TWI427546B - A method for generating an initialization scheme of self-organizing map - Google Patents

Description

Method for generating initial architecture of self-organizing map network

The present invention relates to a method for generating an initial architecture of a self-organizing map network, and more particularly to an initial architecture generation method for a self-organizing map network based on a self-organizing map.

At present, neural networks have been widely used to deal with various applications that require a large amount of data operations, and a Self-Organizing Map (SOM) network is a kind of neural network.

Generally, the self-organizing map network first provides a plurality of data points, and then an initial architecture generation method is used to define an initial structure composed of a plurality of initial points, and then the initial architecture of the self-organizing map is used to perform operations. Convergence to get the final self-organizing map network.

The initial architecture generation method of the self-organizing map network is generally selected in a random manner, and several of the data points are selected as initial points, and the initial points are connected to form an initial structure of the self-organizing map.

However, since the foregoing method selects the data points as the initial points in a random manner to form the initial structure, each execution result is different, which affects the correctness of the execution result.

Referring to FIG. 1 , another method for generating an initial architecture of a self-organizing map network is to generate an initial architecture by using a grid initialization scheme (Grid Initialization Scheme). Taking a two-dimensional plane space as an example, a plurality of data points 91 are scattered in a two-dimensional plane space, and the space scattered by the data points 91 is first divided into a plurality of grids 92, so that the data points are completely distributed in the plane. Within the grid 92. Taking FIG. 1 as an example, the two-dimensional space is selected by a square grid 92; and the positions of the four corners of the periphery of the grids 92 are used as initial points 93, so that the initial points 93 can be Together form the initial architecture of the self-organizing map.

Although the initial structure obtained by this law can fully cover these data points 91. However, as shown in Fig. 1, if the distribution of the data points 91 is not uniform, it is likely that a considerable number of initial points 93 are disposed at positions where the data points 91 are not distributed, resulting in a disadvantage of low execution efficiency.

Therefore, referring to FIG. 2, another initial self-organizing map network initial architecture generation method is improved for the foregoing method. It is mainly that among the data points 91, the four data points 91' located at the outermost edge are selected as the four initial points 93' of the most corner of the initial architecture, and then the network surrounded by the four initial points 93' is used. The division of the format to form a plurality of grids 92, and the positions of the corner edges of the grids 92 as the initial point 93, can be formed by the initial points 93 and the initial points 93' located at the corners. The initial architecture of the organization map. Thereby, the correctness of the distribution of the initial point 93 can be improved.

However, since the four initial points 93' are used as the four corners of the grid distribution, as shown in FIG. 2, some of the edge-located data points 91 are not covered by the grids 92. Therefore, the coverage of data point 91 is low, which may affect the correctness of the initial architecture generated.

For the above reasons, it is necessary to further improve the initial architecture generation method of the above-mentioned conventional self-organizing map network.

The primary object of the present invention is to provide an initial architecture generation method for a self-organizing map network to bring the resulting initial architecture closer to the distribution of data points.

The initial architecture generating method of the self-organizing map network according to the present invention includes: step 1 setting a cutting parameter; and step 2, cutting a data set including a plurality of data points according to the cutting parameter to form a plurality of a grid, and the data points are distributed in the grids; in step 3, the grids are sorted according to the number of data points in the grids; and step 4, according to the sorting of the grids, Select a data point as an initial point in the grid of data points; in step 5, repeat step 4, and take the number of initial points equal to the total number of grids; step 6, each adjacent network The center points of the grid are connected to form a grid area; in step 7, the two initial points that are furthest apart are respectively used as the first initial point and the second initial point, and are located diagonally in the sub-grid area. The two center points of the corner respectively serve as a first center point and a second center point to respectively correspond to the first initial point and the second initial point; and in step 8, the sum of the distances from the first initial point and the second initial point The farthest initial point a reference initial point, and a center point located at another corner of the sub-grid area as a reference center point corresponding to the initial point of the reference; and step 9, the first center point is used as a starting center point, along the sub-network The boundary of the grid area is respectively moved to the second center point by two distance points obtained by a distance, and the two center points are connected to obtain a reference line, and the center point passing through the reference line is used as a layer center point. Step 10: using the first initial point as a midpoint, sequentially selecting a plurality of initial points closest to the midpoint as a layer initial point, and the number of initial points of the layer is equal to the number of center points of the layer; Sorting the initial points of the layers according to the distance between the initial points of the layers and the initial point of the reference, and sorting the center points of the layers according to the distance between the center points of the layers and the reference center point, according to the sorting The center points of the layers respectively correspond to the initial points of the layers; in step 12, the center point of the two lines is used as the starting center point, and the center of the initial points of the layers is used as the midpoint, and steps 9 to 11 are repeated until All initial points Corresponding to a center point respectively; and step 13, connecting the initial points according to the connection relationship of the center points to obtain an initial structure of the self-organizing map network.

The above and other objects, features, and advantages of the present invention will become more apparent from the claims. The network format data grouping method of the present invention is to connect at least one database as an execution architecture by using a computer system, wherein the data library has a data set 1 which is composed of a plurality of data points 11 Cluster.

Step 1 of the present invention sets the cutting parameters. More specifically, the cutting parameter refers to the size of the mesh to be cut. For example, the cutting parameter refers to the length of the side of each grid.

Referring to FIG. 4, step 2 of the present invention cuts a data set 1 including a plurality of data points 11 according to the cutting parameters to form a plurality of meshes 2, and the data points 11 are distributed in These are within grid 2. More specifically, the distribution space of the data points 11 is cut according to the cutting parameter, and a plurality of meshes 2 are obtained, so that the data points 11 are completely distributed in the meshes 2, so that A higher coverage rate for these data points can be obtained, thereby improving the accuracy of the execution results. Taking FIG. 4 as an example, the data points 11 of this embodiment are distributed in a two-dimensional plane space. In this embodiment, after the completion of the embodiment, the distribution of the data points 11 is represented by an initial structure of 3×3. Therefore, the cutting parameters are appropriately adjusted, so that the distribution space of the data points forms 9 grids 2 after cutting, and the formation is performed. The arrangement of the nine palaces. To facilitate the subsequent description, the nine grids are defined from left to right, from top to bottom, into grids 2a, 2b, 2c, 2d, 2e, 2f, 2g, 2h, 2i.

Referring to FIG. 4, step 3 of the present invention sorts the grids 2 according to the number of data points 11 in the grids 2. More specifically, in this embodiment, the number of data points 11 in the grid 2 is selected to be descended and sorted, and the sorting results are sequentially into grids 2f, 2d, 2h, 2e, 2a, 2b, grids 2c, 2g, 2i. There is no data point 11 distribution inside.

Referring to FIG. 5, step 4 of the present invention sequentially selects a data point 11 as an initial point 12 in the grid 2 containing the data points 11 according to the order of the grid 2. More specifically, continuing the above example, a data point is sequentially taken out from the meshes 2f, 2d, 2h, 2e, 2a, 2b containing the data points 11 as the initial point 12. At this point, a total of six initial points 12 are taken out.

Referring to FIG. 6, step 5 of the present invention repeats step 4 until the number of the initial points 12 is equal to the total number of the grids 2. More specifically, since the present embodiment wants to represent the distribution of the data points 11 in a 3×3 initial structure, a total of 9 initial points 12 are required, and the foregoing step 4 only obtains 6 initial points 12, and 3 are still missing. The initial point 12, therefore, repeats step 4, and sequentially takes a data point 11 from the grid 2f, 2d, 2h as the initial point 12. The position of the nine initial points 12 can be initially obtained. Therefore, in the present case, by making the grids 2 completely cover the data points 11, the coverage of the data points 11 can be improved; and, by arranging the grids 2 according to the number of data points 11 included. The method, and sequentially selecting the data point 11 as the initial point 12, can make the grid 2 containing more data points 11 preferentially select more initial points 12 to represent the distribution of the data points 11, and avoid setting the initial point 12 to There is no data point 11 in the distribution grid 2, therefore, the execution efficiency can be effectively improved.

So far, the initial position of the nine initial points 12 can be obtained in this embodiment, but the relative relationship of the initial points 12 has not been connected, so another step is performed to link the nine initial points 12 to obtain The initial architecture.

Referring again to FIG. 7, step 6 of the present invention joins the center points 21 of the two adjacent grids 2 to form the primary grid region 3. More specifically, in the present embodiment, by connecting the center points 21 of any two adjacent meshes 2, a sub-grid region 3 composed of nine center points 21 can be obtained to utilize the time. The mesh area 3 serves as a criterion for judging the connection relationship of the initial architecture. The sub-grid area 3 has a plurality of outer boundaries B. For convenience of the following description, the nine center points 21 are left and right after the eighth picture, and the top-down order is sequentially defined as the center points 21a, 21b, 21c, 21d, 21e, 21f, 21g, 21h, 21i. .

Referring to FIG. 8 again, step 7 of the present invention takes the two initial points 12 which are furthest apart as the first initial point 12a and the second initial point 12i, respectively, and is located in the sub-grid area 3 The two center points 21a, 21i of the diagonal corners serve as a first center point 21a and a second center point 21i, respectively, to correspond to the first initial point 12a and the second initial point 12i, respectively. In more detail, in the present embodiment, by calculating the Achilles distance between the initial points 12, the two initial points 12 farthest from each other are selected as the first initial point 12a and the second initial point 12i; In the sub-grid region 3, the diagonal distance is the farthest, so the center point 21 located at the diagonal corner is also selected as the first center point 21a and the second center point 21i, respectively, to respectively correspond to the first initial point 12a and the first Two initial points 12i.

Referring again to FIG. 9, step 8 of the present invention uses the initial point 12 which is the farthest from the sum of the first initial point 12a and the second initial point 12i as the reference initial point 12c, and the secondary grid The center point 21 in the other corner of the area 3 serves as the reference center point 21c corresponding to the reference initial point 12c. More specifically, the sum of the other seven initial points 12 and the Achilles distance of the first initial point 12a and the second initial point 12i is calculated in the same manner, and the initial point 12 farthest from the sum is selected as the reference initial point 12c. Meanwhile, in the sub-grid region 3, the center point 21c farthest from the first center point 21a and the second center point 21i is used as the reference center point 21c, so that the reference initial point 12c is compared with the reference center point 21c. correspond.

Referring again to FIG. 10, step 9 of the present invention uses the first center point 21a as a starting center point 22, and moves to the second center point 21i along the outer boundary B of the sub-grid area 3, respectively. The two center points 21 obtained are separated by a grid, and the two center points 21 are connected to obtain a reference line L, and the center point 21 through which the reference line L passes is used as the layer center point 23. For example, in FIG. 10, the starting center point 22 is respectively moved to the right and the downward direction along the two outer boundaries B to obtain the center points 21b and 21d, and the center points 21b and 21d are connected. The reference line L is obtained after the line. Since the reference line L passes only the center points 21b and 21d, the center points 21b and 21d are used as the layer center point 23. Thus, the center point 21 through which the reference line L passes is the center point 23 of the layer located in the same layer. If the starting center point 22 is regarded as the first layer, the next layer can be found through this step (second The number of layer center points 23 of the layers is two.

Referring to FIG. 10 again, in step 10 of the present invention, the first initial point 12a is used as the midpoint 13, and the plurality of initial points 12 closest to the midpoint 13 are sequentially selected as the layer initial point 14, and the The number of layer initial points 14 is equal to the number of center points 23 of the layer. More specifically, since the number of the layer center points 23 of the next layer in the step 9 is two, in the initial points 12, the first initial point 12a (the first layer) is the same. Starting at point 13, the two initial points 12 closest to the midpoint 13 are calculated as the layer initial point 14 (the second layer) by the Achilles distance calculation, that is, the second layer initial point 14 will correspond to the second layer center. Points 21b, 21d. However, the next step is required to confirm that the layer initial point 14 is the corresponding layer center point 21b, and that layer initial point 14 is the corresponding layer center point 21d.

Referring to FIG. 10 again, step 11 of the present invention sorts the layer initial points 14 according to the distance between the layer initial points 14 and the reference initial point 12c, and according to the layer center points 23 and The distance from the reference center point 21c ranks the layer center points 23, and the layer center points 23 correspond to the layer initial points 14 respectively according to the ordering. More specifically, in view of the two-layer initial point 14, one of the initial points 14 is closer to the reference initial point 12c; and, at the same time, the center point 21b of the layer is compared with 21d. Next, the layer center point 21b is closer to the reference center point 21c, so that the order of the transmission distance is obtained, and the layer initial point 14 closer to the reference initial point 12c is obtained corresponding to the layer center point 21b, and thus the corresponding layer The initial point 12b of the center point 21b; and the other layer initial point 14 corresponds to the layer center point 21d as the initial point 12d. Thus, the center point 21 corresponding to the initial point 12 can be correctly defined by the distance relationship between the reference initial point 12c and the reference center point 21c. So far, the correspondence between the initial point 12 of the second layer and the center point 21 is completed.

Referring to FIG. 11 again, step 12 of the present invention uses the center point 21 of the connected line as the starting center point 22, and the center of the layer initial point 14 as the midpoint 13 is repeated step 9 ~11, until all the initial points 12 correspond to a center point 21, respectively. More specifically, the foregoing steps have completed the correspondence between the initial point 12 of the second layer and the center point 21. Next, the initial point 12 and the center point 21 of the next layer (third layer) need to be clearly defined. Therefore, continuing the above example, the center point 21b, 21d of the connected line is used as the starting center point 22 (the center point of the second layer), and the step 9 is performed to move one space along the two outer boundaries B, respectively. The distance, that is, the two center points 21c and 21g obtained by moving the distance to the right and the downward direction respectively, and connecting the two center points 21c and 21g to form the reference line L, the reference line L passing through the center point 21c, 21e and 21g have three points (the center point of the third layer), so the number of the third layer is three. However, since the reference initial point 12c corresponds to the reference center point 21c, it is not necessary to incorporate the reference initial point 12c into the calculation, and only the two initial points 12 corresponding to the center point 21e, 21g need to be found in the third layer.

Then, the center of the layer initial points 12b, 12d (the initial point of the layer of the second layer) is used as the midpoint 13'. As shown in FIG. 11, the step 10 is performed, and the closest to the midpoint 13' is selected. The two initial points 12 serve as layer initial points 14' (third layer).

Further, step 11 is performed, so that it can be determined that the layer initial point 14' closer to the reference initial point 12c corresponds to the layer center point 21e closer to the reference center point 21c, and is defined as the initial point 12e; Point 14' corresponds to the layer center point 21g and is the initial point 12g. By repeating the steps 9 to 11 at this point, the correspondence between the initial point 12 of the third layer and the center point 21 is further completed.

Since there are still two initial points 12 corresponding to the center point 21, the steps 9 to 11 need to be repeated. Referring to FIG. 12, the center point 21c, 21g of the connected line is used as the starting center point 22 (the center point of the third layer), and the step 9 is performed to move along the two outer boundaries B, respectively. The grid distance, that is, the two center points 21f, 21h obtained by moving one space downward and downward respectively, and connecting the two center points 21f, 21h to form the reference line L, the reference line L passing through the center point 21f There are two points in 21h (the center point of the fourth layer), so the number of initial points 12 of the fourth layer is two.

Then, the center of the layer initial points 12c, 12e, and 12g (the initial point of the layer of the third layer) is used as the midpoint 13". As shown in FIG. 12, the step 10 is performed to select the closest to the midpoint. Two initial points 12 of 13" are used as layer initial points 14" (third layer).

Further, the step 11 is performed, so that it can be determined that the layer initial point 14" closer to the reference initial point 12c corresponds to the layer center point 21f which is closer to the reference center point 21c, and is defined as the initial point 12f; the other layer initial point 14" corresponds to the center point 21h, and is the initial point 12h. So far, the correspondence between the initial point 12 of the fourth layer and the center point 21 is completed. All of the initial points 12 are respectively associated with a center point 21, so the next step can be performed.

Referring to FIG. 13, step 13 of the present invention connects the initial points 12 according to the connection relationship of the central points 21 to obtain an initial structure of the self-organizing map network. More specifically, the initial structure of the self-organizing map network can be obtained by sequentially connecting the initial points 12 according to the central point connection relationship in the sub-grid area 3. For example, the center point 21a is connected to the center points 21b, 21d, and the initial point 12a corresponding to the center point 21a is connected to the initial points 12b, 12d of the corresponding center points 21b, 21d, and so on. The initial connection of the self-organizing map network can be completed by the initial point connection. As can be seen from Fig. 13, the initial architectural distribution produced by the present invention is quite close to the true distribution of the data points 11. Therefore, the subsequent operation of the self-organizing map network can effectively reduce the execution time.

As described above, the present invention passes through the steps 2 to 5, except that all the data points can be covered in the grids 2, and the initials are sequentially selected by sorting the number of the data points 11 contained in the grid 2. At point 12, the initial point 12 can be prevented from being configured at the position where there is no data point 11, and the subsequent execution efficiency is reduced.

In addition, the present invention determines the two initial points 12a, 12i which are farthest apart as the center points 21a, 21i corresponding to the two diagonal corners of the sub-grid area 3, and determines the reference initial point 12c and the corresponding reference center point. 21c, in order to infer the number of layer initial points 14 of each layer, and compare the distance with the reference initial point 12c to confirm the center point 21 corresponding to the layer initial point 14, and according to the connection relationship of the center points 21 The corresponding initial points 12 are connected to generate an initial architecture, so that the initial architecture can initially conform to the true distribution of the data points 11 to improve execution efficiency and reduce subsequent execution time.

While the invention has been described in connection with the preferred embodiments described above, it is not intended to limit the scope of the invention. The technical scope of the invention is protected, and therefore the scope of the invention is defined by the scope of the appended claims.

[this invention]

1. . . Data set

11. . . Data point

12. . . Initial point

12a. . . First initial point

12b. . . Initial point

12c. . . Baseline initial point

12d~12h. . . Initial point

12i. . . Second initial point

13. . . midpoint

13’. . . midpoint

13"... midpoint

14. . . Layer initial point

14’. . . Layer initial point

14"...layer initial point

2. . . grid

2a. . . grid

2b. . . grid

2c. . . grid

2d. . . grid

2e. . . grid

2f. . . grid

2g. . . grid

2h. . . grid

2i. . . grid

twenty one. . . Center point

21a. . . First center point

21b. . . Center point

21c. . . Reference center point

21d~21h. . . Center point

21i. . . Second center point

twenty two. . . Starting center point

twenty three. . . Layer center point

B. . . Outer boundary

L. . . Baseline

[Prior technology]

91. . . Data point

91’. . . Data point

92. . . grid

93. . . Initial point

93’. . . Initial point

Figure 1: Schematic diagram of the initial architecture generation method for the custom self-organizing map network.

Figure 2: Schematic diagram of another initial architecture generation method for a self-organizing map network.

Figure 3: Schematic representation of the data set of the present invention.

Figure 4: Schematic diagram of step 2 of the present invention.

Figure 5: Schematic diagram of step 4 of the present invention.

Figure 6: Schematic diagram of step 5 of the present invention.

Figure 7: Schematic diagram of step 6 of the present invention.

Figure 8: Schematic diagram of step 7 of the present invention.

Figure 9: Schematic diagram of step 8 of the present invention.

Figure 10: Schematic diagram of steps 9-11 of the present invention.

Figure 11: Schematic diagram of step 12 of the present invention.

Figure 12: Schematic diagram of step 12 of the present invention.

Figure 13: Schematic diagram of step 13 of the present invention.

1. . . Data set

11. . . Data point

12a. . . First initial point

12b. . . Initial point

12c. . . Baseline initial point

12d~12h. . . Initial point

12i. . . Second initial point

Claims (3)

An initial architecture generation method for a self-organizing map network, comprising the steps of: step 1: setting a cutting parameter; and step 2, cutting a data set including a plurality of data points according to the cutting parameter to form a plurality of grids And the data points are distributed in the grids; in step 3, the grids are sorted according to the number of data points in the grids; and step 4, according to the sorting of the grids, in the data points Select a data point as an initial point in the grid; in step 5, repeat step 4, and take the number of initial points equal to the total number of grids; in step 6, the two adjacent grids The center points are connected to form a mesh area; in step 7, the two initial points that are furthest apart are respectively used as the first initial point and the second initial point, and in the sub-grid area, in the diagonal corner The two center points respectively serve as a first center point and a second center point to respectively correspond to the first initial point and the second initial point; and in step 8, the farthest distance from the first initial point and the second initial point Initial point as the base An initial point, and a center point located at another corner of the sub-grid area as a reference center point corresponding to the initial point of the reference; and step 9, the first center point is used as a starting center point along the sub-grid The boundary of the region is respectively moved to the second center point by two distance points obtained by a distance, and the two center points are connected to obtain a reference line, and the center point passing through the reference line is used as a layer center point; Step 10: using the first initial point as a midpoint, sequentially selecting a plurality of initial points closest to the midpoint as a layer initial point, and the number of initial points of the layer is equal to the number of central points of the layer; step 11, according to Sorting the initial points of the layers by the distance between the initial points of the layers and the initial points of the layers, and sorting the center points of the layers according to the distance between the center points of the layers and the reference center point, according to the sorting The layer center points respectively correspond to the layer initial points; in step 12, the two line center points are used as the starting center point, and the center of the layer initial points is used as the midpoint, and steps 9 to 11 are repeated until all Initial points are respectively Corresponding to a center point; and step 13, connecting the initial points according to the connection relationship of the center points to obtain an initial structure of the self-organizing map network. The method for generating an initial architecture of the self-organizing map network according to the first aspect of the patent application scope, wherein in the step 3, the ranking is a descending order. The method for generating an initial structure of the self-organizing map network according to the first aspect of the patent application scope, wherein in step 11, the initial points of the layers are decremented according to the distance between the initial points of the layers and the reference initial point Sorting, and descending the center points of the layers according to the distance between the layer center points and the reference center point.