With this configuration, the learning data can be learned in a plurality of pseudo cells included in the second block corresponding to the learning data. Therefore, for a single cell corresponding to the learning data and a cell in the vicinity region thereof, Unlike the case of learning the learning data, there is no need to set a neighborhood function for setting the range of the neighborhood region for learning the learning data. Thereby, even a person who is not skilled in setting the neighborhood function can easily learn the learning data, so that the classification support map can be easily created.

  In this case, it is preferable that the block selection unit outputs unlearned vector data of a plurality of unlearned cells included in the block for each of a plurality of blocks made up of a collection of unlearned cells extracted from the unlearned map. A block data calculating means for calculating block data as vector data of the block, and comparing the learning data with the block data of the plurality of blocks, and having block data corresponding to the learning data from the plurality of blocks. Block data comparison and selection means for selecting one block. If comprised in this way, the 1st block corresponding to learning data can be easily selected from an unlearned map.

  A program for executing a classification support map creating method according to a third aspect of the present invention is for executing a classification support map creating method for creating a classification support map that supports classification of unknown data consisting of vector data. A step of reading an unlearned map including a plurality of unlearned cells each having unlearned vector data, and a plurality of pseudo cells corresponding to the plurality of unlearned cells and having predetermined vector data A step of reading out a pseudo map including the step of learning the learning data composed of vector data in the pseudo map, and reflecting the vector data included in the pseudo cell of the learned pseudo map in the unlearned cell corresponding to the pseudo cell of the unlearned map And a step of causing.

  In the program for executing the classification support map creation method according to the third aspect, as described above, a pseudo map that corresponds to a plurality of unlearned cells and includes a plurality of pseudo cells having predetermined vector data, After learning the learning data consisting of vector data, the learning data is sequentially learned by reflecting the vector data of the pseudo cell of the learned pseudo map to the unlearned cell corresponding to the pseudo cell of the unlearned map. In this case, the vector data of each unlearned cell in the unlearned map is not changed while all the learning data is learned in the pseudo map. As a result, if a cell in which learning data is to be learned in the unlearned map is selected and then learning data is learned in a cell in the pseudo map corresponding to the cell in the selected unlearned map, Since the vector data of the cells in the unlearned map does not change when selecting the cells in which the learning data is to be learned, there is no inconvenience that the cells selected in the unlearned map differ depending on the order in which the learning data is learned. Thereby, also when learning several learning data, it can suppress that the self-organization map obtained differs by the order which learns learning data. Therefore, a classification support map useful for classification of unknown data that can correctly classify unknown data can be created regardless of the order of learning data.

  Hereinafter, embodiments of the present invention will be described with reference to the drawings. In the following description of the embodiments, a self-organizing map creating method and a self-organizing map creating device as an example of a classification support map creating method and a classification support map creating device according to the present invention will be described as an example.

(First embodiment)
FIG. 1 is a schematic diagram for explaining a configuration of a self-organizing map creating apparatus used in the self-organizing map creating method according to the first embodiment of the present invention. First, a self-organizing map creating apparatus used in the self-organizing map creating method of the first embodiment will be described with reference to FIG.

  As shown in FIG. 1, the self-organizing map creating apparatus 1 used in the self-organizing map creating method of the first embodiment includes an unlearned map storage unit 2a, a pseudo map storage unit 2b, a block selection unit 2c, A pseudo map learning unit 2d and an unlearned map reflection unit 2e are included. The block selection unit 2c includes a block weight data calculation unit 2f and a block weight data comparison selection unit 2g, and the pseudo map learning unit 2d includes a pseudo cell learning unit 2h.

  The unlearned map storage unit 2a is provided for storing an unlearned map described later. The pseudo map storage unit 2b is provided to store a pseudo map described later. The block selection unit 2c is provided for selecting a block made up of an unlearned cell aggregate corresponding to the learning data from the unlearned map. The pseudo map learning unit 2d is provided for causing the pseudo map to learn learning data. The unlearned map reflection unit 2e is provided to reflect the vector data included in the pseudo cell of the pseudo map in which the learning data has been learned to the unlearned cell corresponding to the pseudo cell of the unlearned map. Further, the block weight data calculation unit 2f, for each of the plurality of extracted blocks extracted from the unlearned map, based on the vector data of the plurality of unlearned cells included in the extracted block, the block weight as the vector data of the extracted block It is provided to calculate data. The block weight data comparison / selection unit 2g compares the learning data with the block weight data of the plurality of extracted blocks of the unlearned map, and the block weight having the smallest inter-vector distance from the plurality of extracted blocks to the learning data. It is provided to select an extraction block having data. Further, the pseudo cell learning unit 2h is provided for causing a pseudo cell included in a later-described learning target block of the pseudo map to learn learning data.

  The self-organizing map creation method according to the first embodiment includes a learning algorithm for creating learning data in an unlearned map and creating a self-organizing map (SOM), and classification of cells constituting the self-organizing map. And a classification algorithm for performing In the self-organizing map creation method according to the first embodiment, a self-organizing map (SOM) 3 as shown in FIG. 2 is created. The self-organizing map 3 is composed of a plurality of cells 3a arranged in a grid pattern. The plurality of cells 3a are arranged in the same number in the vertical and horizontal directions (10 in the first embodiment). Each cell 3a is configured to be able to have vector data composed of a plurality of elements (four elements in the first embodiment).

  3 to 8 are diagrams for explaining a learning algorithm of the self-organizing map creation method according to the first embodiment of the present invention. 9 to 13 are diagrams for explaining a classification algorithm of the self-organizing map creation method according to the first embodiment of the present invention. First, a learning algorithm of the self-organizing map creation method according to the first embodiment will be described with reference to FIGS.

(Learning algorithm)
As shown in FIG. 3, the learning algorithm according to the first embodiment includes steps S1 to S9. First, in step S1, an unlearned map is prepared. Here, the unlearned map is a map composed of a plurality of unlearned cells each having unlearned vector data. In preparing an unlearned map, first, a map composed of a plurality of cells having no vector data is prepared. Then, as shown in FIG. 4, predetermined vector data is assigned to each cell of the map made up of cells having no vector data using a random number table. In the first embodiment, vector data consisting of four elements 1 to 4 is assigned to each cell. Thereby, an unlearned map including a plurality of unlearned cells having predetermined unlearned vector data is prepared. In the first embodiment, it is assumed that the unlearned map includes n × n (10 × 10 in the first embodiment) unlearned cells.

Next, it progresses to the global area learning stage after step S2. In the global area learning stage, a relatively large block is extracted from the pseudo map, and learning data is learned for each extracted block, thereby causing the pseudo map to learn learning data relatively coarsely. Thereafter, the vector data of each pseudo cell of the learned pseudo map is reflected on each unlearned cell of the unlearned map. The pseudo map is a map composed of a plurality of pseudo cells corresponding to a plurality of unlearned cells of the unlearned map. As a result, the pseudo map according to the first embodiment includes n × n (10 × 10 in the first embodiment) pseudo cells corresponding to the unlearned map. In addition, each pseudo cell of the pseudo map is configured to have _a learning integrated value w _a (wa ₁ , w _a2 , w _a3 , w _a4 ) each including vector data. The learning integrated value w _a is a value that varies with the pseudo-map to learning data x is learned.

In step S2, the size of the extracted block extracted from the unlearned map and the pseudo map is set, and the quantization error err is initialized to “0”. In this step S2, in order to learn learning data for each relatively large block, the extraction block is composed of m × m = m ² (= M) or more cells, and m = n / 2 (5 = 10/2). That is, in the first embodiment, since the unlearned map and the pseudo map are composed of 10 × 10 cells, the extracted block is set to a relatively large block having 5 × 5 (m × m) or more cells. The quantization error err is an error between the learning data x (x ₁ , x ₂ , x ₃ , x ₄ ) and the block weight data of the selected block selected from the unlearned map described later. It is a value that serves as an index indicating how much (x ₁ , x ₂ , x ₃ , x ₄ ) is reflected in the selected block. Note that the block weight data is an average value of vector data of all cells constituting the block.

  Next, in step S3, the value of the quantization error err set to “0” in step S2 is copied to the previous quantization error err (before) (err (before) = err). After the quantization error err (before) is initialized to 0, the quantization error err is initialized to “0” (err = 0). Then, a subroutine is called. Note that the immediately preceding quantization error err (before) means that the learning operation is performed with the last learning data x of the previous learning cycle when steps S3 to S5 are set as one learning cycle and the learning cycle is repeatedly executed. Means the quantization error calculated in Further, with the processing in step S3, the immediately previous quantization error err (before) = 0 in the first learning cycle.

Next, as shown in FIG. 5, the subroutine of step S4 includes steps S11 to S17. In this subroutine, at step S11, extracts the all relatively large block having the same number of unlearned cells 5 × 5 = 5 ² pieces (= 25) or aspect from unlearned map, all that the extracted Block weight data B _i is calculated for a relatively large block (extracted block). The calculation of the block weight data B _i is performed by adding the vector data of each unlearned cell included in the extracted block for each element and then dividing by the total number of unlearned cells included in the extracted block. Thereby, block weight data B _i which is an average value of weights for each element of all the unlearned cells included in the extracted block is calculated. For example, as shown in FIG. 6, when an extraction block 50 including 5 × 5 = 25 unlearned cells 4a is extracted from the unlearned map 4, the following equations (1) to (4) are obtained. Thus, block weight data B _i (b _i1 , b _i2 , b _i3 , b _i4 ) is calculated. In the following formulas (1) to (4), a1 _{1 to} a1 ₂₅ , a2 _{1 to} a2 ₂₅ , a3 _{1 to} a3 ₂₅ , and a4 _{1 to} a4 ₂₅ are respectively included in the extraction block 50. It is the value of element 1 to element 4 of each unlearned cell 4a included.

Block weight data b _i1 = (a1 ₁ + a1 ₂ + a1 ₃ +... + A1 ₂₅ ) / 25 (1)
Block weight data b _i2 = (a2 ₁ + a2 ₂ + a2 ₃ +... + A2 ₂₅ ) / 25 (2)
Block weight data b _i3 = (a3 ₁ + a3 ₂ + a3 ₃ +... + A3 ₂₅ ) / 25 (3)
Block weight data b _i4 = (a4 ₁ + a4 ₂ + a4 ₃ +... + A4 ₂₅ ) / 25 (4)
Next, in step S12, as shown in FIG. 7, preparation of the pseudo map 5 is performed, and learning integrated values w _a (w _a1 , w _a2 , w _a3 , w of each pseudo cell 5a constituting the pseudo map 5 are prepared. _a4 ) is initialized to “0” ((w _a1 , w _a2 , w _a3 , w _a4 ) = (0, 0, 0, 0)). Next, in step S13, learning data x (x ₁ , x ₂ , x ₃ , x from among relatively large extracted blocks composed of unlearned cells of m × m (5 × 5) or more of the unlearned map described above. ₄ ) The extracted block having the block weight data B _i having the smallest inter-vector distance is selected. For example, as shown in FIG. 6, among all extraction blocks composed of 5 × 5 or more cells that can be extracted from the unlearned map 4, the 5 × 5 extraction block 50 has a vector-to-vector distance from the learning data x. When it has the smallest block weight data B _i , this extraction block 50 is selected as the selected block. The block weight data of the selected block is B _s (b _s1 , b _s2 , b _s3 , b _s4 ).

  Next, in step S14, a quantization error err is calculated. The calculation of the quantization error err is performed by the following equation (5). In the learning calculation of the first learning data x, in the following equation (5), as set in step S3 (see FIG. 3), the immediately preceding quantization error err (before) = 0. .

Quantization error err = err (before) + {(x ₁ −b _s1 ) ² + (x ₂ −b _s2 ) ² + (x ₃ −b _s3 ) ² + (x ₄ −b _s4 ) ² } ^1/2 ... (5)
Next, in step S15, a learning operation for learning the learning data x (x ₁ , x ₂ , x ₃ , x ₄ ) is performed on all the pseudo cells constituting the learning target block of the pseudo map. The learning target block is a block that is selected from the pseudo map and from which the learning data x is learned, and corresponds to the selected block selected in step S13. That is, as shown in FIG. 6, when the extraction block 50 is selected as the selected block from the unlearned map 4, as shown in FIG. 8, the block 51 of the pseudo map 5 corresponding to the extraction block 50 is the learning target block. It becomes. Further, the learning _{operation, w a = w a (before} ) + x / using the formula represented by c, the learning data in the learning before the learning with the pseudo cells included in the target block learning integrated value _w a (before) It carried out by calculating the learned integration value w _a of the pseudo cell by adding the vector data (x / c) based on x. Then, by performing this learning calculation for all the pseudo cells included in the learning target block, the learning integrated value w _a of all the pseudo cells included in the learning target block is calculated. In the above formula, c is the number of pseudo cells included in the learning target block.

For example, as shown in FIG. 8, in a learning target block (block 51) made up of 5 × 5 (= 25) pseudo cells 5a, a learning integrated value w _a (before) (w _ab1 , w _ab2 , When learning operation of learning data x (x ₁ , x ₂ , x ₃ , x ₄ ) is performed on the pseudo cell 52 having w _ab3 , w _ab4 ), the following equations (6) to (9) As shown, the learning integrated value w _a (wa ₁ , w _a2 , w _a3 , w _a4 ) of the pseudo cell 52 is calculated. In the calculation of the learning integrated value w _a corresponding to the first learning data x, w _{ab1 to} w _ab4 are all expressed in the following formulas (6) to (9) by the process of step S12 (see FIG. 5). “0”.

Learning integrated value _{w a1} (element _{1) = w ab1 + x 1} /25 ··· (6)
Learning integrated value _{w a2} (element _{2) = w ab2 + x 2} /25 ··· (7)
Learning integrated value _{w a3} (element _{3) = w ab3 + x 3} /25 ··· (8)
Learning integrated value _{w a4} (element _{4) = w ab4 + x 4} /25 ··· (9)
In step S15, the integrated value (div _a ) of the learning load factor div indicating the degree of learning of the learning data x for each pseudo cell of the learning target block is also calculated. Calculation of the integrated value div _a of the learning load factor, the learning load factor before learning calculation pseudo cell having a div (before), the number of pseudo cells included in the learning target block when the c, div _a = Div (before) + 1 / c. That is, the learning load factor integrated value div _a is a value obtained by integrating the reciprocal of the number c of pseudo cells included in the learning target block every time the learning target block learns the learning data x. Specifically, as shown in FIG. 8, the integrated value div _a of the learning load factor is calculated for the pseudo cell 52 included in the learning target block (block 51) composed of 5 × 5 (= 25) pseudo cells 5a. If so, the calculation is performed as shown in the following equation (10). In the calculation of the integrated value div _a of the learning load factor corresponding to the first learning data x, the learning load factor div (before) before the learning calculation is “0”, so in the following equation (10) Div (before) = 0.

Integrated value of learning load factor div _a = div (before) +1/25 (10)
Next, in step S16, it is determined whether or not the learning calculation for the number of learning data x (all the learning data x) has been completed. When the learning calculation for the number of learning data x is completed, the process proceeds to the next step S17. On the other hand, when the learning calculation for the number of learning data x has not been completed, the above steps S13 to S16 are repeatedly executed.

Next, in step S < _b > 17, an operation is performed to reflect the learning integrated value w _{a included} in the pseudo cell of the pseudo map in the vector data of all the unlearned cells that form the unlearned map. This operation is represented by _w = _{w a} / div a using learning integrated value _{w a (w a1, w a2} , w a3, w a4) and the integrated value div _a learning load factor calculated in step S15 The vector data w (w ₁ , w ₂ , w ₃ , w ₄ ) of the unlearned cell reflecting the learning integrated value w _a of the pseudo cell is calculated by the equation. Specifically, vector data w (w ₁ , w ₂ , w ₃ , w ₄ ) of the unlearned cell reflecting the learning integrated value w _a is calculated as shown in the following equations (11) to (14). To do.

Vector data w ₁ = w _a1 / div _{a of the} unlearned cell reflecting the learning integrated value w _a1 (11)
Vector data w ₂ = wa ₂ / div _{a of the} unlearned cell reflecting the learning integrated value w _a2 (12)
Vector data w ₃ = w _a3 / div _{a of the} unlearned cell reflecting the learning integrated value w _a3 (13)
Vector data w ₄ = w _a4 / div _a (14) of an unlearned cell reflecting the learning integrated value w _a4
From the above formulas (11) to (14), for example, a 5 × 5 learning target block in which the learning data X1 (0.03, 3.1, 0.2, 2.8) is learned, and the learning data X2 3 × 3 learning target blocks learned (0.4, 2.9, 0.2, 2.7) and learning data X3 (0.01, 3.2, 0.1, 2.9) There if to reflect the learned integration value w _a having a predetermined pseudo cells included in the learning object 4 × 4 block learned into vector data unlearned cell, vector data w of the unlearned cells ( w ₁ , w ₂ , w ₃ , w ₄ ) are calculated as in the following equations (15) to (18).

w ₁ = (0.03 × 1/25 + 0.4 × 1/9 + 0.01 × 1/16) / (1/25 + 1/9 + 1/16) = 0.22 (15)
w ₂ = (3.1 × 1/25 + 2.9 × 1/9 + 3.2 × 1/16) / (1/25 + 1/9 + 1/16) = 3.03 (16)
w ₃ = (0.2 × 1/25 + 0.2 × 1/9 + 0.1 × 1/16) / (1/25 + 1/9 + 1/16) = 0.17 (17)
w ₄ = (2.8 × 1/25 + 2.7 × 1/9 + 2.9 × 1/16) / (1/25 + 1/9 + 1/16) = 2.78 (18)
The vector data w of the unlearned cell is determined by dividing the learning integrated value w _a by the learning load factor integrated value div _a for the following reason. That is, since the learning integrated value w _a is obtained by integrating the vector data x / c obtained by dividing the learning data x by the number c of pseudo cells included in the learning target block, the learning integrated value w _a The value of the inverse 1 / c of the number c of pseudo cells included in the block is accumulated. Therefore, in order to calculate the vector data w learning integrated value w _a to the accumulated 1 / c element unlearned cell to remove the, is an integrated value of 1 / c a learned integration value w _a learning load The vector data w of the unlearned cell is obtained by dividing by the integrated value div _a of the rate. In this case as well, the vector data w of the unlearned cell remains while the element of the learning degree of each learning data x (ratio of the influence of the learning data x) corresponding to the number c of pseudo cells included in the learning target block remains. Is required. That is, by changing the above formula (15), it can be expressed as the following formula (19).

w ₁ = (144 × 0.03 + 400 × 0.4 + 225 × 0.01) / 769 (19)
According to this equation (19), 0.4 learned in a learning target block composed of 3 × 3 pseudo cells among the coefficients corresponding to the three learning data (0.03, 0.4 and 0.01), respectively. And the coefficient (225) for 0.01 learning data learned by the learning target block composed of 4 × 4 pseudo cells is the next largest, and the coefficient (400) for the learning data of 5 × 5 × It can be seen that the coefficient (144) applied to the learning data of 0.03 learned by the learning target block consisting of 5 pseudo cells is the smallest. As a result, as the size of the learning target block increases from the smallest 3 × 3 learning target block to 4 × 4 and 5 × 5, the influence ratio (degree of learning) of the learning data x decreases. It can be seen that the vector data w of the unlearned cell is calculated.

By the above equations (11) to (14), vector data w reflecting the learning integrated value w _a is calculated for all the unlearned cells of the unlearned map corresponding to all the pseudo cells constituting the pseudo map. Then, the vector data of each unlearned cell is rewritten with the calculated vector data w.

  Next, in step S5 in FIG. 3, the learning count has reached a preset specified value D1, or the difference between the quantization error err and the previous quantization error err (before) is set to a preset specified value. It is determined whether it is smaller than D2. The number of learnings means the number of times that the learning cycle in steps S3 to S5 has been executed. If the learning count has reached the prescribed value D1, or if the difference between the quantization error err and the previous quantization error err (before) is smaller than the prescribed value D2, the process proceeds to the next step S6. On the other hand, if the number of learnings has not reached the prescribed value D1, and the difference between the quantization error err and the previous quantization error err (before) is not less than the prescribed value D2, the learning cycle in steps S3 to S5 is performed. Run repeatedly.

Next, after step S6, the process proceeds to the normal learning stage, which is the latter half of learning. The normal learning stage includes steps S6 to S9. In the normal learning stage, the learning data x is learned up to a block smaller than the global area learning stage. Thereby, in the normal learning stage, the learning data x is learned in the pseudo map more finely than in the global area learning stage. In the setting of the size of the extracted block (m × m) in step S6, m = 2 is set so that the learning data x is learned for all the blocks that can be extracted from the pseudo map. As a result, in step S11 of the subroutine (see FIG. 5), all blocks composed of 2 × 2 = 4 (= N) or more cells are extracted from the unlearned map as extracted blocks. Then, the block weight data is calculated for all the extracted blocks, and the block having the block weight data having the smallest inter-vector distance from the learning data x is selected as a selected block from all the extracted blocks. Then, the learning calculation of the learning data x is performed on the learning target block of the pseudo map corresponding to the selected block. Then, to reflect the vector data each unlearned cell unlearned map learning integrated value w _a of each pseudo cell included in the learning target block calculated by the learning calculation has.

  In the normal learning stage, in step S9 (see FIG. 3), the number of times of learning has reached a predetermined value D3 set in advance, or the difference between the quantization error err and the previous quantization error err (before) is It is determined whether the value is smaller than a predetermined value D4 set in advance. Thereby, in the learning cycle composed of steps S7 to S9, the number of times of learning reaches the specified value D3, or the difference between the quantization error err and the previous quantization error err (before) becomes smaller than the specified value D4. Done. Note that the prescribed value D3 of the number of learnings in the normal learning stage is set to a value larger than the prescribed value D1 of the number of learnings in the global area learning stage. Thus, more learning cycles are executed in the normal learning stage than in the global area learning stage. Also, the prescribed value D4 of the difference between the quantization error err and the previous quantization error err (before) in the normal learning stage is the difference between the quantization error err and the previous quantization error err (before) in the global area learning stage. A value smaller than the prescribed value D2 is set. Thereby, in the normal learning stage, the learning cycle is executed until the difference between the quantization error err and the previous quantization error err (before) is smaller than in the global area learning stage. The processing method other than the above in the normal learning stage consisting of steps S6 to S9 is the same as the processing method in the global area learning stage consisting of steps S2 to S5 described above.

As described above, the learning data x is learned in the pseudo map, and the learning integrated value w _{a included} in the pseudo cell of the learned pseudo map is reflected in the vector data included in the unlearned cell of the unlearned map. Create an organizational map.

In the first embodiment, as described above, the learning data x corresponds to the plurality of unlearned cells 4a and includes the plurality of pseudo cells 5a having the learning integrated value w _a (before) before learning. , The learned integrated value w _{a included} in the pseudo cell 5a of the learned pseudo map 5 is reflected in the unlearned cell 4a corresponding to the pseudo cell 5a of the unlearned map 4, whereby a plurality of learning data x Are sequentially learned, the vector data of each unlearned cell 4a of the unlearned map 4 is not changed while all the learning data x is learned by the pseudo map 5. Thereby, after selecting the unlearned cell 4a in which the learning data x is to be learned in the unlearned map 4, the pseudo cell of the pseudo map 5 corresponding to the unlearned cell 4a (selected block) of the selected unlearned map 4 Vector data of the unlearned cell 4a of the unlearned map 4 when selecting the unlearned cell 4a in which the learned data x should be learned in the unlearned map 4 by learning the learned data x in 5a (learning target block). Therefore, there is no inconvenience that the unlearned cell 4a selected in the unlearned map 4 differs depending on the order of learning the learning data x. Thereby, also when learning the some learning data x, it can suppress that the self-organization map 3 obtained by the order which learns the learning data x differs. Therefore, a classification support map useful for classification of unknown data that can correctly classify unknown data can be created regardless of the order in which the learning data x is learned.

  In the first embodiment, the selected block corresponding to the learning data x is selected from the extracted blocks of the unlearned map 4, and the pseudo cell 5a included in the learning target block of the pseudo map 5 corresponding to the selected block is selected. Unlike the case where learning data x is learned for a single cell corresponding to learning data x and a cell in the vicinity area by learning learning data x, the range of the neighboring area where learning data x is learned is set. There is no need to set a neighborhood function to Thereby, even a person who is not skilled in setting the neighborhood function can easily learn the learning data x, so the self-organizing map 3 can be easily created.

In the first embodiment, for each learning target block including 5 × 5 = 25 (M) or more pseudo cells 5a, each pseudo cell 5a included in the learning target block learns the learning data x, after learning integrated value w _a to each pseudo cell 5a pseudo map 5 has is reflected in the unlearned cell 4a unlearned map 4, 2 × 2 = 4 pieces (N pieces (N <M)) or pseudo For each learning target block composed of the cells 5a, each of the pseudo cells 5a included in the learning target block learns the learning data x, and the learning integrated value w _a of each pseudo cell 5a of the pseudo map 5 is stored in the unlearned map 4. By reflecting each unlearned cell 4a, the learning data x is relatively coarsely learned for each relatively large learning target block including 25 or more pseudo cells 5a, and then 4 or more pseudo cells 5a are included. ratio Specifically small learning object blocks also each, it is possible to learn the relatively fine learning data x. Thus, unlike the case where learning data x is learned for each relatively small learning target block from the beginning, _a region composed of a plurality of pseudo cells 5 _a having similar learning integrated values w _a is locally clustered on the pseudo map 5. It is possible to suppress the formation of a plurality of particles dispersed in the state of having. For this reason, on the self-organizing map 3 created by reflecting the learning integrated value 5a of the pseudo map 5 on the unlearned map 4, a region composed of a plurality of cells 3a having similar vector data has a local cluster. In this state, it is possible to suppress the formation of a plurality of dispersed particles.

  Next, a classification algorithm of the self-organizing map creation method according to the first embodiment will be described with reference to FIGS.

(Classification algorithm)
In this classification algorithm, the cells constituting the self-organizing map created by the learning algorithm are classified. The class is a classification name given to each learning data x in order to classify the learning data x. In the first embodiment, each learning data x is classified into one of three classes of α, β, and γ, and the cells constituting the self-organizing map are classified into three classes of α, β, and γ. The case of classifying into 2 will be described. Further, each of the cells constituting the self-organizing map includes a class affiliation values R _c indicating the likelihood ratio belonging to each class. In the first embodiment, each cell has three class membership values (R _α , R _β and R _γ ).

First, as shown in FIG. 9, in step S21, it initializes the class affiliation value R _c to "0" in all cells 3a constituting the self-organizing map 3 (see FIG. 10). That is, as shown in FIG. 10, class membership values R _α , R _β and R _γ corresponding to the α, β and γ classes of all the cells 3a constituting the self-organizing map 3 are set to “0”.

Next, in step S22, all the blocks in which the same number of cells 3a are arranged in the vertical and horizontal directions of 2 × 2 or more are extracted from the self-organizing map 3, and the block weight data for all the extracted blocks (extraction blocks) B _j (b _j1 , b _j2 , b _j3 , b _j4 ) is calculated. The block weight data B _j is calculated in the same manner as the block weight data B _i (see formula (1) to formula (4)) based on the learning algorithm. That is, as shown in FIG. 11, when the extraction block 53 consisting of 2 × 2 = 4 cells 3a is extracted, the block weight data B _{j as} shown in the following equations (20) to (23) _: (B _j1 , b _j2 , b _j3 , b _j4 ) are calculated. In the following formulas (20) to (23), a1 _{1 to} a1 ₄ , a2 _{1 to} a2 ₄ , a3 _{1 to} a3 ₄ , and a4 _{1 to} a4 ₄ are respectively included in the extraction block 53. It is the value of element 1 to element 4 of each included cell 3a.

Block weight data b _j1 = (a1 ₁ + a1 ₂ + a1 ₃ + a1 ₄ ) / 4 (20)
Block weight data b _j2 = (a2 ₁ + a2 ₂ + a2 ₃ + a2 ₄ ) / 4 (21)
Block weight data b _j3 = (a3 ₁ + a3 ₂ + a3 ₃ + a3 ₄ ) / 4 (22)
Block weight data b _j4 = (a4 ₁ + a4 ₂ + a4 ₃ + a4 ₄ ) / 4 (23)
Next, in step S23, a classification target block to be classified is determined. This classification target block is determined by determining that the distance between vectors with the learning data x (x ₁ , x ₂ , x ₃ , x ₄ ) is the largest among all the extracted blocks made up of the above 2 × 2 cells 3a. This is done by selecting a block having small block weight data _Bj . For example, as shown in FIG. 12, among all the extracted blocks that can be extracted from the self-organizing map 3, the 3 × 3 block 54 has the block weight data B _j having the smallest inter-vector distance from the learning data x. If so, this block 54 is determined as a classification target block.

Next, in step S24, and calculates the class affiliation values R _d of all cells 3a included in the classification target block. Calculation of this class belongs value R _d is carried out by calculation according to formula (24) below. In the following equation (24), “d” of R _d is the class to which the learning data x used in step S23 belongs. That is, the calculation of Expression (24) is performed only for the class to which the learning data x belongs. R _d (before) is a class belonging value corresponding to the d class of the cell 3a before calculation. U is a coefficient, and is obtained, for example, by u = 1 / (the number of learning data x belonging to the d class). Further, b _n is the number of cells 3a constituting the classification target block.

R _d = R _d (before) + u (1 / b _n ) (24)
Specifically, as shown in FIG. 12, a block 54 consisting of 3 × 3 = 9 cells corresponding to learning data x belonging to the α class is determined as a classification target block, and the classification target block When calculating the class affiliation value R _α of the nine cells 55 in (block 54), it is calculated by the following equation (25). In the calculation of the class affiliation value corresponding to the first learning data x, R _α (before) = 0 in the following equation (25) by the processing of step S21.

R _α = R _α (before) + u (1/9) (25)
Classifying target blocks are determined for all learning data x, and the class belonging value R _d of the class (α, β, or γ) to which each learning data x belongs is assigned to all cells 3a included in the classifying target block. It is obtained by the calculation of the above equation (24). The class belonging values R _α , R _β , and R _{γ of} each class α, β, and _γ are calculated for each cell 3 a by the calculation processing of the class belonging value R _d as described above.

Next, in step S25, whether the calculation of the class affiliation values R _d to the number of all learning data x (all learning data x) has been finished is determined. When the operation to the number of all classes belonging values R _d of the learning data x is completed, the process proceeds to the next step S26. On the other hand, when the operation corresponding to the number of all classes belonging values R _d of the learning data x is not finished, repeatedly executes the steps S22 to S25.

Next, in step S26, the affiliation class of all the cells 3a constituting the self-organizing map 3 is determined. In the determination of the belonging class, the class having the maximum value among the class belonging values R _d (R _α , R _β , R _γ ) possessed by each cell 3a is determined as the belonging class of the cell 3a. For example, as shown in FIG. 13, when the cell 56 has class belonging values of R _α = 3, R _β = 0.1, and R _γ = 0.4, respectively, the class belonging value of the α class Since _Rα is the maximum, the class to which the cell 56 belongs is determined to be the α class. In this way, as shown in FIG. 13, by classifying all the cells 3a, the classification of all the cells 3a constituting the self-organizing map 3 is performed.

(Second Embodiment)
FIG. 14 is a flowchart for explaining a subroutine of the learning algorithm of the self-organizing map creation method according to the second embodiment of the present invention. Next, a learning algorithm of the self-organizing map creation method according to the second embodiment of the present invention will be described with reference to FIG.

In the learning algorithm of the self-organizing map creation method according to the second embodiment, unlike the learning algorithm of the self-organization map creation method according to the first embodiment, learning is performed at the time of learning calculation for the pseudo cell included in the learning target block. Vector data based on the difference between learning data x and vector data w _{b of} unlearned cells included in the selected block is added to the learning integrated value w _a (before) of all the pseudo cells included in the target block. Add each. Thus, the learning integrated value w _a of the pseudo cells included in the learning target block is calculated. That is, in the second embodiment, as shown in FIG. 14, in step _S25, using the formula represented by _{w a = w a (before)} + t · (x-w b) / c, the learning target block The accumulated learning value w _a (before) of the included pseudo cell before learning is calculated based on the difference between the learning data x and the vector data w _b of the unlearned cell included in the selected block (t · (x− The learning integrated value w _a of the pseudo cell is calculated by adding w _b ) / c). Then, by performing this learning calculation for all the pseudo cells included in the learning target block, the learning integrated value w _a of all the pseudo cells included in the learning target block is calculated. In the above equation, t is a constant representing the degree to which the learning data x is learned by the pseudo cell, and c is the number of pseudo cells included in the learning target block.

Specifically, in the learning target block (block 51) composed of 5 × 5 (= 25) pseudo cells 5a (see FIG. 8), the learning integrated value w _a (before) (w _ab1 , w _ab2 ) before learning. , W _ab3 , w _ab4 ), when learning operation of learning data x (x ₁ , x ₂ , x ₃ , x ₄ ) is performed on the pseudo cell 52, the following formulas (26) to (29) ), The learning integrated value w _a (wa ₁ , w _a2 , w _a3 , w _a4 ) of the pseudo cell 52 is calculated. At this time, the unlearned cell in the selected block (extraction block 50) of the unlearned map 4 (see FIG. 6) corresponding to the pseudo cell 52 is represented by vector data w _b (w _b1 , w _b2 , w _b3 , w _b4 ).

Integrated learning value w _a1 (element 1) = w _ab1 + t · (x ₁ −w _b1 ) / 25 (26)
Integrated learning value w _a2 (element 2) = w _ab2 + t · (x ₂ −w _b2 ) / 25 (27)
Integrated learning value w _a3 (element 3) = w _ab3 + t · (x ₃ −w _b3 ) / 25 (28)
Learning integrated value w _a4 (element 4) = w _ab4 + t · (x ₄ −w _b4 ) / 25 (29)
In the second embodiment, in the calculation for reflecting the learning integrated value w _a of the pseudo cell of the pseudo map to the vector data of all the unlearned cells constituting the unlearned map in step 27, w = the formula represented by w _b + v · _{w a,} and calculates the vector data _w unlearned cells reflecting the learned integration value _{w a} pseudo cell _{(w 1, w 2, w} 3, w 4). In this equation, v is a coefficient. Specifically, the vector data w (w ₁ , w ₂ , w ₃ , w ₄ ) of the unlearned cell reflecting the learning integrated value w _a is expressed as shown in the following equations (30) to (33). calculate.

Vector data w ₁ = w _b1 + v · w _a1 (30) of the unlearned cell reflecting the learning integrated value w _a1
Vector data w ₂ = w _b2 + v · w _{a2 of the} unlearned cell reflecting the learning integrated value w _a2 (31)
Vector data w ₃ = w _b3 + v · w _a3 (32) of the unlearned cell reflecting the learning integrated value w _a3
Vector data w ₄ = w _b4 + v · w _a4 (33) of the unlearned cell reflecting the learning integrated value w _a4
In the second embodiment, unlike the first embodiment, the learning load factor integrated value div _a is not obtained, and learning in which the inverse 1 / c of the number c of pseudo cells included in the learning target block is accumulated is stored. The vector data w of the unlearned cell is obtained without dividing the integrated value w _a by the integrated value div _a of the learning load factor.

From the above equations (30) to (33), for example, a 5 × 5 learning target block in which the learning data X1 (0.03, 3.1, 0.2, 2.8) is learned, and the learning data X2 3 × 3 learning target blocks learned (0.4, 2.9, 0.2, 2.7) and learning data X3 (0.01, 3.2, 0.1, 2.9) The learning integrated value w _a included in the predetermined pseudo cell included in the 4 × 4 learning target block in which the learning is performed has the vector data (0.02, 2.9, 0.22, 2.7). When reflecting in the learning cell, the vector data w (w ₁ , w ₂ , w ₃ , w ₄ ) of the unlearned cell is calculated as in the following equations (34) to (37). In this case, the coefficient v = 1.

w ₁ = 0.02 + 1 × (1 × (0.03-0.02) / 25 + 1 × (0.4−0.02) / 9 + 1 × (0.01−0.02) / 16) = 0.06 ... (34)
w ₂ = 2.9 + 1 × (1 × (3.1-2.9) / 25 + 1 × (2.9-2.9) / 9 + 1 × (3.2-2.9) / 16) = 2.93 ... (35)
w ₃ = 0.22 + 1 × (1 × (0.2−0.22) / 25 + 1 × (0.2−0.22) / 9 + 1 × (0.1−0.22) / 16) = 0.21 ... (36)
w ₄ = 2.7 + 1 × (1 × (2.8-2.7) / 25 + 1 × (2.7-2.7) / 9 + 1 × (2.9-2.7) / 16) = 2.72 ... (37)
By the above equations (30) to (33), vector data w reflecting the learning integrated value w _a is calculated for all unlearned cells of the unlearned map corresponding to all the pseudo cells constituting the pseudo map. . Then, the vector data of each unlearned cell is rewritten with the calculated vector data w. The other processing methods of the self-organizing map creating method according to the second embodiment are the same as the processing methods of the self-organizing map creating method according to the first embodiment.

In the second embodiment, as described above, the learning data x corresponds to the plurality of unlearned cells 4a and includes the plurality of pseudo cells 5a having the learning integrated value w _a (before) before learning. , The learned integrated value w _{a included} in the pseudo cell 5a of the learned pseudo map 5 is reflected in the unlearned cell 4a corresponding to the pseudo cell 5a of the unlearned map 4, whereby a plurality of learning data x Are sequentially learned, the vector data of each unlearned cell 4a of the unlearned map 4 is not changed while all the learning data x is learned by the pseudo map 5. Thereby, after selecting the unlearned cell 4a in which the learning data x is to be learned in the unlearned map 4, the pseudo cell of the pseudo map 5 corresponding to the unlearned cell 4a (selected block) of the selected unlearned map 4 Vector data of the unlearned cell 4a of the unlearned map 4 when selecting the unlearned cell 4a in which the learned data x should be learned in the unlearned map 4 by learning the learned data x in 5a (learning target block). Therefore, there is no inconvenience that the unlearned cell 4a selected in the unlearned map 4 differs depending on the order of learning the learning data x. Thereby, also when learning the some learning data x, it can suppress that the self-organization map 3 obtained by the order which learns the learning data x differs. Therefore, a classification support map useful for classification of unknown data that can correctly classify unknown data can be created regardless of the order in which the learning data x is learned.

  The effects of the second embodiment other than those described above are the same as the effects of the first embodiment.

  The embodiment disclosed this time should be considered as illustrative in all points and not restrictive. The scope of the present invention is shown not by the above description of the embodiments but by the scope of claims for patent, and further includes all modifications within the meaning and scope equivalent to the scope of claims for patent.

  For example, in the above embodiment, the classification support map creation method of the present invention, the program for executing the method, the self-organization map creation method as an example of the classification support map creation apparatus, the program for executing the program, and the self-organization Although the present invention is not limited to this, the present invention is not limited to this, and a self-organizing map creating method, a program for executing the method, a classification support map creating method other than the self-organizing map creating device, and the same are executed. The present invention can also be applied to a program for the above and a classification support map creating apparatus.

  In the above embodiment, the learning algorithm prepares an unlearned map by assigning predetermined vector data using a random number table to each cell of a map made up of cells having no vector data. However, the present invention is not limited to this, and an unlearned map may be prepared using a method other than the above. For example, an unlearned map may be prepared by assigning vector data having all elements of “0” to all cells of a map including cells having no vector data.

In the second embodiment, the unlearned cell vector data w is obtained without dividing the learning integrated value w _a by the learning load factor integrated value div _a . However, the present invention is not limited to this, and the second If the method of obtaining the learned integration value w _a by integrating the difference between the vector data of the learning data x and unlearned cell according to the embodiment also divides the learning integrated value w _a learning load factor integration value div _a Thus, the vector data w of the unlearned cell may be obtained.

  Moreover, in the said embodiment, although each cell which comprises a map was represented by the grid, this invention is not restricted to this, You may represent each cell by various shapes other than a grid. For example, each cell may be represented by a circle or a point.

  In the above embodiment, an example in which each cell of the self-organizing map is classified into three classes of α, β, and γ has been described. However, the present invention is not limited to this, and each cell of the self-organizing map is The present invention can also be applied to classification into classes other than three.

  In the above embodiment, vector data composed of four elements is used as vector data. However, the present invention is not limited to this, and any vector data composed of two or more elements may be used.

  In the above embodiment, the average value of the weight of each element of all cells included in the block is used as the block data. However, the present invention is not limited to this, and each cell of all cells included in the block as block data is used. A standard deviation of element weights may be used.

  In the above embodiment, an example in which a map in which the same number of cells are arranged vertically and horizontally is used as an unlearned map and a pseudo map, but the present invention is not limited thereto, and as an unlearned map and a pseudo map, Maps with different numbers of cells arranged vertically and horizontally may be used.

It is a schematic diagram for demonstrating the structure of the self-organization map preparation apparatus used for the self-organization map preparation method by 1st Embodiment of this invention. It is the figure which showed the structure of the self-organization map created by the self-organization map creation apparatus by 1st Embodiment shown in FIG. It is a flowchart for demonstrating the learning algorithm of the self-organizing map creation method by 1st Embodiment of this invention. It is a figure for demonstrating the learning algorithm of the self-organizing map creation method by 1st Embodiment of this invention. It is a flowchart for demonstrating the subroutine of the learning algorithm by 1st Embodiment shown in FIG. It is a figure for demonstrating the learning algorithm of the self-organizing map creation method by 1st Embodiment of this invention. It is a figure for demonstrating the learning algorithm of the self-organizing map creation method by 1st Embodiment of this invention. It is a figure for demonstrating the learning algorithm of the self-organizing map creation method by 1st Embodiment of this invention. It is a flowchart for demonstrating the classification algorithm of the self-organizing map creation method by 1st Embodiment of this invention. It is a figure for demonstrating the classification algorithm of the self-organization map creation method by 1st Embodiment of this invention. It is a figure for demonstrating the classification algorithm of the self-organization map creation method by 1st Embodiment of this invention. It is a figure for demonstrating the classification algorithm of the self-organization map creation method by 1st Embodiment of this invention. It is a figure for demonstrating the classification algorithm of the self-organization map creation method by 1st Embodiment of this invention. It is a flowchart for demonstrating the subroutine of the learning algorithm of the self-organizing map creation method by 2nd Embodiment of this invention.

Explanation of symbols

1 Self-organizing map creation device 2a Unlearned map storage unit (unlearned map storage means)
2b Pseudo map storage unit (pseudo map storage means)
2c Block selection unit (block selection means)
2d pseudo map learning unit (pseudo map learning means)
2e Unlearned map reflection unit (unlearned map reflection means)
2f Block weight data calculation unit (block data calculation means)
2g block weight data comparison / selection unit (block data comparison / selection means)
2h Pseudo cell learning unit (pseudo cell learning means)
3 Self-organizing map (Classification support map)
3a cell 4 unlearned map 4a unlearned cell 5 pseudo map 5a pseudo cell 50 extraction block (block, third block)

Claims (14)

A classification support map creation method for creating a classification support map for supporting classification of unknown data composed of vector data,
Providing an unlearned map including a plurality of unlearned cells each having unlearned vector data;
Preparing a pseudo map that corresponds to the plurality of unlearned cells and includes a plurality of pseudo cells having predetermined vector data;
Learning learning data composed of vector data in the pseudo map;
Reflecting the vector data of the learned pseudo cell of the pseudo map to an unlearned cell corresponding to the pseudo cell of the unlearned map. Selecting a first block consisting of a collection of the unlearned cells from the unlearned map;
The step of causing the pseudo map to learn the learning data includes the step of causing the pseudo cell included in the second block of the pseudo map corresponding to the first block of the unlearned map to learn the learning data. A method for creating a classification support map according to 1. Selecting a first block from the unlearned map comprises:
The vector data of the block is extracted from the unlearned map, for each of a plurality of blocks made up of a collection of the unlearned cells, based on the unlearned vector data of the plurality of unlearned cells included in the block. Calculating block data as:
The learning data is compared with the block data of the plurality of blocks, and a first block having the block data corresponding to the learning data is selected from the plurality of blocks. Classification support map creation method.   The step of learning the learning data in the pseudo cell included in the second block includes the step of adding vector data based on the learning data to the vector data of all the pseudo cells included in the second block. The method for creating a classification support map according to claim 2 or 3. In the step of adding vector data based on the learning data, the vector data of the pseudo cell included in the second block is set as w _a , the learning data is set as x, and the number of the pseudo cells included in the second block is determined as a step of adding the vector data based on the learning data to the vector data of the pseudo cell included in the second block based on an expression represented by w _a + x / c, Item 5. The classification support map creation method according to Item 4.   The step of causing the pseudo cell included in the second block to learn the learning data includes the vector data included in all the pseudo cells included in the second block and the unlearned cell of the first block. 4. The classification support map creation method according to claim 2, further comprising a step of adding vector data based on a difference from unlearned vector data. The step of adding the vector data based on the difference between the unlearned of vector data included in the unlearned cells of the first block and the training data, the vector data included in the pseudo cells included in the second block and w _a, When the constant is t, the learning data is x, the unlearned vector data of the unlearned cells in the first block is w _b, and the number of pseudo cells included in the second block is c, Based on the formula represented by w _a + t · (x−w _b ) / c, the vector data of the pseudo cell included in the second block has the learning data and the unlearned cell of the first block. The classification support map creation method according to claim 6, further comprising the step of adding vector data based on a difference from unlearned vector data. The step of learning the learning data in a pseudo map is performed when the learning load factor indicating the degree of learning of the learning data with respect to the pseudo cell is div and the number of pseudo cells included in the second block is c. calculating an integrated value of the learning load factor based on an expression represented by div + 1 / c;
The step of reflecting the vector data included in the pseudo cell of the pseudo map to the unlearned cell corresponding to the pseudo cell of the unlearned map includes the vector data included in the pseudo cell of the pseudo map in which the learning data has been learned. 8. The method according to claim 2, further comprising: calculating vector data of the unlearned cell reflecting the vector data of the pseudo cell in which the learning data has been learned by dividing by an integrated value of the rate. Classification support map creation method as described in the item.   The step of reflecting the vector data of the pseudo map on the unlearned map is executed after the step of learning the learning data in the pseudo map is executed for all the learning data. Classification support map creation method. Calculating the block data for each of a plurality of third blocks comprising a set of M or more of the unlearned cells extracted from the unlearned map; and the first of the plurality of third blocks. The steps of selecting a block, causing the pseudo cell included in the second block corresponding to the first block to learn learning data, and reflecting the pseudo map vector data in the unlearned map are executed. After
Calculating the block data for each of a plurality of fourth blocks made up of a set of N (N <M) or more unlearned cells extracted from the unlearned map; and Selecting the first block from among the steps, causing the pseudo cell included in the second block corresponding to the first block to learn learning data, and reflecting the pseudo map vector data in the unlearned map The classification support map creation method according to claim 2, wherein the step is executed. A classification support map creation device for creating a classification support map that supports classification of unknown data consisting of vector data,
Unlearned map storage means for storing an unlearned map including a plurality of unlearned cells each having unlearned vector data;
Pseudo map storage means for storing a pseudo map corresponding to the plurality of unlearned cells and including a plurality of pseudo cells having predetermined vector data;
Pseudo map learning means for causing the pseudo map to learn learning data composed of vector data;
A classification support map creating apparatus comprising: an unlearned map reflecting unit that reflects vector data of a learned pseudo cell of the pseudo map to an unlearned cell corresponding to the pseudo cell of the unlearned map. Further comprising a block selection means for selecting a first block composed of a collection of the unlearned cells from the unlearned map;
The pseudo map learning means includes pseudo cell learning means for causing the pseudo cell included in the second block of the pseudo map corresponding to the first block of the unlearned map to learn the learning data. Classification support map creation device of description. The block selection means includes
The vector data of the block is extracted from the unlearned map, for each of a plurality of blocks made up of a collection of the unlearned cells, based on the unlearned vector data of the plurality of unlearned cells included in the block. Block data calculating means for calculating block data as:
Block data comparison and selection means for comparing the learning data with block data of the plurality of blocks and selecting a first block having the block data corresponding to the learning data from the plurality of blocks. Item 13. The classification support map creation device according to Item 12. A program for executing a classification support map creating method for creating a classification support map for supporting classification of unknown data consisting of vector data,
Reading an unlearned map that includes a plurality of unlearned cells each having unlearned vector data;
Reading a pseudo map that corresponds to the plurality of unlearned cells and includes a plurality of pseudo cells having predetermined vector data;
Learning learning data composed of vector data in the pseudo map;
Reflecting a vector data of a learned pseudo cell of the pseudo map to an unlearned cell corresponding to the pseudo cell of the unlearned map, for executing a classification support map creating method.

Images